Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
f30c2269 | 2 | * mm/page-writeback.c |
1da177e4 LT |
3 | * |
4 | * Copyright (C) 2002, Linus Torvalds. | |
04fbfdc1 | 5 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> |
1da177e4 LT |
6 | * |
7 | * Contains functions related to writing back dirty pages at the | |
8 | * address_space level. | |
9 | * | |
e1f8e874 | 10 | * 10Apr2002 Andrew Morton |
1da177e4 LT |
11 | * Initial version |
12 | */ | |
13 | ||
14 | #include <linux/kernel.h> | |
15 | #include <linux/module.h> | |
16 | #include <linux/spinlock.h> | |
17 | #include <linux/fs.h> | |
18 | #include <linux/mm.h> | |
19 | #include <linux/swap.h> | |
20 | #include <linux/slab.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/writeback.h> | |
23 | #include <linux/init.h> | |
24 | #include <linux/backing-dev.h> | |
55e829af | 25 | #include <linux/task_io_accounting_ops.h> |
1da177e4 LT |
26 | #include <linux/blkdev.h> |
27 | #include <linux/mpage.h> | |
d08b3851 | 28 | #include <linux/rmap.h> |
1da177e4 LT |
29 | #include <linux/percpu.h> |
30 | #include <linux/notifier.h> | |
31 | #include <linux/smp.h> | |
32 | #include <linux/sysctl.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/syscalls.h> | |
cf9a2ae8 | 35 | #include <linux/buffer_head.h> |
811d736f | 36 | #include <linux/pagevec.h> |
028c2dd1 | 37 | #include <trace/events/writeback.h> |
1da177e4 | 38 | |
ffd1f609 WF |
39 | /* |
40 | * Sleep at most 200ms at a time in balance_dirty_pages(). | |
41 | */ | |
42 | #define MAX_PAUSE max(HZ/5, 1) | |
43 | ||
e98be2d5 WF |
44 | /* |
45 | * Estimate write bandwidth at 200ms intervals. | |
46 | */ | |
47 | #define BANDWIDTH_INTERVAL max(HZ/5, 1) | |
48 | ||
6c14ae1e WF |
49 | #define RATELIMIT_CALC_SHIFT 10 |
50 | ||
1da177e4 LT |
51 | /* |
52 | * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited | |
53 | * will look to see if it needs to force writeback or throttling. | |
54 | */ | |
55 | static long ratelimit_pages = 32; | |
56 | ||
1da177e4 LT |
57 | /* The following parameters are exported via /proc/sys/vm */ |
58 | ||
59 | /* | |
5b0830cb | 60 | * Start background writeback (via writeback threads) at this percentage |
1da177e4 | 61 | */ |
1b5e62b4 | 62 | int dirty_background_ratio = 10; |
1da177e4 | 63 | |
2da02997 DR |
64 | /* |
65 | * dirty_background_bytes starts at 0 (disabled) so that it is a function of | |
66 | * dirty_background_ratio * the amount of dirtyable memory | |
67 | */ | |
68 | unsigned long dirty_background_bytes; | |
69 | ||
195cf453 BG |
70 | /* |
71 | * free highmem will not be subtracted from the total free memory | |
72 | * for calculating free ratios if vm_highmem_is_dirtyable is true | |
73 | */ | |
74 | int vm_highmem_is_dirtyable; | |
75 | ||
1da177e4 LT |
76 | /* |
77 | * The generator of dirty data starts writeback at this percentage | |
78 | */ | |
1b5e62b4 | 79 | int vm_dirty_ratio = 20; |
1da177e4 | 80 | |
2da02997 DR |
81 | /* |
82 | * vm_dirty_bytes starts at 0 (disabled) so that it is a function of | |
83 | * vm_dirty_ratio * the amount of dirtyable memory | |
84 | */ | |
85 | unsigned long vm_dirty_bytes; | |
86 | ||
1da177e4 | 87 | /* |
704503d8 | 88 | * The interval between `kupdate'-style writebacks |
1da177e4 | 89 | */ |
22ef37ee | 90 | unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */ |
1da177e4 LT |
91 | |
92 | /* | |
704503d8 | 93 | * The longest time for which data is allowed to remain dirty |
1da177e4 | 94 | */ |
22ef37ee | 95 | unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */ |
1da177e4 LT |
96 | |
97 | /* | |
98 | * Flag that makes the machine dump writes/reads and block dirtyings. | |
99 | */ | |
100 | int block_dump; | |
101 | ||
102 | /* | |
ed5b43f1 BS |
103 | * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: |
104 | * a full sync is triggered after this time elapses without any disk activity. | |
1da177e4 LT |
105 | */ |
106 | int laptop_mode; | |
107 | ||
108 | EXPORT_SYMBOL(laptop_mode); | |
109 | ||
110 | /* End of sysctl-exported parameters */ | |
111 | ||
c42843f2 | 112 | unsigned long global_dirty_limit; |
1da177e4 | 113 | |
04fbfdc1 PZ |
114 | /* |
115 | * Scale the writeback cache size proportional to the relative writeout speeds. | |
116 | * | |
117 | * We do this by keeping a floating proportion between BDIs, based on page | |
118 | * writeback completions [end_page_writeback()]. Those devices that write out | |
119 | * pages fastest will get the larger share, while the slower will get a smaller | |
120 | * share. | |
121 | * | |
122 | * We use page writeout completions because we are interested in getting rid of | |
123 | * dirty pages. Having them written out is the primary goal. | |
124 | * | |
125 | * We introduce a concept of time, a period over which we measure these events, | |
126 | * because demand can/will vary over time. The length of this period itself is | |
127 | * measured in page writeback completions. | |
128 | * | |
129 | */ | |
130 | static struct prop_descriptor vm_completions; | |
3e26c149 | 131 | static struct prop_descriptor vm_dirties; |
04fbfdc1 | 132 | |
04fbfdc1 PZ |
133 | /* |
134 | * couple the period to the dirty_ratio: | |
135 | * | |
136 | * period/2 ~ roundup_pow_of_two(dirty limit) | |
137 | */ | |
138 | static int calc_period_shift(void) | |
139 | { | |
140 | unsigned long dirty_total; | |
141 | ||
2da02997 DR |
142 | if (vm_dirty_bytes) |
143 | dirty_total = vm_dirty_bytes / PAGE_SIZE; | |
144 | else | |
145 | dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / | |
146 | 100; | |
04fbfdc1 PZ |
147 | return 2 + ilog2(dirty_total - 1); |
148 | } | |
149 | ||
150 | /* | |
2da02997 | 151 | * update the period when the dirty threshold changes. |
04fbfdc1 | 152 | */ |
2da02997 DR |
153 | static void update_completion_period(void) |
154 | { | |
155 | int shift = calc_period_shift(); | |
156 | prop_change_shift(&vm_completions, shift); | |
157 | prop_change_shift(&vm_dirties, shift); | |
9d823e8f WF |
158 | |
159 | writeback_set_ratelimit(); | |
2da02997 DR |
160 | } |
161 | ||
162 | int dirty_background_ratio_handler(struct ctl_table *table, int write, | |
8d65af78 | 163 | void __user *buffer, size_t *lenp, |
2da02997 DR |
164 | loff_t *ppos) |
165 | { | |
166 | int ret; | |
167 | ||
8d65af78 | 168 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
169 | if (ret == 0 && write) |
170 | dirty_background_bytes = 0; | |
171 | return ret; | |
172 | } | |
173 | ||
174 | int dirty_background_bytes_handler(struct ctl_table *table, int write, | |
8d65af78 | 175 | void __user *buffer, size_t *lenp, |
2da02997 DR |
176 | loff_t *ppos) |
177 | { | |
178 | int ret; | |
179 | ||
8d65af78 | 180 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
181 | if (ret == 0 && write) |
182 | dirty_background_ratio = 0; | |
183 | return ret; | |
184 | } | |
185 | ||
04fbfdc1 | 186 | int dirty_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 187 | void __user *buffer, size_t *lenp, |
04fbfdc1 PZ |
188 | loff_t *ppos) |
189 | { | |
190 | int old_ratio = vm_dirty_ratio; | |
2da02997 DR |
191 | int ret; |
192 | ||
8d65af78 | 193 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
04fbfdc1 | 194 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
2da02997 DR |
195 | update_completion_period(); |
196 | vm_dirty_bytes = 0; | |
197 | } | |
198 | return ret; | |
199 | } | |
200 | ||
201 | ||
202 | int dirty_bytes_handler(struct ctl_table *table, int write, | |
8d65af78 | 203 | void __user *buffer, size_t *lenp, |
2da02997 DR |
204 | loff_t *ppos) |
205 | { | |
fc3501d4 | 206 | unsigned long old_bytes = vm_dirty_bytes; |
2da02997 DR |
207 | int ret; |
208 | ||
8d65af78 | 209 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
210 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
211 | update_completion_period(); | |
212 | vm_dirty_ratio = 0; | |
04fbfdc1 PZ |
213 | } |
214 | return ret; | |
215 | } | |
216 | ||
217 | /* | |
218 | * Increment the BDI's writeout completion count and the global writeout | |
219 | * completion count. Called from test_clear_page_writeback(). | |
220 | */ | |
221 | static inline void __bdi_writeout_inc(struct backing_dev_info *bdi) | |
222 | { | |
f7d2b1ec | 223 | __inc_bdi_stat(bdi, BDI_WRITTEN); |
a42dde04 PZ |
224 | __prop_inc_percpu_max(&vm_completions, &bdi->completions, |
225 | bdi->max_prop_frac); | |
04fbfdc1 PZ |
226 | } |
227 | ||
dd5656e5 MS |
228 | void bdi_writeout_inc(struct backing_dev_info *bdi) |
229 | { | |
230 | unsigned long flags; | |
231 | ||
232 | local_irq_save(flags); | |
233 | __bdi_writeout_inc(bdi); | |
234 | local_irq_restore(flags); | |
235 | } | |
236 | EXPORT_SYMBOL_GPL(bdi_writeout_inc); | |
237 | ||
1cf6e7d8 | 238 | void task_dirty_inc(struct task_struct *tsk) |
3e26c149 PZ |
239 | { |
240 | prop_inc_single(&vm_dirties, &tsk->dirties); | |
241 | } | |
242 | ||
04fbfdc1 PZ |
243 | /* |
244 | * Obtain an accurate fraction of the BDI's portion. | |
245 | */ | |
246 | static void bdi_writeout_fraction(struct backing_dev_info *bdi, | |
247 | long *numerator, long *denominator) | |
248 | { | |
3efaf0fa | 249 | prop_fraction_percpu(&vm_completions, &bdi->completions, |
04fbfdc1 | 250 | numerator, denominator); |
04fbfdc1 PZ |
251 | } |
252 | ||
189d3c4a PZ |
253 | /* |
254 | * | |
255 | */ | |
189d3c4a PZ |
256 | static unsigned int bdi_min_ratio; |
257 | ||
258 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) | |
259 | { | |
260 | int ret = 0; | |
189d3c4a | 261 | |
cfc4ba53 | 262 | spin_lock_bh(&bdi_lock); |
a42dde04 | 263 | if (min_ratio > bdi->max_ratio) { |
189d3c4a | 264 | ret = -EINVAL; |
a42dde04 PZ |
265 | } else { |
266 | min_ratio -= bdi->min_ratio; | |
267 | if (bdi_min_ratio + min_ratio < 100) { | |
268 | bdi_min_ratio += min_ratio; | |
269 | bdi->min_ratio += min_ratio; | |
270 | } else { | |
