blob: bbd396ac9546f5b0625c34facbc0b6d33fe2163a [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.
Wu Fengguang3a2e9a52009-09-23 21:56:00 +080047 * It should be somewhat larger than dirtied pages to ensure that reasonably
Linus Torvalds1da177e2005-04-16 15:20:36 -070048 * large amounts of I/O are submitted.
49 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +080050static inline long sync_writeback_pages(unsigned long dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -070051{
Wu Fengguang3a2e9a52009-09-23 21:56:00 +080052 if (dirtied < ratelimit_pages)
53 dirtied = ratelimit_pages;
54
55 return dirtied + dirtied / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -070056}
57
58/* The following parameters are exported via /proc/sys/vm */
59
60/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020061 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070062 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080063int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
65/*
David Rientjes2da02992009-01-06 14:39:31 -080066 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
67 * dirty_background_ratio * the amount of dirtyable memory
68 */
69unsigned long dirty_background_bytes;
70
71/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080072 * free highmem will not be subtracted from the total free memory
73 * for calculating free ratios if vm_highmem_is_dirtyable is true
74 */
75int vm_highmem_is_dirtyable;
76
77/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070078 * The generator of dirty data starts writeback at this percentage
79 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080080int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070081
82/*
David Rientjes2da02992009-01-06 14:39:31 -080083 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
84 * vm_dirty_ratio * the amount of dirtyable memory
85 */
86unsigned long vm_dirty_bytes;
87
88/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070089 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -070090 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -070091unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -070092
93/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -070094 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -070095 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -070096unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
98/*
99 * Flag that makes the machine dump writes/reads and block dirtyings.
100 */
101int block_dump;
102
103/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800104 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
105 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106 */
107int laptop_mode;
108
109EXPORT_SYMBOL(laptop_mode);
110
111/* End of sysctl-exported parameters */
112
113
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114/*
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700115 * Scale the writeback cache size proportional to the relative writeout speeds.
116 *
117 * We do this by keeping a floating proportion between BDIs, based on page
118 * writeback completions [end_page_writeback()]. Those devices that write out
119 * pages fastest will get the larger share, while the slower will get a smaller
120 * share.
121 *
122 * We use page writeout completions because we are interested in getting rid of
123 * dirty pages. Having them written out is the primary goal.
124 *
125 * We introduce a concept of time, a period over which we measure these events,
126 * because demand can/will vary over time. The length of this period itself is
127 * measured in page writeback completions.
128 *
129 */
130static struct prop_descriptor vm_completions;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700131static struct prop_descriptor vm_dirties;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700132
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700133/*
134 * couple the period to the dirty_ratio:
135 *
136 * period/2 ~ roundup_pow_of_two(dirty limit)
137 */
138static int calc_period_shift(void)
139{
140 unsigned long dirty_total;
141
David Rientjes2da02992009-01-06 14:39:31 -0800142 if (vm_dirty_bytes)
143 dirty_total = vm_dirty_bytes / PAGE_SIZE;
144 else
145 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
146 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700147 return 2 + ilog2(dirty_total - 1);
148}
149
150/*
David Rientjes2da02992009-01-06 14:39:31 -0800151 * update the period when the dirty threshold changes.
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700152 */
David Rientjes2da02992009-01-06 14:39:31 -0800153static void update_completion_period(void)
154{
155 int shift = calc_period_shift();
156 prop_change_shift(&vm_completions, shift);
157 prop_change_shift(&vm_dirties, shift);
158}
159
160int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700161 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800162 loff_t *ppos)
163{
164 int ret;
165
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700166 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800167 if (ret == 0 && write)
168 dirty_background_bytes = 0;
169 return ret;
170}
171
172int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700173 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800174 loff_t *ppos)
175{
176 int ret;
177
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700178 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800179 if (ret == 0 && write)
180 dirty_background_ratio = 0;
181 return ret;
182}
183
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700184int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700185 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700186 loff_t *ppos)
187{
188 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800189 int ret;
190
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700191 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700192 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
David Rientjes2da02992009-01-06 14:39:31 -0800193 update_completion_period();
194 vm_dirty_bytes = 0;
195 }
196 return ret;
197}
198
199
200int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700201 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800202 loff_t *ppos)
203{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800204 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800205 int ret;
206
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700207 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800208 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
209 update_completion_period();
210 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700211 }
212 return ret;
213}
214
215/*
216 * Increment the BDI's writeout completion count and the global writeout
217 * completion count. Called from test_clear_page_writeback().
218 */
219static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
220{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700221 __prop_inc_percpu_max(&vm_completions, &bdi->completions,
222 bdi->max_prop_frac);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700223}
224
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700225void bdi_writeout_inc(struct backing_dev_info *bdi)
226{
227 unsigned long flags;
228
229 local_irq_save(flags);
230 __bdi_writeout_inc(bdi);
231 local_irq_restore(flags);
232}
233EXPORT_SYMBOL_GPL(bdi_writeout_inc);
234
Nick Piggin1cf6e7d2009-02-18 14:48:18 -0800235void task_dirty_inc(struct task_struct *tsk)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700236{
237 prop_inc_single(&vm_dirties, &tsk->dirties);
238}
239
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700240/*
241 * Obtain an accurate fraction of the BDI's portion.
242 */
243static void bdi_writeout_fraction(struct backing_dev_info *bdi,
244 long *numerator, long *denominator)
245{
246 if (bdi_cap_writeback_dirty(bdi)) {
247 prop_fraction_percpu(&vm_completions, &bdi->completions,
248 numerator, denominator);
249 } else {
250 *numerator = 0;
251 *denominator = 1;
252 }
253}
254
255/*
256 * Clip the earned share of dirty pages to that which is actually available.
257 * This avoids exceeding the total dirty_limit when the floating averages
258 * fluctuate too quickly.
