blob: f469e3cd08e82d47ba3a0aa8d9753081e2c7cc99 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Zeisbergerf30c2262006-10-03 23:01:26 +02002 * mm/page-writeback.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
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>
Andrew Morton55e829a2006-12-10 02:19:27 -080024#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/blkdev.h>
26#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070027#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#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>
David Howellscf9a2ae2006-08-29 19:05:54 +010034#include <linux/buffer_head.h>
David Howells811d7362006-08-29 19:06:09 +010035#include <linux/pagevec.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036
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 */
50static long ratelimit_pages = 32;
51
Andrew Mortone236a162006-01-18 17:42:26 -080052static int dirty_exceeded __cacheline_aligned_in_smp; /* Dirty mem may be over limit */
Linus Torvalds1da177e2005-04-16 15:20:36 -070053
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 */
60static 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 */
70int dirty_background_ratio = 10;
71
72/*
73 * The generator of dirty data starts writeback at this percentage
74 */
75int vm_dirty_ratio = 40;
76
77/*
Coywolf Qi Huntfd5403c2006-04-10 22:54:35 -070078 * The interval between `kupdate'-style writebacks, in jiffies
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 */
Bart Samwelf6ef9432006-03-24 03:15:48 -080080int dirty_writeback_interval = 5 * HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -070081
82/*
Coywolf Qi Huntfd5403c2006-04-10 22:54:35 -070083 * The longest number of jiffies for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 */
Bart Samwelf6ef9432006-03-24 03:15:48 -080085int dirty_expire_interval = 30 * HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -070086
87/*
88 * Flag that makes the machine dump writes/reads and block dirtyings.
89 */
90int block_dump;
91
92/*
Bart Samweled5b43f2006-03-24 03:15:49 -080093 * 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070095 */
96int laptop_mode;
97
98EXPORT_SYMBOL(laptop_mode);
99
100/* End of sysctl-exported parameters */
101
102
103static void background_writeout(unsigned long _min_pages);
104
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105/*
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 */
122static void
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700123get_dirty_limits(long *pbackground, long *pdirty,
124 struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125{
126 int background_ratio; /* Percentages */
127 int dirty_ratio;
128 int unmapped_ratio;
129 long background;
130 long dirty;
Chandra Seetharaman40c99aa2006-09-29 02:01:24 -0700131 unsigned long available_memory = vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 struct task_struct *tsk;
133
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134#ifdef CONFIG_HIGHMEM
135 /*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -0800136 * We always exclude high memory from our count.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 */
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -0800138 available_memory -= totalhigh_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139#endif
140
141
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700142 unmapped_ratio = 100 - ((global_page_state(NR_FILE_MAPPED) +
143 global_page_state(NR_ANON_PAGES)) * 100) /
Chandra Seetharaman40c99aa2006-09-29 02:01:24 -0700144 vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
146 dirty_ratio = vm_dirty_ratio;
147 if (dirty_ratio > unmapped_ratio / 2)
148 dirty_ratio = unmapped_ratio / 2;
149
150 if (dirty_ratio < 5)
151 dirty_ratio = 5;
152
153 background_ratio = dirty_background_ratio;
154 if (background_ratio >= dirty_ratio)
155 background_ratio = dirty_ratio / 2;
156
157 background = (background_ratio * available_memory) / 100;
158 dirty = (dirty_ratio * available_memory) / 100;
159 tsk = current;
160 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
161 background += background / 4;
162 dirty += dirty / 4;
163 }
164 *pbackground = background;
165 *pdirty = dirty;
166}
167
168/*
169 * balance_dirty_pages() must be called by processes which are generating dirty
170 * data. It looks at the number of dirty pages in the machine and will force
171 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
172 * If we're over `background_thresh' then pdflush is woken to perform some
173 * writeout.
