]>
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 { | |
e9f07dfd TH |
129 | #ifdef CONFIG_CGROUP_WRITEBACK |
130 | struct wb_domain *dom; | |
9fc3a43e | 131 | struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */ |
e9f07dfd | 132 | #endif |
2bc00aef | 133 | struct bdi_writeback *wb; |
e9770b34 | 134 | struct fprop_local_percpu *wb_completions; |
2bc00aef | 135 | |
9fc3a43e | 136 | unsigned long avail; /* dirtyable */ |
2bc00aef TH |
137 | unsigned long dirty; /* file_dirty + write + nfs */ |
138 | unsigned long thresh; /* dirty threshold */ | |
139 | unsigned long bg_thresh; /* dirty background threshold */ | |
140 | ||
141 | unsigned long wb_dirty; /* per-wb counterparts */ | |
142 | unsigned long wb_thresh; | |
970fb01a | 143 | unsigned long wb_bg_thresh; |
daddfa3c TH |
144 | |
145 | unsigned long pos_ratio; | |
2bc00aef TH |
146 | }; |
147 | ||
e9f07dfd | 148 | #define DTC_INIT_COMMON(__wb) .wb = (__wb), \ |
e9770b34 | 149 | .wb_completions = &(__wb)->completions |
2bc00aef | 150 | |
eb608e3a JK |
151 | /* |
152 | * Length of period for aging writeout fractions of bdis. This is an | |
153 | * arbitrarily chosen number. The longer the period, the slower fractions will | |
154 | * reflect changes in current writeout rate. | |
155 | */ | |
156 | #define VM_COMPLETIONS_PERIOD_LEN (3*HZ) | |
04fbfdc1 | 157 | |
693108a8 TH |
158 | #ifdef CONFIG_CGROUP_WRITEBACK |
159 | ||
e9f07dfd TH |
160 | #define GDTC_INIT(__wb) .dom = &global_wb_domain, \ |
161 | DTC_INIT_COMMON(__wb) | |
9fc3a43e | 162 | #define GDTC_INIT_NO_WB .dom = &global_wb_domain |
e9f07dfd TH |
163 | |
164 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) | |
165 | { | |
166 | return dtc->dom; | |
167 | } | |
168 | ||
9fc3a43e TH |
169 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
170 | { | |
171 | return mdtc->gdtc; | |
172 | } | |
173 | ||
841710aa TH |
174 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
175 | { | |
176 | return &wb->memcg_completions; | |
177 | } | |
178 | ||
693108a8 TH |
179 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
180 | unsigned long *minp, unsigned long *maxp) | |
181 | { | |
182 | unsigned long this_bw = wb->avg_write_bandwidth; | |
183 | unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); | |
184 | unsigned long long min = wb->bdi->min_ratio; | |
185 | unsigned long long max = wb->bdi->max_ratio; | |
186 | ||
187 | /* | |
188 | * @wb may already be clean by the time control reaches here and | |
189 | * the total may not include its bw. | |
190 | */ | |
191 | if (this_bw < tot_bw) { | |
192 | if (min) { | |
193 | min *= this_bw; | |
194 | do_div(min, tot_bw); | |
195 | } | |
196 | if (max < 100) { | |
197 | max *= this_bw; | |
198 | do_div(max, tot_bw); | |
199 | } | |
200 | } | |
201 | ||
202 | *minp = min; | |
203 | *maxp = max; | |
204 | } | |
205 | ||
206 | #else /* CONFIG_CGROUP_WRITEBACK */ | |
207 | ||
e9f07dfd | 208 | #define GDTC_INIT(__wb) DTC_INIT_COMMON(__wb) |
9fc3a43e | 209 | #define GDTC_INIT_NO_WB |
e9f07dfd TH |
210 | |
211 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) | |
212 | { | |
213 | return &global_wb_domain; | |
214 | } | |
215 | ||
9fc3a43e TH |
216 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
217 | { | |
218 | return NULL; | |
219 | } | |
220 | ||
841710aa TH |
221 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
222 | { | |
223 | return NULL; | |
224 | } | |
225 | ||
693108a8 TH |
226 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
227 | unsigned long *minp, unsigned long *maxp) | |
228 | { | |
229 | *minp = wb->bdi->min_ratio; | |
230 | *maxp = wb->bdi->max_ratio; | |
231 | } | |
232 | ||
233 | #endif /* CONFIG_CGROUP_WRITEBACK */ | |
234 | ||
a756cf59 JW |
235 | /* |
236 | * In a memory zone, there is a certain amount of pages we consider | |
237 | * available for the page cache, which is essentially the number of | |
238 | * free and reclaimable pages, minus some zone reserves to protect | |
239 | * lowmem and the ability to uphold the zone's watermarks without | |
240 | * requiring writeback. | |
241 | * | |
242 | * This number of dirtyable pages is the base value of which the | |
243 | * user-configurable dirty ratio is the effictive number of pages that | |
244 | * are allowed to be actually dirtied. Per individual zone, or | |
245 | * globally by using the sum of dirtyable pages over all zones. | |
246 | * | |
247 | * Because the user is allowed to specify the dirty limit globally as | |
248 | * absolute number of bytes, calculating the per-zone dirty limit can | |
249 | * require translating the configured limit into a percentage of | |
250 | * global dirtyable memory first. | |
251 | */ | |
252 | ||
a804552b JW |
253 | /** |
254 | * zone_dirtyable_memory - number of dirtyable pages in a zone | |
255 | * @zone: the zone | |
256 | * | |
257 | * Returns the zone's number of pages potentially available for dirty | |
258 | * page cache. This is the base value for the per-zone dirty limits. | |
259 | */ | |
260 | static unsigned long zone_dirtyable_memory(struct zone *zone) | |
261 | { | |
262 | unsigned long nr_pages; | |
263 | ||
264 | nr_pages = zone_page_state(zone, NR_FREE_PAGES); | |
265 | nr_pages -= min(nr_pages, zone->dirty_balance_reserve); | |
266 | ||
a1c3bfb2 JW |
267 | nr_pages += zone_page_state(zone, NR_INACTIVE_FILE); |
268 | nr_pages += zone_page_state(zone, NR_ACTIVE_FILE); | |
a804552b JW |
269 | |
270 | return nr_pages; | |
271 | } | |
272 | ||
1edf2234 JW |
273 | static unsigned long highmem_dirtyable_memory(unsigned long total) |
274 | { | |
275 | #ifdef CONFIG_HIGHMEM | |
276 | int node; | |
277 | unsigned long x = 0; | |
278 | ||
279 | for_each_node_state(node, N_HIGH_MEMORY) { | |
a804552b | 280 | struct zone *z = &NODE_DATA(node)->node_zones[ZONE_HIGHMEM]; |
1edf2234 | 281 | |
a804552b | 282 | x += zone_dirtyable_memory(z); |
1edf2234 | 283 | } |
c8b74c2f SR |
284 | /* |
285 | * Unreclaimable memory (kernel memory or anonymous memory | |
286 | * without swap) can bring down the dirtyable pages below | |
287 | * the zone's dirty balance reserve and the above calculation | |
288 | * will underflow. However we still want to add in nodes | |
289 | * which are below threshold (negative values) to get a more | |
290 | * accurate calculation but make sure that the total never | |
291 | * underflows. | |
292 | */ | |
293 | if ((long)x < 0) | |
294 | x = 0; | |
295 | ||
1edf2234 JW |
296 | /* |
297 | * Make sure that the number of highmem pages is never larger | |
298 | * than the number of the total dirtyable memory. This can only | |
299 | * occur in very strange VM situations but we want to make sure | |
300 | * that this does not occur. | |
301 | */ | |
302 | return min(x, total); | |
303 | #else | |
304 | return 0; | |
305 | #endif | |
306 | } | |
307 | ||
308 | /** | |
ccafa287 | 309 | * global_dirtyable_memory - number of globally dirtyable pages |
1edf2234 | 310 | * |
ccafa287 JW |
311 | * Returns the global number of pages potentially available for dirty |
312 | * page cache. This is the base value for the global dirty limits. | |
1edf2234 | 313 | */ |
18cf8cf8 | 314 | static unsigned long global_dirtyable_memory(void) |
1edf2234 JW |
315 | { |
316 | unsigned long x; | |
317 | ||
a804552b | 318 | x = global_page_state(NR_FREE_PAGES); |
c8b74c2f | 319 | x -= min(x, dirty_balance_reserve); |
1edf2234 | 320 | |
a1c3bfb2 JW |
321 | x += global_page_state(NR_INACTIVE_FILE); |
322 | x += global_page_state(NR_ACTIVE_FILE); | |
a804552b | 323 | |
1edf2234 JW |
324 | if (!vm_highmem_is_dirtyable) |
325 | x -= highmem_dirtyable_memory(x); | |
326 | ||
327 | return x + 1; /* Ensure that we never return 0 */ | |
328 | } | |
329 | ||
9fc3a43e TH |
330 | /** |
331 | * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain | |
332 | * @dtc: dirty_throttle_control of interest | |
ccafa287 | 333 | * |
9fc3a43e TH |
334 | * Calculate @dtc->thresh and ->bg_thresh considering |
335 | * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller | |
336 | * must ensure that @dtc->avail is set before calling this function. The | |
337 | * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and | |
ccafa287 JW |
338 | * real-time tasks. |
339 | */ | |
9fc3a43e | 340 | static void domain_dirty_limits(struct dirty_throttle_control *dtc) |
ccafa287 | 341 | { |
9fc3a43e TH |
342 | const unsigned long available_memory = dtc->avail; |
343 | struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc); | |
344 | unsigned long bytes = vm_dirty_bytes; | |
345 | unsigned long bg_bytes = dirty_background_bytes; | |
346 | unsigned long ratio = vm_dirty_ratio; | |
347 | unsigned long bg_ratio = dirty_background_ratio; | |
348 | unsigned long thresh; | |
349 | unsigned long bg_thresh; | |
ccafa287 JW |
350 | struct task_struct *tsk; |
351 | ||
9fc3a43e TH |
352 | /* gdtc is !NULL iff @dtc is for memcg domain */ |
353 | if (gdtc) { | |
354 | unsigned long global_avail = gdtc->avail; | |
355 | ||
356 | /* | |
357 | * The byte settings can't be applied directly to memcg | |
358 | * domains. Convert them to ratios by scaling against | |
359 | * globally available memory. | |
360 | */ | |
361 | if (bytes) | |
362 | ratio = min(DIV_ROUND_UP(bytes, PAGE_SIZE) * 100 / | |
363 | global_avail, 100UL); | |
364 | if (bg_bytes) | |
365 | bg_ratio = min(DIV_ROUND_UP(bg_bytes, PAGE_SIZE) * 100 / | |
366 | global_avail, 100UL); | |
367 | bytes = bg_bytes = 0; | |
368 | } | |
369 | ||
370 | if (bytes) | |
371 | thresh = DIV_ROUND_UP(bytes, PAGE_SIZE); | |
ccafa287 | 372 | else |
9fc3a43e | 373 | thresh = (ratio * available_memory) / 100; |
ccafa287 | 374 | |
9fc3a43e TH |
375 | if (bg_bytes) |
376 | bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE); | |
ccafa287 | 377 | else |
9fc3a43e | 378 | bg_thresh = (bg_ratio * available_memory) / 100; |
ccafa287 | 379 | |
9fc3a43e TH |
380 | if (bg_thresh >= thresh) |
381 | bg_thresh = thresh / 2; | |
ccafa287 JW |
382 | tsk = current; |
383 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) { | |
9fc3a43e TH |
384 | bg_thresh += bg_thresh / 4; |
385 | thresh += thresh / 4; | |
ccafa287 | 386 | } |
9fc3a43e TH |
387 | dtc->thresh = thresh; |
388 | dtc->bg_thresh = bg_thresh; | |
389 | ||
390 | /* we should eventually report the domain in the TP */ | |
391 | if (!gdtc) | |
392 | trace_global_dirty_state(bg_thresh, thresh); | |
393 | } | |
394 | ||
395 | /** | |
396 | * global_dirty_limits - background-writeback and dirty-throttling thresholds | |
397 | * @pbackground: out parameter for bg_thresh | |
398 | * @pdirty: out parameter for thresh | |
399 | * | |
400 | * Calculate bg_thresh and thresh for global_wb_domain. See | |
401 | * domain_dirty_limits() for details. | |
402 | */ | |
403 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) | |
404 | { | |
405 | struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB }; | |
406 | ||
407 | gdtc.avail = global_dirtyable_memory(); | |
408 | domain_dirty_limits(&gdtc); | |
409 | ||
410 | *pbackground = gdtc.bg_thresh; | |
411 | *pdirty = gdtc.thresh; | |
ccafa287 JW |
412 | } |
413 | ||
a756cf59 JW |
414 | /** |
415 | * zone_dirty_limit - maximum number of dirty pages allowed in a zone | |
416 | * @zone: the zone | |
417 | * | |
418 | * Returns the maximum number of dirty pages allowed in a zone, based | |
419 | * on the zone's dirtyable memory. | |
420 | */ | |
421 | static unsigned long zone_dirty_limit(struct zone *zone) | |
422 | { | |
423 | unsigned long zone_memory = zone_dirtyable_memory(zone); | |
424 | struct task_struct *tsk = current; | |
425 | unsigned long dirty; | |
426 | ||
427 | if (vm_dirty_bytes) | |
428 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) * | |
429 | zone_memory / global_dirtyable_memory(); | |
430 | else | |
431 | dirty = vm_dirty_ratio * zone_memory / 100; | |
432 | ||
433 | if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) | |
434 | dirty += dirty / 4; | |
435 | ||
436 | return dirty; | |
437 | } | |
438 | ||
439 | /** | |
440 | * zone_dirty_ok - tells whether a zone is within its dirty limits | |
441 | * @zone: the zone to check | |
442 | * | |
443 | * Returns %true when the dirty pages in @zone are within the zone's | |
444 | * dirty limit, %false if the limit is exceeded. | |
445 | */ | |
446 | bool zone_dirty_ok(struct zone *zone) | |
447 | { | |
448 | unsigned long limit = zone_dirty_limit(zone); | |
449 | ||
450 | return zone_page_state(zone, NR_FILE_DIRTY) + | |
451 | zone_page_state(zone, NR_UNSTABLE_NFS) + | |
452 | zone_page_state(zone, NR_WRITEBACK) <= limit; | |
453 | } | |
454 | ||
2da02997 | 455 | int dirty_background_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 456 | void __user *buffer, size_t *lenp, |
2da02997 DR |
457 | loff_t *ppos) |
458 | { | |
459 | int ret; | |
460 | ||
8d65af78 | 461 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
462 | if (ret == 0 && write) |
463 | dirty_background_bytes = 0; | |
464 | return ret; | |
465 | } | |
466 | ||
467 | int dirty_background_bytes_handler(struct ctl_table *table, int write, | |
8d65af78 | 468 | void __user *buffer, size_t *lenp, |
2da02997 DR |
469 | loff_t *ppos) |
470 | { | |
471 | int ret; | |
472 | ||
8d65af78 | 473 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
474 | if (ret == 0 && write) |
475 | dirty_background_ratio = 0; | |
476 | return ret; | |
477 | } | |
478 | ||
04fbfdc1 | 479 | int dirty_ratio_handler(struct ctl_table *table, int write, |
8d65af78 | 480 | void __user *buffer, size_t *lenp, |
04fbfdc1 PZ |
481 | loff_t *ppos) |
482 | { | |
