Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 | 2 | /* |
f30c2269 | 3 | * mm/page-writeback.c |
1da177e4 LT |
4 | * |
5 | * Copyright (C) 2002, Linus Torvalds. | |
90eec103 | 6 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra |
1da177e4 LT |
7 | * |
8 | * Contains functions related to writing back dirty pages at the | |
9 | * address_space level. | |
10 | * | |
e1f8e874 | 11 | * 10Apr2002 Andrew Morton |
1da177e4 LT |
12 | * Initial version |
13 | */ | |
14 | ||
15 | #include <linux/kernel.h> | |
b95f1b31 | 16 | #include <linux/export.h> |
1da177e4 LT |
17 | #include <linux/spinlock.h> |
18 | #include <linux/fs.h> | |
19 | #include <linux/mm.h> | |
20 | #include <linux/swap.h> | |
21 | #include <linux/slab.h> | |
22 | #include <linux/pagemap.h> | |
23 | #include <linux/writeback.h> | |
24 | #include <linux/init.h> | |
25 | #include <linux/backing-dev.h> | |
55e829af | 26 | #include <linux/task_io_accounting_ops.h> |
1da177e4 LT |
27 | #include <linux/blkdev.h> |
28 | #include <linux/mpage.h> | |
d08b3851 | 29 | #include <linux/rmap.h> |
1da177e4 | 30 | #include <linux/percpu.h> |
1da177e4 LT |
31 | #include <linux/smp.h> |
32 | #include <linux/sysctl.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/syscalls.h> | |
811d736f | 35 | #include <linux/pagevec.h> |
eb608e3a | 36 | #include <linux/timer.h> |
8bd75c77 | 37 | #include <linux/sched/rt.h> |
f361bf4a | 38 | #include <linux/sched/signal.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 | 109 | |
1da177e4 | 110 | /* |
ed5b43f1 BS |
111 | * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: |
112 | * a full sync is triggered after this time elapses without any disk activity. | |
1da177e4 LT |
113 | */ |
114 | int laptop_mode; | |
115 | ||
116 | EXPORT_SYMBOL(laptop_mode); | |
117 | ||
118 | /* End of sysctl-exported parameters */ | |
119 | ||
dcc25ae7 | 120 | struct wb_domain global_wb_domain; |
1da177e4 | 121 | |
2bc00aef TH |
122 | /* consolidated parameters for balance_dirty_pages() and its subroutines */ |
123 | struct dirty_throttle_control { | |
e9f07dfd TH |
124 | #ifdef CONFIG_CGROUP_WRITEBACK |
125 | struct wb_domain *dom; | |
9fc3a43e | 126 | struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */ |
e9f07dfd | 127 | #endif |
2bc00aef | 128 | struct bdi_writeback *wb; |
e9770b34 | 129 | struct fprop_local_percpu *wb_completions; |
eb608e3a | 130 | |
9fc3a43e | 131 | unsigned long avail; /* dirtyable */ |
2bc00aef TH |
132 | unsigned long dirty; /* file_dirty + write + nfs */ |
133 | unsigned long thresh; /* dirty threshold */ | |
134 | unsigned long bg_thresh; /* dirty background threshold */ | |
135 | ||
136 | unsigned long wb_dirty; /* per-wb counterparts */ | |
137 | unsigned long wb_thresh; | |
970fb01a | 138 | unsigned long wb_bg_thresh; |
daddfa3c TH |
139 | |
140 | unsigned long pos_ratio; | |
2bc00aef TH |
141 | }; |
142 | ||
eb608e3a JK |
143 | /* |
144 | * Length of period for aging writeout fractions of bdis. This is an | |
145 | * arbitrarily chosen number. The longer the period, the slower fractions will | |
146 | * reflect changes in current writeout rate. | |
147 | */ | |
148 | #define VM_COMPLETIONS_PERIOD_LEN (3*HZ) | |
04fbfdc1 | 149 | |
693108a8 TH |
150 | #ifdef CONFIG_CGROUP_WRITEBACK |
151 | ||
d60d1bdd TH |
152 | #define GDTC_INIT(__wb) .wb = (__wb), \ |
153 | .dom = &global_wb_domain, \ | |
154 | .wb_completions = &(__wb)->completions | |
155 | ||
9fc3a43e | 156 | #define GDTC_INIT_NO_WB .dom = &global_wb_domain |
d60d1bdd TH |
157 | |
158 | #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \ | |
159 | .dom = mem_cgroup_wb_domain(__wb), \ | |
160 | .wb_completions = &(__wb)->memcg_completions, \ | |
161 | .gdtc = __gdtc | |
c2aa723a TH |
162 | |
163 | static bool mdtc_valid(struct dirty_throttle_control *dtc) | |
164 | { | |
165 | return dtc->dom; | |
166 | } | |
e9f07dfd TH |
167 | |
168 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) | |
169 | { | |
170 | return dtc->dom; | |
171 | } | |
172 | ||
9fc3a43e TH |
173 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
174 | { | |
175 | return mdtc->gdtc; | |
176 | } | |
177 | ||
841710aa TH |
178 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
179 | { | |
180 | return &wb->memcg_completions; | |
181 | } | |
182 | ||
693108a8 TH |
183 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
184 | unsigned long *minp, unsigned long *maxp) | |
185 | { | |
20792ebf | 186 | unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth); |
693108a8 TH |
187 | unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); |
188 | unsigned long long min = wb->bdi->min_ratio; | |
189 | unsigned long long max = wb->bdi->max_ratio; | |
190 | ||
191 | /* | |
192 | * @wb may already be clean by the time control reaches here and | |
193 | * the total may not include its bw. | |
194 | */ | |
195 | if (this_bw < tot_bw) { | |
196 | if (min) { | |
197 | min *= this_bw; | |
6d9e8c65 | 198 | min = div64_ul(min, tot_bw); |
693108a8 TH |
199 | } |
200 | if (max < 100) { | |
201 | max *= this_bw; | |
6d9e8c65 | 202 | max = div64_ul(max, tot_bw); |
693108a8 TH |
203 | } |
204 | } | |
205 | ||
206 | *minp = min; | |
207 | *maxp = max; | |
208 | } | |
209 | ||
210 | #else /* CONFIG_CGROUP_WRITEBACK */ | |
211 | ||
d60d1bdd TH |
212 | #define GDTC_INIT(__wb) .wb = (__wb), \ |
213 | .wb_completions = &(__wb)->completions | |
9fc3a43e | 214 | #define GDTC_INIT_NO_WB |
c2aa723a TH |
215 | #define MDTC_INIT(__wb, __gdtc) |
216 | ||
217 | static bool mdtc_valid(struct dirty_throttle_control *dtc) | |
218 | { | |
219 | return false; | |
220 | } | |
e9f07dfd TH |
221 | |
222 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) | |
223 | { | |
224 | return &global_wb_domain; | |
225 | } | |
226 | ||
9fc3a43e TH |
227 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
228 | { | |
229 | return NULL; | |
230 | } | |
231 | ||
841710aa TH |
232 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
233 | { | |
234 | return NULL; | |
235 | } | |
236 | ||
693108a8 TH |
237 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
238 | unsigned long *minp, unsigned long *maxp) | |
239 | { | |
240 | *minp = wb->bdi->min_ratio; | |
241 | *maxp = wb->bdi->max_ratio; | |
242 | } | |
243 | ||
244 | #endif /* CONFIG_CGROUP_WRITEBACK */ | |
245 | ||
a756cf59 JW |
246 | /* |
247 | * In a memory zone, there is a certain amount of pages we consider | |
248 | * available for the page cache, which is essentially the number of | |
249 | * free and reclaimable pages, minus some zone reserves to protect | |
250 | * lowmem and the ability to uphold the zone's watermarks without | |
251 | * requiring writeback. | |
252 | * | |
253 | * This number of dirtyable pages is the base value of which the | |
e0857cf5 | 254 | * user-configurable dirty ratio is the effective number of pages that |
a756cf59 JW |
255 | * are allowed to be actually dirtied. Per individual zone, or |
256 | * globally by using the sum of dirtyable pages over all zones. | |
257 | * | |
258 | * Because the user is allowed to specify the dirty limit globally as | |
259 | * absolute number of bytes, calculating the per-zone dirty limit can | |
260 | * require translating the configured limit into a percentage of | |
261 | * global dirtyable memory first. | |
262 | */ | |
263 | ||
a804552b | 264 | /** |
281e3726 MG |
265 | * node_dirtyable_memory - number of dirtyable pages in a node |
266 | * @pgdat: the node | |
a804552b | 267 | * |
a862f68a | 268 | * Return: the node's number of pages potentially available for dirty |
281e3726 | 269 | * page cache. This is the base value for the per-node dirty limits. |
a804552b | 270 | */ |
281e3726 | 271 | static unsigned long node_dirtyable_memory(struct pglist_data *pgdat) |
a804552b | 272 | { |
281e3726 MG |
273 | unsigned long nr_pages = 0; |
274 | int z; | |
