Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
f30c2269 | 2 | * mm/page-writeback.c |
1da177e4 LT |
3 | * |
4 | * Copyright (C) 2002, Linus Torvalds. | |
04fbfdc1 | 5 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> |
1da177e4 LT |
6 | * |
7 | * Contains functions related to writing back dirty pages at the | |
8 | * address_space level. | |
9 | * | |
e1f8e874 | 10 | * 10Apr2002 Andrew Morton |
1da177e4 LT |
11 | * Initial version |
12 | */ | |
13 | ||
14 | #include <linux/kernel.h> | |
b95f1b31 | 15 | #include <linux/export.h> |
1da177e4 LT |
16 | #include <linux/spinlock.h> |
17 | #include <linux/fs.h> | |
18 | #include <linux/mm.h> | |
19 | #include <linux/swap.h> | |
20 | #include <linux/slab.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/writeback.h> | |
23 | #include <linux/init.h> | |
24 | #include <linux/backing-dev.h> | |
55e829af | 25 | #include <linux/task_io_accounting_ops.h> |
1da177e4 LT |
26 | #include <linux/blkdev.h> |
27 | #include <linux/mpage.h> | |
d08b3851 | 28 | #include <linux/rmap.h> |
1da177e4 LT |
29 | #include <linux/percpu.h> |
30 | #include <linux/notifier.h> | |
31 | #include <linux/smp.h> | |
32 | #include <linux/sysctl.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/syscalls.h> | |
ff01bb48 | 35 | #include <linux/buffer_head.h> /* __set_page_dirty_buffers */ |
811d736f | 36 | #include <linux/pagevec.h> |
eb608e3a | 37 | #include <linux/timer.h> |
8bd75c77 | 38 | #include <linux/sched/rt.h> |
6e543d57 | 39 | #include <linux/mm_inline.h> |
028c2dd1 | 40 | #include <trace/events/writeback.h> |
1da177e4 | 41 | |
6e543d57 LD |
42 | #include "internal.h" |
43 | ||
ffd1f609 WF |
44 | /* |
45 | * Sleep at most 200ms at a time in balance_dirty_pages(). | |
46 | */ | |
47 | #define MAX_PAUSE max(HZ/5, 1) | |
48 | ||
5b9b3574 WF |
49 | /* |
50 | * Try to keep balance_dirty_pages() call intervals higher than this many pages | |
51 | * by raising pause time to max_pause when falls below it. | |
52 | */ | |
53 | #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10)) | |
54 | ||
e98be2d5 WF |
55 | /* |
56 | * Estimate write bandwidth at 200ms intervals. | |
57 | */ | |
58 | #define BANDWIDTH_INTERVAL max(HZ/5, 1) | |
59 | ||
6c14ae1e WF |
60 | #define RATELIMIT_CALC_SHIFT 10 |
61 | ||
1da177e4 LT |
62 | /* |
63 | * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited | |
64 | * will look to see if it needs to force writeback or throttling. | |
65 | */ | |
66 | static long ratelimit_pages = 32; | |
67 | ||
1da177e4 LT |
68 | /* The following parameters are exported via /proc/sys/vm */ |
69 | ||
70 | /* | |
5b0830cb | 71 | * Start background writeback (via writeback threads) at this percentage |
1da177e4 | 72 | */ |
1b5e62b4 | 73 | int dirty_background_ratio = 10; |
1da177e4 | 74 | |
2da02997 DR |
75 | /* |
76 | * dirty_background_bytes starts at 0 (disabled) so that it is a function of | |
77 | * dirty_background_ratio * the amount of dirtyable memory | |
78 | */ | |
79 | unsigned long dirty_background_bytes; | |
80 | ||
195cf453 BG |
81 | /* |
82 | * free highmem will not be subtracted from the total free memory | |
83 | * for calculating free ratios if vm_highmem_is_dirtyable is true | |
84 | */ | |
85 | int vm_highmem_is_dirtyable; | |
86 | ||
1da177e4 LT |
87 | /* |
88 | * The generator of dirty data starts writeback at this percentage | |
89 | */ | |
1b5e62b4 | 90 | int vm_dirty_ratio = 20; |
1da177e4 | 91 | |
2da02997 DR |
92 | /* |
93 | * vm_dirty_bytes starts at 0 (disabled) so that it is a function of | |
94 | * vm_dirty_ratio * the amount of dirtyable memory | |
95 | */ | |
96 | unsigned long vm_dirty_bytes; | |
97 | ||
1da177e4 | 98 | /* |
704503d8 | 99 | * The interval between `kupdate'-style writebacks |
1da177e4 | 100 | */ |
22ef37ee | 101 | unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */ |
1da177e4 | 102 | |
91913a29 AB |
103 | EXPORT_SYMBOL_GPL(dirty_writeback_interval); |
104 | ||
1da177e4 | 105 | /* |
704503d8 | 106 | * The longest time for which data is allowed to remain dirty |
1da177e4 | 107 | */ |
22ef37ee | 108 | unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */ |
1da177e4 LT |
109 | |
110 | /* | |
111 | * Flag that makes the machine dump writes/reads and block dirtyings. | |
112 | */ | |
113 | int block_dump; | |
114 | ||
115 | /* | |
ed5b43f1 BS |
116 | * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: |
117 | * a full sync is triggered after this time elapses without any disk activity. | |
1da177e4 LT |
118 | */ |
119 | int laptop_mode; | |
120 | ||
121 | EXPORT_SYMBOL(laptop_mode); | |
122 | ||
123 | /* End of sysctl-exported parameters */ | |
124 | ||
dcc25ae7 | 125 | struct wb_domain global_wb_domain; |
eb608e3a | 126 | |
2bc00aef TH |
127 | /* consolidated parameters for balance_dirty_pages() and its subroutines */ |
128 | struct dirty_throttle_control { | |
129 | struct bdi_writeback *wb; | |
e9770b34 | 130 | struct fprop_local_percpu *wb_completions; |
2bc00aef TH |
131 | |
132 | unsigned long dirty; /* file_dirty + write + nfs */ | |
133 | unsigned long thresh; /* dirty threshold */ | |
134 | unsigned long bg_thresh; /* dirty background threshold */ | |
135 | ||
136 | unsigned long wb_dirty; /* per-wb counterparts */ | |
137 | unsigned long wb_thresh; | |
970fb01a | 138 | unsigned long wb_bg_thresh; |
daddfa3c TH |
139 | |
140 | unsigned long pos_ratio; | |
2bc00aef TH |
141 | }; |
142 | ||
e9770b34 TH |
143 | #define GDTC_INIT(__wb) .wb = (__wb), \ |
144 | .wb_completions = &(__wb)->completions | |
2bc00aef | 145 | |
eb608e3a JK |
146 | /* |
147 | * Length of period for aging writeout fractions of bdis. This is an | |
148 | * arbitrarily chosen number. The longer the period, the slower fractions will | |
149 | * reflect changes in current writeout rate. | |
150 | */ | |
151 | #define VM_COMPLETIONS_PERIOD_LEN (3*HZ) | |
04fbfdc1 | 152 | |
693108a8 TH |
153 | #ifdef CONFIG_CGROUP_WRITEBACK |
154 | ||
155 | static void wb_min_max_ratio(struct bdi_writeback *wb, | |
156 | unsigned long *minp, unsigned long *maxp) | |
157 | { | |
158 | unsigned long this_bw = wb->avg_write_bandwidth; | |
159 | unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); | |
160 | unsigned long long min = wb->bdi->min_ratio; | |
161 | unsigned long long max = wb->bdi->max_ratio; | |
162 | ||
163 | /* | |
164 | * @wb may already be clean by the time control reaches here and | |
165 | * the total may not include its bw. | |
166 | */ | |
167 | if (this_bw < tot_bw) { | |
168 | if (min) { | |
169 | min *= this_bw; | |
170 | do_div(min, tot_bw); | |
171 | } | |
172 | if (max < 100) { | |
173 | max *= this_bw; | |
174 | do_div(max, tot_bw); | |
175 | } | |
176 | } | |
177 | ||
178 | *minp = min; | |
179 | *maxp = max; | |
180 | } | |
181 | ||
182 | #else /* CONFIG_CGROUP_WRITEBACK */ | |
183 | ||
184 | static void wb_min_max_ratio(struct bdi_writeback *wb, | |
185 | unsigned long *minp, unsigned long *maxp) | |
186 | { | |
187 | *minp = wb->bdi->min_ratio; | |
188 | *maxp = wb->bdi->max_ratio; | |
189 | } | |
190 | ||
191 | #endif /* CONFIG_CGROUP_WRITEBACK */ | |
192 | ||
a756cf59 JW |
193 | /* |
194 | * In a memory zone, there is a certain amount of pages we consider | |
195 | * available for the page cache, which is essentially the number of | |
196 | * free and reclaimable pages, minus some zone reserves to protect | |
197 | * lowmem and the ability to uphold the zone's watermarks without | |
198 | * requiring writeback. | |
199 | * | |
200 | * This number of dirtyable pages is the base value of which the | |
201 | * user-configurable dirty ratio is the effictive number of pages that | |
202 | * are allowed to be actually dirtied. Per individual zone, or | |
203 | * globally by using the sum of dirtyable pages over all zones. | |
204 | * | |
205 | * Because the user is allowed to specify the dirty limit globally as | |
206 | * absolute number of bytes, calculating the per-zone dirty limit can | |
207 | * require translating the configured limit into a percentage of | |
208 | * global dirtyable memory first. | |
209 | */ | |
210 | ||
a804552b JW |
211 | /** |
212 | * zone_dirtyable_memory - number of dirtyable pages in a zone | |
213 | * @zone: the zone | |
214 | * | |
215 | * Returns the zone's number of pages potentially available for dirty | |
216 | * page cache. This is the base value for the per-zone dirty limits. | |
217 | */ | |
218 | static unsigned long zone_dirtyable_memory(struct zone *zone) | |
219 | { | |
220 | unsigned long nr_pages; | |
221 | ||
222 | nr_pages = zone_page_state(zone, NR_FREE_PAGES); | |
223 | nr_pages -= min(nr_pages, zone->dirty_balance_reserve); | |
224 | ||
a1c3bfb2 JW |
225 | nr_pages += zone_page_state(zone, NR_INACTIVE_FILE); |
226 | nr_pages += zone_page_state(zone, NR_ACTIVE_FILE); | |
a804552b JW |
227 | |
228 | return nr_pages; | |
229 | } | |
230 | ||
1edf2234 JW |
231 | static unsigned long highmem_dirtyable_memory(unsigned long total) |
232 | { | |
233 | #ifdef CONFIG_HIGHMEM | |
234 | int node; | |
235 | unsigned long x = 0; | |
236 | ||
237 | for_each_node_state(node, N_HIGH_MEMORY) { | |
a804552b | 238 | struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM]; |
1edf2234 | 239 | |
a804552b | 240 | x += zone_dirtyable_memory(z); |
1edf2234 | 241 | } |
c8b74c2f SR |
242 | /* |
243 | * Unreclaimable memory (kernel memory or anonymous memory | |
244 | * without swap) can bring down the dirtyable pages below | |
245 | * the zone's dirty balance reserve and the above calculation | |
246 | * will underflow. However we still want to add in nodes | |
247 | * which are below threshold (negative values) to get a more | |
248 | * accurate calculation but make sure that the total never | |
249 | * underflows. | |
250 | */ | |
251 | if ((long)x < 0) | |
252 | x = 0; | |
253 | ||
1edf2234 JW |
254 | /* |
255 | * Make sure that the number of highmem pages is never larger | |
256 | * than the number of the total dirtyable memory. This can only | |
257 | * occur in very strange VM situations but we want to make sure | |
258 | * that this does not occur. | |
259 | */ | |
260 | return min(x, total); | |
261 | #else | |
262 | return 0; | |
263 | #endif | |
264 | } | |
265 | ||
266 | /** | |
ccafa287 | 267 | * global_dirtyable_memory - number of globally dirtyable pages |
1edf2234 | 268 | * |
ccafa287 JW |
269 | * Returns the global number of pages potentially available for dirty |
270 | * page cache. This is the base value for the global dirty limits. | |
1edf2234 | 271 | */ |
18cf8cf8 | 272 | static unsigned long global_dirtyable_memory(void) |
1edf2234 JW |
273 | { |
274 | unsigned long x; | |
275 | ||
a804552b | 276 | x = global_page_state(NR_FREE_PAGES); |
c8b74c2f | 277 | x -= min(x, dirty_balance_reserve); |
1edf2234 | 278 | |
a1c3bfb2 JW |
279 | x += global_page_state(NR_INACTIVE_FILE); |
280 | x += global_page_state(NR_ACTIVE_FILE); | |
a804552b | 281 | |
1edf2234 JW |
282 | if (!vm_highmem_is_dirtyable) |
283 | x -= highmem_dirtyable_memory(x); | |
284 | ||
285 | return x + 1; /* Ensure that we never return 0 */ | |
286 | } | |
287 | ||
ccafa287 JW |
288 | /* |
289 | * global_dirty_limits - background-writeback and dirty-throttling thresholds | |
290 | * | |
291 | * Calculate the dirty thresholds based on sysctl parameters | |
292 | * - vm.dirty_background_ratio or vm.dirty_background_bytes | |
293 | * - vm.dirty_ratio or vm.dirty_bytes | |
294 | * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and | |
295 | * real-time tasks. | |
296 | */ | |
297 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) | |
298 | { | |
9ef0a0ff | 299 | const unsigned long available_memory = global_dirtyable_memory(); |
ccafa287 JW |
300 | unsigned long background; |
301 | unsigned long dirty; | |
ccafa287 JW |
302 | struct task_struct *tsk; |
303 | ||
ccafa287 JW |
304 | if (vm_dirty_bytes) |
305 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE); | |
306 | else | |
307 | dirty = (vm_dirty_ratio * available_memory) / 100; | |
308 | ||
309 | if (dirty_background_bytes) | |
310 | background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE); | |
311 | else | |
312 | background = (dirty_background_ratio * available_memory) / 100; | |
313 | ||
314 | if (background >= dirty) | |
315 | background = dirty / 2; | |
316 | tsk = current; | |
317 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) { | |
318 | background += background / 4; | |
319 | dirty += dirty / 4; | |
320 | } | |
321 | *pbackground = background; | |
322 | *pdirty = dirty; | |
323 | trace_global_dirty_state(background, dirty); | |
324 | } | |
325 | ||
a756cf59 JW |
326 | /** |
327 | * zone_dirty_limit - maximum number of dirty pages allowed in a zone | |
328 | * @zone: the zone | |
329 | * | |
330 | * Returns the maximum number of dirty pages allowed in a zone, based | |
331 | * on the zone's dirtyable memory. | |
332 | */ | |
333 | static unsigned long zone_dirty_limit(struct zone *zone) | |
334 | { | |
335 | unsigned long zone_memory = zone_dirtyable_memory(zone); | |
336 | struct task_struct *tsk = current; | |
337 | unsigned long dirty; | |
338 | ||
339 | if (vm_dirty_bytes) | |
340 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) * | |
341 | zone_memory / global_dirtyable_memory(); | |
342 | else | |
343 | dirty = vm_dirty_ratio * zone_memory / 100; | |
344 | ||
345 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) | |
346 | dirty += dirty / 4; | |
347 | ||
348 | return dirty; | |
349 | } | |
350 | ||
