Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
f30c2269 | 2 | * mm/page-writeback.c |
1da177e4 LT |
3 | * |
4 | * Copyright (C) 2002, Linus Torvalds. | |
04fbfdc1 | 5 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> |
1da177e4 LT |
6 | * |
7 | * Contains functions related to writing back dirty pages at the | |
8 | * address_space level. | |
9 | * | |
e1f8e874 | 10 | * 10Apr2002 Andrew Morton |
1da177e4 LT |
11 | * 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> | |
55e829af | 25 | #include <linux/task_io_accounting_ops.h> |
1da177e4 LT |
26 | #include <linux/blkdev.h> |
27 | #include <linux/mpage.h> | |
d08b3851 | 28 | #include <linux/rmap.h> |
1da177e4 LT |
29 | #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> | |
cf9a2ae8 | 35 | #include <linux/buffer_head.h> |
811d736f | 36 | #include <linux/pagevec.h> |
028c2dd1 | 37 | #include <trace/events/writeback.h> |
1da177e4 | 38 | |
e98be2d5 WF |
39 | /* |
40 | * Estimate write bandwidth at 200ms intervals. | |
41 | */ | |
42 | #define BANDWIDTH_INTERVAL max(HZ/5, 1) | |
43 | ||
1da177e4 LT |
44 | /* |
45 | * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited | |
46 | * will look to see if it needs to force writeback or throttling. | |
47 | */ | |
48 | static long ratelimit_pages = 32; | |
49 | ||
1da177e4 LT |
50 | /* |
51 | * When balance_dirty_pages decides that the caller needs to perform some | |
52 | * non-background writeback, this is how many pages it will attempt to write. | |
3a2e9a5a | 53 | * It should be somewhat larger than dirtied pages to ensure that reasonably |
1da177e4 LT |
54 | * large amounts of I/O are submitted. |
55 | */ | |
3a2e9a5a | 56 | static inline long sync_writeback_pages(unsigned long dirtied) |
1da177e4 | 57 | { |
3a2e9a5a WF |
58 | if (dirtied < ratelimit_pages) |
59 | dirtied = ratelimit_pages; | |
60 | ||
61 | return dirtied + dirtied / 2; | |
1da177e4 LT |
62 | } |
63 | ||
64 | /* The following parameters are exported via /proc/sys/vm */ | |
65 | ||
66 | /* | |
5b0830cb | 67 | * Start background writeback (via writeback threads) at this percentage |
1da177e4 | 68 | */ |
1b5e62b4 | 69 | int dirty_background_ratio = 10; |
1da177e4 | 70 | |
2da02997 DR |
71 | /* |
72 | * dirty_background_bytes starts at 0 (disabled) so that it is a function of | |
73 | * dirty_background_ratio * the amount of dirtyable memory | |
74 | */ | |
75 | unsigned long dirty_background_bytes; | |
76 | ||
195cf453 BG |
77 | /* |
78 | * free highmem will not be subtracted from the total free memory | |
79 | * for calculating free ratios if vm_highmem_is_dirtyable is true | |
80 | */ | |
81 | int vm_highmem_is_dirtyable; | |
82 | ||
1da177e4 LT |
83 | /* |
84 | * The generator of dirty data starts writeback at this percentage | |
85 | */ | |
1b5e62b4 | 86 | int vm_dirty_ratio = 20; |
1da177e4 | 87 | |
2da02997 DR |
88 | /* |
89 | * vm_dirty_bytes starts at 0 (disabled) so that it is a function of | |
90 | * vm_dirty_ratio * the amount of dirtyable memory | |
91 | */ | |
92 | unsigned long vm_dirty_bytes; | |
93 | ||
1da177e4 | 94 | /* |
704503d8 | 95 | * The interval between `kupdate'-style writebacks |
1da177e4 | 96 | */ |
22ef37ee | 97 | unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */ |
1da177e4 LT |
98 | |
99 | /* | |
704503d8 | 100 | * The longest time for which data is allowed to remain dirty |
1da177e4 | 101 | */ |
22ef37ee | 102 | unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */ |
1da177e4 LT |
103 | |
104 | /* | |
105 | * Flag that makes the machine dump writes/reads and block dirtyings. | |
106 | */ | |
107 | int block_dump; | |
108 | ||
109 | /* | |
ed5b43f1 BS |
110 | * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: |
111 | * a full sync is triggered after this time elapses without any disk activity. | |
1da177e4 LT |
112 | */ |
113 | int laptop_mode; | |
114 | ||
115 | EXPORT_SYMBOL(laptop_mode); | |
116 | ||
117 | /* End of sysctl-exported parameters */ | |
118 | ||
119 | ||
04fbfdc1 PZ |
120 | /* |
121 | * Scale the writeback cache size proportional to the relative writeout speeds. | |
122 | * | |
123 | * We do this by keeping a floating proportion between BDIs, based on page | |
124 | * writeback completions [end_page_writeback()]. Those devices that write out | |
125 | * pages fastest will get the larger share, while the slower will get a smaller | |
126 | * share. | |
127 | * | |
128 | * We use page writeout completions because we are interested in getting rid of | |
129 | * dirty pages. Having them written out is the primary goal. | |
130 | * | |
131 | * We introduce a concept of time, a period over which we measure these events, | |
132 | * because demand can/will vary over time. The length of this period itself is | |
133 | * measured in page writeback completions. | |
134 | * | |
135 | */ | |
136 | static struct prop_descriptor vm_completions; | |
3e26c149 | 137 | static struct prop_descriptor vm_dirties; |
04fbfdc1 | 138 | |
04fbfdc1 PZ |
139 | /* |
140 | * couple the period to the dirty_ratio: | |
141 | * | |
142 | * period/2 ~ roundup_pow_of_two(dirty limit) | |
143 | */ | |
144 | static int calc_period_shift(void) | |
145 | { | |
146 | unsigned long dirty_total; | |
147 | ||
2da02997 DR |
148 | if (vm_dirty_bytes) |
149 | dirty_total = vm_dirty_bytes / PAGE_SIZE; | |
150 | else | |
151 | dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / | |
152 | 100; | |
04fbfdc1 PZ |
153 | return 2 + ilog2(dirty_total - 1); |
154 | } | |
155 | ||
156 | /* | |
2da02997 | 157 | * update the period when the dirty threshold changes. |
04fbfdc1 | 158 | */ |
2da02997 DR |
159 | static void update_completion_period(void) |
160 | { | |
161 | int shift = calc_period_shift(); | |
162 | prop_change_shift(&vm_completions, shift); | |
163 | prop_change_shift(&vm_dirties, shift); | |
164 | } | |
165 | ||
166 | int dirty_background_ratio_handler(struct ctl_table *table, int write, | |
8d65af78 | 167 | void __user *buffer, size_t *lenp, |
2da02997 DR |
168 | loff_t *ppos) |
169 | { | |
170 | int ret; | |
171 | ||
8d65af78 | 172 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
173 | if (ret == 0 && write) |
174 | dirty_background_bytes = 0; | |
175 | return ret; | |
176 | } | |
177 | ||
178 | int dirty_background_bytes_handler(struct ctl_table *table, int write, | |
8d65af78 | 179 | void __user *buffer, size_t *lenp, |
2da02997 DR |
180 | loff_t *ppos) |
181 | { | |
182 | int ret; | |
183 | ||
8d65af78 | 184 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
185 | if (ret == 0 && write) |
186 | dirty_background_ratio = 0; | |
187 | return ret; | |
188 | } | |
189 | ||
04fbfdc1 | 190 | int dirty_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 191 | void __user *buffer, size_t *lenp, |
04fbfdc1 PZ |
192 | loff_t *ppos) |
193 | { | |
194 | int old_ratio = vm_dirty_ratio; | |
2da02997 DR |
195 | int ret; |
196 | ||
8d65af78 | 197 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
04fbfdc1 | 198 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
2da02997 DR |
199 | update_completion_period(); |
200 | vm_dirty_bytes = 0; | |
201 | } | |
202 | return ret; | |
203 | } | |
204 | ||
205 | ||
206 | int dirty_bytes_handler(struct ctl_table *table, int write, | |
8d65af78 | 207 | void __user *buffer, size_t *lenp, |
2da02997 DR |
208 | loff_t *ppos) |
209 | { | |
fc3501d4 | 210 | unsigned long old_bytes = vm_dirty_bytes; |
2da02997 DR |
211 | int ret; |
212 | ||