259 */
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700260static void clip_bdi_dirty_limit(struct backing_dev_info *bdi,
261 unsigned long dirty, unsigned long *pbdi_dirty)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700262{
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700263 unsigned long avail_dirty;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700264
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700265 avail_dirty = global_page_state(NR_FILE_DIRTY) +
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700266 global_page_state(NR_WRITEBACK) +
Miklos Szeredifc3ba692008-04-30 00:54:38 -0700267 global_page_state(NR_UNSTABLE_NFS) +
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700268 global_page_state(NR_WRITEBACK_TEMP);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700269
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700270 if (avail_dirty < dirty)
271 avail_dirty = dirty - avail_dirty;
272 else
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700273 avail_dirty = 0;
274
275 avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
276 bdi_stat(bdi, BDI_WRITEBACK);
277
278 *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
279}
280
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700281static inline void task_dirties_fraction(struct task_struct *tsk,
282 long *numerator, long *denominator)
283{
284 prop_fraction_single(&vm_dirties, &tsk->dirties,
285 numerator, denominator);
286}
287
288/*
289 * scale the dirty limit
290 *
291 * task specific dirty limit:
292 *
293 * dirty -= (dirty/8) * p_{t}
294 */
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700295static void task_dirty_limit(struct task_struct *tsk, unsigned long *pdirty)
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700296{
297 long numerator, denominator;
H Hartley Sweetendcf975d2009-06-16 15:31:44 -0700298 unsigned long dirty = *pdirty;
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700299 u64 inv = dirty >> 3;
300
301 task_dirties_fraction(tsk, &numerator, &denominator);
302 inv *= numerator;
303 do_div(inv, denominator);
304
305 dirty -= inv;
306 if (dirty < *pdirty/2)
307 dirty = *pdirty/2;
308
309 *pdirty = dirty;
310}
311
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700312/*
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700313 *
314 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700315static unsigned int bdi_min_ratio;
316
317int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
318{
319 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700320
Jens Axboecfc4ba52009-09-14 13:12:40 +0200321 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700322 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700323 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700324 } else {
325 min_ratio -= bdi->min_ratio;
326 if (bdi_min_ratio + min_ratio < 100) {
327 bdi_min_ratio += min_ratio;
328 bdi->min_ratio += min_ratio;
329 } else {
330 ret = -EINVAL;
331 }
332 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200333 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700334
335 return ret;
336}
337
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700338int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
339{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700340 int ret = 0;
341
342 if (max_ratio > 100)
343 return -EINVAL;
344
Jens Axboecfc4ba52009-09-14 13:12:40 +0200345 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700346 if (bdi->min_ratio > max_ratio) {
347 ret = -EINVAL;
348 } else {
349 bdi->max_ratio = max_ratio;
350 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
351 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200352 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700353
354 return ret;
355}
356EXPORT_SYMBOL(bdi_set_max_ratio);
357
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700358/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 * Work out the current dirty-memory clamping and background writeout
360 * thresholds.
361 *
362 * The main aim here is to lower them aggressively if there is a lot of mapped
363 * memory around. To avoid stressing page reclaim with lots of unreclaimable
364 * pages. It is better to clamp down on writers than to start swapping, and
365 * performing lots of scanning.
366 *
367 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
368 *
369 * We don't permit the clamping level to fall below 5% - that is getting rather
370 * excessive.
371 *
372 * We make sure that the background writeout level is below the adjusted
373 * clamping level.
374 */
Christoph Lameter1b424462007-05-06 14:48:59 -0700375
376static unsigned long highmem_dirtyable_memory(unsigned long total)
377{
378#ifdef CONFIG_HIGHMEM
379 int node;
380 unsigned long x = 0;
381
Lee Schermerhorn37b07e42007-10-16 01:25:39 -0700382 for_each_node_state(node, N_HIGH_MEMORY) {
Christoph Lameter1b424462007-05-06 14:48:59 -0700383 struct zone *z =
384 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
385
Wu Fengguangadea02a2009-09-21 17:01:42 -0700386 x += zone_page_state(z, NR_FREE_PAGES) +
387 zone_reclaimable_pages(z);
Christoph Lameter1b424462007-05-06 14:48:59 -0700388 }
389 /*
390 * Make sure that the number of highmem pages is never larger
391 * than the number of the total dirtyable memory. This can only
392 * occur in very strange VM situations but we want to make sure
393 * that this does not occur.
394 */
395 return min(x, total);
396#else
397 return 0;
398#endif
399}
400
Steven Rostedt3eefae92008-05-12 21:21:04 +0200401/**
402 * determine_dirtyable_memory - amount of memory that may be used
403 *
404 * Returns the numebr of pages that can currently be freed and used
405 * by the kernel for direct mappings.
406 */
407unsigned long determine_dirtyable_memory(void)
Christoph Lameter1b424462007-05-06 14:48:59 -0700408{
409 unsigned long x;
410
Wu Fengguangadea02a2009-09-21 17:01:42 -0700411 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
Bron Gondwana195cf4532008-02-04 22:29:20 -0800412
413 if (!vm_highmem_is_dirtyable)
414 x -= highmem_dirtyable_memory(x);
415
Christoph Lameter1b424462007-05-06 14:48:59 -0700416 return x + 1; /* Ensure that we never return 0 */
417}
418
Peter Zijlstracf0ca9f2008-04-30 00:54:32 -0700419void
David Rientjes364aeb22009-01-06 14:39:29 -0800420get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
421 unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422{
David Rientjes364aeb22009-01-06 14:39:29 -0800423 unsigned long background;
424 unsigned long dirty;
Christoph Lameter1b424462007-05-06 14:48:59 -0700425 unsigned long available_memory = determine_dirtyable_memory();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426 struct task_struct *tsk;
427
David Rientjes2da02992009-01-06 14:39:31 -0800428 if (vm_dirty_bytes)
429 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
430 else {
431 int dirty_ratio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432
David Rientjes2da02992009-01-06 14:39:31 -0800433 dirty_ratio = vm_dirty_ratio;
434 if (dirty_ratio < 5)
435 dirty_ratio = 5;
436 dirty = (dirty_ratio * available_memory) / 100;
437 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438
David Rientjes2da02992009-01-06 14:39:31 -0800439 if (dirty_background_bytes)
440 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
441 else
442 background = (dirty_background_ratio * available_memory) / 100;
443
444 if (background >= dirty)
445 background = dirty / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 tsk = current;
447 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
448 background += background / 4;
449 dirty += dirty / 4;
450 }
451 *pbackground = background;
452 *pdirty = dirty;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700453
454 if (bdi) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700455 u64 bdi_dirty;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700456 long numerator, denominator;
457
458 /*
459 * Calculate this BDI's share of the dirty ratio.
