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> |
028c2dd1 | 39 | #include <trace/events/writeback.h> |
1da177e4 | 40 | |
ffd1f609 WF |
41 | /* |
42 | * Sleep at most 200ms at a time in balance_dirty_pages(). | |
43 | */ | |
44 | #define MAX_PAUSE max(HZ/5, 1) | |
45 | ||
5b9b3574 WF |
46 | /* |
47 | * Try to keep balance_dirty_pages() call intervals higher than this many pages | |
48 | * by raising pause time to max_pause when falls below it. | |
49 | */ | |
50 | #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10)) | |
51 | ||
e98be2d5 WF |
52 | /* |
53 | * Estimate write bandwidth at 200ms intervals. | |
54 | */ | |
55 | #define BANDWIDTH_INTERVAL max(HZ/5, 1) | |
56 | ||
6c14ae1e WF |
57 | #define RATELIMIT_CALC_SHIFT 10 |
58 | ||
1da177e4 LT |
59 | /* |
60 | * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited | |
61 | * will look to see if it needs to force writeback or throttling. | |
62 | */ | |
63 | static long ratelimit_pages = 32; | |
64 | ||
1da177e4 LT |
65 | /* The following parameters are exported via /proc/sys/vm */ |
66 | ||
67 | /* | |
5b0830cb | 68 | * Start background writeback (via writeback threads) at this percentage |
1da177e4 | 69 | */ |
1b5e62b4 | 70 | int dirty_background_ratio = 10; |
1da177e4 | 71 | |
2da02997 DR |
72 | /* |
73 | * dirty_background_bytes starts at 0 (disabled) so that it is a function of | |
74 | * dirty_background_ratio * the amount of dirtyable memory | |
75 | */ | |
76 | unsigned long dirty_background_bytes; | |
77 | ||
195cf453 BG |
78 | /* |
79 | * free highmem will not be subtracted from the total free memory | |
80 | * for calculating free ratios if vm_highmem_is_dirtyable is true | |
81 | */ | |
82 | int vm_highmem_is_dirtyable; | |
83 | ||
1da177e4 LT |
84 | /* |
85 | * The generator of dirty data starts writeback at this percentage | |
86 | */ | |
1b5e62b4 | 87 | int vm_dirty_ratio = 20; |
1da177e4 | 88 | |
2da02997 DR |
89 | /* |
90 | * vm_dirty_bytes starts at 0 (disabled) so that it is a function of | |
91 | * vm_dirty_ratio * the amount of dirtyable memory | |
92 | */ | |
93 | unsigned long vm_dirty_bytes; | |
94 | ||
1da177e4 | 95 | /* |
704503d8 | 96 | * The interval between `kupdate'-style writebacks |
1da177e4 | 97 | */ |
22ef37ee | 98 | unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */ |
1da177e4 | 99 | |
91913a29 AB |
100 | EXPORT_SYMBOL_GPL(dirty_writeback_interval); |
101 | ||
1da177e4 | 102 | /* |
704503d8 | 103 | * The longest time for which data is allowed to remain dirty |
1da177e4 | 104 | */ |
22ef37ee | 105 | unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */ |
1da177e4 LT |
106 | |
107 | /* | |
108 | * Flag that makes the machine dump writes/reads and block dirtyings. | |
109 | */ | |
110 | int block_dump; | |
111 | ||
112 | /* | |
ed5b43f1 BS |
113 | * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: |
114 | * a full sync is triggered after this time elapses without any disk activity. | |
1da177e4 LT |
115 | */ |
116 | int laptop_mode; | |
117 | ||
118 | EXPORT_SYMBOL(laptop_mode); | |
119 | ||
120 | /* End of sysctl-exported parameters */ | |
121 | ||
c42843f2 | 122 | unsigned long global_dirty_limit; |
1da177e4 | 123 | |
04fbfdc1 PZ |
124 | /* |
125 | * Scale the writeback cache size proportional to the relative writeout speeds. | |
126 | * | |
127 | * We do this by keeping a floating proportion between BDIs, based on page | |
128 | * writeback completions [end_page_writeback()]. Those devices that write out | |
129 | * pages fastest will get the larger share, while the slower will get a smaller | |
130 | * share. | |
131 | * | |
132 | * We use page writeout completions because we are interested in getting rid of | |
133 | * dirty pages. Having them written out is the primary goal. | |
134 | * | |
135 | * We introduce a concept of time, a period over which we measure these events, | |
136 | * because demand can/will vary over time. The length of this period itself is | |
137 | * measured in page writeback completions. | |
138 | * | |
139 | */ | |
eb608e3a JK |
140 | static struct fprop_global writeout_completions; |
141 | ||
142 | static void writeout_period(unsigned long t); | |
143 | /* Timer for aging of writeout_completions */ | |
144 | static struct timer_list writeout_period_timer = | |
145 | TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0); | |
146 | static unsigned long writeout_period_time = 0; | |
147 | ||
148 | /* | |
149 | * Length of period for aging writeout fractions of bdis. This is an | |
150 | * arbitrarily chosen number. The longer the period, the slower fractions will | |
151 | * reflect changes in current writeout rate. | |
152 | */ | |
153 | #define VM_COMPLETIONS_PERIOD_LEN (3*HZ) | |
04fbfdc1 | 154 | |
1edf2234 JW |
155 | /* |
156 | * Work out the current dirty-memory clamping and background writeout | |
157 | * thresholds. | |
158 | * | |
159 | * The main aim here is to lower them aggressively if there is a lot of mapped | |
160 | * memory around. To avoid stressing page reclaim with lots of unreclaimable | |
161 | * pages. It is better to clamp down on writers than to start swapping, and | |
162 | * performing lots of scanning. | |
163 | * | |
164 | * We only allow 1/2 of the currently-unmapped memory to be dirtied. | |
165 | * | |
166 | * We don't permit the clamping level to fall below 5% - that is getting rather | |
167 | * excessive. | |
168 | * | |
169 | * We make sure that the background writeout level is below the adjusted | |
170 | * clamping level. | |
171 | */ | |
ccafa287 | 172 | |
a756cf59 JW |
173 | /* |
174 | * In a memory zone, there is a certain amount of pages we consider | |
175 | * available for the page cache, which is essentially the number of | |
176 | * free and reclaimable pages, minus some zone reserves to protect | |
177 | * lowmem and the ability to uphold the zone's watermarks without | |
178 | * requiring writeback. | |
179 | * | |
180 | * This number of dirtyable pages is the base value of which the | |
181 | * user-configurable dirty ratio is the effictive number of pages that | |
182 | * are allowed to be actually dirtied. Per individual zone, or | |
183 | * globally by using the sum of dirtyable pages over all zones. | |
184 | * | |
185 | * Because the user is allowed to specify the dirty limit globally as | |
186 | * absolute number of bytes, calculating the per-zone dirty limit can | |
187 | * require translating the configured limit into a percentage of | |
188 | * global dirtyable memory first. | |
189 | */ | |
190 | ||
03381bd2 JW |
191 | /** |
192 | * zone_dirtyable_memory - number of dirtyable pages in a zone | |
193 | * @zone: the zone | |
194 | * | |
195 | * Returns the zone's number of pages potentially available for dirty | |
196 | * page cache. This is the base value for the per-zone dirty limits. | |
197 | */ | |
198 | static unsigned long zone_dirtyable_memory(struct zone *zone) | |
199 | { | |
200 | unsigned long nr_pages; | |
201 | ||
202 | nr_pages = zone_page_state(zone, NR_FREE_PAGES); | |
203 | nr_pages -= min(nr_pages, zone->dirty_balance_reserve); | |
204 | ||
48526149 JW |
205 | nr_pages += zone_page_state(zone, NR_INACTIVE_FILE); |
206 | nr_pages += zone_page_state(zone, NR_ACTIVE_FILE); | |
03381bd2 JW |
207 | |
208 | return nr_pages; | |
209 | } | |
210 | ||
1edf2234 JW |
211 | static unsigned long highmem_dirtyable_memory(unsigned long total) |
212 | { | |
213 | #ifdef CONFIG_HIGHMEM | |
214 | int node; | |
215 | unsigned long x = 0; | |
216 | ||
217 | for_each_node_state(node, N_HIGH_MEMORY) { | |
03381bd2 | 218 | struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM]; |
1edf2234 | 219 | |
03381bd2 | 220 | x += zone_dirtyable_memory(z); |
1edf2234 | 221 | } |
c8b74c2f SR |
222 | /* |
223 | * Unreclaimable memory (kernel memory or anonymous memory | |
224 | * without swap) can bring down the dirtyable pages below | |
225 | * the zone's dirty balance reserve and the above calculation | |
226 | * will underflow. However we still want to add in nodes | |
227 | * which are below threshold (negative values) to get a more | |
228 | * accurate calculation but make sure that the total never | |
229 | * underflows. | |
230 | */ | |
231 | if ((long)x < 0) | |
232 | x = 0; | |
233 | ||
1edf2234 JW |
234 | /* |
235 | * Make sure that the number of highmem pages is never larger | |
236 | * than the number of the total dirtyable memory. This can only | |
237 | * occur in very strange VM situations but we want to make sure | |
238 | * that this does not occur. | |
239 | */ | |
240 | return min(x, total); | |
241 | #else | |
242 | return 0; | |
243 | #endif | |
244 | } | |
245 | ||
246 | /** | |
ccafa287 | 247 | * global_dirtyable_memory - number of globally dirtyable pages |
1edf2234 | 248 | * |
ccafa287 JW |
249 | * Returns the global number of pages potentially available for dirty |
250 | * page cache. This is the base value for the global dirty limits. | |
1edf2234 | 251 | */ |
18cf8cf8 | 252 | static unsigned long global_dirtyable_memory(void) |
1edf2234 JW |
253 | { |
254 | unsigned long x; | |
255 | ||
03381bd2 | 256 | x = global_page_state(NR_FREE_PAGES); |
c8b74c2f | 257 | x -= min(x, dirty_balance_reserve); |
1edf2234 | 258 | |
48526149 JW |
259 | x += global_page_state(NR_INACTIVE_FILE); |
260 | x += global_page_state(NR_ACTIVE_FILE); | |
03381bd2 | 261 | |
1edf2234 JW |
262 | if (!vm_highmem_is_dirtyable) |
263 | x -= highmem_dirtyable_memory(x); | |
264 | ||
75f7ad8e PS |
265 | /* Subtract min_free_kbytes */ |
266 | x -= min_t(unsigned long, x, min_free_kbytes >> (PAGE_SHIFT - 10)); | |
267 | ||
1edf2234 JW |
268 | return x + 1; /* Ensure that we never return 0 */ |
269 | } | |
270 | ||
ccafa287 JW |
271 | /* |
272 | * global_dirty_limits - background-writeback and dirty-throttling thresholds | |
273 | * | |
274 | * Calculate the dirty thresholds based on sysctl parameters | |
275 | * - vm.dirty_background_ratio or vm.dirty_background_bytes | |
276 | * - vm.dirty_ratio or vm.dirty_bytes | |
277 | * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and | |
278 | * real-time tasks. | |
279 | */ | |
280 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) | |
281 | { | |
282 | unsigned long background; | |
283 | unsigned long dirty; | |
284 | unsigned long uninitialized_var(available_memory); | |
285 | struct task_struct *tsk; | |
286 | ||
287 | if (!vm_dirty_bytes || !dirty_background_bytes) | |
288 | available_memory = global_dirtyable_memory(); | |
289 | ||
290 | if (vm_dirty_bytes) | |
291 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE); | |
292 | else | |
293 | dirty = (vm_dirty_ratio * available_memory) / 100; | |
294 | ||
295 | if (dirty_background_bytes) | |
296 | background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE); | |
297 | else | |
298 | background = (dirty_background_ratio * available_memory) / 100; | |
299 | ||
300 | if (background >= dirty) | |
301 | background = dirty / 2; | |
302 | tsk = current; | |
303 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) { | |
304 | background += background / 4; | |
305 | dirty += dirty / 4; | |
306 | } | |
307 | *pbackground = background; | |
308 | *pdirty = dirty; | |
309 | trace_global_dirty_state(background, dirty); | |
310 | } | |
311 | ||
a756cf59 JW |
312 | /** |
313 | * zone_dirty_limit - maximum number of dirty pages allowed in a zone | |
314 | * @zone: the zone | |
315 | * | |
316 | * Returns the maximum number of dirty pages allowed in a zone, based | |
317 | * on the zone's dirtyable memory. | |
318 | */ | |
319 | static unsigned long zone_dirty_limit(struct zone *zone) | |
320 | { | |
321 | unsigned long zone_memory = zone_dirtyable_memory(zone); | |
322 | struct task_struct *tsk = current; | |
323 | unsigned long dirty; | |
324 | ||
325 | if (vm_dirty_bytes) | |