351 | /** | |
352 | * zone_dirty_ok - tells whether a zone is within its dirty limits | |
353 | * @zone: the zone to check | |
354 | * | |
355 | * Returns %true when the dirty pages in @zone are within the zone's | |
356 | * dirty limit, %false if the limit is exceeded. | |
357 | */ | |
358 | bool zone_dirty_ok(struct zone *zone) | |
359 | { | |
360 | unsigned long limit = zone_dirty_limit(zone); | |
361 | ||
362 | return zone_page_state(zone, NR_FILE_DIRTY) + | |
363 | zone_page_state(zone, NR_UNSTABLE_NFS) + | |
364 | zone_page_state(zone, NR_WRITEBACK) <= limit; | |
365 | } | |
366 | ||
2da02997 | 367 | int dirty_background_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 368 | void __user *buffer, size_t *lenp, |
2da02997 DR |
369 | loff_t *ppos) |
370 | { | |
371 | int ret; | |
372 | ||
8d65af78 | 373 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
374 | if (ret == 0 && write) |
375 | dirty_background_bytes = 0; | |
376 | return ret; | |
377 | } | |
378 | ||
379 | int dirty_background_bytes_handler(struct ctl_table *table, int write, | |
8d65af78 | 380 | void __user *buffer, size_t *lenp, |
2da02997 DR |
381 | loff_t *ppos) |
382 | { | |
383 | int ret; | |
384 | ||
8d65af78 | 385 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
386 | if (ret == 0 && write) |
387 | dirty_background_ratio = 0; | |
388 | return ret; | |
389 | } | |
390 | ||
04fbfdc1 | 391 | int dirty_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 392 | void __user *buffer, size_t *lenp, |
04fbfdc1 PZ |
393 | loff_t *ppos) |
394 | { | |
395 | int old_ratio = vm_dirty_ratio; | |
2da02997 DR |
396 | int ret; |
397 | ||
8d65af78 | 398 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
04fbfdc1 | 399 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
eb608e3a | 400 | writeback_set_ratelimit(); |
2da02997 DR |
401 | vm_dirty_bytes = 0; |
402 | } | |
403 | return ret; | |
404 | } | |
405 | ||
2da02997 | 406 | int dirty_bytes_handler(struct ctl_table *table, int write, |
8d65af78 | 407 | void __user *buffer, size_t *lenp, |
2da02997 DR |
408 | loff_t *ppos) |
409 | { | |
fc3501d4 | 410 | unsigned long old_bytes = vm_dirty_bytes; |
2da02997 DR |
411 | int ret; |
412 | ||
8d65af78 | 413 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 | 414 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
eb608e3a | 415 | writeback_set_ratelimit(); |
2da02997 | 416 | vm_dirty_ratio = 0; |
04fbfdc1 PZ |
417 | } |
418 | return ret; | |
419 | } | |
420 | ||
eb608e3a JK |
421 | static unsigned long wp_next_time(unsigned long cur_time) |
422 | { | |
423 | cur_time += VM_COMPLETIONS_PERIOD_LEN; | |
424 | /* 0 has a special meaning... */ | |
425 | if (!cur_time) | |
426 | return 1; | |
427 | return cur_time; | |
428 | } | |
429 | ||
04fbfdc1 | 430 | /* |
380c27ca | 431 | * Increment the wb's writeout completion count and the global writeout |
04fbfdc1 PZ |
432 | * completion count. Called from test_clear_page_writeback(). |
433 | */ | |
93f78d88 | 434 | static inline void __wb_writeout_inc(struct bdi_writeback *wb) |
04fbfdc1 | 435 | { |
380c27ca TH |
436 | struct wb_domain *dom = &global_wb_domain; |
437 | ||
93f78d88 | 438 | __inc_wb_stat(wb, WB_WRITTEN); |
380c27ca | 439 | __fprop_inc_percpu_max(&dom->completions, &wb->completions, |
93f78d88 | 440 | wb->bdi->max_prop_frac); |
eb608e3a | 441 | /* First event after period switching was turned off? */ |
380c27ca | 442 | if (!unlikely(dom->period_time)) { |
eb608e3a JK |
443 | /* |
444 | * We can race with other __bdi_writeout_inc calls here but | |
445 | * it does not cause any harm since the resulting time when | |
446 | * timer will fire and what is in writeout_period_time will be | |
447 | * roughly the same. | |
448 | */ | |
380c27ca TH |
449 | dom->period_time = wp_next_time(jiffies); |
450 | mod_timer(&dom->period_timer, dom->period_time); | |
eb608e3a | 451 | } |
04fbfdc1 PZ |
452 | } |
453 | ||
93f78d88 | 454 | void wb_writeout_inc(struct bdi_writeback *wb) |
dd5656e5 MS |
455 | { |
456 | unsigned long flags; | |
457 | ||
458 | local_irq_save(flags); | |
93f78d88 | 459 | __wb_writeout_inc(wb); |
dd5656e5 MS |
460 | local_irq_restore(flags); |
461 | } | |
93f78d88 | 462 | EXPORT_SYMBOL_GPL(wb_writeout_inc); |
dd5656e5 | 463 | |
eb608e3a JK |
464 | /* |
465 | * On idle system, we can be called long after we scheduled because we use | |
466 | * deferred timers so count with missed periods. | |
467 | */ | |
468 | static void writeout_period(unsigned long t) | |
469 | { | |
380c27ca TH |
470 | struct wb_domain *dom = (void *)t; |
471 | int miss_periods = (jiffies - dom->period_time) / | |
eb608e3a JK |
472 | VM_COMPLETIONS_PERIOD_LEN; |
473 | ||
380c27ca TH |
474 | if (fprop_new_period(&dom->completions, miss_periods + 1)) { |
475 | dom->period_time = wp_next_time(dom->period_time + | |
eb608e3a | 476 | miss_periods * VM_COMPLETIONS_PERIOD_LEN); |
380c27ca | 477 | mod_timer(&dom->period_timer, dom->period_time); |
eb608e3a JK |
478 | } else { |
479 | /* | |
480 | * Aging has zeroed all fractions. Stop wasting CPU on period | |
481 | * updates. | |
482 | */ | |
380c27ca | 483 | dom->period_time = 0; |
eb608e3a JK |
484 | } |
485 | } | |
486 | ||
380c27ca TH |
487 | int wb_domain_init(struct wb_domain *dom, gfp_t gfp) |
488 | { | |
489 | memset(dom, 0, sizeof(*dom)); | |
dcc25ae7 TH |
490 | |
491 | spin_lock_init(&dom->lock); | |
492 | ||
380c27ca TH |
493 | init_timer_deferrable(&dom->period_timer); |
494 | dom->period_timer.function = writeout_period; | |
495 | dom->period_timer.data = (unsigned long)dom; | |
dcc25ae7 TH |
496 | |
497 | dom->dirty_limit_tstamp = jiffies; | |
498 | ||
380c27ca TH |
499 | return fprop_global_init(&dom->completions, gfp); |
500 | } | |
501 | ||
189d3c4a | 502 | /* |
d08c429b JW |
503 | * bdi_min_ratio keeps the sum of the minimum dirty shares of all |
504 | * registered backing devices, which, for obvious reasons, can not | |
505 | * exceed 100%. | |
189d3c4a | 506 | */ |
189d3c4a PZ |
507 | static unsigned int bdi_min_ratio; |
508 | ||
509 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) | |
510 | { | |
511 | int ret = 0; | |
189d3c4a | 512 | |
cfc4ba53 | 513 | spin_lock_bh(&bdi_lock); |
a42dde04 | 514 | if (min_ratio > bdi->max_ratio) { |
189d3c4a | 515 | ret = -EINVAL; |
a42dde04 PZ |
516 | } else { |
517 | min_ratio -= bdi->min_ratio; | |
518 | if (bdi_min_ratio + min_ratio < 100) { | |
519 | bdi_min_ratio += min_ratio; | |
520 | bdi->min_ratio += min_ratio; | |
521 | } else { | |
522 | ret = -EINVAL; | |
523 | } | |
524 | } | |
cfc4ba53 | 525 | spin_unlock_bh(&bdi_lock); |
a42dde04 PZ |
526 | |
527 | return ret; | |
528 | } | |
529 | ||
530 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio) | |
531 | { | |
a42dde04 PZ |
532 | int ret = 0; |
533 | ||
534 | if (max_ratio > 100) | |
535 | return -EINVAL; | |
536 | ||
cfc4ba53 | 537 | spin_lock_bh(&bdi_lock); |
a42dde04 PZ |
538 | if (bdi->min_ratio > max_ratio) { |
539 | ret = -EINVAL; | |
540 | } else { | |
541 | bdi->max_ratio = max_ratio; | |
eb608e3a | 542 | bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100; |
a42dde04 | 543 | } |
cfc4ba53 | 544 | spin_unlock_bh(&bdi_lock); |
189d3c4a PZ |
545 | |
546 | return ret; | |
547 | } | |
a42dde04 | 548 | EXPORT_SYMBOL(bdi_set_max_ratio); |
189d3c4a | 549 | |
6c14ae1e WF |
550 | static unsigned long dirty_freerun_ceiling(unsigned long thresh, |
551 | unsigned long bg_thresh) | |
552 | { | |
553 | return (thresh + bg_thresh) / 2; | |
554 | } | |
555 | ||
ffd1f609 WF |
556 | static unsigned long hard_dirty_limit(unsigned long thresh) |
557 | { | |
dcc25ae7 TH |
558 | struct wb_domain *dom = &global_wb_domain; |
559 | ||
560 | return max(thresh, dom->dirty_limit); | |
ffd1f609 WF |
561 | } |
562 | ||
6f718656 | 563 | /** |
b1cbc6d4 TH |
564 | * __wb_calc_thresh - @wb's share of dirty throttling threshold |
565 | * @dtc: dirty_throttle_context of interest | |
1babe183 | 566 | * |
a88a341a | 567 | * Returns @wb's dirty limit in pages. The term "dirty" in the context of |
6f718656 | 568 | * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages. |
aed21ad2 WF |
569 | * |
570 | * Note that balance_dirty_pages() will only seriously take it as a hard limit | |
571 | * when sleeping max_pause per page is not enough to keep the dirty pages under | |
572 | * control. For example, when the device is completely stalled due to some error | |
573 | * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key. | |
574 | * In the other normal situations, it acts more gently by throttling the tasks | |
a88a341a | 575 | * more (rather than completely block them) when the wb dirty pages go high. |
1babe183 | 576 | * |
6f718656 | 577 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
1babe183 WF |
578 | * - starving fast devices |
579 | * - piling up dirty pages (that will take long time to sync) on slow devices | |
580 | * | |
a88a341a | 581 | * The wb's share of dirty limit will be adapting to its throughput and |
1babe183 WF |
582 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. |
583 | */ | |
b1cbc6d4 | 584 | static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc) |
16c4042f | 585 | { |
380c27ca | 586 | struct wb_domain *dom = &global_wb_domain; |
b1cbc6d4 | 587 | unsigned long thresh = dtc->thresh; |
0d960a38 | 588 | u64 wb_thresh; |
16c4042f | 589 | long numerator, denominator; |
693108a8 | 590 | unsigned long wb_min_ratio, wb_max_ratio; |
04fbfdc1 | 591 | |
16c4042f | 592 | /* |
0d960a38 | 593 | * Calculate this BDI's share of the thresh ratio. |
16c4042f | 594 | */ |
e9770b34 | 595 | fprop_fraction_percpu(&dom->completions, dtc->wb_completions, |
380c27ca | 596 | &numerator, &denominator); |
04fbfdc1 | 597 | |
0d960a38 TH |
598 | wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100; |
599 | wb_thresh *= numerator; | |
600 | do_div(wb_thresh, denominator); | |
04fbfdc1 | 601 | |
b1cbc6d4 | 602 | wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio); |
693108a8 | 603 | |
0d960a38 TH |
604 | wb_thresh += (thresh * wb_min_ratio) / 100; |
605 | if (wb_thresh > (thresh * wb_max_ratio) / 100) | |
606 | wb_thresh = thresh * wb_max_ratio / 100; | |
16c4042f | 607 | |
0d960a38 | 608 | return wb_thresh; |
1da177e4 LT |
609 | } |
610 | ||
b1cbc6d4 TH |
611 | unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh) |
612 | { | |
613 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb), | |
614 | .thresh = thresh }; | |
615 | return __wb_calc_thresh(&gdtc); | |
616 | } | |
617 | ||
5a537485 MP |
618 | /* |
619 | * setpoint - dirty 3 | |
620 | * f(dirty) := 1.0 + (----------------) | |
621 | * limit - setpoint | |
622 | * | |
623 | * it's a 3rd order polynomial that subjects to | |
624 | * | |
625 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast | |
626 | * (2) f(setpoint) = 1.0 => the balance point | |
627 | * (3) f(limit) = 0 => the hard limit | |
628 | * (4) df/dx <= 0 => negative feedback control | |
629 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) | |
630 | * => fast response on large errors; small oscillation near setpoint | |
631 | */ | |
d5c9fde3 | 632 | static long long pos_ratio_polynom(unsigned long setpoint, |
5a537485 MP |
633 | unsigned long dirty, |
634 | unsigned long limit) | |
635 | { | |
636 | long long pos_ratio; | |
637 | long x; | |
638 | ||
d5c9fde3 | 639 | x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT, |
5a537485 MP |
640 | limit - setpoint + 1); |
641 | pos_ratio = x; | |
642 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
643 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
644 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT; | |
645 | ||
646 | return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT); | |
647 | } | |
648 | ||
6c14ae1e WF |
649 | /* |
650 | * Dirty position control. | |
651 | * | |
652 | * (o) global/bdi setpoints | |
653 | * | |
de1fff37 | 654 | * We want the dirty pages be balanced around the global/wb setpoints. |
6c14ae1e WF |
655 | * When the number of dirty pages is higher/lower than the setpoint, the |
656 | * dirty position control ratio (and hence task dirty ratelimit) will be | |
657 | * decreased/increased to bring the dirty pages back to the setpoint. | |
658 | * | |
659 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT | |
660 | * | |
661 | * if (dirty < setpoint) scale up pos_ratio | |
662 | * if (dirty > setpoint) scale down pos_ratio | |
663 | * | |
de1fff37 TH |
664 | * if (wb_dirty < wb_setpoint) scale up pos_ratio |
665 | * if (wb_dirty > wb_setpoint) scale down pos_ratio | |
6c14ae1e WF |
666 | * |
667 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT | |
668 | * | |
669 | * (o) global control line | |
670 | * | |
671 | * ^ pos_ratio | |
672 | * | | |
673 | * | |<===== global dirty control scope ======>| | |
674 | * 2.0 .............* | |
675 | * | .* | |
676 | * | . * | |
677 | * | . * | |
678 | * | . * | |
679 | * | . * | |
680 | * | . * | |
681 | * 1.0 ................................* | |
682 | * | . . * | |
683 | * | . . * | |
684 | * | . . * | |
685 | * | . . * | |
686 | * | . . * | |
687 | * 0 +------------.------------------.----------------------*-------------> | |
688 | * freerun^ setpoint^ limit^ dirty pages | |
689 | * | |
de1fff37 | 690 | * (o) wb control line |
6c14ae1e WF |
691 | * |
692 | * ^ pos_ratio | |
693 | * | | |
694 | * | * | |
695 | * | * | |
696 | * | * | |
697 | * | * | |
698 | * | * |<=========== span ============>| | |
699 | * 1.0 .......................* | |