460 */
461 bdi_writeout_fraction(bdi, &numerator, &denominator);
462
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700463 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700464 bdi_dirty *= numerator;
465 do_div(bdi_dirty, denominator);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700466 bdi_dirty += (dirty * bdi->min_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700467 if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
468 bdi_dirty = dirty * bdi->max_ratio / 100;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700469
470 *pbdi_dirty = bdi_dirty;
471 clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700472 task_dirty_limit(current, pbdi_dirty);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700473 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474}
475
476/*
477 * balance_dirty_pages() must be called by processes which are generating dirty
478 * data. It looks at the number of dirty pages in the machine and will force
479 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
Jens Axboe5b0830c2009-09-23 19:37:09 +0200480 * If we're over `background_thresh' then the writeback threads are woken to
481 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 */
Wu Fengguang3a2e9a52009-09-23 21:56:00 +0800483static void balance_dirty_pages(struct address_space *mapping,
484 unsigned long write_chunk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485{
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800486 long nr_reclaimable, bdi_nr_reclaimable;
487 long nr_writeback, bdi_nr_writeback;
David Rientjes364aeb22009-01-06 14:39:29 -0800488 unsigned long background_thresh;
489 unsigned long dirty_thresh;
490 unsigned long bdi_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 unsigned long pages_written = 0;
Jens Axboe87c6a9b2009-09-17 19:59:14 +0200492 unsigned long pause = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493
494 struct backing_dev_info *bdi = mapping->backing_dev_info;
495
496 for (;;) {
497 struct writeback_control wbc = {
498 .bdi = bdi,
499 .sync_mode = WB_SYNC_NONE,
500 .older_than_this = NULL,
501 .nr_to_write = write_chunk,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700502 .range_cyclic = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 };
504
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700505 get_dirty_limits(&background_thresh, &dirty_thresh,
506 &bdi_thresh, bdi);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800507
508 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
509 global_page_state(NR_UNSTABLE_NFS);
510 nr_writeback = global_page_state(NR_WRITEBACK);
511
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700512 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
513 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800514
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700515 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
516 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517
Peter Zijlstra5fce25a2007-11-14 16:59:15 -0800518 /*
519 * Throttle it only when the background writeback cannot
520 * catch-up. This avoids (excessively) small writeouts
521 * when the bdi limits are ramping up.
522 */
523 if (nr_reclaimable + nr_writeback <
524 (background_thresh + dirty_thresh) / 2)
525 break;
526
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700527 if (!bdi->dirty_exceeded)
528 bdi->dirty_exceeded = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529
530 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
531 * Unstable writes are a feature of certain networked
532 * filesystems (i.e. NFS) in which data may have been
533 * written to the server's write cache, but has not yet
534 * been flushed to permanent storage.
Richard Kennedyd7831a02009-06-30 11:41:35 -0700535 * Only move pages to writeback if this bdi is over its
536 * threshold otherwise wait until the disk writes catch
537 * up.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 */
Richard Kennedyd7831a02009-06-30 11:41:35 -0700539 if (bdi_nr_reclaimable > bdi_thresh) {
Jens Axboe03ba3782009-09-09 09:08:54 +0200540 writeback_inodes_wbc(&wbc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541 pages_written += write_chunk - wbc.nr_to_write;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700542 get_dirty_limits(&background_thresh, &dirty_thresh,
543 &bdi_thresh, bdi);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700545
546 /*
547 * In order to avoid the stacked BDI deadlock we need
548 * to ensure we accurately count the 'dirty' pages when
549 * the threshold is low.
550 *
551 * Otherwise it would be possible to get thresh+n pages
552 * reported dirty, even though there are thresh-m pages
553 * actually dirty; with m+n sitting in the percpu
554 * deltas.
555 */
556 if (bdi_thresh < 2*bdi_stat_error(bdi)) {
557 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
558 bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
559 } else if (bdi_nr_reclaimable) {
560 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
561 bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
562 }
563
564 if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
565 break;
566 if (pages_written >= write_chunk)
567 break; /* We've done our duty */
568
Wu Fengguangd25105e2009-10-09 12:40:42 +0200569 __set_current_state(TASK_INTERRUPTIBLE);
570 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +0200571
572 /*
573 * Increase the delay for each loop, up to our previous
574 * default of taking a 100ms nap.
575 */
576 pause <<= 1;
577 if (pause > HZ / 10)
578 pause = HZ / 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 }
580
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700581 if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
582 bdi->dirty_exceeded)
583 bdi->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584
585 if (writeback_in_progress(bdi))
Jens Axboe5b0830c2009-09-23 19:37:09 +0200586 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587
588 /*
589 * In laptop mode, we wait until hitting the higher threshold before
590 * starting background writeout, and then write out all the way down
591 * to the lower threshold. So slow writers cause minimal disk activity.