326 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) * | |
327 | zone_memory / global_dirtyable_memory(); | |
328 | else | |
329 | dirty = vm_dirty_ratio * zone_memory / 100; | |
330 | ||
331 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) | |
332 | dirty += dirty / 4; | |
333 | ||
334 | return dirty; | |
335 | } | |
336 | ||
337 | /** | |
338 | * zone_dirty_ok - tells whether a zone is within its dirty limits | |
339 | * @zone: the zone to check | |
340 | * | |
341 | * Returns %true when the dirty pages in @zone are within the zone's | |
342 | * dirty limit, %false if the limit is exceeded. | |
343 | */ | |
344 | bool zone_dirty_ok(struct zone *zone) | |
345 | { | |
346 | unsigned long limit = zone_dirty_limit(zone); | |
347 | ||
348 | return zone_page_state(zone, NR_FILE_DIRTY) + | |
349 | zone_page_state(zone, NR_UNSTABLE_NFS) + | |
350 | zone_page_state(zone, NR_WRITEBACK) <= limit; | |
351 | } | |
352 | ||
2da02997 | 353 | int dirty_background_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 354 | void __user *buffer, size_t *lenp, |
2da02997 DR |
355 | loff_t *ppos) |
356 | { | |
357 | int ret; | |
358 | ||
8d65af78 | 359 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
360 | if (ret == 0 && write) |
361 | dirty_background_bytes = 0; | |
362 | return ret; | |
363 | } | |
364 | ||
365 | int dirty_background_bytes_handler(struct ctl_table *table, int write, | |
8d65af78 | 366 | void __user *buffer, size_t *lenp, |
2da02997 DR |
367 | loff_t *ppos) |
368 | { | |
369 | int ret; | |
370 | ||
8d65af78 | 371 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
372 | if (ret == 0 && write) |
373 | dirty_background_ratio = 0; | |
374 | return ret; | |
375 | } | |
376 | ||
04fbfdc1 | 377 | int dirty_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 378 | void __user *buffer, size_t *lenp, |
04fbfdc1 PZ |
379 | loff_t *ppos) |
380 | { | |
381 | int old_ratio = vm_dirty_ratio; | |
2da02997 DR |
382 | int ret; |
383 | ||
8d65af78 | 384 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
04fbfdc1 | 385 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
eb608e3a | 386 | writeback_set_ratelimit(); |
2da02997 DR |
387 | vm_dirty_bytes = 0; |
388 | } | |
389 | return ret; | |
390 | } | |
391 | ||
2da02997 | 392 | int dirty_bytes_handler(struct ctl_table *table, int write, |
8d65af78 | 393 | void __user *buffer, size_t *lenp, |
2da02997 DR |
394 | loff_t *ppos) |
395 | { | |
fc3501d4 | 396 | unsigned long old_bytes = vm_dirty_bytes; |
2da02997 DR |
397 | int ret; |
398 | ||
8d65af78 | 399 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 | 400 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
eb608e3a | 401 | writeback_set_ratelimit(); |
2da02997 | 402 | vm_dirty_ratio = 0; |
04fbfdc1 PZ |
403 | } |
404 | return ret; | |
405 | } | |
406 | ||
eb608e3a JK |
407 | static unsigned long wp_next_time(unsigned long cur_time) |
408 | { | |
409 | cur_time += VM_COMPLETIONS_PERIOD_LEN; | |
410 | /* 0 has a special meaning... */ | |
411 | if (!cur_time) | |
412 | return 1; | |
413 | return cur_time; | |
414 | } | |
415 | ||
04fbfdc1 PZ |
416 | /* |
417 | * Increment the BDI's writeout completion count and the global writeout | |
418 | * completion count. Called from test_clear_page_writeback(). | |
419 | */ | |
420 | static inline void __bdi_writeout_inc(struct backing_dev_info *bdi) | |
421 | { | |
f7d2b1ec | 422 | __inc_bdi_stat(bdi, BDI_WRITTEN); |
eb608e3a JK |
423 | __fprop_inc_percpu_max(&writeout_completions, &bdi->completions, |
424 | bdi->max_prop_frac); | |
425 | /* First event after period switching was turned off? */ | |
426 | if (!unlikely(writeout_period_time)) { | |
427 | /* | |
428 | * We can race with other __bdi_writeout_inc calls here but | |
429 | * it does not cause any harm since the resulting time when | |
430 | * timer will fire and what is in writeout_period_time will be | |
431 | * roughly the same. | |
432 | */ | |
433 | writeout_period_time = wp_next_time(jiffies); | |
434 | mod_timer(&writeout_period_timer, writeout_period_time); | |
435 | } | |
04fbfdc1 PZ |
436 | } |
437 | ||
dd5656e5 MS |
438 | void bdi_writeout_inc(struct backing_dev_info *bdi) |
439 | { | |
440 | unsigned long flags; | |
441 | ||
442 | local_irq_save(flags); | |
443 | __bdi_writeout_inc(bdi); | |
444 | local_irq_restore(flags); | |
445 | } | |
446 | EXPORT_SYMBOL_GPL(bdi_writeout_inc); | |
447 | ||
04fbfdc1 PZ |
448 | /* |
449 | * Obtain an accurate fraction of the BDI's portion. | |
450 | */ | |
451 | static void bdi_writeout_fraction(struct backing_dev_info *bdi, | |
452 | long *numerator, long *denominator) | |
453 | { | |
eb608e3a | 454 | fprop_fraction_percpu(&writeout_completions, &bdi->completions, |
04fbfdc1 | 455 | numerator, denominator); |
04fbfdc1 PZ |
456 | } |
457 | ||
eb608e3a JK |
458 | /* |
459 | * On idle system, we can be called long after we scheduled because we use | |
460 | * deferred timers so count with missed periods. | |
461 | */ | |
462 | static void writeout_period(unsigned long t) | |
463 | { | |
464 | int miss_periods = (jiffies - writeout_period_time) / | |
465 | VM_COMPLETIONS_PERIOD_LEN; | |
466 | ||
467 | if (fprop_new_period(&writeout_completions, miss_periods + 1)) { | |
468 | writeout_period_time = wp_next_time(writeout_period_time + | |
469 | miss_periods * VM_COMPLETIONS_PERIOD_LEN); | |
470 | mod_timer(&writeout_period_timer, writeout_period_time); | |
471 | } else { | |
472 | /* | |
473 | * Aging has zeroed all fractions. Stop wasting CPU on period | |
474 | * updates. | |
475 | */ | |
476 | writeout_period_time = 0; | |
477 | } | |
478 | } | |
479 | ||
189d3c4a | 480 | /* |
d08c429b JW |
481 | * bdi_min_ratio keeps the sum of the minimum dirty shares of all |
482 | * registered backing devices, which, for obvious reasons, can not | |
483 | * exceed 100%. | |
189d3c4a | 484 | */ |
189d3c4a PZ |
485 | static unsigned int bdi_min_ratio; |
486 | ||
487 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) | |
488 | { | |
489 | int ret = 0; | |
189d3c4a | 490 | |
cfc4ba53 | 491 | spin_lock_bh(&bdi_lock); |
a42dde04 | 492 | if (min_ratio > bdi->max_ratio) { |
189d3c4a | 493 | ret = -EINVAL; |
a42dde04 PZ |
494 | } else { |
495 | min_ratio -= bdi->min_ratio; | |
496 | if (bdi_min_ratio + min_ratio < 100) { | |
497 | bdi_min_ratio += min_ratio; | |
498 | bdi->min_ratio += min_ratio; | |
499 | } else { | |
500 | ret = -EINVAL; | |
501 | } | |
502 | } | |
cfc4ba53 | 503 | spin_unlock_bh(&bdi_lock); |
a42dde04 PZ |
504 | |
505 | return ret; | |
506 | } | |
507 | ||
508 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio) | |
509 | { | |
a42dde04 PZ |
510 | int ret = 0; |
511 | ||
512 | if (max_ratio > 100) | |
513 | return -EINVAL; | |
514 | ||
cfc4ba53 | 515 | spin_lock_bh(&bdi_lock); |
a42dde04 PZ |
516 | if (bdi->min_ratio > max_ratio) { |
517 | ret = -EINVAL; | |
518 | } else { | |
519 | bdi->max_ratio = max_ratio; | |
eb608e3a | 520 | bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100; |
a42dde04 | 521 | } |
cfc4ba53 | 522 | spin_unlock_bh(&bdi_lock); |
189d3c4a PZ |
523 | |
524 | return ret; | |
525 | } | |
a42dde04 | 526 | EXPORT_SYMBOL(bdi_set_max_ratio); |
189d3c4a | 527 | |
6c14ae1e WF |
528 | static unsigned long dirty_freerun_ceiling(unsigned long thresh, |
529 | unsigned long bg_thresh) | |
530 | { | |
531 | return (thresh + bg_thresh) / 2; | |
532 | } | |
533 | ||
ffd1f609 WF |
534 | static unsigned long hard_dirty_limit(unsigned long thresh) |
535 | { | |
536 | return max(thresh, global_dirty_limit); | |
537 | } | |
538 | ||
6f718656 | 539 | /** |
1babe183 | 540 | * bdi_dirty_limit - @bdi's share of dirty throttling threshold |
6f718656 WF |
541 | * @bdi: the backing_dev_info to query |
542 | * @dirty: global dirty limit in pages | |
1babe183 | 543 | * |
6f718656 WF |
544 | * Returns @bdi's dirty limit in pages. The term "dirty" in the context of |
545 | * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages. | |
aed21ad2 WF |
546 | * |
547 | * Note that balance_dirty_pages() will only seriously take it as a hard limit | |
548 | * when sleeping max_pause per page is not enough to keep the dirty pages under | |
549 | * control. For example, when the device is completely stalled due to some error | |
550 | * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key. | |
551 | * In the other normal situations, it acts more gently by throttling the tasks | |
552 | * more (rather than completely block them) when the bdi dirty pages go high. | |
1babe183 | 553 | * |
6f718656 | 554 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
1babe183 WF |
555 | * - starving fast devices |
556 | * - piling up dirty pages (that will take long time to sync) on slow devices | |
557 | * | |
558 | * The bdi's share of dirty limit will be adapting to its throughput and | |
559 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. | |
560 | */ | |
561 | unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty) | |
16c4042f WF |
562 | { |
563 | u64 bdi_dirty; | |
564 | long numerator, denominator; | |
04fbfdc1 | 565 | |
16c4042f WF |
566 | /* |
567 | * Calculate this BDI's share of the dirty ratio. | |
568 | */ | |
569 | bdi_writeout_fraction(bdi, &numerator, &denominator); | |
04fbfdc1 | 570 | |
16c4042f WF |
571 | bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100; |
572 | bdi_dirty *= numerator; | |
573 | do_div(bdi_dirty, denominator); | |
04fbfdc1 | 574 | |
16c4042f WF |
575 | bdi_dirty += (dirty * bdi->min_ratio) / 100; |
576 | if (bdi_dirty > (dirty * bdi->max_ratio) / 100) | |
577 | bdi_dirty = dirty * bdi->max_ratio / 100; | |
578 | ||
579 | return bdi_dirty; | |
1da177e4 LT |
580 | } |
581 | ||
6c14ae1e WF |
582 | /* |
583 | * Dirty position control. | |
584 | * | |
585 | * (o) global/bdi setpoints | |
586 | * | |
587 | * We want the dirty pages be balanced around the global/bdi setpoints. | |
588 | * When the number of dirty pages is higher/lower than the setpoint, the | |
589 | * dirty position control ratio (and hence task dirty ratelimit) will be | |
590 | * decreased/increased to bring the dirty pages back to the setpoint. | |
591 | * | |
592 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT | |
593 | * | |
594 | * if (dirty < setpoint) scale up pos_ratio | |
595 | * if (dirty > setpoint) scale down pos_ratio | |
596 | * | |
597 | * if (bdi_dirty < bdi_setpoint) scale up pos_ratio | |
598 | * if (bdi_dirty > bdi_setpoint) scale down pos_ratio | |
599 | * | |
600 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT | |
601 | * | |
602 | * (o) global control line | |
603 | * | |
604 | * ^ pos_ratio | |
605 | * | | |
606 | * | |<===== global dirty control scope ======>| | |
607 | * 2.0 .............* | |
608 | * | .* | |
609 | * | . * | |
610 | * | . * | |
611 | * | . * | |
612 | * | . * | |
613 | * | . * | |
614 | * 1.0 ................................* | |
615 | * | . . * | |
616 | * | . . * | |
617 | * | . . * | |
618 | * | . . * | |
619 | * | . . * | |
620 | * 0 +------------.------------------.----------------------*-------------> | |
621 | * freerun^ setpoint^ limit^ dirty pages | |
622 | * | |
623 | * (o) bdi control line | |
624 | * | |
625 | * ^ pos_ratio | |
626 | * | | |
627 | * | * | |
628 | * | * | |
629 | * | * | |
630 | * | * | |
631 | * | * |<=========== span ============>| | |
632 | * 1.0 .......................* | |
633 | * | . * | |
634 | * | . * | |
635 | * | . * | |
636 | * | . * | |
637 | * | . * | |
638 | * | . * | |
639 | * | . * | |
640 | * | . * | |
641 | * | . * | |
642 | * | . * | |
643 | * | . * | |
644 | * 1/4 ...............................................* * * * * * * * * * * * | |
645 | * | . . | |
646 | * | . . | |
647 | * | . . | |
648 | * 0 +----------------------.-------------------------------.-------------> | |
