blob: d1ba464410536570c899d6b5893c2fe5685e0afa [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.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07005 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
Francois Camie1f8e872008-10-15 22:01:59 -070010 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Initial version
12 */
13
14#include <linux/kernel.h>
15#include <linux/module.h>
16#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080025#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/blkdev.h>
27#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070028#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
David Howellscf9a2ae2006-08-29 19:05:54 +010035#include <linux/buffer_head.h>
David Howells811d7362006-08-29 19:06:09 +010036#include <linux/pagevec.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
38/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070039 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
40 * will look to see if it needs to force writeback or throttling.
41 */
42static long ratelimit_pages = 32;
43
Linus Torvalds1da177e2005-04-16 15:20:36 -070044/*
45 * When balance_dirty_pages decides that the caller needs to perform some
46 * non-background writeback, this is how many pages it will attempt to write.
47 * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
48 * large amounts of I/O are submitted.
49 */
50static inline long sync_writeback_pages(void)
51{
52 return ratelimit_pages + ratelimit_pages / 2;
53}
54
55/* The following parameters are exported via /proc/sys/vm */
56
57/*
58 * Start background writeback (via pdflush) at this percentage
59 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080060int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
62/*
David Rientjes2da02992009-01-06 14:39:31 -080063 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
64 * dirty_background_ratio * the amount of dirtyable memory
65 */
66unsigned long dirty_background_bytes;
67
68/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080069 * free highmem will not be subtracted from the total free memory
70 * for calculating free ratios if vm_highmem_is_dirtyable is true
71 */
72int vm_highmem_is_dirtyable;
73
74/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070075 * The generator of dirty data starts writeback at this percentage
76 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080077int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
79/*
David Rientjes2da02992009-01-06 14:39:31 -080080 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
81 * vm_dirty_ratio * the amount of dirtyable memory
82 */
83unsigned long vm_dirty_bytes;
84
85/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070086 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -070088unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
90/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070091 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -070092 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -070093unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
95/*
96 * Flag that makes the machine dump writes/reads and block dirtyings.
97 */
98int block_dump;
99
100/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800101 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
102 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 */
104int laptop_mode;
105
106EXPORT_SYMBOL(laptop_mode);
107
108/* End of sysctl-exported parameters */
109
110
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700112 * Scale the writeback cache size proportional to the relative writeout speeds.
113 *
114 * We do this by keeping a floating proportion between BDIs, based on page
115 * writeback completions [end_page_writeback()]. Those devices that write out
116 * pages fastest will get the larger share, while the slower will get a smaller
117 * share.
118 *
119 * We use page writeout completions because we are interested in getting rid of
120 * dirty pages. Having them written out is the primary goal.
121 *
122 * We introduce a concept of time, a period over which we measure these events,
123 * because demand can/will vary over time. The length of this period itself is
124 * measured in page writeback completions.
125 *
126 */
127static struct prop_descriptor vm_completions;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700128static struct prop_descriptor vm_dirties;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700129
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700130/*
131 * couple the period to the dirty_ratio:
132 *
133 * period/2 ~ roundup_pow_of_two(dirty limit)
134 */
135static int calc_period_shift(void)
136{
137 unsigned long dirty_total;
138
David Rientjes2da02992009-01-06 14:39:31 -0800139 if (vm_dirty_bytes)
140 dirty_total = vm_dirty_bytes / PAGE_SIZE;
141 else
142 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
143 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700144 return 2 + ilog2(dirty_total - 1);
145}
146
147/*
David Rientjes2da02992009-01-06 14:39:31 -0800148 * update the period when the dirty threshold changes.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700149 */
David Rientjes2da02992009-01-06 14:39:31 -0800150static void update_completion_period(void)
151{
152 int shift = calc_period_shift();
153 prop_change_shift(&vm_completions, shift);
154 prop_change_shift(&vm_dirties, shift);
155}
156
157int dirty_background_ratio_handler(struct ctl_table *table, int write,
158 struct file *filp, void __user *buffer, size_t *lenp,
159 loff_t *ppos)
160{
161 int ret;
162
163 ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
164 if (ret == 0 && write)
165 dirty_background_bytes = 0;
166 return ret;
167}
168
169int dirty_background_bytes_handler(struct ctl_table *table, int write,
170 struct file *filp, void __user *buffer, size_t *lenp,
171 loff_t *ppos)
172{
173 int ret;
174
175 ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
176 if (ret == 0 && write)
177 dirty_background_ratio = 0;
178 return ret;
179}
180
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700181int dirty_ratio_handler(struct ctl_table *table, int write,
182 struct file *filp, void __user *buffer, size_t *lenp,
183 loff_t *ppos)
184{
185 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800186 int ret;
187
188 ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700189 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
David Rientjes2da02992009-01-06 14:39:31 -0800190 update_completion_period();
191 vm_dirty_bytes = 0;
192 }
193 return ret;
194}
195
196
197int dirty_bytes_handler(struct ctl_table *table, int write,
198 struct file *filp, void __user *buffer, size_t *lenp,
199 loff_t *ppos)
200{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800201 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800202 int ret;
203
204 ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
205 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
206 update_completion_period();
207 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700208 }
209 return ret;
210}
211
212/*
213 * Increment the BDI's writeout completion count and the global writeout
214 * completion count. Called from test_clear_page_writeback().