483 | int old_ratio = vm_dirty_ratio; | |
2da02997 DR |
484 | int ret; |
485 | ||
8d65af78 | 486 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
04fbfdc1 | 487 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
eb608e3a | 488 | writeback_set_ratelimit(); |
2da02997 DR |
489 | vm_dirty_bytes = 0; |
490 | } | |
491 | return ret; | |
492 | } | |
493 | ||
2da02997 | 494 | int dirty_bytes_handler(struct ctl_table *table, int write, |
8d65af78 | 495 | void __user *buffer, size_t *lenp, |
2da02997 DR |
496 | loff_t *ppos) |
497 | { | |
fc3501d4 | 498 | unsigned long old_bytes = vm_dirty_bytes; |
2da02997 DR |
499 | int ret; |
500 | ||
8d65af78 | 501 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 | 502 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
eb608e3a | 503 | writeback_set_ratelimit(); |
2da02997 | 504 | vm_dirty_ratio = 0; |
04fbfdc1 PZ |
505 | } |
506 | return ret; | |
507 | } | |
508 | ||
eb608e3a JK |
509 | static unsigned long wp_next_time(unsigned long cur_time) |
510 | { | |
511 | cur_time += VM_COMPLETIONS_PERIOD_LEN; | |
512 | /* 0 has a special meaning... */ | |
513 | if (!cur_time) | |
514 | return 1; | |
515 | return cur_time; | |
516 | } | |
517 | ||
c7981433 TH |
518 | static void wb_domain_writeout_inc(struct wb_domain *dom, |
519 | struct fprop_local_percpu *completions, | |
520 | unsigned int max_prop_frac) | |
04fbfdc1 | 521 | { |
c7981433 TH |
522 | __fprop_inc_percpu_max(&dom->completions, completions, |
523 | max_prop_frac); | |
eb608e3a | 524 | /* First event after period switching was turned off? */ |
380c27ca | 525 | if (!unlikely(dom->period_time)) { |
eb608e3a JK |
526 | /* |
527 | * We can race with other __bdi_writeout_inc calls here but | |
528 | * it does not cause any harm since the resulting time when | |
529 | * timer will fire and what is in writeout_period_time will be | |
530 | * roughly the same. | |
531 | */ | |
380c27ca TH |
532 | dom->period_time = wp_next_time(jiffies); |
533 | mod_timer(&dom->period_timer, dom->period_time); | |
eb608e3a | 534 | } |
04fbfdc1 PZ |
535 | } |
536 | ||
c7981433 TH |
537 | /* |
538 | * Increment @wb's writeout completion count and the global writeout | |
539 | * completion count. Called from test_clear_page_writeback(). | |
540 | */ | |
541 | static inline void __wb_writeout_inc(struct bdi_writeback *wb) | |
542 | { | |
841710aa TH |
543 | struct wb_domain *cgdom; |
544 | ||
c7981433 TH |
545 | __inc_wb_stat(wb, WB_WRITTEN); |
546 | wb_domain_writeout_inc(&global_wb_domain, &wb->completions, | |
547 | wb->bdi->max_prop_frac); | |
841710aa TH |
548 | |
549 | cgdom = mem_cgroup_wb_domain(wb); | |
550 | if (cgdom) | |
551 | wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb), | |
552 | wb->bdi->max_prop_frac); | |
c7981433 TH |
553 | } |
554 | ||
93f78d88 | 555 | void wb_writeout_inc(struct bdi_writeback *wb) |
dd5656e5 MS |
556 | { |
557 | unsigned long flags; | |
558 | ||
559 | local_irq_save(flags); | |
93f78d88 | 560 | __wb_writeout_inc(wb); |
dd5656e5 MS |
561 | local_irq_restore(flags); |
562 | } | |
93f78d88 | 563 | EXPORT_SYMBOL_GPL(wb_writeout_inc); |
dd5656e5 | 564 | |
eb608e3a JK |
565 | /* |
566 | * On idle system, we can be called long after we scheduled because we use | |
567 | * deferred timers so count with missed periods. | |
568 | */ | |
569 | static void writeout_period(unsigned long t) | |
570 | { | |
380c27ca TH |
571 | struct wb_domain *dom = (void *)t; |
572 | int miss_periods = (jiffies - dom->period_time) / | |
eb608e3a JK |
573 | VM_COMPLETIONS_PERIOD_LEN; |
574 | ||
380c27ca TH |
575 | if (fprop_new_period(&dom->completions, miss_periods + 1)) { |
576 | dom->period_time = wp_next_time(dom->period_time + | |
eb608e3a | 577 | miss_periods * VM_COMPLETIONS_PERIOD_LEN); |
380c27ca | 578 | mod_timer(&dom->period_timer, dom->period_time); |
eb608e3a JK |
579 | } else { |
580 | /* | |
581 | * Aging has zeroed all fractions. Stop wasting CPU on period | |
582 | * updates. | |
583 | */ | |
380c27ca | 584 | dom->period_time = 0; |
eb608e3a JK |
585 | } |
586 | } | |
587 | ||
380c27ca TH |
588 | int wb_domain_init(struct wb_domain *dom, gfp_t gfp) |
589 | { | |
590 | memset(dom, 0, sizeof(*dom)); | |
dcc25ae7 TH |
591 | |
592 | spin_lock_init(&dom->lock); | |
593 | ||
380c27ca TH |
594 | init_timer_deferrable(&dom->period_timer); |
595 | dom->period_timer.function = writeout_period; | |
596 | dom->period_timer.data = (unsigned long)dom; | |
dcc25ae7 TH |
597 | |
598 | dom->dirty_limit_tstamp = jiffies; | |
599 | ||
380c27ca TH |
600 | return fprop_global_init(&dom->completions, gfp); |
601 | } | |
602 | ||
841710aa TH |
603 | #ifdef CONFIG_CGROUP_WRITEBACK |
604 | void wb_domain_exit(struct wb_domain *dom) | |
605 | { | |
606 | del_timer_sync(&dom->period_timer); | |
607 | fprop_global_destroy(&dom->completions); | |
608 | } | |
609 | #endif | |
610 | ||
189d3c4a | 611 | /* |
d08c429b JW |
612 | * bdi_min_ratio keeps the sum of the minimum dirty shares of all |
613 | * registered backing devices, which, for obvious reasons, can not | |
614 | * exceed 100%. | |
189d3c4a | 615 | */ |
189d3c4a PZ |
616 | static unsigned int bdi_min_ratio; |
617 | ||
618 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) | |
619 | { | |
620 | int ret = 0; | |
189d3c4a | 621 | |
cfc4ba53 | 622 | spin_lock_bh(&bdi_lock); |
a42dde04 | 623 | if (min_ratio > bdi->max_ratio) { |
189d3c4a | 624 | ret = -EINVAL; |
a42dde04 PZ |
625 | } else { |
626 | min_ratio -= bdi->min_ratio; | |
627 | if (bdi_min_ratio + min_ratio < 100) { | |
628 | bdi_min_ratio += min_ratio; | |
629 | bdi->min_ratio += min_ratio; | |
630 | } else { | |
631 | ret = -EINVAL; | |
632 | } | |
633 | } | |
cfc4ba53 | 634 | spin_unlock_bh(&bdi_lock); |
a42dde04 PZ |
635 | |
636 | return ret; | |
637 | } | |
638 | ||
639 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio) | |
640 | { | |
a42dde04 PZ |
641 | int ret = 0; |
642 | ||
643 | if (max_ratio > 100) | |
644 | return -EINVAL; | |
645 | ||
cfc4ba53 | 646 | spin_lock_bh(&bdi_lock); |
a42dde04 PZ |
647 | if (bdi->min_ratio > max_ratio) { |
648 | ret = -EINVAL; | |
649 | } else { | |
650 | bdi->max_ratio = max_ratio; | |
eb608e3a | 651 | bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100; |
a42dde04 | 652 | } |
cfc4ba53 | 653 | spin_unlock_bh(&bdi_lock); |
189d3c4a PZ |
654 | |
655 | return ret; | |
656 | } | |
a42dde04 | 657 | EXPORT_SYMBOL(bdi_set_max_ratio); |
189d3c4a | 658 | |
6c14ae1e WF |
659 | static unsigned long dirty_freerun_ceiling(unsigned long thresh, |
660 | unsigned long bg_thresh) | |
661 | { | |
662 | return (thresh + bg_thresh) / 2; | |
663 | } | |
664 | ||
c7981433 TH |
665 | static unsigned long hard_dirty_limit(struct wb_domain *dom, |
666 | unsigned long thresh) | |
ffd1f609 | 667 | { |
dcc25ae7 | 668 | return max(thresh, dom->dirty_limit); |
ffd1f609 WF |
669 | } |
670 | ||
6f718656 | 671 | /** |
b1cbc6d4 TH |
672 | * __wb_calc_thresh - @wb's share of dirty throttling threshold |
673 | * @dtc: dirty_throttle_context of interest | |
1babe183 | 674 | * |
a88a341a | 675 | * Returns @wb's dirty limit in pages. The term "dirty" in the context of |
6f718656 | 676 | * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages. |
aed21ad2 WF |
677 | * |
678 | * Note that balance_dirty_pages() will only seriously take it as a hard limit | |
679 | * when sleeping max_pause per page is not enough to keep the dirty pages under | |
680 | * control. For example, when the device is completely stalled due to some error | |
681 | * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key. | |
682 | * In the other normal situations, it acts more gently by throttling the tasks | |
a88a341a | 683 | * more (rather than completely block them) when the wb dirty pages go high. |
1babe183 | 684 | * |
6f718656 | 685 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
1babe183 WF |
686 | * - starving fast devices |
687 | * - piling up dirty pages (that will take long time to sync) on slow devices | |
688 | * | |
a88a341a | 689 | * The wb's share of dirty limit will be adapting to its throughput and |
1babe183 WF |
690 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. |
691 | */ | |
b1cbc6d4 | 692 | static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc) |
16c4042f | 693 | { |
e9f07dfd | 694 | struct wb_domain *dom = dtc_dom(dtc); |
b1cbc6d4 | 695 | unsigned long thresh = dtc->thresh; |
0d960a38 | 696 | u64 wb_thresh; |
16c4042f | 697 | long numerator, denominator; |
693108a8 | 698 | unsigned long wb_min_ratio, wb_max_ratio; |
04fbfdc1 | 699 | |
16c4042f | 700 | /* |
0d960a38 | 701 | * Calculate this BDI's share of the thresh ratio. |
16c4042f | 702 | */ |
e9770b34 | 703 | fprop_fraction_percpu(&dom->completions, dtc->wb_completions, |
380c27ca | 704 | &numerator, &denominator); |
04fbfdc1 | 705 | |
0d960a38 TH |
706 | wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100; |
707 | wb_thresh *= numerator; | |
708 | do_div(wb_thresh, denominator); | |
04fbfdc1 | 709 | |
b1cbc6d4 | 710 | wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio); |
693108a8 | 711 | |
0d960a38 TH |
712 | wb_thresh += (thresh * wb_min_ratio) / 100; |
713 | if (wb_thresh > (thresh * wb_max_ratio) / 100) | |
714 | wb_thresh = thresh * wb_max_ratio / 100; | |
16c4042f | 715 | |
0d960a38 | 716 | return wb_thresh; |
1da177e4 LT |
717 | } |
718 | ||
b1cbc6d4 TH |
719 | unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh) |
720 | { | |
721 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb), | |
722 | .thresh = thresh }; | |
723 | return __wb_calc_thresh(&gdtc); | |
724 | } | |
725 | ||
5a537485 MP |
726 | /* |
727 | * setpoint - dirty 3 | |
728 | * f(dirty) := 1.0 + (----------------) | |
729 | * limit - setpoint | |
730 | * | |
731 | * it's a 3rd order polynomial that subjects to | |
732 | * | |
733 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast | |
734 | * (2) f(setpoint) = 1.0 => the balance point | |
735 | * (3) f(limit) = 0 => the hard limit | |
736 | * (4) df/dx <= 0 => negative feedback control | |
737 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) | |
738 | * => fast response on large errors; small oscillation near setpoint | |
739 | */ | |
d5c9fde3 | 740 | static long long pos_ratio_polynom(unsigned long setpoint, |
5a537485 MP |
741 | unsigned long dirty, |
742 | unsigned long limit) | |
743 | { | |
744 | long long pos_ratio; | |
745 | long x; | |
746 | ||
d5c9fde3 | 747 | x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT, |
5a537485 MP |
748 | limit - setpoint + 1); |
749 | pos_ratio = x; | |
750 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
751 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
752 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT; | |
753 | ||
754 | return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT); | |
755 | } | |
756 | ||
6c14ae1e WF |
757 | /* |
758 | * Dirty position control. | |
759 | * | |
760 | * (o) global/bdi setpoints | |
761 | * | |
de1fff37 | 762 | * We want the dirty pages be balanced around the global/wb setpoints. |
6c14ae1e WF |
763 | * When the number of dirty pages is higher/lower than the setpoint, the |
764 | * dirty position control ratio (and hence task dirty ratelimit) will be | |
765 | * decreased/increased to bring the dirty pages back to the setpoint. | |
766 | * | |
767 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT | |
768 | * | |
769 | * if (dirty < setpoint) scale up pos_ratio | |
770 | * if (dirty > setpoint) scale down pos_ratio | |
771 | * | |
de1fff37 TH |
772 | * if (wb_dirty < wb_setpoint) scale up pos_ratio |
773 | * if (wb_dirty > wb_setpoint) scale down pos_ratio | |
6c14ae1e WF |
774 | * |
775 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT | |
776 | * | |
777 | * (o) global control line | |
778 | * | |
779 | * ^ pos_ratio | |
780 | * | | |
781 | * | |<===== global dirty control scope ======>| | |
782 | * 2.0 .............* | |
783 | * | .* | |
784 | * | . * | |
785 | * | . * | |
786 | * | . * | |
787 | * | . * | |
788 | * | . * | |
789 | * 1.0 ................................* | |
790 | * | . . * | |
791 | * | . . * | |
792 | * | . . * | |
793 | * | . . * | |
794 | * | . . * | |
795 | * 0 +------------.------------------.----------------------*-------------> | |
796 | * freerun^ setpoint^ limit^ dirty pages | |
797 | * | |
de1fff37 | 798 | * (o) wb control line |
6c14ae1e WF |
799 | * |
800 | * ^ pos_ratio | |
801 | * | | |
802 | * | * | |
803 | * | * | |
804 | * | * | |
805 | * | * | |
806 | * | * |<=========== span ============>| | |
807 | * 1.0 .......................* | |
808 | * | . * | |
809 | * | . * | |
810 | * | . * | |
811 | * | . * | |
812 | * | . * | |
813 | * | . * | |
814 | * | . * | |
815 | * | . * | |
816 | * | . * | |
817 | * | . * | |
818 | * | . * | |
819 | * 1/4 ...............................................* * * * * * * * * * * * | |
820 | * | . . | |
821 | * | . . | |
822 | * | . . | |
823 | * 0 +----------------------.-------------------------------.-------------> | |
de1fff37 | 824 | * wb_setpoint^ x_intercept^ |
6c14ae1e | 825 | * |
de1fff37 | 826 | * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can |
6c14ae1e WF |
827 | * be smoothly throttled down to normal if it starts high in situations like |
828 | * - start writing to a slow SD card and a fast disk at the same time. The SD | |
de1fff37 TH |
829 | * card's wb_dirty may rush to many times higher than wb_setpoint. |
830 | * - the wb dirty thresh drops quickly due to change of JBOD workload | |
6c14ae1e | 831 | */ |
