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