215 */
216static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
217{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700218 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
219 bdi->max_prop_frac);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700220}
221
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700222void bdi_writeout_inc(struct backing_dev_info *bdi)
223{
224 unsigned long flags;
225
226 local_irq_save(flags);
227 __bdi_writeout_inc(bdi);
228 local_irq_restore(flags);
229}
230EXPORT_SYMBOL_GPL(bdi_writeout_inc);
231
Nick Piggin1cf6e7d2009-02-18 14:48:18 -0800232void task_dirty_inc(struct task_struct *tsk)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700233{
234 prop_inc_single(&vm_dirties, &tsk->dirties);
235}
236
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700237/*
238 * Obtain an accurate fraction of the BDI's portion.
239 */
240static void bdi_writeout_fraction(struct backing_dev_info *bdi,
241 long *numerator, long *denominator)
242{
243 if (bdi_cap_writeback_dirty(bdi)) {
244 prop_fraction_percpu(&vm_completions, &bdi->completions,
245 numerator, denominator);
246 } else {
247 *numerator = 0;
248 *denominator = 1;
249 }
250}
251
252/*
253 * Clip the earned share of dirty pages to that which is actually available.
254 * This avoids exceeding the total dirty_limit when the floating averages
255 * fluctuate too quickly.
256 */
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700257static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
258 unsigned long dirty, unsigned long *pbdi_dirty)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700259{
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700260 unsigned long avail_dirty;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700261
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700262 avail_dirty = global_page_state(NR_FILE_DIRTY) +
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700263 global_page_state(NR_WRITEBACK) +
Miklos Szeredifc3ba692008-04-30 00:54:38 -0700264 global_page_state(NR_UNSTABLE_NFS) +
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700265 global_page_state(NR_WRITEBACK_TEMP);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700266
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700267 if (avail_dirty < dirty)
268 avail_dirty = dirty - avail_dirty;
269 else
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700270 avail_dirty = 0;
271
272 avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
273 bdi_stat(bdi, BDI_WRITEBACK);
274
275 *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
276}
277
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700278static inline void task_dirties_fraction(struct task_struct *tsk,
279 long *numerator, long *denominator)
280{
281 prop_fraction_single(&vm_dirties, &tsk->dirties,
282 numerator, denominator);
283}
284
285/*
286 * scale the dirty limit
287 *
288 * task specific dirty limit:
289 *
290 * dirty -= (dirty/8) * p_{t}
291 */
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700292static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700293{
294 long numerator, denominator;
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700295 unsigned long dirty = *pdirty;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700296 u64 inv = dirty >> 3;
297
298 task_dirties_fraction(tsk, &numerator, &denominator);
299 inv *= numerator;
300 do_div(inv, denominator);
301
302 dirty -= inv;
303 if (dirty < *pdirty/2)
304 dirty = *pdirty/2;
305
306 *pdirty = dirty;
307}
308
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700309/*
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700310 *
311 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700312static unsigned int bdi_min_ratio;
313
314int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
315{
316 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700317
Jens Axboecfc4ba52009-09-14 13:12:40 +0200318 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700319 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700320 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700321 } else {
322 min_ratio -= bdi->min_ratio;
323 if (bdi_min_ratio + min_ratio < 100) {
324 bdi_min_ratio += min_ratio;
325 bdi->min_ratio += min_ratio;
326 } else {
327 ret = -EINVAL;
328 }
329 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200330 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700331
332 return ret;
333}
334
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700335int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
336{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700337 int ret = 0;
338
339 if (max_ratio > 100)
340 return -EINVAL;
341
Jens Axboecfc4ba52009-09-14 13:12:40 +0200342 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700343 if (bdi->min_ratio > max_ratio) {
344 ret = -EINVAL;
345 } else {
346 bdi->max_ratio = max_ratio;
347 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
348 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200349 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700350
351 return ret;
352}
353EXPORT_SYMBOL(bdi_set_max_ratio);
354
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700355/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356 * Work out the current dirty-memory clamping and background writeout
357 * thresholds.
358 *
359 * The main aim here is to lower them aggressively if there is a lot of mapped
360 * memory around. To avoid stressing page reclaim with lots of unreclaimable
361 * pages. It is better to clamp down on writers than to start swapping, and
362 * performing lots of scanning.
363 *
364 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
365 *
366 * We don't permit the clamping level to fall below 5% - that is getting rather
367 * excessive.
368 *
369 * We make sure that the background writeout level is below the adjusted
370 * clamping level.
371 */
Christoph Lameter1b424462007-05-06 14:48:59 -0700372
373static unsigned long highmem_dirtyable_memory(unsigned long total)
374{
375#ifdef CONFIG_HIGHMEM
376 int node;
377 unsigned long x = 0;
378
Lee Schermerhorn37b07e42007-10-16 01:25:39 -0700379 for_each_node_state(node, N_HIGH_MEMORY) {
Christoph Lameter1b424462007-05-06 14:48:59 -0700380 struct zone *z =
381 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
382
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700383 x += zone_page_state(z, NR_FREE_PAGES) + zone_lru_pages(z);
Christoph Lameter1b424462007-05-06 14:48:59 -0700384 }
385 /*
386 * Make sure that the number of highmem pages is never larger
387 * than the number of the total dirtyable memory. This can only
388 * occur in very strange VM situations but we want to make sure
389 * that this does not occur.