8d65af78 | 213 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
214 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
215 | update_completion_period(); | |
216 | vm_dirty_ratio = 0; | |
04fbfdc1 PZ |
217 | } |
218 | return ret; | |
219 | } | |
220 | ||
221 | /* | |
222 | * Increment the BDI's writeout completion count and the global writeout | |
223 | * completion count. Called from test_clear_page_writeback(). | |
224 | */ | |
225 | static inline void __bdi_writeout_inc(struct backing_dev_info *bdi) | |
226 | { | |
f7d2b1ec | 227 | __inc_bdi_stat(bdi, BDI_WRITTEN); |
a42dde04 PZ |
228 | __prop_inc_percpu_max(&vm_completions, &bdi->completions, |
229 | bdi->max_prop_frac); | |
04fbfdc1 PZ |
230 | } |
231 | ||
dd5656e5 MS |
232 | void bdi_writeout_inc(struct backing_dev_info *bdi) |
233 | { | |
234 | unsigned long flags; | |
235 | ||
236 | local_irq_save(flags); | |
237 | __bdi_writeout_inc(bdi); | |
238 | local_irq_restore(flags); | |
239 | } | |
240 | EXPORT_SYMBOL_GPL(bdi_writeout_inc); | |
241 | ||
1cf6e7d8 | 242 | void task_dirty_inc(struct task_struct *tsk) |
3e26c149 PZ |
243 | { |
244 | prop_inc_single(&vm_dirties, &tsk->dirties); | |
245 | } | |
246 | ||
04fbfdc1 PZ |
247 | /* |
248 | * Obtain an accurate fraction of the BDI's portion. | |
249 | */ | |
250 | static void bdi_writeout_fraction(struct backing_dev_info *bdi, | |
251 | long *numerator, long *denominator) | |
252 | { | |
3efaf0fa | 253 | prop_fraction_percpu(&vm_completions, &bdi->completions, |
04fbfdc1 | 254 | numerator, denominator); |
04fbfdc1 PZ |
255 | } |
256 | ||
3e26c149 PZ |
257 | static inline void task_dirties_fraction(struct task_struct *tsk, |
258 | long *numerator, long *denominator) | |
259 | { | |
260 | prop_fraction_single(&vm_dirties, &tsk->dirties, | |
261 | numerator, denominator); | |
262 | } | |
263 | ||
264 | /* | |
1babe183 | 265 | * task_dirty_limit - scale down dirty throttling threshold for one task |
3e26c149 PZ |
266 | * |
267 | * task specific dirty limit: | |
268 | * | |
269 | * dirty -= (dirty/8) * p_{t} | |
1babe183 WF |
270 | * |
271 | * To protect light/slow dirtying tasks from heavier/fast ones, we start | |
272 | * throttling individual tasks before reaching the bdi dirty limit. | |
273 | * Relatively low thresholds will be allocated to heavy dirtiers. So when | |
274 | * dirty pages grow large, heavy dirtiers will be throttled first, which will | |
275 | * effectively curb the growth of dirty pages. Light dirtiers with high enough | |
276 | * dirty threshold may never get throttled. | |
3e26c149 | 277 | */ |
16c4042f WF |
278 | static unsigned long task_dirty_limit(struct task_struct *tsk, |
279 | unsigned long bdi_dirty) | |
3e26c149 PZ |
280 | { |
281 | long numerator, denominator; | |
16c4042f | 282 | unsigned long dirty = bdi_dirty; |
3e26c149 PZ |
283 | u64 inv = dirty >> 3; |
284 | ||
285 | task_dirties_fraction(tsk, &numerator, &denominator); | |
286 | inv *= numerator; | |
287 | do_div(inv, denominator); | |
288 | ||
289 | dirty -= inv; | |
3e26c149 | 290 | |
16c4042f | 291 | return max(dirty, bdi_dirty/2); |
3e26c149 PZ |
292 | } |
293 | ||
189d3c4a PZ |
294 | /* |
295 | * | |
296 | */ | |
189d3c4a PZ |
297 | static unsigned int bdi_min_ratio; |
298 | ||
299 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) | |
300 | { | |
301 | int ret = 0; | |
189d3c4a | 302 | |
cfc4ba53 | 303 | spin_lock_bh(&bdi_lock); |
a42dde04 | 304 | if (min_ratio > bdi->max_ratio) { |
189d3c4a | 305 | ret = -EINVAL; |
a42dde04 PZ |
306 | } else { |
307 | min_ratio -= bdi->min_ratio; | |
308 | if (bdi_min_ratio + min_ratio < 100) { | |
309 | bdi_min_ratio += min_ratio; | |
310 | bdi->min_ratio += min_ratio; | |
311 | } else { | |
312 | ret = -EINVAL; | |
313 | } | |
314 | } | |
cfc4ba53 | 315 | spin_unlock_bh(&bdi_lock); |
a42dde04 PZ |
316 | |
317 | return ret; | |
318 | } | |
319 | ||
320 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio) | |
321 | { | |
a42dde04 PZ |
322 | int ret = 0; |
323 | ||
324 | if (max_ratio > 100) | |
325 | return -EINVAL; | |
326 | ||
cfc4ba53 | 327 | spin_lock_bh(&bdi_lock); |
a42dde04 PZ |
328 | if (bdi->min_ratio > max_ratio) { |
329 | ret = -EINVAL; | |
330 | } else { | |
331 | bdi->max_ratio = max_ratio; | |
332 | bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100; | |
333 | } | |
cfc4ba53 | 334 | spin_unlock_bh(&bdi_lock); |
189d3c4a PZ |
335 | |
336 | return ret; | |
337 | } | |
a42dde04 | 338 | EXPORT_SYMBOL(bdi_set_max_ratio); |
189d3c4a | 339 | |
1da177e4 LT |
340 | /* |
341 | * Work out the current dirty-memory clamping and background writeout | |
342 | * thresholds. | |
343 | * | |
344 | * The main aim here is to lower them aggressively if there is a lot of mapped | |
345 | * memory around. To avoid stressing page reclaim with lots of unreclaimable | |
346 | * pages. It is better to clamp down on writers than to start swapping, and | |
347 | * performing lots of scanning. | |
348 | * | |
349 | * We only allow 1/2 of the currently-unmapped memory to be dirtied. | |
350 | * | |
351 | * We don't permit the clamping level to fall below 5% - that is getting rather | |
352 | * excessive. | |
353 | * | |
354 | * We make sure that the background writeout level is below the adjusted | |
355 | * clamping level. | |
356 | */ | |
1b424464 CL |
357 | |
358 | static unsigned long highmem_dirtyable_memory(unsigned long total) | |
359 | { | |
360 | #ifdef CONFIG_HIGHMEM | |
361 | int node; | |
362 | unsigned long x = 0; | |
363 | ||
37b07e41 | 364 | for_each_node_state(node, N_HIGH_MEMORY) { |
1b424464 CL |
365 | struct zone *z = |
366 | &NODE_DATA(node)->node_zones[ZONE_HIGHMEM]; | |
367 | ||
adea02a1 WF |
368 | x += zone_page_state(z, NR_FREE_PAGES) + |
369 | zone_reclaimable_pages(z); | |
1b424464 CL |
370 | } |
371 | /* | |
372 | * Make sure that the number of highmem pages is never larger | |
373 | * than the number of the total dirtyable memory. This can only | |
374 | * occur in very strange VM situations but we want to make sure | |
375 | * that this does not occur. | |
376 | */ | |
377 | return min(x, total); | |
378 | #else | |
379 | return 0; | |
380 | #endif | |
381 | } | |
382 | ||
3eefae99 SR |
383 | /** |
384 | * determine_dirtyable_memory - amount of memory that may be used | |
385 | * | |
386 | * Returns the numebr of pages that can currently be freed and used | |
387 | * by the kernel for direct mappings. | |
388 | */ | |
389 | unsigned long determine_dirtyable_memory(void) | |
1b424464 CL |
390 | { |
391 | unsigned long x; | |
392 | ||
adea02a1 | 393 | x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages(); |
195cf453 BG |
394 | |
395 | if (!vm_highmem_is_dirtyable) | |
396 | x -= highmem_dirtyable_memory(x); | |
397 | ||
1b424464 CL |
398 | return x + 1; /* Ensure that we never return 0 */ |
399 | } | |
400 | ||
03ab450f | 401 | /* |
1babe183 WF |
402 | * global_dirty_limits - background-writeback and dirty-throttling thresholds |
403 | * | |
404 | * Calculate the dirty thresholds based on sysctl parameters | |
405 | * - vm.dirty_background_ratio or vm.dirty_background_bytes | |
406 | * - vm.dirty_ratio or vm.dirty_bytes | |
407 | * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and | |
ebd1373d | 408 | * real-time tasks. |
1babe183 | 409 | */ |
16c4042f | 410 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) |
1da177e4 | 411 | { |
364aeb28 DR |
412 | unsigned long background; |
413 | unsigned long dirty; | |
240c879f | 414 | unsigned long uninitialized_var(available_memory); |