649 | * bdi_setpoint^ x_intercept^ | |
650 | * | |
651 | * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can | |
652 | * be smoothly throttled down to normal if it starts high in situations like | |
653 | * - start writing to a slow SD card and a fast disk at the same time. The SD | |
654 | * card's bdi_dirty may rush to many times higher than bdi_setpoint. | |
655 | * - the bdi dirty thresh drops quickly due to change of JBOD workload | |
656 | */ | |
657 | static unsigned long bdi_position_ratio(struct backing_dev_info *bdi, | |
658 | unsigned long thresh, | |
659 | unsigned long bg_thresh, | |
660 | unsigned long dirty, | |
661 | unsigned long bdi_thresh, | |
662 | unsigned long bdi_dirty) | |
663 | { | |
664 | unsigned long write_bw = bdi->avg_write_bandwidth; | |
665 | unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh); | |
666 | unsigned long limit = hard_dirty_limit(thresh); | |
667 | unsigned long x_intercept; | |
668 | unsigned long setpoint; /* dirty pages' target balance point */ | |
669 | unsigned long bdi_setpoint; | |
670 | unsigned long span; | |
671 | long long pos_ratio; /* for scaling up/down the rate limit */ | |
672 | long x; | |
673 | ||
674 | if (unlikely(dirty >= limit)) | |
675 | return 0; | |
676 | ||
677 | /* | |
678 | * global setpoint | |
679 | * | |
680 | * setpoint - dirty 3 | |
681 | * f(dirty) := 1.0 + (----------------) | |
682 | * limit - setpoint | |
683 | * | |
684 | * it's a 3rd order polynomial that subjects to | |
685 | * | |
686 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast | |
687 | * (2) f(setpoint) = 1.0 => the balance point | |
688 | * (3) f(limit) = 0 => the hard limit | |
689 | * (4) df/dx <= 0 => negative feedback control | |
690 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) | |
691 | * => fast response on large errors; small oscillation near setpoint | |
692 | */ | |
693 | setpoint = (freerun + limit) / 2; | |
ed84825b | 694 | x = div_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT, |
6c14ae1e WF |
695 | limit - setpoint + 1); |
696 | pos_ratio = x; | |
697 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
698 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
699 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT; | |
700 | ||
701 | /* | |
702 | * We have computed basic pos_ratio above based on global situation. If | |
703 | * the bdi is over/under its share of dirty pages, we want to scale | |
704 | * pos_ratio further down/up. That is done by the following mechanism. | |
705 | */ | |
706 | ||
707 | /* | |
708 | * bdi setpoint | |
709 | * | |
710 | * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint) | |
711 | * | |
712 | * x_intercept - bdi_dirty | |
713 | * := -------------------------- | |
714 | * x_intercept - bdi_setpoint | |
715 | * | |
716 | * The main bdi control line is a linear function that subjects to | |
717 | * | |
718 | * (1) f(bdi_setpoint) = 1.0 | |
719 | * (2) k = - 1 / (8 * write_bw) (in single bdi case) | |
720 | * or equally: x_intercept = bdi_setpoint + 8 * write_bw | |
721 | * | |
722 | * For single bdi case, the dirty pages are observed to fluctuate | |
723 | * regularly within range | |
724 | * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2] | |
725 | * for various filesystems, where (2) can yield in a reasonable 12.5% | |
726 | * fluctuation range for pos_ratio. | |
727 | * | |
728 | * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its | |
729 | * own size, so move the slope over accordingly and choose a slope that | |
730 | * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh. | |
731 | */ | |
732 | if (unlikely(bdi_thresh > thresh)) | |
733 | bdi_thresh = thresh; | |
aed21ad2 WF |
734 | /* |
735 | * It's very possible that bdi_thresh is close to 0 not because the | |
736 | * device is slow, but that it has remained inactive for long time. | |
737 | * Honour such devices a reasonable good (hopefully IO efficient) | |
738 | * threshold, so that the occasional writes won't be blocked and active | |
739 | * writes can rampup the threshold quickly. | |
740 | */ | |
8927f66c | 741 | bdi_thresh = max(bdi_thresh, (limit - dirty) / 8); |
6c14ae1e WF |
742 | /* |
743 | * scale global setpoint to bdi's: | |
744 | * bdi_setpoint = setpoint * bdi_thresh / thresh | |
745 | */ | |
746 | x = div_u64((u64)bdi_thresh << 16, thresh + 1); | |
747 | bdi_setpoint = setpoint * (u64)x >> 16; | |
748 | /* | |
749 | * Use span=(8*write_bw) in single bdi case as indicated by | |
750 | * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case. | |
751 | * | |
752 | * bdi_thresh thresh - bdi_thresh | |
753 | * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh | |
754 | * thresh thresh | |
755 | */ | |
756 | span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16; | |
757 | x_intercept = bdi_setpoint + span; | |
758 | ||
759 | if (bdi_dirty < x_intercept - span / 4) { | |
50657fc4 WF |
760 | pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty), |
761 | x_intercept - bdi_setpoint + 1); | |
6c14ae1e WF |
762 | } else |
763 | pos_ratio /= 4; | |
764 | ||
8927f66c WF |
765 | /* |
766 | * bdi reserve area, safeguard against dirty pool underrun and disk idle | |
767 | * It may push the desired control point of global dirty pages higher | |
768 | * than setpoint. | |
769 | */ | |
770 | x_intercept = bdi_thresh / 2; | |
771 | if (bdi_dirty < x_intercept) { | |
50657fc4 WF |
772 | if (bdi_dirty > x_intercept / 8) |
773 | pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty); | |
774 | else | |
8927f66c WF |
775 | pos_ratio *= 8; |
776 | } | |
777 | ||
6c14ae1e WF |
778 | return pos_ratio; |
779 | } | |
780 | ||
e98be2d5 WF |
781 | static void bdi_update_write_bandwidth(struct backing_dev_info *bdi, |
782 | unsigned long elapsed, | |
783 | unsigned long written) | |
784 | { | |
785 | const unsigned long period = roundup_pow_of_two(3 * HZ); | |
786 | unsigned long avg = bdi->avg_write_bandwidth; | |
787 | unsigned long old = bdi->write_bandwidth; | |
788 | u64 bw; | |
789 | ||
790 | /* | |
791 | * bw = written * HZ / elapsed | |
792 | * | |
793 | * bw * elapsed + write_bandwidth * (period - elapsed) | |
794 | * write_bandwidth = --------------------------------------------------- | |
795 | * period | |
e58126f5 TH |
796 | * |
797 | * @written may have decreased due to account_page_redirty(). | |
798 | * Avoid underflowing @bw calculation. | |
e98be2d5 | 799 | */ |
e58126f5 | 800 | bw = written - min(written, bdi->written_stamp); |
e98be2d5 WF |
801 | bw *= HZ; |
802 | if (unlikely(elapsed > period)) { | |
803 | do_div(bw, elapsed); | |
804 | avg = bw; | |
805 | goto out; | |
806 | } | |
807 | bw += (u64)bdi->write_bandwidth * (period - elapsed); | |
808 | bw >>= ilog2(period); | |
809 | ||
810 | /* | |
811 | * one more level of smoothing, for filtering out sudden spikes | |
812 | */ | |
813 | if (avg > old && old >= (unsigned long)bw) | |
814 | avg -= (avg - old) >> 3; | |
815 | ||
816 | if (avg < old && old <= (unsigned long)bw) | |
817 | avg += (old - avg) >> 3; | |
818 | ||
819 | out: | |
820 | bdi->write_bandwidth = bw; | |
821 | bdi->avg_write_bandwidth = avg; | |
822 | } | |
823 | ||
c42843f2 WF |
824 | /* |
825 | * The global dirtyable memory and dirty threshold could be suddenly knocked | |
826 | * down by a large amount (eg. on the startup of KVM in a swapless system). | |
827 | * This may throw the system into deep dirty exceeded state and throttle | |
828 | * heavy/light dirtiers alike. To retain good responsiveness, maintain | |
829 | * global_dirty_limit for tracking slowly down to the knocked down dirty | |
830 | * threshold. | |
831 | */ | |
832 | static void update_dirty_limit(unsigned long thresh, unsigned long dirty) | |
833 | { | |
834 | unsigned long limit = global_dirty_limit; | |
835 | ||
836 | /* | |
837 | * Follow up in one step. | |
838 | */ | |
839 | if (limit < thresh) { | |
840 | limit = thresh; | |
841 | goto update; | |
842 | } | |
843 | ||
844 | /* | |
845 | * Follow down slowly. Use the higher one as the target, because thresh | |
846 | * may drop below dirty. This is exactly the reason to introduce | |
847 | * global_dirty_limit which is guaranteed to lie above the dirty pages. | |
848 | */ | |
849 | thresh = max(thresh, dirty); | |
850 | if (limit > thresh) { | |
851 | limit -= (limit - thresh) >> 5; | |
852 | goto update; | |
853 | } | |
854 | return; | |
855 | update: | |
856 | global_dirty_limit = limit; | |
857 | } | |
858 | ||
859 | static void global_update_bandwidth(unsigned long thresh, | |
860 | unsigned long dirty, | |
861 | unsigned long now) | |
862 | { | |
863 | static DEFINE_SPINLOCK(dirty_lock); | |
f1667836 | 864 | static unsigned long update_time = INITIAL_JIFFIES; |
c42843f2 WF |
865 | |
866 | /* | |
867 | * check locklessly first to optimize away locking for the most time | |
868 | */ | |
869 | if (time_before(now, update_time + BANDWIDTH_INTERVAL)) | |
870 | return; | |
871 | ||
872 | spin_lock(&dirty_lock); | |
873 | if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) { | |
874 | update_dirty_limit(thresh, dirty); | |
875 | update_time = now; | |
876 | } | |
877 | spin_unlock(&dirty_lock); | |
878 | } | |
879 | ||
be3ffa27 WF |
880 | /* |
881 | * Maintain bdi->dirty_ratelimit, the base dirty throttle rate. | |
882 | * | |
883 | * Normal bdi tasks will be curbed at or below it in long term. | |
884 | * Obviously it should be around (write_bw / N) when there are N dd tasks. | |
885 | */ | |
886 | static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi, | |
887 | unsigned long thresh, | |
888 | unsigned long bg_thresh, | |
889 | unsigned long dirty, | |
890 | unsigned long bdi_thresh, | |
891 | unsigned long bdi_dirty, | |
892 | unsigned long dirtied, | |
893 | unsigned long elapsed) | |
894 | { | |
7381131c WF |
895 | unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh); |
896 | unsigned long limit = hard_dirty_limit(thresh); | |
897 | unsigned long setpoint = (freerun + limit) / 2; | |
be3ffa27 WF |
898 | unsigned long write_bw = bdi->avg_write_bandwidth; |
899 | unsigned long dirty_ratelimit = bdi->dirty_ratelimit; | |
900 | unsigned long dirty_rate; | |
901 | unsigned long task_ratelimit; | |
902 | unsigned long balanced_dirty_ratelimit; | |
903 | unsigned long pos_ratio; | |
7381131c WF |
904 | unsigned long step; |
905 | unsigned long x; | |
be3ffa27 WF |
906 | |
907 | /* | |
908 | * The dirty rate will match the writeout rate in long term, except | |
909 | * when dirty pages are truncated by userspace or re-dirtied by FS. | |
910 | */ | |
911 | dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed; | |
912 | ||
913 | pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty, | |
914 | bdi_thresh, bdi_dirty); | |
915 | /* | |
916 | * task_ratelimit reflects each dd's dirty rate for the past 200ms. | |
917 | */ | |
918 | task_ratelimit = (u64)dirty_ratelimit * | |
919 | pos_ratio >> RATELIMIT_CALC_SHIFT; | |
920 | task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ | |
921 | ||
922 | /* | |
923 | * A linear estimation of the "balanced" throttle rate. The theory is, | |
924 | * if there are N dd tasks, each throttled at task_ratelimit, the bdi's | |
925 | * dirty_rate will be measured to be (N * task_ratelimit). So the below | |
926 | * formula will yield the balanced rate limit (write_bw / N). | |
927 | * | |
928 | * Note that the expanded form is not a pure rate feedback: | |
929 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) | |
930 | * but also takes pos_ratio into account: | |
931 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) | |
932 | * | |
933 | * (1) is not realistic because pos_ratio also takes part in balancing | |
934 | * the dirty rate. Consider the state | |
935 | * pos_ratio = 0.5 (3) | |
936 | * rate = 2 * (write_bw / N) (4) | |
