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> | |
b95f1b31 | 15 | #include <linux/export.h> |
1da177e4 LT |
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> | |
ff01bb48 | 35 | #include <linux/buffer_head.h> /* __set_page_dirty_buffers */ |
811d736f | 36 | #include <linux/pagevec.h> |
eb608e3a | 37 | #include <linux/timer.h> |
8bd75c77 | 38 | #include <linux/sched/rt.h> |
6e543d57 | 39 | #include <linux/mm_inline.h> |
028c2dd1 | 40 | #include <trace/events/writeback.h> |
1da177e4 | 41 | |
6e543d57 LD |
42 | #include "internal.h" |
43 | ||
ffd1f609 WF |
44 | /* |
45 | * Sleep at most 200ms at a time in balance_dirty_pages(). | |
46 | */ | |
47 | #define MAX_PAUSE max(HZ/5, 1) | |
48 | ||
5b9b3574 WF |
49 | /* |
50 | * Try to keep balance_dirty_pages() call intervals higher than this many pages | |
51 | * by raising pause time to max_pause when falls below it. | |
52 | */ | |
53 | #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10)) | |
54 | ||
e98be2d5 WF |
55 | /* |
56 | * Estimate write bandwidth at 200ms intervals. | |
57 | */ | |
58 | #define BANDWIDTH_INTERVAL max(HZ/5, 1) | |
59 | ||
6c14ae1e WF |
60 | #define RATELIMIT_CALC_SHIFT 10 |
61 | ||
1da177e4 LT |
62 | /* |
63 | * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited | |
64 | * will look to see if it needs to force writeback or throttling. | |
65 | */ | |
66 | static long ratelimit_pages = 32; | |
67 | ||
1da177e4 LT |
68 | /* The following parameters are exported via /proc/sys/vm */ |
69 | ||
70 | /* | |
5b0830cb | 71 | * Start background writeback (via writeback threads) at this percentage |
1da177e4 | 72 | */ |
1b5e62b4 | 73 | int dirty_background_ratio = 10; |
1da177e4 | 74 | |
2da02997 DR |
75 | /* |
76 | * dirty_background_bytes starts at 0 (disabled) so that it is a function of | |
77 | * dirty_background_ratio * the amount of dirtyable memory | |
78 | */ | |
79 | unsigned long dirty_background_bytes; | |
80 | ||
195cf453 BG |
81 | /* |
82 | * free highmem will not be subtracted from the total free memory | |
83 | * for calculating free ratios if vm_highmem_is_dirtyable is true | |
84 | */ | |
85 | int vm_highmem_is_dirtyable; | |
86 | ||
1da177e4 LT |
87 | /* |
88 | * The generator of dirty data starts writeback at this percentage | |
89 | */ | |
1b5e62b4 | 90 | int vm_dirty_ratio = 20; |
1da177e4 | 91 | |
2da02997 DR |
92 | /* |
93 | * vm_dirty_bytes starts at 0 (disabled) so that it is a function of | |
94 | * vm_dirty_ratio * the amount of dirtyable memory | |
95 | */ | |
96 | unsigned long vm_dirty_bytes; | |
97 | ||
1da177e4 | 98 | /* |
704503d8 | 99 | * The interval between `kupdate'-style writebacks |
1da177e4 | 100 | */ |
22ef37ee | 101 | unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */ |
1da177e4 | 102 | |
91913a29 AB |
103 | EXPORT_SYMBOL_GPL(dirty_writeback_interval); |
104 | ||
1da177e4 | 105 | /* |
704503d8 | 106 | * The longest time for which data is allowed to remain dirty |
1da177e4 | 107 | */ |
22ef37ee | 108 | unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */ |
1da177e4 LT |
109 | |
110 | /* | |
111 | * Flag that makes the machine dump writes/reads and block dirtyings. | |
112 | */ | |
113 | int block_dump; | |
114 | ||
115 | /* | |
ed5b43f1 BS |
116 | * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: |
117 | * a full sync is triggered after this time elapses without any disk activity. | |
1da177e4 LT |
118 | */ |
119 | int laptop_mode; | |
120 | ||
121 | EXPORT_SYMBOL(laptop_mode); | |
122 | ||
123 | /* End of sysctl-exported parameters */ | |
124 | ||
c42843f2 | 125 | unsigned long global_dirty_limit; |
1da177e4 | 126 | |
04fbfdc1 PZ |
127 | /* |
128 | * Scale the writeback cache size proportional to the relative writeout speeds. | |
129 | * | |
130 | * We do this by keeping a floating proportion between BDIs, based on page | |
131 | * writeback completions [end_page_writeback()]. Those devices that write out | |
132 | * pages fastest will get the larger share, while the slower will get a smaller | |
133 | * share. | |
134 | * | |
135 | * We use page writeout completions because we are interested in getting rid of | |
136 | * dirty pages. Having them written out is the primary goal. | |
137 | * | |
138 | * We introduce a concept of time, a period over which we measure these events, | |
139 | * because demand can/will vary over time. The length of this period itself is | |
140 | * measured in page writeback completions. | |
141 | * | |
142 | */ | |
eb608e3a JK |
143 | static struct fprop_global writeout_completions; |
144 | ||
145 | static void writeout_period(unsigned long t); | |
146 | /* Timer for aging of writeout_completions */ | |
147 | static struct timer_list writeout_period_timer = | |
148 | TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0); | |
149 | static unsigned long writeout_period_time = 0; | |
150 | ||
151 | /* | |
152 | * Length of period for aging writeout fractions of bdis. This is an | |
153 | * arbitrarily chosen number. The longer the period, the slower fractions will | |
154 | * reflect changes in current writeout rate. | |
155 | */ | |
156 | #define VM_COMPLETIONS_PERIOD_LEN (3*HZ) | |
04fbfdc1 | 157 | |
a756cf59 JW |
158 | /* |
159 | * In a memory zone, there is a certain amount of pages we consider | |
160 | * available for the page cache, which is essentially the number of | |
161 | * free and reclaimable pages, minus some zone reserves to protect | |
162 | * lowmem and the ability to uphold the zone's watermarks without | |
163 | * requiring writeback. | |
164 | * | |
165 | * This number of dirtyable pages is the base value of which the | |
166 | * user-configurable dirty ratio is the effictive number of pages that | |
167 | * are allowed to be actually dirtied. Per individual zone, or | |
168 | * globally by using the sum of dirtyable pages over all zones. | |
169 | * | |
170 | * Because the user is allowed to specify the dirty limit globally as | |
171 | * absolute number of bytes, calculating the per-zone dirty limit can | |
172 | * require translating the configured limit into a percentage of | |
173 | * global dirtyable memory first. | |
174 | */ | |
175 | ||
a804552b JW |
176 | /** |
177 | * zone_dirtyable_memory - number of dirtyable pages in a zone | |
178 | * @zone: the zone | |
179 | * | |
180 | * Returns the zone's number of pages potentially available for dirty | |
181 | * page cache. This is the base value for the per-zone dirty limits. | |
182 | */ | |
183 | static unsigned long zone_dirtyable_memory(struct zone *zone) | |
184 | { | |
185 | unsigned long nr_pages; | |
186 | ||
187 | nr_pages = zone_page_state(zone, NR_FREE_PAGES); | |
188 | nr_pages -= min(nr_pages, zone->dirty_balance_reserve); | |
189 | ||
a1c3bfb2 JW |
190 | nr_pages += zone_page_state(zone, NR_INACTIVE_FILE); |
191 | nr_pages += zone_page_state(zone, NR_ACTIVE_FILE); | |
a804552b JW |
192 | |
193 | return nr_pages; | |
194 | } | |
195 | ||
1edf2234 JW |
196 | static unsigned long highmem_dirtyable_memory(unsigned long total) |
197 | { | |
198 | #ifdef CONFIG_HIGHMEM | |
199 | int node; | |
200 | unsigned long x = 0; | |
201 | ||
202 | for_each_node_state(node, N_HIGH_MEMORY) { | |
a804552b | 203 | struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM]; |
1edf2234 | 204 | |
a804552b | 205 | x += zone_dirtyable_memory(z); |
1edf2234 | 206 | } |
c8b74c2f SR |
207 | /* |
208 | * Unreclaimable memory (kernel memory or anonymous memory | |
209 | * without swap) can bring down the dirtyable pages below | |
210 | * the zone's dirty balance reserve and the above calculation | |
211 | * will underflow. However we still want to add in nodes | |
212 | * which are below threshold (negative values) to get a more | |
213 | * accurate calculation but make sure that the total never | |
214 | * underflows. | |
215 | */ | |
216 | if ((long)x < 0) | |
217 | x = 0; | |
218 | ||
1edf2234 JW |
219 | /* |
220 | * Make sure that the number of highmem pages is never larger | |
221 | * than the number of the total dirtyable memory. This can only | |
222 | * occur in very strange VM situations but we want to make sure | |
223 | * that this does not occur. | |
224 | */ | |
225 | return min(x, total); | |
226 | #else | |
227 | return 0; | |
228 | #endif | |
229 | } | |
230 | ||
231 | /** | |
ccafa287 | 232 | * global_dirtyable_memory - number of globally dirtyable pages |
1edf2234 | 233 | * |
ccafa287 JW |
234 | * Returns the global number of pages potentially available for dirty |
235 | * page cache. This is the base value for the global dirty limits. | |
1edf2234 | 236 | */ |
18cf8cf8 | 237 | static unsigned long global_dirtyable_memory(void) |
1edf2234 JW |
238 | { |
239 | unsigned long x; | |
240 | ||
a804552b | 241 | x = global_page_state(NR_FREE_PAGES); |
c8b74c2f | 242 | x -= min(x, dirty_balance_reserve); |
1edf2234 | 243 | |
a1c3bfb2 JW |
244 | x += global_page_state(NR_INACTIVE_FILE); |
245 | x += global_page_state(NR_ACTIVE_FILE); | |
a804552b | 246 | |
1edf2234 JW |
247 | if (!vm_highmem_is_dirtyable) |
248 | x -= highmem_dirtyable_memory(x); | |
249 | ||
250 | return x + 1; /* Ensure that we never return 0 */ | |
251 | } | |
252 | ||
ccafa287 JW |
253 | /* |
254 | * global_dirty_limits - background-writeback and dirty-throttling thresholds | |
255 | * | |
256 | * Calculate the dirty thresholds based on sysctl parameters | |
257 | * - vm.dirty_background_ratio or vm.dirty_background_bytes | |
258 | * - vm.dirty_ratio or vm.dirty_bytes | |
259 | * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and | |
260 | * real-time tasks. | |
261 | */ | |
262 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) | |
263 | { | |
9ef0a0ff | 264 | const unsigned long available_memory = global_dirtyable_memory(); |
ccafa287 JW |
265 | unsigned long background; |
266 | unsigned long dirty; | |
ccafa287 JW |
267 | struct task_struct *tsk; |
268 | ||
ccafa287 JW |
269 | if (vm_dirty_bytes) |
270 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE); | |
271 | else | |
272 | dirty = (vm_dirty_ratio * available_memory) / 100; | |
273 | ||
274 | if (dirty_background_bytes) | |
275 | background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE); | |
276 | else | |
277 | background = (dirty_background_ratio * available_memory) / 100; | |
278 | ||
279 | if (background >= dirty) | |
280 | background = dirty / 2; | |
281 | tsk = current; | |
282 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) { | |
283 | background += background / 4; | |
284 | dirty += dirty / 4; | |
285 | } | |
286 | *pbackground = background; | |
287 | *pdirty = dirty; | |
288 | trace_global_dirty_state(background, dirty); | |
289 | } | |
290 | ||
a756cf59 JW |
291 | /** |
292 | * zone_dirty_limit - maximum number of dirty pages allowed in a zone | |
293 | * @zone: the zone | |
294 | * | |
295 | * Returns the maximum number of dirty pages allowed in a zone, based | |
296 | * on the zone's dirtyable memory. | |
297 | */ | |
298 | static unsigned long zone_dirty_limit(struct zone *zone) | |
299 | { | |
300 | unsigned long zone_memory = zone_dirtyable_memory(zone); | |
301 | struct task_struct *tsk = current; | |
302 | unsigned long dirty; | |
303 | ||
304 | if (vm_dirty_bytes) | |
305 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) * | |
306 | zone_memory / global_dirtyable_memory(); | |
307 | else | |
308 | dirty = vm_dirty_ratio * zone_memory / 100; | |
309 | ||
310 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) | |
311 | dirty += dirty / 4; | |
312 | ||
313 | return dirty; | |
314 | } | |
315 | ||
316 | /** | |
317 | * zone_dirty_ok - tells whether a zone is within its dirty limits | |
318 | * @zone: the zone to check | |
319 | * | |
320 | * Returns %true when the dirty pages in @zone are within the zone's | |
321 | * dirty limit, %false if the limit is exceeded. | |
322 | */ | |
323 | bool zone_dirty_ok(struct zone *zone) | |
324 | { | |
325 | unsigned long limit = zone_dirty_limit(zone); | |
326 | ||
327 | return zone_page_state(zone, NR_FILE_DIRTY) + | |
328 | zone_page_state(zone, NR_UNSTABLE_NFS) + | |
329 | zone_page_state(zone, NR_WRITEBACK) <= limit; | |
330 | } | |
331 | ||
2da02997 | 332 | int dirty_background_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 333 | void __user *buffer, size_t *lenp, |
2da02997 DR |
334 | loff_t *ppos) |
335 | { | |
336 | int ret; | |
337 | ||
8d65af78 | 338 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
339 | if (ret == 0 && write) |