daddfa3c | 832 | static void wb_position_ratio(struct dirty_throttle_control *dtc) |
6c14ae1e | 833 | { |
2bc00aef | 834 | struct bdi_writeback *wb = dtc->wb; |
a88a341a | 835 | unsigned long write_bw = wb->avg_write_bandwidth; |
2bc00aef | 836 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); |
c7981433 | 837 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
2bc00aef | 838 | unsigned long wb_thresh = dtc->wb_thresh; |
6c14ae1e WF |
839 | unsigned long x_intercept; |
840 | unsigned long setpoint; /* dirty pages' target balance point */ | |
de1fff37 | 841 | unsigned long wb_setpoint; |
6c14ae1e WF |
842 | unsigned long span; |
843 | long long pos_ratio; /* for scaling up/down the rate limit */ | |
844 | long x; | |
845 | ||
daddfa3c TH |
846 | dtc->pos_ratio = 0; |
847 | ||
2bc00aef | 848 | if (unlikely(dtc->dirty >= limit)) |
daddfa3c | 849 | return; |
6c14ae1e WF |
850 | |
851 | /* | |
852 | * global setpoint | |
853 | * | |
5a537485 MP |
854 | * See comment for pos_ratio_polynom(). |
855 | */ | |
856 | setpoint = (freerun + limit) / 2; | |
2bc00aef | 857 | pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit); |
5a537485 MP |
858 | |
859 | /* | |
860 | * The strictlimit feature is a tool preventing mistrusted filesystems | |
861 | * from growing a large number of dirty pages before throttling. For | |
de1fff37 TH |
862 | * such filesystems balance_dirty_pages always checks wb counters |
863 | * against wb limits. Even if global "nr_dirty" is under "freerun". | |
5a537485 MP |
864 | * This is especially important for fuse which sets bdi->max_ratio to |
865 | * 1% by default. Without strictlimit feature, fuse writeback may | |
866 | * consume arbitrary amount of RAM because it is accounted in | |
867 | * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty". | |
6c14ae1e | 868 | * |
a88a341a | 869 | * Here, in wb_position_ratio(), we calculate pos_ratio based on |
de1fff37 | 870 | * two values: wb_dirty and wb_thresh. Let's consider an example: |
5a537485 MP |
871 | * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global |
872 | * limits are set by default to 10% and 20% (background and throttle). | |
de1fff37 | 873 | * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages. |
0d960a38 | 874 | * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is |
de1fff37 | 875 | * about ~6K pages (as the average of background and throttle wb |
5a537485 | 876 | * limits). The 3rd order polynomial will provide positive feedback if |
de1fff37 | 877 | * wb_dirty is under wb_setpoint and vice versa. |
6c14ae1e | 878 | * |
5a537485 | 879 | * Note, that we cannot use global counters in these calculations |
de1fff37 | 880 | * because we want to throttle process writing to a strictlimit wb |
5a537485 MP |
881 | * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB |
882 | * in the example above). | |
6c14ae1e | 883 | */ |
a88a341a | 884 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
de1fff37 | 885 | long long wb_pos_ratio; |
5a537485 | 886 | |
daddfa3c TH |
887 | if (dtc->wb_dirty < 8) { |
888 | dtc->pos_ratio = min_t(long long, pos_ratio * 2, | |
889 | 2 << RATELIMIT_CALC_SHIFT); | |
890 | return; | |
891 | } | |
5a537485 | 892 | |
2bc00aef | 893 | if (dtc->wb_dirty >= wb_thresh) |
daddfa3c | 894 | return; |
5a537485 | 895 | |
970fb01a TH |
896 | wb_setpoint = dirty_freerun_ceiling(wb_thresh, |
897 | dtc->wb_bg_thresh); | |
5a537485 | 898 | |
de1fff37 | 899 | if (wb_setpoint == 0 || wb_setpoint == wb_thresh) |
daddfa3c | 900 | return; |
5a537485 | 901 | |
2bc00aef | 902 | wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty, |
de1fff37 | 903 | wb_thresh); |
5a537485 MP |
904 | |
905 | /* | |
de1fff37 TH |
906 | * Typically, for strictlimit case, wb_setpoint << setpoint |
907 | * and pos_ratio >> wb_pos_ratio. In the other words global | |
5a537485 | 908 | * state ("dirty") is not limiting factor and we have to |
de1fff37 | 909 | * make decision based on wb counters. But there is an |
5a537485 MP |
910 | * important case when global pos_ratio should get precedence: |
911 | * global limits are exceeded (e.g. due to activities on other | |
de1fff37 | 912 | * wb's) while given strictlimit wb is below limit. |
5a537485 | 913 | * |
de1fff37 | 914 | * "pos_ratio * wb_pos_ratio" would work for the case above, |
5a537485 | 915 | * but it would look too non-natural for the case of all |
de1fff37 | 916 | * activity in the system coming from a single strictlimit wb |
5a537485 MP |
917 | * with bdi->max_ratio == 100%. |
918 | * | |
919 | * Note that min() below somewhat changes the dynamics of the | |
920 | * control system. Normally, pos_ratio value can be well over 3 | |
de1fff37 | 921 | * (when globally we are at freerun and wb is well below wb |
5a537485 MP |
922 | * setpoint). Now the maximum pos_ratio in the same situation |
923 | * is 2. We might want to tweak this if we observe the control | |
924 | * system is too slow to adapt. | |
925 | */ | |
daddfa3c TH |
926 | dtc->pos_ratio = min(pos_ratio, wb_pos_ratio); |
927 | return; | |
5a537485 | 928 | } |
6c14ae1e WF |
929 | |
930 | /* | |
931 | * We have computed basic pos_ratio above based on global situation. If | |
de1fff37 | 932 | * the wb is over/under its share of dirty pages, we want to scale |
6c14ae1e WF |
933 | * pos_ratio further down/up. That is done by the following mechanism. |
934 | */ | |
935 | ||
936 | /* | |
de1fff37 | 937 | * wb setpoint |
6c14ae1e | 938 | * |
de1fff37 | 939 | * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint) |
6c14ae1e | 940 | * |
de1fff37 | 941 | * x_intercept - wb_dirty |
6c14ae1e | 942 | * := -------------------------- |
de1fff37 | 943 | * x_intercept - wb_setpoint |
6c14ae1e | 944 | * |
de1fff37 | 945 | * The main wb control line is a linear function that subjects to |
6c14ae1e | 946 | * |
de1fff37 TH |
947 | * (1) f(wb_setpoint) = 1.0 |
948 | * (2) k = - 1 / (8 * write_bw) (in single wb case) | |
949 | * or equally: x_intercept = wb_setpoint + 8 * write_bw | |
6c14ae1e | 950 | * |
de1fff37 | 951 | * For single wb case, the dirty pages are observed to fluctuate |
6c14ae1e | 952 | * regularly within range |
de1fff37 | 953 | * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2] |
6c14ae1e WF |
954 | * for various filesystems, where (2) can yield in a reasonable 12.5% |
955 | * fluctuation range for pos_ratio. | |
956 | * | |
de1fff37 | 957 | * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its |
6c14ae1e | 958 | * own size, so move the slope over accordingly and choose a slope that |
de1fff37 | 959 | * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh. |
6c14ae1e | 960 | */ |
2bc00aef TH |
961 | if (unlikely(wb_thresh > dtc->thresh)) |
962 | wb_thresh = dtc->thresh; | |
aed21ad2 | 963 | /* |
de1fff37 | 964 | * It's very possible that wb_thresh is close to 0 not because the |
aed21ad2 WF |
965 | * device is slow, but that it has remained inactive for long time. |
966 | * Honour such devices a reasonable good (hopefully IO efficient) | |
967 | * threshold, so that the occasional writes won't be blocked and active | |
968 | * writes can rampup the threshold quickly. | |
969 | */ | |
2bc00aef | 970 | wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8); |
6c14ae1e | 971 | /* |
de1fff37 TH |
972 | * scale global setpoint to wb's: |
973 | * wb_setpoint = setpoint * wb_thresh / thresh | |
6c14ae1e | 974 | */ |
2bc00aef | 975 | x = div_u64((u64)wb_thresh << 16, dtc->thresh + 1); |
de1fff37 | 976 | wb_setpoint = setpoint * (u64)x >> 16; |
6c14ae1e | 977 | /* |
de1fff37 TH |
978 | * Use span=(8*write_bw) in single wb case as indicated by |
979 | * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case. | |
6c14ae1e | 980 | * |
de1fff37 TH |
981 | * wb_thresh thresh - wb_thresh |
982 | * span = --------- * (8 * write_bw) + ------------------ * wb_thresh | |
983 | * thresh thresh | |
6c14ae1e | 984 | */ |
2bc00aef | 985 | span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16; |
de1fff37 | 986 | x_intercept = wb_setpoint + span; |
6c14ae1e | 987 | |
2bc00aef TH |
988 | if (dtc->wb_dirty < x_intercept - span / 4) { |
989 | pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty), | |
990 | x_intercept - wb_setpoint + 1); | |
6c14ae1e WF |
991 | } else |
992 | pos_ratio /= 4; | |
993 | ||
8927f66c | 994 | /* |
de1fff37 | 995 | * wb reserve area, safeguard against dirty pool underrun and disk idle |
8927f66c WF |
996 | * It may push the desired control point of global dirty pages higher |
997 | * than setpoint. | |
998 | */ | |
de1fff37 | 999 | x_intercept = wb_thresh / 2; |
2bc00aef TH |
1000 | if (dtc->wb_dirty < x_intercept) { |
1001 | if (dtc->wb_dirty > x_intercept / 8) | |
1002 | pos_ratio = div_u64(pos_ratio * x_intercept, | |
1003 | dtc->wb_dirty); | |
50657fc4 | 1004 | else |
8927f66c WF |
1005 | pos_ratio *= 8; |
1006 | } | |
1007 | ||
daddfa3c | 1008 | dtc->pos_ratio = pos_ratio; |
6c14ae1e WF |
1009 | } |
1010 | ||
a88a341a TH |
1011 | static void wb_update_write_bandwidth(struct bdi_writeback *wb, |
1012 | unsigned long elapsed, | |
1013 | unsigned long written) | |
e98be2d5 WF |
1014 | { |
1015 | const unsigned long period = roundup_pow_of_two(3 * HZ); | |
a88a341a TH |
1016 | unsigned long avg = wb->avg_write_bandwidth; |
1017 | unsigned long old = wb->write_bandwidth; | |
e98be2d5 WF |
1018 | u64 bw; |
1019 | ||
1020 | /* | |
1021 | * bw = written * HZ / elapsed | |
1022 | * | |
1023 | * bw * elapsed + write_bandwidth * (period - elapsed) | |
1024 | * write_bandwidth = --------------------------------------------------- | |
1025 | * period | |
c72efb65 TH |
1026 | * |
1027 | * @written may have decreased due to account_page_redirty(). | |
1028 | * Avoid underflowing @bw calculation. | |
e98be2d5 | 1029 | */ |
a88a341a | 1030 | bw = written - min(written, wb->written_stamp); |
e98be2d5 WF |
1031 | bw *= HZ; |
1032 | if (unlikely(elapsed > period)) { | |
1033 | do_div(bw, elapsed); | |
1034 | avg = bw; | |
1035 | goto out; | |
1036 | } | |
a88a341a | 1037 | bw += (u64)wb->write_bandwidth * (period - elapsed); |
e98be2d5 WF |
1038 | bw >>= ilog2(period); |
1039 | ||
1040 | /* | |
1041 | * one more level of smoothing, for filtering out sudden spikes | |
1042 | */ | |
1043 | if (avg > old && old >= (unsigned long)bw) | |
1044 | avg -= (avg - old) >> 3; | |
1045 | ||
1046 | if (avg < old && old <= (unsigned long)bw) | |
1047 | avg += (old - avg) >> 3; | |
1048 | ||
1049 | out: | |
95a46c65 TH |
1050 | /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */ |
1051 | avg = max(avg, 1LU); | |
1052 | if (wb_has_dirty_io(wb)) { | |
1053 | long delta = avg - wb->avg_write_bandwidth; | |
1054 | WARN_ON_ONCE(atomic_long_add_return(delta, | |
1055 | &wb->bdi->tot_write_bandwidth) <= 0); | |
1056 | } | |
a88a341a TH |
1057 | wb->write_bandwidth = bw; |
1058 | wb->avg_write_bandwidth = avg; | |
e98be2d5 WF |
1059 | } |
1060 | ||
2bc00aef | 1061 | static void update_dirty_limit(struct dirty_throttle_control *dtc) |
c42843f2 | 1062 | { |
e9f07dfd | 1063 | struct wb_domain *dom = dtc_dom(dtc); |
2bc00aef | 1064 | unsigned long thresh = dtc->thresh; |
dcc25ae7 | 1065 | unsigned long limit = dom->dirty_limit; |
c42843f2 WF |
1066 | |
1067 | /* | |
1068 | * Follow up in one step. | |
1069 | */ | |
1070 | if (limit < thresh) { | |
1071 | limit = thresh; | |
1072 | goto update; | |
1073 | } | |
1074 | ||
1075 | /* | |
1076 | * Follow down slowly. Use the higher one as the target, because thresh | |
1077 | * may drop below dirty. This is exactly the reason to introduce | |
dcc25ae7 | 1078 | * dom->dirty_limit which is guaranteed to lie above the dirty pages. |
c42843f2 | 1079 | */ |
2bc00aef | 1080 | thresh = max(thresh, dtc->dirty); |
c42843f2 WF |
1081 | if (limit > thresh) { |
1082 | limit -= (limit - thresh) >> 5; | |
1083 | goto update; | |
1084 | } | |
1085 | return; | |
1086 | update: | |
dcc25ae7 | 1087 | dom->dirty_limit = limit; |
c42843f2 WF |
1088 | } |
1089 | ||
e9f07dfd | 1090 | static void domain_update_bandwidth(struct dirty_throttle_control *dtc, |
c42843f2 WF |
1091 | unsigned long now) |
1092 | { | |
e9f07dfd | 1093 | struct wb_domain *dom = dtc_dom(dtc); |
c42843f2 WF |
1094 | |
1095 | /* | |
1096 | * check locklessly first to optimize away locking for the most time | |
1097 | */ | |
dcc25ae7 | 1098 | if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) |
c42843f2 WF |
1099 | return; |
1100 | ||
dcc25ae7 TH |
1101 | spin_lock(&dom->lock); |
1102 | if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) { | |
2bc00aef | 1103 | update_dirty_limit(dtc); |
dcc25ae7 | 1104 | dom->dirty_limit_tstamp = now; |
c42843f2 | 1105 | } |
dcc25ae7 | 1106 | spin_unlock(&dom->lock); |
c42843f2 WF |
1107 | } |
1108 | ||
be3ffa27 | 1109 | /* |
de1fff37 | 1110 | * Maintain wb->dirty_ratelimit, the base dirty throttle rate. |
be3ffa27 | 1111 | * |
de1fff37 | 1112 | * Normal wb tasks will be curbed at or below it in long term. |
be3ffa27 WF |
1113 | * Obviously it should be around (write_bw / N) when there are N dd tasks. |
1114 | */ | |
2bc00aef | 1115 | static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc, |
a88a341a TH |
1116 | unsigned long dirtied, |
1117 | unsigned long elapsed) | |
be3ffa27 | 1118 | { |
2bc00aef TH |
1119 | struct bdi_writeback *wb = dtc->wb; |
1120 | unsigned long dirty = dtc->dirty; | |
1121 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); | |
c7981433 | 1122 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
7381131c | 1123 | unsigned long setpoint = (freerun + limit) / 2; |
a88a341a TH |
1124 | unsigned long write_bw = wb->avg_write_bandwidth; |