271 | ret = -EINVAL; | |
272 | } | |
273 | } | |
cfc4ba53 | 274 | spin_unlock_bh(&bdi_lock); |
a42dde04 PZ |
275 | |
276 | return ret; | |
277 | } | |
278 | ||
279 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio) | |
280 | { | |
a42dde04 PZ |
281 | int ret = 0; |
282 | ||
283 | if (max_ratio > 100) | |
284 | return -EINVAL; | |
285 | ||
cfc4ba53 | 286 | spin_lock_bh(&bdi_lock); |
a42dde04 PZ |
287 | if (bdi->min_ratio > max_ratio) { |
288 | ret = -EINVAL; | |
289 | } else { | |
290 | bdi->max_ratio = max_ratio; | |
291 | bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100; | |
292 | } | |
cfc4ba53 | 293 | spin_unlock_bh(&bdi_lock); |
189d3c4a PZ |
294 | |
295 | return ret; | |
296 | } | |
a42dde04 | 297 | EXPORT_SYMBOL(bdi_set_max_ratio); |
189d3c4a | 298 | |
1da177e4 LT |
299 | /* |
300 | * Work out the current dirty-memory clamping and background writeout | |
301 | * thresholds. | |
302 | * | |
303 | * The main aim here is to lower them aggressively if there is a lot of mapped | |
304 | * memory around. To avoid stressing page reclaim with lots of unreclaimable | |
305 | * pages. It is better to clamp down on writers than to start swapping, and | |
306 | * performing lots of scanning. | |
307 | * | |
308 | * We only allow 1/2 of the currently-unmapped memory to be dirtied. | |
309 | * | |
310 | * We don't permit the clamping level to fall below 5% - that is getting rather | |
311 | * excessive. | |
312 | * | |
313 | * We make sure that the background writeout level is below the adjusted | |
314 | * clamping level. | |
315 | */ | |
1b424464 CL |
316 | |
317 | static unsigned long highmem_dirtyable_memory(unsigned long total) | |
318 | { | |
319 | #ifdef CONFIG_HIGHMEM | |
320 | int node; | |
321 | unsigned long x = 0; | |
322 | ||
37b07e41 | 323 | for_each_node_state(node, N_HIGH_MEMORY) { |
1b424464 CL |
324 | struct zone *z = |
325 | &NODE_DATA(node)->node_zones[ZONE_HIGHMEM]; | |
326 | ||
adea02a1 WF |
327 | x += zone_page_state(z, NR_FREE_PAGES) + |
328 | zone_reclaimable_pages(z); | |
1b424464 CL |
329 | } |
330 | /* | |
331 | * Make sure that the number of highmem pages is never larger | |
332 | * than the number of the total dirtyable memory. This can only | |
333 | * occur in very strange VM situations but we want to make sure | |
334 | * that this does not occur. | |
335 | */ | |
336 | return min(x, total); | |
337 | #else | |
338 | return 0; | |
339 | #endif | |
340 | } | |
341 | ||
3eefae99 SR |
342 | /** |
343 | * determine_dirtyable_memory - amount of memory that may be used | |
344 | * | |
345 | * Returns the numebr of pages that can currently be freed and used | |
346 | * by the kernel for direct mappings. | |
347 | */ | |
348 | unsigned long determine_dirtyable_memory(void) | |
1b424464 CL |
349 | { |
350 | unsigned long x; | |
351 | ||
adea02a1 | 352 | x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages(); |
195cf453 BG |
353 | |
354 | if (!vm_highmem_is_dirtyable) | |
355 | x -= highmem_dirtyable_memory(x); | |
356 | ||
1b424464 CL |
357 | return x + 1; /* Ensure that we never return 0 */ |
358 | } | |
359 | ||
6c14ae1e WF |
360 | static unsigned long dirty_freerun_ceiling(unsigned long thresh, |
361 | unsigned long bg_thresh) | |
362 | { | |
363 | return (thresh + bg_thresh) / 2; | |
364 | } | |
365 | ||
ffd1f609 WF |
366 | static unsigned long hard_dirty_limit(unsigned long thresh) |
367 | { | |
368 | return max(thresh, global_dirty_limit); | |
369 | } | |
370 | ||
03ab450f | 371 | /* |
1babe183 WF |
372 | * global_dirty_limits - background-writeback and dirty-throttling thresholds |
373 | * | |
374 | * Calculate the dirty thresholds based on sysctl parameters | |
375 | * - vm.dirty_background_ratio or vm.dirty_background_bytes | |
376 | * - vm.dirty_ratio or vm.dirty_bytes | |
377 | * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and | |
ebd1373d | 378 | * real-time tasks. |
1babe183 | 379 | */ |
16c4042f | 380 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) |
1da177e4 | 381 | { |
364aeb28 DR |
382 | unsigned long background; |
383 | unsigned long dirty; | |
240c879f | 384 | unsigned long uninitialized_var(available_memory); |
1da177e4 LT |
385 | struct task_struct *tsk; |
386 | ||
240c879f MK |
387 | if (!vm_dirty_bytes || !dirty_background_bytes) |
388 | available_memory = determine_dirtyable_memory(); | |
389 | ||
2da02997 DR |
390 | if (vm_dirty_bytes) |
391 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE); | |
4cbec4c8 WF |
392 | else |
393 | dirty = (vm_dirty_ratio * available_memory) / 100; | |
1da177e4 | 394 | |
2da02997 DR |
395 | if (dirty_background_bytes) |
396 | background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE); | |
397 | else | |
398 | background = (dirty_background_ratio * available_memory) / 100; | |
1da177e4 | 399 | |
2da02997 DR |
400 | if (background >= dirty) |
401 | background = dirty / 2; | |
1da177e4 LT |
402 | tsk = current; |
403 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) { | |
404 | background += background / 4; | |
405 | dirty += dirty / 4; | |
406 | } | |
407 | *pbackground = background; | |
408 | *pdirty = dirty; | |
e1cbe236 | 409 | trace_global_dirty_state(background, dirty); |
16c4042f | 410 | } |
04fbfdc1 | 411 | |
6f718656 | 412 | /** |
1babe183 | 413 | * bdi_dirty_limit - @bdi's share of dirty throttling threshold |
6f718656 WF |
414 | * @bdi: the backing_dev_info to query |
415 | * @dirty: global dirty limit in pages | |
1babe183 | 416 | * |
6f718656 WF |
417 | * Returns @bdi's dirty limit in pages. The term "dirty" in the context of |
418 | * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages. | |
419 | * And the "limit" in the name is not seriously taken as hard limit in | |
420 | * balance_dirty_pages(). | |
1babe183 | 421 | * |
6f718656 | 422 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
1babe183 WF |
423 | * - starving fast devices |
424 | * - piling up dirty pages (that will take long time to sync) on slow devices | |
425 | * | |
426 | * The bdi's share of dirty limit will be adapting to its throughput and | |
427 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. | |
428 | */ | |
429 | unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty) | |
16c4042f WF |
430 | { |
431 | u64 bdi_dirty; | |
432 | long numerator, denominator; | |
04fbfdc1 | 433 | |
16c4042f WF |
434 | /* |
435 | * Calculate this BDI's share of the dirty ratio. | |
436 | */ | |
437 | bdi_writeout_fraction(bdi, &numerator, &denominator); | |
04fbfdc1 | 438 | |
16c4042f WF |
439 | bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100; |
440 | bdi_dirty *= numerator; | |
441 | do_div(bdi_dirty, denominator); | |
04fbfdc1 | 442 | |
16c4042f WF |
443 | bdi_dirty += (dirty * bdi->min_ratio) / 100; |
444 | if (bdi_dirty > (dirty * bdi->max_ratio) / 100) | |
445 | bdi_dirty = dirty * bdi->max_ratio / 100; | |
446 | ||
447 | return bdi_dirty; | |
1da177e4 LT |
448 | } |
449 | ||
6c14ae1e WF |
450 | /* |
451 | * Dirty position control. | |
452 | * | |
453 | * (o) global/bdi setpoints | |
454 | * | |
455 | * We want the dirty pages be balanced around the global/bdi setpoints. | |
456 | * When the number of dirty pages is higher/lower than the setpoint, the | |
457 | * dirty position control ratio (and hence task dirty ratelimit) will be | |
458 | * decreased/increased to bring the dirty pages back to the setpoint. | |
459 | * | |
460 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT | |
461 | * | |
462 | * if (dirty < setpoint) scale up pos_ratio | |
463 | * if (dirty > setpoint) scale down pos_ratio | |
464 | * | |
465 | * if (bdi_dirty < bdi_setpoint) scale up pos_ratio | |
466 | * if (bdi_dirty > bdi_setpoint) scale down pos_ratio | |
467 | * | |
468 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT | |
469 | * | |
470 | * (o) global control line | |
471 | * | |
472 | * ^ pos_ratio | |
473 | * | | |
474 | * | |<===== global dirty control scope ======>| | |
475 | * 2.0 .............* | |
476 | * | .* | |
477 | * | . * | |
478 | * | . * | |
479 | * | . * | |
480 | * | . * | |
481 | * | . * | |
482 | * 1.0 ................................* | |
483 | * | . . * | |
484 | * | . . * | |
485 | * | . . * | |
486 | * | . . * | |
487 | * | . . * | |
488 | * 0 +------------.------------------.----------------------*-------------> | |
489 | * freerun^ setpoint^ limit^ dirty pages | |
490 | * | |
491 | * (o) bdi control line | |
492 | * | |
493 | * ^ pos_ratio | |
494 | * | | |
495 | * | * | |
496 | * | * | |
497 | * | * | |
498 | * | * | |
499 | * | * |<=========== span ============>| | |
500 | * 1.0 .......................* | |
501 | * | . * | |
502 | * | . * | |
503 | * | . * | |
504 | * | . * | |
505 | * | . * | |
506 | * | . * | |
507 | * | . * | |
508 | * | . * | |
509 | * | . * | |
510 | * | . * | |
511 | * | . * | |
512 | * 1/4 ...............................................* * * * * * * * * * * * | |
513 | * | . . | |
514 | * | . . | |
515 | * | . . | |
516 | * 0 +----------------------.-------------------------------.-------------> | |
517 | * bdi_setpoint^ x_intercept^ | |
518 | * | |
519 | * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can | |
520 | * be smoothly throttled down to normal if it starts high in situations like | |
521 | * - start writing to a slow SD card and a fast disk at the same time. The SD | |
522 | * card's bdi_dirty may rush to many times higher than bdi_setpoint. | |
523 | * - the bdi dirty thresh drops quickly due to change of JBOD workload | |
524 | */ | |
525 | static unsigned long bdi_position_ratio(struct backing_dev_info *bdi, | |
526 | unsigned long thresh, | |
527 | unsigned long bg_thresh, | |
528 | unsigned long dirty, | |
529 | unsigned long bdi_thresh, | |
530 | unsigned long bdi_dirty) | |
531 | { | |
532 | unsigned long write_bw = bdi->avg_write_bandwidth; | |