390 */
391 return min(x, total);
392#else
393 return 0;
394#endif
395}
396
Steven Rostedt3eefae92008-05-12 21:21:04 +0200397/**
398 * determine_dirtyable_memory - amount of memory that may be used
399 *
400 * Returns the numebr of pages that can currently be freed and used
401 * by the kernel for direct mappings.
402 */
403unsigned long determine_dirtyable_memory(void)
Christoph Lameter1b424462007-05-06 14:48:59 -0700404{
405 unsigned long x;
406
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700407 x = global_page_state(NR_FREE_PAGES) + global_lru_pages();
Bron Gondwana195cf4532008-02-04 22:29:20 -0800408
409 if (!vm_highmem_is_dirtyable)
410 x -= highmem_dirtyable_memory(x);
411
Christoph Lameter1b424462007-05-06 14:48:59 -0700412 return x + 1; /* Ensure that we never return 0 */
413}
414
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700415void
David Rientjes364aeb22009-01-06 14:39:29 -0800416get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
417 unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418{
David Rientjes364aeb22009-01-06 14:39:29 -0800419 unsigned long background;
420 unsigned long dirty;
Christoph Lameter1b424462007-05-06 14:48:59 -0700421 unsigned long available_memory = determine_dirtyable_memory();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 struct task_struct *tsk;
423
David Rientjes2da02992009-01-06 14:39:31 -0800424 if (vm_dirty_bytes)
425 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
426 else {
427 int dirty_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428
David Rientjes2da02992009-01-06 14:39:31 -0800429 dirty_ratio = vm_dirty_ratio;
430 if (dirty_ratio < 5)
431 dirty_ratio = 5;
432 dirty = (dirty_ratio * available_memory) / 100;
433 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434
David Rientjes2da02992009-01-06 14:39:31 -0800435 if (dirty_background_bytes)
436 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
437 else
438 background = (dirty_background_ratio * available_memory) / 100;
439
440 if (background >= dirty)
441 background = dirty / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 tsk = current;
443 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
444 background += background / 4;
445 dirty += dirty / 4;
446 }
447 *pbackground = background;
448 *pdirty = dirty;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700449
450 if (bdi) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700451 u64 bdi_dirty;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700452 long numerator, denominator;
453
454 /*
455 * Calculate this BDI's share of the dirty ratio.
456 */
457 bdi_writeout_fraction(bdi, &numerator, &denominator);
458
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700459 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700460 bdi_dirty *= numerator;
461 do_div(bdi_dirty, denominator);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700462 bdi_dirty += (dirty * bdi->min_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700463 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
464 bdi_dirty = dirty * bdi->max_ratio / 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700465
466 *pbdi_dirty = bdi_dirty;
467 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700468 task_dirty_limit(current, pbdi_dirty);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700469 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470}
471
472/*
473 * balance_dirty_pages() must be called by processes which are generating dirty
474 * data. It looks at the number of dirty pages in the machine and will force
475 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
476 * If we're over `background_thresh' then pdflush is woken to perform some
477 * writeout.
478 */
479static void balance_dirty_pages(struct address_space *mapping)
480{
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800481 long nr_reclaimable, bdi_nr_reclaimable;
482 long nr_writeback, bdi_nr_writeback;
David Rientjes364aeb22009-01-06 14:39:29 -0800483 unsigned long background_thresh;
484 unsigned long dirty_thresh;
485 unsigned long bdi_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486 unsigned long pages_written = 0;
487 unsigned long write_chunk = sync_writeback_pages();
Jens Axboe87c6a9b2009-09-17 19:59:14 +0200488 unsigned long pause = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489
490 struct backing_dev_info *bdi = mapping->backing_dev_info;
491
492 for (;;) {
493 struct writeback_control wbc = {
494 .bdi = bdi,
495 .sync_mode = WB_SYNC_NONE,
496 .older_than_this = NULL,
497 .nr_to_write = write_chunk,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700498 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 };
500
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700501 get_dirty_limits(&background_thresh, &dirty_thresh,
502 &bdi_thresh, bdi);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800503
504 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
505 global_page_state(NR_UNSTABLE_NFS);
506 nr_writeback = global_page_state(NR_WRITEBACK);
507
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700508 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
509 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800510
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700511 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
512 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800514 /*
515 * Throttle it only when the background writeback cannot
516 * catch-up. This avoids (excessively) small writeouts
517 * when the bdi limits are ramping up.
518 */
519 if (nr_reclaimable + nr_writeback <
520 (background_thresh + dirty_thresh) / 2)
521 break;
522
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700523 if (!bdi->dirty_exceeded)
524 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525
526 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
527 * Unstable writes are a feature of certain networked
528 * filesystems (i.e. NFS) in which data may have been
529 * written to the server's write cache, but has not yet
530 * been flushed to permanent storage.
Richard Kennedyd7831a02009-06-30 11:41:35 -0700531 * Only move pages to writeback if this bdi is over its
532 * threshold otherwise wait until the disk writes catch
533 * up.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 */
Richard Kennedyd7831a02009-06-30 11:41:35 -0700535 if (bdi_nr_reclaimable > bdi_thresh) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200536 writeback_inodes_wbc(&wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 pages_written += write_chunk - wbc.nr_to_write;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700538 get_dirty_limits(&background_thresh, &dirty_thresh,
539 &bdi_thresh, bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700541
542 /*
543 * In order to avoid the stacked BDI deadlock we need
544 * to ensure we accurately count the 'dirty' pages when
545 * the threshold is low.