700 | * | . * | |
701 | * | . * | |
702 | * | . * | |
703 | * | . * | |
704 | * | . * | |
705 | * | . * | |
706 | * | . * | |
707 | * | . * | |
708 | * | . * | |
709 | * | . * | |
710 | * | . * | |
711 | * 1/4 ...............................................* * * * * * * * * * * * | |
712 | * | . . | |
713 | * | . . | |
714 | * | . . | |
715 | * 0 +----------------------.-------------------------------.-------------> | |
de1fff37 | 716 | * wb_setpoint^ x_intercept^ |
6c14ae1e | 717 | * |
de1fff37 | 718 | * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can |
6c14ae1e WF |
719 | * be smoothly throttled down to normal if it starts high in situations like |
720 | * - start writing to a slow SD card and a fast disk at the same time. The SD | |
de1fff37 TH |
721 | * card's wb_dirty may rush to many times higher than wb_setpoint. |
722 | * - the wb dirty thresh drops quickly due to change of JBOD workload | |
6c14ae1e | 723 | */ |
daddfa3c | 724 | static void wb_position_ratio(struct dirty_throttle_control *dtc) |
6c14ae1e | 725 | { |
2bc00aef | 726 | struct bdi_writeback *wb = dtc->wb; |
a88a341a | 727 | unsigned long write_bw = wb->avg_write_bandwidth; |
2bc00aef TH |
728 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); |
729 | unsigned long limit = hard_dirty_limit(dtc->thresh); | |
730 | unsigned long wb_thresh = dtc->wb_thresh; | |
6c14ae1e WF |
731 | unsigned long x_intercept; |
732 | unsigned long setpoint; /* dirty pages' target balance point */ | |
de1fff37 | 733 | unsigned long wb_setpoint; |
6c14ae1e WF |
734 | unsigned long span; |
735 | long long pos_ratio; /* for scaling up/down the rate limit */ | |
736 | long x; | |
737 | ||
daddfa3c TH |
738 | dtc->pos_ratio = 0; |
739 | ||
2bc00aef | 740 | if (unlikely(dtc->dirty >= limit)) |
daddfa3c | 741 | return; |
6c14ae1e WF |
742 | |
743 | /* | |
744 | * global setpoint | |
745 | * | |
5a537485 MP |
746 | * See comment for pos_ratio_polynom(). |
747 | */ | |
748 | setpoint = (freerun + limit) / 2; | |
2bc00aef | 749 | pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit); |
5a537485 MP |
750 | |
751 | /* | |
752 | * The strictlimit feature is a tool preventing mistrusted filesystems | |
753 | * from growing a large number of dirty pages before throttling. For | |
de1fff37 TH |
754 | * such filesystems balance_dirty_pages always checks wb counters |
755 | * against wb limits. Even if global "nr_dirty" is under "freerun". | |
5a537485 MP |
756 | * This is especially important for fuse which sets bdi->max_ratio to |
757 | * 1% by default. Without strictlimit feature, fuse writeback may | |
758 | * consume arbitrary amount of RAM because it is accounted in | |
759 | * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty". | |
6c14ae1e | 760 | * |
a88a341a | 761 | * Here, in wb_position_ratio(), we calculate pos_ratio based on |
de1fff37 | 762 | * two values: wb_dirty and wb_thresh. Let's consider an example: |
5a537485 MP |
763 | * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global |
764 | * limits are set by default to 10% and 20% (background and throttle). | |
de1fff37 | 765 | * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages. |
0d960a38 | 766 | * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is |
de1fff37 | 767 | * about ~6K pages (as the average of background and throttle wb |
5a537485 | 768 | * limits). The 3rd order polynomial will provide positive feedback if |
de1fff37 | 769 | * wb_dirty is under wb_setpoint and vice versa. |
6c14ae1e | 770 | * |
5a537485 | 771 | * Note, that we cannot use global counters in these calculations |
de1fff37 | 772 | * because we want to throttle process writing to a strictlimit wb |
5a537485 MP |
773 | * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB |
774 | * in the example above). | |
6c14ae1e | 775 | */ |
a88a341a | 776 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
de1fff37 | 777 | long long wb_pos_ratio; |
5a537485 | 778 | |
daddfa3c TH |
779 | if (dtc->wb_dirty < 8) { |
780 | dtc->pos_ratio = min_t(long long, pos_ratio * 2, | |
781 | 2 << RATELIMIT_CALC_SHIFT); | |
782 | return; | |
783 | } | |
5a537485 | 784 | |
2bc00aef | 785 | if (dtc->wb_dirty >= wb_thresh) |
daddfa3c | 786 | return; |
5a537485 | 787 | |
970fb01a TH |
788 | wb_setpoint = dirty_freerun_ceiling(wb_thresh, |
789 | dtc->wb_bg_thresh); | |
5a537485 | 790 | |
de1fff37 | 791 | if (wb_setpoint == 0 || wb_setpoint == wb_thresh) |
daddfa3c | 792 | return; |
5a537485 | 793 | |
2bc00aef | 794 | wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty, |
de1fff37 | 795 | wb_thresh); |
5a537485 MP |
796 | |
797 | /* | |
de1fff37 TH |
798 | * Typically, for strictlimit case, wb_setpoint << setpoint |
799 | * and pos_ratio >> wb_pos_ratio. In the other words global | |
5a537485 | 800 | * state ("dirty") is not limiting factor and we have to |
de1fff37 | 801 | * make decision based on wb counters. But there is an |
5a537485 MP |
802 | * important case when global pos_ratio should get precedence: |
803 | * global limits are exceeded (e.g. due to activities on other | |
de1fff37 | 804 | * wb's) while given strictlimit wb is below limit. |
5a537485 | 805 | * |
de1fff37 | 806 | * "pos_ratio * wb_pos_ratio" would work for the case above, |
5a537485 | 807 | * but it would look too non-natural for the case of all |
de1fff37 | 808 | * activity in the system coming from a single strictlimit wb |
5a537485 MP |
809 | * with bdi->max_ratio == 100%. |
810 | * | |
811 | * Note that min() below somewhat changes the dynamics of the | |
812 | * control system. Normally, pos_ratio value can be well over 3 | |
de1fff37 | 813 | * (when globally we are at freerun and wb is well below wb |
5a537485 MP |
814 | * setpoint). Now the maximum pos_ratio in the same situation |
815 | * is 2. We might want to tweak this if we observe the control | |
816 | * system is too slow to adapt. | |
817 | */ | |
daddfa3c TH |
818 | dtc->pos_ratio = min(pos_ratio, wb_pos_ratio); |
819 | return; | |
5a537485 | 820 | } |
6c14ae1e WF |
821 | |
822 | /* | |
823 | * We have computed basic pos_ratio above based on global situation. If | |
de1fff37 | 824 | * the wb is over/under its share of dirty pages, we want to scale |
6c14ae1e WF |
825 | * pos_ratio further down/up. That is done by the following mechanism. |
826 | */ | |
827 | ||
828 | /* | |
de1fff37 | 829 | * wb setpoint |
6c14ae1e | 830 | * |
de1fff37 | 831 | * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint) |
6c14ae1e | 832 | * |
de1fff37 | 833 | * x_intercept - wb_dirty |
6c14ae1e | 834 | * := -------------------------- |
de1fff37 | 835 | * x_intercept - wb_setpoint |
6c14ae1e | 836 | * |
de1fff37 | 837 | * The main wb control line is a linear function that subjects to |
6c14ae1e | 838 | * |
de1fff37 TH |
839 | * (1) f(wb_setpoint) = 1.0 |
840 | * (2) k = - 1 / (8 * write_bw) (in single wb case) | |
841 | * or equally: x_intercept = wb_setpoint + 8 * write_bw | |
6c14ae1e | 842 | * |
de1fff37 | 843 | * For single wb case, the dirty pages are observed to fluctuate |
6c14ae1e | 844 | * regularly within range |
de1fff37 | 845 | * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2] |
6c14ae1e WF |
846 | * for various filesystems, where (2) can yield in a reasonable 12.5% |
847 | * fluctuation range for pos_ratio. | |
848 | * | |
de1fff37 | 849 | * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its |
6c14ae1e | 850 | * own size, so move the slope over accordingly and choose a slope that |
de1fff37 | 851 | * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh. |
6c14ae1e | 852 | */ |
2bc00aef TH |
853 | if (unlikely(wb_thresh > dtc->thresh)) |
854 | wb_thresh = dtc->thresh; | |
aed21ad2 | 855 | /* |
de1fff37 | 856 | * It's very possible that wb_thresh is close to 0 not because the |
aed21ad2 WF |
857 | * device is slow, but that it has remained inactive for long time. |
858 | * Honour such devices a reasonable good (hopefully IO efficient) | |
859 | * threshold, so that the occasional writes won't be blocked and active | |
860 | * writes can rampup the threshold quickly. | |
861 | */ | |
2bc00aef | 862 | wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8); |
6c14ae1e | 863 | /* |
de1fff37 TH |
864 | * scale global setpoint to wb's: |
865 | * wb_setpoint = setpoint * wb_thresh / thresh | |
6c14ae1e | 866 | */ |
2bc00aef | 867 | x = div_u64((u64)wb_thresh << 16, dtc->thresh + 1); |
de1fff37 | 868 | wb_setpoint = setpoint * (u64)x >> 16; |
6c14ae1e | 869 | /* |
de1fff37 TH |
870 | * Use span=(8*write_bw) in single wb case as indicated by |
871 | * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case. | |
6c14ae1e | 872 | * |
de1fff37 TH |
873 | * wb_thresh thresh - wb_thresh |
874 | * span = --------- * (8 * write_bw) + ------------------ * wb_thresh | |
875 | * thresh thresh | |
6c14ae1e | 876 | */ |
2bc00aef | 877 | span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16; |
de1fff37 | 878 | x_intercept = wb_setpoint + span; |
6c14ae1e | 879 | |
2bc00aef TH |
880 | if (dtc->wb_dirty < x_intercept - span / 4) { |
881 | pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty), | |
882 | x_intercept - wb_setpoint + 1); | |
6c14ae1e WF |
883 | } else |
884 | pos_ratio /= 4; | |
885 | ||
8927f66c | 886 | /* |
de1fff37 | 887 | * wb reserve area, safeguard against dirty pool underrun and disk idle |
8927f66c WF |
888 | * It may push the desired control point of global dirty pages higher |
889 | * than setpoint. | |
890 | */ | |
de1fff37 | 891 | x_intercept = wb_thresh / 2; |
2bc00aef TH |
892 | if (dtc->wb_dirty < x_intercept) { |
893 | if (dtc->wb_dirty > x_intercept / 8) | |
894 | pos_ratio = div_u64(pos_ratio * x_intercept, | |
895 | dtc->wb_dirty); | |
50657fc4 | 896 | else |
8927f66c WF |
897 | pos_ratio *= 8; |
898 | } | |
899 | ||
daddfa3c | 900 | dtc->pos_ratio = pos_ratio; |
6c14ae1e WF |
901 | } |
902 | ||
a88a341a TH |
903 | static void wb_update_write_bandwidth(struct bdi_writeback *wb, |
904 | unsigned long elapsed, | |
905 | unsigned long written) | |
e98be2d5 WF |
906 | { |
907 | const unsigned long period = roundup_pow_of_two(3 * HZ); | |
a88a341a TH |
908 | unsigned long avg = wb->avg_write_bandwidth; |
909 | unsigned long old = wb->write_bandwidth; | |
e98be2d5 WF |
910 | u64 bw; |
911 | ||
912 | /* | |
913 | * bw = written * HZ / elapsed | |
914 | * | |
915 | * bw * elapsed + write_bandwidth * (period - elapsed) | |
916 | * write_bandwidth = --------------------------------------------------- | |
917 | * period | |
c72efb65 TH |
918 | * |
919 | * @written may have decreased due to account_page_redirty(). | |
920 | * Avoid underflowing @bw calculation. | |
e98be2d5 | 921 | */ |
a88a341a | 922 | bw = written - min(written, wb->written_stamp); |
e98be2d5 WF |
923 | bw *= HZ; |
924 | if (unlikely(elapsed > period)) { | |
925 | do_div(bw, elapsed); | |
926 | avg = bw; | |
927 | goto out; | |
928 | } | |
a88a341a | 929 | bw += (u64)wb->write_bandwidth * (period - elapsed); |
e98be2d5 WF |
930 | bw >>= ilog2(period); |
931 | ||
932 | /* | |
933 | * one more level of smoothing, for filtering out sudden spikes | |
934 | */ | |
935 | if (avg > old && old >= (unsigned long)bw) | |
936 | avg -= (avg - old) >> 3; | |
937 | ||
938 | if (avg < old && old <= (unsigned long)bw) | |
939 | avg += (old - avg) >> 3; | |
940 | ||
941 | out: | |
95a46c65 TH |
942 | /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */ |
943 | avg = max(avg, 1LU); | |
944 | if (wb_has_dirty_io(wb)) { | |
945 | long delta = avg - wb->avg_write_bandwidth; | |
946 | WARN_ON_ONCE(atomic_long_add_return(delta, | |
947 | &wb->bdi->tot_write_bandwidth) <= 0); | |
948 | } | |
a88a341a TH |
949 | wb->write_bandwidth = bw; |
950 | wb->avg_write_bandwidth = avg; | |
e98be2d5 WF |
951 | } |
952 | ||
2bc00aef | 953 | static void update_dirty_limit(struct dirty_throttle_control *dtc) |
c42843f2 | 954 | { |
dcc25ae7 | 955 | struct wb_domain *dom = &global_wb_domain; |
2bc00aef | 956 | unsigned long thresh = dtc->thresh; |
dcc25ae7 | 957 | unsigned long limit = dom->dirty_limit; |
c42843f2 WF |
958 | |
959 | /* | |
960 | * Follow up in one step. | |
961 | */ | |
962 | if (limit < thresh) { | |
963 | limit = thresh; | |
964 | goto update; | |
965 | } | |
966 | ||
967 | /* | |
968 | * Follow down slowly. Use the higher one as the target, because thresh | |
969 | * may drop below dirty. This is exactly the reason to introduce | |
dcc25ae7 | 970 | * dom->dirty_limit which is guaranteed to lie above the dirty pages. |
c42843f2 | 971 | */ |
2bc00aef | 972 | thresh = max(thresh, dtc->dirty); |
c42843f2 WF |
973 | if (limit > thresh) { |
974 | limit -= (limit - thresh) >> 5; | |
975 | goto update; | |
976 | } | |
977 | return; | |
978 | update: | |
dcc25ae7 | 979 | dom->dirty_limit = limit; |
c42843f2 WF |
980 | } |
981 | ||
2bc00aef | 982 | static void global_update_bandwidth(struct dirty_throttle_control *dtc, |
c42843f2 WF |
983 | unsigned long now) |
984 | { | |
dcc25ae7 | 985 | struct wb_domain *dom = &global_wb_domain; |
c42843f2 WF |
986 | |
987 | /* | |
988 | * check locklessly first to optimize away locking for the most time | |
989 | */ | |
dcc25ae7 | 990 | if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) |
c42843f2 WF |
991 | return; |
992 | ||
dcc25ae7 TH |
993 | spin_lock(&dom->lock); |
994 | if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) { | |
2bc00aef | 995 | update_dirty_limit(dtc); |
dcc25ae7 | 996 | dom->dirty_limit_tstamp = now; |
c42843f2 | 997 | } |