533 | unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh); | |
534 | unsigned long limit = hard_dirty_limit(thresh); | |
535 | unsigned long x_intercept; | |
536 | unsigned long setpoint; /* dirty pages' target balance point */ | |
537 | unsigned long bdi_setpoint; | |
538 | unsigned long span; | |
539 | long long pos_ratio; /* for scaling up/down the rate limit */ | |
540 | long x; | |
541 | ||
542 | if (unlikely(dirty >= limit)) | |
543 | return 0; | |
544 | ||
545 | /* | |
546 | * global setpoint | |
547 | * | |
548 | * setpoint - dirty 3 | |
549 | * f(dirty) := 1.0 + (----------------) | |
550 | * limit - setpoint | |
551 | * | |
552 | * it's a 3rd order polynomial that subjects to | |
553 | * | |
554 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast | |
555 | * (2) f(setpoint) = 1.0 => the balance point | |
556 | * (3) f(limit) = 0 => the hard limit | |
557 | * (4) df/dx <= 0 => negative feedback control | |
558 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) | |
559 | * => fast response on large errors; small oscillation near setpoint | |
560 | */ | |
561 | setpoint = (freerun + limit) / 2; | |
562 | x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT, | |
563 | limit - setpoint + 1); | |
564 | pos_ratio = x; | |
565 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
566 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
567 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT; | |
568 | ||
569 | /* | |
570 | * We have computed basic pos_ratio above based on global situation. If | |
571 | * the bdi is over/under its share of dirty pages, we want to scale | |
572 | * pos_ratio further down/up. That is done by the following mechanism. | |
573 | */ | |
574 | ||
575 | /* | |
576 | * bdi setpoint | |
577 | * | |
578 | * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint) | |
579 | * | |
580 | * x_intercept - bdi_dirty | |
581 | * := -------------------------- | |
582 | * x_intercept - bdi_setpoint | |
583 | * | |
584 | * The main bdi control line is a linear function that subjects to | |
585 | * | |
586 | * (1) f(bdi_setpoint) = 1.0 | |
587 | * (2) k = - 1 / (8 * write_bw) (in single bdi case) | |
588 | * or equally: x_intercept = bdi_setpoint + 8 * write_bw | |
589 | * | |
590 | * For single bdi case, the dirty pages are observed to fluctuate | |
591 | * regularly within range | |
592 | * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2] | |
593 | * for various filesystems, where (2) can yield in a reasonable 12.5% | |
594 | * fluctuation range for pos_ratio. | |
595 | * | |
596 | * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its | |
597 | * own size, so move the slope over accordingly and choose a slope that | |
598 | * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh. | |
599 | */ | |
600 | if (unlikely(bdi_thresh > thresh)) | |
601 | bdi_thresh = thresh; | |
602 | /* | |
603 | * scale global setpoint to bdi's: | |
604 | * bdi_setpoint = setpoint * bdi_thresh / thresh | |
605 | */ | |
606 | x = div_u64((u64)bdi_thresh << 16, thresh + 1); | |
607 | bdi_setpoint = setpoint * (u64)x >> 16; | |
608 | /* | |
609 | * Use span=(8*write_bw) in single bdi case as indicated by | |
610 | * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case. | |
611 | * | |
612 | * bdi_thresh thresh - bdi_thresh | |
613 | * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh | |
614 | * thresh thresh | |
615 | */ | |
616 | span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16; | |
617 | x_intercept = bdi_setpoint + span; | |
618 | ||
619 | if (bdi_dirty < x_intercept - span / 4) { | |
620 | pos_ratio *= x_intercept - bdi_dirty; | |
621 | do_div(pos_ratio, x_intercept - bdi_setpoint + 1); | |
622 | } else | |
623 | pos_ratio /= 4; | |
624 | ||
625 | return pos_ratio; | |
626 | } | |
627 | ||
e98be2d5 WF |
628 | static void bdi_update_write_bandwidth(struct backing_dev_info *bdi, |
629 | unsigned long elapsed, | |
630 | unsigned long written) | |
631 | { | |
632 | const unsigned long period = roundup_pow_of_two(3 * HZ); | |
633 | unsigned long avg = bdi->avg_write_bandwidth; | |
634 | unsigned long old = bdi->write_bandwidth; | |
635 | u64 bw; | |
636 | ||
637 | /* | |
638 | * bw = written * HZ / elapsed | |
639 | * | |
640 | * bw * elapsed + write_bandwidth * (period - elapsed) | |
641 | * write_bandwidth = --------------------------------------------------- | |
642 | * period | |
643 | */ | |
644 | bw = written - bdi->written_stamp; | |
645 | bw *= HZ; | |
646 | if (unlikely(elapsed > period)) { | |
647 | do_div(bw, elapsed); | |
648 | avg = bw; | |
649 | goto out; | |
650 | } | |
651 | bw += (u64)bdi->write_bandwidth * (period - elapsed); | |
652 | bw >>= ilog2(period); | |
653 | ||
654 | /* | |
655 | * one more level of smoothing, for filtering out sudden spikes | |
656 | */ | |
657 | if (avg > old && old >= (unsigned long)bw) | |
658 | avg -= (avg - old) >> 3; | |
659 | ||
660 | if (avg < old && old <= (unsigned long)bw) | |
661 | avg += (old - avg) >> 3; | |
662 | ||
663 | out: | |
664 | bdi->write_bandwidth = bw; | |
665 | bdi->avg_write_bandwidth = avg; | |
666 | } | |
667 | ||
c42843f2 WF |
668 | /* |
669 | * The global dirtyable memory and dirty threshold could be suddenly knocked | |
670 | * down by a large amount (eg. on the startup of KVM in a swapless system). | |
671 | * This may throw the system into deep dirty exceeded state and throttle | |
672 | * heavy/light dirtiers alike. To retain good responsiveness, maintain | |
673 | * global_dirty_limit for tracking slowly down to the knocked down dirty | |
674 | * threshold. | |
675 | */ | |
676 | static void update_dirty_limit(unsigned long thresh, unsigned long dirty) | |
677 | { | |
678 | unsigned long limit = global_dirty_limit; | |
679 | ||
680 | /* | |
681 | * Follow up in one step. | |
682 | */ | |
683 | if (limit < thresh) { | |
684 | limit = thresh; | |
685 | goto update; | |
686 | } | |
687 | ||
688 | /* | |
689 | * Follow down slowly. Use the higher one as the target, because thresh | |
690 | * may drop below dirty. This is exactly the reason to introduce | |
691 | * global_dirty_limit which is guaranteed to lie above the dirty pages. | |
692 | */ | |
693 | thresh = max(thresh, dirty); | |
694 | if (limit > thresh) { | |
695 | limit -= (limit - thresh) >> 5; | |
696 | goto update; | |
697 | } | |
698 | return; | |
699 | update: | |
700 | global_dirty_limit = limit; | |
701 | } | |
702 | ||
703 | static void global_update_bandwidth(unsigned long thresh, | |
704 | unsigned long dirty, | |
705 | unsigned long now) | |
706 | { | |
707 | static DEFINE_SPINLOCK(dirty_lock); | |
708 | static unsigned long update_time; | |
709 | ||
710 | /* | |
711 | * check locklessly first to optimize away locking for the most time | |
712 | */ | |
713 | if (time_before(now, update_time + BANDWIDTH_INTERVAL)) | |
714 | return; | |
715 | ||
716 | spin_lock(&dirty_lock); | |
717 | if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) { | |
718 | update_dirty_limit(thresh, dirty); | |
719 | update_time = now; | |
720 | } | |
721 | spin_unlock(&dirty_lock); | |
722 | } | |
723 | ||
be3ffa27 WF |
724 | /* |
725 | * Maintain bdi->dirty_ratelimit, the base dirty throttle rate. | |
726 | * | |
727 | * Normal bdi tasks will be curbed at or below it in long term. | |
728 | * Obviously it should be around (write_bw / N) when there are N dd tasks. | |
729 | */ | |
730 | static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi, | |
731 | unsigned long thresh, | |
732 | unsigned long bg_thresh, | |
733 | unsigned long dirty, | |
734 | unsigned long bdi_thresh, | |
735 | unsigned long bdi_dirty, | |
736 | unsigned long dirtied, | |
737 | unsigned long elapsed) | |
738 | { | |
7381131c WF |
739 | unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh); |
740 | unsigned long limit = hard_dirty_limit(thresh); | |
741 | unsigned long setpoint = (freerun + limit) / 2; | |
be3ffa27 WF |
742 | unsigned long write_bw = bdi->avg_write_bandwidth; |
743 | unsigned long dirty_ratelimit = bdi->dirty_ratelimit; | |
744 | unsigned long dirty_rate; | |
745 | unsigned long task_ratelimit; | |
746 | unsigned long balanced_dirty_ratelimit; | |
747 | unsigned long pos_ratio; | |
7381131c WF |
748 | unsigned long step; |
749 | unsigned long x; | |
be3ffa27 WF |
750 | |
751 | /* | |
752 | * The dirty rate will match the writeout rate in long term, except | |
753 | * when dirty pages are truncated by userspace or re-dirtied by FS. | |
754 | */ | |
755 | dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed; | |
756 | ||
757 | pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty, | |
758 | bdi_thresh, bdi_dirty); | |
759 | /* | |
760 | * task_ratelimit reflects each dd's dirty rate for the past 200ms. | |
761 | */ | |
762 | task_ratelimit = (u64)dirty_ratelimit * | |
763 | pos_ratio >> RATELIMIT_CALC_SHIFT; | |
764 | task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ | |
765 | ||
766 | /* | |
767 | * A linear estimation of the "balanced" throttle rate. The theory is, | |
768 | * if there are N dd tasks, each throttled at task_ratelimit, the bdi's | |
769 | * dirty_rate will be measured to be (N * task_ratelimit). So the below | |
770 | * formula will yield the balanced rate limit (write_bw / N). | |
771 | * | |
772 | * Note that the expanded form is not a pure rate feedback: | |
773 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) | |
774 | * but also takes pos_ratio into account: | |
775 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) | |
776 | * | |
777 | * (1) is not realistic because pos_ratio also takes part in balancing | |
778 | * the dirty rate. Consider the state | |
779 | * pos_ratio = 0.5 (3) | |
780 | * rate = 2 * (write_bw / N) (4) | |
781 | * If (1) is used, it will stuck in that state! Because each dd will | |