546 *
547 * Otherwise it would be possible to get thresh+n pages
548 * reported dirty, even though there are thresh-m pages
549 * actually dirty; with m+n sitting in the percpu
550 * deltas.
551 */
552 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
553 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
554 bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
555 } else if (bdi_nr_reclaimable) {
556 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
557 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
558 }
559
560 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
561 break;
562 if (pages_written >= write_chunk)
563 break; /* We've done our duty */
564
Jens Axboe87c6a9b2009-09-17 19:59:14 +0200565 schedule_timeout_interruptible(pause);
566
567 /*
568 * Increase the delay for each loop, up to our previous
569 * default of taking a 100ms nap.
570 */
571 pause <<= 1;
572 if (pause > HZ / 10)
573 pause = HZ / 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 }
575
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700576 if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
577 bdi->dirty_exceeded)
578 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579
580 if (writeback_in_progress(bdi))
581 return; /* pdflush is already working this queue */
582
583 /*
584 * In laptop mode, we wait until hitting the higher threshold before
585 * starting background writeout, and then write out all the way down
586 * to the lower threshold. So slow writers cause minimal disk activity.
587 *
588 * In normal mode, we start background writeout at the lower
589 * background_thresh, to keep the amount of dirty memory low.
590 */
591 if ((laptop_mode && pages_written) ||
Jens Axboe03ba3782009-09-09 09:08:54 +0200592 (!laptop_mode && ((nr_writeback = global_page_state(NR_FILE_DIRTY)
593 + global_page_state(NR_UNSTABLE_NFS))
Jens Axboeb6e51312009-09-16 15:13:54 +0200594 > background_thresh)))
595 bdi_start_writeback(bdi, nr_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700596}
597
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200598void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700599{
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200600 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700601 struct address_space *mapping = page_mapping(page);
602
603 if (mapping)
604 balance_dirty_pages_ratelimited(mapping);
605 }
606}
607
Tejun Heo245b2e72009-06-24 15:13:48 +0900608static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
609
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800611 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -0700612 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +0200613 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 *
615 * Processes which are dirtying memory should call in here once for each page
616 * which was newly dirtied. The function will periodically check the system's
617 * dirty state and will initiate writeback if needed.
618 *
619 * On really big machines, get_writeback_state is expensive, so try to avoid
620 * calling it too often (ratelimiting). But once we're over the dirty memory
621 * limit we decrease the ratelimiting by a lot, to prevent individual processes
622 * from overshooting the limit by (ratelimit_pages) each.
623 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800624void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
625 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626{
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800627 unsigned long ratelimit;
628 unsigned long *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
630 ratelimit = ratelimit_pages;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700631 if (mapping->backing_dev_info->dirty_exceeded)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632 ratelimit = 8;
633
634 /*
635 * Check the rate limiting. Also, we do not want to throttle real-time
636 * tasks in balance_dirty_pages(). Period.
637 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800638 preempt_disable();
Tejun Heo245b2e72009-06-24 15:13:48 +0900639 p = &__get_cpu_var(bdp_ratelimits);
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800640 *p += nr_pages_dirtied;
641 if (unlikely(*p >= ratelimit)) {
642 *p = 0;
643 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 balance_dirty_pages(mapping);
645 return;
646 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800647 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648}
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800649EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650
Andrew Morton232ea4d2007-02-28 20:13:21 -0800651void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652{
David Rientjes364aeb22009-01-06 14:39:29 -0800653 unsigned long background_thresh;
654 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655
656 for ( ; ; ) {
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700657 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658
659 /*
660 * Boost the allowable dirty threshold a bit for page
661 * allocators so they don't get DoS'ed by heavy writers
662 */
663 dirty_thresh += dirty_thresh / 10; /* wheeee... */
664
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700665 if (global_page_state(NR_UNSTABLE_NFS) +
666 global_page_state(NR_WRITEBACK) <= dirty_thresh)
667 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +0200668 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -0700669
670 /*
671 * The caller might hold locks which can prevent IO completion
672 * or progress in the filesystem. So we cannot just sit here
673 * waiting for IO to complete.
674 */
675 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
676 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 }
678}
679
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680static void laptop_timer_fn(unsigned long unused);
681
Ingo Molnar8d06afa2005-09-09 13:10:40 -0700682static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
684/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
686 */
687int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Andrew Morton3e733f02007-07-15 23:41:05 -0700688 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689{
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700690 proc_dointvec(table, write, file, buffer, length, ppos);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 return 0;
692}
693
Jens Axboe03ba3782009-09-09 09:08:54 +0200694static void do_laptop_sync(struct work_struct *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695{
Jens Axboe03ba3782009-09-09 09:08:54 +0200696 wakeup_flusher_threads(0);
697 kfree(work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698}
699
700static void laptop_timer_fn(unsigned long unused)
701{
Jens Axboe03ba3782009-09-09 09:08:54 +0200702 struct work_struct *work;
703
704 work = kmalloc(sizeof(*work), GFP_ATOMIC);
705 if (work) {
706 INIT_WORK(work, do_laptop_sync);
707 schedule_work(work);
708 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709}
710
711/*
712 * We've spun up the disk and we're in laptop mode: schedule writeback
713 * of all dirty data a few seconds from now. If the flush is already scheduled
714 * then push it back - the user is still using the disk.