937 | * If (1) is used, it will stuck in that state! Because each dd will | |
938 | * be throttled at | |
939 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) | |
940 | * yielding | |
941 | * dirty_rate = N * task_ratelimit = write_bw (6) | |
942 | * put (6) into (1) we get | |
943 | * rate_(i+1) = rate_(i) (7) | |
944 | * | |
945 | * So we end up using (2) to always keep | |
946 | * rate_(i+1) ~= (write_bw / N) (8) | |
947 | * regardless of the value of pos_ratio. As long as (8) is satisfied, | |
948 | * pos_ratio is able to drive itself to 1.0, which is not only where | |
949 | * the dirty count meet the setpoint, but also where the slope of | |
950 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated. | |
951 | */ | |
952 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, | |
953 | dirty_rate | 1); | |
bdaac490 WF |
954 | /* |
955 | * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw | |
956 | */ | |
957 | if (unlikely(balanced_dirty_ratelimit > write_bw)) | |
958 | balanced_dirty_ratelimit = write_bw; | |
be3ffa27 | 959 | |
7381131c WF |
960 | /* |
961 | * We could safely do this and return immediately: | |
962 | * | |
963 | * bdi->dirty_ratelimit = balanced_dirty_ratelimit; | |
964 | * | |
965 | * However to get a more stable dirty_ratelimit, the below elaborated | |
331cbdee | 966 | * code makes use of task_ratelimit to filter out singular points and |
7381131c WF |
967 | * limit the step size. |
968 | * | |
969 | * The below code essentially only uses the relative value of | |
970 | * | |
971 | * task_ratelimit - dirty_ratelimit | |
972 | * = (pos_ratio - 1) * dirty_ratelimit | |
973 | * | |
974 | * which reflects the direction and size of dirty position error. | |
975 | */ | |
976 | ||
977 | /* | |
978 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff | |
979 | * task_ratelimit is on the same side of dirty_ratelimit, too. | |
980 | * For example, when | |
981 | * - dirty_ratelimit > balanced_dirty_ratelimit | |
982 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) | |
983 | * lowering dirty_ratelimit will help meet both the position and rate | |
984 | * control targets. Otherwise, don't update dirty_ratelimit if it will | |
985 | * only help meet the rate target. After all, what the users ultimately | |
986 | * feel and care are stable dirty rate and small position error. | |
987 | * | |
988 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size | |
331cbdee | 989 | * and filter out the singular points of balanced_dirty_ratelimit. Which |
7381131c WF |
990 | * keeps jumping around randomly and can even leap far away at times |
991 | * due to the small 200ms estimation period of dirty_rate (we want to | |
992 | * keep that period small to reduce time lags). | |
993 | */ | |
994 | step = 0; | |
995 | if (dirty < setpoint) { | |
996 | x = min(bdi->balanced_dirty_ratelimit, | |
997 | min(balanced_dirty_ratelimit, task_ratelimit)); | |
998 | if (dirty_ratelimit < x) | |
999 | step = x - dirty_ratelimit; | |
1000 | } else { | |
1001 | x = max(bdi->balanced_dirty_ratelimit, | |
1002 | max(balanced_dirty_ratelimit, task_ratelimit)); | |
1003 | if (dirty_ratelimit > x) | |
1004 | step = dirty_ratelimit - x; | |
1005 | } | |
1006 | ||
1007 | /* | |
1008 | * Don't pursue 100% rate matching. It's impossible since the balanced | |
1009 | * rate itself is constantly fluctuating. So decrease the track speed | |
1010 | * when it gets close to the target. Helps eliminate pointless tremors. | |
1011 | */ | |
1012 | step >>= dirty_ratelimit / (2 * step + 1); | |
1013 | /* | |
1014 | * Limit the tracking speed to avoid overshooting. | |
1015 | */ | |
1016 | step = (step + 7) / 8; | |
1017 | ||
1018 | if (dirty_ratelimit < balanced_dirty_ratelimit) | |
1019 | dirty_ratelimit += step; | |
1020 | else | |
1021 | dirty_ratelimit -= step; | |
1022 | ||
1023 | bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL); | |
1024 | bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit; | |
b48c104d WF |
1025 | |
1026 | trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit); | |
be3ffa27 WF |
1027 | } |
1028 | ||
e98be2d5 | 1029 | void __bdi_update_bandwidth(struct backing_dev_info *bdi, |
c42843f2 | 1030 | unsigned long thresh, |
af6a3113 | 1031 | unsigned long bg_thresh, |
c42843f2 WF |
1032 | unsigned long dirty, |
1033 | unsigned long bdi_thresh, | |
1034 | unsigned long bdi_dirty, | |
e98be2d5 WF |
1035 | unsigned long start_time) |
1036 | { | |
1037 | unsigned long now = jiffies; | |
1038 | unsigned long elapsed = now - bdi->bw_time_stamp; | |
be3ffa27 | 1039 | unsigned long dirtied; |
e98be2d5 WF |
1040 | unsigned long written; |
1041 | ||
1042 | /* | |
1043 | * rate-limit, only update once every 200ms. | |
1044 | */ | |
1045 | if (elapsed < BANDWIDTH_INTERVAL) | |
1046 | return; | |
1047 | ||
be3ffa27 | 1048 | dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]); |
e98be2d5 WF |
1049 | written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]); |
1050 | ||
1051 | /* | |
1052 | * Skip quiet periods when disk bandwidth is under-utilized. | |
1053 | * (at least 1s idle time between two flusher runs) | |
1054 | */ | |
1055 | if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time)) | |
1056 | goto snapshot; | |
1057 | ||
be3ffa27 | 1058 | if (thresh) { |
c42843f2 | 1059 | global_update_bandwidth(thresh, dirty, now); |
be3ffa27 WF |
1060 | bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty, |
1061 | bdi_thresh, bdi_dirty, | |
1062 | dirtied, elapsed); | |
1063 | } | |
e98be2d5 WF |
1064 | bdi_update_write_bandwidth(bdi, elapsed, written); |
1065 | ||
1066 | snapshot: | |
be3ffa27 | 1067 | bdi->dirtied_stamp = dirtied; |
e98be2d5 WF |
1068 | bdi->written_stamp = written; |
1069 | bdi->bw_time_stamp = now; | |
1070 | } | |
1071 | ||
1072 | static void bdi_update_bandwidth(struct backing_dev_info *bdi, | |
c42843f2 | 1073 | unsigned long thresh, |
af6a3113 | 1074 | unsigned long bg_thresh, |
c42843f2 WF |
1075 | unsigned long dirty, |
1076 | unsigned long bdi_thresh, | |
1077 | unsigned long bdi_dirty, | |
e98be2d5 WF |
1078 | unsigned long start_time) |
1079 | { | |
1080 | if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL)) | |
1081 | return; | |
1082 | spin_lock(&bdi->wb.list_lock); | |
af6a3113 WF |
1083 | __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty, |
1084 | bdi_thresh, bdi_dirty, start_time); | |
e98be2d5 WF |
1085 | spin_unlock(&bdi->wb.list_lock); |
1086 | } | |
1087 | ||
9d823e8f | 1088 | /* |
d0e1d66b | 1089 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited() |
9d823e8f WF |
1090 | * will look to see if it needs to start dirty throttling. |
1091 | * | |
1092 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive | |
1093 | * global_page_state() too often. So scale it near-sqrt to the safety margin | |
1094 | * (the number of pages we may dirty without exceeding the dirty limits). | |
1095 | */ | |
1096 | static unsigned long dirty_poll_interval(unsigned long dirty, | |
1097 | unsigned long thresh) | |
1098 | { | |
1099 | if (thresh > dirty) | |
1100 | return 1UL << (ilog2(thresh - dirty) >> 1); | |
1101 | ||
1102 | return 1; | |
1103 | } | |
1104 | ||
71c908a0 FW |
1105 | static unsigned long bdi_max_pause(struct backing_dev_info *bdi, |
1106 | unsigned long bdi_dirty) | |
c8462cc9 | 1107 | { |
71c908a0 FW |
1108 | unsigned long bw = bdi->avg_write_bandwidth; |
1109 | unsigned long t; | |
c8462cc9 | 1110 | |
7ccb9ad5 WF |
1111 | /* |
1112 | * Limit pause time for small memory systems. If sleeping for too long | |
1113 | * time, a small pool of dirty/writeback pages may go empty and disk go | |
1114 | * idle. | |
1115 | * | |
1116 | * 8 serves as the safety ratio. | |
1117 | */ | |
1118 | t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8)); | |
1119 | t++; | |
1120 | ||
71c908a0 | 1121 | return min_t(unsigned long, t, MAX_PAUSE); |
7ccb9ad5 WF |
1122 | } |
1123 | ||
1124 | static long bdi_min_pause(struct backing_dev_info *bdi, | |
1125 | long max_pause, | |
1126 | unsigned long task_ratelimit, | |
1127 | unsigned long dirty_ratelimit, | |
1128 | int *nr_dirtied_pause) | |
c8462cc9 | 1129 | { |
7ccb9ad5 WF |
1130 | long hi = ilog2(bdi->avg_write_bandwidth); |
1131 | long lo = ilog2(bdi->dirty_ratelimit); | |
1132 | long t; /* target pause */ | |
1133 | long pause; /* estimated next pause */ | |
1134 | int pages; /* target nr_dirtied_pause */ | |
c8462cc9 | 1135 | |
7ccb9ad5 WF |
1136 | /* target for 10ms pause on 1-dd case */ |
1137 | t = max(1, HZ / 100); | |
c8462cc9 WF |
1138 | |
1139 | /* | |
1140 | * Scale up pause time for concurrent dirtiers in order to reduce CPU | |
1141 | * overheads. | |
1142 | * | |
7ccb9ad5 | 1143 | * (N * 10ms) on 2^N concurrent tasks. |
c8462cc9 WF |
1144 | */ |
1145 | if (hi > lo) | |
7ccb9ad5 | 1146 | t += (hi - lo) * (10 * HZ) / 1024; |
c8462cc9 WF |
1147 | |
1148 | /* | |
7ccb9ad5 WF |
1149 | * This is a bit convoluted. We try to base the next nr_dirtied_pause |
1150 | * on the much more stable dirty_ratelimit. However the next pause time | |
1151 | * will be computed based on task_ratelimit and the two rate limits may | |
1152 | * depart considerably at some time. Especially if task_ratelimit goes | |
1153 | * below dirty_ratelimit/2 and the target pause is max_pause, the next | |
1154 | * pause time will be max_pause*2 _trimmed down_ to max_pause. As a | |
1155 | * result task_ratelimit won't be executed faithfully, which could | |
1156 | * eventually bring down dirty_ratelimit. | |
c8462cc9 | 1157 | * |
7ccb9ad5 WF |
1158 | * We apply two rules to fix it up: |
1159 | * 1) try to estimate the next pause time and if necessary, use a lower | |
1160 | * nr_dirtied_pause so as not to exceed max_pause. When this happens, | |
1161 | * nr_dirtied_pause will be "dancing" with task_ratelimit. | |
1162 | * 2) limit the target pause time to max_pause/2, so that the normal | |
1163 | * small fluctuations of task_ratelimit won't trigger rule (1) and | |
1164 | * nr_dirtied_pause will remain as stable as dirty_ratelimit. | |
c8462cc9 | 1165 | */ |
7ccb9ad5 WF |
1166 | t = min(t, 1 + max_pause / 2); |
1167 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
c8462cc9 WF |
1168 | |
1169 | /* | |
5b9b3574 WF |
1170 | * Tiny nr_dirtied_pause is found to hurt I/O performance in the test |
1171 | * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}. | |
1172 | * When the 16 consecutive reads are often interrupted by some dirty | |
1173 | * throttling pause during the async writes, cfq will go into idles | |
1174 | * (deadline is fine). So push nr_dirtied_pause as high as possible | |
1175 | * until reaches DIRTY_POLL_THRESH=32 pages. | |
c8462cc9 | 1176 | */ |
5b9b3574 WF |
1177 | if (pages < DIRTY_POLL_THRESH) { |
1178 | t = max_pause; | |
1179 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
1180 | if (pages > DIRTY_POLL_THRESH) { | |
1181 | pages = DIRTY_POLL_THRESH; | |
1182 | t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit; | |
1183 | } | |
1184 | } | |
1185 | ||
7ccb9ad5 WF |
1186 | pause = HZ * pages / (task_ratelimit + 1); |
1187 | if (pause > max_pause) { | |
1188 | t = max_pause; | |
1189 | pages = task_ratelimit * t / roundup_pow_of_two(HZ); | |
1190 | } | |
c8462cc9 | 1191 | |
7ccb9ad5 | 1192 | *nr_dirtied_pause = pages; |
c8462cc9 | 1193 | /* |
7ccb9ad5 | 1194 | * The minimal pause time will normally be half the target pause time. |
c8462cc9 | 1195 | */ |
5b9b3574 | 1196 | return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t; |
c8462cc9 WF |
1197 | } |
1198 | ||
1da177e4 LT |
1199 | /* |
1200 | * balance_dirty_pages() must be called by processes which are generating dirty | |
1201 | * data. It looks at the number of dirty pages in the machine and will force | |
143dfe86 | 1202 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. |