dcc25ae7 | 998 | spin_unlock(&dom->lock); |
c42843f2 WF |
999 | } |
1000 | ||
be3ffa27 | 1001 | /* |
de1fff37 | 1002 | * Maintain wb->dirty_ratelimit, the base dirty throttle rate. |
be3ffa27 | 1003 | * |
de1fff37 | 1004 | * Normal wb tasks will be curbed at or below it in long term. |
be3ffa27 WF |
1005 | * Obviously it should be around (write_bw / N) when there are N dd tasks. |
1006 | */ | |
2bc00aef | 1007 | static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc, |
a88a341a TH |
1008 | unsigned long dirtied, |
1009 | unsigned long elapsed) | |
be3ffa27 | 1010 | { |
2bc00aef TH |
1011 | struct bdi_writeback *wb = dtc->wb; |
1012 | unsigned long dirty = dtc->dirty; | |
1013 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); | |
1014 | unsigned long limit = hard_dirty_limit(dtc->thresh); | |
7381131c | 1015 | unsigned long setpoint = (freerun + limit) / 2; |
a88a341a TH |
1016 | unsigned long write_bw = wb->avg_write_bandwidth; |
1017 | unsigned long dirty_ratelimit = wb->dirty_ratelimit; | |
be3ffa27 WF |
1018 | unsigned long dirty_rate; |
1019 | unsigned long task_ratelimit; | |
1020 | unsigned long balanced_dirty_ratelimit; | |
7381131c WF |
1021 | unsigned long step; |
1022 | unsigned long x; | |
be3ffa27 WF |
1023 | |
1024 | /* | |
1025 | * The dirty rate will match the writeout rate in long term, except | |
1026 | * when dirty pages are truncated by userspace or re-dirtied by FS. | |
1027 | */ | |
a88a341a | 1028 | dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed; |
be3ffa27 | 1029 | |
be3ffa27 WF |
1030 | /* |
1031 | * task_ratelimit reflects each dd's dirty rate for the past 200ms. | |
1032 | */ | |
1033 | task_ratelimit = (u64)dirty_ratelimit * | |
daddfa3c | 1034 | dtc->pos_ratio >> RATELIMIT_CALC_SHIFT; |
be3ffa27 WF |
1035 | task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ |
1036 | ||
1037 | /* | |
1038 | * A linear estimation of the "balanced" throttle rate. The theory is, | |
de1fff37 | 1039 | * if there are N dd tasks, each throttled at task_ratelimit, the wb's |
be3ffa27 WF |
1040 | * dirty_rate will be measured to be (N * task_ratelimit). So the below |
1041 | * formula will yield the balanced rate limit (write_bw / N). | |
1042 | * | |
1043 | * Note that the expanded form is not a pure rate feedback: | |
1044 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) | |
1045 | * but also takes pos_ratio into account: | |
1046 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) | |
1047 | * | |
1048 | * (1) is not realistic because pos_ratio also takes part in balancing | |
1049 | * the dirty rate. Consider the state | |
1050 | * pos_ratio = 0.5 (3) | |
1051 | * rate = 2 * (write_bw / N) (4) | |
1052 | * If (1) is used, it will stuck in that state! Because each dd will | |
1053 | * be throttled at | |
1054 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) | |
1055 | * yielding | |
1056 | * dirty_rate = N * task_ratelimit = write_bw (6) | |
1057 | * put (6) into (1) we get | |
1058 | * rate_(i+1) = rate_(i) (7) | |
1059 | * | |
1060 | * So we end up using (2) to always keep | |
1061 | * rate_(i+1) ~= (write_bw / N) (8) | |
1062 | * regardless of the value of pos_ratio. As long as (8) is satisfied, | |
1063 | * pos_ratio is able to drive itself to 1.0, which is not only where | |
1064 | * the dirty count meet the setpoint, but also where the slope of | |
1065 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated. | |
1066 | */ | |
1067 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, | |
1068 | dirty_rate | 1); | |
bdaac490 WF |
1069 | /* |
1070 | * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw | |
1071 | */ | |
1072 | if (unlikely(balanced_dirty_ratelimit > write_bw)) | |
1073 | balanced_dirty_ratelimit = write_bw; | |
be3ffa27 | 1074 | |
7381131c WF |
1075 | /* |
1076 | * We could safely do this and return immediately: | |
1077 | * | |
de1fff37 | 1078 | * wb->dirty_ratelimit = balanced_dirty_ratelimit; |
7381131c WF |
1079 | * |
1080 | * However to get a more stable dirty_ratelimit, the below elaborated | |
331cbdee | 1081 | * code makes use of task_ratelimit to filter out singular points and |
7381131c WF |
1082 | * limit the step size. |
1083 | * | |
1084 | * The below code essentially only uses the relative value of | |
1085 | * | |
1086 | * task_ratelimit - dirty_ratelimit | |
1087 | * = (pos_ratio - 1) * dirty_ratelimit | |
1088 | * | |
1089 | * which reflects the direction and size of dirty position error. | |
1090 | */ | |
1091 | ||
1092 | /* | |
1093 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff | |
1094 | * task_ratelimit is on the same side of dirty_ratelimit, too. | |
1095 | * For example, when | |
1096 | * - dirty_ratelimit > balanced_dirty_ratelimit | |
1097 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) | |
1098 | * lowering dirty_ratelimit will help meet both the position and rate | |
1099 | * control targets. Otherwise, don't update dirty_ratelimit if it will | |
1100 | * only help meet the rate target. After all, what the users ultimately | |
1101 | * feel and care are stable dirty rate and small position error. | |
1102 | * | |
1103 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size | |
331cbdee | 1104 | * and filter out the singular points of balanced_dirty_ratelimit. Which |
7381131c WF |
1105 | * keeps jumping around randomly and can even leap far away at times |
1106 | * due to the small 200ms estimation period of dirty_rate (we want to | |
1107 | * keep that period small to reduce time lags). | |
1108 | */ | |
1109 | step = 0; | |
5a537485 MP |
1110 | |
1111 | /* | |
de1fff37 | 1112 | * For strictlimit case, calculations above were based on wb counters |
a88a341a | 1113 | * and limits (starting from pos_ratio = wb_position_ratio() and up to |
5a537485 | 1114 | * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate). |
de1fff37 TH |
1115 | * Hence, to calculate "step" properly, we have to use wb_dirty as |
1116 | * "dirty" and wb_setpoint as "setpoint". | |
5a537485 | 1117 | * |
de1fff37 TH |
1118 | * We rampup dirty_ratelimit forcibly if wb_dirty is low because |
1119 | * it's possible that wb_thresh is close to zero due to inactivity | |
970fb01a | 1120 | * of backing device. |
5a537485 | 1121 | */ |
a88a341a | 1122 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
2bc00aef TH |
1123 | dirty = dtc->wb_dirty; |
1124 | if (dtc->wb_dirty < 8) | |
1125 | setpoint = dtc->wb_dirty + 1; | |
5a537485 | 1126 | else |
970fb01a | 1127 | setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2; |
5a537485 MP |
1128 | } |
1129 | ||
7381131c | 1130 | if (dirty < setpoint) { |
a88a341a | 1131 | x = min3(wb->balanced_dirty_ratelimit, |
7c809968 | 1132 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1133 | if (dirty_ratelimit < x) |
1134 | step = x - dirty_ratelimit; | |
1135 | } else { | |
a88a341a | 1136 | x = max3(wb->balanced_dirty_ratelimit, |
7c809968 | 1137 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1138 | if (dirty_ratelimit > x) |
1139 | step = dirty_ratelimit - x; | |
1140 | } | |
1141 | ||
1142 | /* | |
1143 | * Don't pursue 100% rate matching. It's impossible since the balanced | |
1144 | * rate itself is constantly fluctuating. So decrease the track speed | |
1145 | * when it gets close to the target. Helps eliminate pointless tremors. | |
1146 | */ | |
1147 | step >>= dirty_ratelimit / (2 * step + 1); | |
1148 | /* | |
1149 | * Limit the tracking speed to avoid overshooting. | |
1150 | */ | |
1151 | step = (step + 7) / 8; | |
1152 | ||
1153 | if (dirty_ratelimit < balanced_dirty_ratelimit) | |
1154 | dirty_ratelimit += step; | |
1155 | else | |
1156 | dirty_ratelimit -= step; | |
1157 | ||
a88a341a TH |
1158 | wb->dirty_ratelimit = max(dirty_ratelimit, 1UL); |
1159 | wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit; | |
b48c104d | 1160 | |
a88a341a | 1161 | trace_bdi_dirty_ratelimit(wb->bdi, dirty_rate, task_ratelimit); |
be3ffa27 WF |
1162 | } |
1163 | ||
2bc00aef | 1164 | static void __wb_update_bandwidth(struct dirty_throttle_control *dtc, |
8a731799 TH |
1165 | unsigned long start_time, |
1166 | bool update_ratelimit) | |
e98be2d5 | 1167 | { |
2bc00aef | 1168 | struct bdi_writeback *wb = dtc->wb; |
e98be2d5 | 1169 | unsigned long now = jiffies; |
a88a341a | 1170 | unsigned long elapsed = now - wb->bw_time_stamp; |
be3ffa27 | 1171 | unsigned long dirtied; |
e98be2d5 WF |
1172 | unsigned long written; |
1173 | ||
8a731799 TH |
1174 | lockdep_assert_held(&wb->list_lock); |
1175 | ||
e98be2d5 WF |
1176 | /* |
1177 | * rate-limit, only update once every 200ms. | |
1178 | */ | |
1179 | if (elapsed < BANDWIDTH_INTERVAL) | |
1180 | return; | |
1181 | ||
a88a341a TH |
1182 | dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]); |
1183 | written = percpu_counter_read(&wb->stat[WB_WRITTEN]); | |
e98be2d5 WF |
1184 | |
1185 | /* | |
1186 | * Skip quiet periods when disk bandwidth is under-utilized. | |
1187 | * (at least 1s idle time between two flusher runs) | |
1188 | */ | |
a88a341a | 1189 | if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time)) |
e98be2d5 WF |
1190 | goto snapshot; |
1191 | ||
8a731799 | 1192 | if (update_ratelimit) { |
2bc00aef TH |
1193 | global_update_bandwidth(dtc, now); |
1194 | wb_update_dirty_ratelimit(dtc, dirtied, elapsed); | |
be3ffa27 | 1195 | } |
a88a341a | 1196 | wb_update_write_bandwidth(wb, elapsed, written); |
e98be2d5 WF |
1197 | |
1198 | snapshot: | |
a88a341a TH |
1199 | wb->dirtied_stamp = dirtied; |
1200 | wb->written_stamp = written; | |
1201 | wb->bw_time_stamp = now; | |
e98be2d5 WF |
1202 | } |
1203 | ||
8a731799 | 1204 | void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time) |
e98be2d5 | 1205 | { |
2bc00aef TH |
1206 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
1207 | ||
1208 | __wb_update_bandwidth(&gdtc, start_time, false); | |
e98be2d5 WF |
1209 | } |
1210 | ||
9d823e8f | 1211 | /* |
d0e1d66b | 1212 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited() |
9d823e8f WF |
1213 | * will look to see if it needs to start dirty throttling. |
1214 | * | |
1215 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive | |
1216 | * global_page_state() too often. So scale it near-sqrt to the safety margin | |
1217 | * (the number of pages we may dirty without exceeding the dirty limits). | |
1218 | */ | |
1219 | static unsigned long dirty_poll_interval(unsigned long dirty, | |
1220 | unsigned long thresh) | |
1221 | { | |
1222 | if (thresh > dirty) | |
1223 | return 1UL << (ilog2(thresh - dirty) >> 1); | |
1224 | ||
1225 | return 1; | |
1226 | } | |
1227 | ||
a88a341a | 1228 | static unsigned long wb_max_pause(struct bdi_writeback *wb, |
de1fff37 | 1229 | unsigned long wb_dirty) |
c8462cc9 | 1230 | { |
a88a341a | 1231 | unsigned long bw = wb->avg_write_bandwidth; |
e3b6c655 | 1232 | unsigned long t; |
c8462cc9 | 1233 | |
7ccb9ad5 WF |
1234 | /* |
1235 | * Limit pause time for small memory systems. If sleeping for too long | |
1236 | * time, a small pool of dirty/writeback pages may go empty and disk go | |
1237 | * idle. | |
1238 | * | |
1239 | * 8 serves as the safety ratio. | |
1240 | */ | |
de1fff37 | 1241 | t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8)); |
7ccb9ad5 WF |
1242 | t++; |
1243 | ||
e3b6c655 | 1244 | return min_t(unsigned long, t, MAX_PAUSE); |
7ccb9ad5 WF |
1245 | } |
1246 | ||
a88a341a TH |
1247 | static long wb_min_pause(struct bdi_writeback *wb, |
1248 | long max_pause, | |
1249 | unsigned long task_ratelimit, | |
1250 | unsigned long dirty_ratelimit, | |
1251 | int *nr_dirtied_pause) | |
c8462cc9 | 1252 | { |
a88a341a TH |
1253 | long hi = ilog2(wb->avg_write_bandwidth); |
1254 | long lo = ilog2(wb->dirty_ratelimit); | |
7ccb9ad5 WF |
1255 | long t; /* target pause */ |
1256 | long pause; /* estimated next pause */ | |
1257 | int pages; /* target nr_dirtied_pause */ | |
c8462cc9 | 1258 | |
7ccb9ad5 WF |
1259 | /* target for 10ms pause on 1-dd case */ |
1260 | t = max(1, HZ / 100); | |
c8462cc9 WF |
1261 | |
1262 | /* | |
1263 | * Scale up pause time for concurrent dirtiers in order to reduce CPU | |
1264 | * overheads. | |
1265 | * | |
7ccb9ad5 | 1266 | * (N * 10ms) on 2^N concurrent tasks. |
c8462cc9 WF |
1267 | */ |
1268 | if (hi > lo) | |
7ccb9ad5 | 1269 | t += (hi - lo) * (10 * HZ) / 1024; |
c8462cc9 WF |
1270 | |
1271 | /* | |
7ccb9ad5 WF |
1272 | * This is a bit convoluted. We try to base the next nr_dirtied_pause |
1273 | * on the much more stable dirty_ratelimit. However the next pause time | |
1274 | * will be computed based on task_ratelimit and the two rate limits may | |
1275 | * depart considerably at some time. Especially if task_ratelimit goes | |
1276 | * below dirty_ratelimit/2 and the target pause is max_pause, the next | |
1277 | * pause time will be max_pause*2 _trimmed down_ to max_pause. As a | |
1278 | * result task_ratelimit won't be executed faithfully, which could | |
1279 | * eventually bring down dirty_ratelimit. | |
c8462cc9 | 1280 | * |
7ccb9ad5 WF |
1281 | * We apply two rules to fix it up: |
1282 | * 1) try to estimate the next pause time and if necessary, use a lower | |
1283 | * nr_dirtied_pause so as not to exceed max_pause. When this happens, | |
1284 | * nr_dirtied_pause will be "dancing" with task_ratelimit. | |
1285 | * 2) limit the target pause time to max_pause/2, so that the normal | |
1286 | * small fluctuations of task_ratelimit won't trigger rule (1) and | |
1287 | * nr_dirtied_pause will remain as stable as dirty_ratelimit. | |