1125 | unsigned long dirty_ratelimit = wb->dirty_ratelimit; | |
be3ffa27 WF |
1126 | unsigned long dirty_rate; |
1127 | unsigned long task_ratelimit; | |
1128 | unsigned long balanced_dirty_ratelimit; | |
7381131c WF |
1129 | unsigned long step; |
1130 | unsigned long x; | |
be3ffa27 WF |
1131 | |
1132 | /* | |
1133 | * The dirty rate will match the writeout rate in long term, except | |
1134 | * when dirty pages are truncated by userspace or re-dirtied by FS. | |
1135 | */ | |
a88a341a | 1136 | dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed; |
be3ffa27 | 1137 | |
be3ffa27 WF |
1138 | /* |
1139 | * task_ratelimit reflects each dd's dirty rate for the past 200ms. | |
1140 | */ | |
1141 | task_ratelimit = (u64)dirty_ratelimit * | |
daddfa3c | 1142 | dtc->pos_ratio >> RATELIMIT_CALC_SHIFT; |
be3ffa27 WF |
1143 | task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ |
1144 | ||
1145 | /* | |
1146 | * A linear estimation of the "balanced" throttle rate. The theory is, | |
de1fff37 | 1147 | * if there are N dd tasks, each throttled at task_ratelimit, the wb's |
be3ffa27 WF |
1148 | * dirty_rate will be measured to be (N * task_ratelimit). So the below |
1149 | * formula will yield the balanced rate limit (write_bw / N). | |
1150 | * | |
1151 | * Note that the expanded form is not a pure rate feedback: | |
1152 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) | |
1153 | * but also takes pos_ratio into account: | |
1154 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) | |
1155 | * | |
1156 | * (1) is not realistic because pos_ratio also takes part in balancing | |
1157 | * the dirty rate. Consider the state | |
1158 | * pos_ratio = 0.5 (3) | |
1159 | * rate = 2 * (write_bw / N) (4) | |
1160 | * If (1) is used, it will stuck in that state! Because each dd will | |
1161 | * be throttled at | |
1162 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) | |
1163 | * yielding | |
1164 | * dirty_rate = N * task_ratelimit = write_bw (6) | |
1165 | * put (6) into (1) we get | |
1166 | * rate_(i+1) = rate_(i) (7) | |
1167 | * | |
1168 | * So we end up using (2) to always keep | |
1169 | * rate_(i+1) ~= (write_bw / N) (8) | |
1170 | * regardless of the value of pos_ratio. As long as (8) is satisfied, | |
1171 | * pos_ratio is able to drive itself to 1.0, which is not only where | |
1172 | * the dirty count meet the setpoint, but also where the slope of | |
1173 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated. | |
1174 | */ | |
1175 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, | |
1176 | dirty_rate | 1); | |
bdaac490 WF |
1177 | /* |
1178 | * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw | |
1179 | */ | |
1180 | if (unlikely(balanced_dirty_ratelimit > write_bw)) | |
1181 | balanced_dirty_ratelimit = write_bw; | |
be3ffa27 | 1182 | |
7381131c WF |
1183 | /* |
1184 | * We could safely do this and return immediately: | |
1185 | * | |
de1fff37 | 1186 | * wb->dirty_ratelimit = balanced_dirty_ratelimit; |
7381131c WF |
1187 | * |
1188 | * However to get a more stable dirty_ratelimit, the below elaborated | |
331cbdee | 1189 | * code makes use of task_ratelimit to filter out singular points and |
7381131c WF |
1190 | * limit the step size. |
1191 | * | |
1192 | * The below code essentially only uses the relative value of | |
1193 | * | |
1194 | * task_ratelimit - dirty_ratelimit | |
1195 | * = (pos_ratio - 1) * dirty_ratelimit | |
1196 | * | |
1197 | * which reflects the direction and size of dirty position error. | |
1198 | */ | |
1199 | ||
1200 | /* | |
1201 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff | |
1202 | * task_ratelimit is on the same side of dirty_ratelimit, too. | |
1203 | * For example, when | |
1204 | * - dirty_ratelimit > balanced_dirty_ratelimit | |
1205 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) | |
1206 | * lowering dirty_ratelimit will help meet both the position and rate | |
1207 | * control targets. Otherwise, don't update dirty_ratelimit if it will | |
1208 | * only help meet the rate target. After all, what the users ultimately | |
1209 | * feel and care are stable dirty rate and small position error. | |
1210 | * | |
1211 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size | |
331cbdee | 1212 | * and filter out the singular points of balanced_dirty_ratelimit. Which |
7381131c WF |
1213 | * keeps jumping around randomly and can even leap far away at times |
1214 | * due to the small 200ms estimation period of dirty_rate (we want to | |
1215 | * keep that period small to reduce time lags). | |
1216 | */ | |
1217 | step = 0; | |
5a537485 MP |
1218 | |
1219 | /* | |
de1fff37 | 1220 | * For strictlimit case, calculations above were based on wb counters |
a88a341a | 1221 | * and limits (starting from pos_ratio = wb_position_ratio() and up to |
5a537485 | 1222 | * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate). |
de1fff37 TH |
1223 | * Hence, to calculate "step" properly, we have to use wb_dirty as |
1224 | * "dirty" and wb_setpoint as "setpoint". | |
5a537485 | 1225 | * |
de1fff37 TH |
1226 | * We rampup dirty_ratelimit forcibly if wb_dirty is low because |
1227 | * it's possible that wb_thresh is close to zero due to inactivity | |
970fb01a | 1228 | * of backing device. |
5a537485 | 1229 | */ |
a88a341a | 1230 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
2bc00aef TH |
1231 | dirty = dtc->wb_dirty; |
1232 | if (dtc->wb_dirty < 8) | |
1233 | setpoint = dtc->wb_dirty + 1; | |
5a537485 | 1234 | else |
970fb01a | 1235 | setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2; |
5a537485 MP |
1236 | } |
1237 | ||
7381131c | 1238 | if (dirty < setpoint) { |
a88a341a | 1239 | x = min3(wb->balanced_dirty_ratelimit, |
7c809968 | 1240 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1241 | if (dirty_ratelimit < x) |
1242 | step = x - dirty_ratelimit; | |
1243 | } else { | |
a88a341a | 1244 | x = max3(wb->balanced_dirty_ratelimit, |
7c809968 | 1245 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1246 | if (dirty_ratelimit > x) |
1247 | step = dirty_ratelimit - x; | |
1248 | } | |
1249 | ||
1250 | /* | |
1251 | * Don't pursue 100% rate matching. It's impossible since the balanced | |
1252 | * rate itself is constantly fluctuating. So decrease the track speed | |
1253 | * when it gets close to the target. Helps eliminate pointless tremors. | |
1254 | */ | |
1255 | step >>= dirty_ratelimit / (2 * step + 1); | |
1256 | /* | |
1257 | * Limit the tracking speed to avoid overshooting. | |
1258 | */ | |
1259 | step = (step + 7) / 8; | |
1260 | ||
1261 | if (dirty_ratelimit < balanced_dirty_ratelimit) | |
1262 | dirty_ratelimit += step; | |
1263 | else | |
1264 | dirty_ratelimit -= step; | |
1265 | ||
a88a341a TH |
1266 | wb->dirty_ratelimit = max(dirty_ratelimit, 1UL); |
1267 | wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit; | |
b48c104d | 1268 | |
a88a341a | 1269 | trace_bdi_dirty_ratelimit(wb->bdi, dirty_rate, task_ratelimit); |
be3ffa27 WF |
1270 | } |
1271 | ||
2bc00aef | 1272 | static void __wb_update_bandwidth(struct dirty_throttle_control *dtc, |
8a731799 TH |
1273 | unsigned long start_time, |
1274 | bool update_ratelimit) | |
e98be2d5 | 1275 | { |
2bc00aef | 1276 | struct bdi_writeback *wb = dtc->wb; |
e98be2d5 | 1277 | unsigned long now = jiffies; |
a88a341a | 1278 | unsigned long elapsed = now - wb->bw_time_stamp; |
be3ffa27 | 1279 | unsigned long dirtied; |
e98be2d5 WF |
1280 | unsigned long written; |
1281 | ||
8a731799 TH |
1282 | lockdep_assert_held(&wb->list_lock); |
1283 | ||
e98be2d5 WF |
1284 | /* |
1285 | * rate-limit, only update once every 200ms. | |
1286 | */ | |
1287 | if (elapsed < BANDWIDTH_INTERVAL) | |
1288 | return; | |
1289 | ||
a88a341a TH |
1290 | dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]); |
1291 | written = percpu_counter_read(&wb->stat[WB_WRITTEN]); | |
e98be2d5 WF |
1292 | |
1293 | /* | |
1294 | * Skip quiet periods when disk bandwidth is under-utilized. | |
1295 | * (at least 1s idle time between two flusher runs) | |
1296 | */ | |
a88a341a | 1297 | if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time)) |
e98be2d5 WF |
1298 | goto snapshot; |
1299 | ||
8a731799 | 1300 | if (update_ratelimit) { |
e9f07dfd | 1301 | domain_update_bandwidth(dtc, now); |
2bc00aef | 1302 | wb_update_dirty_ratelimit(dtc, dirtied, elapsed); |
be3ffa27 | 1303 | } |
a88a341a | 1304 | wb_update_write_bandwidth(wb, elapsed, written); |
e98be2d5 WF |
1305 | |
1306 | snapshot: | |
a88a341a TH |
1307 | wb->dirtied_stamp = dirtied; |
1308 | wb->written_stamp = written; | |
1309 | wb->bw_time_stamp = now; | |
e98be2d5 WF |
1310 | } |
1311 | ||
8a731799 | 1312 | void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time) |
e98be2d5 | 1313 | { |
2bc00aef TH |
1314 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
1315 | ||
1316 | __wb_update_bandwidth(&gdtc, start_time, false); | |
e98be2d5 WF |
1317 | } |
1318 | ||
9d823e8f | 1319 | /* |
d0e1d66b | 1320 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited() |
9d823e8f WF |
1321 | * will look to see if it needs to start dirty throttling. |
1322 | * | |
1323 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive | |
1324 | * global_page_state() too often. So scale it near-sqrt to the safety margin | |
1325 | * (the number of pages we may dirty without exceeding the dirty limits). | |
1326 | */ | |
1327 | static unsigned long dirty_poll_interval(unsigned long dirty, | |
1328 | unsigned long thresh) | |
1329 | { | |
1330 | if (thresh > dirty) | |
1331 | return 1UL << (ilog2(thresh - dirty) >> 1); | |
1332 | ||
1333 | return 1; | |
1334 | } | |
1335 | ||
a88a341a | 1336 | static unsigned long wb_max_pause(struct bdi_writeback *wb, |
de1fff37 | 1337 | unsigned long wb_dirty) |
c8462cc9 | 1338 | { |
a88a341a | 1339 | unsigned long bw = wb->avg_write_bandwidth; |
e3b6c655 | 1340 | unsigned long t; |
c8462cc9 | 1341 | |
7ccb9ad5 WF |
1342 | /* |
1343 | * Limit pause time for small memory systems. If sleeping for too long | |
1344 | * time, a small pool of dirty/writeback pages may go empty and disk go | |
1345 | * idle. | |
1346 | * | |
1347 | * 8 serves as the safety ratio. | |
1348 | */ | |
de1fff37 | 1349 | t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8)); |
7ccb9ad5 WF |
1350 | t++; |
1351 | ||
e3b6c655 | 1352 | return min_t(unsigned long, t, MAX_PAUSE); |
7ccb9ad5 WF |
1353 | } |
1354 | ||
a88a341a TH |
1355 | static long wb_min_pause(struct bdi_writeback *wb, |
1356 | long max_pause, | |
1357 | unsigned long task_ratelimit, | |
1358 | unsigned long dirty_ratelimit, | |
1359 | int *nr_dirtied_pause) | |
c8462cc9 | 1360 | { |
a88a341a TH |
1361 | long hi = ilog2(wb->avg_write_bandwidth); |
1362 | long lo = ilog2(wb->dirty_ratelimit); | |
7ccb9ad5 WF |
1363 | long t; /* target pause */ |
1364 | long pause; /* estimated next pause */ | |
1365 | int pages; /* target nr_dirtied_pause */ | |
c8462cc9 | 1366 | |
7ccb9ad5 WF |
1367 | /* target for 10ms pause on 1-dd case */ |
1368 | t = max(1, HZ / 100); | |
c8462cc9 WF |
1369 | |
1370 | /* | |
1371 | * Scale up pause time for concurrent dirtiers in order to reduce CPU | |
1372 | * overheads. | |
1373 | * | |
7ccb9ad5 | 1374 | * (N * 10ms) on 2^N concurrent tasks. |
c8462cc9 WF |
1375 | */ |
1376 | if (hi > lo) | |
7ccb9ad5 | 1377 | t += (hi - lo) * (10 * HZ) / 1024; |
c8462cc9 WF |
1378 | |
1379 | /* | |
7ccb9ad5 WF |
1380 | * This is a bit convoluted. We try to base the next nr_dirtied_pause |
1381 | * on the much more stable dirty_ratelimit. However the next pause time | |
1382 | * will be computed based on task_ratelimit and the two rate limits may | |
1383 | * depart considerably at some time. Especially if task_ratelimit goes | |
1384 | * below dirty_ratelimit/2 and the target pause is max_pause, the next | |
1385 | * pause time will be max_pause*2 _trimmed down_ to max_pause. As a | |
1386 | * result task_ratelimit won't be executed faithfully, which could | |
1387 | * eventually bring down dirty_ratelimit. | |
c8462cc9 | 1388 | * |
7ccb9ad5 WF |
1389 | * We apply two rules to fix it up: |
1390 | * 1) try to estimate the next pause time and if necessary, use a lower | |
1391 | * nr_dirtied_pause so as not to exceed max_pause. When this happens, | |
1392 | * nr_dirtied_pause will be "dancing" with task_ratelimit. | |
1393 | * 2) limit the target pause time to max_pause/2, so that the normal | |
1394 | * small fluctuations of task_ratelimit won't trigger rule (1) and | |
1395 | * nr_dirtied_pause will remain as stable as dirty_ratelimit. | |
c8462cc9 | 1396 | */ |
7ccb9ad5 WF |
1397 | t = min(t, 1 + max_pause / 2); |
1398 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
c8462cc9 WF |
1399 | |
1400 | /* | |
5b9b3574 WF |
1401 | * Tiny nr_dirtied_pause is found to hurt I/O performance in the test |
1402 | * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}. | |
1403 | * When the 16 consecutive reads are often interrupted by some dirty | |
1404 | * throttling pause during the async writes, cfq will go into idles | |
1405 | * (deadline is fine). So push nr_dirtied_pause as high as possible | |
1406 | * until reaches DIRTY_POLL_THRESH=32 pages. | |
c8462cc9 | 1407 | */ |
5b9b3574 WF |
1408 | if (pages < DIRTY_POLL_THRESH) { |
1409 | t = max_pause; | |
1410 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
1411 | if (pages > DIRTY_POLL_THRESH) { | |
1412 | pages = DIRTY_POLL_THRESH; | |
1413 | t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit; | |
1414 | } | |
1415 | } | |
1416 | ||
7ccb9ad5 WF |
1417 | pause = HZ * pages / (task_ratelimit + 1); |