5b0830cb JA |
1203 | * If we're over `background_thresh' then the writeback threads are woken to |
1204 | * perform some writeout. | |
1da177e4 | 1205 | */ |
3a2e9a5a | 1206 | static void balance_dirty_pages(struct address_space *mapping, |
143dfe86 | 1207 | unsigned long pages_dirtied) |
1da177e4 | 1208 | { |
143dfe86 WF |
1209 | unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */ |
1210 | unsigned long bdi_reclaimable; | |
7762741e WF |
1211 | unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */ |
1212 | unsigned long bdi_dirty; | |
6c14ae1e | 1213 | unsigned long freerun; |
364aeb28 DR |
1214 | unsigned long background_thresh; |
1215 | unsigned long dirty_thresh; | |
1216 | unsigned long bdi_thresh; | |
83712358 | 1217 | long period; |
7ccb9ad5 WF |
1218 | long pause; |
1219 | long max_pause; | |
1220 | long min_pause; | |
1221 | int nr_dirtied_pause; | |
e50e3720 | 1222 | bool dirty_exceeded = false; |
143dfe86 | 1223 | unsigned long task_ratelimit; |
7ccb9ad5 | 1224 | unsigned long dirty_ratelimit; |
143dfe86 | 1225 | unsigned long pos_ratio; |
1da177e4 | 1226 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
e98be2d5 | 1227 | unsigned long start_time = jiffies; |
1da177e4 LT |
1228 | |
1229 | for (;;) { | |
83712358 WF |
1230 | unsigned long now = jiffies; |
1231 | ||
143dfe86 WF |
1232 | /* |
1233 | * Unstable writes are a feature of certain networked | |
1234 | * filesystems (i.e. NFS) in which data may have been | |
1235 | * written to the server's write cache, but has not yet | |
1236 | * been flushed to permanent storage. | |
1237 | */ | |
5fce25a9 PZ |
1238 | nr_reclaimable = global_page_state(NR_FILE_DIRTY) + |
1239 | global_page_state(NR_UNSTABLE_NFS); | |
7762741e | 1240 | nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK); |
5fce25a9 | 1241 | |
16c4042f WF |
1242 | global_dirty_limits(&background_thresh, &dirty_thresh); |
1243 | ||
1244 | /* | |
1245 | * Throttle it only when the background writeback cannot | |
1246 | * catch-up. This avoids (excessively) small writeouts | |
1247 | * when the bdi limits are ramping up. | |
1248 | */ | |
6c14ae1e WF |
1249 | freerun = dirty_freerun_ceiling(dirty_thresh, |
1250 | background_thresh); | |
83712358 WF |
1251 | if (nr_dirty <= freerun) { |
1252 | current->dirty_paused_when = now; | |
1253 | current->nr_dirtied = 0; | |
7ccb9ad5 WF |
1254 | current->nr_dirtied_pause = |
1255 | dirty_poll_interval(nr_dirty, dirty_thresh); | |
16c4042f | 1256 | break; |
83712358 | 1257 | } |
16c4042f | 1258 | |
143dfe86 WF |
1259 | if (unlikely(!writeback_in_progress(bdi))) |
1260 | bdi_start_background_writeback(bdi); | |
1261 | ||
1262 | /* | |
1263 | * bdi_thresh is not treated as some limiting factor as | |
1264 | * dirty_thresh, due to reasons | |
1265 | * - in JBOD setup, bdi_thresh can fluctuate a lot | |
1266 | * - in a system with HDD and USB key, the USB key may somehow | |
1267 | * go into state (bdi_dirty >> bdi_thresh) either because | |
1268 | * bdi_dirty starts high, or because bdi_thresh drops low. | |
1269 | * In this case we don't want to hard throttle the USB key | |
1270 | * dirtiers for 100 seconds until bdi_dirty drops under | |
1271 | * bdi_thresh. Instead the auxiliary bdi control line in | |
1272 | * bdi_position_ratio() will let the dirtier task progress | |
1273 | * at some rate <= (write_bw / 2) for bringing down bdi_dirty. | |
1274 | */ | |
16c4042f | 1275 | bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh); |
16c4042f | 1276 | |
e50e3720 WF |
1277 | /* |
1278 | * In order to avoid the stacked BDI deadlock we need | |
1279 | * to ensure we accurately count the 'dirty' pages when | |
1280 | * the threshold is low. | |
1281 | * | |
1282 | * Otherwise it would be possible to get thresh+n pages | |
1283 | * reported dirty, even though there are thresh-m pages | |
1284 | * actually dirty; with m+n sitting in the percpu | |
1285 | * deltas. | |
1286 | */ | |
143dfe86 WF |
1287 | if (bdi_thresh < 2 * bdi_stat_error(bdi)) { |
1288 | bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE); | |
1289 | bdi_dirty = bdi_reclaimable + | |
7762741e | 1290 | bdi_stat_sum(bdi, BDI_WRITEBACK); |
e50e3720 | 1291 | } else { |
143dfe86 WF |
1292 | bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE); |
1293 | bdi_dirty = bdi_reclaimable + | |
7762741e | 1294 | bdi_stat(bdi, BDI_WRITEBACK); |
e50e3720 | 1295 | } |
5fce25a9 | 1296 | |
82791940 | 1297 | dirty_exceeded = (bdi_dirty > bdi_thresh) && |
7762741e | 1298 | (nr_dirty > dirty_thresh); |
143dfe86 | 1299 | if (dirty_exceeded && !bdi->dirty_exceeded) |
04fbfdc1 | 1300 | bdi->dirty_exceeded = 1; |
1da177e4 | 1301 | |
af6a3113 WF |
1302 | bdi_update_bandwidth(bdi, dirty_thresh, background_thresh, |
1303 | nr_dirty, bdi_thresh, bdi_dirty, | |
1304 | start_time); | |
e98be2d5 | 1305 | |
143dfe86 WF |
1306 | dirty_ratelimit = bdi->dirty_ratelimit; |
1307 | pos_ratio = bdi_position_ratio(bdi, dirty_thresh, | |
1308 | background_thresh, nr_dirty, | |
1309 | bdi_thresh, bdi_dirty); | |
3a73dbbc WF |
1310 | task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >> |
1311 | RATELIMIT_CALC_SHIFT; | |
7ccb9ad5 WF |
1312 | max_pause = bdi_max_pause(bdi, bdi_dirty); |
1313 | min_pause = bdi_min_pause(bdi, max_pause, | |
1314 | task_ratelimit, dirty_ratelimit, | |
1315 | &nr_dirtied_pause); | |
1316 | ||
3a73dbbc | 1317 | if (unlikely(task_ratelimit == 0)) { |
83712358 | 1318 | period = max_pause; |
c8462cc9 | 1319 | pause = max_pause; |
143dfe86 | 1320 | goto pause; |
04fbfdc1 | 1321 | } |
83712358 WF |
1322 | period = HZ * pages_dirtied / task_ratelimit; |
1323 | pause = period; | |
1324 | if (current->dirty_paused_when) | |
1325 | pause -= now - current->dirty_paused_when; | |
1326 | /* | |
1327 | * For less than 1s think time (ext3/4 may block the dirtier | |
1328 | * for up to 800ms from time to time on 1-HDD; so does xfs, | |
1329 | * however at much less frequency), try to compensate it in | |
1330 | * future periods by updating the virtual time; otherwise just | |
1331 | * do a reset, as it may be a light dirtier. | |
1332 | */ | |
7ccb9ad5 | 1333 | if (pause < min_pause) { |
ece13ac3 WF |
1334 | trace_balance_dirty_pages(bdi, |
1335 | dirty_thresh, | |
1336 | background_thresh, | |
1337 | nr_dirty, | |
1338 | bdi_thresh, | |
1339 | bdi_dirty, | |
1340 | dirty_ratelimit, | |
1341 | task_ratelimit, | |
1342 | pages_dirtied, | |
83712358 | 1343 | period, |
7ccb9ad5 | 1344 | min(pause, 0L), |
ece13ac3 | 1345 | start_time); |
83712358 WF |
1346 | if (pause < -HZ) { |
1347 | current->dirty_paused_when = now; | |
1348 | current->nr_dirtied = 0; | |
1349 | } else if (period) { | |
1350 | current->dirty_paused_when += period; | |
1351 | current->nr_dirtied = 0; | |
7ccb9ad5 WF |
1352 | } else if (current->nr_dirtied_pause <= pages_dirtied) |
1353 | current->nr_dirtied_pause += pages_dirtied; | |
57fc978c | 1354 | break; |
04fbfdc1 | 1355 | } |
7ccb9ad5 WF |
1356 | if (unlikely(pause > max_pause)) { |
1357 | /* for occasional dropped task_ratelimit */ | |
1358 | now += min(pause - max_pause, max_pause); | |
1359 | pause = max_pause; | |
1360 | } | |
143dfe86 WF |
1361 | |
1362 | pause: | |
ece13ac3 WF |
1363 | trace_balance_dirty_pages(bdi, |
1364 | dirty_thresh, | |
1365 | background_thresh, | |
1366 | nr_dirty, | |
1367 | bdi_thresh, | |
1368 | bdi_dirty, | |
1369 | dirty_ratelimit, | |
1370 | task_ratelimit, | |
1371 | pages_dirtied, | |
83712358 | 1372 | period, |
ece13ac3 WF |
1373 | pause, |
1374 | start_time); | |
499d05ec | 1375 | __set_current_state(TASK_KILLABLE); |
d25105e8 | 1376 | io_schedule_timeout(pause); |
87c6a9b2 | 1377 | |
83712358 WF |
1378 | current->dirty_paused_when = now + pause; |
1379 | current->nr_dirtied = 0; | |
7ccb9ad5 | 1380 | current->nr_dirtied_pause = nr_dirtied_pause; |
83712358 | 1381 | |
ffd1f609 | 1382 | /* |
1df64719 WF |
1383 | * This is typically equal to (nr_dirty < dirty_thresh) and can |
1384 | * also keep "1000+ dd on a slow USB stick" under control. | |
ffd1f609 | 1385 | */ |
1df64719 | 1386 | if (task_ratelimit) |
ffd1f609 | 1387 | break; |
499d05ec | 1388 | |
c5c6343c WF |
1389 | /* |
1390 | * In the case of an unresponding NFS server and the NFS dirty | |
1391 | * pages exceeds dirty_thresh, give the other good bdi's a pipe | |
1392 | * to go through, so that tasks on them still remain responsive. | |
1393 | * | |
1394 | * In theory 1 page is enough to keep the comsumer-producer | |
1395 | * pipe going: the flusher cleans 1 page => the task dirties 1 | |
1396 | * more page. However bdi_dirty has accounting errors. So use | |
1397 | * the larger and more IO friendly bdi_stat_error. | |
1398 | */ | |
1399 | if (bdi_dirty <= bdi_stat_error(bdi)) | |
1400 | break; | |
1401 | ||
499d05ec JK |
1402 | if (fatal_signal_pending(current)) |
1403 | break; | |
1da177e4 LT |
1404 | } |
1405 | ||
143dfe86 | 1406 | if (!dirty_exceeded && bdi->dirty_exceeded) |
04fbfdc1 | 1407 | bdi->dirty_exceeded = 0; |
1da177e4 LT |
1408 | |
1409 | if (writeback_in_progress(bdi)) | |
5b0830cb | 1410 | return; |
1da177e4 LT |
1411 | |
1412 | /* | |
1413 | * In laptop mode, we wait until hitting the higher threshold before | |
1414 | * starting background writeout, and then write out all the way down | |
1415 | * to the lower threshold. So slow writers cause minimal disk activity. | |
1416 | * | |
1417 | * In normal mode, we start background writeout at the lower | |
1418 | * background_thresh, to keep the amount of dirty memory low. | |
1419 | */ | |
143dfe86 WF |
1420 | if (laptop_mode) |
1421 | return; | |
1422 | ||
1423 | if (nr_reclaimable > background_thresh) | |
c5444198 | 1424 | bdi_start_background_writeback(bdi); |
1da177e4 LT |
1425 | } |
1426 | ||
a200ee18 | 1427 | void set_page_dirty_balance(struct page *page, int page_mkwrite) |
edc79b2a | 1428 | { |
a200ee18 | 1429 | if (set_page_dirty(page) || page_mkwrite) { |
edc79b2a PZ |
1430 | struct address_space *mapping = page_mapping(page); |
1431 | ||
1432 | if (mapping) | |
1433 | balance_dirty_pages_ratelimited(mapping); | |
1434 | } | |
1435 | } | |
1436 | ||
9d823e8f | 1437 | static DEFINE_PER_CPU(int, bdp_ratelimits); |
245b2e70 | 1438 | |
54848d73 WF |
1439 | /* |
1440 | * Normal tasks are throttled by | |
1441 | * loop { | |
1442 | * dirty tsk->nr_dirtied_pause pages; | |
1443 | * take a snap in balance_dirty_pages(); | |
1444 | * } | |
1445 | * However there is a worst case. If every task exit immediately when dirtied | |
1446 | * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be | |
1447 | * called to throttle the page dirties. The solution is to save the not yet | |
1448 | * throttled page dirties in dirty_throttle_leaks on task exit and charge them | |
1449 | * randomly into the running tasks. This works well for the above worst case, | |
1450 | * as the new task will pick up and accumulate the old task's leaked dirty | |
1451 | * count and eventually get throttled. | |
1452 | */ | |
1453 | DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0; | |
1454 | ||
1da177e4 | 1455 | /** |
d0e1d66b | 1456 | * balance_dirty_pages_ratelimited - balance dirty memory state |
67be2dd1 | 1457 | * @mapping: address_space which was dirtied |
1da177e4 LT |
1458 | * |
1459 | * Processes which are dirtying memory should call in here once for each page | |
1460 | * which was newly dirtied. The function will periodically check the system's | |
1461 | * dirty state and will initiate writeback if needed. | |
1462 | * | |
1463 | * On really big machines, get_writeback_state is expensive, so try to avoid | |
1464 | * calling it too often (ratelimiting). But once we're over the dirty memory | |
1465 | * limit we decrease the ratelimiting by a lot, to prevent individual processes | |
1466 | * from overshooting the limit by (ratelimit_pages) each. | |