c8462cc9 | 1288 | */ |
7ccb9ad5 WF |
1289 | t = min(t, 1 + max_pause / 2); |
1290 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
c8462cc9 WF |
1291 | |
1292 | /* | |
5b9b3574 WF |
1293 | * Tiny nr_dirtied_pause is found to hurt I/O performance in the test |
1294 | * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}. | |
1295 | * When the 16 consecutive reads are often interrupted by some dirty | |
1296 | * throttling pause during the async writes, cfq will go into idles | |
1297 | * (deadline is fine). So push nr_dirtied_pause as high as possible | |
1298 | * until reaches DIRTY_POLL_THRESH=32 pages. | |
c8462cc9 | 1299 | */ |
5b9b3574 WF |
1300 | if (pages < DIRTY_POLL_THRESH) { |
1301 | t = max_pause; | |
1302 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
1303 | if (pages > DIRTY_POLL_THRESH) { | |
1304 | pages = DIRTY_POLL_THRESH; | |
1305 | t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit; | |
1306 | } | |
1307 | } | |
1308 | ||
7ccb9ad5 WF |
1309 | pause = HZ * pages / (task_ratelimit + 1); |
1310 | if (pause > max_pause) { | |
1311 | t = max_pause; | |
1312 | pages = task_ratelimit * t / roundup_pow_of_two(HZ); | |
1313 | } | |
c8462cc9 | 1314 | |
7ccb9ad5 | 1315 | *nr_dirtied_pause = pages; |
c8462cc9 | 1316 | /* |
7ccb9ad5 | 1317 | * The minimal pause time will normally be half the target pause time. |
c8462cc9 | 1318 | */ |
5b9b3574 | 1319 | return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t; |
c8462cc9 WF |
1320 | } |
1321 | ||
970fb01a | 1322 | static inline void wb_dirty_limits(struct dirty_throttle_control *dtc) |
5a537485 | 1323 | { |
2bc00aef | 1324 | struct bdi_writeback *wb = dtc->wb; |
93f78d88 | 1325 | unsigned long wb_reclaimable; |
5a537485 MP |
1326 | |
1327 | /* | |
de1fff37 | 1328 | * wb_thresh is not treated as some limiting factor as |
5a537485 | 1329 | * dirty_thresh, due to reasons |
de1fff37 | 1330 | * - in JBOD setup, wb_thresh can fluctuate a lot |
5a537485 | 1331 | * - in a system with HDD and USB key, the USB key may somehow |
de1fff37 TH |
1332 | * go into state (wb_dirty >> wb_thresh) either because |
1333 | * wb_dirty starts high, or because wb_thresh drops low. | |
5a537485 | 1334 | * In this case we don't want to hard throttle the USB key |
de1fff37 TH |
1335 | * dirtiers for 100 seconds until wb_dirty drops under |
1336 | * wb_thresh. Instead the auxiliary wb control line in | |
a88a341a | 1337 | * wb_position_ratio() will let the dirtier task progress |
de1fff37 | 1338 | * at some rate <= (write_bw / 2) for bringing down wb_dirty. |
5a537485 | 1339 | */ |
b1cbc6d4 | 1340 | dtc->wb_thresh = __wb_calc_thresh(dtc); |
970fb01a TH |
1341 | dtc->wb_bg_thresh = dtc->thresh ? |
1342 | div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0; | |
5a537485 MP |
1343 | |
1344 | /* | |
1345 | * In order to avoid the stacked BDI deadlock we need | |
1346 | * to ensure we accurately count the 'dirty' pages when | |
1347 | * the threshold is low. | |
1348 | * | |
1349 | * Otherwise it would be possible to get thresh+n pages | |
1350 | * reported dirty, even though there are thresh-m pages | |
1351 | * actually dirty; with m+n sitting in the percpu | |
1352 | * deltas. | |
1353 | */ | |
2bc00aef | 1354 | if (dtc->wb_thresh < 2 * wb_stat_error(wb)) { |
93f78d88 | 1355 | wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); |
2bc00aef | 1356 | dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK); |
5a537485 | 1357 | } else { |
93f78d88 | 1358 | wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE); |
2bc00aef | 1359 | dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK); |
5a537485 MP |
1360 | } |
1361 | } | |
1362 | ||
1da177e4 LT |
1363 | /* |
1364 | * balance_dirty_pages() must be called by processes which are generating dirty | |
1365 | * data. It looks at the number of dirty pages in the machine and will force | |
143dfe86 | 1366 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. |
5b0830cb JA |
1367 | * If we're over `background_thresh' then the writeback threads are woken to |
1368 | * perform some writeout. | |
1da177e4 | 1369 | */ |
3a2e9a5a | 1370 | static void balance_dirty_pages(struct address_space *mapping, |
dfb8ae56 | 1371 | struct bdi_writeback *wb, |
143dfe86 | 1372 | unsigned long pages_dirtied) |
1da177e4 | 1373 | { |
2bc00aef TH |
1374 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) }; |
1375 | struct dirty_throttle_control * const gdtc = &gdtc_stor; | |
143dfe86 | 1376 | unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */ |
83712358 | 1377 | long period; |
7ccb9ad5 WF |
1378 | long pause; |
1379 | long max_pause; | |
1380 | long min_pause; | |
1381 | int nr_dirtied_pause; | |
e50e3720 | 1382 | bool dirty_exceeded = false; |
143dfe86 | 1383 | unsigned long task_ratelimit; |
7ccb9ad5 | 1384 | unsigned long dirty_ratelimit; |
dfb8ae56 | 1385 | struct backing_dev_info *bdi = wb->bdi; |
5a537485 | 1386 | bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT; |
e98be2d5 | 1387 | unsigned long start_time = jiffies; |
1da177e4 LT |
1388 | |
1389 | for (;;) { | |
83712358 | 1390 | unsigned long now = jiffies; |
2bc00aef | 1391 | unsigned long dirty, thresh, bg_thresh; |
83712358 | 1392 | |
143dfe86 WF |
1393 | /* |
1394 | * Unstable writes are a feature of certain networked | |
1395 | * filesystems (i.e. NFS) in which data may have been | |
1396 | * written to the server's write cache, but has not yet | |
1397 | * been flushed to permanent storage. | |
1398 | */ | |
5fce25a9 PZ |
1399 | nr_reclaimable = global_page_state(NR_FILE_DIRTY) + |
1400 | global_page_state(NR_UNSTABLE_NFS); | |
2bc00aef | 1401 | gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK); |
5fce25a9 | 1402 | |
2bc00aef | 1403 | global_dirty_limits(&gdtc->bg_thresh, &gdtc->thresh); |
16c4042f | 1404 | |
5a537485 | 1405 | if (unlikely(strictlimit)) { |
970fb01a | 1406 | wb_dirty_limits(gdtc); |
5a537485 | 1407 | |
2bc00aef TH |
1408 | dirty = gdtc->wb_dirty; |
1409 | thresh = gdtc->wb_thresh; | |
970fb01a | 1410 | bg_thresh = gdtc->wb_bg_thresh; |
5a537485 | 1411 | } else { |
2bc00aef TH |
1412 | dirty = gdtc->dirty; |
1413 | thresh = gdtc->thresh; | |
1414 | bg_thresh = gdtc->bg_thresh; | |
5a537485 MP |
1415 | } |
1416 | ||
16c4042f WF |
1417 | /* |
1418 | * Throttle it only when the background writeback cannot | |
1419 | * catch-up. This avoids (excessively) small writeouts | |
de1fff37 | 1420 | * when the wb limits are ramping up in case of !strictlimit. |
5a537485 | 1421 | * |
de1fff37 TH |
1422 | * In strictlimit case make decision based on the wb counters |
1423 | * and limits. Small writeouts when the wb limits are ramping | |
5a537485 | 1424 | * up are the price we consciously pay for strictlimit-ing. |
16c4042f | 1425 | */ |
5a537485 | 1426 | if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) { |
83712358 WF |
1427 | current->dirty_paused_when = now; |
1428 | current->nr_dirtied = 0; | |
7ccb9ad5 | 1429 | current->nr_dirtied_pause = |
5a537485 | 1430 | dirty_poll_interval(dirty, thresh); |
16c4042f | 1431 | break; |
83712358 | 1432 | } |
16c4042f | 1433 | |
bc05873d | 1434 | if (unlikely(!writeback_in_progress(wb))) |
9ecf4866 | 1435 | wb_start_background_writeback(wb); |
143dfe86 | 1436 | |
5a537485 | 1437 | if (!strictlimit) |
970fb01a | 1438 | wb_dirty_limits(gdtc); |
5fce25a9 | 1439 | |
2bc00aef TH |
1440 | dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) && |
1441 | ((gdtc->dirty > gdtc->thresh) || strictlimit); | |
daddfa3c TH |
1442 | |
1443 | wb_position_ratio(gdtc); | |
1444 | ||
a88a341a TH |
1445 | if (dirty_exceeded && !wb->dirty_exceeded) |
1446 | wb->dirty_exceeded = 1; | |
1da177e4 | 1447 | |
8a731799 TH |
1448 | if (time_is_before_jiffies(wb->bw_time_stamp + |
1449 | BANDWIDTH_INTERVAL)) { | |
1450 | spin_lock(&wb->list_lock); | |
2bc00aef | 1451 | __wb_update_bandwidth(gdtc, start_time, true); |
8a731799 TH |
1452 | spin_unlock(&wb->list_lock); |
1453 | } | |
e98be2d5 | 1454 | |
a88a341a | 1455 | dirty_ratelimit = wb->dirty_ratelimit; |
daddfa3c | 1456 | task_ratelimit = ((u64)dirty_ratelimit * gdtc->pos_ratio) >> |
3a73dbbc | 1457 | RATELIMIT_CALC_SHIFT; |
2bc00aef | 1458 | max_pause = wb_max_pause(wb, gdtc->wb_dirty); |
a88a341a TH |
1459 | min_pause = wb_min_pause(wb, max_pause, |
1460 | task_ratelimit, dirty_ratelimit, | |
1461 | &nr_dirtied_pause); | |
7ccb9ad5 | 1462 | |
3a73dbbc | 1463 | if (unlikely(task_ratelimit == 0)) { |
83712358 | 1464 | period = max_pause; |
c8462cc9 | 1465 | pause = max_pause; |
143dfe86 | 1466 | goto pause; |
04fbfdc1 | 1467 | } |
83712358 WF |
1468 | period = HZ * pages_dirtied / task_ratelimit; |
1469 | pause = period; | |
1470 | if (current->dirty_paused_when) | |
1471 | pause -= now - current->dirty_paused_when; | |
1472 | /* | |
1473 | * For less than 1s think time (ext3/4 may block the dirtier | |
1474 | * for up to 800ms from time to time on 1-HDD; so does xfs, | |
1475 | * however at much less frequency), try to compensate it in | |
1476 | * future periods by updating the virtual time; otherwise just | |
1477 | * do a reset, as it may be a light dirtier. | |
1478 | */ | |
7ccb9ad5 | 1479 | if (pause < min_pause) { |
ece13ac3 | 1480 | trace_balance_dirty_pages(bdi, |
2bc00aef TH |
1481 | gdtc->thresh, |
1482 | gdtc->bg_thresh, | |
1483 | gdtc->dirty, | |
1484 | gdtc->wb_thresh, | |
1485 | gdtc->wb_dirty, | |
ece13ac3 WF |
1486 | dirty_ratelimit, |
1487 | task_ratelimit, | |
1488 | pages_dirtied, | |
83712358 | 1489 | period, |
7ccb9ad5 | 1490 | min(pause, 0L), |
ece13ac3 | 1491 | start_time); |
83712358 WF |
1492 | if (pause < -HZ) { |
1493 | current->dirty_paused_when = now; | |
1494 | current->nr_dirtied = 0; | |
1495 | } else if (period) { | |
1496 | current->dirty_paused_when += period; | |
1497 | current->nr_dirtied = 0; | |
7ccb9ad5 WF |
1498 | } else if (current->nr_dirtied_pause <= pages_dirtied) |
1499 | current->nr_dirtied_pause += pages_dirtied; | |
57fc978c | 1500 | break; |
04fbfdc1 | 1501 | } |
7ccb9ad5 WF |
1502 | if (unlikely(pause > max_pause)) { |
1503 | /* for occasional dropped task_ratelimit */ | |
1504 | now += min(pause - max_pause, max_pause); | |
1505 | pause = max_pause; | |
1506 | } | |
143dfe86 WF |
1507 | |
1508 | pause: | |
ece13ac3 | 1509 | trace_balance_dirty_pages(bdi, |
2bc00aef TH |
1510 | gdtc->thresh, |
1511 | gdtc->bg_thresh, | |
1512 | gdtc->dirty, | |
1513 | gdtc->wb_thresh, | |
1514 | gdtc->wb_dirty, | |
ece13ac3 WF |
1515 | dirty_ratelimit, |
1516 | task_ratelimit, | |
1517 | pages_dirtied, | |
83712358 | 1518 | period, |
ece13ac3 WF |
1519 | pause, |
1520 | start_time); | |
499d05ec | 1521 | __set_current_state(TASK_KILLABLE); |
d25105e8 | 1522 | io_schedule_timeout(pause); |
87c6a9b2 | 1523 | |
83712358 WF |
1524 | current->dirty_paused_when = now + pause; |
1525 | current->nr_dirtied = 0; | |
7ccb9ad5 | 1526 | current->nr_dirtied_pause = nr_dirtied_pause; |
83712358 | 1527 | |
ffd1f609 | 1528 | /* |
2bc00aef TH |
1529 | * This is typically equal to (dirty < thresh) and can also |
1530 | * keep "1000+ dd on a slow USB stick" under control. | |
ffd1f609 | 1531 | */ |
1df64719 | 1532 | if (task_ratelimit) |
ffd1f609 | 1533 | break; |
499d05ec | 1534 | |
c5c6343c WF |
1535 | /* |
1536 | * In the case of an unresponding NFS server and the NFS dirty | |
de1fff37 | 1537 | * pages exceeds dirty_thresh, give the other good wb's a pipe |
c5c6343c WF |
1538 | * to go through, so that tasks on them still remain responsive. |
1539 | * | |
1540 | * In theory 1 page is enough to keep the comsumer-producer | |
1541 | * pipe going: the flusher cleans 1 page => the task dirties 1 | |
de1fff37 | 1542 | * more page. However wb_dirty has accounting errors. So use |
93f78d88 | 1543 | * the larger and more IO friendly wb_stat_error. |
c5c6343c | 1544 | */ |
2bc00aef | 1545 | if (gdtc->wb_dirty <= wb_stat_error(wb)) |
c5c6343c WF |
1546 | break; |
1547 | ||
499d05ec JK |
1548 | if (fatal_signal_pending(current)) |
1549 | break; | |
1da177e4 LT |
1550 | } |
1551 | ||
a88a341a TH |
1552 | if (!dirty_exceeded && wb->dirty_exceeded) |
1553 | wb->dirty_exceeded = 0; | |
1da177e4 | 1554 | |
bc05873d | 1555 | if (writeback_in_progress(wb)) |
5b0830cb | 1556 | return; |
1da177e4 LT |
1557 | |
1558 | /* | |
1559 | * In laptop mode, we wait until hitting the higher threshold before | |
1560 | * starting background writeout, and then write out all the way down | |
1561 | * to the lower threshold. So slow writers cause minimal disk activity. | |
1562 | * | |
1563 | * In normal mode, we start background writeout at the lower | |
1564 | * background_thresh, to keep the amount of dirty memory low. | |
1565 | */ | |
143dfe86 WF |
1566 | if (laptop_mode) |
1567 | return; | |
1568 | ||
2bc00aef | 1569 | if (nr_reclaimable > gdtc->bg_thresh) |
9ecf4866 | 1570 | wb_start_background_writeback(wb); |
1da177e4 LT |
1571 | } |
1572 | ||
9d823e8f | 1573 | static DEFINE_PER_CPU(int, bdp_ratelimits); |
245b2e70 | 1574 | |
54848d73 WF |
1575 | /* |
1576 | * Normal tasks are throttled by | |
1577 | * loop { | |
1578 | * dirty tsk->nr_dirtied_pause pages; | |
1579 | * take a snap in balance_dirty_pages(); | |
1580 | * } | |
1581 | * However there is a worst case. If every task exit immediately when dirtied | |
1582 | * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be | |
1583 | * called to throttle the page dirties. The solution is to save the not yet | |
1584 | * throttled page dirties in dirty_throttle_leaks on task exit and charge them | |