1da177e4 LT |
415 | struct task_struct *tsk; |
416 | ||
240c879f MK |
417 | if (!vm_dirty_bytes || !dirty_background_bytes) |
418 | available_memory = determine_dirtyable_memory(); | |
419 | ||
2da02997 DR |
420 | if (vm_dirty_bytes) |
421 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE); | |
4cbec4c8 WF |
422 | else |
423 | dirty = (vm_dirty_ratio * available_memory) / 100; | |
1da177e4 | 424 | |
2da02997 DR |
425 | if (dirty_background_bytes) |
426 | background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE); | |
427 | else | |
428 | background = (dirty_background_ratio * available_memory) / 100; | |
1da177e4 | 429 | |
2da02997 DR |
430 | if (background >= dirty) |
431 | background = dirty / 2; | |
1da177e4 LT |
432 | tsk = current; |
433 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) { | |
434 | background += background / 4; | |
435 | dirty += dirty / 4; | |
436 | } | |
437 | *pbackground = background; | |
438 | *pdirty = dirty; | |
16c4042f | 439 | } |
04fbfdc1 | 440 | |
6f718656 | 441 | /** |
1babe183 | 442 | * bdi_dirty_limit - @bdi's share of dirty throttling threshold |
6f718656 WF |
443 | * @bdi: the backing_dev_info to query |
444 | * @dirty: global dirty limit in pages | |
445 | * | |
446 | * Returns @bdi's dirty limit in pages. The term "dirty" in the context of | |
447 | * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages. | |
448 | * And the "limit" in the name is not seriously taken as hard limit in | |
449 | * balance_dirty_pages(). | |
1babe183 | 450 | * |
6f718656 | 451 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
1babe183 WF |
452 | * - starving fast devices |
453 | * - piling up dirty pages (that will take long time to sync) on slow devices | |
454 | * | |
455 | * The bdi's share of dirty limit will be adapting to its throughput and | |
456 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. | |
457 | */ | |
458 | unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty) | |
16c4042f WF |
459 | { |
460 | u64 bdi_dirty; | |
461 | long numerator, denominator; | |
04fbfdc1 | 462 | |
16c4042f WF |
463 | /* |
464 | * Calculate this BDI's share of the dirty ratio. | |
465 | */ | |
466 | bdi_writeout_fraction(bdi, &numerator, &denominator); | |
04fbfdc1 | 467 | |
16c4042f WF |
468 | bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100; |
469 | bdi_dirty *= numerator; | |
470 | do_div(bdi_dirty, denominator); | |
04fbfdc1 | 471 | |
16c4042f WF |
472 | bdi_dirty += (dirty * bdi->min_ratio) / 100; |
473 | if (bdi_dirty > (dirty * bdi->max_ratio) / 100) | |
474 | bdi_dirty = dirty * bdi->max_ratio / 100; | |
475 | ||
476 | return bdi_dirty; | |
1da177e4 LT |
477 | } |
478 | ||
e98be2d5 WF |
479 | static void bdi_update_write_bandwidth(struct backing_dev_info *bdi, |
480 | unsigned long elapsed, | |
481 | unsigned long written) | |
482 | { | |
483 | const unsigned long period = roundup_pow_of_two(3 * HZ); | |
484 | unsigned long avg = bdi->avg_write_bandwidth; | |
485 | unsigned long old = bdi->write_bandwidth; | |
486 | u64 bw; | |
487 | ||
488 | /* | |
489 | * bw = written * HZ / elapsed | |
490 | * | |
491 | * bw * elapsed + write_bandwidth * (period - elapsed) | |
492 | * write_bandwidth = --------------------------------------------------- | |
493 | * period | |
494 | */ | |
495 | bw = written - bdi->written_stamp; | |
496 | bw *= HZ; | |
497 | if (unlikely(elapsed > period)) { | |
498 | do_div(bw, elapsed); | |
499 | avg = bw; | |
500 | goto out; | |
501 | } | |
502 | bw += (u64)bdi->write_bandwidth * (period - elapsed); | |
503 | bw >>= ilog2(period); | |
504 | ||
505 | /* | |
506 | * one more level of smoothing, for filtering out sudden spikes | |
507 | */ | |
508 | if (avg > old && old >= (unsigned long)bw) | |
509 | avg -= (avg - old) >> 3; | |
510 | ||
511 | if (avg < old && old <= (unsigned long)bw) | |
512 | avg += (old - avg) >> 3; | |
513 | ||
514 | out: | |
515 | bdi->write_bandwidth = bw; | |
516 | bdi->avg_write_bandwidth = avg; | |
517 | } | |
518 | ||
519 | void __bdi_update_bandwidth(struct backing_dev_info *bdi, | |
520 | unsigned long start_time) | |
521 | { | |
522 | unsigned long now = jiffies; | |
523 | unsigned long elapsed = now - bdi->bw_time_stamp; | |
524 | unsigned long written; | |
525 | ||
526 | /* | |
527 | * rate-limit, only update once every 200ms. | |
528 | */ | |
529 | if (elapsed < BANDWIDTH_INTERVAL) | |
530 | return; | |
531 | ||
532 | written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]); | |
533 | ||
534 | /* | |
535 | * Skip quiet periods when disk bandwidth is under-utilized. | |
536 | * (at least 1s idle time between two flusher runs) | |
537 | */ | |
538 | if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time)) | |
539 | goto snapshot; | |
540 | ||
541 | bdi_update_write_bandwidth(bdi, elapsed, written); | |
542 | ||
543 | snapshot: | |
544 | bdi->written_stamp = written; | |
545 | bdi->bw_time_stamp = now; | |
546 | } | |
547 | ||
548 | static void bdi_update_bandwidth(struct backing_dev_info *bdi, | |
549 | unsigned long start_time) | |
550 | { | |
551 | if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL)) | |
552 | return; | |
553 | spin_lock(&bdi->wb.list_lock); | |
554 | __bdi_update_bandwidth(bdi, start_time); | |
555 | spin_unlock(&bdi->wb.list_lock); | |
556 | } | |
557 | ||
1da177e4 LT |
558 | /* |
559 | * balance_dirty_pages() must be called by processes which are generating dirty | |
560 | * data. It looks at the number of dirty pages in the machine and will force | |
561 | * the caller to perform writeback if the system is over `vm_dirty_ratio'. | |
5b0830cb JA |
562 | * If we're over `background_thresh' then the writeback threads are woken to |
563 | * perform some writeout. | |
1da177e4 | 564 | */ |
3a2e9a5a WF |
565 | static void balance_dirty_pages(struct address_space *mapping, |
566 | unsigned long write_chunk) | |
1da177e4 | 567 | { |
5fce25a9 PZ |
568 | long nr_reclaimable, bdi_nr_reclaimable; |
569 | long nr_writeback, bdi_nr_writeback; | |
364aeb28 DR |
570 | unsigned long background_thresh; |
571 | unsigned long dirty_thresh; | |
572 | unsigned long bdi_thresh; | |
1da177e4 | 573 | unsigned long pages_written = 0; |
87c6a9b2 | 574 | unsigned long pause = 1; |
e50e3720 | 575 | bool dirty_exceeded = false; |
1da177e4 | 576 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
e98be2d5 | 577 | unsigned long start_time = jiffies; |
1da177e4 LT |
578 | |
579 | for (;;) { | |
5fce25a9 PZ |
580 | nr_reclaimable = global_page_state(NR_FILE_DIRTY) + |
581 | global_page_state(NR_UNSTABLE_NFS); | |
582 | nr_writeback = global_page_state(NR_WRITEBACK); | |
583 | ||
16c4042f WF |
584 | global_dirty_limits(&background_thresh, &dirty_thresh); |
585 | ||
586 | /* | |
587 | * Throttle it only when the background writeback cannot | |
588 | * catch-up. This avoids (excessively) small writeouts | |
589 | * when the bdi limits are ramping up. | |
590 | */ | |
4cbec4c8 | 591 | if (nr_reclaimable + nr_writeback <= |
16c4042f WF |
592 | (background_thresh + dirty_thresh) / 2) |
593 | break; | |
594 | ||
595 | bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh); | |
596 | bdi_thresh = task_dirty_limit(current, bdi_thresh); | |
597 | ||
e50e3720 WF |
598 | /* |
599 | * In order to avoid the stacked BDI deadlock we need | |
600 | * to ensure we accurately count the 'dirty' pages when | |
601 | * the threshold is low. | |
602 | * | |