275 | ||
276 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
277 | struct zone *zone = pgdat->node_zones + z; | |
278 | ||
279 | if (!populated_zone(zone)) | |
280 | continue; | |
281 | ||
282 | nr_pages += zone_page_state(zone, NR_FREE_PAGES); | |
283 | } | |
a804552b | 284 | |
a8d01437 JW |
285 | /* |
286 | * Pages reserved for the kernel should not be considered | |
287 | * dirtyable, to prevent a situation where reclaim has to | |
288 | * clean pages in order to balance the zones. | |
289 | */ | |
281e3726 | 290 | nr_pages -= min(nr_pages, pgdat->totalreserve_pages); |
a804552b | 291 | |
281e3726 MG |
292 | nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE); |
293 | nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE); | |
a804552b JW |
294 | |
295 | return nr_pages; | |
296 | } | |
297 | ||
1edf2234 JW |
298 | static unsigned long highmem_dirtyable_memory(unsigned long total) |
299 | { | |
300 | #ifdef CONFIG_HIGHMEM | |
301 | int node; | |
bb4cc2be | 302 | unsigned long x = 0; |
09b4ab3c | 303 | int i; |
1edf2234 JW |
304 | |
305 | for_each_node_state(node, N_HIGH_MEMORY) { | |
281e3726 MG |
306 | for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) { |
307 | struct zone *z; | |
9cb937e2 | 308 | unsigned long nr_pages; |
281e3726 MG |
309 | |
310 | if (!is_highmem_idx(i)) | |
311 | continue; | |
312 | ||
313 | z = &NODE_DATA(node)->node_zones[i]; | |
9cb937e2 MK |
314 | if (!populated_zone(z)) |
315 | continue; | |
1edf2234 | 316 | |
9cb937e2 | 317 | nr_pages = zone_page_state(z, NR_FREE_PAGES); |
281e3726 | 318 | /* watch for underflows */ |
9cb937e2 | 319 | nr_pages -= min(nr_pages, high_wmark_pages(z)); |
bb4cc2be MG |
320 | nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE); |
321 | nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE); | |
322 | x += nr_pages; | |
09b4ab3c | 323 | } |
1edf2234 | 324 | } |
281e3726 | 325 | |
c8b74c2f SR |
326 | /* |
327 | * Unreclaimable memory (kernel memory or anonymous memory | |
328 | * without swap) can bring down the dirtyable pages below | |
329 | * the zone's dirty balance reserve and the above calculation | |
330 | * will underflow. However we still want to add in nodes | |
331 | * which are below threshold (negative values) to get a more | |
332 | * accurate calculation but make sure that the total never | |
333 | * underflows. | |
334 | */ | |
335 | if ((long)x < 0) | |
336 | x = 0; | |
337 | ||
1edf2234 JW |
338 | /* |
339 | * Make sure that the number of highmem pages is never larger | |
340 | * than the number of the total dirtyable memory. This can only | |
341 | * occur in very strange VM situations but we want to make sure | |
342 | * that this does not occur. | |
343 | */ | |
344 | return min(x, total); | |
345 | #else | |
346 | return 0; | |
347 | #endif | |
348 | } | |
349 | ||
350 | /** | |
ccafa287 | 351 | * global_dirtyable_memory - number of globally dirtyable pages |
1edf2234 | 352 | * |
a862f68a | 353 | * Return: the global number of pages potentially available for dirty |
ccafa287 | 354 | * page cache. This is the base value for the global dirty limits. |
1edf2234 | 355 | */ |
18cf8cf8 | 356 | static unsigned long global_dirtyable_memory(void) |
1edf2234 JW |
357 | { |
358 | unsigned long x; | |
359 | ||
c41f012a | 360 | x = global_zone_page_state(NR_FREE_PAGES); |
a8d01437 JW |
361 | /* |
362 | * Pages reserved for the kernel should not be considered | |
363 | * dirtyable, to prevent a situation where reclaim has to | |
364 | * clean pages in order to balance the zones. | |
365 | */ | |
366 | x -= min(x, totalreserve_pages); | |
1edf2234 | 367 | |
599d0c95 MG |
368 | x += global_node_page_state(NR_INACTIVE_FILE); |
369 | x += global_node_page_state(NR_ACTIVE_FILE); | |
a804552b | 370 | |
1edf2234 JW |
371 | if (!vm_highmem_is_dirtyable) |
372 | x -= highmem_dirtyable_memory(x); | |
373 | ||
374 | return x + 1; /* Ensure that we never return 0 */ | |
375 | } | |
376 | ||
9fc3a43e TH |
377 | /** |
378 | * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain | |
379 | * @dtc: dirty_throttle_control of interest | |
ccafa287 | 380 | * |
9fc3a43e TH |
381 | * Calculate @dtc->thresh and ->bg_thresh considering |
382 | * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller | |
383 | * must ensure that @dtc->avail is set before calling this function. The | |
a37b0715 | 384 | * dirty limits will be lifted by 1/4 for real-time tasks. |
ccafa287 | 385 | */ |
9fc3a43e | 386 | static void domain_dirty_limits(struct dirty_throttle_control *dtc) |
ccafa287 | 387 | { |
9fc3a43e TH |
388 | const unsigned long available_memory = dtc->avail; |
389 | struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc); | |
390 | unsigned long bytes = vm_dirty_bytes; | |
391 | unsigned long bg_bytes = dirty_background_bytes; | |
62a584fe TH |
392 | /* convert ratios to per-PAGE_SIZE for higher precision */ |
393 | unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100; | |
394 | unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100; | |
9fc3a43e TH |
395 | unsigned long thresh; |
396 | unsigned long bg_thresh; | |
ccafa287 JW |
397 | struct task_struct *tsk; |
398 | ||
9fc3a43e TH |
399 | /* gdtc is !NULL iff @dtc is for memcg domain */ |
400 | if (gdtc) { | |
401 | unsigned long global_avail = gdtc->avail; | |
402 | ||
403 | /* | |
404 | * The byte settings can't be applied directly to memcg | |
405 | * domains. Convert them to ratios by scaling against | |
62a584fe TH |
406 | * globally available memory. As the ratios are in |
407 | * per-PAGE_SIZE, they can be obtained by dividing bytes by | |
408 | * number of pages. | |
9fc3a43e TH |
409 | */ |
410 | if (bytes) | |
62a584fe TH |
411 | ratio = min(DIV_ROUND_UP(bytes, global_avail), |
412 | PAGE_SIZE); | |
9fc3a43e | 413 | if (bg_bytes) |
62a584fe TH |
414 | bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail), |
415 | PAGE_SIZE); | |
9fc3a43e TH |
416 | bytes = bg_bytes = 0; |
417 | } | |
418 | ||
419 | if (bytes) | |
420 | thresh = DIV_ROUND_UP(bytes, PAGE_SIZE); | |
ccafa287 | 421 | else |
62a584fe | 422 | thresh = (ratio * available_memory) / PAGE_SIZE; |
ccafa287 | 423 | |
9fc3a43e TH |
424 | if (bg_bytes) |
425 | bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE); | |
ccafa287 | 426 | else |
62a584fe | 427 | bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE; |
ccafa287 | 428 | |
90daf306 | 429 | if (bg_thresh >= thresh) |
9fc3a43e | 430 | bg_thresh = thresh / 2; |
ccafa287 | 431 | tsk = current; |
a37b0715 | 432 | if (rt_task(tsk)) { |
a53eaff8 N |
433 | bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32; |
434 | thresh += thresh / 4 + global_wb_domain.dirty_limit / 32; | |
ccafa287 | 435 | } |
9fc3a43e TH |
436 | dtc->thresh = thresh; |
437 | dtc->bg_thresh = bg_thresh; | |
438 | ||
439 | /* we should eventually report the domain in the TP */ | |
440 | if (!gdtc) | |
441 | trace_global_dirty_state(bg_thresh, thresh); | |
442 | } | |
443 | ||
444 | /** | |
445 | * global_dirty_limits - background-writeback and dirty-throttling thresholds | |
446 | * @pbackground: out parameter for bg_thresh | |
447 | * @pdirty: out parameter for thresh | |
448 | * | |
449 | * Calculate bg_thresh and thresh for global_wb_domain. See | |
450 | * domain_dirty_limits() for details. | |
451 | */ | |
452 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) | |
453 | { | |
454 | struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB }; | |
455 | ||
456 | gdtc.avail = global_dirtyable_memory(); | |
457 | domain_dirty_limits(&gdtc); | |
458 | ||
459 | *pbackground = gdtc.bg_thresh; | |
460 | *pdirty = gdtc.thresh; | |
ccafa287 JW |
461 | } |
462 | ||
a756cf59 | 463 | /** |
281e3726 MG |
464 | * node_dirty_limit - maximum number of dirty pages allowed in a node |
465 | * @pgdat: the node | |
a756cf59 | 466 | * |