1585 | * randomly into the running tasks. This works well for the above worst case, | |
1586 | * as the new task will pick up and accumulate the old task's leaked dirty | |
1587 | * count and eventually get throttled. | |
1588 | */ | |
1589 | DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0; | |
1590 | ||
1da177e4 | 1591 | /** |
d0e1d66b | 1592 | * balance_dirty_pages_ratelimited - balance dirty memory state |
67be2dd1 | 1593 | * @mapping: address_space which was dirtied |
1da177e4 LT |
1594 | * |
1595 | * Processes which are dirtying memory should call in here once for each page | |
1596 | * which was newly dirtied. The function will periodically check the system's | |
1597 | * dirty state and will initiate writeback if needed. | |
1598 | * | |
1599 | * On really big machines, get_writeback_state is expensive, so try to avoid | |
1600 | * calling it too often (ratelimiting). But once we're over the dirty memory | |
1601 | * limit we decrease the ratelimiting by a lot, to prevent individual processes | |
1602 | * from overshooting the limit by (ratelimit_pages) each. | |
1603 | */ | |
d0e1d66b | 1604 | void balance_dirty_pages_ratelimited(struct address_space *mapping) |
1da177e4 | 1605 | { |
dfb8ae56 TH |
1606 | struct inode *inode = mapping->host; |
1607 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1608 | struct bdi_writeback *wb = NULL; | |
9d823e8f WF |
1609 | int ratelimit; |
1610 | int *p; | |
1da177e4 | 1611 | |
36715cef WF |
1612 | if (!bdi_cap_account_dirty(bdi)) |
1613 | return; | |
1614 | ||
dfb8ae56 TH |
1615 | if (inode_cgwb_enabled(inode)) |
1616 | wb = wb_get_create_current(bdi, GFP_KERNEL); | |
1617 | if (!wb) | |
1618 | wb = &bdi->wb; | |
1619 | ||
9d823e8f | 1620 | ratelimit = current->nr_dirtied_pause; |
a88a341a | 1621 | if (wb->dirty_exceeded) |
9d823e8f WF |
1622 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); |
1623 | ||
9d823e8f | 1624 | preempt_disable(); |
1da177e4 | 1625 | /* |
9d823e8f WF |
1626 | * This prevents one CPU to accumulate too many dirtied pages without |
1627 | * calling into balance_dirty_pages(), which can happen when there are | |
1628 | * 1000+ tasks, all of them start dirtying pages at exactly the same | |
1629 | * time, hence all honoured too large initial task->nr_dirtied_pause. | |
1da177e4 | 1630 | */ |
7c8e0181 | 1631 | p = this_cpu_ptr(&bdp_ratelimits); |
9d823e8f | 1632 | if (unlikely(current->nr_dirtied >= ratelimit)) |
fa5a734e | 1633 | *p = 0; |
d3bc1fef WF |
1634 | else if (unlikely(*p >= ratelimit_pages)) { |
1635 | *p = 0; | |
1636 | ratelimit = 0; | |
1da177e4 | 1637 | } |
54848d73 WF |
1638 | /* |
1639 | * Pick up the dirtied pages by the exited tasks. This avoids lots of | |
1640 | * short-lived tasks (eg. gcc invocations in a kernel build) escaping | |
1641 | * the dirty throttling and livelock other long-run dirtiers. | |
1642 | */ | |
7c8e0181 | 1643 | p = this_cpu_ptr(&dirty_throttle_leaks); |
54848d73 | 1644 | if (*p > 0 && current->nr_dirtied < ratelimit) { |
d0e1d66b | 1645 | unsigned long nr_pages_dirtied; |
54848d73 WF |
1646 | nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied); |
1647 | *p -= nr_pages_dirtied; | |
1648 | current->nr_dirtied += nr_pages_dirtied; | |
1da177e4 | 1649 | } |
fa5a734e | 1650 | preempt_enable(); |
9d823e8f WF |
1651 | |
1652 | if (unlikely(current->nr_dirtied >= ratelimit)) | |
dfb8ae56 TH |
1653 | balance_dirty_pages(mapping, wb, current->nr_dirtied); |
1654 | ||
1655 | wb_put(wb); | |
1da177e4 | 1656 | } |
d0e1d66b | 1657 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited); |
1da177e4 | 1658 | |
232ea4d6 | 1659 | void throttle_vm_writeout(gfp_t gfp_mask) |
1da177e4 | 1660 | { |
364aeb28 DR |
1661 | unsigned long background_thresh; |
1662 | unsigned long dirty_thresh; | |
1da177e4 LT |
1663 | |
1664 | for ( ; ; ) { | |
16c4042f | 1665 | global_dirty_limits(&background_thresh, &dirty_thresh); |
47a13333 | 1666 | dirty_thresh = hard_dirty_limit(dirty_thresh); |
1da177e4 LT |
1667 | |
1668 | /* | |
1669 | * Boost the allowable dirty threshold a bit for page | |
1670 | * allocators so they don't get DoS'ed by heavy writers | |
1671 | */ | |
1672 | dirty_thresh += dirty_thresh / 10; /* wheeee... */ | |
1673 | ||
c24f21bd CL |
1674 | if (global_page_state(NR_UNSTABLE_NFS) + |
1675 | global_page_state(NR_WRITEBACK) <= dirty_thresh) | |
1676 | break; | |
8aa7e847 | 1677 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
369f2389 FW |
1678 | |
1679 | /* | |
1680 | * The caller might hold locks which can prevent IO completion | |
1681 | * or progress in the filesystem. So we cannot just sit here | |
1682 | * waiting for IO to complete. | |
1683 | */ | |
1684 | if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) | |
1685 | break; | |
1da177e4 LT |
1686 | } |
1687 | } | |
1688 | ||
1da177e4 LT |
1689 | /* |
1690 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
1691 | */ | |
cccad5b9 | 1692 | int dirty_writeback_centisecs_handler(struct ctl_table *table, int write, |
8d65af78 | 1693 | void __user *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 1694 | { |
8d65af78 | 1695 | proc_dointvec(table, write, buffer, length, ppos); |
1da177e4 LT |
1696 | return 0; |
1697 | } | |
1698 | ||
c2c4986e | 1699 | #ifdef CONFIG_BLOCK |
31373d09 | 1700 | void laptop_mode_timer_fn(unsigned long data) |
1da177e4 | 1701 | { |
31373d09 MG |
1702 | struct request_queue *q = (struct request_queue *)data; |
1703 | int nr_pages = global_page_state(NR_FILE_DIRTY) + | |
1704 | global_page_state(NR_UNSTABLE_NFS); | |
a06fd6b1 TH |
1705 | struct bdi_writeback *wb; |
1706 | struct wb_iter iter; | |
1da177e4 | 1707 | |
31373d09 MG |
1708 | /* |
1709 | * We want to write everything out, not just down to the dirty | |
1710 | * threshold | |
1711 | */ | |
a06fd6b1 TH |
1712 | if (!bdi_has_dirty_io(&q->backing_dev_info)) |
1713 | return; | |
1714 | ||
1715 | bdi_for_each_wb(wb, &q->backing_dev_info, &iter, 0) | |
1716 | if (wb_has_dirty_io(wb)) | |
1717 | wb_start_writeback(wb, nr_pages, true, | |
1718 | WB_REASON_LAPTOP_TIMER); | |
1da177e4 LT |
1719 | } |
1720 | ||
1721 | /* | |
1722 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
1723 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
1724 | * then push it back - the user is still using the disk. | |
1725 | */ | |
31373d09 | 1726 | void laptop_io_completion(struct backing_dev_info *info) |
1da177e4 | 1727 | { |
31373d09 | 1728 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
1729 | } |
1730 | ||
1731 | /* | |
1732 | * We're in laptop mode and we've just synced. The sync's writes will have | |
1733 | * caused another writeback to be scheduled by laptop_io_completion. | |
1734 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
1735 | */ | |
1736 | void laptop_sync_completion(void) | |
1737 | { | |
31373d09 MG |
1738 | struct backing_dev_info *bdi; |
1739 | ||
1740 | rcu_read_lock(); | |
1741 | ||
1742 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) | |
1743 | del_timer(&bdi->laptop_mode_wb_timer); | |
1744 | ||
1745 | rcu_read_unlock(); | |
1da177e4 | 1746 | } |
c2c4986e | 1747 | #endif |
1da177e4 LT |
1748 | |
1749 | /* | |
1750 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
1751 | * if a large number of processes all perform writes at the same time. | |
1752 | * If it is too low then SMP machines will call the (expensive) | |
1753 | * get_writeback_state too often. | |
1754 | * | |
1755 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
1756 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
9d823e8f | 1757 | * thresholds. |
1da177e4 LT |
1758 | */ |
1759 | ||
2d1d43f6 | 1760 | void writeback_set_ratelimit(void) |
1da177e4 | 1761 | { |
dcc25ae7 | 1762 | struct wb_domain *dom = &global_wb_domain; |
9d823e8f WF |
1763 | unsigned long background_thresh; |
1764 | unsigned long dirty_thresh; | |
dcc25ae7 | 1765 | |
9d823e8f | 1766 | global_dirty_limits(&background_thresh, &dirty_thresh); |
dcc25ae7 | 1767 | dom->dirty_limit = dirty_thresh; |
9d823e8f | 1768 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); |
1da177e4 LT |
1769 | if (ratelimit_pages < 16) |
1770 | ratelimit_pages = 16; | |
1da177e4 LT |
1771 | } |
1772 | ||
0db0628d | 1773 | static int |
2f60d628 SB |
1774 | ratelimit_handler(struct notifier_block *self, unsigned long action, |
1775 | void *hcpu) | |
1da177e4 | 1776 | { |
2f60d628 SB |
1777 | |
1778 | switch (action & ~CPU_TASKS_FROZEN) { | |
1779 | case CPU_ONLINE: | |
1780 | case CPU_DEAD: | |
1781 | writeback_set_ratelimit(); | |
1782 | return NOTIFY_OK; | |
1783 | default: | |
1784 | return NOTIFY_DONE; | |
1785 | } | |
1da177e4 LT |
1786 | } |
1787 | ||
0db0628d | 1788 | static struct notifier_block ratelimit_nb = { |
1da177e4 LT |
1789 | .notifier_call = ratelimit_handler, |
1790 | .next = NULL, | |
1791 | }; | |
1792 | ||
1793 | /* | |
dc6e29da LT |
1794 | * Called early on to tune the page writeback dirty limits. |
1795 | * | |
1796 | * We used to scale dirty pages according to how total memory | |
1797 | * related to pages that could be allocated for buffers (by | |
1798 | * comparing nr_free_buffer_pages() to vm_total_pages. | |
1799 | * | |
1800 | * However, that was when we used "dirty_ratio" to scale with | |
1801 | * all memory, and we don't do that any more. "dirty_ratio" | |
1802 | * is now applied to total non-HIGHPAGE memory (by subtracting | |
1803 | * totalhigh_pages from vm_total_pages), and as such we can't | |
1804 | * get into the old insane situation any more where we had | |
1805 | * large amounts of dirty pages compared to a small amount of | |
1806 | * non-HIGHMEM memory. | |
1807 | * | |
1808 | * But we might still want to scale the dirty_ratio by how | |
1809 | * much memory the box has.. | |
1da177e4 LT |
1810 | */ |
1811 | void __init page_writeback_init(void) | |
1812 | { | |
2d1d43f6 | 1813 | writeback_set_ratelimit(); |
1da177e4 | 1814 | register_cpu_notifier(&ratelimit_nb); |
04fbfdc1 | 1815 | |
380c27ca | 1816 | BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL)); |
1da177e4 LT |
1817 | } |
1818 | ||
f446daae JK |
1819 | /** |
1820 | * tag_pages_for_writeback - tag pages to be written by write_cache_pages | |
1821 | * @mapping: address space structure to write | |
1822 | * @start: starting page index | |
1823 | * @end: ending page index (inclusive) | |
1824 | * | |
1825 | * This function scans the page range from @start to @end (inclusive) and tags | |
1826 | * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is | |
1827 | * that write_cache_pages (or whoever calls this function) will then use | |
1828 | * TOWRITE tag to identify pages eligible for writeback. This mechanism is | |
1829 | * used to avoid livelocking of writeback by a process steadily creating new | |
1830 | * dirty pages in the file (thus it is important for this function to be quick | |
1831 | * so that it can tag pages faster than a dirtying process can create them). | |
1832 | */ | |
1833 | /* | |
1834 | * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency. | |
1835 | */ | |
f446daae JK |
1836 | void tag_pages_for_writeback(struct address_space *mapping, |
1837 | pgoff_t start, pgoff_t end) | |
1838 | { | |
3c111a07 | 1839 | #define WRITEBACK_TAG_BATCH 4096 |
f446daae JK |
1840 | unsigned long tagged; |
1841 | ||
1842 | do { | |
1843 | spin_lock_irq(&mapping->tree_lock); | |
1844 | tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree, | |
1845 | &start, end, WRITEBACK_TAG_BATCH, | |
1846 | PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE); | |
1847 | spin_unlock_irq(&mapping->tree_lock); | |
1848 | WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH); | |
1849 | cond_resched(); | |
d5ed3a4a JK |
1850 | /* We check 'start' to handle wrapping when end == ~0UL */ |
1851 | } while (tagged >= WRITEBACK_TAG_BATCH && start); | |
f446daae JK |
1852 | } |
1853 | EXPORT_SYMBOL(tag_pages_for_writeback); | |
1854 | ||
811d736f | 1855 | /** |
0ea97180 | 1856 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. |
811d736f DH |
1857 | * @mapping: address space structure to write |
1858 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
0ea97180 MS |
1859 | * @writepage: function called for each page |
1860 | * @data: data passed to writepage function | |
811d736f | 1861 | * |
0ea97180 | 1862 | * If a page is already under I/O, write_cache_pages() skips it, even |
811d736f DH |
1863 | * if it's dirty. This is desirable behaviour for memory-cleaning writeback, |
1864 | * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() | |
1865 | * and msync() need to guarantee that all the data which was dirty at the time | |
1866 | * the call was made get new I/O started against them. If wbc->sync_mode is | |
1867 | * WB_SYNC_ALL then we were called for data integrity and we must wait for | |
1868 | * existing IO to complete. | |
f446daae JK |
1869 | * |
1870 | * To avoid livelocks (when other process dirties new pages), we first tag | |
1871 | * pages which should be written back with TOWRITE tag and only then start | |
1872 | * writing them. For data-integrity sync we have to be careful so that we do | |
1873 | * not miss some pages (e.g., because some other process has cleared TOWRITE | |
1874 | * tag we set). The rule we follow is that TOWRITE tag can be cleared only | |
1875 | * by the process clearing the DIRTY tag (and submitting the page for IO). | |
811d736f | 1876 | */ |
0ea97180 MS |
1877 | int write_cache_pages(struct address_space *mapping, |