715 */
716void laptop_io_completion(void)
717{
Bart Samweled5b43f2006-03-24 03:15:49 -0800718 mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719}
720
721/*
722 * We're in laptop mode and we've just synced. The sync's writes will have
723 * caused another writeback to be scheduled by laptop_io_completion.
724 * Nothing needs to be written back anymore, so we unschedule the writeback.
725 */
726void laptop_sync_completion(void)
727{
728 del_timer(&laptop_mode_wb_timer);
729}
730
731/*
732 * If ratelimit_pages is too high then we can get into dirty-data overload
733 * if a large number of processes all perform writes at the same time.
734 * If it is too low then SMP machines will call the (expensive)
735 * get_writeback_state too often.
736 *
737 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
738 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
739 * thresholds before writeback cuts in.
740 *
741 * But the limit should not be set too high. Because it also controls the
742 * amount of memory which the balance_dirty_pages() caller has to write back.
743 * If this is too large then the caller will block on the IO queue all the
744 * time. So limit it to four megabytes - the balance_dirty_pages() caller
745 * will write six megabyte chunks, max.
746 */
747
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700748void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749{
Chandra Seetharaman40c99aa2006-09-29 02:01:24 -0700750 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 if (ratelimit_pages < 16)
752 ratelimit_pages = 16;
753 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
754 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
755}
756
Chandra Seetharaman26c21432006-06-27 02:54:10 -0700757static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
759{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700760 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -0800761 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762}
763
Chandra Seetharaman74b85f32006-06-27 02:54:09 -0700764static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 .notifier_call = ratelimit_handler,
766 .next = NULL,
767};
768
769/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -0800770 * Called early on to tune the page writeback dirty limits.
771 *
772 * We used to scale dirty pages according to how total memory
773 * related to pages that could be allocated for buffers (by
774 * comparing nr_free_buffer_pages() to vm_total_pages.
775 *
776 * However, that was when we used "dirty_ratio" to scale with
777 * all memory, and we don't do that any more. "dirty_ratio"
778 * is now applied to total non-HIGHPAGE memory (by subtracting
779 * totalhigh_pages from vm_total_pages), and as such we can't
780 * get into the old insane situation any more where we had
781 * large amounts of dirty pages compared to a small amount of
782 * non-HIGHMEM memory.
783 *
784 * But we might still want to scale the dirty_ratio by how
785 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 */
787void __init page_writeback_init(void)
788{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700789 int shift;
790
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700791 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700793
794 shift = calc_period_shift();
795 prop_descriptor_init(&vm_completions, shift);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700796 prop_descriptor_init(&vm_dirties, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797}
798
David Howells811d7362006-08-29 19:06:09 +0100799/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700800 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
David Howells811d7362006-08-29 19:06:09 +0100801 * @mapping: address space structure to write
802 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700803 * @writepage: function called for each page
804 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +0100805 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700806 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +0100807 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
808 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
809 * and msync() need to guarantee that all the data which was dirty at the time
810 * the call was made get new I/O started against them. If wbc->sync_mode is
811 * WB_SYNC_ALL then we were called for data integrity and we must wait for
812 * existing IO to complete.
David Howells811d7362006-08-29 19:06:09 +0100813 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700814int write_cache_pages(struct address_space *mapping,
815 struct writeback_control *wbc, writepage_t writepage,
816 void *data)
David Howells811d7362006-08-29 19:06:09 +0100817{
818 struct backing_dev_info *bdi = mapping->backing_dev_info;
819 int ret = 0;
820 int done = 0;
David Howells811d7362006-08-29 19:06:09 +0100821 struct pagevec pvec;
822 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -0800823 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +0100824 pgoff_t index;
825 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -0800826 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -0800827 int cycled;
David Howells811d7362006-08-29 19:06:09 +0100828 int range_whole = 0;
Aneesh Kumar K.V17bc6c32008-10-16 10:09:17 -0400829 long nr_to_write = wbc->nr_to_write;
David Howells811d7362006-08-29 19:06:09 +0100830
831 if (wbc->nonblocking && bdi_write_congested(bdi)) {
832 wbc->encountered_congestion = 1;
833 return 0;
834 }
835
David Howells811d7362006-08-29 19:06:09 +0100836 pagevec_init(&pvec, 0);
837 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -0800838 writeback_index = mapping->writeback_index; /* prev offset */
839 index = writeback_index;
840 if (index == 0)
841 cycled = 1;
842 else
843 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +0100844 end = -1;
845 } else {
846 index = wbc->range_start >> PAGE_CACHE_SHIFT;
847 end = wbc->range_end >> PAGE_CACHE_SHIFT;
848 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
849 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -0800850 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +0100851 }
852retry:
Nick Pigginbd19e012009-01-06 14:39:06 -0800853 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800854 while (!done && (index <= end)) {
855 int i;
856
857 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
858 PAGECACHE_TAG_DIRTY,
859 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
860 if (nr_pages == 0)
861 break;
David Howells811d7362006-08-29 19:06:09 +0100862
David Howells811d7362006-08-29 19:06:09 +0100863 for (i = 0; i < nr_pages; i++) {
864 struct page *page = pvec.pages[i];
865
Nick Piggind5482cd2009-01-06 14:39:11 -0800866 /*
867 * At this point, the page may be truncated or
868 * invalidated (changing page->mapping to NULL), or
869 * even swizzled back from swapper_space to tmpfs file
870 * mapping. However, page->index will not change
871 * because we have a reference on the page.