592 *
593 * In normal mode, we start background writeout at the lower
594 * background_thresh, to keep the amount of dirty memory low.
595 */
596 if ((laptop_mode && pages_written) ||
Wu Fengguangd3ddec72009-09-23 20:33:40 +0800597 (!laptop_mode && ((global_page_state(NR_FILE_DIRTY)
598 + global_page_state(NR_UNSTABLE_NFS))
Jens Axboeb6e51312009-09-16 15:13:54 +0200599 > background_thresh)))
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200600 bdi_start_writeback(bdi, NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601}
602
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200603void set_page_dirty_balance(struct page *page, int page_mkwrite)
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700604{
Peter Zijlstraa200ee12007-10-08 18:54:37 +0200605 if (set_page_dirty(page) || page_mkwrite) {
Peter Zijlstraedc79b22006-09-25 23:30:58 -0700606 struct address_space *mapping = page_mapping(page);
607
608 if (mapping)
609 balance_dirty_pages_ratelimited(mapping);
610 }
611}
612
Tejun Heo245b2e72009-06-24 15:13:48 +0900613static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
614
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615/**
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800616 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -0700617 * @mapping: address_space which was dirtied
Martin Waitza5802902006-04-02 13:59:55 +0200618 * @nr_pages_dirtied: number of pages which the caller has just dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 *
620 * Processes which are dirtying memory should call in here once for each page
621 * which was newly dirtied. The function will periodically check the system's
622 * dirty state and will initiate writeback if needed.
623 *
624 * On really big machines, get_writeback_state is expensive, so try to avoid
625 * calling it too often (ratelimiting). But once we're over the dirty memory
626 * limit we decrease the ratelimiting by a lot, to prevent individual processes
627 * from overshooting the limit by (ratelimit_pages) each.
628 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800629void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
630 unsigned long nr_pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631{
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800632 unsigned long ratelimit;
633 unsigned long *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634
635 ratelimit = ratelimit_pages;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700636 if (mapping->backing_dev_info->dirty_exceeded)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 ratelimit = 8;
638
639 /*
640 * Check the rate limiting. Also, we do not want to throttle real-time
641 * tasks in balance_dirty_pages(). Period.
642 */
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800643 preempt_disable();
Tejun Heo245b2e72009-06-24 15:13:48 +0900644 p = &__get_cpu_var(bdp_ratelimits);
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800645 *p += nr_pages_dirtied;
646 if (unlikely(*p >= ratelimit)) {
Wu Fengguang3a2e9a52009-09-23 21:56:00 +0800647 ratelimit = sync_writeback_pages(*p);
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800648 *p = 0;
649 preempt_enable();
Wu Fengguang3a2e9a52009-09-23 21:56:00 +0800650 balance_dirty_pages(mapping, ratelimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 return;
652 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800653 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654}
Andrew Mortonfa5a7342006-03-24 03:18:10 -0800655EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656
Andrew Morton232ea4d2007-02-28 20:13:21 -0800657void throttle_vm_writeout(gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658{
David Rientjes364aeb22009-01-06 14:39:29 -0800659 unsigned long background_thresh;
660 unsigned long dirty_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661
662 for ( ; ; ) {
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700663 get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665 /*
666 * Boost the allowable dirty threshold a bit for page
667 * allocators so they don't get DoS'ed by heavy writers
668 */
669 dirty_thresh += dirty_thresh / 10; /* wheeee... */
670
Christoph Lameterc24f21b2006-06-30 01:55:42 -0700671 if (global_page_state(NR_UNSTABLE_NFS) +
672 global_page_state(NR_WRITEBACK) <= dirty_thresh)
673 break;
Jens Axboe8aa7e842009-07-09 14:52:32 +0200674 congestion_wait(BLK_RW_ASYNC, HZ/10);
Fengguang Wu369f2382007-10-16 23:30:45 -0700675
676 /*
677 * The caller might hold locks which can prevent IO completion
678 * or progress in the filesystem. So we cannot just sit here
679 * waiting for IO to complete.
680 */
681 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
682 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 }
684}
685
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
688 */
689int dirty_writeback_centisecs_handler(ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700690 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700692 proc_dointvec(table, write, buffer, length, ppos);
Jens Axboe64231042010-05-21 20:00:35 +0200693 bdi_arm_supers_timer();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 return 0;
695}
696
Jens Axboec2c49862010-05-20 09:18:47 +0200697#ifdef CONFIG_BLOCK
Matthew Garrett31373d02010-04-06 14:25:14 +0200698void laptop_mode_timer_fn(unsigned long data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699{
Matthew Garrett31373d02010-04-06 14:25:14 +0200700 struct request_queue *q = (struct request_queue *)data;
701 int nr_pages = global_page_state(NR_FILE_DIRTY) +
702 global_page_state(NR_UNSTABLE_NFS);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
Matthew Garrett31373d02010-04-06 14:25:14 +0200704 /*
705 * We want to write everything out, not just down to the dirty
706 * threshold
707 */
Jens Axboe03ba3782009-09-09 09:08:54 +0200708
Matthew Garrett31373d02010-04-06 14:25:14 +0200709 if (bdi_has_dirty_io(&q->backing_dev_info))
Jens Axboe0e3c9a22010-06-01 11:08:43 +0200710 bdi_start_writeback(&q->backing_dev_info, NULL, nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711}
712
713/*
714 * We've spun up the disk and we're in laptop mode: schedule writeback
715 * of all dirty data a few seconds from now. If the flush is already scheduled
716 * then push it back - the user is still using the disk.
717 */
Matthew Garrett31373d02010-04-06 14:25:14 +0200718void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719{
Matthew Garrett31373d02010-04-06 14:25:14 +0200720 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721}
722
723/*
724 * We're in laptop mode and we've just synced. The sync's writes will have
725 * caused another writeback to be scheduled by laptop_io_completion.