a862f68a | 467 | * Return: the maximum number of dirty pages allowed in a node, based |
281e3726 | 468 | * on the node's dirtyable memory. |
a756cf59 | 469 | */ |
281e3726 | 470 | static unsigned long node_dirty_limit(struct pglist_data *pgdat) |
a756cf59 | 471 | { |
281e3726 | 472 | unsigned long node_memory = node_dirtyable_memory(pgdat); |
a756cf59 JW |
473 | struct task_struct *tsk = current; |
474 | unsigned long dirty; | |
475 | ||
476 | if (vm_dirty_bytes) | |
477 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) * | |
281e3726 | 478 | node_memory / global_dirtyable_memory(); |
a756cf59 | 479 | else |
281e3726 | 480 | dirty = vm_dirty_ratio * node_memory / 100; |
a756cf59 | 481 | |
a37b0715 | 482 | if (rt_task(tsk)) |
a756cf59 JW |
483 | dirty += dirty / 4; |
484 | ||
485 | return dirty; | |
486 | } | |
487 | ||
488 | /** | |
281e3726 MG |
489 | * node_dirty_ok - tells whether a node is within its dirty limits |
490 | * @pgdat: the node to check | |
a756cf59 | 491 | * |
a862f68a | 492 | * Return: %true when the dirty pages in @pgdat are within the node's |
a756cf59 JW |
493 | * dirty limit, %false if the limit is exceeded. |
494 | */ | |
281e3726 | 495 | bool node_dirty_ok(struct pglist_data *pgdat) |
a756cf59 | 496 | { |
281e3726 MG |
497 | unsigned long limit = node_dirty_limit(pgdat); |
498 | unsigned long nr_pages = 0; | |
499 | ||
11fb9989 | 500 | nr_pages += node_page_state(pgdat, NR_FILE_DIRTY); |
11fb9989 | 501 | nr_pages += node_page_state(pgdat, NR_WRITEBACK); |
a756cf59 | 502 | |
281e3726 | 503 | return nr_pages <= limit; |
a756cf59 JW |
504 | } |
505 | ||
2da02997 | 506 | int dirty_background_ratio_handler(struct ctl_table *table, int write, |
32927393 | 507 | void *buffer, size_t *lenp, loff_t *ppos) |
2da02997 DR |
508 | { |
509 | int ret; | |
510 | ||
8d65af78 | 511 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
512 | if (ret == 0 && write) |
513 | dirty_background_bytes = 0; | |
514 | return ret; | |
515 | } | |
516 | ||
517 | int dirty_background_bytes_handler(struct ctl_table *table, int write, | |
32927393 | 518 | void *buffer, size_t *lenp, loff_t *ppos) |
2da02997 DR |
519 | { |
520 | int ret; | |
521 | ||
8d65af78 | 522 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
523 | if (ret == 0 && write) |
524 | dirty_background_ratio = 0; | |
525 | return ret; | |
526 | } | |
527 | ||
32927393 CH |
528 | int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer, |
529 | size_t *lenp, loff_t *ppos) | |
04fbfdc1 PZ |
530 | { |
531 | int old_ratio = vm_dirty_ratio; | |
2da02997 DR |
532 | int ret; |
533 | ||
8d65af78 | 534 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
04fbfdc1 | 535 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
eb608e3a | 536 | writeback_set_ratelimit(); |
2da02997 DR |
537 | vm_dirty_bytes = 0; |
538 | } | |
539 | return ret; | |
540 | } | |
541 | ||
2da02997 | 542 | int dirty_bytes_handler(struct ctl_table *table, int write, |
32927393 | 543 | void *buffer, size_t *lenp, loff_t *ppos) |
2da02997 | 544 | { |
fc3501d4 | 545 | unsigned long old_bytes = vm_dirty_bytes; |
2da02997 DR |
546 | int ret; |
547 | ||
8d65af78 | 548 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 | 549 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
eb608e3a | 550 | writeback_set_ratelimit(); |
2da02997 | 551 | vm_dirty_ratio = 0; |
04fbfdc1 PZ |
552 | } |
553 | return ret; | |
554 | } | |
555 | ||
eb608e3a JK |
556 | static unsigned long wp_next_time(unsigned long cur_time) |
557 | { | |
558 | cur_time += VM_COMPLETIONS_PERIOD_LEN; | |
559 | /* 0 has a special meaning... */ | |
560 | if (!cur_time) | |
561 | return 1; | |
562 | return cur_time; | |
563 | } | |
564 | ||
cc24df4c | 565 | static void wb_domain_writeout_add(struct wb_domain *dom, |
c7981433 | 566 | struct fprop_local_percpu *completions, |
cc24df4c | 567 | unsigned int max_prop_frac, long nr) |
04fbfdc1 | 568 | { |
be5f1797 | 569 | __fprop_add_percpu_max(&dom->completions, completions, |
cc24df4c | 570 | max_prop_frac, nr); |
eb608e3a | 571 | /* First event after period switching was turned off? */ |
517663ed | 572 | if (unlikely(!dom->period_time)) { |
eb608e3a JK |
573 | /* |
574 | * We can race with other __bdi_writeout_inc calls here but | |
575 | * it does not cause any harm since the resulting time when | |
576 | * timer will fire and what is in writeout_period_time will be | |
577 | * roughly the same. | |
578 | */ | |
380c27ca TH |
579 | dom->period_time = wp_next_time(jiffies); |
580 | mod_timer(&dom->period_timer, dom->period_time); | |
eb608e3a | 581 | } |
04fbfdc1 PZ |
582 | } |
583 | ||
c7981433 TH |
584 | /* |
585 | * Increment @wb's writeout completion count and the global writeout | |
269ccca3 | 586 | * completion count. Called from __folio_end_writeback(). |
c7981433 | 587 | */ |
cc24df4c | 588 | static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr) |
dd5656e5 | 589 | { |
841710aa | 590 | struct wb_domain *cgdom; |
dd5656e5 | 591 | |
cc24df4c MWO |
592 | wb_stat_mod(wb, WB_WRITTEN, nr); |
593 | wb_domain_writeout_add(&global_wb_domain, &wb->completions, | |
594 | wb->bdi->max_prop_frac, nr); | |
841710aa TH |
595 | |
596 | cgdom = mem_cgroup_wb_domain(wb); | |
597 | if (cgdom) | |
cc24df4c MWO |
598 | wb_domain_writeout_add(cgdom, wb_memcg_completions(wb), |
599 | wb->bdi->max_prop_frac, nr); | |
dd5656e5 | 600 | } |
dd5656e5 | 601 | |
93f78d88 | 602 | void wb_writeout_inc(struct bdi_writeback *wb) |
04fbfdc1 | 603 | { |
dd5656e5 MS |
604 | unsigned long flags; |
605 | ||
606 | local_irq_save(flags); | |
cc24df4c | 607 | __wb_writeout_add(wb, 1); |
dd5656e5 | 608 | local_irq_restore(flags); |
04fbfdc1 | 609 | } |
93f78d88 | 610 | EXPORT_SYMBOL_GPL(wb_writeout_inc); |
04fbfdc1 | 611 | |
eb608e3a JK |
612 | /* |
613 | * On idle system, we can be called long after we scheduled because we use | |
614 | * deferred timers so count with missed periods. | |
615 | */ | |
9823e51b | 616 | static void writeout_period(struct timer_list *t) |
eb608e3a | 617 | { |
9823e51b | 618 | struct wb_domain *dom = from_timer(dom, t, period_timer); |
380c27ca | 619 | int miss_periods = (jiffies - dom->period_time) / |
eb608e3a JK |
620 | VM_COMPLETIONS_PERIOD_LEN; |
621 | ||
380c27ca TH |
622 | if (fprop_new_period(&dom->completions, miss_periods + 1)) { |
623 | dom->period_time = wp_next_time(dom->period_time + | |
eb608e3a | 624 | miss_periods * VM_COMPLETIONS_PERIOD_LEN); |
380c27ca | 625 | mod_timer(&dom->period_timer, dom->period_time); |
eb608e3a JK |
626 | } else { |
627 | /* | |
628 | * Aging has zeroed all fractions. Stop wasting CPU on period | |
629 | * updates. | |
630 | */ | |
380c27ca | 631 | dom->period_time = 0; |
eb608e3a JK |
632 | } |
633 | } | |
634 | ||
380c27ca TH |
635 | int wb_domain_init(struct wb_domain *dom, gfp_t gfp) |
636 | { | |
637 | memset(dom, 0, sizeof(*dom)); | |
dcc25ae7 TH |
638 | |
639 | spin_lock_init(&dom->lock); | |
640 | ||
9823e51b | 641 | timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE); |
dcc25ae7 TH |
642 | |
643 | dom->dirty_limit_tstamp = jiffies; | |
644 | ||
380c27ca TH |
645 | return fprop_global_init(&dom->completions, gfp); |
646 | } | |
647 | ||
841710aa TH |
648 | #ifdef CONFIG_CGROUP_WRITEBACK |
649 | void wb_domain_exit(struct wb_domain *dom) | |
650 | { | |
651 | del_timer_sync(&dom->period_timer); | |
652 | fprop_global_destroy(&dom->completions); | |
653 | } | |
654 | #endif | |
655 | ||
189d3c4a | 656 | /* |
d08c429b JW |
657 | * bdi_min_ratio keeps the sum of the minimum dirty shares of all |
658 | * registered backing devices, which, for obvious reasons, can not | |
659 | * exceed 100%. | |
189d3c4a | 660 | */ |
189d3c4a PZ |
661 | static unsigned int bdi_min_ratio; |
662 | ||
663 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) | |
664 | { | |
665 | int ret = 0; | |