782 | * be throttled at | |
783 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) | |
784 | * yielding | |
785 | * dirty_rate = N * task_ratelimit = write_bw (6) | |
786 | * put (6) into (1) we get | |
787 | * rate_(i+1) = rate_(i) (7) | |
788 | * | |
789 | * So we end up using (2) to always keep | |
790 | * rate_(i+1) ~= (write_bw / N) (8) | |
791 | * regardless of the value of pos_ratio. As long as (8) is satisfied, | |
792 | * pos_ratio is able to drive itself to 1.0, which is not only where | |
793 | * the dirty count meet the setpoint, but also where the slope of | |
794 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated. | |
795 | */ | |
796 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, | |
797 | dirty_rate | 1); | |
798 | ||
7381131c WF |
799 | /* |
800 | * We could safely do this and return immediately: | |
801 | * | |
802 | * bdi->dirty_ratelimit = balanced_dirty_ratelimit; | |
803 | * | |
804 | * However to get a more stable dirty_ratelimit, the below elaborated | |
805 | * code makes use of task_ratelimit to filter out sigular points and | |
806 | * limit the step size. | |
807 | * | |
808 | * The below code essentially only uses the relative value of | |
809 | * | |
810 | * task_ratelimit - dirty_ratelimit | |
811 | * = (pos_ratio - 1) * dirty_ratelimit | |
812 | * | |
813 | * which reflects the direction and size of dirty position error. | |
814 | */ | |
815 | ||
816 | /* | |
817 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff | |
818 | * task_ratelimit is on the same side of dirty_ratelimit, too. | |
819 | * For example, when | |
820 | * - dirty_ratelimit > balanced_dirty_ratelimit | |
821 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) | |
822 | * lowering dirty_ratelimit will help meet both the position and rate | |
823 | * control targets. Otherwise, don't update dirty_ratelimit if it will | |
824 | * only help meet the rate target. After all, what the users ultimately | |
825 | * feel and care are stable dirty rate and small position error. | |
826 | * | |
827 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size | |
828 | * and filter out the sigular points of balanced_dirty_ratelimit. Which | |
829 | * keeps jumping around randomly and can even leap far away at times | |
830 | * due to the small 200ms estimation period of dirty_rate (we want to | |
831 | * keep that period small to reduce time lags). | |
832 | */ | |
833 | step = 0; | |
834 | if (dirty < setpoint) { | |
835 | x = min(bdi->balanced_dirty_ratelimit, | |
836 | min(balanced_dirty_ratelimit, task_ratelimit)); | |
837 | if (dirty_ratelimit < x) | |
838 | step = x - dirty_ratelimit; | |
839 | } else { | |
840 | x = max(bdi->balanced_dirty_ratelimit, | |
841 | max(balanced_dirty_ratelimit, task_ratelimit)); | |
842 | if (dirty_ratelimit > x) | |
843 | step = dirty_ratelimit - x; | |
844 | } | |
845 | ||
846 | /* | |
847 | * Don't pursue 100% rate matching. It's impossible since the balanced | |
848 | * rate itself is constantly fluctuating. So decrease the track speed | |
849 | * when it gets close to the target. Helps eliminate pointless tremors. | |
850 | */ | |
851 | step >>= dirty_ratelimit / (2 * step + 1); | |
852 | /* | |
853 | * Limit the tracking speed to avoid overshooting. | |
854 | */ | |
855 | step = (step + 7) / 8; | |
856 | ||
857 | if (dirty_ratelimit < balanced_dirty_ratelimit) | |
858 | dirty_ratelimit += step; | |
859 | else | |
860 | dirty_ratelimit -= step; | |
861 | ||
862 | bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL); | |
863 | bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit; | |
be3ffa27 WF |
864 | } |
865 | ||
e98be2d5 | 866 | void __bdi_update_bandwidth(struct backing_dev_info *bdi, |
c42843f2 | 867 | unsigned long thresh, |
af6a3113 | 868 | unsigned long bg_thresh, |
c42843f2 WF |
869 | unsigned long dirty, |
870 | unsigned long bdi_thresh, | |
871 | unsigned long bdi_dirty, | |
e98be2d5 WF |
872 | unsigned long start_time) |
873 | { | |
874 | unsigned long now = jiffies; | |
875 | unsigned long elapsed = now - bdi->bw_time_stamp; | |
be3ffa27 | 876 | unsigned long dirtied; |
e98be2d5 WF |
877 | unsigned long written; |
878 | ||
879 | /* | |
880 | * rate-limit, only update once every 200ms. | |
881 | */ | |
882 | if (elapsed < BANDWIDTH_INTERVAL) | |
883 | return; | |
884 | ||
be3ffa27 | 885 | dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]); |
e98be2d5 WF |
886 | written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]); |
887 | ||
888 | /* | |
889 | * Skip quiet periods when disk bandwidth is under-utilized. | |
890 | * (at least 1s idle time between two flusher runs) | |
891 | */ | |
892 | if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time)) | |
893 | goto snapshot; | |
894 | ||
be3ffa27 | 895 | if (thresh) { |
c42843f2 | 896 | global_update_bandwidth(thresh, dirty, now); |
be3ffa27 WF |
897 | bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty, |
898 | bdi_thresh, bdi_dirty, | |
899 | dirtied, elapsed); | |
900 | } | |
e98be2d5 WF |
901 | bdi_update_write_bandwidth(bdi, elapsed, written); |
902 | ||
903 | snapshot: | |
be3ffa27 | 904 | bdi->dirtied_stamp = dirtied; |
e98be2d5 WF |
905 | bdi->written_stamp = written; |
906 | bdi->bw_time_stamp = now; | |
907 | } | |
908 | ||
909 | static void bdi_update_bandwidth(struct backing_dev_info *bdi, | |
c42843f2 | 910 | unsigned long thresh, |
af6a3113 | 911 | unsigned long bg_thresh, |
c42843f2 WF |
912 | unsigned long dirty, |
913 | unsigned long bdi_thresh, | |
914 | unsigned long bdi_dirty, | |
e98be2d5 WF |
915 | unsigned long start_time) |
916 | { | |
917 | if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL)) | |
918 | return; | |
919 | spin_lock(&bdi->wb.list_lock); | |
af6a3113 WF |
920 | __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty, |
921 | bdi_thresh, bdi_dirty, start_time); | |
e98be2d5 WF |
922 | spin_unlock(&bdi->wb.list_lock); |
923 | } | |
924 | ||
9d823e8f WF |
925 | /* |
926 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr() | |
927 | * will look to see if it needs to start dirty throttling. | |
928 | * | |
929 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive | |
930 | * global_page_state() too often. So scale it near-sqrt to the safety margin | |
931 | * (the number of pages we may dirty without exceeding the dirty limits). | |
932 | */ | |
933 | static unsigned long dirty_poll_interval(unsigned long dirty, | |
934 | unsigned long thresh) | |
935 | { | |
936 | if (thresh > dirty) | |
937 | return 1UL << (ilog2(thresh - dirty) >> 1); | |
938 | ||
939 | return 1; | |
940 | } | |
941 | ||
c8462cc9 WF |
942 | static unsigned long bdi_max_pause(struct backing_dev_info *bdi, |
943 | unsigned long bdi_dirty) | |
944 | { | |
945 | unsigned long bw = bdi->avg_write_bandwidth; | |
946 | unsigned long hi = ilog2(bw); | |
947 | unsigned long lo = ilog2(bdi->dirty_ratelimit); | |
948 | unsigned long t; | |
949 | ||
950 | /* target for 20ms max pause on 1-dd case */ | |
951 | t = HZ / 50; | |
952 | ||
953 | /* | |
954 | * Scale up pause time for concurrent dirtiers in order to reduce CPU | |
955 | * overheads. | |
956 | * | |
957 | * (N * 20ms) on 2^N concurrent tasks. | |
958 | */ | |
959 | if (hi > lo) | |
960 | t += (hi - lo) * (20 * HZ) / 1024; | |
961 | ||
962 | /* | |
963 | * Limit pause time for small memory systems. If sleeping for too long | |
964 | * time, a small pool of dirty/writeback pages may go empty and disk go | |
965 | * idle. | |
966 | * | |
967 | * 8 serves as the safety ratio. | |
968 | */ | |
969 | if (bdi_dirty) | |
970 | t = min(t, bdi_dirty * HZ / (8 * bw + 1)); | |
971 | ||
972 | /* | |
973 | * The pause time will be settled within range (max_pause/4, max_pause). | |
974 | * Apply a minimal value of 4 to get a non-zero max_pause/4. | |
975 | */ | |
976 | return clamp_val(t, 4, MAX_PAUSE); | |
977 | } | |
978 | ||
1da177e4 LT |
979 | /* |
980 | * balance_dirty_pages() must be called by processes which are generating dirty | |
981 | * data. It looks at the number of dirty pages in the machine and will force | |
143dfe86 | 982 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. |
5b0830cb JA |
983 | * If we're over `background_thresh' then the writeback threads are woken to |
984 | * perform some writeout. | |
1da177e4 | 985 | */ |
3a2e9a5a | 986 | static void balance_dirty_pages(struct address_space *mapping, |
143dfe86 | 987 | unsigned long pages_dirtied) |
1da177e4 | 988 | { |
143dfe86 WF |
989 | unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */ |
990 | unsigned long bdi_reclaimable; | |
7762741e WF |
991 | unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */ |
992 | unsigned long bdi_dirty; | |
6c14ae1e | 993 | unsigned long freerun; |
364aeb28 DR |
994 | unsigned long background_thresh; |
995 | unsigned long dirty_thresh; | |
996 | unsigned long bdi_thresh; | |
143dfe86 | 997 | long pause = 0; |
c8462cc9 | 998 | long max_pause; |
e50e3720 | 999 | bool dirty_exceeded = false; |
143dfe86 WF |
1000 | unsigned long task_ratelimit; |
1001 | unsigned long dirty_ratelimit; | |
1002 | unsigned long pos_ratio; | |
1da177e4 | 1003 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
e98be2d5 | 1004 | unsigned long start_time = jiffies; |
1da177e4 LT |
1005 | |
1006 | for (;;) { | |
143dfe86 WF |
1007 | /* |
1008 | * Unstable writes are a feature of certain networked | |
1009 | * filesystems (i.e. NFS) in which data may have been | |
1010 | * written to the server's write cache, but has not yet | |
1011 | * been flushed to permanent storage. | |
1012 | */ | |
5fce25a9 PZ |
1013 | nr_reclaimable = global_page_state(NR_FILE_DIRTY) + |
1014 | global_page_state(NR_UNSTABLE_NFS); | |
7762741e | 1015 | nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK); |
5fce25a9 | 1016 | |