726 * Nothing needs to be written back anymore, so we unschedule the writeback.
727 */
728void laptop_sync_completion(void)
729{
Matthew Garrett31373d02010-04-06 14:25:14 +0200730 struct backing_dev_info *bdi;
731
732 rcu_read_lock();
733
734 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
735 del_timer(&bdi->laptop_mode_wb_timer);
736
737 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738}
Jens Axboec2c49862010-05-20 09:18:47 +0200739#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740
741/*
742 * If ratelimit_pages is too high then we can get into dirty-data overload
743 * if a large number of processes all perform writes at the same time.
744 * If it is too low then SMP machines will call the (expensive)
745 * get_writeback_state too often.
746 *
747 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
748 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
749 * thresholds before writeback cuts in.
750 *
751 * But the limit should not be set too high. Because it also controls the
752 * amount of memory which the balance_dirty_pages() caller has to write back.
753 * If this is too large then the caller will block on the IO queue all the
754 * time. So limit it to four megabytes - the balance_dirty_pages() caller
755 * will write six megabyte chunks, max.
756 */
757
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700758void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759{
Chandra Seetharaman40c99aa2006-09-29 02:01:24 -0700760 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 if (ratelimit_pages < 16)
762 ratelimit_pages = 16;
763 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
764 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
765}
766
Chandra Seetharaman26c21432006-06-27 02:54:10 -0700767static int __cpuinit
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
769{
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700770 writeback_set_ratelimit();
Paul E. McKenneyaa0f0302007-02-10 01:46:37 -0800771 return NOTIFY_DONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772}
773
Chandra Seetharaman74b85f32006-06-27 02:54:09 -0700774static struct notifier_block __cpuinitdata ratelimit_nb = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 .notifier_call = ratelimit_handler,
776 .next = NULL,
777};
778
779/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -0800780 * Called early on to tune the page writeback dirty limits.
781 *
782 * We used to scale dirty pages according to how total memory
783 * related to pages that could be allocated for buffers (by
784 * comparing nr_free_buffer_pages() to vm_total_pages.
785 *
786 * However, that was when we used "dirty_ratio" to scale with
787 * all memory, and we don't do that any more. "dirty_ratio"
788 * is now applied to total non-HIGHPAGE memory (by subtracting
789 * totalhigh_pages from vm_total_pages), and as such we can't
790 * get into the old insane situation any more where we had
791 * large amounts of dirty pages compared to a small amount of
792 * non-HIGHMEM memory.
793 *
794 * But we might still want to scale the dirty_ratio by how
795 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 */
797void __init page_writeback_init(void)
798{
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700799 int shift;
800
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -0700801 writeback_set_ratelimit();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 register_cpu_notifier(&ratelimit_nb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700803
804 shift = calc_period_shift();
805 prop_descriptor_init(&vm_completions, shift);
Peter Zijlstra3e26c142007-10-16 23:25:50 -0700806 prop_descriptor_init(&vm_dirties, shift);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807}
808
David Howells811d7362006-08-29 19:06:09 +0100809/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700810 * 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 +0100811 * @mapping: address space structure to write
812 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700813 * @writepage: function called for each page
814 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +0100815 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700816 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +0100817 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
818 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
819 * and msync() need to guarantee that all the data which was dirty at the time
820 * the call was made get new I/O started against them. If wbc->sync_mode is
821 * WB_SYNC_ALL then we were called for data integrity and we must wait for
822 * existing IO to complete.
David Howells811d7362006-08-29 19:06:09 +0100823 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700824int write_cache_pages(struct address_space *mapping,
825 struct writeback_control *wbc, writepage_t writepage,
826 void *data)
David Howells811d7362006-08-29 19:06:09 +0100827{
David Howells811d7362006-08-29 19:06:09 +0100828 int ret = 0;
829 int done = 0;
David Howells811d7362006-08-29 19:06:09 +0100830 struct pagevec pvec;
831 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -0800832 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +0100833 pgoff_t index;
834 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -0800835 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -0800836 int cycled;
David Howells811d7362006-08-29 19:06:09 +0100837 int range_whole = 0;
838
David Howells811d7362006-08-29 19:06:09 +0100839 pagevec_init(&pvec, 0);
840 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -0800841 writeback_index = mapping->writeback_index; /* prev offset */
842 index = writeback_index;
843 if (index == 0)
844 cycled = 1;
845 else
846 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +0100847 end = -1;
848 } else {
849 index = wbc->range_start >> PAGE_CACHE_SHIFT;
850 end = wbc->range_end >> PAGE_CACHE_SHIFT;
851 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
852 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -0800853 cycled = 1; /* ignore range_cyclic tests */
Dave Chinnerd87815c2010-06-09 10:37:20 +1000854
855 /*
856 * If this is a data integrity sync, cap the writeback to the
857 * current end of file. Any extension to the file that occurs
858 * after this is a new write and we don't need to write those
859 * pages out to fulfil our data integrity requirements. If we
860 * try to write them out, we can get stuck in this scan until
861 * the concurrent writer stops adding dirty pages and extending
862 * EOF.
863 */
864 if (wbc->sync_mode == WB_SYNC_ALL &&
865 wbc->range_end == LLONG_MAX) {
866 end = i_size_read(mapping->host) >> PAGE_CACHE_SHIFT;
867 }
David Howells811d7362006-08-29 19:06:09 +0100868 }
Dave Chinnerd87815c2010-06-09 10:37:20 +1000869
David Howells811d7362006-08-29 19:06:09 +0100870retry:
Nick Pigginbd19e012009-01-06 14:39:06 -0800871 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800872 while (!done && (index <= end)) {
873 int i;
874
875 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
876 PAGECACHE_TAG_DIRTY,
877 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
878 if (nr_pages == 0)
879 break;
David Howells811d7362006-08-29 19:06:09 +0100880
David Howells811d7362006-08-29 19:06:09 +0100881 for (i = 0; i < nr_pages; i++) {
882 struct page *page = pvec.pages[i];
883
Nick Piggind5482cd2009-01-06 14:39:11 -0800884 /*
885 * At this point, the page may be truncated or
886 * invalidated (changing page->mapping to NULL), or
887 * even swizzled back from swapper_space to tmpfs file
888 * mapping. However, page->index will not change
889 * because we have a reference on the page.