189d3c4a | 666 | |
cfc4ba53 | 667 | spin_lock_bh(&bdi_lock); |
a42dde04 | 668 | if (min_ratio > bdi->max_ratio) { |
189d3c4a | 669 | ret = -EINVAL; |
a42dde04 PZ |
670 | } else { |
671 | min_ratio -= bdi->min_ratio; | |
672 | if (bdi_min_ratio + min_ratio < 100) { | |
673 | bdi_min_ratio += min_ratio; | |
674 | bdi->min_ratio += min_ratio; | |
675 | } else { | |
676 | ret = -EINVAL; | |
677 | } | |
678 | } | |
cfc4ba53 | 679 | spin_unlock_bh(&bdi_lock); |
a42dde04 PZ |
680 | |
681 | return ret; | |
682 | } | |
683 | ||
684 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio) | |
685 | { | |
a42dde04 PZ |
686 | int ret = 0; |
687 | ||
688 | if (max_ratio > 100) | |
689 | return -EINVAL; | |
690 | ||
cfc4ba53 | 691 | spin_lock_bh(&bdi_lock); |
a42dde04 PZ |
692 | if (bdi->min_ratio > max_ratio) { |
693 | ret = -EINVAL; | |
694 | } else { | |
695 | bdi->max_ratio = max_ratio; | |
eb608e3a | 696 | bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100; |
a42dde04 | 697 | } |
cfc4ba53 | 698 | spin_unlock_bh(&bdi_lock); |
189d3c4a PZ |
699 | |
700 | return ret; | |
701 | } | |
a42dde04 | 702 | EXPORT_SYMBOL(bdi_set_max_ratio); |
189d3c4a | 703 | |
6c14ae1e WF |
704 | static unsigned long dirty_freerun_ceiling(unsigned long thresh, |
705 | unsigned long bg_thresh) | |
706 | { | |
707 | return (thresh + bg_thresh) / 2; | |
708 | } | |
709 | ||
c7981433 TH |
710 | static unsigned long hard_dirty_limit(struct wb_domain *dom, |
711 | unsigned long thresh) | |
ffd1f609 | 712 | { |
dcc25ae7 | 713 | return max(thresh, dom->dirty_limit); |
ffd1f609 WF |
714 | } |
715 | ||
c5edf9cd TH |
716 | /* |
717 | * Memory which can be further allocated to a memcg domain is capped by | |
718 | * system-wide clean memory excluding the amount being used in the domain. | |
719 | */ | |
720 | static void mdtc_calc_avail(struct dirty_throttle_control *mdtc, | |
721 | unsigned long filepages, unsigned long headroom) | |
c2aa723a TH |
722 | { |
723 | struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc); | |
c5edf9cd TH |
724 | unsigned long clean = filepages - min(filepages, mdtc->dirty); |
725 | unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty); | |
726 | unsigned long other_clean = global_clean - min(global_clean, clean); | |
c2aa723a | 727 | |
c5edf9cd | 728 | mdtc->avail = filepages + min(headroom, other_clean); |
ffd1f609 WF |
729 | } |
730 | ||
6f718656 | 731 | /** |
b1cbc6d4 TH |
732 | * __wb_calc_thresh - @wb's share of dirty throttling threshold |
733 | * @dtc: dirty_throttle_context of interest | |
1babe183 | 734 | * |
aed21ad2 WF |
735 | * Note that balance_dirty_pages() will only seriously take it as a hard limit |
736 | * when sleeping max_pause per page is not enough to keep the dirty pages under | |
737 | * control. For example, when the device is completely stalled due to some error | |
738 | * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key. | |
739 | * In the other normal situations, it acts more gently by throttling the tasks | |
a88a341a | 740 | * more (rather than completely block them) when the wb dirty pages go high. |
1babe183 | 741 | * |
6f718656 | 742 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
1babe183 WF |
743 | * - starving fast devices |
744 | * - piling up dirty pages (that will take long time to sync) on slow devices | |
745 | * | |
a88a341a | 746 | * The wb's share of dirty limit will be adapting to its throughput and |
1babe183 | 747 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. |
a862f68a MR |
748 | * |
749 | * Return: @wb's dirty limit in pages. The term "dirty" in the context of | |
8d92890b | 750 | * dirty balancing includes all PG_dirty and PG_writeback pages. |
1babe183 | 751 | */ |
b1cbc6d4 | 752 | static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc) |
16c4042f | 753 | { |
e9f07dfd | 754 | struct wb_domain *dom = dtc_dom(dtc); |
b1cbc6d4 | 755 | unsigned long thresh = dtc->thresh; |
0d960a38 | 756 | u64 wb_thresh; |
d3ac946e | 757 | unsigned long numerator, denominator; |
693108a8 | 758 | unsigned long wb_min_ratio, wb_max_ratio; |
04fbfdc1 | 759 | |
16c4042f | 760 | /* |
0d960a38 | 761 | * Calculate this BDI's share of the thresh ratio. |
16c4042f | 762 | */ |
e9770b34 | 763 | fprop_fraction_percpu(&dom->completions, dtc->wb_completions, |
380c27ca | 764 | &numerator, &denominator); |
04fbfdc1 | 765 | |
0d960a38 TH |
766 | wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100; |
767 | wb_thresh *= numerator; | |
d3ac946e | 768 | wb_thresh = div64_ul(wb_thresh, denominator); |
04fbfdc1 | 769 | |
b1cbc6d4 | 770 | wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio); |
04fbfdc1 | 771 | |
0d960a38 TH |
772 | wb_thresh += (thresh * wb_min_ratio) / 100; |
773 | if (wb_thresh > (thresh * wb_max_ratio) / 100) | |
774 | wb_thresh = thresh * wb_max_ratio / 100; | |
16c4042f | 775 | |
0d960a38 | 776 | return wb_thresh; |
1da177e4 LT |
777 | } |
778 | ||
b1cbc6d4 TH |
779 | unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh) |
780 | { | |
781 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb), | |
782 | .thresh = thresh }; | |
783 | return __wb_calc_thresh(&gdtc); | |
1da177e4 LT |
784 | } |
785 | ||
5a537485 MP |
786 | /* |
787 | * setpoint - dirty 3 | |
788 | * f(dirty) := 1.0 + (----------------) | |
789 | * limit - setpoint | |
790 | * | |
791 | * it's a 3rd order polynomial that subjects to | |
792 | * | |
793 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast | |
794 | * (2) f(setpoint) = 1.0 => the balance point | |
795 | * (3) f(limit) = 0 => the hard limit | |
796 | * (4) df/dx <= 0 => negative feedback control | |
797 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) | |
798 | * => fast response on large errors; small oscillation near setpoint | |
799 | */ | |
d5c9fde3 | 800 | static long long pos_ratio_polynom(unsigned long setpoint, |
5a537485 MP |
801 | unsigned long dirty, |
802 | unsigned long limit) | |
803 | { | |
804 | long long pos_ratio; | |
805 | long x; | |
806 | ||
d5c9fde3 | 807 | x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT, |
464d1387 | 808 | (limit - setpoint) | 1); |
5a537485 MP |
809 | pos_ratio = x; |
810 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
811 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
812 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT; | |
813 | ||
814 | return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT); | |
815 | } | |
816 | ||
6c14ae1e WF |
817 | /* |
818 | * Dirty position control. | |
819 | * | |
820 | * (o) global/bdi setpoints | |
821 | * | |
de1fff37 | 822 | * We want the dirty pages be balanced around the global/wb setpoints. |
6c14ae1e WF |
823 | * When the number of dirty pages is higher/lower than the setpoint, the |
824 | * dirty position control ratio (and hence task dirty ratelimit) will be | |
825 | * decreased/increased to bring the dirty pages back to the setpoint. | |
826 | * | |
827 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT | |
828 | * | |
829 | * if (dirty < setpoint) scale up pos_ratio | |
830 | * if (dirty > setpoint) scale down pos_ratio | |
831 | * | |
de1fff37 TH |
832 | * if (wb_dirty < wb_setpoint) scale up pos_ratio |
833 | * if (wb_dirty > wb_setpoint) scale down pos_ratio | |
6c14ae1e WF |
834 | * |
835 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT | |
836 | * | |
837 | * (o) global control line | |
838 | * | |
839 | * ^ pos_ratio | |
840 | * | | |
841 | * | |<===== global dirty control scope ======>| | |
03231554 | 842 | * 2.0 * * * * * * * |
6c14ae1e WF |
843 | * | .* |
844 | * | . * | |
845 | * | . * | |
846 | * | . * | |
847 | * | . * | |
848 | * | . * | |
849 | * 1.0 ................................* | |
850 | * | . . * | |
851 | * | . . * | |
852 | * | . . * | |
853 | * | . . * | |
854 | * | . . * | |
855 | * 0 +------------.------------------.----------------------*-------------> | |