174 */
175static void balance_dirty_pages(struct address_space *mapping)
176{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 long nr_reclaimable;
178 long background_thresh;
179 long dirty_thresh;
180 unsigned long pages_written = 0;
181 unsigned long write_chunk = sync_writeback_pages();
182
183 struct backing_dev_info *bdi = mapping->backing_dev_info;
184
185 for (;;) {
186 struct writeback_control wbc = {
187 .bdi = bdi,
188 .sync_mode = WB_SYNC_NONE,
189 .older_than_this = NULL,
190 .nr_to_write = write_chunk,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700191 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 };
193
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700194 get_dirty_limits(&background_thresh, &dirty_thresh, mapping);
195 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
196 global_page_state(NR_UNSTABLE_NFS);
197 if (nr_reclaimable + global_page_state(NR_WRITEBACK) <=
198 dirty_thresh)
199 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200
Andrew Mortone236a162006-01-18 17:42:26 -0800201 if (!dirty_exceeded)
202 dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
204 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
205 * Unstable writes are a feature of certain networked
206 * filesystems (i.e. NFS) in which data may have been
207 * written to the server's write cache, but has not yet
208 * been flushed to permanent storage.
209 */
210 if (nr_reclaimable) {
211 writeback_inodes(&wbc);
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700212 get_dirty_limits(&background_thresh,
213 &dirty_thresh, mapping);
214 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
215 global_page_state(NR_UNSTABLE_NFS);
216 if (nr_reclaimable +
217 global_page_state(NR_WRITEBACK)
218 <= dirty_thresh)
219 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 pages_written += write_chunk - wbc.nr_to_write;
221 if (pages_written >= write_chunk)
222 break; /* We've done our duty */
223 }
Andrew Morton3fcfab12006-10-19 23:28:16 -0700224 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 }
226
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700227 if (nr_reclaimable + global_page_state(NR_WRITEBACK)
228 <= dirty_thresh && dirty_exceeded)
229 dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230
231 if (writeback_in_progress(bdi))
232 return; /* pdflush is already working this queue */
233
234 /*
235 * In laptop mode, we wait until hitting the higher threshold before
236 * starting background writeout, and then write out all the way down
237 * to the lower threshold. So slow writers cause minimal disk activity.
238 *
239 * In normal mode, we start background writeout at the lower
240 * background_thresh, to keep the amount of dirty memory low.
241 */
242 if ((laptop_mode && pages_written) ||
243 (!laptop_mode && (nr_reclaimable > background_thresh)))
244 pdflush_operation(background_writeout, 0);
245}
246
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700247void set_page_dirty_balance(struct page *page)
248{
249 if (set_page_dirty(page)) {
250 struct address_space *mapping = page_mapping(page);
251
252 if (mapping)
253 balance_dirty_pages_ratelimited(mapping);
254 }
255}
256
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800258 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -0700259 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +0200260 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 *
262 * Processes which are dirtying memory should call in here once for each page
263 * which was newly dirtied. The function will periodically check the system's
264 * dirty state and will initiate writeback if needed.
265 *
266 * On really big machines, get_writeback_state is expensive, so try to avoid
267 * calling it too often (ratelimiting). But once we're over the dirty memory
268 * limit we decrease the ratelimiting by a lot, to prevent individual processes
269 * from overshooting the limit by (ratelimit_pages) each.
270 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800271void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
272 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273{
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800274 static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
275 unsigned long ratelimit;
276 unsigned long *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277
278 ratelimit = ratelimit_pages;
279 if (dirty_exceeded)
280 ratelimit = 8;
281
282 /*
283 * Check the rate limiting. Also, we do not want to throttle real-time
284 * tasks in balance_dirty_pages(). Period.