1418 | if (pause > max_pause) { | |
1419 | t = max_pause; | |
1420 | pages = task_ratelimit * t / roundup_pow_of_two(HZ); | |
1421 | } | |
c8462cc9 | 1422 | |
7ccb9ad5 | 1423 | *nr_dirtied_pause = pages; |
c8462cc9 | 1424 | /* |
7ccb9ad5 | 1425 | * The minimal pause time will normally be half the target pause time. |
c8462cc9 | 1426 | */ |
5b9b3574 | 1427 | return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t; |
c8462cc9 WF |
1428 | } |
1429 | ||
970fb01a | 1430 | static inline void wb_dirty_limits(struct dirty_throttle_control *dtc) |
5a537485 | 1431 | { |
2bc00aef | 1432 | struct bdi_writeback *wb = dtc->wb; |
93f78d88 | 1433 | unsigned long wb_reclaimable; |
5a537485 MP |
1434 | |
1435 | /* | |
de1fff37 | 1436 | * wb_thresh is not treated as some limiting factor as |
5a537485 | 1437 | * dirty_thresh, due to reasons |
de1fff37 | 1438 | * - in JBOD setup, wb_thresh can fluctuate a lot |
5a537485 | 1439 | * - in a system with HDD and USB key, the USB key may somehow |
de1fff37 TH |
1440 | * go into state (wb_dirty >> wb_thresh) either because |
1441 | * wb_dirty starts high, or because wb_thresh drops low. | |
5a537485 | 1442 | * In this case we don't want to hard throttle the USB key |
de1fff37 TH |
1443 | * dirtiers for 100 seconds until wb_dirty drops under |
1444 | * wb_thresh. Instead the auxiliary wb control line in | |
a88a341a | 1445 | * wb_position_ratio() will let the dirtier task progress |
de1fff37 | 1446 | * at some rate <= (write_bw / 2) for bringing down wb_dirty. |
5a537485 | 1447 | */ |
b1cbc6d4 | 1448 | dtc->wb_thresh = __wb_calc_thresh(dtc); |
970fb01a TH |
1449 | dtc->wb_bg_thresh = dtc->thresh ? |
1450 | div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0; | |
5a537485 MP |
1451 | |
1452 | /* | |
1453 | * In order to avoid the stacked BDI deadlock we need | |
1454 | * to ensure we accurately count the 'dirty' pages when | |
1455 | * the threshold is low. | |
1456 | * | |
1457 | * Otherwise it would be possible to get thresh+n pages | |
1458 | * reported dirty, even though there are thresh-m pages | |
1459 | * actually dirty; with m+n sitting in the percpu | |
1460 | * deltas. | |
1461 | */ | |
2bc00aef | 1462 | if (dtc->wb_thresh < 2 * wb_stat_error(wb)) { |
93f78d88 | 1463 | wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); |
2bc00aef | 1464 | dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK); |
5a537485 | 1465 | } else { |
93f78d88 | 1466 | wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE); |
2bc00aef | 1467 | dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK); |
5a537485 MP |
1468 | } |
1469 | } | |
1470 | ||
1da177e4 LT |
1471 | /* |
1472 | * balance_dirty_pages() must be called by processes which are generating dirty | |
1473 | * data. It looks at the number of dirty pages in the machine and will force | |
143dfe86 | 1474 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. |
5b0830cb JA |
1475 | * If we're over `background_thresh' then the writeback threads are woken to |
1476 | * perform some writeout. | |
1da177e4 | 1477 | */ |
3a2e9a5a | 1478 | static void balance_dirty_pages(struct address_space *mapping, |
dfb8ae56 | 1479 | struct bdi_writeback *wb, |
143dfe86 | 1480 | unsigned long pages_dirtied) |
1da177e4 | 1481 | { |
2bc00aef TH |
1482 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) }; |
1483 | struct dirty_throttle_control * const gdtc = &gdtc_stor; | |
143dfe86 | 1484 | unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */ |
83712358 | 1485 | long period; |
7ccb9ad5 WF |
1486 | long pause; |
1487 | long max_pause; | |
1488 | long min_pause; | |
1489 | int nr_dirtied_pause; | |
e50e3720 | 1490 | bool dirty_exceeded = false; |
143dfe86 | 1491 | unsigned long task_ratelimit; |
7ccb9ad5 | 1492 | unsigned long dirty_ratelimit; |
dfb8ae56 | 1493 | struct backing_dev_info *bdi = wb->bdi; |
5a537485 | 1494 | bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT; |
e98be2d5 | 1495 | unsigned long start_time = jiffies; |
1da177e4 LT |
1496 | |
1497 | for (;;) { | |
83712358 | 1498 | unsigned long now = jiffies; |
2bc00aef | 1499 | unsigned long dirty, thresh, bg_thresh; |
83712358 | 1500 | |
143dfe86 WF |
1501 | /* |
1502 | * Unstable writes are a feature of certain networked | |
1503 | * filesystems (i.e. NFS) in which data may have been | |
1504 | * written to the server's write cache, but has not yet | |
1505 | * been flushed to permanent storage. | |
1506 | */ | |
5fce25a9 PZ |
1507 | nr_reclaimable = global_page_state(NR_FILE_DIRTY) + |
1508 | global_page_state(NR_UNSTABLE_NFS); | |
9fc3a43e | 1509 | gdtc->avail = global_dirtyable_memory(); |
2bc00aef | 1510 | gdtc->dirty = nr_reclaimable + global_page_state(NR_WRITEBACK); |
5fce25a9 | 1511 | |
9fc3a43e | 1512 | domain_dirty_limits(gdtc); |
16c4042f | 1513 | |
5a537485 | 1514 | if (unlikely(strictlimit)) { |
970fb01a | 1515 | wb_dirty_limits(gdtc); |
5a537485 | 1516 | |
2bc00aef TH |
1517 | dirty = gdtc->wb_dirty; |
1518 | thresh = gdtc->wb_thresh; | |
970fb01a | 1519 | bg_thresh = gdtc->wb_bg_thresh; |
5a537485 | 1520 | } else { |
2bc00aef TH |
1521 | dirty = gdtc->dirty; |
1522 | thresh = gdtc->thresh; | |
1523 | bg_thresh = gdtc->bg_thresh; | |
5a537485 MP |
1524 | } |
1525 | ||
16c4042f WF |
1526 | /* |
1527 | * Throttle it only when the background writeback cannot | |
1528 | * catch-up. This avoids (excessively) small writeouts | |
de1fff37 | 1529 | * when the wb limits are ramping up in case of !strictlimit. |
5a537485 | 1530 | * |
de1fff37 TH |
1531 | * In strictlimit case make decision based on the wb counters |
1532 | * and limits. Small writeouts when the wb limits are ramping | |
5a537485 | 1533 | * up are the price we consciously pay for strictlimit-ing. |
16c4042f | 1534 | */ |
5a537485 | 1535 | if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) { |
83712358 WF |
1536 | current->dirty_paused_when = now; |
1537 | current->nr_dirtied = 0; | |
7ccb9ad5 | 1538 | current->nr_dirtied_pause = |
5a537485 | 1539 | dirty_poll_interval(dirty, thresh); |
16c4042f | 1540 | break; |
83712358 | 1541 | } |
16c4042f | 1542 | |
bc05873d | 1543 | if (unlikely(!writeback_in_progress(wb))) |
9ecf4866 | 1544 | wb_start_background_writeback(wb); |
143dfe86 | 1545 | |
5a537485 | 1546 | if (!strictlimit) |
970fb01a | 1547 | wb_dirty_limits(gdtc); |
5fce25a9 | 1548 | |
2bc00aef TH |
1549 | dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) && |
1550 | ((gdtc->dirty > gdtc->thresh) || strictlimit); | |
daddfa3c TH |
1551 | |
1552 | wb_position_ratio(gdtc); | |
1553 | ||
a88a341a TH |
1554 | if (dirty_exceeded && !wb->dirty_exceeded) |
1555 | wb->dirty_exceeded = 1; | |
1da177e4 | 1556 | |
8a731799 TH |
1557 | if (time_is_before_jiffies(wb->bw_time_stamp + |
1558 | BANDWIDTH_INTERVAL)) { | |
1559 | spin_lock(&wb->list_lock); | |
2bc00aef | 1560 | __wb_update_bandwidth(gdtc, start_time, true); |
8a731799 TH |
1561 | spin_unlock(&wb->list_lock); |
1562 | } | |
e98be2d5 | 1563 | |
a88a341a | 1564 | dirty_ratelimit = wb->dirty_ratelimit; |
daddfa3c | 1565 | task_ratelimit = ((u64)dirty_ratelimit * gdtc->pos_ratio) >> |
3a73dbbc | 1566 | RATELIMIT_CALC_SHIFT; |
2bc00aef | 1567 | max_pause = wb_max_pause(wb, gdtc->wb_dirty); |
a88a341a TH |
1568 | min_pause = wb_min_pause(wb, max_pause, |
1569 | task_ratelimit, dirty_ratelimit, | |
1570 | &nr_dirtied_pause); | |
7ccb9ad5 | 1571 | |
3a73dbbc | 1572 | if (unlikely(task_ratelimit == 0)) { |
83712358 | 1573 | period = max_pause; |
c8462cc9 | 1574 | pause = max_pause; |
143dfe86 | 1575 | goto pause; |
04fbfdc1 | 1576 | } |
83712358 WF |
1577 | period = HZ * pages_dirtied / task_ratelimit; |
1578 | pause = period; | |
1579 | if (current->dirty_paused_when) | |
1580 | pause -= now - current->dirty_paused_when; | |
1581 | /* | |
1582 | * For less than 1s think time (ext3/4 may block the dirtier | |
1583 | * for up to 800ms from time to time on 1-HDD; so does xfs, | |
1584 | * however at much less frequency), try to compensate it in | |
1585 | * future periods by updating the virtual time; otherwise just | |
1586 | * do a reset, as it may be a light dirtier. | |
1587 | */ | |
7ccb9ad5 | 1588 | if (pause < min_pause) { |
ece13ac3 | 1589 | trace_balance_dirty_pages(bdi, |
2bc00aef TH |
1590 | gdtc->thresh, |
1591 | gdtc->bg_thresh, | |
1592 | gdtc->dirty, | |
1593 | gdtc->wb_thresh, | |
1594 | gdtc->wb_dirty, | |
ece13ac3 WF |
1595 | dirty_ratelimit, |
1596 | task_ratelimit, | |
1597 | pages_dirtied, | |
83712358 | 1598 | period, |
7ccb9ad5 | 1599 | min(pause, 0L), |
ece13ac3 | 1600 | start_time); |
83712358 WF |
1601 | if (pause < -HZ) { |
1602 | current->dirty_paused_when = now; | |
1603 | current->nr_dirtied = 0; | |
1604 | } else if (period) { | |
1605 | current->dirty_paused_when += period; | |
1606 | current->nr_dirtied = 0; | |
7ccb9ad5 WF |
1607 | } else if (current->nr_dirtied_pause <= pages_dirtied) |
1608 | current->nr_dirtied_pause += pages_dirtied; | |
57fc978c | 1609 | break; |
04fbfdc1 | 1610 | } |
7ccb9ad5 WF |
1611 | if (unlikely(pause > max_pause)) { |
1612 | /* for occasional dropped task_ratelimit */ | |
1613 | now += min(pause - max_pause, max_pause); | |
1614 | pause = max_pause; | |
1615 | } | |
143dfe86 WF |
1616 | |
1617 | pause: | |
ece13ac3 | 1618 | trace_balance_dirty_pages(bdi, |
2bc00aef TH |
1619 | gdtc->thresh, |
1620 | gdtc->bg_thresh, | |
1621 | gdtc->dirty, | |
1622 | gdtc->wb_thresh, | |
1623 | gdtc->wb_dirty, | |
ece13ac3 WF |
1624 | dirty_ratelimit, |
1625 | task_ratelimit, | |
1626 | pages_dirtied, | |
83712358 | 1627 | period, |
ece13ac3 WF |
1628 | pause, |
1629 | start_time); | |
499d05ec | 1630 | __set_current_state(TASK_KILLABLE); |
d25105e8 | 1631 | io_schedule_timeout(pause); |
87c6a9b2 | 1632 | |
83712358 WF |
1633 | current->dirty_paused_when = now + pause; |
1634 | current->nr_dirtied = 0; | |
7ccb9ad5 | 1635 | current->nr_dirtied_pause = nr_dirtied_pause; |
83712358 | 1636 | |
ffd1f609 | 1637 | /* |
2bc00aef TH |
1638 | * This is typically equal to (dirty < thresh) and can also |
1639 | * keep "1000+ dd on a slow USB stick" under control. | |
ffd1f609 | 1640 | */ |
1df64719 | 1641 | if (task_ratelimit) |
ffd1f609 | 1642 | break; |
499d05ec | 1643 | |
c5c6343c WF |
1644 | /* |
1645 | * In the case of an unresponding NFS server and the NFS dirty | |
de1fff37 | 1646 | * pages exceeds dirty_thresh, give the other good wb's a pipe |
c5c6343c WF |
1647 | * to go through, so that tasks on them still remain responsive. |
1648 | * | |
1649 | * In theory 1 page is enough to keep the comsumer-producer | |
1650 | * pipe going: the flusher cleans 1 page => the task dirties 1 | |
de1fff37 | 1651 | * more page. However wb_dirty has accounting errors. So use |
93f78d88 | 1652 | * the larger and more IO friendly wb_stat_error. |
c5c6343c | 1653 | */ |
2bc00aef | 1654 | if (gdtc->wb_dirty <= wb_stat_error(wb)) |
c5c6343c WF |
1655 | break; |
1656 | ||
499d05ec JK |
1657 | if (fatal_signal_pending(current)) |
1658 | break; | |
1da177e4 LT |
1659 | } |
1660 | ||
a88a341a TH |
1661 | if (!dirty_exceeded && wb->dirty_exceeded) |
1662 | wb->dirty_exceeded = 0; | |
1da177e4 | 1663 | |
bc05873d | 1664 | if (writeback_in_progress(wb)) |
5b0830cb | 1665 | return; |
1da177e4 LT |
1666 | |
1667 | /* | |
1668 | * In laptop mode, we wait until hitting the higher threshold before | |
1669 | * starting background writeout, and then write out all the way down | |
1670 | * to the lower threshold. So slow writers cause minimal disk activity. | |
1671 | * | |
1672 | * In normal mode, we start background writeout at the lower | |
1673 | * background_thresh, to keep the amount of dirty memory low. | |
1674 | */ | |
143dfe86 WF |
1675 | if (laptop_mode) |
1676 | return; | |
1677 | ||
2bc00aef | 1678 | if (nr_reclaimable > gdtc->bg_thresh) |
9ecf4866 | 1679 | wb_start_background_writeback(wb); |
1da177e4 LT |
1680 | } |
1681 | ||
9d823e8f | 1682 | static DEFINE_PER_CPU(int, bdp_ratelimits); |
245b2e70 | 1683 | |
54848d73 WF |
1684 | /* |
1685 | * Normal tasks are throttled by | |
1686 | * loop { | |
1687 | * dirty tsk->nr_dirtied_pause pages; | |
1688 | * take a snap in balance_dirty_pages(); | |
1689 | * } | |
1690 | * However there is a worst case. If every task exit immediately when dirtied | |
1691 | * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be | |
1692 | * called to throttle the page dirties. The solution is to save the not yet | |
1693 | * throttled page dirties in dirty_throttle_leaks on task exit and charge them | |
1694 | * randomly into the running tasks. This works well for the above worst case, | |
1695 | * as the new task will pick up and accumulate the old task's leaked dirty | |
1696 | * count and eventually get throttled. | |
1697 | */ | |
1698 | DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0; | |
1699 | ||
1da177e4 | 1700 | /** |
d0e1d66b | 1701 | * balance_dirty_pages_ratelimited - balance dirty memory state |
67be2dd1 | 1702 | * @mapping: address_space which was dirtied |
1da177e4 LT |
1703 | * |
1704 | * Processes which are dirtying memory should call in here once for each page | |
1705 | * which was newly dirtied. The function will periodically check the system's | |
1706 | * dirty state and will initiate writeback if needed. | |
1707 | * | |
1708 | * On really big machines, get_writeback_state is expensive, so try to avoid | |
1709 | * calling it too often (ratelimiting). But once we're over the dirty memory | |
1710 | * limit we decrease the ratelimiting by a lot, to prevent individual processes | |