16c4042f WF |
1017 | global_dirty_limits(&background_thresh, &dirty_thresh); |
1018 | ||
1019 | /* | |
1020 | * Throttle it only when the background writeback cannot | |
1021 | * catch-up. This avoids (excessively) small writeouts | |
1022 | * when the bdi limits are ramping up. | |
1023 | */ | |
6c14ae1e WF |
1024 | freerun = dirty_freerun_ceiling(dirty_thresh, |
1025 | background_thresh); | |
1026 | if (nr_dirty <= freerun) | |
16c4042f WF |
1027 | break; |
1028 | ||
143dfe86 WF |
1029 | if (unlikely(!writeback_in_progress(bdi))) |
1030 | bdi_start_background_writeback(bdi); | |
1031 | ||
1032 | /* | |
1033 | * bdi_thresh is not treated as some limiting factor as | |
1034 | * dirty_thresh, due to reasons | |
1035 | * - in JBOD setup, bdi_thresh can fluctuate a lot | |
1036 | * - in a system with HDD and USB key, the USB key may somehow | |
1037 | * go into state (bdi_dirty >> bdi_thresh) either because | |
1038 | * bdi_dirty starts high, or because bdi_thresh drops low. | |
1039 | * In this case we don't want to hard throttle the USB key | |
1040 | * dirtiers for 100 seconds until bdi_dirty drops under | |
1041 | * bdi_thresh. Instead the auxiliary bdi control line in | |
1042 | * bdi_position_ratio() will let the dirtier task progress | |
1043 | * at some rate <= (write_bw / 2) for bringing down bdi_dirty. | |
1044 | */ | |
16c4042f | 1045 | bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh); |
16c4042f | 1046 | |
e50e3720 WF |
1047 | /* |
1048 | * In order to avoid the stacked BDI deadlock we need | |
1049 | * to ensure we accurately count the 'dirty' pages when | |
1050 | * the threshold is low. | |
1051 | * | |
1052 | * Otherwise it would be possible to get thresh+n pages | |
1053 | * reported dirty, even though there are thresh-m pages | |
1054 | * actually dirty; with m+n sitting in the percpu | |
1055 | * deltas. | |
1056 | */ | |
143dfe86 WF |
1057 | if (bdi_thresh < 2 * bdi_stat_error(bdi)) { |
1058 | bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE); | |
1059 | bdi_dirty = bdi_reclaimable + | |
7762741e | 1060 | bdi_stat_sum(bdi, BDI_WRITEBACK); |
e50e3720 | 1061 | } else { |
143dfe86 WF |
1062 | bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE); |
1063 | bdi_dirty = bdi_reclaimable + | |
7762741e | 1064 | bdi_stat(bdi, BDI_WRITEBACK); |
e50e3720 | 1065 | } |
5fce25a9 | 1066 | |
143dfe86 | 1067 | dirty_exceeded = (bdi_dirty > bdi_thresh) || |
7762741e | 1068 | (nr_dirty > dirty_thresh); |
143dfe86 | 1069 | if (dirty_exceeded && !bdi->dirty_exceeded) |
04fbfdc1 | 1070 | bdi->dirty_exceeded = 1; |
1da177e4 | 1071 | |
af6a3113 WF |
1072 | bdi_update_bandwidth(bdi, dirty_thresh, background_thresh, |
1073 | nr_dirty, bdi_thresh, bdi_dirty, | |
1074 | start_time); | |
e98be2d5 | 1075 | |
c8462cc9 WF |
1076 | max_pause = bdi_max_pause(bdi, bdi_dirty); |
1077 | ||
143dfe86 WF |
1078 | dirty_ratelimit = bdi->dirty_ratelimit; |
1079 | pos_ratio = bdi_position_ratio(bdi, dirty_thresh, | |
1080 | background_thresh, nr_dirty, | |
1081 | bdi_thresh, bdi_dirty); | |
1082 | if (unlikely(pos_ratio == 0)) { | |
c8462cc9 | 1083 | pause = max_pause; |
143dfe86 | 1084 | goto pause; |
04fbfdc1 | 1085 | } |
143dfe86 WF |
1086 | task_ratelimit = (u64)dirty_ratelimit * |
1087 | pos_ratio >> RATELIMIT_CALC_SHIFT; | |
1088 | pause = (HZ * pages_dirtied) / (task_ratelimit | 1); | |
c8462cc9 | 1089 | pause = min(pause, max_pause); |
143dfe86 WF |
1090 | |
1091 | pause: | |
d153ba64 | 1092 | __set_current_state(TASK_UNINTERRUPTIBLE); |
d25105e8 | 1093 | io_schedule_timeout(pause); |
87c6a9b2 | 1094 | |
ffd1f609 WF |
1095 | dirty_thresh = hard_dirty_limit(dirty_thresh); |
1096 | /* | |
1097 | * max-pause area. If dirty exceeded but still within this | |
1098 | * area, no need to sleep for more than 200ms: (a) 8 pages per | |
1099 | * 200ms is typically more than enough to curb heavy dirtiers; | |
1100 | * (b) the pause time limit makes the dirtiers more responsive. | |
1101 | */ | |
143dfe86 | 1102 | if (nr_dirty < dirty_thresh) |
ffd1f609 | 1103 | break; |
1da177e4 LT |
1104 | } |
1105 | ||
143dfe86 | 1106 | if (!dirty_exceeded && bdi->dirty_exceeded) |
04fbfdc1 | 1107 | bdi->dirty_exceeded = 0; |
1da177e4 | 1108 | |
9d823e8f WF |
1109 | current->nr_dirtied = 0; |
1110 | current->nr_dirtied_pause = dirty_poll_interval(nr_dirty, dirty_thresh); | |
1111 | ||
1da177e4 | 1112 | if (writeback_in_progress(bdi)) |
5b0830cb | 1113 | return; |
1da177e4 LT |
1114 | |
1115 | /* | |
1116 | * In laptop mode, we wait until hitting the higher threshold before | |
1117 | * starting background writeout, and then write out all the way down | |
1118 | * to the lower threshold. So slow writers cause minimal disk activity. | |
1119 | * | |
1120 | * In normal mode, we start background writeout at the lower | |
1121 | * background_thresh, to keep the amount of dirty memory low. | |
1122 | */ | |
143dfe86 WF |
1123 | if (laptop_mode) |
1124 | return; | |
1125 | ||
1126 | if (nr_reclaimable > background_thresh) | |
c5444198 | 1127 | bdi_start_background_writeback(bdi); |
1da177e4 LT |
1128 | } |
1129 | ||
a200ee18 | 1130 | void set_page_dirty_balance(struct page *page, int page_mkwrite) |
edc79b2a | 1131 | { |
a200ee18 | 1132 | if (set_page_dirty(page) || page_mkwrite) { |
edc79b2a PZ |
1133 | struct address_space *mapping = page_mapping(page); |
1134 | ||
1135 | if (mapping) | |
1136 | balance_dirty_pages_ratelimited(mapping); | |
1137 | } | |
1138 | } | |
1139 | ||
9d823e8f | 1140 | static DEFINE_PER_CPU(int, bdp_ratelimits); |
245b2e70 | 1141 | |
1da177e4 | 1142 | /** |
fa5a734e | 1143 | * balance_dirty_pages_ratelimited_nr - balance dirty memory state |
67be2dd1 | 1144 | * @mapping: address_space which was dirtied |
a580290c | 1145 | * @nr_pages_dirtied: number of pages which the caller has just dirtied |
1da177e4 LT |
1146 | * |
1147 | * Processes which are dirtying memory should call in here once for each page | |
1148 | * which was newly dirtied. The function will periodically check the system's | |
1149 | * dirty state and will initiate writeback if needed. | |
1150 | * | |
1151 | * On really big machines, get_writeback_state is expensive, so try to avoid | |
1152 | * calling it too often (ratelimiting). But once we're over the dirty memory | |
1153 | * limit we decrease the ratelimiting by a lot, to prevent individual processes | |
1154 | * from overshooting the limit by (ratelimit_pages) each. | |
1155 | */ | |
fa5a734e AM |
1156 | void balance_dirty_pages_ratelimited_nr(struct address_space *mapping, |
1157 | unsigned long nr_pages_dirtied) | |
1da177e4 | 1158 | { |
36715cef | 1159 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
9d823e8f WF |
1160 | int ratelimit; |
1161 | int *p; | |
1da177e4 | 1162 | |
36715cef WF |
1163 | if (!bdi_cap_account_dirty(bdi)) |
1164 | return; | |
1165 | ||
9d823e8f WF |
1166 | ratelimit = current->nr_dirtied_pause; |
1167 | if (bdi->dirty_exceeded) | |
1168 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); | |
1169 | ||
1170 | current->nr_dirtied += nr_pages_dirtied; | |
1da177e4 | 1171 | |
9d823e8f | 1172 | preempt_disable(); |
1da177e4 | 1173 | /* |
9d823e8f WF |
1174 | * This prevents one CPU to accumulate too many dirtied pages without |
1175 | * calling into balance_dirty_pages(), which can happen when there are | |
1176 | * 1000+ tasks, all of them start dirtying pages at exactly the same | |
1177 | * time, hence all honoured too large initial task->nr_dirtied_pause. | |
1da177e4 | 1178 | */ |
245b2e70 | 1179 | p = &__get_cpu_var(bdp_ratelimits); |
9d823e8f | 1180 | if (unlikely(current->nr_dirtied >= ratelimit)) |
fa5a734e | 1181 | *p = 0; |
9d823e8f WF |
1182 | else { |
1183 | *p += nr_pages_dirtied; | |
1184 | if (unlikely(*p >= ratelimit_pages)) { | |
1185 | *p = 0; | |
1186 | ratelimit = 0; | |
1187 | } | |
1da177e4 | 1188 | } |
fa5a734e | 1189 | preempt_enable(); |
9d823e8f WF |
1190 | |
1191 | if (unlikely(current->nr_dirtied >= ratelimit)) | |
1192 | balance_dirty_pages(mapping, current->nr_dirtied); | |
1da177e4 | 1193 | } |
fa5a734e | 1194 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr); |
1da177e4 | 1195 | |
232ea4d6 | 1196 | void throttle_vm_writeout(gfp_t gfp_mask) |
1da177e4 | 1197 | { |
364aeb28 DR |
1198 | unsigned long background_thresh; |
1199 | unsigned long dirty_thresh; | |
1da177e4 LT |
1200 | |
1201 | for ( ; ; ) { | |
16c4042f | 1202 | global_dirty_limits(&background_thresh, &dirty_thresh); |
1da177e4 LT |
1203 | |
1204 | /* | |
1205 | * Boost the allowable dirty threshold a bit for page | |
1206 | * allocators so they don't get DoS'ed by heavy writers | |
1207 | */ | |
1208 | dirty_thresh += dirty_thresh / 10; /* wheeee... */ | |
1209 | ||
c24f21bd CL |
1210 | if (global_page_state(NR_UNSTABLE_NFS) + |
1211 | global_page_state(NR_WRITEBACK) <= dirty_thresh) | |
1212 | break; | |
8aa7e847 | 1213 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
369f2389 FW |
1214 | |
1215 | /* | |
1216 | * The caller might hold locks which can prevent IO completion | |
1217 | * or progress in the filesystem. So we cannot just sit here | |
1218 | * waiting for IO to complete. | |
1219 | */ | |
1220 | if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) | |
1221 | break; | |
1da177e4 LT |
1222 | } |
1223 | } | |
1224 | ||
1da177e4 LT |
1225 | /* |
1226 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
1227 | */ | |
1228 | int dirty_writeback_centisecs_handler(ctl_table *table, int write, | |
8d65af78 | 1229 | void __user *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 1230 | { |
8d65af78 | 1231 | proc_dointvec(table, write, buffer, length, ppos); |
6423104b | 1232 | bdi_arm_supers_timer(); |
1da177e4 LT |
1233 | return 0; |
1234 | } | |
1235 | ||
c2c4986e | 1236 | #ifdef CONFIG_BLOCK |
31373d09 | 1237 | void laptop_mode_timer_fn(unsigned long data) |