856 | * freerun^ setpoint^ limit^ dirty pages | |
857 | * | |
de1fff37 | 858 | * (o) wb control line |
6c14ae1e WF |
859 | * |
860 | * ^ pos_ratio | |
861 | * | | |
862 | * | * | |
863 | * | * | |
864 | * | * | |
865 | * | * | |
866 | * | * |<=========== span ============>| | |
867 | * 1.0 .......................* | |
868 | * | . * | |
869 | * | . * | |
870 | * | . * | |
871 | * | . * | |
872 | * | . * | |
873 | * | . * | |
874 | * | . * | |
875 | * | . * | |
876 | * | . * | |
877 | * | . * | |
878 | * | . * | |
879 | * 1/4 ...............................................* * * * * * * * * * * * | |
880 | * | . . | |
881 | * | . . | |
882 | * | . . | |
883 | * 0 +----------------------.-------------------------------.-------------> | |
de1fff37 | 884 | * wb_setpoint^ x_intercept^ |
6c14ae1e | 885 | * |
de1fff37 | 886 | * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can |
6c14ae1e WF |
887 | * be smoothly throttled down to normal if it starts high in situations like |
888 | * - start writing to a slow SD card and a fast disk at the same time. The SD | |
de1fff37 TH |
889 | * card's wb_dirty may rush to many times higher than wb_setpoint. |
890 | * - the wb dirty thresh drops quickly due to change of JBOD workload | |
6c14ae1e | 891 | */ |
daddfa3c | 892 | static void wb_position_ratio(struct dirty_throttle_control *dtc) |
6c14ae1e | 893 | { |
2bc00aef | 894 | struct bdi_writeback *wb = dtc->wb; |
20792ebf | 895 | unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth); |
2bc00aef | 896 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); |
c7981433 | 897 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
2bc00aef | 898 | unsigned long wb_thresh = dtc->wb_thresh; |
6c14ae1e WF |
899 | unsigned long x_intercept; |
900 | unsigned long setpoint; /* dirty pages' target balance point */ | |
de1fff37 | 901 | unsigned long wb_setpoint; |
6c14ae1e WF |
902 | unsigned long span; |
903 | long long pos_ratio; /* for scaling up/down the rate limit */ | |
904 | long x; | |
905 | ||
daddfa3c TH |
906 | dtc->pos_ratio = 0; |
907 | ||
2bc00aef | 908 | if (unlikely(dtc->dirty >= limit)) |
daddfa3c | 909 | return; |
6c14ae1e WF |
910 | |
911 | /* | |
912 | * global setpoint | |
913 | * | |
5a537485 MP |
914 | * See comment for pos_ratio_polynom(). |
915 | */ | |
916 | setpoint = (freerun + limit) / 2; | |
2bc00aef | 917 | pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit); |
5a537485 MP |
918 | |
919 | /* | |
920 | * The strictlimit feature is a tool preventing mistrusted filesystems | |
921 | * from growing a large number of dirty pages before throttling. For | |
de1fff37 TH |
922 | * such filesystems balance_dirty_pages always checks wb counters |
923 | * against wb limits. Even if global "nr_dirty" is under "freerun". | |
5a537485 MP |
924 | * This is especially important for fuse which sets bdi->max_ratio to |
925 | * 1% by default. Without strictlimit feature, fuse writeback may | |
926 | * consume arbitrary amount of RAM because it is accounted in | |
927 | * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty". | |
6c14ae1e | 928 | * |
a88a341a | 929 | * Here, in wb_position_ratio(), we calculate pos_ratio based on |
de1fff37 | 930 | * two values: wb_dirty and wb_thresh. Let's consider an example: |
5a537485 MP |
931 | * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global |
932 | * limits are set by default to 10% and 20% (background and throttle). | |
de1fff37 | 933 | * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages. |
0d960a38 | 934 | * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is |
de1fff37 | 935 | * about ~6K pages (as the average of background and throttle wb |
5a537485 | 936 | * limits). The 3rd order polynomial will provide positive feedback if |
de1fff37 | 937 | * wb_dirty is under wb_setpoint and vice versa. |
6c14ae1e | 938 | * |
5a537485 | 939 | * Note, that we cannot use global counters in these calculations |
de1fff37 | 940 | * because we want to throttle process writing to a strictlimit wb |
5a537485 MP |
941 | * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB |
942 | * in the example above). | |
6c14ae1e | 943 | */ |
a88a341a | 944 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
de1fff37 | 945 | long long wb_pos_ratio; |
5a537485 | 946 | |
daddfa3c TH |
947 | if (dtc->wb_dirty < 8) { |
948 | dtc->pos_ratio = min_t(long long, pos_ratio * 2, | |
949 | 2 << RATELIMIT_CALC_SHIFT); | |
950 | return; | |
951 | } | |
5a537485 | 952 | |
2bc00aef | 953 | if (dtc->wb_dirty >= wb_thresh) |
daddfa3c | 954 | return; |
5a537485 | 955 | |
970fb01a TH |
956 | wb_setpoint = dirty_freerun_ceiling(wb_thresh, |
957 | dtc->wb_bg_thresh); | |
5a537485 | 958 | |
de1fff37 | 959 | if (wb_setpoint == 0 || wb_setpoint == wb_thresh) |
daddfa3c | 960 | return; |
5a537485 | 961 | |
2bc00aef | 962 | wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty, |
de1fff37 | 963 | wb_thresh); |
5a537485 MP |
964 | |
965 | /* | |
de1fff37 TH |
966 | * Typically, for strictlimit case, wb_setpoint << setpoint |
967 | * and pos_ratio >> wb_pos_ratio. In the other words global | |
5a537485 | 968 | * state ("dirty") is not limiting factor and we have to |
de1fff37 | 969 | * make decision based on wb counters. But there is an |
5a537485 MP |
970 | * important case when global pos_ratio should get precedence: |
971 | * global limits are exceeded (e.g. due to activities on other | |
de1fff37 | 972 | * wb's) while given strictlimit wb is below limit. |
5a537485 | 973 | * |
de1fff37 | 974 | * "pos_ratio * wb_pos_ratio" would work for the case above, |
5a537485 | 975 | * but it would look too non-natural for the case of all |
de1fff37 | 976 | * activity in the system coming from a single strictlimit wb |
5a537485 MP |
977 | * with bdi->max_ratio == 100%. |
978 | * | |
979 | * Note that min() below somewhat changes the dynamics of the | |
980 | * control system. Normally, pos_ratio value can be well over 3 | |
de1fff37 | 981 | * (when globally we are at freerun and wb is well below wb |
5a537485 MP |
982 | * setpoint). Now the maximum pos_ratio in the same situation |
983 | * is 2. We might want to tweak this if we observe the control | |
984 | * system is too slow to adapt. | |
985 | */ | |
daddfa3c TH |
986 | dtc->pos_ratio = min(pos_ratio, wb_pos_ratio); |
987 | return; | |
5a537485 | 988 | } |
6c14ae1e WF |
989 | |
990 | /* | |
991 | * We have computed basic pos_ratio above based on global situation. If | |
de1fff37 | 992 | * the wb is over/under its share of dirty pages, we want to scale |
6c14ae1e WF |
993 | * pos_ratio further down/up. That is done by the following mechanism. |
994 | */ | |
995 | ||
996 | /* | |
de1fff37 | 997 | * wb setpoint |
6c14ae1e | 998 | * |
de1fff37 | 999 | * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint) |
6c14ae1e | 1000 | * |
de1fff37 | 1001 | * x_intercept - wb_dirty |
6c14ae1e | 1002 | * := -------------------------- |
de1fff37 | 1003 | * x_intercept - wb_setpoint |
6c14ae1e | 1004 | * |
de1fff37 | 1005 | * The main wb control line is a linear function that subjects to |
6c14ae1e | 1006 | * |
de1fff37 TH |
1007 | * (1) f(wb_setpoint) = 1.0 |
1008 | * (2) k = - 1 / (8 * write_bw) (in single wb case) | |
1009 | * or equally: x_intercept = wb_setpoint + 8 * write_bw | |
6c14ae1e | 1010 | * |
de1fff37 | 1011 | * For single wb case, the dirty pages are observed to fluctuate |
6c14ae1e | 1012 | * regularly within range |
de1fff37 | 1013 | * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2] |
6c14ae1e WF |
1014 | * for various filesystems, where (2) can yield in a reasonable 12.5% |
1015 | * fluctuation range for pos_ratio. | |
1016 | * | |
de1fff37 | 1017 | * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its |
6c14ae1e | 1018 | * own size, so move the slope over accordingly and choose a slope that |