1467 | */ | |
d0e1d66b | 1468 | void balance_dirty_pages_ratelimited(struct address_space *mapping) |
1da177e4 | 1469 | { |
36715cef | 1470 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
9d823e8f WF |
1471 | int ratelimit; |
1472 | int *p; | |
1da177e4 | 1473 | |
36715cef WF |
1474 | if (!bdi_cap_account_dirty(bdi)) |
1475 | return; | |
1476 | ||
9d823e8f WF |
1477 | ratelimit = current->nr_dirtied_pause; |
1478 | if (bdi->dirty_exceeded) | |
1479 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); | |
1480 | ||
9d823e8f | 1481 | preempt_disable(); |
1da177e4 | 1482 | /* |
9d823e8f WF |
1483 | * This prevents one CPU to accumulate too many dirtied pages without |
1484 | * calling into balance_dirty_pages(), which can happen when there are | |
1485 | * 1000+ tasks, all of them start dirtying pages at exactly the same | |
1486 | * time, hence all honoured too large initial task->nr_dirtied_pause. | |
1da177e4 | 1487 | */ |
245b2e70 | 1488 | p = &__get_cpu_var(bdp_ratelimits); |
9d823e8f | 1489 | if (unlikely(current->nr_dirtied >= ratelimit)) |
fa5a734e | 1490 | *p = 0; |
d3bc1fef WF |
1491 | else if (unlikely(*p >= ratelimit_pages)) { |
1492 | *p = 0; | |
1493 | ratelimit = 0; | |
1da177e4 | 1494 | } |
54848d73 WF |
1495 | /* |
1496 | * Pick up the dirtied pages by the exited tasks. This avoids lots of | |
1497 | * short-lived tasks (eg. gcc invocations in a kernel build) escaping | |
1498 | * the dirty throttling and livelock other long-run dirtiers. | |
1499 | */ | |
1500 | p = &__get_cpu_var(dirty_throttle_leaks); | |
1501 | if (*p > 0 && current->nr_dirtied < ratelimit) { | |
d0e1d66b | 1502 | unsigned long nr_pages_dirtied; |
54848d73 WF |
1503 | nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied); |
1504 | *p -= nr_pages_dirtied; | |
1505 | current->nr_dirtied += nr_pages_dirtied; | |
1da177e4 | 1506 | } |
fa5a734e | 1507 | preempt_enable(); |
9d823e8f WF |
1508 | |
1509 | if (unlikely(current->nr_dirtied >= ratelimit)) | |
1510 | balance_dirty_pages(mapping, current->nr_dirtied); | |
1da177e4 | 1511 | } |
d0e1d66b | 1512 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited); |
1da177e4 | 1513 | |
232ea4d6 | 1514 | void throttle_vm_writeout(gfp_t gfp_mask) |
1da177e4 | 1515 | { |
364aeb28 DR |
1516 | unsigned long background_thresh; |
1517 | unsigned long dirty_thresh; | |
1da177e4 LT |
1518 | |
1519 | for ( ; ; ) { | |
16c4042f | 1520 | global_dirty_limits(&background_thresh, &dirty_thresh); |
47a13333 | 1521 | dirty_thresh = hard_dirty_limit(dirty_thresh); |
1da177e4 LT |
1522 | |
1523 | /* | |
1524 | * Boost the allowable dirty threshold a bit for page | |
1525 | * allocators so they don't get DoS'ed by heavy writers | |
1526 | */ | |
1527 | dirty_thresh += dirty_thresh / 10; /* wheeee... */ | |
1528 | ||
c24f21bd CL |
1529 | if (global_page_state(NR_UNSTABLE_NFS) + |
1530 | global_page_state(NR_WRITEBACK) <= dirty_thresh) | |
1531 | break; | |
8aa7e847 | 1532 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
369f2389 FW |
1533 | |
1534 | /* | |
1535 | * The caller might hold locks which can prevent IO completion | |
1536 | * or progress in the filesystem. So we cannot just sit here | |
1537 | * waiting for IO to complete. | |
1538 | */ | |
1539 | if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) | |
1540 | break; | |
1da177e4 LT |
1541 | } |
1542 | } | |
1543 | ||
1da177e4 LT |
1544 | /* |
1545 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
1546 | */ | |
1547 | int dirty_writeback_centisecs_handler(ctl_table *table, int write, | |
8d65af78 | 1548 | void __user *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 1549 | { |
8d65af78 | 1550 | proc_dointvec(table, write, buffer, length, ppos); |
1da177e4 LT |
1551 | return 0; |
1552 | } | |
1553 | ||
c2c4986e | 1554 | #ifdef CONFIG_BLOCK |
31373d09 | 1555 | void laptop_mode_timer_fn(unsigned long data) |
1da177e4 | 1556 | { |
31373d09 MG |
1557 | struct request_queue *q = (struct request_queue *)data; |
1558 | int nr_pages = global_page_state(NR_FILE_DIRTY) + | |
1559 | global_page_state(NR_UNSTABLE_NFS); | |
1da177e4 | 1560 | |
31373d09 MG |
1561 | /* |
1562 | * We want to write everything out, not just down to the dirty | |
1563 | * threshold | |
1564 | */ | |
31373d09 | 1565 | if (bdi_has_dirty_io(&q->backing_dev_info)) |
0e175a18 CW |
1566 | bdi_start_writeback(&q->backing_dev_info, nr_pages, |
1567 | WB_REASON_LAPTOP_TIMER); | |
1da177e4 LT |
1568 | } |
1569 | ||
1570 | /* | |
1571 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
1572 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
1573 | * then push it back - the user is still using the disk. | |
1574 | */ | |
31373d09 | 1575 | void laptop_io_completion(struct backing_dev_info *info) |
1da177e4 | 1576 | { |
31373d09 | 1577 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
1578 | } |
1579 | ||
1580 | /* | |
1581 | * We're in laptop mode and we've just synced. The sync's writes will have | |
1582 | * caused another writeback to be scheduled by laptop_io_completion. | |
1583 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
1584 | */ | |
1585 | void laptop_sync_completion(void) | |
1586 | { | |
31373d09 MG |
1587 | struct backing_dev_info *bdi; |
1588 | ||
1589 | rcu_read_lock(); | |
1590 | ||
1591 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) | |
1592 | del_timer(&bdi->laptop_mode_wb_timer); | |
1593 | ||
1594 | rcu_read_unlock(); | |
1da177e4 | 1595 | } |
c2c4986e | 1596 | #endif |
1da177e4 LT |
1597 | |
1598 | /* | |
1599 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
1600 | * if a large number of processes all perform writes at the same time. | |
1601 | * If it is too low then SMP machines will call the (expensive) | |
1602 | * get_writeback_state too often. | |
1603 | * | |
1604 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
1605 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
9d823e8f | 1606 | * thresholds. |
1da177e4 LT |
1607 | */ |
1608 | ||
2d1d43f6 | 1609 | void writeback_set_ratelimit(void) |
1da177e4 | 1610 | { |
9d823e8f WF |
1611 | unsigned long background_thresh; |
1612 | unsigned long dirty_thresh; | |
1613 | global_dirty_limits(&background_thresh, &dirty_thresh); | |
68809c71 | 1614 | global_dirty_limit = dirty_thresh; |
9d823e8f | 1615 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); |
1da177e4 LT |
1616 | if (ratelimit_pages < 16) |
1617 | ratelimit_pages = 16; | |
1da177e4 LT |
1618 | } |
1619 | ||
26c2143b | 1620 | static int __cpuinit |
2f60d628 SB |
1621 | ratelimit_handler(struct notifier_block *self, unsigned long action, |
1622 | void *hcpu) | |
1da177e4 | 1623 | { |
2f60d628 SB |
1624 | |
1625 | switch (action & ~CPU_TASKS_FROZEN) { | |
1626 | case CPU_ONLINE: | |
1627 | case CPU_DEAD: | |
1628 | writeback_set_ratelimit(); | |
1629 | return NOTIFY_OK; | |
1630 | default: | |
1631 | return NOTIFY_DONE; | |
1632 | } | |
1da177e4 LT |
1633 | } |
1634 | ||
74b85f37 | 1635 | static struct notifier_block __cpuinitdata ratelimit_nb = { |
1da177e4 LT |
1636 | .notifier_call = ratelimit_handler, |
1637 | .next = NULL, | |
1638 | }; | |
1639 | ||
1640 | /* | |
dc6e29da LT |
1641 | * Called early on to tune the page writeback dirty limits. |
1642 | * | |
1643 | * We used to scale dirty pages according to how total memory | |
1644 | * related to pages that could be allocated for buffers (by | |
1645 | * comparing nr_free_buffer_pages() to vm_total_pages. | |
1646 | * | |
1647 | * However, that was when we used "dirty_ratio" to scale with | |
1648 | * all memory, and we don't do that any more. "dirty_ratio" | |
1649 | * is now applied to total non-HIGHPAGE memory (by subtracting | |
1650 | * totalhigh_pages from vm_total_pages), and as such we can't | |
1651 | * get into the old insane situation any more where we had | |
1652 | * large amounts of dirty pages compared to a small amount of | |
1653 | * non-HIGHMEM memory. | |
1654 | * | |
1655 | * But we might still want to scale the dirty_ratio by how | |
1656 | * much memory the box has.. | |
1da177e4 LT |
1657 | */ |
1658 | void __init page_writeback_init(void) | |
1659 | { | |
2d1d43f6 | 1660 | writeback_set_ratelimit(); |
1da177e4 | 1661 | register_cpu_notifier(&ratelimit_nb); |
04fbfdc1 | 1662 | |
eb608e3a | 1663 | fprop_global_init(&writeout_completions); |
1da177e4 LT |
1664 | } |
1665 | ||
f446daae JK |
1666 | /** |
1667 | * tag_pages_for_writeback - tag pages to be written by write_cache_pages | |
1668 | * @mapping: address space structure to write | |
1669 | * @start: starting page index | |
1670 | * @end: ending page index (inclusive) | |
1671 | * | |
1672 | * This function scans the page range from @start to @end (inclusive) and tags | |
1673 | * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is | |
1674 | * that write_cache_pages (or whoever calls this function) will then use | |
1675 | * TOWRITE tag to identify pages eligible for writeback. This mechanism is | |
1676 | * used to avoid livelocking of writeback by a process steadily creating new | |
1677 | * dirty pages in the file (thus it is important for this function to be quick | |
1678 | * so that it can tag pages faster than a dirtying process can create them). | |
1679 | */ | |
1680 | /* | |
1681 | * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency. | |
1682 | */ | |
f446daae JK |
1683 | void tag_pages_for_writeback(struct address_space *mapping, |
1684 | pgoff_t start, pgoff_t end) | |
1685 | { | |
3c111a07 | 1686 | #define WRITEBACK_TAG_BATCH 4096 |
f446daae JK |
1687 | unsigned long tagged; |
1688 | ||
1689 | do { | |
1690 | spin_lock_irq(&mapping->tree_lock); | |
1691 | tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree, | |
1692 | &start, end, WRITEBACK_TAG_BATCH, | |
1693 | PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE); | |
1694 | spin_unlock_irq(&mapping->tree_lock); | |
1695 | WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH); | |
1696 | cond_resched(); | |
d5ed3a4a JK |
1697 | /* We check 'start' to handle wrapping when end == ~0UL */ |
1698 | } while (tagged >= WRITEBACK_TAG_BATCH && start); | |
f446daae JK |
1699 | } |
1700 | EXPORT_SYMBOL(tag_pages_for_writeback); | |
1701 | ||
811d736f | 1702 | /** |
0ea97180 | 1703 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. |
811d736f DH |
1704 | * @mapping: address space structure to write |
1705 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
0ea97180 MS |
1706 | * @writepage: function called for each page |
1707 | * @data: data passed to writepage function | |
811d736f | 1708 | * |
0ea97180 | 1709 | * If a page is already under I/O, write_cache_pages() skips it, even |
811d736f DH |
1710 | * if it's dirty. This is desirable behaviour for memory-cleaning writeback, |
1711 | * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() | |
1712 | * and msync() need to guarantee that all the data which was dirty at the time | |
1713 | * the call was made get new I/O started against them. If wbc->sync_mode is | |
1714 | * WB_SYNC_ALL then we were called for data integrity and we must wait for | |
1715 | * existing IO to complete. | |
f446daae JK |
1716 | * |
1717 | * To avoid livelocks (when other process dirties new pages), we first tag | |
1718 | * pages which should be written back with TOWRITE tag and only then start | |
1719 | * writing them. For data-integrity sync we have to be careful so that we do | |
1720 | * not miss some pages (e.g., because some other process has cleared TOWRITE | |
1721 | * tag we set). The rule we follow is that TOWRITE tag can be cleared only | |
1722 | * by the process clearing the DIRTY tag (and submitting the page for IO). | |
811d736f | 1723 | */ |
0ea97180 MS |
1724 | int write_cache_pages(struct address_space *mapping, |