340 | dirty_background_bytes = 0; | |
341 | return ret; | |
342 | } | |
343 | ||
344 | int dirty_background_bytes_handler(struct ctl_table *table, int write, | |
8d65af78 | 345 | void __user *buffer, size_t *lenp, |
2da02997 DR |
346 | loff_t *ppos) |
347 | { | |
348 | int ret; | |
349 | ||
8d65af78 | 350 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
351 | if (ret == 0 && write) |
352 | dirty_background_ratio = 0; | |
353 | return ret; | |
354 | } | |
355 | ||
04fbfdc1 | 356 | int dirty_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 357 | void __user *buffer, size_t *lenp, |
04fbfdc1 PZ |
358 | loff_t *ppos) |
359 | { | |
360 | int old_ratio = vm_dirty_ratio; | |
2da02997 DR |
361 | int ret; |
362 | ||
8d65af78 | 363 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
04fbfdc1 | 364 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
eb608e3a | 365 | writeback_set_ratelimit(); |
2da02997 DR |
366 | vm_dirty_bytes = 0; |
367 | } | |
368 | return ret; | |
369 | } | |
370 | ||
2da02997 | 371 | int dirty_bytes_handler(struct ctl_table *table, int write, |
8d65af78 | 372 | void __user *buffer, size_t *lenp, |
2da02997 DR |
373 | loff_t *ppos) |
374 | { | |
fc3501d4 | 375 | unsigned long old_bytes = vm_dirty_bytes; |
2da02997 DR |
376 | int ret; |
377 | ||
8d65af78 | 378 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 | 379 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
eb608e3a | 380 | writeback_set_ratelimit(); |
2da02997 | 381 | vm_dirty_ratio = 0; |
04fbfdc1 PZ |
382 | } |
383 | return ret; | |
384 | } | |
385 | ||
eb608e3a JK |
386 | static unsigned long wp_next_time(unsigned long cur_time) |
387 | { | |
388 | cur_time += VM_COMPLETIONS_PERIOD_LEN; | |
389 | /* 0 has a special meaning... */ | |
390 | if (!cur_time) | |
391 | return 1; | |
392 | return cur_time; | |
393 | } | |
394 | ||
04fbfdc1 PZ |
395 | /* |
396 | * Increment the BDI's writeout completion count and the global writeout | |
397 | * completion count. Called from test_clear_page_writeback(). | |
398 | */ | |
399 | static inline void __bdi_writeout_inc(struct backing_dev_info *bdi) | |
400 | { | |
f7d2b1ec | 401 | __inc_bdi_stat(bdi, BDI_WRITTEN); |
eb608e3a JK |
402 | __fprop_inc_percpu_max(&writeout_completions, &bdi->completions, |
403 | bdi->max_prop_frac); | |
404 | /* First event after period switching was turned off? */ | |
405 | if (!unlikely(writeout_period_time)) { | |
406 | /* | |
407 | * We can race with other __bdi_writeout_inc calls here but | |
408 | * it does not cause any harm since the resulting time when | |
409 | * timer will fire and what is in writeout_period_time will be | |
410 | * roughly the same. | |
411 | */ | |
412 | writeout_period_time = wp_next_time(jiffies); | |
413 | mod_timer(&writeout_period_timer, writeout_period_time); | |
414 | } | |
04fbfdc1 PZ |
415 | } |
416 | ||
dd5656e5 MS |
417 | void bdi_writeout_inc(struct backing_dev_info *bdi) |
418 | { | |
419 | unsigned long flags; | |
420 | ||
421 | local_irq_save(flags); | |
422 | __bdi_writeout_inc(bdi); | |
423 | local_irq_restore(flags); | |
424 | } | |
425 | EXPORT_SYMBOL_GPL(bdi_writeout_inc); | |
426 | ||
04fbfdc1 PZ |
427 | /* |
428 | * Obtain an accurate fraction of the BDI's portion. | |
429 | */ | |
430 | static void bdi_writeout_fraction(struct backing_dev_info *bdi, | |
431 | long *numerator, long *denominator) | |
432 | { | |
eb608e3a | 433 | fprop_fraction_percpu(&writeout_completions, &bdi->completions, |
04fbfdc1 | 434 | numerator, denominator); |
04fbfdc1 PZ |
435 | } |
436 | ||
eb608e3a JK |
437 | /* |
438 | * On idle system, we can be called long after we scheduled because we use | |
439 | * deferred timers so count with missed periods. | |
440 | */ | |
441 | static void writeout_period(unsigned long t) | |
442 | { | |
443 | int miss_periods = (jiffies - writeout_period_time) / | |
444 | VM_COMPLETIONS_PERIOD_LEN; | |
445 | ||
446 | if (fprop_new_period(&writeout_completions, miss_periods + 1)) { | |
447 | writeout_period_time = wp_next_time(writeout_period_time + | |
448 | miss_periods * VM_COMPLETIONS_PERIOD_LEN); | |
449 | mod_timer(&writeout_period_timer, writeout_period_time); | |
450 | } else { | |
451 | /* | |
452 | * Aging has zeroed all fractions. Stop wasting CPU on period | |
453 | * updates. | |
454 | */ | |
455 | writeout_period_time = 0; | |
456 | } | |
457 | } | |
458 | ||
189d3c4a | 459 | /* |
d08c429b JW |
460 | * bdi_min_ratio keeps the sum of the minimum dirty shares of all |
461 | * registered backing devices, which, for obvious reasons, can not | |
462 | * exceed 100%. | |
189d3c4a | 463 | */ |
189d3c4a PZ |
464 | static unsigned int bdi_min_ratio; |
465 | ||
466 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) | |
467 | { | |
468 | int ret = 0; | |
189d3c4a | 469 | |
cfc4ba53 | 470 | spin_lock_bh(&bdi_lock); |
a42dde04 | 471 | if (min_ratio > bdi->max_ratio) { |
189d3c4a | 472 | ret = -EINVAL; |
a42dde04 PZ |
473 | } else { |
474 | min_ratio -= bdi->min_ratio; | |
475 | if (bdi_min_ratio + min_ratio < 100) { | |
476 | bdi_min_ratio += min_ratio; | |
477 | bdi->min_ratio += min_ratio; | |
478 | } else { | |
479 | ret = -EINVAL; | |
480 | } | |
481 | } | |
cfc4ba53 | 482 | spin_unlock_bh(&bdi_lock); |
a42dde04 PZ |
483 | |
484 | return ret; | |
485 | } | |
486 | ||
487 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio) | |
488 | { | |
a42dde04 PZ |
489 | int ret = 0; |
490 | ||
491 | if (max_ratio > 100) | |
492 | return -EINVAL; | |
493 | ||
cfc4ba53 | 494 | spin_lock_bh(&bdi_lock); |
a42dde04 PZ |
495 | if (bdi->min_ratio > max_ratio) { |
496 | ret = -EINVAL; | |
497 | } else { | |
498 | bdi->max_ratio = max_ratio; | |
eb608e3a | 499 | bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100; |
a42dde04 | 500 | } |
cfc4ba53 | 501 | spin_unlock_bh(&bdi_lock); |
189d3c4a PZ |
502 | |
503 | return ret; | |
504 | } | |
a42dde04 | 505 | EXPORT_SYMBOL(bdi_set_max_ratio); |
189d3c4a | 506 | |
6c14ae1e WF |
507 | static unsigned long dirty_freerun_ceiling(unsigned long thresh, |
508 | unsigned long bg_thresh) | |
509 | { | |
510 | return (thresh + bg_thresh) / 2; | |
511 | } | |
512 | ||
ffd1f609 WF |
513 | static unsigned long hard_dirty_limit(unsigned long thresh) |
514 | { | |
515 | return max(thresh, global_dirty_limit); | |
516 | } | |
517 | ||
6f718656 | 518 | /** |
1babe183 | 519 | * bdi_dirty_limit - @bdi's share of dirty throttling threshold |
6f718656 WF |
520 | * @bdi: the backing_dev_info to query |
521 | * @dirty: global dirty limit in pages | |
1babe183 | 522 | * |
6f718656 WF |
523 | * Returns @bdi's dirty limit in pages. The term "dirty" in the context of |
524 | * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages. | |
aed21ad2 WF |
525 | * |
526 | * Note that balance_dirty_pages() will only seriously take it as a hard limit | |
527 | * when sleeping max_pause per page is not enough to keep the dirty pages under | |
528 | * control. For example, when the device is completely stalled due to some error | |
529 | * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key. | |
530 | * In the other normal situations, it acts more gently by throttling the tasks | |
531 | * more (rather than completely block them) when the bdi dirty pages go high. | |
1babe183 | 532 | * |
6f718656 | 533 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
1babe183 WF |
534 | * - starving fast devices |
535 | * - piling up dirty pages (that will take long time to sync) on slow devices | |
536 | * | |
537 | * The bdi's share of dirty limit will be adapting to its throughput and | |
538 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. | |
539 | */ | |
540 | unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty) | |
16c4042f WF |
541 | { |
542 | u64 bdi_dirty; | |
543 | long numerator, denominator; | |
04fbfdc1 | 544 | |
16c4042f WF |
545 | /* |
546 | * Calculate this BDI's share of the dirty ratio. | |
547 | */ | |
548 | bdi_writeout_fraction(bdi, &numerator, &denominator); | |
04fbfdc1 | 549 | |
16c4042f WF |
550 | bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100; |
551 | bdi_dirty *= numerator; | |
552 | do_div(bdi_dirty, denominator); | |
04fbfdc1 | 553 | |
16c4042f WF |
554 | bdi_dirty += (dirty * bdi->min_ratio) / 100; |
555 | if (bdi_dirty > (dirty * bdi->max_ratio) / 100) | |
556 | bdi_dirty = dirty * bdi->max_ratio / 100; | |
557 | ||
558 | return bdi_dirty; | |
1da177e4 LT |
559 | } |
560 | ||
5a537485 MP |
561 | /* |
562 | * setpoint - dirty 3 | |
563 | * f(dirty) := 1.0 + (----------------) | |
564 | * limit - setpoint | |
565 | * | |
566 | * it's a 3rd order polynomial that subjects to | |
567 | * | |
568 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast | |
569 | * (2) f(setpoint) = 1.0 => the balance point | |
570 | * (3) f(limit) = 0 => the hard limit | |
571 | * (4) df/dx <= 0 => negative feedback control | |
572 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) | |
573 | * => fast response on large errors; small oscillation near setpoint | |
574 | */ | |
d5c9fde3 | 575 | static long long pos_ratio_polynom(unsigned long setpoint, |
5a537485 MP |
576 | unsigned long dirty, |
577 | unsigned long limit) | |
578 | { | |
579 | long long pos_ratio; | |
580 | long x; | |
581 | ||
d5c9fde3 | 582 | x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT, |
5a537485 MP |
583 | limit - setpoint + 1); |
584 | pos_ratio = x; | |
585 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
586 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
587 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT; | |
588 | ||
589 | return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT); | |
590 | } | |
591 | ||
6c14ae1e WF |
592 | /* |
593 | * Dirty position control. | |
594 | * | |
595 | * (o) global/bdi setpoints | |
596 | * | |
597 | * We want the dirty pages be balanced around the global/bdi setpoints. | |
598 | * When the number of dirty pages is higher/lower than the setpoint, the | |
599 | * dirty position control ratio (and hence task dirty ratelimit) will be | |
600 | * decreased/increased to bring the dirty pages back to the setpoint. | |
601 | * | |
602 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT | |
603 | * | |
604 | * if (dirty < setpoint) scale up pos_ratio | |
605 | * if (dirty > setpoint) scale down pos_ratio | |
606 | * | |
607 | * if (bdi_dirty < bdi_setpoint) scale up pos_ratio | |
608 | * if (bdi_dirty > bdi_setpoint) scale down pos_ratio | |
609 | * | |
610 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT | |
611 | * | |
612 | * (o) global control line | |
613 | * | |
614 | * ^ pos_ratio | |
615 | * | | |
616 | * | |<===== global dirty control scope ======>| | |
617 | * 2.0 .............* | |
618 | * | .* | |
619 | * | . * | |
620 | * | . * | |
621 | * | . * | |
622 | * | . * | |
623 | * | . * | |
624 | * 1.0 ................................* | |
625 | * | . . * | |
626 | * | . . * | |
627 | * | . . * | |
628 | * | . . * | |
629 | * | . . * | |
630 | * 0 +------------.------------------.----------------------*-------------> | |
631 | * freerun^ setpoint^ limit^ dirty pages | |
632 | * | |
633 | * (o) bdi control line | |
634 | * | |
635 | * ^ pos_ratio | |
636 | * | | |
637 | * | * | |
638 | * | * | |
639 | * | * | |
640 | * | * | |
641 | * | * |<=========== span ============>| | |
642 | * 1.0 .......................* | |
643 | * | . * | |
644 | * | . * | |
645 | * | . * | |
646 | * | . * | |
647 | * | . * | |
648 | * | . * | |
649 | * | . * | |
650 | * | . * | |
651 | * | . * | |
652 | * | . * | |
653 | * | . * | |
654 | * 1/4 ...............................................* * * * * * * * * * * * | |
655 | * | . . | |
656 | * | . . | |
657 | * | . . | |
658 | * 0 +----------------------.-------------------------------.-------------> | |
659 | * bdi_setpoint^ x_intercept^ | |
660 | * | |
661 | * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can | |
662 | * be smoothly throttled down to normal if it starts high in situations like | |
663 | * - start writing to a slow SD card and a fast disk at the same time. The SD | |
664 | * card's bdi_dirty may rush to many times higher than bdi_setpoint. | |
665 | * - the bdi dirty thresh drops quickly due to change of JBOD workload | |
666 | */ | |
667 | static unsigned long bdi_position_ratio(struct backing_dev_info *bdi, | |
668 | unsigned long thresh, | |
669 | unsigned long bg_thresh, | |
670 | unsigned long dirty, | |
671 | unsigned long bdi_thresh, | |
672 | unsigned long bdi_dirty) | |