1711 | * from overshooting the limit by (ratelimit_pages) each. | |
1712 | */ | |
d0e1d66b | 1713 | void balance_dirty_pages_ratelimited(struct address_space *mapping) |
1da177e4 | 1714 | { |
dfb8ae56 TH |
1715 | struct inode *inode = mapping->host; |
1716 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1717 | struct bdi_writeback *wb = NULL; | |
9d823e8f WF |
1718 | int ratelimit; |
1719 | int *p; | |
1da177e4 | 1720 | |
36715cef WF |
1721 | if (!bdi_cap_account_dirty(bdi)) |
1722 | return; | |
1723 | ||
dfb8ae56 TH |
1724 | if (inode_cgwb_enabled(inode)) |
1725 | wb = wb_get_create_current(bdi, GFP_KERNEL); | |
1726 | if (!wb) | |
1727 | wb = &bdi->wb; | |
1728 | ||
9d823e8f | 1729 | ratelimit = current->nr_dirtied_pause; |
a88a341a | 1730 | if (wb->dirty_exceeded) |
9d823e8f WF |
1731 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); |
1732 | ||
9d823e8f | 1733 | preempt_disable(); |
1da177e4 | 1734 | /* |
9d823e8f WF |
1735 | * This prevents one CPU to accumulate too many dirtied pages without |
1736 | * calling into balance_dirty_pages(), which can happen when there are | |
1737 | * 1000+ tasks, all of them start dirtying pages at exactly the same | |
1738 | * time, hence all honoured too large initial task->nr_dirtied_pause. | |
1da177e4 | 1739 | */ |
7c8e0181 | 1740 | p = this_cpu_ptr(&bdp_ratelimits); |
9d823e8f | 1741 | if (unlikely(current->nr_dirtied >= ratelimit)) |
fa5a734e | 1742 | *p = 0; |
d3bc1fef WF |
1743 | else if (unlikely(*p >= ratelimit_pages)) { |
1744 | *p = 0; | |
1745 | ratelimit = 0; | |
1da177e4 | 1746 | } |
54848d73 WF |
1747 | /* |
1748 | * Pick up the dirtied pages by the exited tasks. This avoids lots of | |
1749 | * short-lived tasks (eg. gcc invocations in a kernel build) escaping | |
1750 | * the dirty throttling and livelock other long-run dirtiers. | |
1751 | */ | |
7c8e0181 | 1752 | p = this_cpu_ptr(&dirty_throttle_leaks); |
54848d73 | 1753 | if (*p > 0 && current->nr_dirtied < ratelimit) { |
d0e1d66b | 1754 | unsigned long nr_pages_dirtied; |
54848d73 WF |
1755 | nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied); |
1756 | *p -= nr_pages_dirtied; | |
1757 | current->nr_dirtied += nr_pages_dirtied; | |
1da177e4 | 1758 | } |
fa5a734e | 1759 | preempt_enable(); |
9d823e8f WF |
1760 | |
1761 | if (unlikely(current->nr_dirtied >= ratelimit)) | |
dfb8ae56 TH |
1762 | balance_dirty_pages(mapping, wb, current->nr_dirtied); |
1763 | ||
1764 | wb_put(wb); | |
1da177e4 | 1765 | } |
d0e1d66b | 1766 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited); |
1da177e4 | 1767 | |
aa661bbe TH |
1768 | /** |
1769 | * wb_over_bg_thresh - does @wb need to be written back? | |
1770 | * @wb: bdi_writeback of interest | |
1771 | * | |
1772 | * Determines whether background writeback should keep writing @wb or it's | |
1773 | * clean enough. Returns %true if writeback should continue. | |
1774 | */ | |
1775 | bool wb_over_bg_thresh(struct bdi_writeback *wb) | |
1776 | { | |
947e9762 TH |
1777 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) }; |
1778 | struct dirty_throttle_control * const gdtc = &gdtc_stor; | |
aa661bbe | 1779 | |
947e9762 TH |
1780 | /* |
1781 | * Similar to balance_dirty_pages() but ignores pages being written | |
1782 | * as we're trying to decide whether to put more under writeback. | |
1783 | */ | |
1784 | gdtc->avail = global_dirtyable_memory(); | |
1785 | gdtc->dirty = global_page_state(NR_FILE_DIRTY) + | |
1786 | global_page_state(NR_UNSTABLE_NFS); | |
1787 | domain_dirty_limits(gdtc); | |
aa661bbe | 1788 | |
947e9762 | 1789 | if (gdtc->dirty > gdtc->bg_thresh) |
aa661bbe TH |
1790 | return true; |
1791 | ||
947e9762 | 1792 | if (wb_stat(wb, WB_RECLAIMABLE) > __wb_calc_thresh(gdtc)) |
aa661bbe TH |
1793 | return true; |
1794 | ||
1795 | return false; | |
1796 | } | |
1797 | ||
232ea4d6 | 1798 | void throttle_vm_writeout(gfp_t gfp_mask) |
1da177e4 | 1799 | { |
364aeb28 DR |
1800 | unsigned long background_thresh; |
1801 | unsigned long dirty_thresh; | |
1da177e4 LT |
1802 | |
1803 | for ( ; ; ) { | |
16c4042f | 1804 | global_dirty_limits(&background_thresh, &dirty_thresh); |
c7981433 | 1805 | dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh); |
1da177e4 LT |
1806 | |
1807 | /* | |
1808 | * Boost the allowable dirty threshold a bit for page | |
1809 | * allocators so they don't get DoS'ed by heavy writers | |
1810 | */ | |
1811 | dirty_thresh += dirty_thresh / 10; /* wheeee... */ | |
1812 | ||
c24f21bd CL |
1813 | if (global_page_state(NR_UNSTABLE_NFS) + |
1814 | global_page_state(NR_WRITEBACK) <= dirty_thresh) | |
1815 | break; | |
8aa7e847 | 1816 | congestion_wait(BLK_RW_ASYNC, HZ/10); |
369f2389 FW |
1817 | |
1818 | /* | |
1819 | * The caller might hold locks which can prevent IO completion | |
1820 | * or progress in the filesystem. So we cannot just sit here | |
1821 | * waiting for IO to complete. | |
1822 | */ | |
1823 | if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) | |
1824 | break; | |
1da177e4 LT |
1825 | } |
1826 | } | |
1827 | ||
1da177e4 LT |
1828 | /* |
1829 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
1830 | */ | |
cccad5b9 | 1831 | int dirty_writeback_centisecs_handler(struct ctl_table *table, int write, |
8d65af78 | 1832 | void __user *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 1833 | { |
8d65af78 | 1834 | proc_dointvec(table, write, buffer, length, ppos); |
1da177e4 LT |
1835 | return 0; |
1836 | } | |
1837 | ||
c2c4986e | 1838 | #ifdef CONFIG_BLOCK |
31373d09 | 1839 | void laptop_mode_timer_fn(unsigned long data) |
1da177e4 | 1840 | { |
31373d09 MG |
1841 | struct request_queue *q = (struct request_queue *)data; |
1842 | int nr_pages = global_page_state(NR_FILE_DIRTY) + | |
1843 | global_page_state(NR_UNSTABLE_NFS); | |
a06fd6b1 TH |
1844 | struct bdi_writeback *wb; |
1845 | struct wb_iter iter; | |
1da177e4 | 1846 | |
31373d09 MG |
1847 | /* |
1848 | * We want to write everything out, not just down to the dirty | |
1849 | * threshold | |
1850 | */ | |
a06fd6b1 TH |
1851 | if (!bdi_has_dirty_io(&q->backing_dev_info)) |
1852 | return; | |
1853 | ||
1854 | bdi_for_each_wb(wb, &q->backing_dev_info, &iter, 0) | |
1855 | if (wb_has_dirty_io(wb)) | |
1856 | wb_start_writeback(wb, nr_pages, true, | |
1857 | WB_REASON_LAPTOP_TIMER); | |
1da177e4 LT |
1858 | } |
1859 | ||
1860 | /* | |
1861 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
1862 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
1863 | * then push it back - the user is still using the disk. | |
1864 | */ | |
31373d09 | 1865 | void laptop_io_completion(struct backing_dev_info *info) |
1da177e4 | 1866 | { |
31373d09 | 1867 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
1868 | } |
1869 | ||
1870 | /* | |
1871 | * We're in laptop mode and we've just synced. The sync's writes will have | |
1872 | * caused another writeback to be scheduled by laptop_io_completion. | |
1873 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
1874 | */ | |
1875 | void laptop_sync_completion(void) | |
1876 | { | |
31373d09 MG |
1877 | struct backing_dev_info *bdi; |
1878 | ||
1879 | rcu_read_lock(); | |
1880 | ||
1881 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) | |
1882 | del_timer(&bdi->laptop_mode_wb_timer); | |
1883 | ||
1884 | rcu_read_unlock(); | |
1da177e4 | 1885 | } |
c2c4986e | 1886 | #endif |
1da177e4 LT |
1887 | |
1888 | /* | |
1889 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
1890 | * if a large number of processes all perform writes at the same time. | |
1891 | * If it is too low then SMP machines will call the (expensive) | |
1892 | * get_writeback_state too often. | |
1893 | * | |
1894 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
1895 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
9d823e8f | 1896 | * thresholds. |
1da177e4 LT |
1897 | */ |
1898 | ||
2d1d43f6 | 1899 | void writeback_set_ratelimit(void) |
1da177e4 | 1900 | { |
dcc25ae7 | 1901 | struct wb_domain *dom = &global_wb_domain; |
9d823e8f WF |
1902 | unsigned long background_thresh; |
1903 | unsigned long dirty_thresh; | |
dcc25ae7 | 1904 | |
9d823e8f | 1905 | global_dirty_limits(&background_thresh, &dirty_thresh); |
dcc25ae7 | 1906 | dom->dirty_limit = dirty_thresh; |
9d823e8f | 1907 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); |
1da177e4 LT |
1908 | if (ratelimit_pages < 16) |
1909 | ratelimit_pages = 16; | |
1da177e4 LT |
1910 | } |
1911 | ||
0db0628d | 1912 | static int |
2f60d628 SB |
1913 | ratelimit_handler(struct notifier_block *self, unsigned long action, |
1914 | void *hcpu) | |
1da177e4 | 1915 | { |
2f60d628 SB |
1916 | |
1917 | switch (action & ~CPU_TASKS_FROZEN) { | |
1918 | case CPU_ONLINE: | |
1919 | case CPU_DEAD: | |
1920 | writeback_set_ratelimit(); | |
1921 | return NOTIFY_OK; | |
1922 | default: | |
1923 | return NOTIFY_DONE; | |
1924 | } | |
1da177e4 LT |
1925 | } |
1926 | ||
0db0628d | 1927 | static struct notifier_block ratelimit_nb = { |
1da177e4 LT |
1928 | .notifier_call = ratelimit_handler, |
1929 | .next = NULL, | |
1930 | }; | |
1931 | ||
1932 | /* | |
dc6e29da LT |
1933 | * Called early on to tune the page writeback dirty limits. |
1934 | * | |
1935 | * We used to scale dirty pages according to how total memory | |
1936 | * related to pages that could be allocated for buffers (by | |
1937 | * comparing nr_free_buffer_pages() to vm_total_pages. | |
1938 | * | |
1939 | * However, that was when we used "dirty_ratio" to scale with | |
1940 | * all memory, and we don't do that any more. "dirty_ratio" | |
1941 | * is now applied to total non-HIGHPAGE memory (by subtracting | |
1942 | * totalhigh_pages from vm_total_pages), and as such we can't | |
1943 | * get into the old insane situation any more where we had | |
1944 | * large amounts of dirty pages compared to a small amount of | |
1945 | * non-HIGHMEM memory. | |
1946 | * | |
1947 | * But we might still want to scale the dirty_ratio by how | |
1948 | * much memory the box has.. | |
1da177e4 LT |
1949 | */ |
1950 | void __init page_writeback_init(void) | |
1951 | { | |
2d1d43f6 | 1952 | writeback_set_ratelimit(); |
1da177e4 | 1953 | register_cpu_notifier(&ratelimit_nb); |
04fbfdc1 | 1954 | |
380c27ca | 1955 | BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL)); |
1da177e4 LT |
1956 | } |
1957 | ||
f446daae JK |
1958 | /** |
1959 | * tag_pages_for_writeback - tag pages to be written by write_cache_pages | |
1960 | * @mapping: address space structure to write | |
1961 | * @start: starting page index | |
1962 | * @end: ending page index (inclusive) | |
1963 | * | |
1964 | * This function scans the page range from @start to @end (inclusive) and tags | |
1965 | * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is | |
1966 | * that write_cache_pages (or whoever calls this function) will then use | |
1967 | * TOWRITE tag to identify pages eligible for writeback. This mechanism is | |
1968 | * used to avoid livelocking of writeback by a process steadily creating new | |
1969 | * dirty pages in the file (thus it is important for this function to be quick | |
1970 | * so that it can tag pages faster than a dirtying process can create them). | |
1971 | */ | |
1972 | /* | |
1973 | * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency. | |
1974 | */ | |
f446daae JK |
1975 | void tag_pages_for_writeback(struct address_space *mapping, |
1976 | pgoff_t start, pgoff_t end) | |
1977 | { | |
3c111a07 | 1978 | #define WRITEBACK_TAG_BATCH 4096 |
f446daae JK |
1979 | unsigned long tagged; |
1980 | ||
1981 | do { | |
1982 | spin_lock_irq(&mapping->tree_lock); | |
1983 | tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree, | |
1984 | &start, end, WRITEBACK_TAG_BATCH, | |
1985 | PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE); | |
1986 | spin_unlock_irq(&mapping->tree_lock); | |
1987 | WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH); | |
1988 | cond_resched(); | |
d5ed3a4a JK |
1989 | /* We check 'start' to handle wrapping when end == ~0UL */ |
1990 | } while (tagged >= WRITEBACK_TAG_BATCH && start); | |
f446daae JK |
1991 | } |
1992 | EXPORT_SYMBOL(tag_pages_for_writeback); | |
1993 | ||
811d736f | 1994 | /** |
0ea97180 | 1995 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. |
811d736f DH |
1996 | * @mapping: address space structure to write |
1997 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
0ea97180 MS |
1998 | * @writepage: function called for each page |
1999 | * @data: data passed to writepage function | |
811d736f | 2000 | * |
0ea97180 | 2001 | * If a page is already under I/O, write_cache_pages() skips it, even |
811d736f DH |
2002 | * if it's dirty. This is desirable behaviour for memory-cleaning writeback, |
2003 | * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() | |
2004 | * and msync() need to guarantee that all the data which was dirty at the time | |
2005 | * the call was made get new I/O started against them. If wbc->sync_mode is | |
2006 | * WB_SYNC_ALL then we were called for data integrity and we must wait for | |
2007 | * existing IO to complete. | |
f446daae JK |
2008 | * |
2009 | * To avoid livelocks (when other process dirties new pages), we first tag | |
2010 | * pages which should be written back with TOWRITE tag and only then start | |
2011 | * writing them. For data-integrity sync we have to be careful so that we do | |
2012 | * not miss some pages (e.g., because some other process has cleared TOWRITE | |
2013 | * tag we set). The rule we follow is that TOWRITE tag can be cleared only | |
2014 | * by the process clearing the DIRTY tag (and submitting the page for IO). | |