de1fff37 | 1019 | * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh. |
6c14ae1e | 1020 | */ |
2bc00aef TH |
1021 | if (unlikely(wb_thresh > dtc->thresh)) |
1022 | wb_thresh = dtc->thresh; | |
aed21ad2 | 1023 | /* |
de1fff37 | 1024 | * It's very possible that wb_thresh is close to 0 not because the |
aed21ad2 WF |
1025 | * device is slow, but that it has remained inactive for long time. |
1026 | * Honour such devices a reasonable good (hopefully IO efficient) | |
1027 | * threshold, so that the occasional writes won't be blocked and active | |
1028 | * writes can rampup the threshold quickly. | |
1029 | */ | |
2bc00aef | 1030 | wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8); |
6c14ae1e | 1031 | /* |
de1fff37 TH |
1032 | * scale global setpoint to wb's: |
1033 | * wb_setpoint = setpoint * wb_thresh / thresh | |
6c14ae1e | 1034 | */ |
e4bc13ad | 1035 | x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1); |
de1fff37 | 1036 | wb_setpoint = setpoint * (u64)x >> 16; |
6c14ae1e | 1037 | /* |
de1fff37 TH |
1038 | * Use span=(8*write_bw) in single wb case as indicated by |
1039 | * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case. | |
6c14ae1e | 1040 | * |
de1fff37 TH |
1041 | * wb_thresh thresh - wb_thresh |
1042 | * span = --------- * (8 * write_bw) + ------------------ * wb_thresh | |
1043 | * thresh thresh | |
6c14ae1e | 1044 | */ |
2bc00aef | 1045 | span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16; |
de1fff37 | 1046 | x_intercept = wb_setpoint + span; |
6c14ae1e | 1047 | |
2bc00aef TH |
1048 | if (dtc->wb_dirty < x_intercept - span / 4) { |
1049 | pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty), | |
e4bc13ad | 1050 | (x_intercept - wb_setpoint) | 1); |
6c14ae1e WF |
1051 | } else |
1052 | pos_ratio /= 4; | |
1053 | ||
8927f66c | 1054 | /* |
de1fff37 | 1055 | * wb reserve area, safeguard against dirty pool underrun and disk idle |
8927f66c WF |
1056 | * It may push the desired control point of global dirty pages higher |
1057 | * than setpoint. | |
1058 | */ | |
de1fff37 | 1059 | x_intercept = wb_thresh / 2; |
2bc00aef TH |
1060 | if (dtc->wb_dirty < x_intercept) { |
1061 | if (dtc->wb_dirty > x_intercept / 8) | |
1062 | pos_ratio = div_u64(pos_ratio * x_intercept, | |
1063 | dtc->wb_dirty); | |
50657fc4 | 1064 | else |
8927f66c WF |
1065 | pos_ratio *= 8; |
1066 | } | |
1067 | ||
daddfa3c | 1068 | dtc->pos_ratio = pos_ratio; |
6c14ae1e WF |
1069 | } |
1070 | ||
a88a341a TH |
1071 | static void wb_update_write_bandwidth(struct bdi_writeback *wb, |
1072 | unsigned long elapsed, | |
1073 | unsigned long written) | |
e98be2d5 WF |
1074 | { |
1075 | const unsigned long period = roundup_pow_of_two(3 * HZ); | |
a88a341a TH |
1076 | unsigned long avg = wb->avg_write_bandwidth; |
1077 | unsigned long old = wb->write_bandwidth; | |
e98be2d5 WF |
1078 | u64 bw; |
1079 | ||
1080 | /* | |
1081 | * bw = written * HZ / elapsed | |
1082 | * | |
1083 | * bw * elapsed + write_bandwidth * (period - elapsed) | |
1084 | * write_bandwidth = --------------------------------------------------- | |
1085 | * period | |
c72efb65 | 1086 | * |
25ff8b15 | 1087 | * @written may have decreased due to folio_account_redirty(). |
c72efb65 | 1088 | * Avoid underflowing @bw calculation. |
e98be2d5 | 1089 | */ |
a88a341a | 1090 | bw = written - min(written, wb->written_stamp); |
e98be2d5 WF |
1091 | bw *= HZ; |
1092 | if (unlikely(elapsed > period)) { | |
0a5d1a7f | 1093 | bw = div64_ul(bw, elapsed); |
e98be2d5 WF |
1094 | avg = bw; |
1095 | goto out; | |
1096 | } | |
a88a341a | 1097 | bw += (u64)wb->write_bandwidth * (period - elapsed); |
e98be2d5 WF |
1098 | bw >>= ilog2(period); |
1099 | ||
1100 | /* | |
1101 | * one more level of smoothing, for filtering out sudden spikes | |
1102 | */ | |
1103 | if (avg > old && old >= (unsigned long)bw) | |
1104 | avg -= (avg - old) >> 3; | |
1105 | ||
1106 | if (avg < old && old <= (unsigned long)bw) | |
1107 | avg += (old - avg) >> 3; | |
1108 | ||
1109 | out: | |
95a46c65 TH |
1110 | /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */ |
1111 | avg = max(avg, 1LU); | |
1112 | if (wb_has_dirty_io(wb)) { | |
1113 | long delta = avg - wb->avg_write_bandwidth; | |
1114 | WARN_ON_ONCE(atomic_long_add_return(delta, | |
1115 | &wb->bdi->tot_write_bandwidth) <= 0); | |
1116 | } | |
a88a341a | 1117 | wb->write_bandwidth = bw; |
20792ebf | 1118 | WRITE_ONCE(wb->avg_write_bandwidth, avg); |
e98be2d5 WF |
1119 | } |
1120 | ||
2bc00aef | 1121 | static void update_dirty_limit(struct dirty_throttle_control *dtc) |
c42843f2 | 1122 | { |
e9f07dfd | 1123 | struct wb_domain *dom = dtc_dom(dtc); |
2bc00aef | 1124 | unsigned long thresh = dtc->thresh; |
dcc25ae7 | 1125 | unsigned long limit = dom->dirty_limit; |
c42843f2 WF |
1126 | |
1127 | /* | |
1128 | * Follow up in one step. | |
1129 | */ | |
1130 | if (limit < thresh) { | |
1131 | limit = thresh; | |
1132 | goto update; | |
1133 | } | |
1134 | ||
1135 | /* | |
1136 | * Follow down slowly. Use the higher one as the target, because thresh | |
1137 | * may drop below dirty. This is exactly the reason to introduce | |
dcc25ae7 | 1138 | * dom->dirty_limit which is guaranteed to lie above the dirty pages. |
c42843f2 | 1139 | */ |
2bc00aef | 1140 | thresh = max(thresh, dtc->dirty); |
c42843f2 WF |
1141 | if (limit > thresh) { |
1142 | limit -= (limit - thresh) >> 5; | |
1143 | goto update; | |
1144 | } | |
1145 | return; | |
1146 | update: | |
dcc25ae7 | 1147 | dom->dirty_limit = limit; |
c42843f2 WF |
1148 | } |
1149 | ||
42dd235c JK |
1150 | static void domain_update_dirty_limit(struct dirty_throttle_control *dtc, |
1151 | unsigned long now) | |
c42843f2 | 1152 | { |
e9f07dfd | 1153 | struct wb_domain *dom = dtc_dom(dtc); |
c42843f2 WF |
1154 | |
1155 | /* | |
1156 | * check locklessly first to optimize away locking for the most time | |
1157 | */ | |
dcc25ae7 | 1158 | if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) |
c42843f2 WF |
1159 | return; |
1160 | ||
dcc25ae7 TH |
1161 | spin_lock(&dom->lock); |
1162 | if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) { | |
2bc00aef | 1163 | update_dirty_limit(dtc); |
dcc25ae7 | 1164 | dom->dirty_limit_tstamp = now; |
c42843f2 | 1165 | } |
dcc25ae7 | 1166 | spin_unlock(&dom->lock); |
c42843f2 WF |
1167 | } |
1168 | ||
be3ffa27 | 1169 | /* |
de1fff37 | 1170 | * Maintain wb->dirty_ratelimit, the base dirty throttle rate. |
be3ffa27 | 1171 | * |
de1fff37 | 1172 | * Normal wb tasks will be curbed at or below it in long term. |
be3ffa27 WF |
1173 | * Obviously it should be around (write_bw / N) when there are N dd tasks. |
1174 | */ | |
2bc00aef | 1175 | static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc, |
a88a341a TH |
1176 | unsigned long dirtied, |
1177 | unsigned long elapsed) | |
be3ffa27 | 1178 | { |
2bc00aef TH |
1179 | struct bdi_writeback *wb = dtc->wb; |
1180 | unsigned long dirty = dtc->dirty; | |
1181 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); | |
c7981433 | 1182 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
7381131c | 1183 | unsigned long setpoint = (freerun + limit) / 2; |
a88a341a TH |
1184 | unsigned long write_bw = wb->avg_write_bandwidth; |
1185 | unsigned long dirty_ratelimit = wb->dirty_ratelimit; | |
be3ffa27 WF |
1186 | unsigned long dirty_rate; |
1187 | unsigned long task_ratelimit; | |
1188 | unsigned long balanced_dirty_ratelimit; | |
7381131c WF |
1189 | unsigned long step; |
1190 | unsigned long x; | |
d59b1087 | 1191 | unsigned long shift; |
be3ffa27 WF |
1192 | |
1193 | /* | |
1194 | * The dirty rate will match the writeout rate in long term, except | |
1195 | * when dirty pages are truncated by userspace or re-dirtied by FS. | |
1196 | */ | |
a88a341a | 1197 | dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed; |
be3ffa27 | 1198 | |