673 | { | |
674 | unsigned long write_bw = bdi->avg_write_bandwidth; | |
675 | unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh); | |
676 | unsigned long limit = hard_dirty_limit(thresh); | |
677 | unsigned long x_intercept; | |
678 | unsigned long setpoint; /* dirty pages' target balance point */ | |
679 | unsigned long bdi_setpoint; | |
680 | unsigned long span; | |
681 | long long pos_ratio; /* for scaling up/down the rate limit */ | |
682 | long x; | |
683 | ||
684 | if (unlikely(dirty >= limit)) | |
685 | return 0; | |
686 | ||
687 | /* | |
688 | * global setpoint | |
689 | * | |
5a537485 MP |
690 | * See comment for pos_ratio_polynom(). |
691 | */ | |
692 | setpoint = (freerun + limit) / 2; | |
693 | pos_ratio = pos_ratio_polynom(setpoint, dirty, limit); | |
694 | ||
695 | /* | |
696 | * The strictlimit feature is a tool preventing mistrusted filesystems | |
697 | * from growing a large number of dirty pages before throttling. For | |
698 | * such filesystems balance_dirty_pages always checks bdi counters | |
699 | * against bdi limits. Even if global "nr_dirty" is under "freerun". | |
700 | * This is especially important for fuse which sets bdi->max_ratio to | |
701 | * 1% by default. Without strictlimit feature, fuse writeback may | |
702 | * consume arbitrary amount of RAM because it is accounted in | |
703 | * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty". | |
6c14ae1e | 704 | * |
5a537485 MP |
705 | * Here, in bdi_position_ratio(), we calculate pos_ratio based on |
706 | * two values: bdi_dirty and bdi_thresh. Let's consider an example: | |
707 | * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global | |
708 | * limits are set by default to 10% and 20% (background and throttle). | |
709 | * Then bdi_thresh is 1% of 20% of 16GB. This amounts to ~8K pages. | |
710 | * bdi_dirty_limit(bdi, bg_thresh) is about ~4K pages. bdi_setpoint is | |
711 | * about ~6K pages (as the average of background and throttle bdi | |
712 | * limits). The 3rd order polynomial will provide positive feedback if | |
713 | * bdi_dirty is under bdi_setpoint and vice versa. | |
6c14ae1e | 714 | * |
5a537485 MP |
715 | * Note, that we cannot use global counters in these calculations |
716 | * because we want to throttle process writing to a strictlimit BDI | |
717 | * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB | |
718 | * in the example above). | |
6c14ae1e | 719 | */ |
5a537485 MP |
720 | if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
721 | long long bdi_pos_ratio; | |
722 | unsigned long bdi_bg_thresh; | |
723 | ||
724 | if (bdi_dirty < 8) | |
725 | return min_t(long long, pos_ratio * 2, | |
726 | 2 << RATELIMIT_CALC_SHIFT); | |
727 | ||
728 | if (bdi_dirty >= bdi_thresh) | |
729 | return 0; | |
730 | ||
731 | bdi_bg_thresh = div_u64((u64)bdi_thresh * bg_thresh, thresh); | |
732 | bdi_setpoint = dirty_freerun_ceiling(bdi_thresh, | |
733 | bdi_bg_thresh); | |
734 | ||
735 | if (bdi_setpoint == 0 || bdi_setpoint == bdi_thresh) | |
736 | return 0; | |
737 | ||
738 | bdi_pos_ratio = pos_ratio_polynom(bdi_setpoint, bdi_dirty, | |
739 | bdi_thresh); | |
740 | ||
741 | /* | |
742 | * Typically, for strictlimit case, bdi_setpoint << setpoint | |
743 | * and pos_ratio >> bdi_pos_ratio. In the other words global | |
744 | * state ("dirty") is not limiting factor and we have to | |
745 | * make decision based on bdi counters. But there is an | |
746 | * important case when global pos_ratio should get precedence: | |
747 | * global limits are exceeded (e.g. due to activities on other | |
748 | * BDIs) while given strictlimit BDI is below limit. | |
749 | * | |
750 | * "pos_ratio * bdi_pos_ratio" would work for the case above, | |
751 | * but it would look too non-natural for the case of all | |
752 | * activity in the system coming from a single strictlimit BDI | |
753 | * with bdi->max_ratio == 100%. | |
754 | * | |
755 | * Note that min() below somewhat changes the dynamics of the | |
756 | * control system. Normally, pos_ratio value can be well over 3 | |
757 | * (when globally we are at freerun and bdi is well below bdi | |
758 | * setpoint). Now the maximum pos_ratio in the same situation | |
759 | * is 2. We might want to tweak this if we observe the control | |
760 | * system is too slow to adapt. | |
761 | */ | |
762 | return min(pos_ratio, bdi_pos_ratio); | |
763 | } | |
6c14ae1e WF |
764 | |
765 | /* | |
766 | * We have computed basic pos_ratio above based on global situation. If | |
767 | * the bdi is over/under its share of dirty pages, we want to scale | |
768 | * pos_ratio further down/up. That is done by the following mechanism. | |
769 | */ | |
770 | ||
771 | /* | |
772 | * bdi setpoint | |
773 | * | |
774 | * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint) | |
775 | * | |
776 | * x_intercept - bdi_dirty | |
777 | * := -------------------------- | |
778 | * x_intercept - bdi_setpoint | |
779 | * | |
780 | * The main bdi control line is a linear function that subjects to | |
781 | * | |
782 | * (1) f(bdi_setpoint) = 1.0 | |
783 | * (2) k = - 1 / (8 * write_bw) (in single bdi case) | |
784 | * or equally: x_intercept = bdi_setpoint + 8 * write_bw | |
785 | * | |
786 | * For single bdi case, the dirty pages are observed to fluctuate | |
787 | * regularly within range | |
788 | * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2] | |
789 | * for various filesystems, where (2) can yield in a reasonable 12.5% | |
790 | * fluctuation range for pos_ratio. | |
791 | * | |
792 | * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its | |
793 | * own size, so move the slope over accordingly and choose a slope that | |
794 | * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh. | |
795 | */ | |
796 | if (unlikely(bdi_thresh > thresh)) | |
797 | bdi_thresh = thresh; | |
aed21ad2 WF |
798 | /* |
799 | * It's very possible that bdi_thresh is close to 0 not because the | |
800 | * device is slow, but that it has remained inactive for long time. | |
801 | * Honour such devices a reasonable good (hopefully IO efficient) | |
802 | * threshold, so that the occasional writes won't be blocked and active | |
803 | * writes can rampup the threshold quickly. | |
804 | */ | |
8927f66c | 805 | bdi_thresh = max(bdi_thresh, (limit - dirty) / 8); |
6c14ae1e WF |
806 | /* |
807 | * scale global setpoint to bdi's: | |
808 | * bdi_setpoint = setpoint * bdi_thresh / thresh | |
809 | */ | |
810 | x = div_u64((u64)bdi_thresh << 16, thresh + 1); | |
811 | bdi_setpoint = setpoint * (u64)x >> 16; | |
812 | /* | |
813 | * Use span=(8*write_bw) in single bdi case as indicated by | |
814 | * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case. | |
815 | * | |
816 | * bdi_thresh thresh - bdi_thresh | |
817 | * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh | |
818 | * thresh thresh | |
819 | */ | |
820 | span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16; | |
821 | x_intercept = bdi_setpoint + span; | |
822 | ||
823 | if (bdi_dirty < x_intercept - span / 4) { | |
d5c9fde3 | 824 | pos_ratio = div64_u64(pos_ratio * (x_intercept - bdi_dirty), |
50657fc4 | 825 | x_intercept - bdi_setpoint + 1); |
6c14ae1e WF |
826 | } else |
827 | pos_ratio /= 4; | |
828 | ||
8927f66c WF |
829 | /* |
830 | * bdi reserve area, safeguard against dirty pool underrun and disk idle | |
831 | * It may push the desired control point of global dirty pages higher | |
832 | * than setpoint. | |
833 | */ | |
834 | x_intercept = bdi_thresh / 2; | |
835 | if (bdi_dirty < x_intercept) { | |
50657fc4 WF |
836 | if (bdi_dirty > x_intercept / 8) |
837 | pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty); | |
838 | else | |
8927f66c WF |
839 | pos_ratio *= 8; |
840 | } | |
841 | ||
6c14ae1e WF |
842 | return pos_ratio; |
843 | } | |
844 | ||
e98be2d5 WF |
845 | static void bdi_update_write_bandwidth(struct backing_dev_info *bdi, |
846 | unsigned long elapsed, | |
847 | unsigned long written) | |
848 | { | |
849 | const unsigned long period = roundup_pow_of_two(3 * HZ); | |
850 | unsigned long avg = bdi->avg_write_bandwidth; | |
851 | unsigned long old = bdi->write_bandwidth; | |
852 | u64 bw; | |
853 | ||
854 | /* | |
855 | * bw = written * HZ / elapsed | |
856 | * | |
857 | * bw * elapsed + write_bandwidth * (period - elapsed) | |
858 | * write_bandwidth = --------------------------------------------------- | |
859 | * period | |
860 | */ | |
861 | bw = written - bdi->written_stamp; | |
862 | bw *= HZ; | |
863 | if (unlikely(elapsed > period)) { | |
864 | do_div(bw, elapsed); | |
865 | avg = bw; | |
866 | goto out; | |
867 | } | |
868 | bw += (u64)bdi->write_bandwidth * (period - elapsed); | |
869 | bw >>= ilog2(period); | |
870 | ||
871 | /* | |
872 | * one more level of smoothing, for filtering out sudden spikes | |
873 | */ | |
874 | if (avg > old && old >= (unsigned long)bw) | |
875 | avg -= (avg - old) >> 3; | |
876 | ||
877 | if (avg < old && old <= (unsigned long)bw) | |
878 | avg += (old - avg) >> 3; | |
879 | ||
880 | out: | |
881 | bdi->write_bandwidth = bw; | |
882 | bdi->avg_write_bandwidth = avg; | |
883 | } | |
884 | ||
c42843f2 WF |
885 | /* |
886 | * The global dirtyable memory and dirty threshold could be suddenly knocked | |
887 | * down by a large amount (eg. on the startup of KVM in a swapless system). | |
888 | * This may throw the system into deep dirty exceeded state and throttle | |
889 | * heavy/light dirtiers alike. To retain good responsiveness, maintain | |
890 | * global_dirty_limit for tracking slowly down to the knocked down dirty | |
891 | * threshold. | |
892 | */ | |
893 | static void update_dirty_limit(unsigned long thresh, unsigned long dirty) | |
894 | { | |
895 | unsigned long limit = global_dirty_limit; | |
896 | ||
897 | /* | |
898 | * Follow up in one step. | |
899 | */ | |
900 | if (limit < thresh) { | |
901 | limit = thresh; | |
902 | goto update; | |
903 | } | |
904 | ||
905 | /* | |
906 | * Follow down slowly. Use the higher one as the target, because thresh | |
907 | * may drop below dirty. This is exactly the reason to introduce | |
908 | * global_dirty_limit which is guaranteed to lie above the dirty pages. | |
909 | */ | |
910 | thresh = max(thresh, dirty); | |
911 | if (limit > thresh) { | |
912 | limit -= (limit - thresh) >> 5; | |
913 | goto update; | |
914 | } | |
915 | return; | |
916 | update: | |
917 | global_dirty_limit = limit; | |
918 | } | |
919 | ||
920 | static void global_update_bandwidth(unsigned long thresh, | |
921 | unsigned long dirty, | |
922 | unsigned long now) | |
923 | { | |
924 | static DEFINE_SPINLOCK(dirty_lock); | |
925 | static unsigned long update_time; | |
926 | ||
927 | /* | |
928 | * check locklessly first to optimize away locking for the most time | |
929 | */ | |
930 | if (time_before(now, update_time + BANDWIDTH_INTERVAL)) | |
931 | return; | |
932 | ||
933 | spin_lock(&dirty_lock); | |
934 | if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) { | |
935 | update_dirty_limit(thresh, dirty); | |
936 | update_time = now; | |
937 | } | |
938 | spin_unlock(&dirty_lock); | |
939 | } | |
940 | ||
be3ffa27 WF |
941 | /* |
942 | * Maintain bdi->dirty_ratelimit, the base dirty throttle rate. | |
943 | * | |
944 | * Normal bdi tasks will be curbed at or below it in long term. | |
945 | * Obviously it should be around (write_bw / N) when there are N dd tasks. | |
946 | */ | |
947 | static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi, | |
948 | unsigned long thresh, | |
949 | unsigned long bg_thresh, | |
950 | unsigned long dirty, | |
951 | unsigned long bdi_thresh, | |
952 | unsigned long bdi_dirty, | |
953 | unsigned long dirtied, | |
954 | unsigned long elapsed) | |
955 | { | |
7381131c WF |
956 | unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh); |
957 | unsigned long limit = hard_dirty_limit(thresh); | |
958 | unsigned long setpoint = (freerun + limit) / 2; | |
be3ffa27 WF |
959 | unsigned long write_bw = bdi->avg_write_bandwidth; |
960 | unsigned long dirty_ratelimit = bdi->dirty_ratelimit; | |
961 | unsigned long dirty_rate; | |
962 | unsigned long task_ratelimit; | |
963 | unsigned long balanced_dirty_ratelimit; | |
964 | unsigned long pos_ratio; | |
7381131c WF |
965 | unsigned long step; |
966 | unsigned long x; | |
be3ffa27 WF |
967 | |
968 | /* | |
969 | * The dirty rate will match the writeout rate in long term, except | |
970 | * when dirty pages are truncated by userspace or re-dirtied by FS. | |
971 | */ | |