603 | * Otherwise it would be possible to get thresh+n pages | |
604 | * reported dirty, even though there are thresh-m pages | |
605 | * actually dirty; with m+n sitting in the percpu | |
606 | * deltas. | |
607 | */ | |
608 | if (bdi_thresh < 2*bdi_stat_error(bdi)) { | |
609 | bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE); | |
610 | bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK); | |
611 | } else { | |
612 | bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE); | |
613 | bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK); | |
614 | } | |
5fce25a9 | 615 | |
e50e3720 WF |
616 | /* |
617 | * The bdi thresh is somehow "soft" limit derived from the | |
618 | * global "hard" limit. The former helps to prevent heavy IO | |
619 | * bdi or process from holding back light ones; The latter is | |
620 | * the last resort safeguard. | |
621 | */ | |
622 | dirty_exceeded = | |
4cbec4c8 WF |
623 | (bdi_nr_reclaimable + bdi_nr_writeback > bdi_thresh) |
624 | || (nr_reclaimable + nr_writeback > dirty_thresh); | |
e50e3720 WF |
625 | |
626 | if (!dirty_exceeded) | |
04fbfdc1 | 627 | break; |
1da177e4 | 628 | |
04fbfdc1 PZ |
629 | if (!bdi->dirty_exceeded) |
630 | bdi->dirty_exceeded = 1; | |
1da177e4 | 631 | |
e98be2d5 WF |
632 | bdi_update_bandwidth(bdi, start_time); |
633 | ||
1da177e4 LT |
634 | /* Note: nr_reclaimable denotes nr_dirty + nr_unstable. |
635 | * Unstable writes are a feature of certain networked | |
636 | * filesystems (i.e. NFS) in which data may have been | |
637 | * written to the server's write cache, but has not yet | |
638 | * been flushed to permanent storage. | |
d7831a0b RK |
639 | * Only move pages to writeback if this bdi is over its |
640 | * threshold otherwise wait until the disk writes catch | |
641 | * up. | |
1da177e4 | 642 | */ |
d46db3d5 | 643 | trace_balance_dirty_start(bdi); |
d7831a0b | 644 | if (bdi_nr_reclaimable > bdi_thresh) { |
d46db3d5 WF |
645 | pages_written += writeback_inodes_wb(&bdi->wb, |
646 | write_chunk); | |
647 | trace_balance_dirty_written(bdi, pages_written); | |
e50e3720 WF |
648 | if (pages_written >= write_chunk) |
649 | break; /* We've done our duty */ | |
04fbfdc1 | 650 | } |
d153ba64 | 651 | __set_current_state(TASK_UNINTERRUPTIBLE); |
d25105e8 | 652 | io_schedule_timeout(pause); |
d46db3d5 | 653 | trace_balance_dirty_wait(bdi); |
87c6a9b2 JA |
654 | |
655 | /* | |
656 | * Increase the delay for each loop, up to our previous | |
657 | * default of taking a 100ms nap. | |
658 | */ | |
659 | pause <<= 1; | |
660 | if (pause > HZ / 10) | |
661 | pause = HZ / 10; | |
1da177e4 LT |
662 | } |
663 | ||
e50e3720 | 664 | if (!dirty_exceeded && bdi->dirty_exceeded) |
04fbfdc1 | 665 | bdi->dirty_exceeded = 0; |
1da177e4 LT |
666 | |
667 | if (writeback_in_progress(bdi)) | |
5b0830cb | 668 | return; |
1da177e4 LT |
669 | |
670 | /* | |
671 | * In laptop mode, we wait until hitting the higher threshold before | |
672 | * starting background writeout, and then write out all the way down | |
673 | * to the lower threshold. So slow writers cause minimal disk activity. | |
674 | * | |
675 | * In normal mode, we start background writeout at the lower | |
676 | * background_thresh, to keep the amount of dirty memory low. | |
677 | */ | |
678 | if ((laptop_mode && pages_written) || | |
e50e3720 | 679 | (!laptop_mode && (nr_reclaimable > background_thresh))) |
c5444198 | 680 | bdi_start_background_writeback(bdi); |
1da177e4 LT |
681 | } |
682 | ||
a200ee18 | 683 | void set_page_dirty_balance(struct page *page, int page_mkwrite) |
edc79b2a | 684 | { |
a200ee18 | 685 | if (set_page_dirty(page) || page_mkwrite) { |
edc79b2a PZ |
686 | struct address_space *mapping = page_mapping(page); |
687 | ||
688 | if (mapping) | |
689 | balance_dirty_pages_ratelimited(mapping); | |
690 | } | |
691 | } | |
692 | ||
245b2e70 TH |
693 | static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0; |
694 | ||
1da177e4 | 695 | /** |
fa5a734e | 696 | * balance_dirty_pages_ratelimited_nr - balance dirty memory state |
67be2dd1 | 697 | * @mapping: address_space which was dirtied |
a580290c | 698 | * @nr_pages_dirtied: number of pages which the caller has just dirtied |
1da177e4 LT |
699 | * |
700 | * Processes which are dirtying memory should call in here once for each page | |
701 | * which was newly dirtied. The function will periodically check the system's | |
702 | * dirty state and will initiate writeback if needed. | |
703 | * | |
704 | * On really big machines, get_writeback_state is expensive, so try to avoid | |
705 | * calling it too often (ratelimiting). But once we're over the dirty memory | |
706 | * limit we decrease the ratelimiting by a lot, to prevent individual processes | |
707 | * from overshooting the limit by (ratelimit_pages) each. | |
708 | */ | |
fa5a734e AM |
709 | void balance_dirty_pages_ratelimited_nr(struct address_space *mapping, |
710 | unsigned long nr_pages_dirtied) | |
1da177e4 | 711 | { |
36715cef | 712 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
fa5a734e AM |
713 | unsigned long ratelimit; |
714 | unsigned long *p; | |
1da177e4 | 715 | |
36715cef WF |
716 | if (!bdi_cap_account_dirty(bdi)) |
717 | return; | |
718 | ||
1da177e4 | 719 | ratelimit = ratelimit_pages; |
04fbfdc1 | 720 | if (mapping->backing_dev_info->dirty_exceeded) |
1da177e4 LT |
721 | ratelimit = 8; |
722 | ||
723 | /* | |
724 | * Check the rate limiting. Also, we do not want to throttle real-time | |
725 | * tasks in balance_dirty_pages(). Period. | |
726 | */ | |
fa5a734e | 727 | preempt_disable(); |
245b2e70 | 728 | p = &__get_cpu_var(bdp_ratelimits); |
fa5a734e AM |
729 | *p += nr_pages_dirtied; |
730 | if (unlikely(*p >= ratelimit)) { | |
3a2e9a5a | 731 | ratelimit = sync_writeback_pages(*p); |
fa5a734e AM |
732 | *p = 0; |
733 | preempt_enable(); | |
3a2e9a5a | 734 | balance_dirty_pages(mapping, ratelimit); |
1da177e4 LT |
735 | return; |
736 | } | |
fa5a734e | 737 | preempt_enable(); |
1da177e4 | 738 | } |
fa5a734e | 739 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr); |
1da177e4 | 740 | |
232ea4d6 | 741 | void throttle_vm_writeout(gfp_t gfp_mask) |
1da177e4 | 742 | { |
364aeb28 DR |
743 | unsigned long background_thresh; |
744 | unsigned long dirty_thresh; | |
1da177e4 LT |
745 | |
746 | for ( ; ; ) { | |
16c4042f | 747 | global_dirty_limits(&background_thresh, &dirty_thresh); |
1da177e4 LT |
748 | |
749 | /* | |
750 | * Boost the allowable dirty threshold a bit for page | |
751 | * allocators so they don't get DoS'ed by heavy writers | |
752 | */ | |
753 | dirty_thresh += dirty_thresh / 10; /* wheeee... */ | |
754 | ||
c24f21bd CL |
755 | if (global_page_state(NR_UNSTABLE_NFS) + |
756 | global_page_state(NR_WRITEBACK) <= dirty_thresh) | |
757 | break; | |
8aa7e847 | 758 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
369f2389 FW |
759 | |
760 | /* | |
761 | * The caller might hold locks which can prevent IO completion | |
762 | * or progress in the filesystem. So we cannot just sit here | |
763 | * waiting for IO to complete. | |
764 | */ | |
765 | if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) | |
766 | break; | |
1da177e4 LT |
767 | } |
768 | } | |
769 | ||
1da177e4 LT |
770 | /* |
771 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
772 | */ | |
773 | int dirty_writeback_centisecs_handler(ctl_table *table, int write, | |
8d65af78 | 774 | void __user *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 775 | { |