872 */
873 if (page->index > end) {
874 /*
875 * can't be range_cyclic (1st pass) because
876 * end == -1 in that case.
877 */
878 done = 1;
879 break;
880 }
881
Nick Pigginbd19e012009-01-06 14:39:06 -0800882 done_index = page->index + 1;
883
David Howells811d7362006-08-29 19:06:09 +0100884 lock_page(page);
885
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800886 /*
887 * Page truncated or invalidated. We can freely skip it
888 * then, even for data integrity operations: the page
889 * has disappeared concurrently, so there could be no
890 * real expectation of this data interity operation
891 * even if there is now a new, dirty page at the same
892 * pagecache address.
893 */
David Howells811d7362006-08-29 19:06:09 +0100894 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800895continue_unlock:
David Howells811d7362006-08-29 19:06:09 +0100896 unlock_page(page);
897 continue;
898 }
899
Nick Piggin515f4a02009-01-06 14:39:10 -0800900 if (!PageDirty(page)) {
901 /* someone wrote it for us */
902 goto continue_unlock;
903 }
David Howells811d7362006-08-29 19:06:09 +0100904
Nick Piggin515f4a02009-01-06 14:39:10 -0800905 if (PageWriteback(page)) {
906 if (wbc->sync_mode != WB_SYNC_NONE)
907 wait_on_page_writeback(page);
908 else
909 goto continue_unlock;
910 }
911
912 BUG_ON(PageWriteback(page));
913 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800914 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +0100915
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700916 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -0800917 if (unlikely(ret)) {
918 if (ret == AOP_WRITEPAGE_ACTIVATE) {
919 unlock_page(page);
920 ret = 0;
921 } else {
922 /*
923 * done_index is set past this page,
924 * so media errors will not choke
925 * background writeout for the entire
926 * file. This has consequences for
927 * range_cyclic semantics (ie. it may
928 * not be suitable for data integrity
929 * writeout).
930 */
931 done = 1;
932 break;
933 }
934 }
David Howells811d7362006-08-29 19:06:09 +0100935
Federico Cuello89e12192009-02-11 13:04:39 -0800936 if (nr_to_write > 0) {
Artem Bityutskiydcf6a792009-02-02 18:33:49 +0200937 nr_to_write--;
Federico Cuello89e12192009-02-11 13:04:39 -0800938 if (nr_to_write == 0 &&
939 wbc->sync_mode == WB_SYNC_NONE) {
940 /*
941 * We stop writing back only if we are
942 * not doing integrity sync. In case of
943 * integrity sync we have to keep going
944 * because someone may be concurrently
945 * dirtying pages, and we might have
946 * synced a lot of newly appeared dirty
947 * pages, but have not synced all of the
948 * old dirty pages.
949 */
950 done = 1;
951 break;
952 }
Nick Piggin05fe4782009-01-06 14:39:08 -0800953 }
Artem Bityutskiydcf6a792009-02-02 18:33:49 +0200954
David Howells811d7362006-08-29 19:06:09 +0100955 if (wbc->nonblocking && bdi_write_congested(bdi)) {
956 wbc->encountered_congestion = 1;
957 done = 1;
Andrew Morton82fd1a92009-01-06 14:39:11 -0800958 break;
David Howells811d7362006-08-29 19:06:09 +0100959 }
960 }
961 pagevec_release(&pvec);
962 cond_resched();
963 }
Nick Piggin3a4c6802009-02-12 04:34:23 +0100964 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +0100965 /*
Nick Piggin31a12662009-01-06 14:39:04 -0800966 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +0100967 * We hit the last page and there is more work to be done: wrap
968 * back to the start of the file
969 */
Nick Piggin31a12662009-01-06 14:39:04 -0800970 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +0100971 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -0800972 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +0100973 goto retry;
974 }
Aneesh Kumar K.V17bc6c32008-10-16 10:09:17 -0400975 if (!wbc->no_nrwrite_index_update) {
976 if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
Nick Pigginbd19e012009-01-06 14:39:06 -0800977 mapping->writeback_index = done_index;
Aneesh Kumar K.V17bc6c32008-10-16 10:09:17 -0400978 wbc->nr_to_write = nr_to_write;
979 }
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -0400980
David Howells811d7362006-08-29 19:06:09 +0100981 return ret;
982}
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700983EXPORT_SYMBOL(write_cache_pages);
984
985/*
986 * Function used by generic_writepages to call the real writepage
987 * function and set the mapping flags on error
988 */
989static int __writepage(struct page *page, struct writeback_control *wbc,
990 void *data)
991{
992 struct address_space *mapping = data;
993 int ret = mapping->a_ops->writepage(page, wbc);
994 mapping_set_error(mapping, ret);
995 return ret;
996}
997
998/**
999 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1000 * @mapping: address space structure to write
1001 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1002 *
1003 * This is a library function, which implements the writepages()
1004 * address_space_operation.