972 | dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed; | |
973 | ||
974 | pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty, | |
975 | bdi_thresh, bdi_dirty); | |
976 | /* | |
977 | * task_ratelimit reflects each dd's dirty rate for the past 200ms. | |
978 | */ | |
979 | task_ratelimit = (u64)dirty_ratelimit * | |
980 | pos_ratio >> RATELIMIT_CALC_SHIFT; | |
981 | task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ | |
982 | ||
983 | /* | |
984 | * A linear estimation of the "balanced" throttle rate. The theory is, | |
985 | * if there are N dd tasks, each throttled at task_ratelimit, the bdi's | |
986 | * dirty_rate will be measured to be (N * task_ratelimit). So the below | |
987 | * formula will yield the balanced rate limit (write_bw / N). | |
988 | * | |
989 | * Note that the expanded form is not a pure rate feedback: | |
990 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) | |
991 | * but also takes pos_ratio into account: | |
992 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) | |
993 | * | |
994 | * (1) is not realistic because pos_ratio also takes part in balancing | |
995 | * the dirty rate. Consider the state | |
996 | * pos_ratio = 0.5 (3) | |
997 | * rate = 2 * (write_bw / N) (4) | |
998 | * If (1) is used, it will stuck in that state! Because each dd will | |
999 | * be throttled at | |
1000 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) | |
1001 | * yielding | |
1002 | * dirty_rate = N * task_ratelimit = write_bw (6) | |
1003 | * put (6) into (1) we get | |
1004 | * rate_(i+1) = rate_(i) (7) | |
1005 | * | |
1006 | * So we end up using (2) to always keep | |
1007 | * rate_(i+1) ~= (write_bw / N) (8) | |
1008 | * regardless of the value of pos_ratio. As long as (8) is satisfied, | |
1009 | * pos_ratio is able to drive itself to 1.0, which is not only where | |
1010 | * the dirty count meet the setpoint, but also where the slope of | |
1011 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated. | |
1012 | */ | |
1013 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, | |
1014 | dirty_rate | 1); | |
bdaac490 WF |
1015 | /* |
1016 | * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw | |
1017 | */ | |
1018 | if (unlikely(balanced_dirty_ratelimit > write_bw)) | |
1019 | balanced_dirty_ratelimit = write_bw; | |
be3ffa27 | 1020 | |
7381131c WF |
1021 | /* |
1022 | * We could safely do this and return immediately: | |
1023 | * | |
1024 | * bdi->dirty_ratelimit = balanced_dirty_ratelimit; | |
1025 | * | |
1026 | * However to get a more stable dirty_ratelimit, the below elaborated | |
331cbdee | 1027 | * code makes use of task_ratelimit to filter out singular points and |
7381131c WF |
1028 | * limit the step size. |
1029 | * | |
1030 | * The below code essentially only uses the relative value of | |
1031 | * | |
1032 | * task_ratelimit - dirty_ratelimit | |
1033 | * = (pos_ratio - 1) * dirty_ratelimit | |
1034 | * | |
1035 | * which reflects the direction and size of dirty position error. | |
1036 | */ | |
1037 | ||
1038 | /* | |
1039 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff | |
1040 | * task_ratelimit is on the same side of dirty_ratelimit, too. | |
1041 | * For example, when | |
1042 | * - dirty_ratelimit > balanced_dirty_ratelimit | |
1043 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) | |
1044 | * lowering dirty_ratelimit will help meet both the position and rate | |
1045 | * control targets. Otherwise, don't update dirty_ratelimit if it will | |
1046 | * only help meet the rate target. After all, what the users ultimately | |
1047 | * feel and care are stable dirty rate and small position error. | |
1048 | * | |
1049 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size | |
331cbdee | 1050 | * and filter out the singular points of balanced_dirty_ratelimit. Which |
7381131c WF |
1051 | * keeps jumping around randomly and can even leap far away at times |
1052 | * due to the small 200ms estimation period of dirty_rate (we want to | |
1053 | * keep that period small to reduce time lags). | |
1054 | */ | |
1055 | step = 0; | |
5a537485 MP |
1056 | |
1057 | /* | |
1058 | * For strictlimit case, calculations above were based on bdi counters | |
1059 | * and limits (starting from pos_ratio = bdi_position_ratio() and up to | |
1060 | * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate). | |
1061 | * Hence, to calculate "step" properly, we have to use bdi_dirty as | |
1062 | * "dirty" and bdi_setpoint as "setpoint". | |
1063 | * | |
1064 | * We rampup dirty_ratelimit forcibly if bdi_dirty is low because | |
1065 | * it's possible that bdi_thresh is close to zero due to inactivity | |
1066 | * of backing device (see the implementation of bdi_dirty_limit()). | |
1067 | */ | |
1068 | if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) { | |
1069 | dirty = bdi_dirty; | |
1070 | if (bdi_dirty < 8) | |
1071 | setpoint = bdi_dirty + 1; | |
1072 | else | |
1073 | setpoint = (bdi_thresh + | |
1074 | bdi_dirty_limit(bdi, bg_thresh)) / 2; | |
1075 | } | |
1076 | ||
7381131c | 1077 | if (dirty < setpoint) { |
7c809968 MR |
1078 | x = min3(bdi->balanced_dirty_ratelimit, |
1079 | balanced_dirty_ratelimit, task_ratelimit); | |
7381131c WF |
1080 | if (dirty_ratelimit < x) |
1081 | step = x - dirty_ratelimit; | |
1082 | } else { | |
7c809968 MR |
1083 | x = max3(bdi->balanced_dirty_ratelimit, |
1084 | balanced_dirty_ratelimit, task_ratelimit); | |
7381131c WF |
1085 | if (dirty_ratelimit > x) |
1086 | step = dirty_ratelimit - x; | |
1087 | } | |
1088 | ||
1089 | /* | |
1090 | * Don't pursue 100% rate matching. It's impossible since the balanced | |
1091 | * rate itself is constantly fluctuating. So decrease the track speed | |
1092 | * when it gets close to the target. Helps eliminate pointless tremors. | |
1093 | */ | |
1094 | step >>= dirty_ratelimit / (2 * step + 1); | |
1095 | /* | |
1096 | * Limit the tracking speed to avoid overshooting. | |
1097 | */ | |
1098 | step = (step + 7) / 8; | |
1099 | ||
1100 | if (dirty_ratelimit < balanced_dirty_ratelimit) | |
1101 | dirty_ratelimit += step; | |
1102 | else | |
1103 | dirty_ratelimit -= step; | |
1104 | ||
1105 | bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL); | |
1106 | bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit; | |
b48c104d WF |
1107 | |
1108 | trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit); | |
be3ffa27 WF |
1109 | } |
1110 | ||
e98be2d5 | 1111 | void __bdi_update_bandwidth(struct backing_dev_info *bdi, |
c42843f2 | 1112 | unsigned long thresh, |
af6a3113 | 1113 | unsigned long bg_thresh, |
c42843f2 WF |
1114 | unsigned long dirty, |
1115 | unsigned long bdi_thresh, | |
1116 | unsigned long bdi_dirty, | |
e98be2d5 WF |
1117 | unsigned long start_time) |
1118 | { | |
1119 | unsigned long now = jiffies; | |
1120 | unsigned long elapsed = now - bdi->bw_time_stamp; | |
be3ffa27 | 1121 | unsigned long dirtied; |
e98be2d5 WF |
1122 | unsigned long written; |
1123 | ||
1124 | /* | |
1125 | * rate-limit, only update once every 200ms. | |
1126 | */ | |
1127 | if (elapsed < BANDWIDTH_INTERVAL) | |
1128 | return; | |
1129 | ||
be3ffa27 | 1130 | dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]); |
e98be2d5 WF |
1131 | written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]); |
1132 | ||
1133 | /* | |
1134 | * Skip quiet periods when disk bandwidth is under-utilized. | |
1135 | * (at least 1s idle time between two flusher runs) | |
1136 | */ | |
1137 | if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time)) | |
1138 | goto snapshot; | |
1139 | ||
be3ffa27 | 1140 | if (thresh) { |
c42843f2 | 1141 | global_update_bandwidth(thresh, dirty, now); |
be3ffa27 WF |
1142 | bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty, |
1143 | bdi_thresh, bdi_dirty, | |
1144 | dirtied, elapsed); | |
1145 | } | |
e98be2d5 WF |
1146 | bdi_update_write_bandwidth(bdi, elapsed, written); |
1147 | ||
1148 | snapshot: | |
be3ffa27 | 1149 | bdi->dirtied_stamp = dirtied; |
e98be2d5 WF |
1150 | bdi->written_stamp = written; |
1151 | bdi->bw_time_stamp = now; | |
1152 | } | |
1153 | ||
1154 | static void bdi_update_bandwidth(struct backing_dev_info *bdi, | |
c42843f2 | 1155 | unsigned long thresh, |
af6a3113 | 1156 | unsigned long bg_thresh, |
c42843f2 WF |
1157 | unsigned long dirty, |
1158 | unsigned long bdi_thresh, | |
1159 | unsigned long bdi_dirty, | |
e98be2d5 WF |
1160 | unsigned long start_time) |
1161 | { | |
1162 | if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL)) | |
1163 | return; | |
1164 | spin_lock(&bdi->wb.list_lock); | |
af6a3113 WF |
1165 | __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty, |
1166 | bdi_thresh, bdi_dirty, start_time); | |
e98be2d5 WF |
1167 | spin_unlock(&bdi->wb.list_lock); |
1168 | } | |
1169 | ||
9d823e8f | 1170 | /* |
d0e1d66b | 1171 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited() |
9d823e8f WF |
1172 | * will look to see if it needs to start dirty throttling. |
1173 | * | |
1174 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive | |
1175 | * global_page_state() too often. So scale it near-sqrt to the safety margin | |
1176 | * (the number of pages we may dirty without exceeding the dirty limits). | |
1177 | */ | |
1178 | static unsigned long dirty_poll_interval(unsigned long dirty, | |
1179 | unsigned long thresh) | |
1180 | { | |
1181 | if (thresh > dirty) | |
1182 | return 1UL << (ilog2(thresh - dirty) >> 1); | |
1183 | ||
1184 | return 1; | |
1185 | } | |
1186 | ||
e3b6c655 FW |
1187 | static unsigned long bdi_max_pause(struct backing_dev_info *bdi, |
1188 | unsigned long bdi_dirty) | |
c8462cc9 | 1189 | { |
e3b6c655 FW |
1190 | unsigned long bw = bdi->avg_write_bandwidth; |
1191 | unsigned long t; | |
c8462cc9 | 1192 | |
7ccb9ad5 WF |
1193 | /* |
1194 | * Limit pause time for small memory systems. If sleeping for too long | |
1195 | * time, a small pool of dirty/writeback pages may go empty and disk go | |
1196 | * idle. | |
1197 | * | |
1198 | * 8 serves as the safety ratio. | |
1199 | */ | |
1200 | t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8)); | |
1201 | t++; | |
1202 | ||
e3b6c655 | 1203 | return min_t(unsigned long, t, MAX_PAUSE); |
7ccb9ad5 WF |
1204 | } |
1205 | ||
1206 | static long bdi_min_pause(struct backing_dev_info *bdi, | |
1207 | long max_pause, | |
1208 | unsigned long task_ratelimit, | |
1209 | unsigned long dirty_ratelimit, | |
1210 | int *nr_dirtied_pause) | |
c8462cc9 | 1211 | { |
7ccb9ad5 WF |
1212 | long hi = ilog2(bdi->avg_write_bandwidth); |
1213 | long lo = ilog2(bdi->dirty_ratelimit); | |
1214 | long t; /* target pause */ | |
1215 | long pause; /* estimated next pause */ | |
1216 | int pages; /* target nr_dirtied_pause */ | |
c8462cc9 | 1217 | |
7ccb9ad5 WF |
1218 | /* target for 10ms pause on 1-dd case */ |
1219 | t = max(1, HZ / 100); | |
c8462cc9 WF |
1220 | |
1221 | /* | |
1222 | * Scale up pause time for concurrent dirtiers in order to reduce CPU | |
1223 | * overheads. | |
1224 | * | |
7ccb9ad5 | 1225 | * (N * 10ms) on 2^N concurrent tasks. |
c8462cc9 WF |
1226 | */ |
1227 | if (hi > lo) | |
7ccb9ad5 | 1228 | t += (hi - lo) * (10 * HZ) / 1024; |
c8462cc9 WF |
1229 | |
1230 | /* | |
7ccb9ad5 WF |
1231 | * This is a bit convoluted. We try to base the next nr_dirtied_pause |
1232 | * on the much more stable dirty_ratelimit. However the next pause time | |
1233 | * will be computed based on task_ratelimit and the two rate limits may | |
1234 | * depart considerably at some time. Especially if task_ratelimit goes | |
1235 | * below dirty_ratelimit/2 and the target pause is max_pause, the next | |
1236 | * pause time will be max_pause*2 _trimmed down_ to max_pause. As a | |
1237 | * result task_ratelimit won't be executed faithfully, which could | |
1238 | * eventually bring down dirty_ratelimit. | |
c8462cc9 | 1239 | * |
7ccb9ad5 WF |
1240 | * We apply two rules to fix it up: |
1241 | * 1) try to estimate the next pause time and if necessary, use a lower | |
1242 | * nr_dirtied_pause so as not to exceed max_pause. When this happens, | |
1243 | * nr_dirtied_pause will be "dancing" with task_ratelimit. | |
1244 | * 2) limit the target pause time to max_pause/2, so that the normal | |
1245 | * small fluctuations of task_ratelimit won't trigger rule (1) and | |
1246 | * nr_dirtied_pause will remain as stable as dirty_ratelimit. | |
c8462cc9 | 1247 | */ |
7ccb9ad5 WF |
1248 | t = min(t, 1 + max_pause / 2); |
1249 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