be3ffa27 WF |
1199 | /* |
1200 | * task_ratelimit reflects each dd's dirty rate for the past 200ms. | |
1201 | */ | |
1202 | task_ratelimit = (u64)dirty_ratelimit * | |
daddfa3c | 1203 | dtc->pos_ratio >> RATELIMIT_CALC_SHIFT; |
be3ffa27 WF |
1204 | task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ |
1205 | ||
1206 | /* | |
1207 | * A linear estimation of the "balanced" throttle rate. The theory is, | |
de1fff37 | 1208 | * if there are N dd tasks, each throttled at task_ratelimit, the wb's |
be3ffa27 WF |
1209 | * dirty_rate will be measured to be (N * task_ratelimit). So the below |
1210 | * formula will yield the balanced rate limit (write_bw / N). | |
1211 | * | |
1212 | * Note that the expanded form is not a pure rate feedback: | |
1213 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) | |
1214 | * but also takes pos_ratio into account: | |
1215 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) | |
1216 | * | |
1217 | * (1) is not realistic because pos_ratio also takes part in balancing | |
1218 | * the dirty rate. Consider the state | |
1219 | * pos_ratio = 0.5 (3) | |
1220 | * rate = 2 * (write_bw / N) (4) | |
1221 | * If (1) is used, it will stuck in that state! Because each dd will | |
1222 | * be throttled at | |
1223 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) | |
1224 | * yielding | |
1225 | * dirty_rate = N * task_ratelimit = write_bw (6) | |
1226 | * put (6) into (1) we get | |
1227 | * rate_(i+1) = rate_(i) (7) | |
1228 | * | |
1229 | * So we end up using (2) to always keep | |
1230 | * rate_(i+1) ~= (write_bw / N) (8) | |
1231 | * regardless of the value of pos_ratio. As long as (8) is satisfied, | |
1232 | * pos_ratio is able to drive itself to 1.0, which is not only where | |
1233 | * the dirty count meet the setpoint, but also where the slope of | |
1234 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated. | |
1235 | */ | |
1236 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, | |
1237 | dirty_rate | 1); | |
bdaac490 WF |
1238 | /* |
1239 | * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw | |
1240 | */ | |
1241 | if (unlikely(balanced_dirty_ratelimit > write_bw)) | |
1242 | balanced_dirty_ratelimit = write_bw; | |
be3ffa27 | 1243 | |
7381131c WF |
1244 | /* |
1245 | * We could safely do this and return immediately: | |
1246 | * | |
de1fff37 | 1247 | * wb->dirty_ratelimit = balanced_dirty_ratelimit; |
7381131c WF |
1248 | * |
1249 | * However to get a more stable dirty_ratelimit, the below elaborated | |
331cbdee | 1250 | * code makes use of task_ratelimit to filter out singular points and |
7381131c WF |
1251 | * limit the step size. |
1252 | * | |
1253 | * The below code essentially only uses the relative value of | |
1254 | * | |
1255 | * task_ratelimit - dirty_ratelimit | |
1256 | * = (pos_ratio - 1) * dirty_ratelimit | |
1257 | * | |
1258 | * which reflects the direction and size of dirty position error. | |
1259 | */ | |
1260 | ||
1261 | /* | |
1262 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff | |
1263 | * task_ratelimit is on the same side of dirty_ratelimit, too. | |
1264 | * For example, when | |
1265 | * - dirty_ratelimit > balanced_dirty_ratelimit | |
1266 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) | |
1267 | * lowering dirty_ratelimit will help meet both the position and rate | |
1268 | * control targets. Otherwise, don't update dirty_ratelimit if it will | |
1269 | * only help meet the rate target. After all, what the users ultimately | |
1270 | * feel and care are stable dirty rate and small position error. | |
1271 | * | |
1272 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size | |
331cbdee | 1273 | * and filter out the singular points of balanced_dirty_ratelimit. Which |
7381131c WF |
1274 | * keeps jumping around randomly and can even leap far away at times |
1275 | * due to the small 200ms estimation period of dirty_rate (we want to | |
1276 | * keep that period small to reduce time lags). | |
1277 | */ | |
1278 | step = 0; | |
5a537485 MP |
1279 | |
1280 | /* | |
de1fff37 | 1281 | * For strictlimit case, calculations above were based on wb counters |
a88a341a | 1282 | * and limits (starting from pos_ratio = wb_position_ratio() and up to |
5a537485 | 1283 | * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate). |
de1fff37 TH |
1284 | * Hence, to calculate "step" properly, we have to use wb_dirty as |
1285 | * "dirty" and wb_setpoint as "setpoint". | |
5a537485 | 1286 | * |
de1fff37 TH |
1287 | * We rampup dirty_ratelimit forcibly if wb_dirty is low because |
1288 | * it's possible that wb_thresh is close to zero due to inactivity | |
970fb01a | 1289 | * of backing device. |
5a537485 | 1290 | */ |
a88a341a | 1291 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
2bc00aef TH |
1292 | dirty = dtc->wb_dirty; |
1293 | if (dtc->wb_dirty < 8) | |
1294 | setpoint = dtc->wb_dirty + 1; | |
5a537485 | 1295 | else |
970fb01a | 1296 | setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2; |
5a537485 MP |
1297 | } |
1298 | ||
7381131c | 1299 | if (dirty < setpoint) { |
a88a341a | 1300 | x = min3(wb->balanced_dirty_ratelimit, |
7c809968 | 1301 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1302 | if (dirty_ratelimit < x) |
1303 | step = x - dirty_ratelimit; | |
1304 | } else { | |
a88a341a | 1305 | x = max3(wb->balanced_dirty_ratelimit, |
7c809968 | 1306 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1307 | if (dirty_ratelimit > x) |
1308 | step = dirty_ratelimit - x; | |
1309 | } | |
1310 | ||
1311 | /* | |
1312 | * Don't pursue 100% rate matching. It's impossible since the balanced | |
1313 | * rate itself is constantly fluctuating. So decrease the track speed | |
1314 | * when it gets close to the target. Helps eliminate pointless tremors. | |
1315 | */ | |
d59b1087 AR |
1316 | shift = dirty_ratelimit / (2 * step + 1); |
1317 | if (shift < BITS_PER_LONG) | |
1318 | step = DIV_ROUND_UP(step >> shift, 8); | |
1319 | else | |
1320 | step = 0; | |
7381131c WF |
1321 | |
1322 | if (dirty_ratelimit < balanced_dirty_ratelimit) | |
1323 | dirty_ratelimit += step; | |
1324 | else | |
1325 | dirty_ratelimit -= step; | |
1326 | ||
20792ebf | 1327 | WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL)); |
a88a341a | 1328 | wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit; |
b48c104d | 1329 | |
5634cc2a | 1330 | trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit); |
be3ffa27 WF |
1331 | } |
1332 | ||
c2aa723a TH |
1333 | static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc, |
1334 | struct dirty_throttle_control *mdtc, | |
8a731799 | 1335 | bool update_ratelimit) |
e98be2d5 | 1336 | { |
c2aa723a | 1337 | struct bdi_writeback *wb = gdtc->wb; |
e98be2d5 | 1338 | unsigned long now = jiffies; |
45a2966f | 1339 | unsigned long elapsed; |
be3ffa27 | 1340 | unsigned long dirtied; |
e98be2d5 WF |
1341 | unsigned long written; |
1342 | ||
45a2966f | 1343 | spin_lock(&wb->list_lock); |
8a731799 | 1344 | |
e98be2d5 | 1345 | /* |
45a2966f JK |
1346 | * Lockless checks for elapsed time are racy and delayed update after |
1347 | * IO completion doesn't do it at all (to make sure written pages are | |
1348 | * accounted reasonably quickly). Make sure elapsed >= 1 to avoid | |
1349 | * division errors. | |
e98be2d5 | 1350 | */ |
45a2966f | 1351 | elapsed = max(now - wb->bw_time_stamp, 1UL); |
a88a341a TH |
1352 | dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]); |
1353 | written = percpu_counter_read(&wb->stat[WB_WRITTEN]); | |
e98be2d5 | 1354 | |
8a731799 | 1355 | if (update_ratelimit) { |
42dd235c | 1356 | domain_update_dirty_limit(gdtc, now); |
c2aa723a TH |
1357 | wb_update_dirty_ratelimit(gdtc, dirtied, elapsed); |
1358 | ||
1359 | /* | |
1360 | * @mdtc is always NULL if !CGROUP_WRITEBACK but the | |
1361 | * compiler has no way to figure that out. Help it. | |