890 */
891 if (page->index > end) {
892 /*
893 * can't be range_cyclic (1st pass) because
894 * end == -1 in that case.
895 */
896 done = 1;
897 break;
898 }
899
Nick Pigginbd19e012009-01-06 14:39:06 -0800900 done_index = page->index + 1;
901
David Howells811d7362006-08-29 19:06:09 +0100902 lock_page(page);
903
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800904 /*
905 * Page truncated or invalidated. We can freely skip it
906 * then, even for data integrity operations: the page
907 * has disappeared concurrently, so there could be no
908 * real expectation of this data interity operation
909 * even if there is now a new, dirty page at the same
910 * pagecache address.
911 */
David Howells811d7362006-08-29 19:06:09 +0100912 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800913continue_unlock:
David Howells811d7362006-08-29 19:06:09 +0100914 unlock_page(page);
915 continue;
916 }
917
Nick Piggin515f4a02009-01-06 14:39:10 -0800918 if (!PageDirty(page)) {
919 /* someone wrote it for us */
920 goto continue_unlock;
921 }
David Howells811d7362006-08-29 19:06:09 +0100922
Nick Piggin515f4a02009-01-06 14:39:10 -0800923 if (PageWriteback(page)) {
924 if (wbc->sync_mode != WB_SYNC_NONE)
925 wait_on_page_writeback(page);
926 else
927 goto continue_unlock;
928 }
929
930 BUG_ON(PageWriteback(page));
931 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -0800932 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +0100933
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700934 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -0800935 if (unlikely(ret)) {
936 if (ret == AOP_WRITEPAGE_ACTIVATE) {
937 unlock_page(page);
938 ret = 0;
939 } else {
940 /*
941 * done_index is set past this page,
942 * so media errors will not choke
943 * background writeout for the entire
944 * file. This has consequences for
945 * range_cyclic semantics (ie. it may
946 * not be suitable for data integrity
947 * writeout).
948 */
949 done = 1;
950 break;
951 }
Dave Chinner0b564922010-06-09 10:37:18 +1000952 }
David Howells811d7362006-08-29 19:06:09 +0100953
Dave Chinner0b564922010-06-09 10:37:18 +1000954 if (wbc->nr_to_write > 0) {
955 if (--wbc->nr_to_write == 0 &&
Federico Cuello89e12192009-02-11 13:04:39 -0800956 wbc->sync_mode == WB_SYNC_NONE) {
957 /*
958 * We stop writing back only if we are
959 * not doing integrity sync. In case of
960 * integrity sync we have to keep going
961 * because someone may be concurrently
962 * dirtying pages, and we might have
963 * synced a lot of newly appeared dirty
964 * pages, but have not synced all of the
965 * old dirty pages.
966 */
967 done = 1;
968 break;
969 }
Nick Piggin05fe4782009-01-06 14:39:08 -0800970 }
David Howells811d7362006-08-29 19:06:09 +0100971 }
972 pagevec_release(&pvec);
973 cond_resched();
974 }
Nick Piggin3a4c6802009-02-12 04:34:23 +0100975 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +0100976 /*
Nick Piggin31a12662009-01-06 14:39:04 -0800977 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +0100978 * We hit the last page and there is more work to be done: wrap
979 * back to the start of the file
980 */
Nick Piggin31a12662009-01-06 14:39:04 -0800981 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +0100982 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -0800983 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +0100984 goto retry;
985 }
Dave Chinner0b564922010-06-09 10:37:18 +1000986 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
987 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -0400988
David Howells811d7362006-08-29 19:06:09 +0100989 return ret;
990}
Miklos Szeredi0ea97182007-05-10 22:22:51 -0700991EXPORT_SYMBOL(write_cache_pages);
992
993/*
994 * Function used by generic_writepages to call the real writepage
995 * function and set the mapping flags on error
996 */
997static int __writepage(struct page *page, struct writeback_control *wbc,
998 void *data)
999{
1000 struct address_space *mapping = data;
1001 int ret = mapping->a_ops->writepage(page, wbc);
1002 mapping_set_error(mapping, ret);
1003 return ret;
1004}
1005
1006/**
1007 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1008 * @mapping: address space structure to write
1009 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1010 *
1011 * This is a library function, which implements the writepages()
1012 * address_space_operation.
1013 */
1014int generic_writepages(struct address_space *mapping,
1015 struct writeback_control *wbc)
1016{
1017 /* deal with chardevs and other special file */
1018 if (!mapping->a_ops->writepage)
1019 return 0;
1020
1021 return write_cache_pages(mapping, wbc, __writepage, mapping);
1022}
David Howells811d7362006-08-29 19:06:09 +01001023
1024EXPORT_SYMBOL(generic_writepages);
1025
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1027{
Andrew Morton22905f72005-11-16 15:07:01 -08001028 int ret;
1029
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 if (wbc->nr_to_write <= 0)
1031 return 0;
1032 if (mapping->a_ops->writepages)
Peter Zijlstrad08b3852006-09-25 23:30:57 -07001033 ret = mapping->a_ops->writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001034 else
1035 ret = generic_writepages(mapping, wbc);
Andrew Morton22905f72005-11-16 15:07:01 -08001036 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037}
1038
1039/**
1040 * write_one_page - write out a single page and optionally wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07001041 * @page: the page to write
1042 * @wait: if true, wait on writeout
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 *
1044 * The page must be locked by the caller and will be unlocked upon return.