811d736f | 2015 | */ |
0ea97180 MS |
2016 | int write_cache_pages(struct address_space *mapping, |
2017 | struct writeback_control *wbc, writepage_t writepage, | |
2018 | void *data) | |
811d736f | 2019 | { |
811d736f DH |
2020 | int ret = 0; |
2021 | int done = 0; | |
811d736f DH |
2022 | struct pagevec pvec; |
2023 | int nr_pages; | |
31a12666 | 2024 | pgoff_t uninitialized_var(writeback_index); |
811d736f DH |
2025 | pgoff_t index; |
2026 | pgoff_t end; /* Inclusive */ | |
bd19e012 | 2027 | pgoff_t done_index; |
31a12666 | 2028 | int cycled; |
811d736f | 2029 | int range_whole = 0; |
f446daae | 2030 | int tag; |
811d736f | 2031 | |
811d736f DH |
2032 | pagevec_init(&pvec, 0); |
2033 | if (wbc->range_cyclic) { | |
31a12666 NP |
2034 | writeback_index = mapping->writeback_index; /* prev offset */ |
2035 | index = writeback_index; | |
2036 | if (index == 0) | |
2037 | cycled = 1; | |
2038 | else | |
2039 | cycled = 0; | |
811d736f DH |
2040 | end = -1; |
2041 | } else { | |
2042 | index = wbc->range_start >> PAGE_CACHE_SHIFT; | |
2043 | end = wbc->range_end >> PAGE_CACHE_SHIFT; | |
2044 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) | |
2045 | range_whole = 1; | |
31a12666 | 2046 | cycled = 1; /* ignore range_cyclic tests */ |
811d736f | 2047 | } |
6e6938b6 | 2048 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae JK |
2049 | tag = PAGECACHE_TAG_TOWRITE; |
2050 | else | |
2051 | tag = PAGECACHE_TAG_DIRTY; | |
811d736f | 2052 | retry: |
6e6938b6 | 2053 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
f446daae | 2054 | tag_pages_for_writeback(mapping, index, end); |
bd19e012 | 2055 | done_index = index; |
5a3d5c98 NP |
2056 | while (!done && (index <= end)) { |
2057 | int i; | |
2058 | ||
f446daae | 2059 | nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag, |
5a3d5c98 NP |
2060 | min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1); |
2061 | if (nr_pages == 0) | |
2062 | break; | |
811d736f | 2063 | |
811d736f DH |
2064 | for (i = 0; i < nr_pages; i++) { |
2065 | struct page *page = pvec.pages[i]; | |
2066 | ||
2067 | /* | |
d5482cdf NP |
2068 | * At this point, the page may be truncated or |
2069 | * invalidated (changing page->mapping to NULL), or | |
2070 | * even swizzled back from swapper_space to tmpfs file | |
2071 | * mapping. However, page->index will not change | |
2072 | * because we have a reference on the page. | |
811d736f | 2073 | */ |
d5482cdf NP |
2074 | if (page->index > end) { |
2075 | /* | |
2076 | * can't be range_cyclic (1st pass) because | |
2077 | * end == -1 in that case. | |
2078 | */ | |
2079 | done = 1; | |
2080 | break; | |
2081 | } | |
2082 | ||
cf15b07c | 2083 | done_index = page->index; |
d5482cdf | 2084 | |
811d736f DH |
2085 | lock_page(page); |
2086 | ||
5a3d5c98 NP |
2087 | /* |
2088 | * Page truncated or invalidated. We can freely skip it | |
2089 | * then, even for data integrity operations: the page | |
2090 | * has disappeared concurrently, so there could be no | |
2091 | * real expectation of this data interity operation | |
2092 | * even if there is now a new, dirty page at the same | |
2093 | * pagecache address. | |
2094 | */ | |
811d736f | 2095 | if (unlikely(page->mapping != mapping)) { |
5a3d5c98 | 2096 | continue_unlock: |
811d736f DH |
2097 | unlock_page(page); |
2098 | continue; | |
2099 | } | |
2100 | ||
515f4a03 NP |
2101 | if (!PageDirty(page)) { |
2102 | /* someone wrote it for us */ | |
2103 | goto continue_unlock; | |
2104 | } | |
2105 | ||
2106 | if (PageWriteback(page)) { | |
2107 | if (wbc->sync_mode != WB_SYNC_NONE) | |
2108 | wait_on_page_writeback(page); | |
2109 | else | |
2110 | goto continue_unlock; | |
2111 | } | |
811d736f | 2112 | |
515f4a03 NP |
2113 | BUG_ON(PageWriteback(page)); |
2114 | if (!clear_page_dirty_for_io(page)) | |
5a3d5c98 | 2115 | goto continue_unlock; |
811d736f | 2116 | |
de1414a6 | 2117 | trace_wbc_writepage(wbc, inode_to_bdi(mapping->host)); |
0ea97180 | 2118 | ret = (*writepage)(page, wbc, data); |
00266770 NP |
2119 | if (unlikely(ret)) { |
2120 | if (ret == AOP_WRITEPAGE_ACTIVATE) { | |
2121 | unlock_page(page); | |
2122 | ret = 0; | |
2123 | } else { | |
2124 | /* | |
2125 | * done_index is set past this page, | |
2126 | * so media errors will not choke | |
2127 | * background writeout for the entire | |
2128 | * file. This has consequences for | |
2129 | * range_cyclic semantics (ie. it may | |
2130 | * not be suitable for data integrity | |
2131 | * writeout). | |
2132 | */ | |
cf15b07c | 2133 | done_index = page->index + 1; |
00266770 NP |
2134 | done = 1; |
2135 | break; | |
2136 | } | |
0b564927 | 2137 | } |
00266770 | 2138 | |
546a1924 DC |
2139 | /* |
2140 | * We stop writing back only if we are not doing | |
2141 | * integrity sync. In case of integrity sync we have to | |
2142 | * keep going until we have written all the pages | |
2143 | * we tagged for writeback prior to entering this loop. | |
2144 | */ | |
2145 | if (--wbc->nr_to_write <= 0 && | |
2146 | wbc->sync_mode == WB_SYNC_NONE) { | |
2147 | done = 1; | |
2148 | break; | |
05fe478d | 2149 | } |
811d736f DH |
2150 | } |
2151 | pagevec_release(&pvec); | |
2152 | cond_resched(); | |
2153 | } | |
3a4c6800 | 2154 | if (!cycled && !done) { |
811d736f | 2155 | /* |
31a12666 | 2156 | * range_cyclic: |
811d736f DH |
2157 | * We hit the last page and there is more work to be done: wrap |
2158 | * back to the start of the file | |
2159 | */ | |
31a12666 | 2160 | cycled = 1; |
811d736f | 2161 | index = 0; |
31a12666 | 2162 | end = writeback_index - 1; |
811d736f DH |
2163 | goto retry; |
2164 | } | |
0b564927 DC |
2165 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) |
2166 | mapping->writeback_index = done_index; | |
06d6cf69 | 2167 | |
811d736f DH |
2168 | return ret; |
2169 | } | |
0ea97180 MS |
2170 | EXPORT_SYMBOL(write_cache_pages); |
2171 | ||
2172 | /* | |
2173 | * Function used by generic_writepages to call the real writepage | |
2174 | * function and set the mapping flags on error | |
2175 | */ | |
2176 | static int __writepage(struct page *page, struct writeback_control *wbc, | |
2177 | void *data) | |
2178 | { | |
2179 | struct address_space *mapping = data; | |
2180 | int ret = mapping->a_ops->writepage(page, wbc); | |
2181 | mapping_set_error(mapping, ret); | |
2182 | return ret; | |
2183 | } | |
2184 | ||
2185 | /** | |
2186 | * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. | |
2187 | * @mapping: address space structure to write | |
2188 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
2189 | * | |
2190 | * This is a library function, which implements the writepages() | |
2191 | * address_space_operation. | |
2192 | */ | |
2193 | int generic_writepages(struct address_space *mapping, | |
2194 | struct writeback_control *wbc) | |
2195 | { | |
9b6096a6 SL |
2196 | struct blk_plug plug; |
2197 | int ret; | |
2198 | ||
0ea97180 MS |
2199 | /* deal with chardevs and other special file */ |
2200 | if (!mapping->a_ops->writepage) | |
2201 | return 0; | |
2202 | ||
9b6096a6 SL |
2203 | blk_start_plug(&plug); |
2204 | ret = write_cache_pages(mapping, wbc, __writepage, mapping); | |
2205 | blk_finish_plug(&plug); | |
2206 | return ret; | |
0ea97180 | 2207 | } |
811d736f DH |
2208 | |
2209 | EXPORT_SYMBOL(generic_writepages); | |
2210 | ||
1da177e4 LT |
2211 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
2212 | { | |
22905f77 AM |
2213 | int ret; |
2214 | ||
1da177e4 LT |
2215 | if (wbc->nr_to_write <= 0) |
2216 | return 0; | |
2217 | if (mapping->a_ops->writepages) | |
d08b3851 | 2218 | ret = mapping->a_ops->writepages(mapping, wbc); |
22905f77 AM |
2219 | else |
2220 | ret = generic_writepages(mapping, wbc); | |
22905f77 | 2221 | return ret; |
1da177e4 LT |
2222 | } |
2223 | ||
2224 | /** | |
2225 | * write_one_page - write out a single page and optionally wait on I/O | |
67be2dd1 MW |
2226 | * @page: the page to write |
2227 | * @wait: if true, wait on writeout | |
1da177e4 LT |
2228 | * |
2229 | * The page must be locked by the caller and will be unlocked upon return. | |
2230 | * | |
2231 | * write_one_page() returns a negative error code if I/O failed. | |
2232 | */ | |
2233 | int write_one_page(struct page *page, int wait) | |
2234 | { | |
2235 | struct address_space *mapping = page->mapping; | |
2236 | int ret = 0; | |
2237 | struct writeback_control wbc = { | |
2238 | .sync_mode = WB_SYNC_ALL, | |
2239 | .nr_to_write = 1, | |
2240 | }; | |
2241 | ||
2242 | BUG_ON(!PageLocked(page)); | |
2243 | ||
2244 | if (wait) | |
2245 | wait_on_page_writeback(page); | |
2246 | ||
2247 | if (clear_page_dirty_for_io(page)) { | |
2248 | page_cache_get(page); | |
2249 | ret = mapping->a_ops->writepage(page, &wbc); | |
2250 | if (ret == 0 && wait) { | |
2251 | wait_on_page_writeback(page); | |
2252 | if (PageError(page)) | |
2253 | ret = -EIO; | |
2254 | } | |
2255 | page_cache_release(page); | |
2256 | } else { | |
2257 | unlock_page(page); | |
2258 | } | |
2259 | return ret; | |
2260 | } | |
2261 | EXPORT_SYMBOL(write_one_page); | |
2262 | ||
76719325 KC |
2263 | /* |
2264 | * For address_spaces which do not use buffers nor write back. | |
2265 | */ | |
2266 | int __set_page_dirty_no_writeback(struct page *page) | |
2267 | { | |
2268 | if (!PageDirty(page)) | |
c3f0da63 | 2269 | return !TestSetPageDirty(page); |
76719325 KC |
2270 | return 0; |
2271 | } | |
2272 | ||
e3a7cca1 ES |
2273 | /* |
2274 | * Helper function for set_page_dirty family. | |
c4843a75 GT |
2275 | * |
2276 | * Caller must hold mem_cgroup_begin_page_stat(). | |
2277 | * | |
e3a7cca1 ES |
2278 | * NOTE: This relies on being atomic wrt interrupts. |
2279 | */ | |
c4843a75 GT |
2280 | void account_page_dirtied(struct page *page, struct address_space *mapping, |
2281 | struct mem_cgroup *memcg) | |
e3a7cca1 | 2282 | { |
52ebea74 TH |
2283 | struct inode *inode = mapping->host; |
2284 | ||
9fb0a7da TH |
2285 | trace_writeback_dirty_page(page, mapping); |
2286 | ||
e3a7cca1 | 2287 | if (mapping_cap_account_dirty(mapping)) { |
52ebea74 TH |
2288 | struct bdi_writeback *wb; |
2289 | ||
2290 | inode_attach_wb(inode, page); | |
2291 | wb = inode_to_wb(inode); | |
de1414a6 | 2292 | |
c4843a75 | 2293 | mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_DIRTY); |
e3a7cca1 | 2294 | __inc_zone_page_state(page, NR_FILE_DIRTY); |
ea941f0e | 2295 | __inc_zone_page_state(page, NR_DIRTIED); |
52ebea74 TH |
2296 | __inc_wb_stat(wb, WB_RECLAIMABLE); |
2297 | __inc_wb_stat(wb, WB_DIRTIED); | |
e3a7cca1 | 2298 | task_io_account_write(PAGE_CACHE_SIZE); |
d3bc1fef WF |
2299 | current->nr_dirtied++; |
2300 | this_cpu_inc(bdp_ratelimits); | |
e3a7cca1 ES |
2301 | } |
2302 | } | |
679ceace | 2303 | EXPORT_SYMBOL(account_page_dirtied); |
e3a7cca1 | 2304 | |
b9ea2515 KK |
2305 | /* |
2306 | * Helper function for deaccounting dirty page without writeback. | |
c4843a75 GT |
2307 | * |
2308 | * Caller must hold mem_cgroup_begin_page_stat(). | |
b9ea2515 | 2309 | */ |
c4843a75 GT |
2310 | void account_page_cleaned(struct page *page, struct address_space *mapping, |
2311 | struct mem_cgroup *memcg) | |
b9ea2515 KK |
2312 | { |
2313 | if (mapping_cap_account_dirty(mapping)) { | |
c4843a75 | 2314 | mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY); |
b9ea2515 | 2315 | dec_zone_page_state(page, NR_FILE_DIRTY); |
91018134 | 2316 | dec_wb_stat(inode_to_wb(mapping->host), WB_RECLAIMABLE); |
b9ea2515 KK |
2317 | task_io_account_cancelled_write(PAGE_CACHE_SIZE); |
2318 | } | |
2319 | } | |
b9ea2515 | 2320 | |
1da177e4 LT |
2321 | /* |
2322 | * For address_spaces which do not use buffers. Just tag the page as dirty in | |
2323 | * its radix tree. | |
2324 | * | |
2325 | * This is also used when a single buffer is being dirtied: we want to set the | |
2326 | * page dirty in that case, but not all the buffers. This is a "bottom-up" | |
2327 | * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying. | |
2328 | * | |
2d6d7f98 JW |
2329 | * The caller must ensure this doesn't race with truncation. Most will simply |
2330 | * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and | |
2331 | * the pte lock held, which also locks out truncation. | |
1da177e4 LT |
2332 | */ |
2333 | int __set_page_dirty_nobuffers(struct page *page) | |
2334 | { | |
c4843a75 GT |
2335 | struct mem_cgroup *memcg; |
2336 | ||
2337 | memcg = mem_cgroup_begin_page_stat(page); | |
1da177e4 LT |
2338 | if (!TestSetPageDirty(page)) { |
2339 | struct address_space *mapping = page_mapping(page); | |
a85d9df1 | 2340 | unsigned long flags; |
1da177e4 | 2341 | |
c4843a75 GT |
2342 | if (!mapping) { |
2343 | mem_cgroup_end_page_stat(memcg); | |
8c08540f | 2344 | return 1; |
c4843a75 | 2345 | } |
8c08540f | 2346 | |
a85d9df1 | 2347 | spin_lock_irqsave(&mapping->tree_lock, flags); |
2d6d7f98 JW |
2348 | BUG_ON(page_mapping(page) != mapping); |
2349 | WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page)); | |
c4843a75 | 2350 | account_page_dirtied(page, mapping, memcg); |
2d6d7f98 JW |
2351 | radix_tree_tag_set(&mapping->page_tree, page_index(page), |
2352 | PAGECACHE_TAG_DIRTY); | |
a85d9df1 | 2353 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
c4843a75 GT |
2354 | mem_cgroup_end_page_stat(memcg); |
2355 | ||
8c08540f AM |
2356 | if (mapping->host) { |
2357 | /* !PageAnon && !swapper_space */ | |
2358 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 2359 | } |
4741c9fd | 2360 | return 1; |
1da177e4 | 2361 | } |
c4843a75 | 2362 | mem_cgroup_end_page_stat(memcg); |
4741c9fd | 2363 | return 0; |
1da177e4 LT |
2364 | } |