8d65af78 | 776 | proc_dointvec(table, write, buffer, length, ppos); |
6423104b | 777 | bdi_arm_supers_timer(); |
1da177e4 LT |
778 | return 0; |
779 | } | |
780 | ||
c2c4986e | 781 | #ifdef CONFIG_BLOCK |
31373d09 | 782 | void laptop_mode_timer_fn(unsigned long data) |
1da177e4 | 783 | { |
31373d09 MG |
784 | struct request_queue *q = (struct request_queue *)data; |
785 | int nr_pages = global_page_state(NR_FILE_DIRTY) + | |
786 | global_page_state(NR_UNSTABLE_NFS); | |
1da177e4 | 787 | |
31373d09 MG |
788 | /* |
789 | * We want to write everything out, not just down to the dirty | |
790 | * threshold | |
791 | */ | |
31373d09 | 792 | if (bdi_has_dirty_io(&q->backing_dev_info)) |
c5444198 | 793 | bdi_start_writeback(&q->backing_dev_info, nr_pages); |
1da177e4 LT |
794 | } |
795 | ||
796 | /* | |
797 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
798 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
799 | * then push it back - the user is still using the disk. | |
800 | */ | |
31373d09 | 801 | void laptop_io_completion(struct backing_dev_info *info) |
1da177e4 | 802 | { |
31373d09 | 803 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
804 | } |
805 | ||
806 | /* | |
807 | * We're in laptop mode and we've just synced. The sync's writes will have | |
808 | * caused another writeback to be scheduled by laptop_io_completion. | |
809 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
810 | */ | |
811 | void laptop_sync_completion(void) | |
812 | { | |
31373d09 MG |
813 | struct backing_dev_info *bdi; |
814 | ||
815 | rcu_read_lock(); | |
816 | ||
817 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) | |
818 | del_timer(&bdi->laptop_mode_wb_timer); | |
819 | ||
820 | rcu_read_unlock(); | |
1da177e4 | 821 | } |
c2c4986e | 822 | #endif |
1da177e4 LT |
823 | |
824 | /* | |
825 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
826 | * if a large number of processes all perform writes at the same time. | |
827 | * If it is too low then SMP machines will call the (expensive) | |
828 | * get_writeback_state too often. | |
829 | * | |
830 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
831 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
832 | * thresholds before writeback cuts in. | |
833 | * | |
834 | * But the limit should not be set too high. Because it also controls the | |
835 | * amount of memory which the balance_dirty_pages() caller has to write back. | |
836 | * If this is too large then the caller will block on the IO queue all the | |
837 | * time. So limit it to four megabytes - the balance_dirty_pages() caller | |
838 | * will write six megabyte chunks, max. | |
839 | */ | |
840 | ||
2d1d43f6 | 841 | void writeback_set_ratelimit(void) |
1da177e4 | 842 | { |
40c99aae | 843 | ratelimit_pages = vm_total_pages / (num_online_cpus() * 32); |
1da177e4 LT |
844 | if (ratelimit_pages < 16) |
845 | ratelimit_pages = 16; | |
846 | if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024) | |
847 | ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE; | |
848 | } | |
849 | ||
26c2143b | 850 | static int __cpuinit |
1da177e4 LT |
851 | ratelimit_handler(struct notifier_block *self, unsigned long u, void *v) |
852 | { | |
2d1d43f6 | 853 | writeback_set_ratelimit(); |
aa0f0303 | 854 | return NOTIFY_DONE; |
1da177e4 LT |
855 | } |
856 | ||
74b85f37 | 857 | static struct notifier_block __cpuinitdata ratelimit_nb = { |
1da177e4 LT |
858 | .notifier_call = ratelimit_handler, |
859 | .next = NULL, | |
860 | }; | |
861 | ||
862 | /* | |
dc6e29da LT |
863 | * Called early on to tune the page writeback dirty limits. |
864 | * | |
865 | * We used to scale dirty pages according to how total memory | |
866 | * related to pages that could be allocated for buffers (by | |
867 | * comparing nr_free_buffer_pages() to vm_total_pages. | |
868 | * | |
869 | * However, that was when we used "dirty_ratio" to scale with | |
870 | * all memory, and we don't do that any more. "dirty_ratio" | |
871 | * is now applied to total non-HIGHPAGE memory (by subtracting | |
872 | * totalhigh_pages from vm_total_pages), and as such we can't | |
873 | * get into the old insane situation any more where we had | |
874 | * large amounts of dirty pages compared to a small amount of | |
875 | * non-HIGHMEM memory. | |
876 | * | |
877 | * But we might still want to scale the dirty_ratio by how | |
878 | * much memory the box has.. | |
1da177e4 LT |
879 | */ |
880 | void __init page_writeback_init(void) | |
881 | { | |
04fbfdc1 PZ |
882 | int shift; |
883 | ||
2d1d43f6 | 884 | writeback_set_ratelimit(); |
1da177e4 | 885 | register_cpu_notifier(&ratelimit_nb); |
04fbfdc1 PZ |
886 | |
887 | shift = calc_period_shift(); | |
888 | prop_descriptor_init(&vm_completions, shift); | |
3e26c149 | 889 | prop_descriptor_init(&vm_dirties, shift); |
1da177e4 LT |
890 | } |
891 | ||
f446daae JK |
892 | /** |
893 | * tag_pages_for_writeback - tag pages to be written by write_cache_pages | |
894 | * @mapping: address space structure to write | |
895 | * @start: starting page index | |
896 | * @end: ending page index (inclusive) | |
897 | * | |
898 | * This function scans the page range from @start to @end (inclusive) and tags | |
899 | * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is | |
900 | * that write_cache_pages (or whoever calls this function) will then use | |
901 | * TOWRITE tag to identify pages eligible for writeback. This mechanism is | |
902 | * used to avoid livelocking of writeback by a process steadily creating new | |
903 | * dirty pages in the file (thus it is important for this function to be quick | |
904 | * so that it can tag pages faster than a dirtying process can create them). | |
905 | */ | |
906 | /* | |
907 | * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency. | |
908 | */ | |
f446daae JK |
909 | void tag_pages_for_writeback(struct address_space *mapping, |
910 | pgoff_t start, pgoff_t end) | |
911 | { | |
3c111a07 | 912 | #define WRITEBACK_TAG_BATCH 4096 |
f446daae JK |
913 | unsigned long tagged; |
914 | ||
915 | do { | |
916 | spin_lock_irq(&mapping->tree_lock); | |
917 | tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree, | |
918 | &start, end, WRITEBACK_TAG_BATCH, | |
919 | PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE); | |
920 | spin_unlock_irq(&mapping->tree_lock); | |
921 | WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH); | |
922 | cond_resched(); | |
d5ed3a4a JK |
923 | /* We check 'start' to handle wrapping when end == ~0UL */ |
924 | } while (tagged >= WRITEBACK_TAG_BATCH && start); | |
f446daae JK |
925 | } |
926 | EXPORT_SYMBOL(tag_pages_for_writeback); | |
927 | ||
811d736f | 928 | /** |
0ea97180 | 929 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. |
811d736f DH |
930 | * @mapping: address space structure to write |
931 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
0ea97180 MS |
932 | * @writepage: function called for each page |
933 | * @data: data passed to writepage function | |
811d736f | 934 | * |
0ea97180 | 935 | * If a page is already under I/O, write_cache_pages() skips it, even |
811d736f DH |
936 | * if it's dirty. This is desirable behaviour for memory-cleaning writeback, |
937 | * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() | |
938 | * and msync() need to guarantee that all the data which was dirty at the time | |
939 | * the call was made get new I/O started against them. If wbc->sync_mode is | |
940 | * WB_SYNC_ALL then we were called for data integrity and we must wait for | |