1045 *
1046 * write_one_page() returns a negative error code if I/O failed.
1047 */
1048int write_one_page(struct page *page, int wait)
1049{
1050 struct address_space *mapping = page->mapping;
1051 int ret = 0;
1052 struct writeback_control wbc = {
1053 .sync_mode = WB_SYNC_ALL,
1054 .nr_to_write = 1,
1055 };
1056
1057 BUG_ON(!PageLocked(page));
1058
1059 if (wait)
1060 wait_on_page_writeback(page);
1061
1062 if (clear_page_dirty_for_io(page)) {
1063 page_cache_get(page);
1064 ret = mapping->a_ops->writepage(page, &wbc);
1065 if (ret == 0 && wait) {
1066 wait_on_page_writeback(page);
1067 if (PageError(page))
1068 ret = -EIO;
1069 }
1070 page_cache_release(page);
1071 } else {
1072 unlock_page(page);
1073 }
1074 return ret;
1075}
1076EXPORT_SYMBOL(write_one_page);
1077
1078/*
Ken Chen76719322007-02-10 01:43:15 -08001079 * For address_spaces which do not use buffers nor write back.
1080 */
1081int __set_page_dirty_no_writeback(struct page *page)
1082{
1083 if (!PageDirty(page))
1084 SetPageDirty(page);
1085 return 0;
1086}
1087
1088/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001089 * Helper function for set_page_dirty family.
1090 * NOTE: This relies on being atomic wrt interrupts.
1091 */
1092void account_page_dirtied(struct page *page, struct address_space *mapping)
1093{
1094 if (mapping_cap_account_dirty(mapping)) {
1095 __inc_zone_page_state(page, NR_FILE_DIRTY);
1096 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1097 task_dirty_inc(current);
1098 task_io_account_write(PAGE_CACHE_SIZE);
1099 }
1100}
1101
1102/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 * For address_spaces which do not use buffers. Just tag the page as dirty in
1104 * its radix tree.
1105 *
1106 * This is also used when a single buffer is being dirtied: we want to set the
1107 * page dirty in that case, but not all the buffers. This is a "bottom-up"
1108 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1109 *
1110 * Most callers have locked the page, which pins the address_space in memory.
1111 * But zap_pte_range() does not lock the page, however in that case the
1112 * mapping is pinned by the vma's ->vm_file reference.
1113 *
1114 * We take care to handle the case where the page was truncated from the
Simon Arlott183ff222007-10-20 01:27:18 +02001115 * mapping by re-checking page_mapping() inside tree_lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 */
1117int __set_page_dirty_nobuffers(struct page *page)
1118{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 if (!TestSetPageDirty(page)) {
1120 struct address_space *mapping = page_mapping(page);
1121 struct address_space *mapping2;
1122
Andrew Morton8c085402006-12-10 02:19:24 -08001123 if (!mapping)
1124 return 1;
1125
Nick Piggin19fd6232008-07-25 19:45:32 -07001126 spin_lock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001127 mapping2 = page_mapping(page);
1128 if (mapping2) { /* Race with truncate? */
1129 BUG_ON(mapping2 != mapping);
Nick Piggin787d2212007-07-17 04:03:34 -07001130 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Edward Shishkine3a7cca2009-03-31 15:19:39 -07001131 account_page_dirtied(page, mapping);
Andrew Morton8c085402006-12-10 02:19:24 -08001132 radix_tree_tag_set(&mapping->page_tree,
1133 page_index(page), PAGECACHE_TAG_DIRTY);
1134 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001135 spin_unlock_irq(&mapping->tree_lock);
Andrew Morton8c085402006-12-10 02:19:24 -08001136 if (mapping->host) {
1137 /* !PageAnon && !swapper_space */
1138 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001140 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001142 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143}
1144EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1145
1146/*
1147 * When a writepage implementation decides that it doesn't want to write this
1148 * page for some reason, it should redirty the locked page via
1149 * redirty_page_for_writepage() and it should then unlock the page and return 0
1150 */
1151int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1152{
1153 wbc->pages_skipped++;
1154 return __set_page_dirty_nobuffers(page);
1155}
1156EXPORT_SYMBOL(redirty_page_for_writepage);
1157
1158/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02001159 * Dirty a page.
1160 *
1161 * For pages with a mapping this should be done under the page lock
1162 * for the benefit of asynchronous memory errors who prefer a consistent
1163 * dirty state. This rule can be broken in some special cases,
1164 * but should be better not to.
1165 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 * If the mapping doesn't provide a set_page_dirty a_op, then
1167 * just fall through and assume that it wants buffer_heads.
1168 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08001169int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170{
1171 struct address_space *mapping = page_mapping(page);
1172
1173 if (likely(mapping)) {
1174 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
David Howells93614012006-09-30 20:45:40 +02001175#ifdef CONFIG_BLOCK
1176 if (!spd)
1177 spd = __set_page_dirty_buffers;
1178#endif
1179 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08001181 if (!PageDirty(page)) {
1182 if (!TestSetPageDirty(page))
1183 return 1;
1184 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185 return 0;
1186}
1187EXPORT_SYMBOL(set_page_dirty);
1188
1189/*
1190 * set_page_dirty() is racy if the caller has no reference against
1191 * page->mapping->host, and if the page is unlocked. This is because another
1192 * CPU could truncate the page off the mapping and then free the mapping.
1193 *
1194 * Usually, the page _is_ locked, or the caller is a user-space process which
1195 * holds a reference on the inode by having an open file.
1196 *
1197 * In other cases, the page should be locked before running set_page_dirty().