1005 */
1006int generic_writepages(struct address_space *mapping,
1007 struct writeback_control *wbc)
1008{
1009 /* deal with chardevs and other special file */
1010 if (!mapping->a_ops->writepage)
1011 return 0;
1012
1013 return write_cache_pages(mapping, wbc, __writepage, mapping);
1014}
David Howells811d7362006-08-29 19:06:09 +01001015
1016EXPORT_SYMBOL(generic_writepages);
1017
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1019{
Andrew Morton22905f72005-11-16 15:07:01 -08001020 int ret;
1021
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 if (wbc->nr_to_write <= 0)
1023 return 0;
1024 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001025 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001026 else
1027 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001028 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029}
1030
1031/**
1032 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001033 * @page: the page to write
1034 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 *
1036 * The page must be locked by the caller and will be unlocked upon return.
1037 *
1038 * write_one_page() returns a negative error code if I/O failed.
1039 */
1040int write_one_page(struct page *page, int wait)
1041{
1042 struct address_space *mapping = page->mapping;
1043 int ret = 0;
1044 struct writeback_control wbc = {
1045 .sync_mode = WB_SYNC_ALL,
1046 .nr_to_write = 1,
1047 };
1048
1049 BUG_ON(!PageLocked(page));
1050
1051 if (wait)
1052 wait_on_page_writeback(page);
1053
1054 if (clear_page_dirty_for_io(page)) {
1055 page_cache_get(page);
1056 ret = mapping->a_ops->writepage(page, &wbc);
1057 if (ret == 0 && wait) {
1058 wait_on_page_writeback(page);
1059 if (PageError(page))
1060 ret = -EIO;
1061 }
1062 page_cache_release(page);
1063 } else {
1064 unlock_page(page);
1065 }
1066 return ret;
1067}
1068EXPORT_SYMBOL(write_one_page);
1069
1070/*
Ken Chen76719322007-02-10 01:43:15 -08001071 * For address_spaces which do not use buffers nor write back.
1072 */
1073int __set_page_dirty_no_writeback(struct page *page)
1074{
1075 if (!PageDirty(page))
1076 SetPageDirty(page);
1077 return 0;
1078}
1079
1080/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001081 * Helper function for set_page_dirty family.
1082 * NOTE: This relies on being atomic wrt interrupts.
1083 */
1084void account_page_dirtied(struct page *page, struct address_space *mapping)
1085{
1086 if (mapping_cap_account_dirty(mapping)) {
1087 __inc_zone_page_state(page, NR_FILE_DIRTY);
1088 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1089 task_dirty_inc(current);
1090 task_io_account_write(PAGE_CACHE_SIZE);
1091 }
1092}
1093
1094/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 * For address_spaces which do not use buffers. Just tag the page as dirty in
1096 * its radix tree.
1097 *
1098 * This is also used when a single buffer is being dirtied: we want to set the
1099 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1100 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1101 *
1102 * Most callers have locked the page, which pins the address_space in memory.
1103 * But zap_pte_range() does not lock the page, however in that case the
1104 * mapping is pinned by the vma's ->vm_file reference.
1105 *
1106 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02001107 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 */
1109int __set_page_dirty_nobuffers(struct page *page)
1110{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 if (!TestSetPageDirty(page)) {
1112 struct address_space *mapping = page_mapping(page);
1113 struct address_space *mapping2;
1114
Andrew Morton8c085402006-12-10 02:19:24 -08001115 if (!mapping)
1116 return 1;
1117
Nick Piggin19fd6232008-07-25 19:45:32 -07001118 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001119 mapping2 = page_mapping(page);
1120 if (mapping2) { /* Race with truncate? */
1121 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001122 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001123 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08001124 radix_tree_tag_set(&mapping->page_tree,
1125 page_index(page), PAGECACHE_TAG_DIRTY);
1126 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001127 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001128 if (mapping->host) {
1129 /* !PageAnon && !swapper_space */
1130 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001132 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001134 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135}
1136EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1137
1138/*
1139 * When a writepage implementation decides that it doesn't want to write this
1140 * page for some reason, it should redirty the locked page via
1141 * redirty_page_for_writepage() and it should then unlock the page and return 0
1142 */
1143int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1144{
1145 wbc->pages_skipped++;
1146 return __set_page_dirty_nobuffers(page);
1147}
1148EXPORT_SYMBOL(redirty_page_for_writepage);
1149
1150/*
1151 * If the mapping doesn't provide a set_page_dirty a_op, then
1152 * just fall through and assume that it wants buffer_heads.
1153 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08001154int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155{
1156 struct address_space *mapping = page_mapping(page);
1157
1158 if (likely(mapping)) {
1159 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
David Howells93614012006-09-30 20:45:40 +02001160#ifdef CONFIG_BLOCK
1161 if (!spd)
1162 spd = __set_page_dirty_buffers;
1163#endif
1164 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001166 if (!PageDirty(page)) {
1167 if (!TestSetPageDirty(page))
1168 return 1;
1169 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 return 0;
1171}
1172EXPORT_SYMBOL(set_page_dirty);
1173
1174/*
1175 * set_page_dirty() is racy if the caller has no reference against
1176 * page->mapping->host, and if the page is unlocked. This is because another
1177 * CPU could truncate the page off the mapping and then free the mapping.