285 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800286 preempt_disable();
287 p = &__get_cpu_var(ratelimits);
288 *p += nr_pages_dirtied;
289 if (unlikely(*p >= ratelimit)) {
290 *p = 0;
291 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 balance_dirty_pages(mapping);
293 return;
294 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800295 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296}
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800297EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298
Andrew Morton232ea4d2007-02-28 20:13:21 -0800299void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301 long background_thresh;
302 long dirty_thresh;
303
Andrew Morton232ea4d2007-02-28 20:13:21 -0800304 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) {
305 /*
306 * The caller might hold locks which can prevent IO completion
307 * or progress in the filesystem. So we cannot just sit here
308 * waiting for IO to complete.
309 */
310 congestion_wait(WRITE, HZ/10);
311 return;
312 }
313
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 for ( ; ; ) {
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700315 get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316
317 /*
318 * Boost the allowable dirty threshold a bit for page
319 * allocators so they don't get DoS'ed by heavy writers
320 */
321 dirty_thresh += dirty_thresh / 10; /* wheeee... */
322
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700323 if (global_page_state(NR_UNSTABLE_NFS) +
324 global_page_state(NR_WRITEBACK) <= dirty_thresh)
325 break;
Andrew Morton3fcfab12006-10-19 23:28:16 -0700326 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327 }
328}
329
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330/*
331 * writeback at least _min_pages, and keep writing until the amount of dirty
332 * memory is less than the background threshold, or until we're all clean.
333 */
334static void background_writeout(unsigned long _min_pages)
335{
336 long min_pages = _min_pages;
337 struct writeback_control wbc = {
338 .bdi = NULL,
339 .sync_mode = WB_SYNC_NONE,
340 .older_than_this = NULL,
341 .nr_to_write = 0,
342 .nonblocking = 1,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700343 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 };
345
346 for ( ; ; ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 long background_thresh;
348 long dirty_thresh;
349
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700350 get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
351 if (global_page_state(NR_FILE_DIRTY) +
352 global_page_state(NR_UNSTABLE_NFS) < background_thresh
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353 && min_pages <= 0)
354 break;
355 wbc.encountered_congestion = 0;
356 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
357 wbc.pages_skipped = 0;
358 writeback_inodes(&wbc);
359 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
360 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
361 /* Wrote less than expected */
Andrew Morton3fcfab12006-10-19 23:28:16 -0700362 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 if (!wbc.encountered_congestion)
364 break;
365 }
366 }
367}
368
369/*
370 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
371 * the whole world. Returns 0 if a pdflush thread was dispatched. Returns
372 * -1 if all pdflush threads were busy.
373 */
Pekka J Enberg687a21c2005-06-28 20:44:55 -0700374int wakeup_pdflush(long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375{
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700376 if (nr_pages == 0)
377 nr_pages = global_page_state(NR_FILE_DIRTY) +
378 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 return pdflush_operation(background_writeout, nr_pages);
380}
381
382static void wb_timer_fn(unsigned long unused);
383static void laptop_timer_fn(unsigned long unused);
384
Ingo Molnar8d06afa2005-09-09 13:10:40 -0700385static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
386static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387
388/*
389 * Periodic writeback of "old" data.
390 *
391 * Define "old": the first time one of an inode's pages is dirtied, we mark the
392 * dirtying-time in the inode's address_space. So this periodic writeback code
393 * just walks the superblock inode list, writing back any inodes which are
394 * older than a specific point in time.
395 *
Bart Samwelf6ef9432006-03-24 03:15:48 -0800396 * Try to run once per dirty_writeback_interval. But if a writeback event
397 * takes longer than a dirty_writeback_interval interval, then leave a
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 * one-second gap.
399 *
400 * older_than_this takes precedence over nr_to_write. So we'll only write back
401 * all dirty pages if they are all attached to "old" mappings.