1198 */
1199int set_page_dirty_lock(struct page *page)
1200{
1201 int ret;
1202
Nick Piggindb376482006-09-25 23:31:24 -07001203 lock_page_nosync(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204 ret = set_page_dirty(page);
1205 unlock_page(page);
1206 return ret;
1207}
1208EXPORT_SYMBOL(set_page_dirty_lock);
1209
1210/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 * Clear a page's dirty flag, while caring for dirty memory accounting.
1212 * Returns true if the page was previously dirty.
1213 *
1214 * This is for preparing to put the page under writeout. We leave the page
1215 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1216 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
1217 * implementation will run either set_page_writeback() or set_page_dirty(),
1218 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1219 * back into sync.
1220 *
1221 * This incoherency between the page's dirty flag and radix-tree tag is
1222 * unfortunate, but it only exists while the page is locked.
1223 */
1224int clear_page_dirty_for_io(struct page *page)
1225{
1226 struct address_space *mapping = page_mapping(page);
1227
Nick Piggin79352892007-07-19 01:47:22 -07001228 BUG_ON(!PageLocked(page));
1229
Fengguang Wufe3cba12007-07-19 01:48:07 -07001230 ClearPageReclaim(page);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001231 if (mapping && mapping_cap_account_dirty(mapping)) {
1232 /*
1233 * Yes, Virginia, this is indeed insane.
1234 *
1235 * We use this sequence to make sure that
1236 * (a) we account for dirty stats properly
1237 * (b) we tell the low-level filesystem to
1238 * mark the whole page dirty if it was
1239 * dirty in a pagetable. Only to then
1240 * (c) clean the page again and return 1 to
1241 * cause the writeback.
1242 *
1243 * This way we avoid all nasty races with the
1244 * dirty bit in multiple places and clearing
1245 * them concurrently from different threads.
1246 *
1247 * Note! Normally the "set_page_dirty(page)"
1248 * has no effect on the actual dirty bit - since
1249 * that will already usually be set. But we
1250 * need the side effects, and it can help us
1251 * avoid races.
1252 *
1253 * We basically use the page "master dirty bit"
1254 * as a serialization point for all the different
1255 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08001256 */
1257 if (page_mkclean(page))
1258 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07001259 /*
1260 * We carefully synchronise fault handlers against
1261 * installing a dirty pte and marking the page dirty
1262 * at this point. We do this by having them hold the
1263 * page lock at some point after installing their
1264 * pte, but before marking the page dirty.
1265 * Pages are always locked coming in here, so we get
1266 * the desired exclusion. See mm/memory.c:do_wp_page()
1267 * for more comments.
1268 */
Linus Torvalds7658cc22006-12-29 10:00:58 -08001269 if (TestClearPageDirty(page)) {
Andrew Morton8c085402006-12-10 02:19:24 -08001270 dec_zone_page_state(page, NR_FILE_DIRTY);
Peter Zijlstrac9e51e42007-10-16 23:25:47 -07001271 dec_bdi_stat(mapping->backing_dev_info,
1272 BDI_RECLAIMABLE);
Linus Torvalds7658cc22006-12-29 10:00:58 -08001273 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001275 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08001277 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278}
Hans Reiser58bb01a2005-11-18 01:10:53 -08001279EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280
1281int test_clear_page_writeback(struct page *page)
1282{
1283 struct address_space *mapping = page_mapping(page);
1284 int ret;
1285
1286 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001287 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 unsigned long flags;
1289
Nick Piggin19fd6232008-07-25 19:45:32 -07001290 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001292 if (ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 radix_tree_tag_clear(&mapping->page_tree,
1294 page_index(page),
1295 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001296 if (bdi_cap_account_writeback(bdi)) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001297 __dec_bdi_stat(bdi, BDI_WRITEBACK);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001298 __bdi_writeout_inc(bdi);
1299 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001300 }
Nick Piggin19fd6232008-07-25 19:45:32 -07001301 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 } else {
1303 ret = TestClearPageWriteback(page);
1304 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001305 if (ret)
1306 dec_zone_page_state(page, NR_WRITEBACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 return ret;
1308}
1309
1310int test_set_page_writeback(struct page *page)
1311{
1312 struct address_space *mapping = page_mapping(page);
1313 int ret;
1314
1315 if (mapping) {
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001316 struct backing_dev_info *bdi = mapping->backing_dev_info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 unsigned long flags;
1318
Nick Piggin19fd6232008-07-25 19:45:32 -07001319 spin_lock_irqsave(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001321 if (!ret) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 radix_tree_tag_set(&mapping->page_tree,
1323 page_index(page),
1324 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07001325 if (bdi_cap_account_writeback(bdi))
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07001326 __inc_bdi_stat(bdi, BDI_WRITEBACK);
1327 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001328 if (!PageDirty(page))
1329 radix_tree_tag_clear(&mapping->page_tree,
1330 page_index(page),
1331 PAGECACHE_TAG_DIRTY);
Nick Piggin19fd6232008-07-25 19:45:32 -07001332 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 } else {
1334 ret = TestSetPageWriteback(page);
1335 }
Andrew Mortond688abf2007-07-19 01:49:17 -07001336 if (!ret)
1337 inc_zone_page_state(page, NR_WRITEBACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 return ret;
1339
1340}
1341EXPORT_SYMBOL(test_set_page_writeback);
1342
1343/*
Nick Piggin00128182007-10-16 01:24:40 -07001344 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 * passed tag.
1346 */
1347int mapping_tagged(struct address_space *mapping, int tag)
1348{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 int ret;
Nick Piggin00128182007-10-16 01:24:40 -07001350 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 ret = radix_tree_tagged(&mapping->page_tree, tag);
Nick Piggin00128182007-10-16 01:24:40 -07001352 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 return ret;
1354}
1355EXPORT_SYMBOL(mapping_tagged);