c8462cc9 WF |
1250 | |
1251 | /* | |
5b9b3574 WF |
1252 | * Tiny nr_dirtied_pause is found to hurt I/O performance in the test |
1253 | * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}. | |
1254 | * When the 16 consecutive reads are often interrupted by some dirty | |
1255 | * throttling pause during the async writes, cfq will go into idles | |
1256 | * (deadline is fine). So push nr_dirtied_pause as high as possible | |
1257 | * until reaches DIRTY_POLL_THRESH=32 pages. | |
c8462cc9 | 1258 | */ |
5b9b3574 WF |
1259 | if (pages < DIRTY_POLL_THRESH) { |
1260 | t = max_pause; | |
1261 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
1262 | if (pages > DIRTY_POLL_THRESH) { | |
1263 | pages = DIRTY_POLL_THRESH; | |
1264 | t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit; | |
1265 | } | |
1266 | } | |
1267 | ||
7ccb9ad5 WF |
1268 | pause = HZ * pages / (task_ratelimit + 1); |
1269 | if (pause > max_pause) { | |
1270 | t = max_pause; | |
1271 | pages = task_ratelimit * t / roundup_pow_of_two(HZ); | |
1272 | } | |
c8462cc9 | 1273 | |
7ccb9ad5 | 1274 | *nr_dirtied_pause = pages; |
c8462cc9 | 1275 | /* |
7ccb9ad5 | 1276 | * The minimal pause time will normally be half the target pause time. |
c8462cc9 | 1277 | */ |
5b9b3574 | 1278 | return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t; |
c8462cc9 WF |
1279 | } |
1280 | ||
5a537485 MP |
1281 | static inline void bdi_dirty_limits(struct backing_dev_info *bdi, |
1282 | unsigned long dirty_thresh, | |
1283 | unsigned long background_thresh, | |
1284 | unsigned long *bdi_dirty, | |
1285 | unsigned long *bdi_thresh, | |
1286 | unsigned long *bdi_bg_thresh) | |
1287 | { | |
1288 | unsigned long bdi_reclaimable; | |
1289 | ||
1290 | /* | |
1291 | * bdi_thresh is not treated as some limiting factor as | |
1292 | * dirty_thresh, due to reasons | |
1293 | * - in JBOD setup, bdi_thresh can fluctuate a lot | |
1294 | * - in a system with HDD and USB key, the USB key may somehow | |
1295 | * go into state (bdi_dirty >> bdi_thresh) either because | |
1296 | * bdi_dirty starts high, or because bdi_thresh drops low. | |
1297 | * In this case we don't want to hard throttle the USB key | |
1298 | * dirtiers for 100 seconds until bdi_dirty drops under | |
1299 | * bdi_thresh. Instead the auxiliary bdi control line in | |
1300 | * bdi_position_ratio() will let the dirtier task progress | |
1301 | * at some rate <= (write_bw / 2) for bringing down bdi_dirty. | |
1302 | */ | |
1303 | *bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh); | |
1304 | ||
1305 | if (bdi_bg_thresh) | |
f6789593 MP |
1306 | *bdi_bg_thresh = dirty_thresh ? div_u64((u64)*bdi_thresh * |
1307 | background_thresh, | |
1308 | dirty_thresh) : 0; | |
5a537485 MP |
1309 | |
1310 | /* | |
1311 | * In order to avoid the stacked BDI deadlock we need | |
1312 | * to ensure we accurately count the 'dirty' pages when | |
1313 | * the threshold is low. | |
1314 | * | |
1315 | * Otherwise it would be possible to get thresh+n pages | |
1316 | * reported dirty, even though there are thresh-m pages | |
1317 | * actually dirty; with m+n sitting in the percpu | |
1318 | * deltas. | |
1319 | */ | |
1320 | if (*bdi_thresh < 2 * bdi_stat_error(bdi)) { | |
1321 | bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE); | |
1322 | *bdi_dirty = bdi_reclaimable + | |
1323 | bdi_stat_sum(bdi, BDI_WRITEBACK); | |
1324 | } else { | |
1325 | bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE); | |
1326 | *bdi_dirty = bdi_reclaimable + | |
1327 | bdi_stat(bdi, BDI_WRITEBACK); | |
1328 | } | |
1329 | } | |
1330 | ||
1da177e4 LT |
1331 | /* |
1332 | * balance_dirty_pages() must be called by processes which are generating dirty | |
1333 | * data. It looks at the number of dirty pages in the machine and will force | |
143dfe86 | 1334 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. |
5b0830cb JA |
1335 | * If we're over `background_thresh' then the writeback threads are woken to |
1336 | * perform some writeout. | |
1da177e4 | 1337 | */ |
3a2e9a5a | 1338 | static void balance_dirty_pages(struct address_space *mapping, |
143dfe86 | 1339 | unsigned long pages_dirtied) |
1da177e4 | 1340 | { |
143dfe86 | 1341 | unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */ |
7762741e | 1342 | unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */ |
364aeb28 DR |
1343 | unsigned long background_thresh; |
1344 | unsigned long dirty_thresh; | |
83712358 | 1345 | long period; |
7ccb9ad5 WF |
1346 | long pause; |
1347 | long max_pause; | |
1348 | long min_pause; | |
1349 | int nr_dirtied_pause; | |
e50e3720 | 1350 | bool dirty_exceeded = false; |
143dfe86 | 1351 | unsigned long task_ratelimit; |
7ccb9ad5 | 1352 | unsigned long dirty_ratelimit; |
143dfe86 | 1353 | unsigned long pos_ratio; |
1da177e4 | 1354 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
5a537485 | 1355 | bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT; |
e98be2d5 | 1356 | unsigned long start_time = jiffies; |
1da177e4 LT |
1357 | |
1358 | for (;;) { | |
83712358 | 1359 | unsigned long now = jiffies; |
5a537485 MP |
1360 | unsigned long uninitialized_var(bdi_thresh); |
1361 | unsigned long thresh; | |
1362 | unsigned long uninitialized_var(bdi_dirty); | |
1363 | unsigned long dirty; | |
1364 | unsigned long bg_thresh; | |
83712358 | 1365 | |
143dfe86 WF |
1366 | /* |
1367 | * Unstable writes are a feature of certain networked | |
1368 | * filesystems (i.e. NFS) in which data may have been | |
1369 | * written to the server's write cache, but has not yet | |
1370 | * been flushed to permanent storage. | |
1371 | */ | |
5fce25a9 PZ |
1372 | nr_reclaimable = global_page_state(NR_FILE_DIRTY) + |
1373 | global_page_state(NR_UNSTABLE_NFS); | |
7762741e | 1374 | nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK); |
5fce25a9 | 1375 | |
16c4042f WF |
1376 | global_dirty_limits(&background_thresh, &dirty_thresh); |
1377 | ||
5a537485 MP |
1378 | if (unlikely(strictlimit)) { |
1379 | bdi_dirty_limits(bdi, dirty_thresh, background_thresh, | |
1380 | &bdi_dirty, &bdi_thresh, &bg_thresh); | |
1381 | ||
1382 | dirty = bdi_dirty; | |
1383 | thresh = bdi_thresh; | |
1384 | } else { | |
1385 | dirty = nr_dirty; | |
1386 | thresh = dirty_thresh; | |
1387 | bg_thresh = background_thresh; | |
1388 | } | |
1389 | ||
16c4042f WF |
1390 | /* |
1391 | * Throttle it only when the background writeback cannot | |
1392 | * catch-up. This avoids (excessively) small writeouts | |
5a537485 MP |
1393 | * when the bdi limits are ramping up in case of !strictlimit. |
1394 | * | |
1395 | * In strictlimit case make decision based on the bdi counters | |
1396 | * and limits. Small writeouts when the bdi limits are ramping | |
1397 | * up are the price we consciously pay for strictlimit-ing. | |
16c4042f | 1398 | */ |
5a537485 | 1399 | if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) { |
83712358 WF |
1400 | current->dirty_paused_when = now; |
1401 | current->nr_dirtied = 0; | |
7ccb9ad5 | 1402 | current->nr_dirtied_pause = |
5a537485 | 1403 | dirty_poll_interval(dirty, thresh); |
16c4042f | 1404 | break; |
83712358 | 1405 | } |
16c4042f | 1406 | |
143dfe86 WF |
1407 | if (unlikely(!writeback_in_progress(bdi))) |
1408 | bdi_start_background_writeback(bdi); | |
1409 | ||
5a537485 MP |
1410 | if (!strictlimit) |
1411 | bdi_dirty_limits(bdi, dirty_thresh, background_thresh, | |
1412 | &bdi_dirty, &bdi_thresh, NULL); | |
5fce25a9 | 1413 | |
82791940 | 1414 | dirty_exceeded = (bdi_dirty > bdi_thresh) && |
5a537485 | 1415 | ((nr_dirty > dirty_thresh) || strictlimit); |
143dfe86 | 1416 | if (dirty_exceeded && !bdi->dirty_exceeded) |
04fbfdc1 | 1417 | bdi->dirty_exceeded = 1; |
1da177e4 | 1418 | |
af6a3113 WF |
1419 | bdi_update_bandwidth(bdi, dirty_thresh, background_thresh, |
1420 | nr_dirty, bdi_thresh, bdi_dirty, | |
1421 | start_time); | |
e98be2d5 | 1422 | |
143dfe86 WF |
1423 | dirty_ratelimit = bdi->dirty_ratelimit; |
1424 | pos_ratio = bdi_position_ratio(bdi, dirty_thresh, | |
1425 | background_thresh, nr_dirty, | |
1426 | bdi_thresh, bdi_dirty); | |
3a73dbbc WF |
1427 | task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >> |
1428 | RATELIMIT_CALC_SHIFT; | |
7ccb9ad5 WF |
1429 | max_pause = bdi_max_pause(bdi, bdi_dirty); |
1430 | min_pause = bdi_min_pause(bdi, max_pause, | |
1431 | task_ratelimit, dirty_ratelimit, | |
1432 | &nr_dirtied_pause); | |
1433 | ||
3a73dbbc | 1434 | if (unlikely(task_ratelimit == 0)) { |
83712358 | 1435 | period = max_pause; |
c8462cc9 | 1436 | pause = max_pause; |
143dfe86 | 1437 | goto pause; |
04fbfdc1 | 1438 | } |
83712358 WF |
1439 | period = HZ * pages_dirtied / task_ratelimit; |
1440 | pause = period; | |
1441 | if (current->dirty_paused_when) | |
1442 | pause -= now - current->dirty_paused_when; | |
1443 | /* | |
1444 | * For less than 1s think time (ext3/4 may block the dirtier | |
1445 | * for up to 800ms from time to time on 1-HDD; so does xfs, | |
1446 | * however at much less frequency), try to compensate it in | |
1447 | * future periods by updating the virtual time; otherwise just | |
1448 | * do a reset, as it may be a light dirtier. | |
1449 | */ | |
7ccb9ad5 | 1450 | if (pause < min_pause) { |
ece13ac3 WF |
1451 | trace_balance_dirty_pages(bdi, |
1452 | dirty_thresh, | |
1453 | background_thresh, | |
1454 | nr_dirty, | |
1455 | bdi_thresh, | |
1456 | bdi_dirty, | |
1457 | dirty_ratelimit, | |
1458 | task_ratelimit, | |
1459 | pages_dirtied, | |
83712358 | 1460 | period, |
7ccb9ad5 | 1461 | min(pause, 0L), |
ece13ac3 | 1462 | start_time); |
83712358 WF |
1463 | if (pause < -HZ) { |
1464 | current->dirty_paused_when = now; | |
1465 | current->nr_dirtied = 0; | |
1466 | } else if (period) { | |
1467 | current->dirty_paused_when += period; | |
1468 | current->nr_dirtied = 0; | |
7ccb9ad5 WF |
1469 | } else if (current->nr_dirtied_pause <= pages_dirtied) |
1470 | current->nr_dirtied_pause += pages_dirtied; | |
57fc978c | 1471 | break; |
04fbfdc1 | 1472 | } |
7ccb9ad5 WF |
1473 | if (unlikely(pause > max_pause)) { |
1474 | /* for occasional dropped task_ratelimit */ | |
1475 | now += min(pause - max_pause, max_pause); | |
1476 | pause = max_pause; | |
1477 | } | |
143dfe86 WF |
1478 | |
1479 | pause: | |
ece13ac3 WF |
1480 | trace_balance_dirty_pages(bdi, |
1481 | dirty_thresh, | |
1482 | background_thresh, | |
1483 | nr_dirty, | |
1484 | bdi_thresh, | |
1485 | bdi_dirty, | |
1486 | dirty_ratelimit, | |
1487 | task_ratelimit, | |
1488 | pages_dirtied, | |
83712358 | 1489 | period, |
ece13ac3 WF |
1490 | pause, |
1491 | start_time); | |
499d05ec | 1492 | __set_current_state(TASK_KILLABLE); |
d25105e8 | 1493 | io_schedule_timeout(pause); |
87c6a9b2 | 1494 | |
83712358 WF |
1495 | current->dirty_paused_when = now + pause; |
1496 | current->nr_dirtied = 0; | |
7ccb9ad5 | 1497 | current->nr_dirtied_pause = nr_dirtied_pause; |
83712358 | 1498 | |
ffd1f609 | 1499 | /* |
1df64719 WF |
1500 | * This is typically equal to (nr_dirty < dirty_thresh) and can |
1501 | * also keep "1000+ dd on a slow USB stick" under control. | |
ffd1f609 | 1502 | */ |
1df64719 | 1503 | if (task_ratelimit) |
ffd1f609 | 1504 | break; |
499d05ec | 1505 | |
c5c6343c WF |
1506 | /* |
1507 | * In the case of an unresponding NFS server and the NFS dirty | |
1508 | * pages exceeds dirty_thresh, give the other good bdi's a pipe | |
1509 | * to go through, so that tasks on them still remain responsive. | |
1510 | * | |
1511 | * In theory 1 page is enough to keep the comsumer-producer | |
1512 | * pipe going: the flusher cleans 1 page => the task dirties 1 | |
1513 | * more page. However bdi_dirty has accounting errors. So use | |
1514 | * the larger and more IO friendly bdi_stat_error. | |
1515 | */ | |
1516 | if (bdi_dirty <= bdi_stat_error(bdi)) | |
1517 | break; | |
1518 | ||
499d05ec JK |
1519 | if (fatal_signal_pending(current)) |
1520 | break; | |
1da177e4 LT |
1521 | } |
1522 | ||
143dfe86 | 1523 | if (!dirty_exceeded && bdi->dirty_exceeded) |
04fbfdc1 | 1524 | bdi->dirty_exceeded = 0; |
1da177e4 LT |
1525 | |
1526 | if (writeback_in_progress(bdi)) | |
5b0830cb | 1527 | return; |
1da177e4 LT |
1528 | |
1529 | /* | |
1530 | * In laptop mode, we wait until hitting the higher threshold before | |
1531 | * starting background writeout, and then write out all the way down | |
1532 | * to the lower threshold. So slow writers cause minimal disk activity. | |
1533 | * | |
1534 | * In normal mode, we start background writeout at the lower | |