402 */
403static void wb_kupdate(unsigned long arg)
404{
405 unsigned long oldest_jif;
406 unsigned long start_jif;
407 unsigned long next_jif;
408 long nr_to_write;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409 struct writeback_control wbc = {
410 .bdi = NULL,
411 .sync_mode = WB_SYNC_NONE,
412 .older_than_this = &oldest_jif,
413 .nr_to_write = 0,
414 .nonblocking = 1,
415 .for_kupdate = 1,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700416 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 };
418
419 sync_supers();
420
Bart Samwelf6ef9432006-03-24 03:15:48 -0800421 oldest_jif = jiffies - dirty_expire_interval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 start_jif = jiffies;
Bart Samwelf6ef9432006-03-24 03:15:48 -0800423 next_jif = start_jif + dirty_writeback_interval;
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700424 nr_to_write = global_page_state(NR_FILE_DIRTY) +
425 global_page_state(NR_UNSTABLE_NFS) +
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 (inodes_stat.nr_inodes - inodes_stat.nr_unused);
427 while (nr_to_write > 0) {
428 wbc.encountered_congestion = 0;
429 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
430 writeback_inodes(&wbc);
431 if (wbc.nr_to_write > 0) {
432 if (wbc.encountered_congestion)
Andrew Morton3fcfab12006-10-19 23:28:16 -0700433 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 else
435 break; /* All the old data is written */
436 }
437 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
438 }
439 if (time_before(next_jif, jiffies + HZ))
440 next_jif = jiffies + HZ;
Bart Samwelf6ef9432006-03-24 03:15:48 -0800441 if (dirty_writeback_interval)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 mod_timer(&wb_timer, next_jif);
443}
444
445/*
446 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
447 */
448int dirty_writeback_centisecs_handler(ctl_table *table, int write,
449 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
450{
Bart Samwelf6ef9432006-03-24 03:15:48 -0800451 proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
452 if (dirty_writeback_interval) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453 mod_timer(&wb_timer,
Bart Samwelf6ef9432006-03-24 03:15:48 -0800454 jiffies + dirty_writeback_interval);
455 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 del_timer(&wb_timer);
457 }
458 return 0;
459}
460
461static void wb_timer_fn(unsigned long unused)
462{
463 if (pdflush_operation(wb_kupdate, 0) < 0)
464 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
465}
466
467static void laptop_flush(unsigned long unused)
468{
469 sys_sync();
470}
471
472static void laptop_timer_fn(unsigned long unused)
473{
474 pdflush_operation(laptop_flush, 0);
475}
476
477/*
478 * We've spun up the disk and we're in laptop mode: schedule writeback
479 * of all dirty data a few seconds from now. If the flush is already scheduled
480 * then push it back - the user is still using the disk.
481 */
482void laptop_io_completion(void)
483{
Bart Samweled5b43f2006-03-24 03:15:49 -0800484 mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485}
486
487/*
488 * We're in laptop mode and we've just synced. The sync's writes will have
489 * caused another writeback to be scheduled by laptop_io_completion.
490 * Nothing needs to be written back anymore, so we unschedule the writeback.
491 */
492void laptop_sync_completion(void)
493{
494 del_timer(&laptop_mode_wb_timer);
495}
496
497/*
498 * If ratelimit_pages is too high then we can get into dirty-data overload
499 * if a large number of processes all perform writes at the same time.
500 * If it is too low then SMP machines will call the (expensive)
501 * get_writeback_state too often.
502 *
503 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
504 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
505 * thresholds before writeback cuts in.
506 *
507 * But the limit should not be set too high. Because it also controls the
508 * amount of memory which the balance_dirty_pages() caller has to write back.
509 * If this is too large then the caller will block on the IO queue all the
510 * time. So limit it to four megabytes - the balance_dirty_pages() caller
511 * will write six megabyte chunks, max.
512 */
513
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700514void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515{
Chandra Seetharaman40c99aa2006-09-29 02:01:24 -0700516 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 if (ratelimit_pages < 16)
518 ratelimit_pages = 16;
519 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
520 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
521}
522
Chandra Seetharaman26c21432006-06-27 02:54:10 -0700523static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
525{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700526 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -0800527 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528}
529
Chandra Seetharaman74b85f32006-06-27 02:54:09 -0700530static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531 .notifier_call = ratelimit_handler,
532 .next = NULL,
533};
534
535/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -0800536 * Called early on to tune the page writeback dirty limits.