1178 *
1179 * Usually, the page _is_ locked, or the caller is a user-space process which
1180 * holds a reference on the inode by having an open file.
1181 *
1182 * In other cases, the page should be locked before running set_page_dirty().
1183 */
1184int set_page_dirty_lock(struct page *page)
1185{
1186 int ret;
1187
Nick Piggindb376482006-09-25 23:31:24 -07001188 lock_page_nosync(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 ret = set_page_dirty(page);
1190 unlock_page(page);
1191 return ret;
1192}
1193EXPORT_SYMBOL(set_page_dirty_lock);
1194
1195/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 * Clear a page's dirty flag, while caring for dirty memory accounting.
1197 * Returns true if the page was previously dirty.
1198 *
1199 * This is for preparing to put the page under writeout. We leave the page
1200 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1201 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1202 * implementation will run either set_page_writeback() or set_page_dirty(),
1203 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1204 * back into sync.
1205 *
1206 * This incoherency between the page's dirty flag and radix-tree tag is
1207 * unfortunate, but it only exists while the page is locked.
1208 */
1209int clear_page_dirty_for_io(struct page *page)
1210{
1211 struct address_space *mapping = page_mapping(page);
1212
Nick Piggin79352892007-07-19 01:47:22 -07001213 BUG_ON(!PageLocked(page));
1214
Fengguang Wufe3cba12007-07-19 01:48:07 -07001215 ClearPageReclaim(page);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001216 if (mapping && mapping_cap_account_dirty(mapping)) {
1217 /*
1218 * Yes, Virginia, this is indeed insane.
1219 *
1220 * We use this sequence to make sure that
1221 * (a) we account for dirty stats properly
1222 * (b) we tell the low-level filesystem to
1223 * mark the whole page dirty if it was
1224 * dirty in a pagetable. Only to then
1225 * (c) clean the page again and return 1 to
1226 * cause the writeback.
1227 *
1228 * This way we avoid all nasty races with the
1229 * dirty bit in multiple places and clearing
1230 * them concurrently from different threads.
1231 *
1232 * Note! Normally the "set_page_dirty(page)"
1233 * has no effect on the actual dirty bit - since
1234 * that will already usually be set. But we
1235 * need the side effects, and it can help us
1236 * avoid races.
1237 *
1238 * We basically use the page "master dirty bit"
1239 * as a serialization point for all the different
1240 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08001241 */
1242 if (page_mkclean(page))
1243 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07001244 /*
1245 * We carefully synchronise fault handlers against
1246 * installing a dirty pte and marking the page dirty
1247 * at this point. We do this by having them hold the
1248 * page lock at some point after installing their
1249 * pte, but before marking the page dirty.
1250 * Pages are always locked coming in here, so we get
1251 * the desired exclusion. See mm/memory.c:do_wp_page()
1252 * for more comments.
1253 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08001254 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001255 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001256 dec_bdi_stat(mapping->backing_dev_info,
1257 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001258 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001260 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001262 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263}
Hans Reiser58bb01a2005-11-18 01:10:53 -08001264EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265
1266int test_clear_page_writeback(struct page *page)
1267{
1268 struct address_space *mapping = page_mapping(page);
1269 int ret;
1270
1271 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001272 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273 unsigned long flags;
1274
Nick Piggin19fd6232008-07-25 19:45:32 -07001275 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001277 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 radix_tree_tag_clear(&mapping->page_tree,
1279 page_index(page),
1280 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001281 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001282 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001283 __bdi_writeout_inc(bdi);
1284 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001285 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001286 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 } else {
1288 ret = TestClearPageWriteback(page);
1289 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001290 if (ret)
1291 dec_zone_page_state(page, NR_WRITEBACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 return ret;
1293}
1294
1295int test_set_page_writeback(struct page *page)
1296{
1297 struct address_space *mapping = page_mapping(page);
1298 int ret;
1299
1300 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001301 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 unsigned long flags;
1303
Nick Piggin19fd6232008-07-25 19:45:32 -07001304 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001306 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 radix_tree_tag_set(&mapping->page_tree,
1308 page_index(page),
1309 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001310 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001311 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1312 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313 if (!PageDirty(page))
1314 radix_tree_tag_clear(&mapping->page_tree,
1315 page_index(page),
1316 PAGECACHE_TAG_DIRTY);
Nick Piggin19fd6232008-07-25 19:45:32 -07001317 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318 } else {
1319 ret = TestSetPageWriteback(page);
1320 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001321 if (!ret)
1322 inc_zone_page_state(page, NR_WRITEBACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 return ret;
1324
1325}
1326EXPORT_SYMBOL(test_set_page_writeback);
1327
1328/*
Nick Piggin00128182007-10-16 01:24:40 -07001329 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 * passed tag.
1331 */
1332int mapping_tagged(struct address_space *mapping, int tag)
1333{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 int ret;
Nick Piggin00128182007-10-16 01:24:40 -07001335 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 ret = radix_tree_tagged(&mapping->page_tree, tag);
Nick Piggin00128182007-10-16 01:24:40 -07001337 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 return ret;
1339}
1340EXPORT_SYMBOL(mapping_tagged);