1da177e4 | 1238 | { |
31373d09 MG |
1239 | struct request_queue *q = (struct request_queue *)data; |
1240 | int nr_pages = global_page_state(NR_FILE_DIRTY) + | |
1241 | global_page_state(NR_UNSTABLE_NFS); | |
1da177e4 | 1242 | |
31373d09 MG |
1243 | /* |
1244 | * We want to write everything out, not just down to the dirty | |
1245 | * threshold | |
1246 | */ | |
31373d09 | 1247 | if (bdi_has_dirty_io(&q->backing_dev_info)) |
c5444198 | 1248 | bdi_start_writeback(&q->backing_dev_info, nr_pages); |
1da177e4 LT |
1249 | } |
1250 | ||
1251 | /* | |
1252 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
1253 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
1254 | * then push it back - the user is still using the disk. | |
1255 | */ | |
31373d09 | 1256 | void laptop_io_completion(struct backing_dev_info *info) |
1da177e4 | 1257 | { |
31373d09 | 1258 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
1259 | } |
1260 | ||
1261 | /* | |
1262 | * We're in laptop mode and we've just synced. The sync's writes will have | |
1263 | * caused another writeback to be scheduled by laptop_io_completion. | |
1264 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
1265 | */ | |
1266 | void laptop_sync_completion(void) | |
1267 | { | |
31373d09 MG |
1268 | struct backing_dev_info *bdi; |
1269 | ||
1270 | rcu_read_lock(); | |
1271 | ||
1272 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) | |
1273 | del_timer(&bdi->laptop_mode_wb_timer); | |
1274 | ||
1275 | rcu_read_unlock(); | |
1da177e4 | 1276 | } |
c2c4986e | 1277 | #endif |
1da177e4 LT |
1278 | |
1279 | /* | |
1280 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
1281 | * if a large number of processes all perform writes at the same time. | |
1282 | * If it is too low then SMP machines will call the (expensive) | |
1283 | * get_writeback_state too often. | |
1284 | * | |
1285 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
1286 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
9d823e8f | 1287 | * thresholds. |
1da177e4 LT |
1288 | */ |
1289 | ||
2d1d43f6 | 1290 | void writeback_set_ratelimit(void) |
1da177e4 | 1291 | { |
9d823e8f WF |
1292 | unsigned long background_thresh; |
1293 | unsigned long dirty_thresh; | |
1294 | global_dirty_limits(&background_thresh, &dirty_thresh); | |
1295 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); | |
1da177e4 LT |
1296 | if (ratelimit_pages < 16) |
1297 | ratelimit_pages = 16; | |
1da177e4 LT |
1298 | } |
1299 | ||
26c2143b | 1300 | static int __cpuinit |
1da177e4 LT |
1301 | ratelimit_handler(struct notifier_block *self, unsigned long u, void *v) |
1302 | { | |
2d1d43f6 | 1303 | writeback_set_ratelimit(); |
aa0f0303 | 1304 | return NOTIFY_DONE; |
1da177e4 LT |
1305 | } |
1306 | ||
74b85f37 | 1307 | static struct notifier_block __cpuinitdata ratelimit_nb = { |
1da177e4 LT |
1308 | .notifier_call = ratelimit_handler, |
1309 | .next = NULL, | |
1310 | }; | |
1311 | ||
1312 | /* | |
dc6e29da LT |
1313 | * Called early on to tune the page writeback dirty limits. |
1314 | * | |
1315 | * We used to scale dirty pages according to how total memory | |
1316 | * related to pages that could be allocated for buffers (by | |
1317 | * comparing nr_free_buffer_pages() to vm_total_pages. | |
1318 | * | |
1319 | * However, that was when we used "dirty_ratio" to scale with | |
1320 | * all memory, and we don't do that any more. "dirty_ratio" | |
1321 | * is now applied to total non-HIGHPAGE memory (by subtracting | |
1322 | * totalhigh_pages from vm_total_pages), and as such we can't | |
1323 | * get into the old insane situation any more where we had | |
1324 | * large amounts of dirty pages compared to a small amount of | |
1325 | * non-HIGHMEM memory. | |
1326 | * | |
1327 | * But we might still want to scale the dirty_ratio by how | |
1328 | * much memory the box has.. | |
1da177e4 LT |
1329 | */ |
1330 | void __init page_writeback_init(void) | |
1331 | { | |
04fbfdc1 PZ |
1332 | int shift; |
1333 | ||
2d1d43f6 | 1334 | writeback_set_ratelimit(); |
1da177e4 | 1335 | register_cpu_notifier(&ratelimit_nb); |
04fbfdc1 PZ |
1336 | |
1337 | shift = calc_period_shift(); | |
1338 | prop_descriptor_init(&vm_completions, shift); | |
3e26c149 | 1339 | prop_descriptor_init(&vm_dirties, shift); |
1da177e4 LT |
1340 | } |
1341 | ||
f446daae JK |
1342 | /** |
1343 | * tag_pages_for_writeback - tag pages to be written by write_cache_pages | |
1344 | * @mapping: address space structure to write | |
1345 | * @start: starting page index | |
1346 | * @end: ending page index (inclusive) | |
1347 | * | |
1348 | * This function scans the page range from @start to @end (inclusive) and tags | |
1349 | * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is | |
1350 | * that write_cache_pages (or whoever calls this function) will then use | |
1351 | * TOWRITE tag to identify pages eligible for writeback. This mechanism is | |
1352 | * used to avoid livelocking of writeback by a process steadily creating new | |
1353 | * dirty pages in the file (thus it is important for this function to be quick | |
1354 | * so that it can tag pages faster than a dirtying process can create them). | |
1355 | */ | |
1356 | /* | |
1357 | * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency. | |
1358 | */ | |
f446daae JK |
1359 | void tag_pages_for_writeback(struct address_space *mapping, |
1360 | pgoff_t start, pgoff_t end) | |
1361 | { | |
3c111a07 | 1362 | #define WRITEBACK_TAG_BATCH 4096 |
f446daae JK |
1363 | unsigned long tagged; |
1364 | ||
1365 | do { | |
1366 | spin_lock_irq(&mapping->tree_lock); | |
1367 | tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree, | |
1368 | &start, end, WRITEBACK_TAG_BATCH, | |
1369 | PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE); | |
1370 | spin_unlock_irq(&mapping->tree_lock); | |
1371 | WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH); | |
1372 | cond_resched(); | |
d5ed3a4a JK |
1373 | /* We check 'start' to handle wrapping when end == ~0UL */ |
1374 | } while (tagged >= WRITEBACK_TAG_BATCH && start); | |
f446daae JK |
1375 | } |
1376 | EXPORT_SYMBOL(tag_pages_for_writeback); | |
1377 | ||
811d736f | 1378 | /** |
0ea97180 | 1379 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. |
811d736f DH |
1380 | * @mapping: address space structure to write |
1381 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
0ea97180 MS |
1382 | * @writepage: function called for each page |
1383 | * @data: data passed to writepage function | |
811d736f | 1384 | * |
0ea97180 | 1385 | * If a page is already under I/O, write_cache_pages() skips it, even |
811d736f DH |
1386 | * if it's dirty. This is desirable behaviour for memory-cleaning writeback, |
1387 | * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() | |
1388 | * and msync() need to guarantee that all the data which was dirty at the time | |
1389 | * the call was made get new I/O started against them. If wbc->sync_mode is | |
1390 | * WB_SYNC_ALL then we were called for data integrity and we must wait for | |
1391 | * existing IO to complete. | |
f446daae JK |
1392 | * |
1393 | * To avoid livelocks (when other process dirties new pages), we first tag | |
1394 | * pages which should be written back with TOWRITE tag and only then start | |
1395 | * writing them. For data-integrity sync we have to be careful so that we do | |
1396 | * not miss some pages (e.g., because some other process has cleared TOWRITE | |
1397 | * tag we set). The rule we follow is that TOWRITE tag can be cleared only | |
1398 | * by the process clearing the DIRTY tag (and submitting the page for IO). | |
811d736f | 1399 | */ |
0ea97180 MS |
1400 | int write_cache_pages(struct address_space *mapping, |
1401 | struct writeback_control *wbc, writepage_t writepage, | |
1402 | void *data) | |
811d736f | 1403 | { |
811d736f DH |
1404 | int ret = 0; |
1405 | int done = 0; | |
811d736f DH |
1406 | struct pagevec pvec; |
1407 | int nr_pages; | |
31a12666 | 1408 | pgoff_t uninitialized_var(writeback_index); |
811d736f DH |
1409 | pgoff_t index; |
1410 | pgoff_t end; /* Inclusive */ | |
bd19e012 | 1411 | pgoff_t done_index; |
31a12666 | 1412 | int cycled; |
811d736f | 1413 | int range_whole = 0; |
f446daae | 1414 | int tag; |
811d736f | 1415 | |
811d736f DH |
1416 | pagevec_init(&pvec, 0); |
1417 | if (wbc->range_cyclic) { | |
31a12666 NP |
1418 | writeback_index = mapping->writeback_index; /* prev offset */ |
1419 | index = writeback_index; | |
1420 | if (index == 0) | |
1421 | cycled = 1; | |
1422 | else | |
1423 | cycled = 0; | |
811d736f DH |
1424 | end = -1; |
1425 | } else { | |
1426 | index = wbc->range_start >> PAGE_CACHE_SHIFT; | |
1427 | end = wbc->range_end >> PAGE_CACHE_SHIFT; | |
1428 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) | |
1429 | range_whole = 1; | |
31a12666 | 1430 | cycled = 1; /* ignore range_cyclic tests */ |
811d736f | 1431 | } |
6e6938b6 | 1432 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae JK |
1433 | tag = PAGECACHE_TAG_TOWRITE; |
1434 | else | |
1435 | tag = PAGECACHE_TAG_DIRTY; | |
811d736f | 1436 | retry: |
6e6938b6 | 1437 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae | 1438 | tag_pages_for_writeback(mapping, index, end); |
bd19e012 | 1439 | done_index = index; |
5a3d5c98 NP |
1440 | while (!done && (index <= end)) { |
1441 | int i; | |
1442 | ||
f446daae | 1443 | nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag, |
5a3d5c98 NP |
1444 | min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1); |
1445 | if (nr_pages == 0) | |
1446 | break; | |
811d736f | 1447 | |
811d736f DH |
1448 | for (i = 0; i < nr_pages; i++) { |
1449 | struct page *page = pvec.pages[i]; | |
1450 | ||
1451 | /* | |
d5482cdf NP |
1452 | * At this point, the page may be truncated or |