1725 | struct writeback_control *wbc, writepage_t writepage, | |
1726 | void *data) | |
811d736f | 1727 | { |
811d736f DH |
1728 | int ret = 0; |
1729 | int done = 0; | |
811d736f DH |
1730 | struct pagevec pvec; |
1731 | int nr_pages; | |
31a12666 | 1732 | pgoff_t uninitialized_var(writeback_index); |
811d736f DH |
1733 | pgoff_t index; |
1734 | pgoff_t end; /* Inclusive */ | |
bd19e012 | 1735 | pgoff_t done_index; |
31a12666 | 1736 | int cycled; |
811d736f | 1737 | int range_whole = 0; |
f446daae | 1738 | int tag; |
811d736f | 1739 | |
811d736f DH |
1740 | pagevec_init(&pvec, 0); |
1741 | if (wbc->range_cyclic) { | |
31a12666 NP |
1742 | writeback_index = mapping->writeback_index; /* prev offset */ |
1743 | index = writeback_index; | |
1744 | if (index == 0) | |
1745 | cycled = 1; | |
1746 | else | |
1747 | cycled = 0; | |
811d736f DH |
1748 | end = -1; |
1749 | } else { | |
1750 | index = wbc->range_start >> PAGE_CACHE_SHIFT; | |
1751 | end = wbc->range_end >> PAGE_CACHE_SHIFT; | |
1752 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) | |
1753 | range_whole = 1; | |
31a12666 | 1754 | cycled = 1; /* ignore range_cyclic tests */ |
811d736f | 1755 | } |
6e6938b6 | 1756 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae JK |
1757 | tag = PAGECACHE_TAG_TOWRITE; |
1758 | else | |
1759 | tag = PAGECACHE_TAG_DIRTY; | |
811d736f | 1760 | retry: |
6e6938b6 | 1761 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae | 1762 | tag_pages_for_writeback(mapping, index, end); |
bd19e012 | 1763 | done_index = index; |
5a3d5c98 NP |
1764 | while (!done && (index <= end)) { |
1765 | int i; | |
1766 | ||
f446daae | 1767 | nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag, |
5a3d5c98 NP |
1768 | min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1); |
1769 | if (nr_pages == 0) | |
1770 | break; | |
811d736f | 1771 | |
811d736f DH |
1772 | for (i = 0; i < nr_pages; i++) { |
1773 | struct page *page = pvec.pages[i]; | |
1774 | ||
1775 | /* | |
d5482cdf NP |
1776 | * At this point, the page may be truncated or |
1777 | * invalidated (changing page->mapping to NULL), or | |
1778 | * even swizzled back from swapper_space to tmpfs file | |
1779 | * mapping. However, page->index will not change | |
1780 | * because we have a reference on the page. | |
811d736f | 1781 | */ |
d5482cdf NP |
1782 | if (page->index > end) { |
1783 | /* | |
1784 | * can't be range_cyclic (1st pass) because | |
1785 | * end == -1 in that case. | |
1786 | */ | |
1787 | done = 1; | |
1788 | break; | |
1789 | } | |
1790 | ||
cf15b07c | 1791 | done_index = page->index; |
d5482cdf | 1792 | |
811d736f DH |
1793 | lock_page(page); |
1794 | ||
5a3d5c98 NP |
1795 | /* |
1796 | * Page truncated or invalidated. We can freely skip it | |
1797 | * then, even for data integrity operations: the page | |
1798 | * has disappeared concurrently, so there could be no | |
1799 | * real expectation of this data interity operation | |
1800 | * even if there is now a new, dirty page at the same | |
1801 | * pagecache address. | |
1802 | */ | |
811d736f | 1803 | if (unlikely(page->mapping != mapping)) { |
5a3d5c98 | 1804 | continue_unlock: |
811d736f DH |
1805 | unlock_page(page); |
1806 | continue; | |
1807 | } | |
1808 | ||
515f4a03 NP |
1809 | if (!PageDirty(page)) { |
1810 | /* someone wrote it for us */ | |
1811 | goto continue_unlock; | |
1812 | } | |
1813 | ||
1814 | if (PageWriteback(page)) { | |
1815 | if (wbc->sync_mode != WB_SYNC_NONE) | |
1816 | wait_on_page_writeback(page); | |
1817 | else | |
1818 | goto continue_unlock; | |
1819 | } | |
811d736f | 1820 | |
515f4a03 NP |
1821 | BUG_ON(PageWriteback(page)); |
1822 | if (!clear_page_dirty_for_io(page)) | |
5a3d5c98 | 1823 | goto continue_unlock; |
811d736f | 1824 | |
9e094383 | 1825 | trace_wbc_writepage(wbc, mapping->backing_dev_info); |
0ea97180 | 1826 | ret = (*writepage)(page, wbc, data); |
00266770 NP |
1827 | if (unlikely(ret)) { |
1828 | if (ret == AOP_WRITEPAGE_ACTIVATE) { | |
1829 | unlock_page(page); | |
1830 | ret = 0; | |
1831 | } else { | |
1832 | /* | |
1833 | * done_index is set past this page, | |
1834 | * so media errors will not choke | |
1835 | * background writeout for the entire | |
1836 | * file. This has consequences for | |
1837 | * range_cyclic semantics (ie. it may | |
1838 | * not be suitable for data integrity | |
1839 | * writeout). | |
1840 | */ | |
cf15b07c | 1841 | done_index = page->index + 1; |
00266770 NP |
1842 | done = 1; |
1843 | break; | |
1844 | } | |
0b564927 | 1845 | } |
00266770 | 1846 | |
546a1924 DC |
1847 | /* |
1848 | * We stop writing back only if we are not doing | |
1849 | * integrity sync. In case of integrity sync we have to | |
1850 | * keep going until we have written all the pages | |
1851 | * we tagged for writeback prior to entering this loop. | |
1852 | */ | |
1853 | if (--wbc->nr_to_write <= 0 && | |
1854 | wbc->sync_mode == WB_SYNC_NONE) { | |
1855 | done = 1; | |
1856 | break; | |
05fe478d | 1857 | } |
811d736f DH |
1858 | } |
1859 | pagevec_release(&pvec); | |
1860 | cond_resched(); | |
1861 | } | |
3a4c6800 | 1862 | if (!cycled && !done) { |
811d736f | 1863 | /* |
31a12666 | 1864 | * range_cyclic: |
811d736f DH |
1865 | * We hit the last page and there is more work to be done: wrap |
1866 | * back to the start of the file | |
1867 | */ | |
31a12666 | 1868 | cycled = 1; |
811d736f | 1869 | index = 0; |
31a12666 | 1870 | end = writeback_index - 1; |
811d736f DH |
1871 | goto retry; |
1872 | } | |
0b564927 DC |
1873 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) |
1874 | mapping->writeback_index = done_index; | |
06d6cf69 | 1875 | |
811d736f DH |
1876 | return ret; |
1877 | } | |
0ea97180 MS |
1878 | EXPORT_SYMBOL(write_cache_pages); |
1879 | ||
1880 | /* | |
1881 | * Function used by generic_writepages to call the real writepage | |
1882 | * function and set the mapping flags on error | |
1883 | */ | |
1884 | static int __writepage(struct page *page, struct writeback_control *wbc, | |
1885 | void *data) | |
1886 | { | |
1887 | struct address_space *mapping = data; | |
1888 | int ret = mapping->a_ops->writepage(page, wbc); | |
1889 | mapping_set_error(mapping, ret); | |
1890 | return ret; | |
1891 | } | |
1892 | ||
1893 | /** | |
1894 | * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. | |
1895 | * @mapping: address space structure to write | |
1896 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
1897 | * | |
1898 | * This is a library function, which implements the writepages() | |
1899 | * address_space_operation. | |
1900 | */ | |
1901 | int generic_writepages(struct address_space *mapping, | |
1902 | struct writeback_control *wbc) | |
1903 | { | |
9b6096a6 SL |
1904 | struct blk_plug plug; |
1905 | int ret; | |
1906 | ||
0ea97180 MS |
1907 | /* deal with chardevs and other special file */ |
1908 | if (!mapping->a_ops->writepage) | |
1909 | return 0; | |
1910 | ||
9b6096a6 SL |
1911 | blk_start_plug(&plug); |
1912 | ret = write_cache_pages(mapping, wbc, __writepage, mapping); | |
1913 | blk_finish_plug(&plug); | |
1914 | return ret; | |
0ea97180 | 1915 | } |
811d736f DH |
1916 | |
1917 | EXPORT_SYMBOL(generic_writepages); | |
1918 | ||
1da177e4 LT |
1919 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
1920 | { | |
22905f77 AM |
1921 | int ret; |
1922 | ||
1da177e4 LT |
1923 | if (wbc->nr_to_write <= 0) |
1924 | return 0; | |
1925 | if (mapping->a_ops->writepages) | |
d08b3851 | 1926 | ret = mapping->a_ops->writepages(mapping, wbc); |
22905f77 AM |
1927 | else |
1928 | ret = generic_writepages(mapping, wbc); | |
22905f77 | 1929 | return ret; |
1da177e4 LT |
1930 | } |
1931 | ||
1932 | /** | |
1933 | * write_one_page - write out a single page and optionally wait on I/O | |
67be2dd1 MW |
1934 | * @page: the page to write |
1935 | * @wait: if true, wait on writeout | |
1da177e4 LT |
1936 | * |
1937 | * The page must be locked by the caller and will be unlocked upon return. | |
1938 | * | |
1939 | * write_one_page() returns a negative error code if I/O failed. | |
1940 | */ | |
1941 | int write_one_page(struct page *page, int wait) | |
1942 | { | |
1943 | struct address_space *mapping = page->mapping; | |
1944 | int ret = 0; | |
1945 | struct writeback_control wbc = { | |
1946 | .sync_mode = WB_SYNC_ALL, | |
1947 | .nr_to_write = 1, | |
1948 | }; | |
1949 | ||
1950 | BUG_ON(!PageLocked(page)); | |
1951 | ||
1952 | if (wait) | |
1953 | wait_on_page_writeback(page); | |
1954 | ||
1955 | if (clear_page_dirty_for_io(page)) { | |
1956 | page_cache_get(page); | |
1957 | ret = mapping->a_ops->writepage(page, &wbc); | |
1958 | if (ret == 0 && wait) { | |
1959 | wait_on_page_writeback(page); | |
1960 | if (PageError(page)) | |
1961 | ret = -EIO; | |
1962 | } | |
1963 | page_cache_release(page); | |
1964 | } else { | |
1965 | unlock_page(page); | |
1966 | } | |
1967 | return ret; | |
1968 | } | |
1969 | EXPORT_SYMBOL(write_one_page); | |
1970 | ||
76719325 KC |
1971 | /* |
1972 | * For address_spaces which do not use buffers nor write back. | |
1973 | */ | |
1974 | int __set_page_dirty_no_writeback(struct page *page) | |
1975 | { | |
1976 | if (!PageDirty(page)) | |
c3f0da63 | 1977 | return !TestSetPageDirty(page); |
76719325 KC |
1978 | return 0; |
1979 | } | |
1980 | ||
e3a7cca1 ES |
1981 | /* |
1982 | * Helper function for set_page_dirty family. | |
1983 | * NOTE: This relies on being atomic wrt interrupts. | |
1984 | */ | |
1985 | void account_page_dirtied(struct page *page, struct address_space *mapping) | |
1986 | { | |
9fb0a7da TH |
1987 | trace_writeback_dirty_page(page, mapping); |
1988 | ||
e3a7cca1 ES |
1989 | if (mapping_cap_account_dirty(mapping)) { |
1990 | __inc_zone_page_state(page, NR_FILE_DIRTY); | |
ea941f0e | 1991 | __inc_zone_page_state(page, NR_DIRTIED); |
e3a7cca1 | 1992 | __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE); |
c8e28ce0 | 1993 | __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED); |
e3a7cca1 | 1994 | task_io_account_write(PAGE_CACHE_SIZE); |
d3bc1fef WF |
1995 | current->nr_dirtied++; |
1996 | this_cpu_inc(bdp_ratelimits); | |
e3a7cca1 ES |
1997 | } |
1998 | } | |
679ceace | 1999 | EXPORT_SYMBOL(account_page_dirtied); |
e3a7cca1 | 2000 | |
f629d1c9 MR |
2001 | /* |
2002 | * Helper function for set_page_writeback family. | |
2003 | * NOTE: Unlike account_page_dirtied this does not rely on being atomic | |
2004 | * wrt interrupts. | |
2005 | */ | |
2006 | void account_page_writeback(struct page *page) | |
2007 | { | |
2008 | inc_zone_page_state(page, NR_WRITEBACK); | |
2009 | } | |
2010 | EXPORT_SYMBOL(account_page_writeback); | |
2011 | ||
1da177e4 LT |
2012 | /* |
2013 | * For address_spaces which do not use buffers. Just tag the page as dirty in | |
2014 | * its radix tree. | |
2015 | * | |
2016 | * This is also used when a single buffer is being dirtied: we want to set the | |
2017 | * page dirty in that case, but not all the buffers. This is a "bottom-up" | |
2018 | * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying. | |
2019 | * | |
2020 | * Most callers have locked the page, which pins the address_space in memory. | |
2021 | * But zap_pte_range() does not lock the page, however in that case the | |
2022 | * mapping is pinned by the vma's ->vm_file reference. | |
2023 | * | |
2024 | * We take care to handle the case where the page was truncated from the | |
183ff22b | 2025 | * mapping by re-checking page_mapping() inside tree_lock. |
1da177e4 LT |
2026 | */ |
2027 | int __set_page_dirty_nobuffers(struct page *page) | |
2028 | { | |
1da177e4 LT |
2029 | if (!TestSetPageDirty(page)) { |
2030 | struct address_space *mapping = page_mapping(page); | |
2031 | struct address_space *mapping2; | |
dce0b4fc | 2032 | unsigned long flags; |
1da177e4 | 2033 | |
8c08540f AM |
2034 | if (!mapping) |
2035 | return 1; | |
2036 | ||