941 | * existing IO to complete. | |
f446daae JK |
942 | * |
943 | * To avoid livelocks (when other process dirties new pages), we first tag | |
944 | * pages which should be written back with TOWRITE tag and only then start | |
945 | * writing them. For data-integrity sync we have to be careful so that we do | |
946 | * not miss some pages (e.g., because some other process has cleared TOWRITE | |
947 | * tag we set). The rule we follow is that TOWRITE tag can be cleared only | |
948 | * by the process clearing the DIRTY tag (and submitting the page for IO). | |
811d736f | 949 | */ |
0ea97180 MS |
950 | int write_cache_pages(struct address_space *mapping, |
951 | struct writeback_control *wbc, writepage_t writepage, | |
952 | void *data) | |
811d736f | 953 | { |
811d736f DH |
954 | int ret = 0; |
955 | int done = 0; | |
811d736f DH |
956 | struct pagevec pvec; |
957 | int nr_pages; | |
31a12666 | 958 | pgoff_t uninitialized_var(writeback_index); |
811d736f DH |
959 | pgoff_t index; |
960 | pgoff_t end; /* Inclusive */ | |
bd19e012 | 961 | pgoff_t done_index; |
31a12666 | 962 | int cycled; |
811d736f | 963 | int range_whole = 0; |
f446daae | 964 | int tag; |
811d736f | 965 | |
811d736f DH |
966 | pagevec_init(&pvec, 0); |
967 | if (wbc->range_cyclic) { | |
31a12666 NP |
968 | writeback_index = mapping->writeback_index; /* prev offset */ |
969 | index = writeback_index; | |
970 | if (index == 0) | |
971 | cycled = 1; | |
972 | else | |
973 | cycled = 0; | |
811d736f DH |
974 | end = -1; |
975 | } else { | |
976 | index = wbc->range_start >> PAGE_CACHE_SHIFT; | |
977 | end = wbc->range_end >> PAGE_CACHE_SHIFT; | |
978 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) | |
979 | range_whole = 1; | |
31a12666 | 980 | cycled = 1; /* ignore range_cyclic tests */ |
811d736f | 981 | } |
6e6938b6 | 982 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae JK |
983 | tag = PAGECACHE_TAG_TOWRITE; |
984 | else | |
985 | tag = PAGECACHE_TAG_DIRTY; | |
811d736f | 986 | retry: |
6e6938b6 | 987 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae | 988 | tag_pages_for_writeback(mapping, index, end); |
bd19e012 | 989 | done_index = index; |
5a3d5c98 NP |
990 | while (!done && (index <= end)) { |
991 | int i; | |
992 | ||
f446daae | 993 | nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag, |
5a3d5c98 NP |
994 | min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1); |
995 | if (nr_pages == 0) | |
996 | break; | |
811d736f | 997 | |
811d736f DH |
998 | for (i = 0; i < nr_pages; i++) { |
999 | struct page *page = pvec.pages[i]; | |
1000 | ||
1001 | /* | |
d5482cdf NP |
1002 | * At this point, the page may be truncated or |
1003 | * invalidated (changing page->mapping to NULL), or | |
1004 | * even swizzled back from swapper_space to tmpfs file | |
1005 | * mapping. However, page->index will not change | |
1006 | * because we have a reference on the page. | |
811d736f | 1007 | */ |
d5482cdf NP |
1008 | if (page->index > end) { |
1009 | /* | |
1010 | * can't be range_cyclic (1st pass) because | |
1011 | * end == -1 in that case. | |
1012 | */ | |
1013 | done = 1; | |
1014 | break; | |
1015 | } | |
1016 | ||
cf15b07c | 1017 | done_index = page->index; |
d5482cdf | 1018 | |
811d736f DH |
1019 | lock_page(page); |
1020 | ||
5a3d5c98 NP |
1021 | /* |
1022 | * Page truncated or invalidated. We can freely skip it | |
1023 | * then, even for data integrity operations: the page | |
1024 | * has disappeared concurrently, so there could be no | |
1025 | * real expectation of this data interity operation | |
1026 | * even if there is now a new, dirty page at the same | |
1027 | * pagecache address. | |
1028 | */ | |
811d736f | 1029 | if (unlikely(page->mapping != mapping)) { |
5a3d5c98 | 1030 | continue_unlock: |
811d736f DH |
1031 | unlock_page(page); |
1032 | continue; | |
1033 | } | |
1034 | ||
515f4a03 NP |
1035 | if (!PageDirty(page)) { |
1036 | /* someone wrote it for us */ | |
1037 | goto continue_unlock; | |
1038 | } | |
1039 | ||
1040 | if (PageWriteback(page)) { | |
1041 | if (wbc->sync_mode != WB_SYNC_NONE) | |
1042 | wait_on_page_writeback(page); | |
1043 | else | |
1044 | goto continue_unlock; | |
1045 | } | |
811d736f | 1046 | |
515f4a03 NP |
1047 | BUG_ON(PageWriteback(page)); |
1048 | if (!clear_page_dirty_for_io(page)) | |
5a3d5c98 | 1049 | goto continue_unlock; |
811d736f | 1050 | |
9e094383 | 1051 | trace_wbc_writepage(wbc, mapping->backing_dev_info); |
0ea97180 | 1052 | ret = (*writepage)(page, wbc, data); |
00266770 NP |
1053 | if (unlikely(ret)) { |
1054 | if (ret == AOP_WRITEPAGE_ACTIVATE) { | |
1055 | unlock_page(page); | |
1056 | ret = 0; | |
1057 | } else { | |
1058 | /* | |
1059 | * done_index is set past this page, | |
1060 | * so media errors will not choke | |
1061 | * background writeout for the entire | |
1062 | * file. This has consequences for | |
1063 | * range_cyclic semantics (ie. it may | |
1064 | * not be suitable for data integrity | |
1065 | * writeout). | |
1066 | */ | |
cf15b07c | 1067 | done_index = page->index + 1; |
00266770 NP |
1068 | done = 1; |
1069 | break; | |
1070 | } | |
0b564927 | 1071 | } |
00266770 | 1072 | |
546a1924 DC |
1073 | /* |
1074 | * We stop writing back only if we are not doing | |
1075 | * integrity sync. In case of integrity sync we have to | |
1076 | * keep going until we have written all the pages | |
1077 | * we tagged for writeback prior to entering this loop. | |
1078 | */ | |
1079 | if (--wbc->nr_to_write <= 0 && | |
1080 | wbc->sync_mode == WB_SYNC_NONE) { | |
1081 | done = 1; | |
1082 | break; | |
05fe478d | 1083 | } |
811d736f DH |
1084 | } |
1085 | pagevec_release(&pvec); | |
1086 | cond_resched(); | |
1087 | } | |
3a4c6800 | 1088 | if (!cycled && !done) { |
811d736f | 1089 | /* |
31a12666 | 1090 | * range_cyclic: |
811d736f DH |
1091 | * We hit the last page and there is more work to be done: wrap |
1092 | * back to the start of the file | |
1093 | */ | |
31a12666 | 1094 | cycled = 1; |
811d736f | 1095 | index = 0; |
31a12666 | 1096 | end = writeback_index - 1; |
811d736f DH |
1097 | goto retry; |
1098 | } | |
0b564927 DC |
1099 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) |
1100 | mapping->writeback_index = done_index; | |
06d6cf69 | 1101 | |
811d736f DH |
1102 | return ret; |
1103 | } | |
0ea97180 MS |
1104 | EXPORT_SYMBOL(write_cache_pages); |
1105 | ||
1106 | /* | |
1107 | * Function used by generic_writepages to call the real writepage | |
1108 | * function and set the mapping flags on error | |
1109 | */ | |
1110 | static int __writepage(struct page *page, struct writeback_control *wbc, | |
1111 | void *data) | |
1112 | { | |
1113 | struct address_space *mapping = data; | |
1114 | int ret = mapping->a_ops->writepage(page, wbc); | |
1115 | mapping_set_error(mapping, ret); | |
1116 | return ret; | |
1117 | } | |
1118 | ||
1119 | /** | |
1120 | * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. | |
1121 | * @mapping: address space structure to write | |
1122 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
1123 | * | |
1124 | * This is a library function, which implements the writepages() | |
1125 | * address_space_operation. | |
1126 | */ | |
1127 | int generic_writepages(struct address_space *mapping, | |
1128 | struct writeback_control *wbc) | |
1129 | { | |
9b6096a6 SL |
1130 | struct blk_plug plug; |
1131 | int ret; | |
1132 | ||
0ea97180 MS |
1133 | /* deal with chardevs and other special file */ |