1453 | * invalidated (changing page->mapping to NULL), or | |
1454 | * even swizzled back from swapper_space to tmpfs file | |
1455 | * mapping. However, page->index will not change | |
1456 | * because we have a reference on the page. | |
811d736f | 1457 | */ |
d5482cdf NP |
1458 | if (page->index > end) { |
1459 | /* | |
1460 | * can't be range_cyclic (1st pass) because | |
1461 | * end == -1 in that case. | |
1462 | */ | |
1463 | done = 1; | |
1464 | break; | |
1465 | } | |
1466 | ||
cf15b07c | 1467 | done_index = page->index; |
d5482cdf | 1468 | |
811d736f DH |
1469 | lock_page(page); |
1470 | ||
5a3d5c98 NP |
1471 | /* |
1472 | * Page truncated or invalidated. We can freely skip it | |
1473 | * then, even for data integrity operations: the page | |
1474 | * has disappeared concurrently, so there could be no | |
1475 | * real expectation of this data interity operation | |
1476 | * even if there is now a new, dirty page at the same | |
1477 | * pagecache address. | |
1478 | */ | |
811d736f | 1479 | if (unlikely(page->mapping != mapping)) { |
5a3d5c98 | 1480 | continue_unlock: |
811d736f DH |
1481 | unlock_page(page); |
1482 | continue; | |
1483 | } | |
1484 | ||
515f4a03 NP |
1485 | if (!PageDirty(page)) { |
1486 | /* someone wrote it for us */ | |
1487 | goto continue_unlock; | |
1488 | } | |
1489 | ||
1490 | if (PageWriteback(page)) { | |
1491 | if (wbc->sync_mode != WB_SYNC_NONE) | |
1492 | wait_on_page_writeback(page); | |
1493 | else | |
1494 | goto continue_unlock; | |
1495 | } | |
811d736f | 1496 | |
515f4a03 NP |
1497 | BUG_ON(PageWriteback(page)); |
1498 | if (!clear_page_dirty_for_io(page)) | |
5a3d5c98 | 1499 | goto continue_unlock; |
811d736f | 1500 | |
9e094383 | 1501 | trace_wbc_writepage(wbc, mapping->backing_dev_info); |
0ea97180 | 1502 | ret = (*writepage)(page, wbc, data); |
00266770 NP |
1503 | if (unlikely(ret)) { |
1504 | if (ret == AOP_WRITEPAGE_ACTIVATE) { | |
1505 | unlock_page(page); | |
1506 | ret = 0; | |
1507 | } else { | |
1508 | /* | |
1509 | * done_index is set past this page, | |
1510 | * so media errors will not choke | |
1511 | * background writeout for the entire | |
1512 | * file. This has consequences for | |
1513 | * range_cyclic semantics (ie. it may | |
1514 | * not be suitable for data integrity | |
1515 | * writeout). | |
1516 | */ | |
cf15b07c | 1517 | done_index = page->index + 1; |
00266770 NP |
1518 | done = 1; |
1519 | break; | |
1520 | } | |
0b564927 | 1521 | } |
00266770 | 1522 | |
546a1924 DC |
1523 | /* |
1524 | * We stop writing back only if we are not doing | |
1525 | * integrity sync. In case of integrity sync we have to | |
1526 | * keep going until we have written all the pages | |
1527 | * we tagged for writeback prior to entering this loop. | |
1528 | */ | |
1529 | if (--wbc->nr_to_write <= 0 && | |
1530 | wbc->sync_mode == WB_SYNC_NONE) { | |
1531 | done = 1; | |
1532 | break; | |
05fe478d | 1533 | } |
811d736f DH |
1534 | } |
1535 | pagevec_release(&pvec); | |
1536 | cond_resched(); | |
1537 | } | |
3a4c6800 | 1538 | if (!cycled && !done) { |
811d736f | 1539 | /* |
31a12666 | 1540 | * range_cyclic: |
811d736f DH |
1541 | * We hit the last page and there is more work to be done: wrap |
1542 | * back to the start of the file | |
1543 | */ | |
31a12666 | 1544 | cycled = 1; |
811d736f | 1545 | index = 0; |
31a12666 | 1546 | end = writeback_index - 1; |
811d736f DH |
1547 | goto retry; |
1548 | } | |
0b564927 DC |
1549 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) |
1550 | mapping->writeback_index = done_index; | |
06d6cf69 | 1551 | |
811d736f DH |
1552 | return ret; |
1553 | } | |
0ea97180 MS |
1554 | EXPORT_SYMBOL(write_cache_pages); |
1555 | ||
1556 | /* | |
1557 | * Function used by generic_writepages to call the real writepage | |
1558 | * function and set the mapping flags on error | |
1559 | */ | |
1560 | static int __writepage(struct page *page, struct writeback_control *wbc, | |
1561 | void *data) | |
1562 | { | |
1563 | struct address_space *mapping = data; | |
1564 | int ret = mapping->a_ops->writepage(page, wbc); | |
1565 | mapping_set_error(mapping, ret); | |
1566 | return ret; | |
1567 | } | |
1568 | ||
1569 | /** | |
1570 | * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. | |
1571 | * @mapping: address space structure to write | |
1572 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
1573 | * | |
1574 | * This is a library function, which implements the writepages() | |
1575 | * address_space_operation. | |
1576 | */ | |
1577 | int generic_writepages(struct address_space *mapping, | |
1578 | struct writeback_control *wbc) | |
1579 | { | |
9b6096a6 SL |
1580 | struct blk_plug plug; |
1581 | int ret; | |
1582 | ||
0ea97180 MS |
1583 | /* deal with chardevs and other special file */ |
1584 | if (!mapping->a_ops->writepage) | |
1585 | return 0; | |
1586 | ||
9b6096a6 SL |
1587 | blk_start_plug(&plug); |
1588 | ret = write_cache_pages(mapping, wbc, __writepage, mapping); | |
1589 | blk_finish_plug(&plug); | |
1590 | return ret; | |
0ea97180 | 1591 | } |
811d736f DH |
1592 | |
1593 | EXPORT_SYMBOL(generic_writepages); | |
1594 | ||
1da177e4 LT |
1595 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
1596 | { | |
22905f77 AM |
1597 | int ret; |
1598 | ||
1da177e4 LT |
1599 | if (wbc->nr_to_write <= 0) |
1600 | return 0; | |
1601 | if (mapping->a_ops->writepages) | |
d08b3851 | 1602 | ret = mapping->a_ops->writepages(mapping, wbc); |
22905f77 AM |
1603 | else |
1604 | ret = generic_writepages(mapping, wbc); | |
22905f77 | 1605 | return ret; |
1da177e4 LT |
1606 | } |
1607 | ||
1608 | /** | |
1609 | * write_one_page - write out a single page and optionally wait on I/O | |
67be2dd1 MW |
1610 | * @page: the page to write |
1611 | * @wait: if true, wait on writeout | |
1da177e4 LT |
1612 | * |
1613 | * The page must be locked by the caller and will be unlocked upon return. | |
1614 | * | |
1615 | * write_one_page() returns a negative error code if I/O failed. | |
1616 | */ | |
1617 | int write_one_page(struct page *page, int wait) | |
1618 | { | |
1619 | struct address_space *mapping = page->mapping; | |
1620 | int ret = 0; | |
1621 | struct writeback_control wbc = { | |
1622 | .sync_mode = WB_SYNC_ALL, | |
1623 | .nr_to_write = 1, | |
1624 | }; | |
1625 | ||
1626 | BUG_ON(!PageLocked(page)); | |
1627 | ||
1628 | if (wait) | |
1629 | wait_on_page_writeback(page); | |
1630 | ||
1631 | if (clear_page_dirty_for_io(page)) { | |
1632 | page_cache_get(page); | |
1633 | ret = mapping->a_ops->writepage(page, &wbc); | |
1634 | if (ret == 0 && wait) { | |
1635 | wait_on_page_writeback(page); | |
1636 | if (PageError(page)) | |
1637 | ret = -EIO; | |
1638 | } | |
1639 | page_cache_release(page); | |
1640 | } else { | |
1641 | unlock_page(page); | |
1642 | } | |
1643 | return ret; | |
1644 | } | |
1645 | EXPORT_SYMBOL(write_one_page); | |
1646 | ||
76719325 KC |
1647 | /* |
1648 | * For address_spaces which do not use buffers nor write back. | |
1649 | */ | |
1650 | int __set_page_dirty_no_writeback(struct page *page) | |
1651 | { | |
1652 | if (!PageDirty(page)) | |
c3f0da63 | 1653 | return !TestSetPageDirty(page); |
76719325 KC |
1654 | return 0; |
1655 | } | |
1656 | ||
e3a7cca1 ES |
1657 | /* |
1658 | * Helper function for set_page_dirty family. | |
1659 | * NOTE: This relies on being atomic wrt interrupts. | |
1660 | */ | |
1661 | void account_page_dirtied(struct page *page, struct address_space *mapping) | |
1662 | { | |
1663 | if (mapping_cap_account_dirty(mapping)) { | |
1664 | __inc_zone_page_state(page, NR_FILE_DIRTY); | |
ea941f0e | 1665 | __inc_zone_page_state(page, NR_DIRTIED); |
e3a7cca1 | 1666 | __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE); |
c8e28ce0 | 1667 | __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED); |
e3a7cca1 ES |
1668 | task_dirty_inc(current); |
1669 | task_io_account_write(PAGE_CACHE_SIZE); | |
1670 | } | |
1671 | } | |
679ceace | 1672 | EXPORT_SYMBOL(account_page_dirtied); |
e3a7cca1 | 1673 | |
f629d1c9 MR |
1674 | /* |
1675 | * Helper function for set_page_writeback family. | |
1676 | * NOTE: Unlike account_page_dirtied this does not rely on being atomic | |
1677 | * wrt interrupts. | |
1678 | */ | |
1679 | void account_page_writeback(struct page *page) | |
1680 | { | |
1681 | inc_zone_page_state(page, NR_WRITEBACK); | |
1682 | } | |
1683 | EXPORT_SYMBOL(account_page_writeback); | |
1684 | ||
1da177e4 LT |
1685 | /* |
1686 | * For address_spaces which do not use buffers. Just tag the page as dirty in | |
1687 | * its radix tree. | |
1688 | * | |
1689 | * This is also used when a single buffer is being dirtied: we want to set the | |
1690 | * page dirty in that case, but not all the buffers. This is a "bottom-up" | |
1691 | * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying. | |
1692 | * | |
1693 | * Most callers have locked the page, which pins the address_space in memory. | |
1694 | * But zap_pte_range() does not lock the page, however in that case the | |
1695 | * mapping is pinned by the vma's ->vm_file reference. | |
1696 | * | |
1697 | * We take care to handle the case where the page was truncated from the | |
183ff22b | 1698 | * mapping by re-checking page_mapping() inside tree_lock. |
1da177e4 LT |
1699 | */ |
1700 | int __set_page_dirty_nobuffers(struct page *page) | |
1701 | { | |
1da177e4 LT |
1702 | if (!TestSetPageDirty(page)) { |
1703 | struct address_space *mapping = page_mapping(page); | |
1704 | struct address_space *mapping2; | |
1705 | ||
8c08540f AM |
1706 | if (!mapping) |
1707 | return 1; | |
1708 | ||
19fd6231 | 1709 | spin_lock_irq(&mapping->tree_lock); |
8c08540f AM |
1710 | mapping2 = page_mapping(page); |
1711 | if (mapping2) { /* Race with truncate? */ | |
1712 | BUG_ON(mapping2 != mapping); | |
787d2214 | 1713 | WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page)); |