2365 | EXPORT_SYMBOL(__set_page_dirty_nobuffers); | |
2366 | ||
2f800fbd WF |
2367 | /* |
2368 | * Call this whenever redirtying a page, to de-account the dirty counters | |
2369 | * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written | |
2370 | * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to | |
2371 | * systematic errors in balanced_dirty_ratelimit and the dirty pages position | |
2372 | * control. | |
2373 | */ | |
2374 | void account_page_redirty(struct page *page) | |
2375 | { | |
2376 | struct address_space *mapping = page->mapping; | |
91018134 | 2377 | |
2f800fbd | 2378 | if (mapping && mapping_cap_account_dirty(mapping)) { |
91018134 TH |
2379 | struct bdi_writeback *wb = inode_to_wb(mapping->host); |
2380 | ||
2f800fbd WF |
2381 | current->nr_dirtied--; |
2382 | dec_zone_page_state(page, NR_DIRTIED); | |
91018134 | 2383 | dec_wb_stat(wb, WB_DIRTIED); |
2f800fbd WF |
2384 | } |
2385 | } | |
2386 | EXPORT_SYMBOL(account_page_redirty); | |
2387 | ||
1da177e4 LT |
2388 | /* |
2389 | * When a writepage implementation decides that it doesn't want to write this | |
2390 | * page for some reason, it should redirty the locked page via | |
2391 | * redirty_page_for_writepage() and it should then unlock the page and return 0 | |
2392 | */ | |
2393 | int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page) | |
2394 | { | |
8d38633c KK |
2395 | int ret; |
2396 | ||
1da177e4 | 2397 | wbc->pages_skipped++; |
8d38633c | 2398 | ret = __set_page_dirty_nobuffers(page); |
2f800fbd | 2399 | account_page_redirty(page); |
8d38633c | 2400 | return ret; |
1da177e4 LT |
2401 | } |
2402 | EXPORT_SYMBOL(redirty_page_for_writepage); | |
2403 | ||
2404 | /* | |
6746aff7 WF |
2405 | * Dirty a page. |
2406 | * | |
2407 | * For pages with a mapping this should be done under the page lock | |
2408 | * for the benefit of asynchronous memory errors who prefer a consistent | |
2409 | * dirty state. This rule can be broken in some special cases, | |
2410 | * but should be better not to. | |
2411 | * | |
1da177e4 LT |
2412 | * If the mapping doesn't provide a set_page_dirty a_op, then |
2413 | * just fall through and assume that it wants buffer_heads. | |
2414 | */ | |
1cf6e7d8 | 2415 | int set_page_dirty(struct page *page) |
1da177e4 LT |
2416 | { |
2417 | struct address_space *mapping = page_mapping(page); | |
2418 | ||
2419 | if (likely(mapping)) { | |
2420 | int (*spd)(struct page *) = mapping->a_ops->set_page_dirty; | |
278df9f4 MK |
2421 | /* |
2422 | * readahead/lru_deactivate_page could remain | |
2423 | * PG_readahead/PG_reclaim due to race with end_page_writeback | |
2424 | * About readahead, if the page is written, the flags would be | |
2425 | * reset. So no problem. | |
2426 | * About lru_deactivate_page, if the page is redirty, the flag | |
2427 | * will be reset. So no problem. but if the page is used by readahead | |
2428 | * it will confuse readahead and make it restart the size rampup | |
2429 | * process. But it's a trivial problem. | |
2430 | */ | |
a4bb3ecd NH |
2431 | if (PageReclaim(page)) |
2432 | ClearPageReclaim(page); | |
9361401e DH |
2433 | #ifdef CONFIG_BLOCK |
2434 | if (!spd) | |
2435 | spd = __set_page_dirty_buffers; | |
2436 | #endif | |
2437 | return (*spd)(page); | |
1da177e4 | 2438 | } |
4741c9fd AM |
2439 | if (!PageDirty(page)) { |
2440 | if (!TestSetPageDirty(page)) | |
2441 | return 1; | |
2442 | } | |
1da177e4 LT |
2443 | return 0; |
2444 | } | |
2445 | EXPORT_SYMBOL(set_page_dirty); | |
2446 | ||
2447 | /* | |
2448 | * set_page_dirty() is racy if the caller has no reference against | |
2449 | * page->mapping->host, and if the page is unlocked. This is because another | |
2450 | * CPU could truncate the page off the mapping and then free the mapping. | |
2451 | * | |
2452 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
2453 | * holds a reference on the inode by having an open file. | |
2454 | * | |
2455 | * In other cases, the page should be locked before running set_page_dirty(). | |
2456 | */ | |
2457 | int set_page_dirty_lock(struct page *page) | |
2458 | { | |
2459 | int ret; | |
2460 | ||
7eaceacc | 2461 | lock_page(page); |
1da177e4 LT |
2462 | ret = set_page_dirty(page); |
2463 | unlock_page(page); | |
2464 | return ret; | |
2465 | } | |
2466 | EXPORT_SYMBOL(set_page_dirty_lock); | |
2467 | ||
11f81bec TH |
2468 | /* |
2469 | * This cancels just the dirty bit on the kernel page itself, it does NOT | |
2470 | * actually remove dirty bits on any mmap's that may be around. It also | |
2471 | * leaves the page tagged dirty, so any sync activity will still find it on | |
2472 | * the dirty lists, and in particular, clear_page_dirty_for_io() will still | |
2473 | * look at the dirty bits in the VM. | |
2474 | * | |
2475 | * Doing this should *normally* only ever be done when a page is truncated, | |
2476 | * and is not actually mapped anywhere at all. However, fs/buffer.c does | |
2477 | * this when it notices that somebody has cleaned out all the buffers on a | |
2478 | * page without actually doing it through the VM. Can you say "ext3 is | |
2479 | * horribly ugly"? Thought you could. | |
2480 | */ | |
2481 | void cancel_dirty_page(struct page *page) | |
2482 | { | |
c4843a75 GT |
2483 | struct address_space *mapping = page_mapping(page); |
2484 | ||
2485 | if (mapping_cap_account_dirty(mapping)) { | |
2486 | struct mem_cgroup *memcg; | |
2487 | ||
2488 | memcg = mem_cgroup_begin_page_stat(page); | |
2489 | ||
2490 | if (TestClearPageDirty(page)) | |
2491 | account_page_cleaned(page, mapping, memcg); | |
2492 | ||
2493 | mem_cgroup_end_page_stat(memcg); | |
2494 | } else { | |
2495 | ClearPageDirty(page); | |
2496 | } | |
11f81bec TH |
2497 | } |
2498 | EXPORT_SYMBOL(cancel_dirty_page); | |
2499 | ||
1da177e4 LT |
2500 | /* |
2501 | * Clear a page's dirty flag, while caring for dirty memory accounting. | |
2502 | * Returns true if the page was previously dirty. | |
2503 | * | |
2504 | * This is for preparing to put the page under writeout. We leave the page | |
2505 | * tagged as dirty in the radix tree so that a concurrent write-for-sync | |
2506 | * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage | |
2507 | * implementation will run either set_page_writeback() or set_page_dirty(), | |
2508 | * at which stage we bring the page's dirty flag and radix-tree dirty tag | |
2509 | * back into sync. | |
2510 | * | |
2511 | * This incoherency between the page's dirty flag and radix-tree tag is | |
2512 | * unfortunate, but it only exists while the page is locked. | |
2513 | */ | |
2514 | int clear_page_dirty_for_io(struct page *page) | |
2515 | { | |
2516 | struct address_space *mapping = page_mapping(page); | |
c4843a75 GT |
2517 | struct mem_cgroup *memcg; |
2518 | int ret = 0; | |
1da177e4 | 2519 | |
79352894 NP |
2520 | BUG_ON(!PageLocked(page)); |
2521 | ||
7658cc28 LT |
2522 | if (mapping && mapping_cap_account_dirty(mapping)) { |
2523 | /* | |
2524 | * Yes, Virginia, this is indeed insane. | |
2525 | * | |
2526 | * We use this sequence to make sure that | |
2527 | * (a) we account for dirty stats properly | |
2528 | * (b) we tell the low-level filesystem to | |
2529 | * mark the whole page dirty if it was | |
2530 | * dirty in a pagetable. Only to then | |
2531 | * (c) clean the page again and return 1 to | |
2532 | * cause the writeback. | |
2533 | * | |
2534 | * This way we avoid all nasty races with the | |
2535 | * dirty bit in multiple places and clearing | |
2536 | * them concurrently from different threads. | |
2537 | * | |
2538 | * Note! Normally the "set_page_dirty(page)" | |
2539 | * has no effect on the actual dirty bit - since | |
2540 | * that will already usually be set. But we | |
2541 | * need the side effects, and it can help us | |
2542 | * avoid races. | |
2543 | * | |
2544 | * We basically use the page "master dirty bit" | |
2545 | * as a serialization point for all the different | |
2546 | * threads doing their things. | |
7658cc28 LT |
2547 | */ |
2548 | if (page_mkclean(page)) | |
2549 | set_page_dirty(page); | |
79352894 NP |
2550 | /* |
2551 | * We carefully synchronise fault handlers against | |
2552 | * installing a dirty pte and marking the page dirty | |
2d6d7f98 JW |
2553 | * at this point. We do this by having them hold the |
2554 | * page lock while dirtying the page, and pages are | |
2555 | * always locked coming in here, so we get the desired | |
2556 | * exclusion. | |
79352894 | 2557 | */ |
c4843a75 | 2558 | memcg = mem_cgroup_begin_page_stat(page); |
7658cc28 | 2559 | if (TestClearPageDirty(page)) { |
c4843a75 | 2560 | mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_DIRTY); |
8c08540f | 2561 | dec_zone_page_state(page, NR_FILE_DIRTY); |
91018134 | 2562 | dec_wb_stat(inode_to_wb(mapping->host), WB_RECLAIMABLE); |
c4843a75 | 2563 | ret = 1; |
1da177e4 | 2564 | } |
c4843a75 GT |
2565 | mem_cgroup_end_page_stat(memcg); |
2566 | return ret; | |
1da177e4 | 2567 | } |
7658cc28 | 2568 | return TestClearPageDirty(page); |
1da177e4 | 2569 | } |
58bb01a9 | 2570 | EXPORT_SYMBOL(clear_page_dirty_for_io); |
1da177e4 LT |
2571 | |
2572 | int test_clear_page_writeback(struct page *page) | |
2573 | { | |
2574 | struct address_space *mapping = page_mapping(page); | |
d7365e78 | 2575 | struct mem_cgroup *memcg; |
d7365e78 | 2576 | int ret; |
1da177e4 | 2577 | |
6de22619 | 2578 | memcg = mem_cgroup_begin_page_stat(page); |
1da177e4 | 2579 | if (mapping) { |
91018134 TH |
2580 | struct inode *inode = mapping->host; |
2581 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 LT |
2582 | unsigned long flags; |
2583 | ||
19fd6231 | 2584 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 2585 | ret = TestClearPageWriteback(page); |
69cb51d1 | 2586 | if (ret) { |
1da177e4 LT |
2587 | radix_tree_tag_clear(&mapping->page_tree, |
2588 | page_index(page), | |
2589 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 2590 | if (bdi_cap_account_writeback(bdi)) { |
91018134 TH |
2591 | struct bdi_writeback *wb = inode_to_wb(inode); |
2592 | ||
2593 | __dec_wb_stat(wb, WB_WRITEBACK); | |
2594 | __wb_writeout_inc(wb); | |
04fbfdc1 | 2595 | } |
69cb51d1 | 2596 | } |
19fd6231 | 2597 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
2598 | } else { |
2599 | ret = TestClearPageWriteback(page); | |
2600 | } | |
99b12e3d | 2601 | if (ret) { |
d7365e78 | 2602 | mem_cgroup_dec_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK); |
d688abf5 | 2603 | dec_zone_page_state(page, NR_WRITEBACK); |
99b12e3d WF |
2604 | inc_zone_page_state(page, NR_WRITTEN); |
2605 | } | |
6de22619 | 2606 | mem_cgroup_end_page_stat(memcg); |
1da177e4 LT |
2607 | return ret; |
2608 | } | |
2609 | ||
1c8349a1 | 2610 | int __test_set_page_writeback(struct page *page, bool keep_write) |
1da177e4 LT |
2611 | { |
2612 | struct address_space *mapping = page_mapping(page); | |
d7365e78 | 2613 | struct mem_cgroup *memcg; |
d7365e78 | 2614 | int ret; |
1da177e4 | 2615 | |
6de22619 | 2616 | memcg = mem_cgroup_begin_page_stat(page); |
1da177e4 | 2617 | if (mapping) { |
91018134 TH |
2618 | struct inode *inode = mapping->host; |
2619 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 LT |
2620 | unsigned long flags; |
2621 | ||
19fd6231 | 2622 | spin_lock_irqsave(&mapping->tree_lock, flags); |
1da177e4 | 2623 | ret = TestSetPageWriteback(page); |
69cb51d1 | 2624 | if (!ret) { |
1da177e4 LT |
2625 | radix_tree_tag_set(&mapping->page_tree, |
2626 | page_index(page), | |
2627 | PAGECACHE_TAG_WRITEBACK); | |
e4ad08fe | 2628 | if (bdi_cap_account_writeback(bdi)) |
91018134 | 2629 | __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK); |
69cb51d1 | 2630 | } |
1da177e4 LT |
2631 | if (!PageDirty(page)) |
2632 | radix_tree_tag_clear(&mapping->page_tree, | |
2633 | page_index(page), | |
2634 | PAGECACHE_TAG_DIRTY); | |
1c8349a1 NJ |
2635 | if (!keep_write) |
2636 | radix_tree_tag_clear(&mapping->page_tree, | |
2637 | page_index(page), | |
2638 | PAGECACHE_TAG_TOWRITE); | |
19fd6231 | 2639 | spin_unlock_irqrestore(&mapping->tree_lock, flags); |
1da177e4 LT |
2640 | } else { |
2641 | ret = TestSetPageWriteback(page); | |
2642 | } | |
3a3c02ec | 2643 | if (!ret) { |
d7365e78 | 2644 | mem_cgroup_inc_page_stat(memcg, MEM_CGROUP_STAT_WRITEBACK); |
3a3c02ec JW |
2645 | inc_zone_page_state(page, NR_WRITEBACK); |
2646 | } | |
6de22619 | 2647 | mem_cgroup_end_page_stat(memcg); |
1da177e4 LT |
2648 | return ret; |
2649 | ||
2650 | } | |
1c8349a1 | 2651 | EXPORT_SYMBOL(__test_set_page_writeback); |
1da177e4 LT |
2652 | |
2653 | /* | |
00128188 | 2654 | * Return true if any of the pages in the mapping are marked with the |
1da177e4 LT |
2655 | * passed tag. |
2656 | */ | |
2657 | int mapping_tagged(struct address_space *mapping, int tag) | |
2658 | { | |
72c47832 | 2659 | return radix_tree_tagged(&mapping->page_tree, tag); |
1da177e4 LT |
2660 | } |
2661 | EXPORT_SYMBOL(mapping_tagged); | |
1d1d1a76 DW |
2662 | |
2663 | /** | |
2664 | * wait_for_stable_page() - wait for writeback to finish, if necessary. | |
2665 | * @page: The page to wait on. | |
2666 | * | |
2667 | * This function determines if the given page is related to a backing device | |
2668 | * that requires page contents to be held stable during writeback. If so, then | |
2669 | * it will wait for any pending writeback to complete. | |
2670 | */ | |
2671 | void wait_for_stable_page(struct page *page) | |
2672 | { | |
de1414a6 CH |
2673 | if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host))) |
2674 | wait_on_page_writeback(page); | |
1d1d1a76 DW |
2675 | } |
2676 | EXPORT_SYMBOL_GPL(wait_for_stable_page); |