dce0b4fc | 2037 | spin_lock_irqsave(&mapping->tree_lock, flags); |
8c08540f AM |
2038 | mapping2 = page_mapping(page); |
2039 | if (mapping2) { /* Race with truncate? */ | |
2040 | BUG_ON(mapping2 != mapping); | |
787d2214 | 2041 | WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page)); |
e3a7cca1 | 2042 | account_page_dirtied(page, mapping); |
8c08540f AM |
2043 | radix_tree_tag_set(&mapping->page_tree, |
2044 | page_index(page), PAGECACHE_TAG_DIRTY); | |
2045 | } | |
dce0b4fc | 2046 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
8c08540f AM |
2047 | if (mapping->host) { |
2048 | /* !PageAnon && !swapper_space */ | |
2049 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 2050 | } |
4741c9fd | 2051 | return 1; |
1da177e4 | 2052 | } |
4741c9fd | 2053 | return 0; |
1da177e4 LT |
2054 | } |
2055 | EXPORT_SYMBOL(__set_page_dirty_nobuffers); | |
2056 | ||
2f800fbd WF |
2057 | /* |
2058 | * Call this whenever redirtying a page, to de-account the dirty counters | |
2059 | * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written | |
2060 | * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to | |
2061 | * systematic errors in balanced_dirty_ratelimit and the dirty pages position | |
2062 | * control. | |
2063 | */ | |
2064 | void account_page_redirty(struct page *page) | |
2065 | { | |
2066 | struct address_space *mapping = page->mapping; | |
2067 | if (mapping && mapping_cap_account_dirty(mapping)) { | |
2068 | current->nr_dirtied--; | |
2069 | dec_zone_page_state(page, NR_DIRTIED); | |
2070 | dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED); | |
2071 | } | |
2072 | } | |
2073 | EXPORT_SYMBOL(account_page_redirty); | |
2074 | ||
1da177e4 LT |
2075 | /* |
2076 | * When a writepage implementation decides that it doesn't want to write this | |
2077 | * page for some reason, it should redirty the locked page via | |
2078 | * redirty_page_for_writepage() and it should then unlock the page and return 0 | |
2079 | */ | |
2080 | int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page) | |
2081 | { | |
2082 | wbc->pages_skipped++; | |
2f800fbd | 2083 | account_page_redirty(page); |
1da177e4 LT |
2084 | return __set_page_dirty_nobuffers(page); |
2085 | } | |
2086 | EXPORT_SYMBOL(redirty_page_for_writepage); | |
2087 | ||
2088 | /* | |
6746aff7 WF |
2089 | * Dirty a page. |
2090 | * | |
2091 | * For pages with a mapping this should be done under the page lock | |
2092 | * for the benefit of asynchronous memory errors who prefer a consistent | |
2093 | * dirty state. This rule can be broken in some special cases, | |
2094 | * but should be better not to. | |
2095 | * | |
1da177e4 LT |
2096 | * If the mapping doesn't provide a set_page_dirty a_op, then |
2097 | * just fall through and assume that it wants buffer_heads. | |
2098 | */ | |
1cf6e7d8 | 2099 | int set_page_dirty(struct page *page) |
1da177e4 LT |
2100 | { |
2101 | struct address_space *mapping = page_mapping(page); | |
2102 | ||
2103 | if (likely(mapping)) { | |
2104 | int (*spd)(struct page *) = mapping->a_ops->set_page_dirty; | |
278df9f4 MK |
2105 | /* |
2106 | * readahead/lru_deactivate_page could remain | |
2107 | * PG_readahead/PG_reclaim due to race with end_page_writeback | |
2108 | * About readahead, if the page is written, the flags would be | |
2109 | * reset. So no problem. | |
2110 | * About lru_deactivate_page, if the page is redirty, the flag | |
2111 | * will be reset. So no problem. but if the page is used by readahead | |
2112 | * it will confuse readahead and make it restart the size rampup | |
2113 | * process. But it's a trivial problem. | |
2114 | */ | |
2115 | ClearPageReclaim(page); | |
9361401e DH |
2116 | #ifdef CONFIG_BLOCK |
2117 | if (!spd) | |
2118 | spd = __set_page_dirty_buffers; | |
2119 | #endif | |
2120 | return (*spd)(page); | |
1da177e4 | 2121 | } |
4741c9fd AM |
2122 | if (!PageDirty(page)) { |
2123 | if (!TestSetPageDirty(page)) | |
2124 | return 1; | |
2125 | } | |
1da177e4 LT |
2126 | return 0; |
2127 | } | |
2128 | EXPORT_SYMBOL(set_page_dirty); | |
2129 | ||
2130 | /* | |
2131 | * set_page_dirty() is racy if the caller has no reference against | |
2132 | * page->mapping->host, and if the page is unlocked. This is because another | |
2133 | * CPU could truncate the page off the mapping and then free the mapping. | |
2134 | * | |
2135 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
2136 | * holds a reference on the inode by having an open file. | |
2137 | * | |
2138 | * In other cases, the page should be locked before running set_page_dirty(). | |
2139 | */ | |
2140 | int set_page_dirty_lock(struct page *page) | |
2141 | { | |
2142 | int ret; | |
2143 | ||
7eaceacc | 2144 | lock_page(page); |
1da177e4 LT |
2145 | ret = set_page_dirty(page); |
2146 | unlock_page(page); | |
2147 | return ret; | |
2148 | } | |
2149 | EXPORT_SYMBOL(set_page_dirty_lock); | |
2150 | ||
1da177e4 LT |
2151 | /* |
2152 | * Clear a page's dirty flag, while caring for dirty memory accounting. | |
2153 | * Returns true if the page was previously dirty. | |
2154 | * | |
2155 | * This is for preparing to put the page under writeout. We leave the page | |
2156 | * tagged as dirty in the radix tree so that a concurrent write-for-sync | |
2157 | * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage | |
2158 | * implementation will run either set_page_writeback() or set_page_dirty(), | |
2159 | * at which stage we bring the page's dirty flag and radix-tree dirty tag | |
2160 | * back into sync. | |
2161 | * | |
2162 | * This incoherency between the page's dirty flag and radix-tree tag is | |
2163 | * unfortunate, but it only exists while the page is locked. | |
2164 | */ | |
2165 | int clear_page_dirty_for_io(struct page *page) | |
2166 | { | |
2167 | struct address_space *mapping = page_mapping(page); | |
2168 | ||
79352894 NP |
2169 | BUG_ON(!PageLocked(page)); |
2170 | ||
7658cc28 LT |
2171 | if (mapping && mapping_cap_account_dirty(mapping)) { |
2172 | /* | |
2173 | * Yes, Virginia, this is indeed insane. | |
2174 | * | |
2175 | * We use this sequence to make sure that | |
2176 | * (a) we account for dirty stats properly | |
2177 | * (b) we tell the low-level filesystem to | |
2178 | * mark the whole page dirty if it was | |
2179 | * dirty in a pagetable. Only to then | |
2180 | * (c) clean the page again and return 1 to | |
2181 | * cause the writeback. | |
2182 | * | |
2183 | * This way we avoid all nasty races with the | |
2184 | * dirty bit in multiple places and clearing | |
2185 | * them concurrently from different threads. | |
2186 | * | |
2187 | * Note! Normally the "set_page_dirty(page)" | |
2188 | * has no effect on the actual dirty bit - since | |
2189 | * that will already usually be set. But we | |
2190 | * need the side effects, and it can help us | |
2191 | * avoid races. | |
2192 | * | |
2193 | * We basically use the page "master dirty bit" | |
2194 | * as a serialization point for all the different | |
2195 | * threads doing their things. | |
7658cc28 LT |
2196 | */ |
2197 | if (page_mkclean(page)) | |
2198 | set_page_dirty(page); | |
79352894 NP |
2199 | /* |
2200 | * We carefully synchronise fault handlers against | |
2201 | * installing a dirty pte and marking the page dirty | |
2202 | * at this point. We do this by having them hold the | |
2203 | * page lock at some point after installing their | |
2204 | * pte, but before marking the page dirty. | |
2205 | * Pages are always locked coming in here, so we get | |
2206 | * the desired exclusion. See mm/memory.c:do_wp_page() | |
2207 | * for more comments. | |
2208 | */ | |
7658cc28 | 2209 | if (TestClearPageDirty(page)) { |
8c08540f | 2210 | dec_zone_page_state(page, NR_FILE_DIRTY); |
c9e51e41 PZ |
2211 | dec_bdi_stat(mapping->backing_dev_info, |
2212 | BDI_RECLAIMABLE); | |
7658cc28 | 2213 | return 1; |
1da177e4 | 2214 | } |
7658cc28 | 2215 | return 0; |
1da177e4 | 2216 | } |
7658cc28 | 2217 | return TestClearPageDirty(page); |
1da177e4 | 2218 | } |
58bb01a9 | 2219 | EXPORT_SYMBOL(clear_page_dirty_for_io); |
1da177e4 LT |
2220 | |
2221 | int test_clear_page_writeback(struct page *page) | |
2222 | { | |
2223 | struct address_space *mapping = page_mapping(page); | |
2224 | int ret; | |
2225 | ||
2226 | if (mapping) { | |
69cb51d1 | 2227 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
1da177e4 LT |
2228 | unsigned long flags; |
2229 | ||
19fd6231 | 2230 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 2231 | ret = TestClearPageWriteback(page); |
69cb51d1 | 2232 | if (ret) { |
1da177e4 LT |
2233 | radix_tree_tag_clear(&mapping->page_tree, |
2234 | page_index(page), | |
2235 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 2236 | if (bdi_cap_account_writeback(bdi)) { |
69cb51d1 | 2237 | __dec_bdi_stat(bdi, BDI_WRITEBACK); |
04fbfdc1 PZ |
2238 | __bdi_writeout_inc(bdi); |
2239 | } | |
69cb51d1 | 2240 | } |
19fd6231 | 2241 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
2242 | } else { |
2243 | ret = TestClearPageWriteback(page); | |
2244 | } | |
99b12e3d | 2245 | if (ret) { |
d688abf5 | 2246 | dec_zone_page_state(page, NR_WRITEBACK); |
99b12e3d WF |
2247 | inc_zone_page_state(page, NR_WRITTEN); |
2248 | } | |
1da177e4 LT |
2249 | return ret; |
2250 | } | |
2251 | ||
2252 | int test_set_page_writeback(struct page *page) | |
2253 | { | |
2254 | struct address_space *mapping = page_mapping(page); | |
2255 | int ret; | |
2256 | ||
2257 | if (mapping) { | |
69cb51d1 | 2258 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
1da177e4 LT |
2259 | unsigned long flags; |
2260 | ||
19fd6231 | 2261 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 2262 | ret = TestSetPageWriteback(page); |
69cb51d1 | 2263 | if (!ret) { |
1da177e4 LT |
2264 | radix_tree_tag_set(&mapping->page_tree, |
2265 | page_index(page), | |
2266 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 2267 | if (bdi_cap_account_writeback(bdi)) |
69cb51d1 PZ |
2268 | __inc_bdi_stat(bdi, BDI_WRITEBACK); |
2269 | } | |
1da177e4 LT |
2270 | if (!PageDirty(page)) |
2271 | radix_tree_tag_clear(&mapping->page_tree, | |
2272 | page_index(page), | |
2273 | PAGECACHE_TAG_DIRTY); | |
f446daae JK |
2274 | radix_tree_tag_clear(&mapping->page_tree, |
2275 | page_index(page), | |
2276 | PAGECACHE_TAG_TOWRITE); | |
19fd6231 | 2277 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
2278 | } else { |
2279 | ret = TestSetPageWriteback(page); | |
2280 | } | |
d688abf5 | 2281 | if (!ret) |
f629d1c9 | 2282 | account_page_writeback(page); |
1da177e4 LT |
2283 | return ret; |
2284 | ||
2285 | } | |
2286 | EXPORT_SYMBOL(test_set_page_writeback); | |
2287 | ||
2288 | /* | |
00128188 | 2289 | * Return true if any of the pages in the mapping are marked with the |
1da177e4 LT |
2290 | * passed tag. |
2291 | */ | |
2292 | int mapping_tagged(struct address_space *mapping, int tag) | |
2293 | { | |
72c47832 | 2294 | return radix_tree_tagged(&mapping->page_tree, tag); |
1da177e4 LT |
2295 | } |
2296 | EXPORT_SYMBOL(mapping_tagged); | |
1d1d1a76 DW |
2297 | |
2298 | /** | |
2299 | * wait_for_stable_page() - wait for writeback to finish, if necessary. | |
2300 | * @page: The page to wait on. | |
2301 | * | |
2302 | * This function determines if the given page is related to a backing device | |
2303 | * that requires page contents to be held stable during writeback. If so, then | |
2304 | * it will wait for any pending writeback to complete. | |
2305 | */ | |
2306 | void wait_for_stable_page(struct page *page) | |
2307 | { | |
2308 | struct address_space *mapping = page_mapping(page); | |
2309 | struct backing_dev_info *bdi = mapping->backing_dev_info; | |
2310 | ||
2311 | if (!bdi_cap_stable_pages_required(bdi)) | |
2312 | return; | |
2313 | ||
2314 | wait_on_page_writeback(page); | |
2315 | } | |
2316 | EXPORT_SYMBOL_GPL(wait_for_stable_page); |