e3a7cca1 | 1714 | account_page_dirtied(page, mapping); |
8c08540f AM |
1715 | radix_tree_tag_set(&mapping->page_tree, |
1716 | page_index(page), PAGECACHE_TAG_DIRTY); | |
1717 | } | |
19fd6231 | 1718 | spin_unlock_irq(&mapping->tree_lock); |
8c08540f AM |
1719 | if (mapping->host) { |
1720 | /* !PageAnon && !swapper_space */ | |
1721 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 1722 | } |
4741c9fd | 1723 | return 1; |
1da177e4 | 1724 | } |
4741c9fd | 1725 | return 0; |
1da177e4 LT |
1726 | } |
1727 | EXPORT_SYMBOL(__set_page_dirty_nobuffers); | |
1728 | ||
1729 | /* | |
1730 | * When a writepage implementation decides that it doesn't want to write this | |
1731 | * page for some reason, it should redirty the locked page via | |
1732 | * redirty_page_for_writepage() and it should then unlock the page and return 0 | |
1733 | */ | |
1734 | int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page) | |
1735 | { | |
1736 | wbc->pages_skipped++; | |
1737 | return __set_page_dirty_nobuffers(page); | |
1738 | } | |
1739 | EXPORT_SYMBOL(redirty_page_for_writepage); | |
1740 | ||
1741 | /* | |
6746aff7 WF |
1742 | * Dirty a page. |
1743 | * | |
1744 | * For pages with a mapping this should be done under the page lock | |
1745 | * for the benefit of asynchronous memory errors who prefer a consistent | |
1746 | * dirty state. This rule can be broken in some special cases, | |
1747 | * but should be better not to. | |
1748 | * | |
1da177e4 LT |
1749 | * If the mapping doesn't provide a set_page_dirty a_op, then |
1750 | * just fall through and assume that it wants buffer_heads. | |
1751 | */ | |
1cf6e7d8 | 1752 | int set_page_dirty(struct page *page) |
1da177e4 LT |
1753 | { |
1754 | struct address_space *mapping = page_mapping(page); | |
1755 | ||
1756 | if (likely(mapping)) { | |
1757 | int (*spd)(struct page *) = mapping->a_ops->set_page_dirty; | |
278df9f4 MK |
1758 | /* |
1759 | * readahead/lru_deactivate_page could remain | |
1760 | * PG_readahead/PG_reclaim due to race with end_page_writeback | |
1761 | * About readahead, if the page is written, the flags would be | |
1762 | * reset. So no problem. | |
1763 | * About lru_deactivate_page, if the page is redirty, the flag | |
1764 | * will be reset. So no problem. but if the page is used by readahead | |
1765 | * it will confuse readahead and make it restart the size rampup | |
1766 | * process. But it's a trivial problem. | |
1767 | */ | |
1768 | ClearPageReclaim(page); | |
9361401e DH |
1769 | #ifdef CONFIG_BLOCK |
1770 | if (!spd) | |
1771 | spd = __set_page_dirty_buffers; | |
1772 | #endif | |
1773 | return (*spd)(page); | |
1da177e4 | 1774 | } |
4741c9fd AM |
1775 | if (!PageDirty(page)) { |
1776 | if (!TestSetPageDirty(page)) | |
1777 | return 1; | |
1778 | } | |
1da177e4 LT |
1779 | return 0; |
1780 | } | |
1781 | EXPORT_SYMBOL(set_page_dirty); | |
1782 | ||
1783 | /* | |
1784 | * set_page_dirty() is racy if the caller has no reference against | |
1785 | * page->mapping->host, and if the page is unlocked. This is because another | |
1786 | * CPU could truncate the page off the mapping and then free the mapping. | |
1787 | * | |
1788 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
1789 | * holds a reference on the inode by having an open file. | |
1790 | * | |
1791 | * In other cases, the page should be locked before running set_page_dirty(). | |
1792 | */ | |
1793 | int set_page_dirty_lock(struct page *page) | |
1794 | { | |
1795 | int ret; | |
1796 | ||
7eaceacc | 1797 | lock_page(page); |
1da177e4 LT |
1798 | ret = set_page_dirty(page); |
1799 | unlock_page(page); | |
1800 | return ret; | |
1801 | } | |
1802 | EXPORT_SYMBOL(set_page_dirty_lock); | |
1803 | ||
1da177e4 LT |
1804 | /* |
1805 | * Clear a page's dirty flag, while caring for dirty memory accounting. | |
1806 | * Returns true if the page was previously dirty. | |
1807 | * | |
1808 | * This is for preparing to put the page under writeout. We leave the page | |
1809 | * tagged as dirty in the radix tree so that a concurrent write-for-sync | |
1810 | * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage | |
1811 | * implementation will run either set_page_writeback() or set_page_dirty(), | |
1812 | * at which stage we bring the page's dirty flag and radix-tree dirty tag | |
1813 | * back into sync. | |
1814 | * | |
1815 | * This incoherency between the page's dirty flag and radix-tree tag is | |
1816 | * unfortunate, but it only exists while the page is locked. | |
1817 | */ | |
1818 | int clear_page_dirty_for_io(struct page *page) | |
1819 | { | |
1820 | struct address_space *mapping = page_mapping(page); | |
1821 | ||
79352894 NP |
1822 | BUG_ON(!PageLocked(page)); |
1823 | ||
7658cc28 LT |
1824 | if (mapping && mapping_cap_account_dirty(mapping)) { |
1825 | /* | |
1826 | * Yes, Virginia, this is indeed insane. | |
1827 | * | |
1828 | * We use this sequence to make sure that | |
1829 | * (a) we account for dirty stats properly | |
1830 | * (b) we tell the low-level filesystem to | |
1831 | * mark the whole page dirty if it was | |
1832 | * dirty in a pagetable. Only to then | |
1833 | * (c) clean the page again and return 1 to | |
1834 | * cause the writeback. | |
1835 | * | |
1836 | * This way we avoid all nasty races with the | |
1837 | * dirty bit in multiple places and clearing | |
1838 | * them concurrently from different threads. | |
1839 | * | |
1840 | * Note! Normally the "set_page_dirty(page)" | |
1841 | * has no effect on the actual dirty bit - since | |
1842 | * that will already usually be set. But we | |
1843 | * need the side effects, and it can help us | |
1844 | * avoid races. | |
1845 | * | |
1846 | * We basically use the page "master dirty bit" | |
1847 | * as a serialization point for all the different | |
1848 | * threads doing their things. | |
7658cc28 LT |
1849 | */ |
1850 | if (page_mkclean(page)) | |
1851 | set_page_dirty(page); | |
79352894 NP |
1852 | /* |
1853 | * We carefully synchronise fault handlers against | |
1854 | * installing a dirty pte and marking the page dirty | |
1855 | * at this point. We do this by having them hold the | |
1856 | * page lock at some point after installing their | |
1857 | * pte, but before marking the page dirty. | |
1858 | * Pages are always locked coming in here, so we get | |
1859 | * the desired exclusion. See mm/memory.c:do_wp_page() | |
1860 | * for more comments. | |
1861 | */ | |
7658cc28 | 1862 | if (TestClearPageDirty(page)) { |
8c08540f | 1863 | dec_zone_page_state(page, NR_FILE_DIRTY); |
c9e51e41 PZ |
1864 | dec_bdi_stat(mapping->backing_dev_info, |
1865 | BDI_RECLAIMABLE); | |
7658cc28 | 1866 | return 1; |
1da177e4 | 1867 | } |
7658cc28 | 1868 | return 0; |
1da177e4 | 1869 | } |
7658cc28 | 1870 | return TestClearPageDirty(page); |
1da177e4 | 1871 | } |
58bb01a9 | 1872 | EXPORT_SYMBOL(clear_page_dirty_for_io); |
1da177e4 LT |
1873 | |
1874 | int test_clear_page_writeback(struct page *page) | |
1875 | { | |
1876 | struct address_space *mapping = page_mapping(page); | |
1877 | int ret; | |
1878 | ||
1879 | if (mapping) { | |
69cb51d1 | 1880 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
1da177e4 LT |
1881 | unsigned long flags; |
1882 | ||
19fd6231 | 1883 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 1884 | ret = TestClearPageWriteback(page); |
69cb51d1 | 1885 | if (ret) { |
1da177e4 LT |
1886 | radix_tree_tag_clear(&mapping->page_tree, |
1887 | page_index(page), | |
1888 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 1889 | if (bdi_cap_account_writeback(bdi)) { |
69cb51d1 | 1890 | __dec_bdi_stat(bdi, BDI_WRITEBACK); |
04fbfdc1 PZ |
1891 | __bdi_writeout_inc(bdi); |
1892 | } | |
69cb51d1 | 1893 | } |
19fd6231 | 1894 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
1895 | } else { |
1896 | ret = TestClearPageWriteback(page); | |
1897 | } | |
99b12e3d | 1898 | if (ret) { |
d688abf5 | 1899 | dec_zone_page_state(page, NR_WRITEBACK); |
99b12e3d WF |
1900 | inc_zone_page_state(page, NR_WRITTEN); |
1901 | } | |
1da177e4 LT |
1902 | return ret; |
1903 | } | |
1904 | ||
1905 | int test_set_page_writeback(struct page *page) | |
1906 | { | |
1907 | struct address_space *mapping = page_mapping(page); | |
1908 | int ret; | |
1909 | ||
1910 | if (mapping) { | |
69cb51d1 | 1911 | struct backing_dev_info *bdi = mapping->backing_dev_info; |
1da177e4 LT |
1912 | unsigned long flags; |
1913 | ||
19fd6231 | 1914 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 1915 | ret = TestSetPageWriteback(page); |
69cb51d1 | 1916 | if (!ret) { |
1da177e4 LT |
1917 | radix_tree_tag_set(&mapping->page_tree, |
1918 | page_index(page), | |
1919 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 1920 | if (bdi_cap_account_writeback(bdi)) |
69cb51d1 PZ |
1921 | __inc_bdi_stat(bdi, BDI_WRITEBACK); |
1922 | } | |
1da177e4 LT |
1923 | if (!PageDirty(page)) |
1924 | radix_tree_tag_clear(&mapping->page_tree, | |
1925 | page_index(page), | |
1926 | PAGECACHE_TAG_DIRTY); | |
f446daae JK |
1927 | radix_tree_tag_clear(&mapping->page_tree, |
1928 | page_index(page), | |
1929 | PAGECACHE_TAG_TOWRITE); | |
19fd6231 | 1930 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
1931 | } else { |
1932 | ret = TestSetPageWriteback(page); | |
1933 | } | |
d688abf5 | 1934 | if (!ret) |
f629d1c9 | 1935 | account_page_writeback(page); |
1da177e4 LT |
1936 | return ret; |
1937 | ||
1938 | } | |
1939 | EXPORT_SYMBOL(test_set_page_writeback); | |
1940 | ||
1941 | /* | |
00128188 | 1942 | * Return true if any of the pages in the mapping are marked with the |
1da177e4 LT |
1943 | * passed tag. |
1944 | */ | |
1945 | int mapping_tagged(struct address_space *mapping, int tag) | |
1946 | { | |
72c47832 | 1947 | return radix_tree_tagged(&mapping->page_tree, tag); |
1da177e4 LT |
1948 | } |
1949 | EXPORT_SYMBOL(mapping_tagged); |