1134 | if (!mapping->a_ops->writepage) | |
1135 | return 0; | |
1136 | ||
9b6096a6 SL |
1137 | blk_start_plug(&plug); |
1138 | ret = write_cache_pages(mapping, wbc, __writepage, mapping); | |
1139 | blk_finish_plug(&plug); | |
1140 | return ret; | |
0ea97180 | 1141 | } |
811d736f DH |
1142 | |
1143 | EXPORT_SYMBOL(generic_writepages); | |
1144 | ||
1da177e4 LT |
1145 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
1146 | { | |
22905f77 AM |
1147 | int ret; |
1148 | ||
1da177e4 LT |
1149 | if (wbc->nr_to_write <= 0) |
1150 | return 0; | |
1151 | if (mapping->a_ops->writepages) | |
d08b3851 | 1152 | ret = mapping->a_ops->writepages(mapping, wbc); |
22905f77 AM |
1153 | else |
1154 | ret = generic_writepages(mapping, wbc); | |
22905f77 | 1155 | return ret; |
1da177e4 LT |
1156 | } |
1157 | ||
1158 | /** | |
1159 | * write_one_page - write out a single page and optionally wait on I/O | |
67be2dd1 MW |
1160 | * @page: the page to write |
1161 | * @wait: if true, wait on writeout | |
1da177e4 LT |
1162 | * |
1163 | * The page must be locked by the caller and will be unlocked upon return. | |
1164 | * | |
1165 | * write_one_page() returns a negative error code if I/O failed. | |
1166 | */ | |
1167 | int write_one_page(struct page *page, int wait) | |
1168 | { | |
1169 | struct address_space *mapping = page->mapping; | |
1170 | int ret = 0; | |
1171 | struct writeback_control wbc = { | |
1172 | .sync_mode = WB_SYNC_ALL, | |
1173 | .nr_to_write = 1, | |
1174 | }; | |
1175 | ||
1176 | BUG_ON(!PageLocked(page)); | |
1177 | ||
1178 | if (wait) | |
1179 | wait_on_page_writeback(page); | |
1180 | ||
1181 | if (clear_page_dirty_for_io(page)) { | |
1182 | page_cache_get(page); | |
1183 | ret = mapping->a_ops->writepage(page, &wbc); | |
1184 | if (ret == 0 && wait) { | |
1185 | wait_on_page_writeback(page); | |
1186 | if (PageError(page)) | |
1187 | ret = -EIO; | |
1188 | } | |
1189 | page_cache_release(page); | |
1190 | } else { | |
1191 | unlock_page(page); | |
1192 | } | |
1193 | return ret; | |
1194 | } | |
1195 | EXPORT_SYMBOL(write_one_page); | |
1196 | ||
76719325 KC |
1197 | /* |
1198 | * For address_spaces which do not use buffers nor write back. | |
1199 | */ | |
1200 | int __set_page_dirty_no_writeback(struct page *page) | |
1201 | { | |
1202 | if (!PageDirty(page)) | |
c3f0da63 | 1203 | return !TestSetPageDirty(page); |
76719325 KC |
1204 | return 0; |
1205 | } | |
1206 | ||
e3a7cca1 ES |
1207 | /* |
1208 | * Helper function for set_page_dirty family. | |
1209 | * NOTE: This relies on being atomic wrt interrupts. | |
1210 | */ | |
1211 | void account_page_dirtied(struct page *page, struct address_space *mapping) | |
1212 | { | |
1213 | if (mapping_cap_account_dirty(mapping)) { | |
1214 | __inc_zone_page_state(page, NR_FILE_DIRTY); | |
ea941f0e | 1215 | __inc_zone_page_state(page, NR_DIRTIED); |
e3a7cca1 ES |
1216 | __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE); |
1217 | task_dirty_inc(current); | |
1218 | task_io_account_write(PAGE_CACHE_SIZE); | |
1219 | } | |
1220 | } | |
679ceace | 1221 | EXPORT_SYMBOL(account_page_dirtied); |
e3a7cca1 | 1222 | |
f629d1c9 MR |
1223 | /* |
1224 | * Helper function for set_page_writeback family. | |
1225 | * NOTE: Unlike account_page_dirtied this does not rely on being atomic | |
1226 | * wrt interrupts. | |
1227 | */ | |
1228 | void account_page_writeback(struct page *page) | |
1229 | { | |
1230 | inc_zone_page_state(page, NR_WRITEBACK); | |
ea941f0e | 1231 | inc_zone_page_state(page, NR_WRITTEN); |
f629d1c9 MR |
1232 | } |
1233 | EXPORT_SYMBOL(account_page_writeback); | |
1234 | ||
1da177e4 LT |
1235 | /* |
1236 | * For address_spaces which do not use buffers. Just tag the page as dirty in | |
1237 | * its radix tree. | |
1238 | * | |
1239 | * This is also used when a single buffer is being dirtied: we want to set the | |
1240 | * page dirty in that case, but not all the buffers. This is a "bottom-up" | |
1241 | * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying. | |
1242 | * | |
1243 | * Most callers have locked the page, which pins the address_space in memory. | |
1244 | * But zap_pte_range() does not lock the page, however in that case the | |
1245 | * mapping is pinned by the vma's ->vm_file reference. | |
1246 | * | |
1247 | * We take care to handle the case where the page was truncated from the | |
183ff22b | 1248 | * mapping by re-checking page_mapping() inside tree_lock. |
1da177e4 LT |
1249 | */ |
1250 | int __set_page_dirty_nobuffers(struct page *page) | |
1251 | { | |
1da177e4 LT |
1252 | if (!TestSetPageDirty(page)) { |
1253 | struct address_space *mapping = page_mapping(page); | |
1254 | struct address_space *mapping2; | |
1255 | ||
8c08540f AM |
1256 | if (!mapping) |
1257 | return 1; | |
1258 | ||
19fd6231 | 1259 | spin_lock_irq(&mapping->tree_lock); |
8c08540f AM |
1260 | mapping2 = page_mapping(page); |
1261 | if (mapping2) { /* Race with truncate? */ | |
1262 | BUG_ON(mapping2 != mapping); | |
787d2214 | 1263 | WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page)); |
e3a7cca1 | 1264 | account_page_dirtied(page, mapping); |
8c08540f AM |
1265 | radix_tree_tag_set(&mapping->page_tree, |
1266 | page_index(page), PAGECACHE_TAG_DIRTY); | |
1267 | } | |
19fd6231 | 1268 | spin_unlock_irq(&mapping->tree_lock); |
8c08540f AM |
1269 | if (mapping->host) { |
1270 | /* !PageAnon && !swapper_space */ | |
1271 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 1272 | } |
4741c9fd | 1273 | return 1; |
1da177e4 | 1274 | } |
4741c9fd | 1275 | return 0; |
1da177e4 LT |
1276 | } |
1277 | EXPORT_SYMBOL(__set_page_dirty_nobuffers); | |
1278 | ||
1279 | /* | |
1280 | * When a writepage implementation decides that it doesn't want to write this | |
1281 | * page for some reason, it should redirty the locked page via | |
1282 | * redirty_page_for_writepage() and it should then unlock the page and return 0 | |
1283 | */ | |
1284 | int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page) | |
1285 | { | |
1286 | wbc->pages_skipped++; | |
1287 | return __set_page_dirty_nobuffers(page); | |
1288 | } | |
1289 | EXPORT_SYMBOL(redirty_page_for_writepage); | |
1290 | ||
1291 | /* | |
6746aff7 WF |
1292 | * Dirty a page. |
1293 | * | |
1294 | * For pages with a mapping this should be done under the page lock | |
1295 | * for the benefit of asynchronous memory errors who prefer a consistent | |
1296 | * dirty state. This rule can be broken in some special cases, | |
1297 | * but should be better not to. | |
1298 | * | |
1da177e4 LT |
1299 | * If the mapping doesn't provide a set_page_dirty a_op, then |
1300 | * just fall through and assume that it wants buffer_heads. | |
1301 | */ | |
1cf6e7d8 | 1302 | int set_page_dirty(struct page *page) |
1da177e4 LT |
1303 | { |
1304 | struct address_space *mapping = page_mapping(page); | |
1305 | ||
1306 | if (likely(mapping)) { | |
1307 | int (*spd)(struct page *) = mapping->a_ops->set_page_dirty; | |
278df9f4 MK |
1308 | /* |
1309 | * readahead/lru_deactivate_page could remain | |
1310 | * PG_readahead/PG_reclaim due to race with end_page_writeback | |
1311 | * About readahead, if the page is written, the flags would be | |
1312 | * reset. So no problem. | |
1313 | * About lru_deactivate_page, if the page is redirty, the flag | |
1314 | * will be reset. So no problem. but if the page is used by readahead | |
1315 | * it will confuse readahead and make it restart the size rampup | |