1878 | struct writeback_control *wbc, writepage_t writepage, | |
1879 | void *data) | |
811d736f | 1880 | { |
811d736f DH |
1881 | int ret = 0; |
1882 | int done = 0; | |
811d736f DH |
1883 | struct pagevec pvec; |
1884 | int nr_pages; | |
31a12666 | 1885 | pgoff_t uninitialized_var(writeback_index); |
811d736f DH |
1886 | pgoff_t index; |
1887 | pgoff_t end; /* Inclusive */ | |
bd19e012 | 1888 | pgoff_t done_index; |
31a12666 | 1889 | int cycled; |
811d736f | 1890 | int range_whole = 0; |
f446daae | 1891 | int tag; |
811d736f | 1892 | |
811d736f DH |
1893 | pagevec_init(&pvec, 0); |
1894 | if (wbc->range_cyclic) { | |
31a12666 NP |
1895 | writeback_index = mapping->writeback_index; /* prev offset */ |
1896 | index = writeback_index; | |
1897 | if (index == 0) | |
1898 | cycled = 1; | |
1899 | else | |
1900 | cycled = 0; | |
811d736f DH |
1901 | end = -1; |
1902 | } else { | |
1903 | index = wbc->range_start >> PAGE_CACHE_SHIFT; | |
1904 | end = wbc->range_end >> PAGE_CACHE_SHIFT; | |
1905 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) | |
1906 | range_whole = 1; | |
31a12666 | 1907 | cycled = 1; /* ignore range_cyclic tests */ |
811d736f | 1908 | } |
6e6938b6 | 1909 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae JK |
1910 | tag = PAGECACHE_TAG_TOWRITE; |
1911 | else | |
1912 | tag = PAGECACHE_TAG_DIRTY; | |
811d736f | 1913 | retry: |
6e6938b6 | 1914 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae | 1915 | tag_pages_for_writeback(mapping, index, end); |
bd19e012 | 1916 | done_index = index; |
5a3d5c98 NP |
1917 | while (!done && (index <= end)) { |
1918 | int i; | |
1919 | ||
f446daae | 1920 | nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag, |
5a3d5c98 NP |
1921 | min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1); |
1922 | if (nr_pages == 0) | |
1923 | break; | |
811d736f | 1924 | |
811d736f DH |
1925 | for (i = 0; i < nr_pages; i++) { |
1926 | struct page *page = pvec.pages[i]; | |
1927 | ||
1928 | /* | |
d5482cdf NP |
1929 | * At this point, the page may be truncated or |
1930 | * invalidated (changing page->mapping to NULL), or | |
1931 | * even swizzled back from swapper_space to tmpfs file | |
1932 | * mapping. However, page->index will not change | |
1933 | * because we have a reference on the page. | |
811d736f | 1934 | */ |
d5482cdf NP |
1935 | if (page->index > end) { |
1936 | /* | |
1937 | * can't be range_cyclic (1st pass) because | |
1938 | * end == -1 in that case. | |
1939 | */ | |
1940 | done = 1; | |
1941 | break; | |
1942 | } | |
1943 | ||
cf15b07c | 1944 | done_index = page->index; |
d5482cdf | 1945 | |
811d736f DH |
1946 | lock_page(page); |
1947 | ||
5a3d5c98 NP |
1948 | /* |
1949 | * Page truncated or invalidated. We can freely skip it | |
1950 | * then, even for data integrity operations: the page | |
1951 | * has disappeared concurrently, so there could be no | |
1952 | * real expectation of this data interity operation | |
1953 | * even if there is now a new, dirty page at the same | |
1954 | * pagecache address. | |
1955 | */ | |
811d736f | 1956 | if (unlikely(page->mapping != mapping)) { |
5a3d5c98 | 1957 | continue_unlock: |
811d736f DH |
1958 | unlock_page(page); |
1959 | continue; | |
1960 | } | |
1961 | ||
515f4a03 NP |
1962 | if (!PageDirty(page)) { |
1963 | /* someone wrote it for us */ | |
1964 | goto continue_unlock; | |
1965 | } | |
1966 | ||
1967 | if (PageWriteback(page)) { | |
1968 | if (wbc->sync_mode != WB_SYNC_NONE) | |
1969 | wait_on_page_writeback(page); | |
1970 | else | |
1971 | goto continue_unlock; | |
1972 | } | |
811d736f | 1973 | |
515f4a03 NP |
1974 | BUG_ON(PageWriteback(page)); |
1975 | if (!clear_page_dirty_for_io(page)) | |
5a3d5c98 | 1976 | goto continue_unlock; |
811d736f | 1977 | |
de1414a6 | 1978 | trace_wbc_writepage(wbc, inode_to_bdi(mapping->host)); |
0ea97180 | 1979 | ret = (*writepage)(page, wbc, data); |
00266770 NP |
1980 | if (unlikely(ret)) { |
1981 | if (ret == AOP_WRITEPAGE_ACTIVATE) { | |
1982 | unlock_page(page); | |
1983 | ret = 0; | |
1984 | } else { | |
1985 | /* | |
1986 | * done_index is set past this page, | |
1987 | * so media errors will not choke | |
1988 | * background writeout for the entire | |
1989 | * file. This has consequences for | |
1990 | * range_cyclic semantics (ie. it may | |
1991 | * not be suitable for data integrity | |
1992 | * writeout). | |
1993 | */ | |
cf15b07c | 1994 | done_index = page->index + 1; |
00266770 NP |
1995 | done = 1; |
1996 | break; | |
1997 | } | |
0b564927 | 1998 | } |
00266770 | 1999 | |
546a1924 DC |
2000 | /* |
2001 | * We stop writing back only if we are not doing | |
2002 | * integrity sync. In case of integrity sync we have to | |
2003 | * keep going until we have written all the pages | |
2004 | * we tagged for writeback prior to entering this loop. | |
2005 | */ | |
2006 | if (--wbc->nr_to_write <= 0 && | |
2007 | wbc->sync_mode == WB_SYNC_NONE) { | |
2008 | done = 1; | |
2009 | break; | |
05fe478d | 2010 | } |
811d736f DH |
2011 | } |
2012 | pagevec_release(&pvec); | |
2013 | cond_resched(); | |
2014 | } | |
3a4c6800 | 2015 | if (!cycled && !done) { |
811d736f | 2016 | /* |
31a12666 | 2017 | * range_cyclic: |
811d736f DH |
2018 | * We hit the last page and there is more work to be done: wrap |
2019 | * back to the start of the file | |
2020 | */ | |
31a12666 | 2021 | cycled = 1; |
811d736f | 2022 | index = 0; |
31a12666 | 2023 | end = writeback_index - 1; |
811d736f DH |
2024 | goto retry; |
2025 | } | |
0b564927 DC |
2026 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) |
2027 | mapping->writeback_index = done_index; | |
06d6cf69 | 2028 | |
811d736f DH |
2029 | return ret; |
2030 | } | |
0ea97180 MS |
2031 | EXPORT_SYMBOL(write_cache_pages); |
2032 | ||
2033 | /* | |
2034 | * Function used by generic_writepages to call the real writepage | |
2035 | * function and set the mapping flags on error | |
2036 | */ | |
2037 | static int __writepage(struct page *page, struct writeback_control *wbc, | |
2038 | void *data) | |
2039 | { | |
2040 | struct address_space *mapping = data; | |
2041 | int ret = mapping->a_ops->writepage(page, wbc); | |
2042 | mapping_set_error(mapping, ret); | |
2043 | return ret; | |
2044 | } | |
2045 | ||
2046 | /** | |
2047 | * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. | |
2048 | * @mapping: address space structure to write | |
2049 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
2050 | * | |
2051 | * This is a library function, which implements the writepages() | |
2052 | * address_space_operation. | |
2053 | */ | |
2054 | int generic_writepages(struct address_space *mapping, | |
2055 | struct writeback_control *wbc) | |
2056 | { | |
9b6096a6 SL |
2057 | struct blk_plug plug; |
2058 | int ret; | |
2059 | ||
0ea97180 MS |
2060 | /* deal with chardevs and other special file */ |
2061 | if (!mapping->a_ops->writepage) | |
2062 | return 0; | |
2063 | ||
9b6096a6 SL |
2064 | blk_start_plug(&plug); |
2065 | ret = write_cache_pages(mapping, wbc, __writepage, mapping); | |
2066 | blk_finish_plug(&plug); | |
2067 | return ret; | |
0ea97180 | 2068 | } |
811d736f DH |
2069 | |
2070 | EXPORT_SYMBOL(generic_writepages); | |
2071 | ||
1da177e4 LT |
2072 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
2073 | { | |
22905f77 AM |
2074 | int ret; |
2075 | ||
1da177e4 LT |
2076 | if (wbc->nr_to_write <= 0) |
2077 | return 0; | |
2078 | if (mapping->a_ops->writepages) | |
d08b3851 | 2079 | ret = mapping->a_ops->writepages(mapping, wbc); |
22905f77 AM |
2080 | else |
2081 | ret = generic_writepages(mapping, wbc); | |
22905f77 | 2082 | return ret; |
1da177e4 LT |
2083 | } |
2084 | ||
2085 | /** | |
2086 | * write_one_page - write out a single page and optionally wait on I/O | |
67be2dd1 MW |
2087 | * @page: the page to write |
2088 | * @wait: if true, wait on writeout | |
1da177e4 LT |
2089 | * |
2090 | * The page must be locked by the caller and will be unlocked upon return. | |
2091 | * | |
2092 | * write_one_page() returns a negative error code if I/O failed. | |
2093 | */ | |
2094 | int write_one_page(struct page *page, int wait) | |
2095 | { | |
2096 | struct address_space *mapping = page->mapping; | |
2097 | int ret = 0; | |
2098 | struct writeback_control wbc = { | |
2099 | .sync_mode = WB_SYNC_ALL, | |
2100 | .nr_to_write = 1, | |
2101 | }; | |
2102 | ||
2103 | BUG_ON(!PageLocked(page)); | |
2104 | ||
2105 | if (wait) | |
2106 | wait_on_page_writeback(page); | |
2107 | ||
2108 | if (clear_page_dirty_for_io(page)) { | |
2109 | page_cache_get(page); | |
2110 | ret = mapping->a_ops->writepage(page, &wbc); | |
2111 | if (ret == 0 && wait) { | |
2112 | wait_on_page_writeback(page); | |
2113 | if (PageError(page)) | |
2114 | ret = -EIO; | |
2115 | } | |
2116 | page_cache_release(page); | |
2117 | } else { | |
2118 | unlock_page(page); | |
2119 | } | |
2120 | return ret; | |
2121 | } | |
2122 | EXPORT_SYMBOL(write_one_page); | |
2123 | ||
76719325 KC |
2124 | /* |
2125 | * For address_spaces which do not use buffers nor write back. | |
2126 | */ | |
2127 | int __set_page_dirty_no_writeback(struct page *page) | |
2128 | { | |
2129 | if (!PageDirty(page)) | |
c3f0da63 | 2130 | return !TestSetPageDirty(page); |
76719325 KC |
2131 | return 0; |
2132 | } | |
2133 | ||
e3a7cca1 ES |
2134 | /* |
2135 | * Helper function for set_page_dirty family. | |
c4843a75 GT |
2136 | * |
2137 | * Caller must hold mem_cgroup_begin_page_stat(). | |
2138 | * | |
e3a7cca1 ES |
2139 | * NOTE: This relies on being atomic wrt interrupts. |
2140 | */ | |
c4843a75 GT |
2141 | void account_page_dirtied(struct page *page, struct address_space *mapping, |
2142 | struct mem_cgroup *memcg) | |
e3a7cca1 | 2143 | { |
52ebea74 TH |
2144 | struct inode *inode = mapping->host; |
2145 | ||
9fb0a7da TH |
2146 | trace_writeback_dirty_page(page, mapping); |
2147 | ||
e3a7cca1 | 2148 | if (mapping_cap_account_dirty(mapping)) { |
52ebea74 TH |
2149 | struct bdi_writeback *wb; |
2150 | ||
2151 | inode_attach_wb(inode, page); | |
2152 | wb = inode_to_wb(inode); | |
de1414a6 | 2153 | |
c4843a75 | 2154 | mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_DIRTY); |
e3a7cca1 | 2155 | __inc_zone_page_state(page, NR_FILE_DIRTY); |
ea941f0e | 2156 | __inc_zone_page_state(page, NR_DIRTIED); |
52ebea74 TH |
2157 | __inc_wb_stat(wb, WB_RECLAIMABLE); |
2158 | __inc_wb_stat(wb, WB_DIRTIED); | |
e3a7cca1 | 2159 | task_io_account_write(PAGE_CACHE_SIZE); |
d3bc1fef WF |
2160 | current->nr_dirtied++; |
2161 | this_cpu_inc(bdp_ratelimits); | |
e3a7cca1 ES |
2162 | } |
2163 | } | |
679ceace | 2164 | EXPORT_SYMBOL(account_page_dirtied); |
e3a7cca1 | 2165 | |
b9ea2515 KK |
2166 | /* |
2167 | * Helper function for deaccounting dirty page without writeback. | |
c4843a75 GT |
2168 | * |
2169 | * Caller must hold mem_cgroup_begin_page_stat(). | |
b9ea2515 | 2170 | */ |
c4843a75 GT |
2171 | void account_page_cleaned(struct page *page, struct address_space *mapping, |
2172 | struct mem_cgroup *memcg) | |
b9ea2515 KK |
2173 | { |
2174 | if (mapping_cap_account_dirty(mapping)) { | |
c4843a75 | 2175 | mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY); |
b9ea2515 | 2176 | dec_zone_page_state(page, NR_FILE_DIRTY); |
91018134 | 2177 | dec_wb_stat(inode_to_wb(mapping->host), WB_RECLAIMABLE); |
b9ea2515 KK |
2178 | task_io_account_cancelled_write(PAGE_CACHE_SIZE); |
2179 | } | |
2180 | } | |
b9ea2515 | 2181 | |
1da177e4 LT |
2182 | /* |
2183 | * For address_spaces which do not use buffers. Just tag the page as dirty in | |
2184 | * its radix tree. | |
2185 | * | |
2186 | * This is also used when a single buffer is being dirtied: we want to set the | |
2187 | * page dirty in that case, but not all the buffers. This is a "bottom-up" | |
2188 | * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying. | |
2189 | * | |
2d6d7f98 JW |
2190 | * The caller must ensure this doesn't race with truncation. Most will simply |
2191 | * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and | |
2192 | * the pte lock held, which also locks out truncation. | |
1da177e4 LT |
2193 | */ |
2194 | int __set_page_dirty_nobuffers(struct page *page) | |
2195 | { | |
c4843a75 GT |
2196 | struct mem_cgroup *memcg; |
2197 | ||
2198 | memcg = mem_cgroup_begin_page_stat(page); | |
1da177e4 LT |
2199 | if (!TestSetPageDirty(page)) { |
2200 | struct address_space *mapping = page_mapping(page); | |
a85d9df1 | 2201 | unsigned long flags; |
1da177e4 | 2202 | |
c4843a75 GT |
2203 | if (!mapping) { |
2204 | mem_cgroup_end_page_stat(memcg); | |
8c08540f | 2205 | return 1; |
c4843a75 | 2206 | } |
8c08540f | 2207 | |
a85d9df1 | 2208 | spin_lock_irqsave(&mapping->tree_lock, flags); |
2d6d7f98 JW |
2209 | BUG_ON(page_mapping(page) != mapping); |
2210 | WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page)); | |
c4843a75 | 2211 | account_page_dirtied(page, mapping, memcg); |
2d6d7f98 JW |
2212 | radix_tree_tag_set(&mapping->page_tree, page_index(page), |
2213 | PAGECACHE_TAG_DIRTY); | |
a85d9df1 | 2214 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
c4843a75 GT |
2215 | mem_cgroup_end_page_stat(memcg); |
2216 | ||
8c08540f AM |
2217 | if (mapping->host) { |
2218 | /* !PageAnon && !swapper_space */ | |
2219 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 2220 | } |
4741c9fd | 2221 | return 1; |
1da177e4 | 2222 | } |
c4843a75 | 2223 | mem_cgroup_end_page_stat(memcg); |
4741c9fd | 2224 | return 0; |
1da177e4 LT |
2225 | } |
2226 | EXPORT_SYMBOL(__set_page_dirty_nobuffers); | |
2227 | ||