537 *
538 * We used to scale dirty pages according to how total memory
539 * related to pages that could be allocated for buffers (by
540 * comparing nr_free_buffer_pages() to vm_total_pages.
541 *
542 * However, that was when we used "dirty_ratio" to scale with
543 * all memory, and we don't do that any more. "dirty_ratio"
544 * is now applied to total non-HIGHPAGE memory (by subtracting
545 * totalhigh_pages from vm_total_pages), and as such we can't
546 * get into the old insane situation any more where we had
547 * large amounts of dirty pages compared to a small amount of
548 * non-HIGHMEM memory.
549 *
550 * But we might still want to scale the dirty_ratio by how
551 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 */
553void __init page_writeback_init(void)
554{
Bart Samwelf6ef9432006-03-24 03:15:48 -0800555 mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700556 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557 register_cpu_notifier(&ratelimit_nb);
558}
559
David Howells811d7362006-08-29 19:06:09 +0100560/**
Robert P. J. Day72fd4a32007-02-10 01:45:59 -0800561 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
David Howells811d7362006-08-29 19:06:09 +0100562 * @mapping: address space structure to write
563 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
564 *
565 * This is a library function, which implements the writepages()
566 * address_space_operation.
567 *
568 * If a page is already under I/O, generic_writepages() skips it, even
569 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
570 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
571 * and msync() need to guarantee that all the data which was dirty at the time
572 * the call was made get new I/O started against them. If wbc->sync_mode is
573 * WB_SYNC_ALL then we were called for data integrity and we must wait for
574 * existing IO to complete.
575 *
576 * Derived from mpage_writepages() - if you fix this you should check that
577 * also!
578 */
579int generic_writepages(struct address_space *mapping,
580 struct writeback_control *wbc)
581{
582 struct backing_dev_info *bdi = mapping->backing_dev_info;
583 int ret = 0;
584 int done = 0;
585 int (*writepage)(struct page *page, struct writeback_control *wbc);
586 struct pagevec pvec;
587 int nr_pages;
588 pgoff_t index;
589 pgoff_t end; /* Inclusive */
590 int scanned = 0;
591 int range_whole = 0;
592
593 if (wbc->nonblocking && bdi_write_congested(bdi)) {
594 wbc->encountered_congestion = 1;
595 return 0;
596 }
597
598 writepage = mapping->a_ops->writepage;
599
600 /* deal with chardevs and other special file */
601 if (!writepage)
602 return 0;
603
604 pagevec_init(&pvec, 0);
605 if (wbc->range_cyclic) {
606 index = mapping->writeback_index; /* Start from prev offset */
607 end = -1;
608 } else {
609 index = wbc->range_start >> PAGE_CACHE_SHIFT;
610 end = wbc->range_end >> PAGE_CACHE_SHIFT;
611 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
612 range_whole = 1;
613 scanned = 1;
614 }
615retry:
616 while (!done && (index <= end) &&
617 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
618 PAGECACHE_TAG_DIRTY,
619 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
620 unsigned i;
621
622 scanned = 1;
623 for (i = 0; i < nr_pages; i++) {
624 struct page *page = pvec.pages[i];
625
626 /*
627 * At this point we hold neither mapping->tree_lock nor
628 * lock on the page itself: the page may be truncated or
629 * invalidated (changing page->mapping to NULL), or even
630 * swizzled back from swapper_space to tmpfs file
631 * mapping
632 */
633 lock_page(page);
634
635 if (unlikely(page->mapping != mapping)) {
636 unlock_page(page);
637 continue;
638 }
639