1535 | * background_thresh, to keep the amount of dirty memory low. | |
1536 | */ | |
143dfe86 WF |
1537 | if (laptop_mode) |
1538 | return; | |
1539 | ||
1540 | if (nr_reclaimable > background_thresh) | |
c5444198 | 1541 | bdi_start_background_writeback(bdi); |
1da177e4 LT |
1542 | } |
1543 | ||
9d823e8f | 1544 | static DEFINE_PER_CPU(int, bdp_ratelimits); |
245b2e70 | 1545 | |
54848d73 WF |
1546 | /* |
1547 | * Normal tasks are throttled by | |
1548 | * loop { | |
1549 | * dirty tsk->nr_dirtied_pause pages; | |
1550 | * take a snap in balance_dirty_pages(); | |
1551 | * } | |
1552 | * However there is a worst case. If every task exit immediately when dirtied | |
1553 | * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be | |
1554 | * called to throttle the page dirties. The solution is to save the not yet | |
1555 | * throttled page dirties in dirty_throttle_leaks on task exit and charge them | |
1556 | * randomly into the running tasks. This works well for the above worst case, | |
1557 | * as the new task will pick up and accumulate the old task's leaked dirty | |
1558 | * count and eventually get throttled. | |
1559 | */ | |
1560 | DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0; | |
1561 | ||
1da177e4 | 1562 | /** |
d0e1d66b | 1563 | * balance_dirty_pages_ratelimited - balance dirty memory state |
67be2dd1 | 1564 | * @mapping: address_space which was dirtied |
1da177e4 LT |
1565 | * |
1566 | * Processes which are dirtying memory should call in here once for each page | |
1567 | * which was newly dirtied. The function will periodically check the system's | |
1568 | * dirty state and will initiate writeback if needed. | |
1569 | * | |
1570 | * On really big machines, get_writeback_state is expensive, so try to avoid | |
1571 | * calling it too often (ratelimiting). But once we're over the dirty memory | |
1572 | * limit we decrease the ratelimiting by a lot, to prevent individual processes | |
1573 | * from overshooting the limit by (ratelimit_pages) each. | |
1574 | */ | |
d0e1d66b | 1575 | void balance_dirty_pages_ratelimited(struct address_space *mapping) |
1da177e4 | 1576 | { |
36715cef | 1577 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
9d823e8f WF |
1578 | int ratelimit; |
1579 | int *p; | |
1da177e4 | 1580 | |
36715cef WF |
1581 | if (!bdi_cap_account_dirty(bdi)) |
1582 | return; | |
1583 | ||
9d823e8f WF |
1584 | ratelimit = current->nr_dirtied_pause; |
1585 | if (bdi->dirty_exceeded) | |
1586 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); | |
1587 | ||
9d823e8f | 1588 | preempt_disable(); |
1da177e4 | 1589 | /* |
9d823e8f WF |
1590 | * This prevents one CPU to accumulate too many dirtied pages without |
1591 | * calling into balance_dirty_pages(), which can happen when there are | |
1592 | * 1000+ tasks, all of them start dirtying pages at exactly the same | |
1593 | * time, hence all honoured too large initial task->nr_dirtied_pause. | |
1da177e4 | 1594 | */ |
7c8e0181 | 1595 | p = this_cpu_ptr(&bdp_ratelimits); |
9d823e8f | 1596 | if (unlikely(current->nr_dirtied >= ratelimit)) |
fa5a734e | 1597 | *p = 0; |
d3bc1fef WF |
1598 | else if (unlikely(*p >= ratelimit_pages)) { |
1599 | *p = 0; | |
1600 | ratelimit = 0; | |
1da177e4 | 1601 | } |
54848d73 WF |
1602 | /* |
1603 | * Pick up the dirtied pages by the exited tasks. This avoids lots of | |
1604 | * short-lived tasks (eg. gcc invocations in a kernel build) escaping | |
1605 | * the dirty throttling and livelock other long-run dirtiers. | |
1606 | */ | |
7c8e0181 | 1607 | p = this_cpu_ptr(&dirty_throttle_leaks); |
54848d73 | 1608 | if (*p > 0 && current->nr_dirtied < ratelimit) { |
d0e1d66b | 1609 | unsigned long nr_pages_dirtied; |
54848d73 WF |
1610 | nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied); |
1611 | *p -= nr_pages_dirtied; | |
1612 | current->nr_dirtied += nr_pages_dirtied; | |
1da177e4 | 1613 | } |
fa5a734e | 1614 | preempt_enable(); |
9d823e8f WF |
1615 | |
1616 | if (unlikely(current->nr_dirtied >= ratelimit)) | |
1617 | balance_dirty_pages(mapping, current->nr_dirtied); | |
1da177e4 | 1618 | } |
d0e1d66b | 1619 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited); |
1da177e4 | 1620 | |
232ea4d6 | 1621 | void throttle_vm_writeout(gfp_t gfp_mask) |
1da177e4 | 1622 | { |
364aeb28 DR |
1623 | unsigned long background_thresh; |
1624 | unsigned long dirty_thresh; | |
1da177e4 LT |
1625 | |
1626 | for ( ; ; ) { | |
16c4042f | 1627 | global_dirty_limits(&background_thresh, &dirty_thresh); |
47a13333 | 1628 | dirty_thresh = hard_dirty_limit(dirty_thresh); |
1da177e4 LT |
1629 | |
1630 | /* | |
1631 | * Boost the allowable dirty threshold a bit for page | |
1632 | * allocators so they don't get DoS'ed by heavy writers | |
1633 | */ | |
1634 | dirty_thresh += dirty_thresh / 10; /* wheeee... */ | |
1635 | ||
c24f21bd CL |
1636 | if (global_page_state(NR_UNSTABLE_NFS) + |
1637 | global_page_state(NR_WRITEBACK) <= dirty_thresh) | |
1638 | break; | |
8aa7e847 | 1639 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
369f2389 FW |
1640 | |
1641 | /* | |
1642 | * The caller might hold locks which can prevent IO completion | |
1643 | * or progress in the filesystem. So we cannot just sit here | |
1644 | * waiting for IO to complete. | |
1645 | */ | |
1646 | if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) | |
1647 | break; | |
1da177e4 LT |
1648 | } |
1649 | } | |
1650 | ||
1da177e4 LT |
1651 | /* |
1652 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
1653 | */ | |
cccad5b9 | 1654 | int dirty_writeback_centisecs_handler(struct ctl_table *table, int write, |
8d65af78 | 1655 | void __user *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 1656 | { |
8d65af78 | 1657 | proc_dointvec(table, write, buffer, length, ppos); |
1da177e4 LT |
1658 | return 0; |
1659 | } | |
1660 | ||
c2c4986e | 1661 | #ifdef CONFIG_BLOCK |
31373d09 | 1662 | void laptop_mode_timer_fn(unsigned long data) |
1da177e4 | 1663 | { |
31373d09 MG |
1664 | struct request_queue *q = (struct request_queue *)data; |
1665 | int nr_pages = global_page_state(NR_FILE_DIRTY) + | |
1666 | global_page_state(NR_UNSTABLE_NFS); | |
1da177e4 | 1667 | |
31373d09 MG |
1668 | /* |
1669 | * We want to write everything out, not just down to the dirty | |
1670 | * threshold | |
1671 | */ | |
31373d09 | 1672 | if (bdi_has_dirty_io(&q->backing_dev_info)) |
0e175a18 CW |
1673 | bdi_start_writeback(&q->backing_dev_info, nr_pages, |
1674 | WB_REASON_LAPTOP_TIMER); | |
1da177e4 LT |
1675 | } |
1676 | ||
1677 | /* | |
1678 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
1679 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
1680 | * then push it back - the user is still using the disk. | |
1681 | */ | |
31373d09 | 1682 | void laptop_io_completion(struct backing_dev_info *info) |
1da177e4 | 1683 | { |
31373d09 | 1684 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
1685 | } |
1686 | ||
1687 | /* | |
1688 | * We're in laptop mode and we've just synced. The sync's writes will have | |
1689 | * caused another writeback to be scheduled by laptop_io_completion. | |
1690 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
1691 | */ | |
1692 | void laptop_sync_completion(void) | |
1693 | { | |
31373d09 MG |
1694 | struct backing_dev_info *bdi; |
1695 | ||
1696 | rcu_read_lock(); | |
1697 | ||
1698 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) | |
1699 | del_timer(&bdi->laptop_mode_wb_timer); | |
1700 | ||
1701 | rcu_read_unlock(); | |
1da177e4 | 1702 | } |
c2c4986e | 1703 | #endif |
1da177e4 LT |
1704 | |
1705 | /* | |
1706 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
1707 | * if a large number of processes all perform writes at the same time. | |
1708 | * If it is too low then SMP machines will call the (expensive) | |
1709 | * get_writeback_state too often. | |
1710 | * | |
1711 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
1712 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
9d823e8f | 1713 | * thresholds. |
1da177e4 LT |
1714 | */ |
1715 | ||
2d1d43f6 | 1716 | void writeback_set_ratelimit(void) |
1da177e4 | 1717 | { |
9d823e8f WF |
1718 | unsigned long background_thresh; |
1719 | unsigned long dirty_thresh; | |
1720 | global_dirty_limits(&background_thresh, &dirty_thresh); | |
68809c71 | 1721 | global_dirty_limit = dirty_thresh; |
9d823e8f | 1722 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); |
1da177e4 LT |
1723 | if (ratelimit_pages < 16) |
1724 | ratelimit_pages = 16; | |
1da177e4 LT |
1725 | } |
1726 | ||
0db0628d | 1727 | static int |
2f60d628 SB |
1728 | ratelimit_handler(struct notifier_block *self, unsigned long action, |
1729 | void *hcpu) | |
1da177e4 | 1730 | { |
2f60d628 SB |
1731 | |
1732 | switch (action & ~CPU_TASKS_FROZEN) { | |
1733 | case CPU_ONLINE: | |
1734 | case CPU_DEAD: | |
1735 | writeback_set_ratelimit(); | |
1736 | return NOTIFY_OK; | |
1737 | default: | |
1738 | return NOTIFY_DONE; | |
1739 | } | |
1da177e4 LT |
1740 | } |
1741 | ||
0db0628d | 1742 | static struct notifier_block ratelimit_nb = { |
1da177e4 LT |
1743 | .notifier_call = ratelimit_handler, |
1744 | .next = NULL, | |
1745 | }; | |
1746 | ||
1747 | /* | |
dc6e29da LT |
1748 | * Called early on to tune the page writeback dirty limits. |
1749 | * | |
1750 | * We used to scale dirty pages according to how total memory | |
1751 | * related to pages that could be allocated for buffers (by | |
1752 | * comparing nr_free_buffer_pages() to vm_total_pages. | |
1753 | * | |
1754 | * However, that was when we used "dirty_ratio" to scale with | |
1755 | * all memory, and we don't do that any more. "dirty_ratio" | |
1756 | * is now applied to total non-HIGHPAGE memory (by subtracting | |
1757 | * totalhigh_pages from vm_total_pages), and as such we can't | |
1758 | * get into the old insane situation any more where we had | |
1759 | * large amounts of dirty pages compared to a small amount of | |
1760 | * non-HIGHMEM memory. | |
1761 | * | |
1762 | * But we might still want to scale the dirty_ratio by how | |
1763 | * much memory the box has.. | |
1da177e4 LT |
1764 | */ |
1765 | void __init page_writeback_init(void) | |
1766 | { | |
2d1d43f6 | 1767 | writeback_set_ratelimit(); |
1da177e4 | 1768 | register_cpu_notifier(&ratelimit_nb); |
04fbfdc1 | 1769 | |
20ae0079 | 1770 | fprop_global_init(&writeout_completions, GFP_KERNEL); |
1da177e4 LT |
1771 | } |
1772 | ||
f446daae JK |
1773 | /** |
1774 | * tag_pages_for_writeback - tag pages to be written by write_cache_pages | |
1775 | * @mapping: address space structure to write | |
1776 | * @start: starting page index | |
1777 | * @end: ending page index (inclusive) | |
1778 | * | |
1779 | * This function scans the page range from @start to @end (inclusive) and tags | |
1780 | * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is | |
1781 | * that write_cache_pages (or whoever calls this function) will then use | |
1782 | * TOWRITE tag to identify pages eligible for writeback. This mechanism is | |
1783 | * used to avoid livelocking of writeback by a process steadily creating new | |
1784 | * dirty pages in the file (thus it is important for this function to be quick | |
1785 | * so that it can tag pages faster than a dirtying process can create them). | |
1786 | */ | |
1787 | /* | |
1788 | * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency. | |
1789 | */ | |
f446daae JK |
1790 | void tag_pages_for_writeback(struct address_space *mapping, |
1791 | pgoff_t start, pgoff_t end) | |
1792 | { | |
3c111a07 | 1793 | #define WRITEBACK_TAG_BATCH 4096 |
f446daae JK |
1794 | unsigned long tagged; |
1795 | ||
1796 | do { | |
1797 | spin_lock_irq(&mapping->tree_lock); | |
1798 | tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree, | |
1799 | &start, end, WRITEBACK_TAG_BATCH, | |
1800 | PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE); | |
1801 | spin_unlock_irq(&mapping->tree_lock); | |
1802 | WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH); | |
1803 | cond_resched(); | |
d5ed3a4a JK |
1804 | /* We check 'start' to handle wrapping when end == ~0UL */ |
1805 | } while (tagged >= WRITEBACK_TAG_BATCH && start); | |
f446daae JK |
1806 | } |
1807 | EXPORT_SYMBOL(tag_pages_for_writeback); | |
1808 | ||
811d736f | 1809 | /** |
0ea97180 | 1810 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. |
811d736f DH |
1811 | * @mapping: address space structure to write |
1812 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
0ea97180 MS |