2f800fbd WF |
2228 | /* |
2229 | * Call this whenever redirtying a page, to de-account the dirty counters | |
2230 | * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written | |
2231 | * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to | |
2232 | * systematic errors in balanced_dirty_ratelimit and the dirty pages position | |
2233 | * control. | |
2234 | */ | |
2235 | void account_page_redirty(struct page *page) | |
2236 | { | |
2237 | struct address_space *mapping = page->mapping; | |
91018134 | 2238 | |
2f800fbd | 2239 | if (mapping && mapping_cap_account_dirty(mapping)) { |
91018134 TH |
2240 | struct bdi_writeback *wb = inode_to_wb(mapping->host); |
2241 | ||
2f800fbd WF |
2242 | current->nr_dirtied--; |
2243 | dec_zone_page_state(page, NR_DIRTIED); | |
91018134 | 2244 | dec_wb_stat(wb, WB_DIRTIED); |
2f800fbd WF |
2245 | } |
2246 | } | |
2247 | EXPORT_SYMBOL(account_page_redirty); | |
2248 | ||
1da177e4 LT |
2249 | /* |
2250 | * When a writepage implementation decides that it doesn't want to write this | |
2251 | * page for some reason, it should redirty the locked page via | |
2252 | * redirty_page_for_writepage() and it should then unlock the page and return 0 | |
2253 | */ | |
2254 | int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page) | |
2255 | { | |
8d38633c KK |
2256 | int ret; |
2257 | ||
1da177e4 | 2258 | wbc->pages_skipped++; |
8d38633c | 2259 | ret = __set_page_dirty_nobuffers(page); |
2f800fbd | 2260 | account_page_redirty(page); |
8d38633c | 2261 | return ret; |
1da177e4 LT |
2262 | } |
2263 | EXPORT_SYMBOL(redirty_page_for_writepage); | |
2264 | ||
2265 | /* | |
6746aff7 WF |
2266 | * Dirty a page. |
2267 | * | |
2268 | * For pages with a mapping this should be done under the page lock | |
2269 | * for the benefit of asynchronous memory errors who prefer a consistent | |
2270 | * dirty state. This rule can be broken in some special cases, | |
2271 | * but should be better not to. | |
2272 | * | |
1da177e4 LT |
2273 | * If the mapping doesn't provide a set_page_dirty a_op, then |
2274 | * just fall through and assume that it wants buffer_heads. | |
2275 | */ | |
1cf6e7d8 | 2276 | int set_page_dirty(struct page *page) |
1da177e4 LT |
2277 | { |
2278 | struct address_space *mapping = page_mapping(page); | |
2279 | ||
2280 | if (likely(mapping)) { | |
2281 | int (*spd)(struct page *) = mapping->a_ops->set_page_dirty; | |
278df9f4 MK |
2282 | /* |
2283 | * readahead/lru_deactivate_page could remain | |
2284 | * PG_readahead/PG_reclaim due to race with end_page_writeback | |
2285 | * About readahead, if the page is written, the flags would be | |
2286 | * reset. So no problem. | |
2287 | * About lru_deactivate_page, if the page is redirty, the flag | |
2288 | * will be reset. So no problem. but if the page is used by readahead | |
2289 | * it will confuse readahead and make it restart the size rampup | |
2290 | * process. But it's a trivial problem. | |
2291 | */ | |
a4bb3ecd NH |
2292 | if (PageReclaim(page)) |
2293 | ClearPageReclaim(page); | |
9361401e DH |
2294 | #ifdef CONFIG_BLOCK |
2295 | if (!spd) | |
2296 | spd = __set_page_dirty_buffers; | |
2297 | #endif | |
2298 | return (*spd)(page); | |
1da177e4 | 2299 | } |
4741c9fd AM |
2300 | if (!PageDirty(page)) { |
2301 | if (!TestSetPageDirty(page)) | |
2302 | return 1; | |
2303 | } | |
1da177e4 LT |
2304 | return 0; |
2305 | } | |
2306 | EXPORT_SYMBOL(set_page_dirty); | |
2307 | ||
2308 | /* | |
2309 | * set_page_dirty() is racy if the caller has no reference against | |
2310 | * page->mapping->host, and if the page is unlocked. This is because another | |
2311 | * CPU could truncate the page off the mapping and then free the mapping. | |
2312 | * | |
2313 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
2314 | * holds a reference on the inode by having an open file. | |
2315 | * | |
2316 | * In other cases, the page should be locked before running set_page_dirty(). | |
2317 | */ | |
2318 | int set_page_dirty_lock(struct page *page) | |
2319 | { | |
2320 | int ret; | |
2321 | ||
7eaceacc | 2322 | lock_page(page); |
1da177e4 LT |
2323 | ret = set_page_dirty(page); |
2324 | unlock_page(page); | |
2325 | return ret; | |
2326 | } | |
2327 | EXPORT_SYMBOL(set_page_dirty_lock); | |
2328 | ||
11f81bec TH |
2329 | /* |
2330 | * This cancels just the dirty bit on the kernel page itself, it does NOT | |
2331 | * actually remove dirty bits on any mmap's that may be around. It also | |
2332 | * leaves the page tagged dirty, so any sync activity will still find it on | |
2333 | * the dirty lists, and in particular, clear_page_dirty_for_io() will still | |
2334 | * look at the dirty bits in the VM. | |
2335 | * | |
2336 | * Doing this should *normally* only ever be done when a page is truncated, | |
2337 | * and is not actually mapped anywhere at all. However, fs/buffer.c does | |
2338 | * this when it notices that somebody has cleaned out all the buffers on a | |
2339 | * page without actually doing it through the VM. Can you say "ext3 is | |
2340 | * horribly ugly"? Thought you could. | |
2341 | */ | |
2342 | void cancel_dirty_page(struct page *page) | |
2343 | { | |
c4843a75 GT |
2344 | struct address_space *mapping = page_mapping(page); |
2345 | ||
2346 | if (mapping_cap_account_dirty(mapping)) { | |
2347 | struct mem_cgroup *memcg; | |
2348 | ||
2349 | memcg = mem_cgroup_begin_page_stat(page); | |
2350 | ||
2351 | if (TestClearPageDirty(page)) | |
2352 | account_page_cleaned(page, mapping, memcg); | |
2353 | ||
2354 | mem_cgroup_end_page_stat(memcg); | |
2355 | } else { | |
2356 | ClearPageDirty(page); | |
2357 | } | |
11f81bec TH |
2358 | } |
2359 | EXPORT_SYMBOL(cancel_dirty_page); | |
2360 | ||
1da177e4 LT |
2361 | /* |
2362 | * Clear a page's dirty flag, while caring for dirty memory accounting. | |
2363 | * Returns true if the page was previously dirty. | |
2364 | * | |
2365 | * This is for preparing to put the page under writeout. We leave the page | |
2366 | * tagged as dirty in the radix tree so that a concurrent write-for-sync | |
2367 | * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage | |
2368 | * implementation will run either set_page_writeback() or set_page_dirty(), | |
2369 | * at which stage we bring the page's dirty flag and radix-tree dirty tag | |
2370 | * back into sync. | |
2371 | * | |
2372 | * This incoherency between the page's dirty flag and radix-tree tag is | |
2373 | * unfortunate, but it only exists while the page is locked. | |
2374 | */ | |
2375 | int clear_page_dirty_for_io(struct page *page) | |
2376 | { | |
2377 | struct address_space *mapping = page_mapping(page); | |
c4843a75 GT |
2378 | struct mem_cgroup *memcg; |
2379 | int ret = 0; | |
1da177e4 | 2380 | |
79352894 NP |
2381 | BUG_ON(!PageLocked(page)); |
2382 | ||
7658cc28 LT |
2383 | if (mapping && mapping_cap_account_dirty(mapping)) { |
2384 | /* | |
2385 | * Yes, Virginia, this is indeed insane. | |
2386 | * | |
2387 | * We use this sequence to make sure that | |
2388 | * (a) we account for dirty stats properly | |
2389 | * (b) we tell the low-level filesystem to | |
2390 | * mark the whole page dirty if it was | |
2391 | * dirty in a pagetable. Only to then | |
2392 | * (c) clean the page again and return 1 to | |
2393 | * cause the writeback. | |
2394 | * | |
2395 | * This way we avoid all nasty races with the | |
2396 | * dirty bit in multiple places and clearing | |
2397 | * them concurrently from different threads. | |
2398 | * | |
2399 | * Note! Normally the "set_page_dirty(page)" | |
2400 | * has no effect on the actual dirty bit - since | |
2401 | * that will already usually be set. But we | |
2402 | * need the side effects, and it can help us | |
2403 | * avoid races. | |
2404 | * | |
2405 | * We basically use the page "master dirty bit" | |
2406 | * as a serialization point for all the different | |
2407 | * threads doing their things. | |
7658cc28 LT |
2408 | */ |
2409 | if (page_mkclean(page)) | |
2410 | set_page_dirty(page); | |
79352894 NP |
2411 | /* |
2412 | * We carefully synchronise fault handlers against | |
2413 | * installing a dirty pte and marking the page dirty | |
2d6d7f98 JW |
2414 | * at this point. We do this by having them hold the |
2415 | * page lock while dirtying the page, and pages are | |
2416 | * always locked coming in here, so we get the desired | |
2417 | * exclusion. | |
79352894 | 2418 | */ |
c4843a75 | 2419 | memcg = mem_cgroup_begin_page_stat(page); |
7658cc28 | 2420 | if (TestClearPageDirty(page)) { |
c4843a75 | 2421 | mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY); |
8c08540f | 2422 | dec_zone_page_state(page, NR_FILE_DIRTY); |
91018134 | 2423 | dec_wb_stat(inode_to_wb(mapping->host), WB_RECLAIMABLE); |
c4843a75 | 2424 | ret = 1; |
1da177e4 | 2425 | } |
c4843a75 GT |
2426 | mem_cgroup_end_page_stat(memcg); |
2427 | return ret; | |
1da177e4 | 2428 | } |
7658cc28 | 2429 | return TestClearPageDirty(page); |
1da177e4 | 2430 | } |
58bb01a9 | 2431 | EXPORT_SYMBOL(clear_page_dirty_for_io); |
1da177e4 LT |
2432 | |
2433 | int test_clear_page_writeback(struct page *page) | |
2434 | { | |
2435 | struct address_space *mapping = page_mapping(page); | |
d7365e78 | 2436 | struct mem_cgroup *memcg; |
d7365e78 | 2437 | int ret; |
1da177e4 | 2438 | |
6de22619 | 2439 | memcg = mem_cgroup_begin_page_stat(page); |
1da177e4 | 2440 | if (mapping) { |
91018134 TH |
2441 | struct inode *inode = mapping->host; |
2442 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 LT |
2443 | unsigned long flags; |
2444 | ||
19fd6231 | 2445 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 2446 | ret = TestClearPageWriteback(page); |
69cb51d1 | 2447 | if (ret) { |
1da177e4 LT |
2448 | radix_tree_tag_clear(&mapping->page_tree, |
2449 | page_index(page), | |
2450 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 2451 | if (bdi_cap_account_writeback(bdi)) { |
91018134 TH |
2452 | struct bdi_writeback *wb = inode_to_wb(inode); |
2453 | ||
2454 | __dec_wb_stat(wb, WB_WRITEBACK); | |
2455 | __wb_writeout_inc(wb); | |
04fbfdc1 | 2456 | } |
69cb51d1 | 2457 | } |
19fd6231 | 2458 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
2459 | } else { |
2460 | ret = TestClearPageWriteback(page); | |
2461 | } | |
99b12e3d | 2462 | if (ret) { |
d7365e78 | 2463 | mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK); |
d688abf5 | 2464 | dec_zone_page_state(page, NR_WRITEBACK); |
99b12e3d WF |
2465 | inc_zone_page_state(page, NR_WRITTEN); |
2466 | } | |
6de22619 | 2467 | mem_cgroup_end_page_stat(memcg); |
1da177e4 LT |
2468 | return ret; |
2469 | } | |
2470 | ||
1c8349a1 | 2471 | int __test_set_page_writeback(struct page *page, bool keep_write) |
1da177e4 LT |
2472 | { |
2473 | struct address_space *mapping = page_mapping(page); | |
d7365e78 | 2474 | struct mem_cgroup *memcg; |
d7365e78 | 2475 | int ret; |
1da177e4 | 2476 | |
6de22619 | 2477 | memcg = mem_cgroup_begin_page_stat(page); |
1da177e4 | 2478 | if (mapping) { |
91018134 TH |
2479 | struct inode *inode = mapping->host; |
2480 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 LT |
2481 | unsigned long flags; |
2482 | ||
19fd6231 | 2483 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 2484 | ret = TestSetPageWriteback(page); |
69cb51d1 | 2485 | if (!ret) { |
1da177e4 LT |
2486 | radix_tree_tag_set(&mapping->page_tree, |
2487 | page_index(page), | |
2488 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 2489 | if (bdi_cap_account_writeback(bdi)) |
91018134 | 2490 | __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK); |
69cb51d1 | 2491 | } |
1da177e4 LT |
2492 | if (!PageDirty(page)) |
2493 | radix_tree_tag_clear(&mapping->page_tree, | |
2494 | page_index(page), | |
2495 | PAGECACHE_TAG_DIRTY); | |
1c8349a1 NJ |
2496 | if (!keep_write) |
2497 | radix_tree_tag_clear(&mapping->page_tree, | |
2498 | page_index(page), | |
2499 | PAGECACHE_TAG_TOWRITE); | |
19fd6231 | 2500 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
2501 | } else { |
2502 | ret = TestSetPageWriteback(page); | |
2503 | } | |
3a3c02ec | 2504 | if (!ret) { |
d7365e78 | 2505 | mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK); |
3a3c02ec JW |
2506 | inc_zone_page_state(page, NR_WRITEBACK); |
2507 | } | |
6de22619 | 2508 | mem_cgroup_end_page_stat(memcg); |
1da177e4 LT |
2509 | return ret; |
2510 | ||
2511 | } | |
1c8349a1 | 2512 | EXPORT_SYMBOL(__test_set_page_writeback); |
1da177e4 LT |
2513 | |
2514 | /* | |
00128188 | 2515 | * Return true if any of the pages in the mapping are marked with the |
1da177e4 LT |
2516 | * passed tag. |
2517 | */ | |
2518 | int mapping_tagged(struct address_space *mapping, int tag) | |
2519 | { | |
72c47832 | 2520 | return radix_tree_tagged(&mapping->page_tree, tag); |
1da177e4 LT |
2521 | } |
2522 | EXPORT_SYMBOL(mapping_tagged); | |
1d1d1a76 DW |
2523 | |
2524 | /** | |
2525 | * wait_for_stable_page() - wait for writeback to finish, if necessary. | |
2526 | * @page: The page to wait on. | |
2527 | * | |
2528 | * This function determines if the given page is related to a backing device | |
2529 | * that requires page contents to be held stable during writeback. If so, then | |
2530 | * it will wait for any pending writeback to complete. | |
2531 | */ | |
2532 | void wait_for_stable_page(struct page *page) | |
2533 | { | |
de1414a6 CH |
2534 | if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host))) |
2535 | wait_on_page_writeback(page); | |
1d1d1a76 DW |
2536 | } |
2537 | EXPORT_SYMBOL_GPL(wait_for_stable_page); |