640 if (!wbc->range_cyclic && page->index > end) {
641 done = 1;
642 unlock_page(page);
643 continue;
644 }
645
646 if (wbc->sync_mode != WB_SYNC_NONE)
647 wait_on_page_writeback(page);
648
649 if (PageWriteback(page) ||
650 !clear_page_dirty_for_io(page)) {
651 unlock_page(page);
652 continue;
653 }
654
655 ret = (*writepage)(page, wbc);
656 if (ret) {
657 if (ret == -ENOSPC)
658 set_bit(AS_ENOSPC, &mapping->flags);
659 else
660 set_bit(AS_EIO, &mapping->flags);
661 }
662
663 if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE))
664 unlock_page(page);
665 if (ret || (--(wbc->nr_to_write) <= 0))
666 done = 1;
667 if (wbc->nonblocking && bdi_write_congested(bdi)) {
668 wbc->encountered_congestion = 1;
669 done = 1;
670 }
671 }
672 pagevec_release(&pvec);
673 cond_resched();
674 }
675 if (!scanned && !done) {
676 /*
677 * We hit the last page and there is more work to be done: wrap
678 * back to the start of the file
679 */
680 scanned = 1;
681 index = 0;
682 goto retry;
683 }
684 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
685 mapping->writeback_index = index;
686 return ret;
687}
688
689EXPORT_SYMBOL(generic_writepages);
690
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
692{
Andrew Morton22905f72005-11-16 15:07:01 -0800693 int ret;
694
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 if (wbc->nr_to_write <= 0)
696 return 0;
Andrew Morton22905f72005-11-16 15:07:01 -0800697 wbc->for_writepages = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -0700699 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -0800700 else
701 ret = generic_writepages(mapping, wbc);
702 wbc->for_writepages = 0;
703 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704}
705
706/**
707 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -0700708 * @page: the page to write
709 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 *
711 * The page must be locked by the caller and will be unlocked upon return.
712 *
713 * write_one_page() returns a negative error code if I/O failed.
714 */
715int write_one_page(struct page *page, int wait)
716{
717 struct address_space *mapping = page->mapping;
718 int ret = 0;
719 struct writeback_control wbc = {
720 .sync_mode = WB_SYNC_ALL,
721 .nr_to_write = 1,
722 };
723
724 BUG_ON(!PageLocked(page));
725
726 if (wait)
727 wait_on_page_writeback(page);
728
729 if (clear_page_dirty_for_io(page)) {
730 page_cache_get(page);
731 ret = mapping->a_ops->writepage(page, &wbc);
732 if (ret == 0 && wait) {
733 wait_on_page_writeback(page);
734 if (PageError(page))
735 ret = -EIO;
736 }
737 page_cache_release(page);
738 } else {
739 unlock_page(page);
740 }
741 return ret;
742}
743EXPORT_SYMBOL(write_one_page);
744
745/*
Ken Chen76719322007-02-10 01:43:15 -0800746 * For address_spaces which do not use buffers nor write back.
747 */
748int __set_page_dirty_no_writeback(struct page *page)
749{
750 if (!PageDirty(page))
751 SetPageDirty(page);
752 return 0;
753}
754
755/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 * For address_spaces which do not use buffers. Just tag the page as dirty in
757 * its radix tree.
758 *
759 * This is also used when a single buffer is being dirtied: we want to set the
760 * page dirty in that case, but not all the buffers. This is a "bottom-up"
761 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
762 *
763 * Most callers have locked the page, which pins the address_space in memory.
764 * But zap_pte_range() does not lock the page, however in that case the
765 * mapping is pinned by the vma's ->vm_file reference.
766 *
767 * We take care to handle the case where the page was truncated from the
768 * mapping by re-checking page_mapping() insode tree_lock.