1813 | * @writepage: function called for each page |
1814 | * @data: data passed to writepage function | |
811d736f | 1815 | * |
0ea97180 | 1816 | * If a page is already under I/O, write_cache_pages() skips it, even |
811d736f DH |
1817 | * if it's dirty. This is desirable behaviour for memory-cleaning writeback, |
1818 | * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() | |
1819 | * and msync() need to guarantee that all the data which was dirty at the time | |
1820 | * the call was made get new I/O started against them. If wbc->sync_mode is | |
1821 | * WB_SYNC_ALL then we were called for data integrity and we must wait for | |
1822 | * existing IO to complete. | |
f446daae JK |
1823 | * |
1824 | * To avoid livelocks (when other process dirties new pages), we first tag | |
1825 | * pages which should be written back with TOWRITE tag and only then start | |
1826 | * writing them. For data-integrity sync we have to be careful so that we do | |
1827 | * not miss some pages (e.g., because some other process has cleared TOWRITE | |
1828 | * tag we set). The rule we follow is that TOWRITE tag can be cleared only | |
1829 | * by the process clearing the DIRTY tag (and submitting the page for IO). | |
811d736f | 1830 | */ |
0ea97180 MS |
1831 | int write_cache_pages(struct address_space *mapping, |
1832 | struct writeback_control *wbc, writepage_t writepage, | |
1833 | void *data) | |
811d736f | 1834 | { |
811d736f DH |
1835 | int ret = 0; |
1836 | int done = 0; | |
811d736f DH |
1837 | struct pagevec pvec; |
1838 | int nr_pages; | |
31a12666 | 1839 | pgoff_t uninitialized_var(writeback_index); |
811d736f DH |
1840 | pgoff_t index; |
1841 | pgoff_t end; /* Inclusive */ | |
bd19e012 | 1842 | pgoff_t done_index; |
31a12666 | 1843 | int cycled; |
811d736f | 1844 | int range_whole = 0; |
f446daae | 1845 | int tag; |
811d736f | 1846 | |
811d736f DH |
1847 | pagevec_init(&pvec, 0); |
1848 | if (wbc->range_cyclic) { | |
31a12666 NP |
1849 | writeback_index = mapping->writeback_index; /* prev offset */ |
1850 | index = writeback_index; | |
1851 | if (index == 0) | |
1852 | cycled = 1; | |
1853 | else | |
1854 | cycled = 0; | |
811d736f DH |
1855 | end = -1; |
1856 | } else { | |
1857 | index = wbc->range_start >> PAGE_CACHE_SHIFT; | |
1858 | end = wbc->range_end >> PAGE_CACHE_SHIFT; | |
1859 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) | |
1860 | range_whole = 1; | |
31a12666 | 1861 | cycled = 1; /* ignore range_cyclic tests */ |
811d736f | 1862 | } |
6e6938b6 | 1863 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae JK |
1864 | tag = PAGECACHE_TAG_TOWRITE; |
1865 | else | |
1866 | tag = PAGECACHE_TAG_DIRTY; | |
811d736f | 1867 | retry: |
6e6938b6 | 1868 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae | 1869 | tag_pages_for_writeback(mapping, index, end); |
bd19e012 | 1870 | done_index = index; |
5a3d5c98 NP |
1871 | while (!done && (index <= end)) { |
1872 | int i; | |
1873 | ||
f446daae | 1874 | nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag, |
5a3d5c98 NP |
1875 | min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1); |
1876 | if (nr_pages == 0) | |
1877 | break; | |
811d736f | 1878 | |
811d736f DH |
1879 | for (i = 0; i < nr_pages; i++) { |
1880 | struct page *page = pvec.pages[i]; | |
1881 | ||
1882 | /* | |
d5482cdf NP |
1883 | * At this point, the page may be truncated or |
1884 | * invalidated (changing page->mapping to NULL), or | |
1885 | * even swizzled back from swapper_space to tmpfs file | |
1886 | * mapping. However, page->index will not change | |
1887 | * because we have a reference on the page. | |
811d736f | 1888 | */ |
d5482cdf NP |
1889 | if (page->index > end) { |
1890 | /* | |
1891 | * can't be range_cyclic (1st pass) because | |
1892 | * end == -1 in that case. | |
1893 | */ | |
1894 | done = 1; | |
1895 | break; | |
1896 | } | |
1897 | ||
cf15b07c | 1898 | done_index = page->index; |
d5482cdf | 1899 | |
811d736f DH |
1900 | lock_page(page); |
1901 | ||
5a3d5c98 NP |
1902 | /* |
1903 | * Page truncated or invalidated. We can freely skip it | |
1904 | * then, even for data integrity operations: the page | |
1905 | * has disappeared concurrently, so there could be no | |
1906 | * real expectation of this data interity operation | |
1907 | * even if there is now a new, dirty page at the same | |
1908 | * pagecache address. | |
1909 | */ | |
811d736f | 1910 | if (unlikely(page->mapping != mapping)) { |
5a3d5c98 | 1911 | continue_unlock: |
811d736f DH |
1912 | unlock_page(page); |
1913 | continue; | |
1914 | } | |
1915 | ||
515f4a03 NP |
1916 | if (!PageDirty(page)) { |
1917 | /* someone wrote it for us */ | |
1918 | goto continue_unlock; | |
1919 | } | |
1920 | ||
1921 | if (PageWriteback(page)) { | |
1922 | if (wbc->sync_mode != WB_SYNC_NONE) | |
1923 | wait_on_page_writeback(page); | |
1924 | else | |
1925 | goto continue_unlock; | |
1926 | } | |
811d736f | 1927 | |
515f4a03 NP |
1928 | BUG_ON(PageWriteback(page)); |
1929 | if (!clear_page_dirty_for_io(page)) | |
5a3d5c98 | 1930 | goto continue_unlock; |
811d736f | 1931 | |
9e094383 | 1932 | trace_wbc_writepage(wbc, mapping->backing_dev_info); |
0ea97180 | 1933 | ret = (*writepage)(page, wbc, data); |
00266770 NP |
1934 | if (unlikely(ret)) { |
1935 | if (ret == AOP_WRITEPAGE_ACTIVATE) { | |
1936 | unlock_page(page); | |
1937 | ret = 0; | |
1938 | } else { | |
1939 | /* | |
1940 | * done_index is set past this page, | |
1941 | * so media errors will not choke | |
1942 | * background writeout for the entire | |
1943 | * file. This has consequences for | |
1944 | * range_cyclic semantics (ie. it may | |
1945 | * not be suitable for data integrity | |
1946 | * writeout). | |
1947 | */ | |
cf15b07c | 1948 | done_index = page->index + 1; |
00266770 NP |
1949 | done = 1; |
1950 | break; | |
1951 | } | |
0b564927 | 1952 | } |
00266770 | 1953 | |
546a1924 DC |
1954 | /* |
1955 | * We stop writing back only if we are not doing | |
1956 | * integrity sync. In case of integrity sync we have to | |
1957 | * keep going until we have written all the pages | |
1958 | * we tagged for writeback prior to entering this loop. | |
1959 | */ | |
1960 | if (--wbc->nr_to_write <= 0 && | |
1961 | wbc->sync_mode == WB_SYNC_NONE) { | |
1962 | done = 1; | |
1963 | break; | |
05fe478d | 1964 | } |
811d736f DH |
1965 | } |
1966 | pagevec_release(&pvec); | |
1967 | cond_resched(); | |
1968 | } | |
3a4c6800 | 1969 | if (!cycled && !done) { |
811d736f | 1970 | /* |
31a12666 | 1971 | * range_cyclic: |
811d736f DH |
1972 | * We hit the last page and there is more work to be done: wrap |
1973 | * back to the start of the file | |
1974 | */ | |
31a12666 | 1975 | cycled = 1; |
811d736f | 1976 | index = 0; |
31a12666 | 1977 | end = writeback_index - 1; |
811d736f DH |
1978 | goto retry; |
1979 | } | |
0b564927 DC |
1980 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) |
1981 | mapping->writeback_index = done_index; | |
06d6cf69 | 1982 | |
811d736f DH |
1983 | return ret; |
1984 | } | |
0ea97180 MS |
1985 | EXPORT_SYMBOL(write_cache_pages); |
1986 | ||
1987 | /* | |
1988 | * Function used by generic_writepages to call the real writepage | |
1989 | * function and set the mapping flags on error | |
1990 | */ | |
1991 | static int __writepage(struct page *page, struct writeback_control *wbc, | |
1992 | void *data) | |
1993 | { | |
1994 | struct address_space *mapping = data; | |
1995 | int ret = mapping->a_ops->writepage(page, wbc); | |
1996 | mapping_set_error(mapping, ret); | |
1997 | return ret; | |
1998 | } | |
1999 | ||
2000 | /** | |
2001 | * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. | |
2002 | * @mapping: address space structure to write | |
2003 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
2004 | * | |
2005 | * This is a library function, which implements the writepages() | |
2006 | * address_space_operation. | |
2007 | */ | |
2008 | int generic_writepages(struct address_space *mapping, | |
2009 | struct writeback_control *wbc) | |
2010 | { | |
9b6096a6 SL |
2011 | struct blk_plug plug; |
2012 | int ret; | |
2013 | ||
0ea97180 MS |
2014 | /* deal with chardevs and other special file */ |
2015 | if (!mapping->a_ops->writepage) | |
2016 | return 0; | |
2017 | ||
9b6096a6 SL |
2018 | blk_start_plug(&plug); |
2019 | ret = write_cache_pages(mapping, wbc, __writepage, mapping); | |
2020 | blk_finish_plug(&plug); | |
2021 | return ret; | |
0ea97180 | 2022 | } |
811d736f DH |
2023 | |
2024 | EXPORT_SYMBOL(generic_writepages); | |
2025 | ||
1da177e4 LT |
2026 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
2027 | { | |
22905f77 AM |
2028 | int ret; |
2029 | ||
1da177e4 LT |
2030 | if (wbc->nr_to_write <= 0) |
2031 | return 0; | |
2032 | if (mapping->a_ops->writepages) | |
d08b3851 | 2033 | ret = mapping->a_ops->writepages(mapping, wbc); |
22905f77 AM |
2034 | else |
2035 | ret = generic_writepages(mapping, wbc); | |
22905f77 | 2036 | return ret; |
1da177e4 LT |
2037 | } |
2038 | ||
2039 | /** | |
2040 | * write_one_page - write out a single page and optionally wait on I/O | |
67be2dd1 MW |
2041 | * @page: the page to write |
2042 | * @wait: if true, wait on writeout | |
1da177e4 LT |
2043 | * |
2044 | * The page must be locked by the caller and will be unlocked upon return. | |
2045 | * | |
2046 | * write_one_page() returns a negative error code if I/O failed. | |
2047 | */ | |
2048 | int write_one_page(struct page *page, int wait) | |
2049 | { | |
2050 | struct address_space *mapping = page->mapping; | |
2051 | int ret = 0; | |
2052 | struct writeback_control wbc = { | |
2053 | .sync_mode = WB_SYNC_ALL, | |
2054 | .nr_to_write = 1, | |
2055 | }; | |
2056 | ||
2057 | BUG_ON(!PageLocked(page)); | |
2058 | ||
2059 | if (wait) | |
2060 | wait_on_page_writeback(page); | |
2061 | ||
2062 | if (clear_page_dirty_for_io(page)) { | |
2063 | page_cache_get(page); | |
2064 | ret = mapping->a_ops->writepage(page, &wbc); | |
2065 | if (ret == 0 && wait) { | |
2066 | wait_on_page_writeback(page); | |
2067 | if (PageError(page)) | |
2068 | ret = -EIO; | |
2069 | } | |
2070 | page_cache_release(page); | |
2071 | } else { | |
2072 | unlock_page(page); | |
2073 | } | |
2074 | return ret; | |
2075 | } | |
2076 | EXPORT_SYMBOL(write_one_page); | |
2077 | ||
76719325 KC |
2078 | /* |
2079 | * For address_spaces which do not use buffers nor write back. | |
2080 | */ | |
2081 | int __set_page_dirty_no_writeback(struct page *page) | |
2082 | { | |
2083 | if (!PageDirty(page)) | |
c3f0da63 | 2084 | return !TestSetPageDirty(page); |
76719325 KC |
2085 | return 0; |
2086 | } | |
2087 | ||
e3a7cca1 ES |
2088 | /* |
2089 | * Helper function for set_page_dirty family. | |
2090 | * NOTE: This relies on being atomic wrt interrupts. | |
2091 | */ | |
2092 | void account_page_dirtied(struct page *page, struct address_space *mapping) | |
2093 | { | |
9fb0a7da TH |
2094 | trace_writeback_dirty_page(page, mapping); |
2095 | ||
e3a7cca1 ES |
2096 | if (mapping_cap_account_dirty(mapping)) { |
2097 | __inc_zone_page_state(page, NR_FILE_DIRTY); | |
ea941f0e | 2098 | __inc_zone_page_state(page, NR_DIRTIED); |
e3a7cca1 | 2099 | __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE); |
c8e28ce0 | 2100 | __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED); |
e3a7cca1 | 2101 | task_io_account_write(PAGE_CACHE_SIZE); |
d3bc1fef WF |
2102 | current->nr_dirtied++; |
2103 | this_cpu_inc(bdp_ratelimits); | |
e3a7cca1 ES |
2104 | } |
2105 | } | |
679ceace | 2106 | EXPORT_SYMBOL(account_page_dirtied); |
e3a7cca1 | 2107 | |
1da177e4 LT |
2108 | /* |
2109 | * For address_spaces which do not use buffers. Just tag the page as dirty in | |
2110 | * its radix tree. | |
2111 | * | |
2112 | * This is also used when a single buffer is being dirtied: we want to set the | |
2113 | * page dirty in that case, but not all the buffers. This is a "bottom-up" | |
2114 | * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying. | |
2115 | * | |
2d6d7f98 JW |
2116 | * The caller must ensure this doesn't race with truncation. Most will simply |
2117 | * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and | |
2118 | * the pte lock held, which also locks out truncation. | |
1da177e4 LT |
2119 | */ |
2120 | int __set_page_dirty_nobuffers(struct page *page) | |
2121 | { | |
1da177e4 LT |
2122 | if (!TestSetPageDirty(page)) { |
2123 | struct address_space *mapping = page_mapping(page); | |
a85d9df1 | 2124 | unsigned long flags; |
1da177e4 | 2125 | |
8c08540f AM |
2126 | if (!mapping) |
2127 | return 1; | |