1316 | * process. But it's a trivial problem. | |
1317 | */ | |
1318 | ClearPageReclaim(page); | |
9361401e DH |
1319 | #ifdef CONFIG_BLOCK |
1320 | if (!spd) | |
1321 | spd = __set_page_dirty_buffers; | |
1322 | #endif | |
1323 | return (*spd)(page); | |
1da177e4 | 1324 | } |
4741c9fd AM |
1325 | if (!PageDirty(page)) { |
1326 | if (!TestSetPageDirty(page)) | |
1327 | return 1; | |
1328 | } | |
1da177e4 LT |
1329 | return 0; |
1330 | } | |
1331 | EXPORT_SYMBOL(set_page_dirty); | |
1332 | ||
1333 | /* | |
1334 | * set_page_dirty() is racy if the caller has no reference against | |
1335 | * page->mapping->host, and if the page is unlocked. This is because another | |
1336 | * CPU could truncate the page off the mapping and then free the mapping. | |
1337 | * | |
1338 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
1339 | * holds a reference on the inode by having an open file. | |
1340 | * | |
1341 | * In other cases, the page should be locked before running set_page_dirty(). | |
1342 | */ | |
1343 | int set_page_dirty_lock(struct page *page) | |
1344 | { | |
1345 | int ret; | |
1346 | ||
7eaceacc | 1347 | lock_page(page); |
1da177e4 LT |
1348 | ret = set_page_dirty(page); |
1349 | unlock_page(page); | |
1350 | return ret; | |
1351 | } | |
1352 | EXPORT_SYMBOL(set_page_dirty_lock); | |
1353 | ||
1da177e4 LT |
1354 | /* |
1355 | * Clear a page's dirty flag, while caring for dirty memory accounting. | |
1356 | * Returns true if the page was previously dirty. | |
1357 | * | |
1358 | * This is for preparing to put the page under writeout. We leave the page | |
1359 | * tagged as dirty in the radix tree so that a concurrent write-for-sync | |
1360 | * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage | |
1361 | * implementation will run either set_page_writeback() or set_page_dirty(), | |
1362 | * at which stage we bring the page's dirty flag and radix-tree dirty tag | |
1363 | * back into sync. | |
1364 | * | |
1365 | * This incoherency between the page's dirty flag and radix-tree tag is | |
1366 | * unfortunate, but it only exists while the page is locked. | |
1367 | */ | |
1368 | int clear_page_dirty_for_io(struct page *page) | |
1369 | { | |
1370 | struct address_space *mapping = page_mapping(page); | |
1371 | ||
79352894 NP |
1372 | BUG_ON(!PageLocked(page)); |
1373 | ||
7658cc28 LT |
1374 | if (mapping && mapping_cap_account_dirty(mapping)) { |
1375 | /* | |
1376 | * Yes, Virginia, this is indeed insane. | |
1377 | * | |
1378 | * We use this sequence to make sure that | |
1379 | * (a) we account for dirty stats properly | |
1380 | * (b) we tell the low-level filesystem to | |
1381 | * mark the whole page dirty if it was | |
1382 | * dirty in a pagetable. Only to then | |
1383 | * (c) clean the page again and return 1 to | |
1384 | * cause the writeback. | |
1385 | * | |
1386 | * This way we avoid all nasty races with the | |
1387 | * dirty bit in multiple places and clearing | |
1388 | * them concurrently from different threads. | |
1389 | * | |
1390 | * Note! Normally the "set_page_dirty(page)" | |
1391 | * has no effect on the actual dirty bit - since | |
1392 | * that will already usually be set. But we | |
1393 | * need the side effects, and it can help us | |
1394 | * avoid races. | |
1395 | * | |
1396 | * We basically use the page "master dirty bit" | |
1397 | * as a serialization point for all the different | |
1398 | * threads doing their things. | |
7658cc28 LT |
1399 | */ |
1400 | if (page_mkclean(page)) | |
1401 | set_page_dirty(page); | |
79352894 NP |
1402 | /* |
1403 | * We carefully synchronise fault handlers against | |
1404 | * installing a dirty pte and marking the page dirty | |
1405 | * at this point. We do this by having them hold the | |
1406 | * page lock at some point after installing their | |
1407 | * pte, but before marking the page dirty. | |
1408 | * Pages are always locked coming in here, so we get | |
1409 | * the desired exclusion. See mm/memory.c:do_wp_page() | |
1410 | * for more comments. | |
1411 | */ | |
7658cc28 | 1412 | if (TestClearPageDirty(page)) { |
8c08540f | 1413 | dec_zone_page_state(page, NR_FILE_DIRTY); |
c9e51e41 PZ |
1414 | dec_bdi_stat(mapping->backing_dev_info, |
1415 | BDI_RECLAIMABLE); | |
7658cc28 | 1416 | return 1; |
1da177e4 | 1417 | } |
7658cc28 | 1418 | return 0; |
1da177e4 | 1419 | } |
7658cc28 | 1420 | return TestClearPageDirty(page); |
1da177e4 | 1421 | } |
58bb01a9 | 1422 | EXPORT_SYMBOL(clear_page_dirty_for_io); |
1da177e4 LT |
1423 | |
1424 | int test_clear_page_writeback(struct page *page) | |
1425 | { | |
1426 | struct address_space *mapping = page_mapping(page); | |
1427 | int ret; | |
1428 | ||
1429 | if (mapping) { | |
69cb51d1 | 1430 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
1da177e4 LT |
1431 | unsigned long flags; |
1432 | ||
19fd6231 | 1433 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 1434 | ret = TestClearPageWriteback(page); |
69cb51d1 | 1435 | if (ret) { |
1da177e4 LT |
1436 | radix_tree_tag_clear(&mapping->page_tree, |
1437 | page_index(page), | |
1438 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 1439 | if (bdi_cap_account_writeback(bdi)) { |
69cb51d1 | 1440 | __dec_bdi_stat(bdi, BDI_WRITEBACK); |
04fbfdc1 PZ |
1441 | __bdi_writeout_inc(bdi); |
1442 | } | |
69cb51d1 | 1443 | } |
19fd6231 | 1444 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
1445 | } else { |
1446 | ret = TestClearPageWriteback(page); | |
1447 | } | |
d688abf5 AM |
1448 | if (ret) |
1449 | dec_zone_page_state(page, NR_WRITEBACK); | |
1da177e4 LT |
1450 | return ret; |
1451 | } | |
1452 | ||
1453 | int test_set_page_writeback(struct page *page) | |
1454 | { | |
1455 | struct address_space *mapping = page_mapping(page); | |
1456 | int ret; | |
1457 | ||
1458 | if (mapping) { | |
69cb51d1 | 1459 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
1da177e4 LT |
1460 | unsigned long flags; |
1461 | ||
19fd6231 | 1462 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 1463 | ret = TestSetPageWriteback(page); |
69cb51d1 | 1464 | if (!ret) { |
1da177e4 LT |
1465 | radix_tree_tag_set(&mapping->page_tree, |
1466 | page_index(page), | |
1467 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 1468 | if (bdi_cap_account_writeback(bdi)) |
69cb51d1 PZ |
1469 | __inc_bdi_stat(bdi, BDI_WRITEBACK); |
1470 | } | |
1da177e4 LT |
1471 | if (!PageDirty(page)) |
1472 | radix_tree_tag_clear(&mapping->page_tree, | |
1473 | page_index(page), | |
1474 | PAGECACHE_TAG_DIRTY); | |
f446daae JK |
1475 | radix_tree_tag_clear(&mapping->page_tree, |
1476 | page_index(page), | |
1477 | PAGECACHE_TAG_TOWRITE); | |
19fd6231 | 1478 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
1479 | } else { |
1480 | ret = TestSetPageWriteback(page); | |
1481 | } | |
d688abf5 | 1482 | if (!ret) |
f629d1c9 | 1483 | account_page_writeback(page); |
1da177e4 LT |
1484 | return ret; |
1485 | ||
1486 | } | |
1487 | EXPORT_SYMBOL(test_set_page_writeback); | |
1488 | ||
1489 | /* | |
00128188 | 1490 | * Return true if any of the pages in the mapping are marked with the |
1da177e4 LT |
1491 | * passed tag. |
1492 | */ | |
1493 | int mapping_tagged(struct address_space *mapping, int tag) | |
1494 | { | |
1da177e4 | 1495 | int ret; |
00128188 | 1496 | rcu_read_lock(); |
1da177e4 | 1497 | ret = radix_tree_tagged(&mapping->page_tree, tag); |
00128188 | 1498 | rcu_read_unlock(); |
1da177e4 LT |
1499 | return ret; |
1500 | } | |
1501 | EXPORT_SYMBOL(mapping_tagged); |