769 */
770int __set_page_dirty_nobuffers(struct page *page)
771{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 if (!TestSetPageDirty(page)) {
773 struct address_space *mapping = page_mapping(page);
774 struct address_space *mapping2;
775
Andrew Morton8c085402006-12-10 02:19:24 -0800776 if (!mapping)
777 return 1;
778
779 write_lock_irq(&mapping->tree_lock);
780 mapping2 = page_mapping(page);
781 if (mapping2) { /* Race with truncate? */
782 BUG_ON(mapping2 != mapping);
Andrew Morton55e829a2006-12-10 02:19:27 -0800783 if (mapping_cap_account_dirty(mapping)) {
Andrew Morton8c085402006-12-10 02:19:24 -0800784 __inc_zone_page_state(page, NR_FILE_DIRTY);
Andrew Morton55e829a2006-12-10 02:19:27 -0800785 task_io_account_write(PAGE_CACHE_SIZE);
786 }
Andrew Morton8c085402006-12-10 02:19:24 -0800787 radix_tree_tag_set(&mapping->page_tree,
788 page_index(page), PAGECACHE_TAG_DIRTY);
789 }
790 write_unlock_irq(&mapping->tree_lock);
791 if (mapping->host) {
792 /* !PageAnon && !swapper_space */
793 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 }
Andrew Morton4741c9f2006-03-24 03:18:11 -0800795 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 }
Andrew Morton4741c9f2006-03-24 03:18:11 -0800797 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798}
799EXPORT_SYMBOL(__set_page_dirty_nobuffers);
800
801/*
802 * When a writepage implementation decides that it doesn't want to write this
803 * page for some reason, it should redirty the locked page via
804 * redirty_page_for_writepage() and it should then unlock the page and return 0
805 */
806int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
807{
808 wbc->pages_skipped++;
809 return __set_page_dirty_nobuffers(page);
810}
811EXPORT_SYMBOL(redirty_page_for_writepage);
812
813/*
814 * If the mapping doesn't provide a set_page_dirty a_op, then
815 * just fall through and assume that it wants buffer_heads.
816 */
817int fastcall set_page_dirty(struct page *page)
818{
819 struct address_space *mapping = page_mapping(page);
820
821 if (likely(mapping)) {
822 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
David Howells93614012006-09-30 20:45:40 +0200823#ifdef CONFIG_BLOCK
824 if (!spd)
825 spd = __set_page_dirty_buffers;
826#endif
827 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 }
Andrew Morton4741c9f2006-03-24 03:18:11 -0800829 if (!PageDirty(page)) {
830 if (!TestSetPageDirty(page))
831 return 1;
832 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 return 0;
834}
835EXPORT_SYMBOL(set_page_dirty);
836
837/*
838 * set_page_dirty() is racy if the caller has no reference against
839 * page->mapping->host, and if the page is unlocked. This is because another
840 * CPU could truncate the page off the mapping and then free the mapping.
841 *
842 * Usually, the page _is_ locked, or the caller is a user-space process which
843 * holds a reference on the inode by having an open file.
844 *
845 * In other cases, the page should be locked before running set_page_dirty().
846 */
847int set_page_dirty_lock(struct page *page)
848{
849 int ret;
850
Nick Piggindb376482006-09-25 23:31:24 -0700851 lock_page_nosync(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 ret = set_page_dirty(page);
853 unlock_page(page);
854 return ret;
855}
856EXPORT_SYMBOL(set_page_dirty_lock);
857
858/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 * Clear a page's dirty flag, while caring for dirty memory accounting.
860 * Returns true if the page was previously dirty.
861 *
862 * This is for preparing to put the page under writeout. We leave the page
863 * tagged as dirty in the radix tree so that a concurrent write-for-sync
864 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
865 * implementation will run either set_page_writeback() or set_page_dirty(),
866 * at which stage we bring the page's dirty flag and radix-tree dirty tag
867 * back into sync.
868 *
869 * This incoherency between the page's dirty flag and radix-tree tag is
870 * unfortunate, but it only exists while the page is locked.