2128 | ||
a85d9df1 | 2129 | spin_lock_irqsave(&mapping->tree_lock, flags); |
2d6d7f98 JW |
2130 | BUG_ON(page_mapping(page) != mapping); |
2131 | WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page)); | |
2132 | account_page_dirtied(page, mapping); | |
2133 | radix_tree_tag_set(&mapping->page_tree, page_index(page), | |
2134 | PAGECACHE_TAG_DIRTY); | |
a85d9df1 | 2135 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
8c08540f AM |
2136 | if (mapping->host) { |
2137 | /* !PageAnon && !swapper_space */ | |
2138 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 2139 | } |
4741c9fd | 2140 | return 1; |
1da177e4 | 2141 | } |
4741c9fd | 2142 | return 0; |
1da177e4 LT |
2143 | } |
2144 | EXPORT_SYMBOL(__set_page_dirty_nobuffers); | |
2145 | ||
2f800fbd WF |
2146 | /* |
2147 | * Call this whenever redirtying a page, to de-account the dirty counters | |
2148 | * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written | |
2149 | * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to | |
2150 | * systematic errors in balanced_dirty_ratelimit and the dirty pages position | |
2151 | * control. | |
2152 | */ | |
2153 | void account_page_redirty(struct page *page) | |
2154 | { | |
2155 | struct address_space *mapping = page->mapping; | |
2156 | if (mapping && mapping_cap_account_dirty(mapping)) { | |
2157 | current->nr_dirtied--; | |
2158 | dec_zone_page_state(page, NR_DIRTIED); | |
2159 | dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED); | |
2160 | } | |
2161 | } | |
2162 | EXPORT_SYMBOL(account_page_redirty); | |
2163 | ||
1da177e4 LT |
2164 | /* |
2165 | * When a writepage implementation decides that it doesn't want to write this | |
2166 | * page for some reason, it should redirty the locked page via | |
2167 | * redirty_page_for_writepage() and it should then unlock the page and return 0 | |
2168 | */ | |
2169 | int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page) | |
2170 | { | |
2171 | wbc->pages_skipped++; | |
2f800fbd | 2172 | account_page_redirty(page); |
1da177e4 LT |
2173 | return __set_page_dirty_nobuffers(page); |
2174 | } | |
2175 | EXPORT_SYMBOL(redirty_page_for_writepage); | |
2176 | ||
2177 | /* | |
6746aff7 WF |
2178 | * Dirty a page. |
2179 | * | |
2180 | * For pages with a mapping this should be done under the page lock | |
2181 | * for the benefit of asynchronous memory errors who prefer a consistent | |
2182 | * dirty state. This rule can be broken in some special cases, | |
2183 | * but should be better not to. | |
2184 | * | |
1da177e4 LT |
2185 | * If the mapping doesn't provide a set_page_dirty a_op, then |
2186 | * just fall through and assume that it wants buffer_heads. | |
2187 | */ | |
1cf6e7d8 | 2188 | int set_page_dirty(struct page *page) |
1da177e4 LT |
2189 | { |
2190 | struct address_space *mapping = page_mapping(page); | |
2191 | ||
2192 | if (likely(mapping)) { | |
2193 | int (*spd)(struct page *) = mapping->a_ops->set_page_dirty; | |
278df9f4 MK |
2194 | /* |
2195 | * readahead/lru_deactivate_page could remain | |
2196 | * PG_readahead/PG_reclaim due to race with end_page_writeback | |
2197 | * About readahead, if the page is written, the flags would be | |
2198 | * reset. So no problem. | |
2199 | * About lru_deactivate_page, if the page is redirty, the flag | |
2200 | * will be reset. So no problem. but if the page is used by readahead | |
2201 | * it will confuse readahead and make it restart the size rampup | |
2202 | * process. But it's a trivial problem. | |
2203 | */ | |
2204 | ClearPageReclaim(page); | |
9361401e DH |
2205 | #ifdef CONFIG_BLOCK |
2206 | if (!spd) | |
2207 | spd = __set_page_dirty_buffers; | |
2208 | #endif | |
2209 | return (*spd)(page); | |
1da177e4 | 2210 | } |
4741c9fd AM |
2211 | if (!PageDirty(page)) { |
2212 | if (!TestSetPageDirty(page)) | |
2213 | return 1; | |
2214 | } | |
1da177e4 LT |
2215 | return 0; |
2216 | } | |
2217 | EXPORT_SYMBOL(set_page_dirty); | |
2218 | ||
2219 | /* | |
2220 | * set_page_dirty() is racy if the caller has no reference against | |
2221 | * page->mapping->host, and if the page is unlocked. This is because another | |
2222 | * CPU could truncate the page off the mapping and then free the mapping. | |
2223 | * | |
2224 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
2225 | * holds a reference on the inode by having an open file. | |
2226 | * | |
2227 | * In other cases, the page should be locked before running set_page_dirty(). | |
2228 | */ | |
2229 | int set_page_dirty_lock(struct page *page) | |
2230 | { | |
2231 | int ret; | |
2232 | ||
7eaceacc | 2233 | lock_page(page); |
1da177e4 LT |
2234 | ret = set_page_dirty(page); |
2235 | unlock_page(page); | |
2236 | return ret; | |
2237 | } | |
2238 | EXPORT_SYMBOL(set_page_dirty_lock); | |
2239 | ||
1da177e4 LT |
2240 | /* |
2241 | * Clear a page's dirty flag, while caring for dirty memory accounting. | |
2242 | * Returns true if the page was previously dirty. | |
2243 | * | |
2244 | * This is for preparing to put the page under writeout. We leave the page | |
2245 | * tagged as dirty in the radix tree so that a concurrent write-for-sync | |
2246 | * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage | |
2247 | * implementation will run either set_page_writeback() or set_page_dirty(), | |
2248 | * at which stage we bring the page's dirty flag and radix-tree dirty tag | |
2249 | * back into sync. | |
2250 | * | |
2251 | * This incoherency between the page's dirty flag and radix-tree tag is | |
2252 | * unfortunate, but it only exists while the page is locked. | |
2253 | */ | |
2254 | int clear_page_dirty_for_io(struct page *page) | |
2255 | { | |
2256 | struct address_space *mapping = page_mapping(page); | |
2257 | ||
79352894 NP |
2258 | BUG_ON(!PageLocked(page)); |
2259 | ||
7658cc28 LT |
2260 | if (mapping && mapping_cap_account_dirty(mapping)) { |
2261 | /* | |
2262 | * Yes, Virginia, this is indeed insane. | |
2263 | * | |
2264 | * We use this sequence to make sure that | |
2265 | * (a) we account for dirty stats properly | |
2266 | * (b) we tell the low-level filesystem to | |
2267 | * mark the whole page dirty if it was | |
2268 | * dirty in a pagetable. Only to then | |
2269 | * (c) clean the page again and return 1 to | |
2270 | * cause the writeback. | |
2271 | * | |
2272 | * This way we avoid all nasty races with the | |
2273 | * dirty bit in multiple places and clearing | |
2274 | * them concurrently from different threads. | |
2275 | * | |
2276 | * Note! Normally the "set_page_dirty(page)" | |
2277 | * has no effect on the actual dirty bit - since | |
2278 | * that will already usually be set. But we | |
2279 | * need the side effects, and it can help us | |
2280 | * avoid races. | |
2281 | * | |
2282 | * We basically use the page "master dirty bit" | |
2283 | * as a serialization point for all the different | |
2284 | * threads doing their things. | |
7658cc28 LT |
2285 | */ |
2286 | if (page_mkclean(page)) | |
2287 | set_page_dirty(page); | |
79352894 NP |
2288 | /* |
2289 | * We carefully synchronise fault handlers against | |
2290 | * installing a dirty pte and marking the page dirty | |
2d6d7f98 JW |
2291 | * at this point. We do this by having them hold the |
2292 | * page lock while dirtying the page, and pages are | |
2293 | * always locked coming in here, so we get the desired | |
2294 | * exclusion. | |
79352894 | 2295 | */ |
7658cc28 | 2296 | if (TestClearPageDirty(page)) { |
8c08540f | 2297 | dec_zone_page_state(page, NR_FILE_DIRTY); |
c9e51e41 PZ |
2298 | dec_bdi_stat(mapping->backing_dev_info, |
2299 | BDI_RECLAIMABLE); | |
7658cc28 | 2300 | return 1; |
1da177e4 | 2301 | } |
7658cc28 | 2302 | return 0; |
1da177e4 | 2303 | } |
7658cc28 | 2304 | return TestClearPageDirty(page); |
1da177e4 | 2305 | } |
58bb01a9 | 2306 | EXPORT_SYMBOL(clear_page_dirty_for_io); |
1da177e4 LT |
2307 | |
2308 | int test_clear_page_writeback(struct page *page) | |
2309 | { | |
2310 | struct address_space *mapping = page_mapping(page); | |
3ea67d06 | 2311 | unsigned long memcg_flags; |
d7365e78 JW |
2312 | struct mem_cgroup *memcg; |
2313 | bool locked; | |
2314 | int ret; | |
1da177e4 | 2315 | |
d7365e78 | 2316 | memcg = mem_cgroup_begin_page_stat(page, &locked, &memcg_flags); |
1da177e4 | 2317 | if (mapping) { |
69cb51d1 | 2318 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
1da177e4 LT |
2319 | unsigned long flags; |
2320 | ||
19fd6231 | 2321 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 2322 | ret = TestClearPageWriteback(page); |
69cb51d1 | 2323 | if (ret) { |
1da177e4 LT |
2324 | radix_tree_tag_clear(&mapping->page_tree, |
2325 | page_index(page), | |
2326 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 2327 | if (bdi_cap_account_writeback(bdi)) { |
69cb51d1 | 2328 | __dec_bdi_stat(bdi, BDI_WRITEBACK); |
04fbfdc1 PZ |
2329 | __bdi_writeout_inc(bdi); |
2330 | } | |
69cb51d1 | 2331 | } |
19fd6231 | 2332 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
2333 | } else { |
2334 | ret = TestClearPageWriteback(page); | |
2335 | } | |
99b12e3d | 2336 | if (ret) { |
d7365e78 | 2337 | mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK); |
d688abf5 | 2338 | dec_zone_page_state(page, NR_WRITEBACK); |
99b12e3d WF |
2339 | inc_zone_page_state(page, NR_WRITTEN); |
2340 | } | |
e4bd6a02 | 2341 | mem_cgroup_end_page_stat(memcg, &locked, &memcg_flags); |
1da177e4 LT |
2342 | return ret; |
2343 | } | |
2344 | ||
1c8349a1 | 2345 | int __test_set_page_writeback(struct page *page, bool keep_write) |
1da177e4 LT |
2346 | { |
2347 | struct address_space *mapping = page_mapping(page); | |
3ea67d06 | 2348 | unsigned long memcg_flags; |
d7365e78 JW |
2349 | struct mem_cgroup *memcg; |
2350 | bool locked; | |
2351 | int ret; | |
1da177e4 | 2352 | |
d7365e78 | 2353 | memcg = mem_cgroup_begin_page_stat(page, &locked, &memcg_flags); |
1da177e4 | 2354 | if (mapping) { |
69cb51d1 | 2355 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
1da177e4 LT |
2356 | unsigned long flags; |
2357 | ||
19fd6231 | 2358 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 2359 | ret = TestSetPageWriteback(page); |
69cb51d1 | 2360 | if (!ret) { |
1da177e4 LT |
2361 | radix_tree_tag_set(&mapping->page_tree, |
2362 | page_index(page), | |
2363 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 2364 | if (bdi_cap_account_writeback(bdi)) |
69cb51d1 PZ |
2365 | __inc_bdi_stat(bdi, BDI_WRITEBACK); |
2366 | } | |
1da177e4 LT |
2367 | if (!PageDirty(page)) |
2368 | radix_tree_tag_clear(&mapping->page_tree, | |
2369 | page_index(page), | |
2370 | PAGECACHE_TAG_DIRTY); | |
1c8349a1 NJ |
2371 | if (!keep_write) |
2372 | radix_tree_tag_clear(&mapping->page_tree, | |
2373 | page_index(page), | |
2374 | PAGECACHE_TAG_TOWRITE); | |
19fd6231 | 2375 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
2376 | } else { |
2377 | ret = TestSetPageWriteback(page); | |
2378 | } | |
3a3c02ec | 2379 | if (!ret) { |
d7365e78 | 2380 | mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK); |
3a3c02ec JW |
2381 | inc_zone_page_state(page, NR_WRITEBACK); |
2382 | } | |
e4bd6a02 | 2383 | mem_cgroup_end_page_stat(memcg, &locked, &memcg_flags); |
1da177e4 LT |
2384 | return ret; |
2385 | ||
2386 | } | |
1c8349a1 | 2387 | EXPORT_SYMBOL(__test_set_page_writeback); |
1da177e4 LT |
2388 | |
2389 | /* | |
00128188 | 2390 | * Return true if any of the pages in the mapping are marked with the |
1da177e4 LT |
2391 | * passed tag. |
2392 | */ | |
2393 | int mapping_tagged(struct address_space *mapping, int tag) | |
2394 | { | |
72c47832 | 2395 | return radix_tree_tagged(&mapping->page_tree, tag); |
1da177e4 LT |
2396 | } |
2397 | EXPORT_SYMBOL(mapping_tagged); | |
1d1d1a76 DW |
2398 | |
2399 | /** | |
2400 | * wait_for_stable_page() - wait for writeback to finish, if necessary. | |
2401 | * @page: The page to wait on. | |
2402 | * | |
2403 | * This function determines if the given page is related to a backing device | |
2404 | * that requires page contents to be held stable during writeback. If so, then | |
2405 | * it will wait for any pending writeback to complete. | |
2406 | */ | |
2407 | void wait_for_stable_page(struct page *page) | |
2408 | { | |
2409 | struct address_space *mapping = page_mapping(page); | |
2410 | struct backing_dev_info *bdi = mapping->backing_dev_info; | |
2411 | ||
2412 | if (!bdi_cap_stable_pages_required(bdi)) | |
2413 | return; | |
2414 | ||
2415 | wait_on_page_writeback(page); | |
2416 | } | |
2417 | EXPORT_SYMBOL_GPL(wait_for_stable_page); |