871 */
872int clear_page_dirty_for_io(struct page *page)
873{
874 struct address_space *mapping = page_mapping(page);
875
Linus Torvalds7658cc22006-12-29 10:00:58 -0800876 if (mapping && mapping_cap_account_dirty(mapping)) {
877 /*
878 * Yes, Virginia, this is indeed insane.
879 *
880 * We use this sequence to make sure that
881 * (a) we account for dirty stats properly
882 * (b) we tell the low-level filesystem to
883 * mark the whole page dirty if it was
884 * dirty in a pagetable. Only to then
885 * (c) clean the page again and return 1 to
886 * cause the writeback.
887 *
888 * This way we avoid all nasty races with the
889 * dirty bit in multiple places and clearing
890 * them concurrently from different threads.
891 *
892 * Note! Normally the "set_page_dirty(page)"
893 * has no effect on the actual dirty bit - since
894 * that will already usually be set. But we
895 * need the side effects, and it can help us
896 * avoid races.
897 *
898 * We basically use the page "master dirty bit"
899 * as a serialization point for all the different
900 * threads doing their things.
901 *
902 * FIXME! We still have a race here: if somebody
903 * adds the page back to the page tables in
904 * between the "page_mkclean()" and the "TestClearPageDirty()",
905 * we might have it mapped without the dirty bit set.
906 */
907 if (page_mkclean(page))
908 set_page_dirty(page);
909 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -0800910 dec_zone_page_state(page, NR_FILE_DIRTY);
Linus Torvalds7658cc22006-12-29 10:00:58 -0800911 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 }
Linus Torvalds7658cc22006-12-29 10:00:58 -0800913 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 }
Linus Torvalds7658cc22006-12-29 10:00:58 -0800915 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916}
Hans Reiser58bb01a2005-11-18 01:10:53 -0800917EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
919int test_clear_page_writeback(struct page *page)
920{
921 struct address_space *mapping = page_mapping(page);
922 int ret;
923
924 if (mapping) {
925 unsigned long flags;
926
927 write_lock_irqsave(&mapping->tree_lock, flags);
928 ret = TestClearPageWriteback(page);
929 if (ret)
930 radix_tree_tag_clear(&mapping->page_tree,
931 page_index(page),
932 PAGECACHE_TAG_WRITEBACK);
933 write_unlock_irqrestore(&mapping->tree_lock, flags);
934 } else {
935 ret = TestClearPageWriteback(page);
936 }
937 return ret;
938}
939
940int test_set_page_writeback(struct page *page)
941{
942 struct address_space *mapping = page_mapping(page);
943 int ret;
944
945 if (mapping) {
946 unsigned long flags;
947
948 write_lock_irqsave(&mapping->tree_lock, flags);
949 ret = TestSetPageWriteback(page);
950 if (!ret)
951 radix_tree_tag_set(&mapping->page_tree,
952 page_index(page),
953 PAGECACHE_TAG_WRITEBACK);
954 if (!PageDirty(page))
955 radix_tree_tag_clear(&mapping->page_tree,
956 page_index(page),
957 PAGECACHE_TAG_DIRTY);
958 write_unlock_irqrestore(&mapping->tree_lock, flags);
959 } else {
960 ret = TestSetPageWriteback(page);
961 }
962 return ret;
963
964}
965EXPORT_SYMBOL(test_set_page_writeback);
966
967/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 * Return true if any of the pages in the mapping are marged with the
969 * passed tag.
970 */
971int mapping_tagged(struct address_space *mapping, int tag)
972{
973 unsigned long flags;
974 int ret;
975
976 read_lock_irqsave(&mapping->tree_lock, flags);
977 ret = radix_tree_tagged(&mapping->page_tree, tag);
978 read_unlock_irqrestore(&mapping->tree_lock, flags);
979 return ret;
980}
981EXPORT_SYMBOL(mapping_tagged);