1362 | */ | |
1363 | if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) { | |
42dd235c | 1364 | domain_update_dirty_limit(mdtc, now); |
c2aa723a TH |
1365 | wb_update_dirty_ratelimit(mdtc, dirtied, elapsed); |
1366 | } | |
be3ffa27 | 1367 | } |
a88a341a | 1368 | wb_update_write_bandwidth(wb, elapsed, written); |
e98be2d5 | 1369 | |
a88a341a TH |
1370 | wb->dirtied_stamp = dirtied; |
1371 | wb->written_stamp = written; | |
20792ebf | 1372 | WRITE_ONCE(wb->bw_time_stamp, now); |
45a2966f | 1373 | spin_unlock(&wb->list_lock); |
e98be2d5 WF |
1374 | } |
1375 | ||
45a2966f | 1376 | void wb_update_bandwidth(struct bdi_writeback *wb) |
e98be2d5 | 1377 | { |
2bc00aef TH |
1378 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
1379 | ||
fee468fd | 1380 | __wb_update_bandwidth(&gdtc, NULL, false); |
fee468fd JK |
1381 | } |
1382 | ||
1383 | /* Interval after which we consider wb idle and don't estimate bandwidth */ | |
1384 | #define WB_BANDWIDTH_IDLE_JIF (HZ) | |
1385 | ||
1386 | static void wb_bandwidth_estimate_start(struct bdi_writeback *wb) | |
1387 | { | |
1388 | unsigned long now = jiffies; | |
1389 | unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp); | |
1390 | ||
1391 | if (elapsed > WB_BANDWIDTH_IDLE_JIF && | |
1392 | !atomic_read(&wb->writeback_inodes)) { | |
1393 | spin_lock(&wb->list_lock); | |
1394 | wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED); | |
1395 | wb->written_stamp = wb_stat(wb, WB_WRITTEN); | |
20792ebf | 1396 | WRITE_ONCE(wb->bw_time_stamp, now); |
fee468fd JK |
1397 | spin_unlock(&wb->list_lock); |
1398 | } | |
e98be2d5 WF |
1399 | } |
1400 | ||
9d823e8f | 1401 | /* |
d0e1d66b | 1402 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited() |
9d823e8f WF |
1403 | * will look to see if it needs to start dirty throttling. |
1404 | * | |
1405 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive | |
c41f012a | 1406 | * global_zone_page_state() too often. So scale it near-sqrt to the safety margin |
9d823e8f WF |
1407 | * (the number of pages we may dirty without exceeding the dirty limits). |
1408 | */ | |
1409 | static unsigned long dirty_poll_interval(unsigned long dirty, | |
1410 | unsigned long thresh) | |
1411 | { | |
1412 | if (thresh > dirty) | |
1413 | return 1UL << (ilog2(thresh - dirty) >> 1); | |
1414 | ||
1415 | return 1; | |
1416 | } | |
1417 | ||
a88a341a | 1418 | static unsigned long wb_max_pause(struct bdi_writeback *wb, |
de1fff37 | 1419 | unsigned long wb_dirty) |
c8462cc9 | 1420 | { |
20792ebf | 1421 | unsigned long bw = READ_ONCE(wb->avg_write_bandwidth); |
e3b6c655 | 1422 | unsigned long t; |
c8462cc9 | 1423 | |
7ccb9ad5 WF |
1424 | /* |
1425 | * Limit pause time for small memory systems. If sleeping for too long | |
1426 | * time, a small pool of dirty/writeback pages may go empty and disk go | |
1427 | * idle. | |
1428 | * | |
1429 | * 8 serves as the safety ratio. | |
1430 | */ | |
de1fff37 | 1431 | t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8)); |
7ccb9ad5 WF |
1432 | t++; |
1433 | ||
e3b6c655 | 1434 | return min_t(unsigned long, t, MAX_PAUSE); |
7ccb9ad5 WF |
1435 | } |
1436 | ||
a88a341a TH |
1437 | static long wb_min_pause(struct bdi_writeback *wb, |
1438 | long max_pause, | |
1439 | unsigned long task_ratelimit, | |
1440 | unsigned long dirty_ratelimit, | |
1441 | int *nr_dirtied_pause) | |
c8462cc9 | 1442 | { |
20792ebf JK |
1443 | long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth)); |
1444 | long lo = ilog2(READ_ONCE(wb->dirty_ratelimit)); | |
7ccb9ad5 WF |
1445 | long t; /* target pause */ |
1446 | long pause; /* estimated next pause */ | |
1447 | int pages; /* target nr_dirtied_pause */ | |
c8462cc9 | 1448 | |
7ccb9ad5 WF |
1449 | /* target for 10ms pause on 1-dd case */ |
1450 | t = max(1, HZ / 100); | |
c8462cc9 WF |
1451 | |
1452 | /* | |
1453 | * Scale up pause time for concurrent dirtiers in order to reduce CPU | |
1454 | * overheads. | |
1455 | * | |
7ccb9ad5 | 1456 | * (N * 10ms) on 2^N concurrent tasks. |
c8462cc9 WF |
1457 | */ |
1458 | if (hi > lo) | |
7ccb9ad5 | 1459 | t += (hi - lo) * (10 * HZ) / 1024; |
c8462cc9 WF |
1460 | |
1461 | /* | |
7ccb9ad5 WF |
1462 | * This is a bit convoluted. We try to base the next nr_dirtied_pause |
1463 | * on the much more stable dirty_ratelimit. However the next pause time | |
1464 | * will be computed based on task_ratelimit and the two rate limits may | |
1465 | * depart considerably at some time. Especially if task_ratelimit goes | |
1466 | * below dirty_ratelimit/2 and the target pause is max_pause, the next | |
1467 | * pause time will be max_pause*2 _trimmed down_ to max_pause. As a | |
1468 | * result task_ratelimit won't be executed faithfully, which could | |
1469 | * eventually bring down dirty_ratelimit. | |
c8462cc9 | 1470 | * |
7ccb9ad5 WF |
1471 | * We apply two rules to fix it up: |
1472 | * 1) try to estimate the next pause time and if necessary, use a lower | |
1473 | * nr_dirtied_pause so as not to exceed max_pause. When this happens, | |
1474 | * nr_dirtied_pause will be "dancing" with task_ratelimit. | |
1475 | * 2) limit the target pause time to max_pause/2, so that the normal | |
1476 | * small fluctuations of task_ratelimit won't trigger rule (1) and | |
1477 | * nr_dirtied_pause will remain as stable as dirty_ratelimit. | |
c8462cc9 | 1478 | */ |
7ccb9ad5 WF |
1479 | t = min(t, 1 + max_pause / 2); |
1480 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
c8462cc9 WF |
1481 | |
1482 | /* | |
5b9b3574 WF |
1483 | * Tiny nr_dirtied_pause is found to hurt I/O performance in the test |
1484 | * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}. | |
1485 | * When the 16 consecutive reads are often interrupted by some dirty | |
1486 | * throttling pause during the async writes, cfq will go into idles | |
1487 | * (deadline is fine). So push nr_dirtied_pause as high as possible | |
1488 | * until reaches DIRTY_POLL_THRESH=32 pages. | |
c8462cc9 | 1489 | */ |
5b9b3574 WF |
1490 | if (pages < DIRTY_POLL_THRESH) { |
1491 | t = max_pause; | |
1492 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
1493 | if (pages > DIRTY_POLL_THRESH) { | |
1494 | pages = DIRTY_POLL_THRESH; | |
1495 | t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit; | |
1496 | } | |
1497 | } | |
1498 | ||
7ccb9ad5 WF |
1499 | pause = HZ * pages / (task_ratelimit + 1); |
1500 | if (pause > max_pause) { | |
1501 | t = max_pause; | |
1502 | pages = task_ratelimit * t / roundup_pow_of_two(HZ); | |
1503 | } | |
c8462cc9 | 1504 | |
7ccb9ad5 | 1505 | *nr_dirtied_pause = pages; |
c8462cc9 | 1506 | /* |
7ccb9ad5 | 1507 | * The minimal pause time will normally be half the target pause time. |
c8462cc9 | 1508 | */ |
5b9b3574 | 1509 | return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t; |
c8462cc9 WF |
1510 | } |
1511 | ||
970fb01a | 1512 | static inline void wb_dirty_limits(struct dirty_throttle_control *dtc) |
5a537485 | 1513 | { |
2bc00aef | 1514 | struct bdi_writeback *wb = dtc->wb; |
93f78d88 | 1515 | unsigned long wb_reclaimable; |
5a537485 MP |
1516 | |
1517 | /* | |
de1fff37 | 1518 | * wb_thresh is not treated as some limiting factor as |
5a537485 | 1519 | * dirty_thresh, due to reasons |
de1fff37 | 1520 | * - in JBOD setup, wb_thresh can fluctuate a lot |
5a537485 | 1521 | * - in a system with HDD and USB key, the USB key may somehow |
de1fff37 TH |
1522 | * go into state (wb_dirty >> wb_thresh) either because |
1523 | * wb_dirty starts high, or because wb_thresh drops low. | |
5a537485 | 1524 | * In this case we don't want to hard throttle the USB key |
de1fff37 TH |
1525 | * dirtiers for 100 seconds until wb_dirty drops under |
1526 | * wb_thresh. Instead the auxiliary wb control line in | |
a88a341a | 1527 | * wb_position_ratio() will let the dirtier task progress |
de1fff37 | 1528 | * at some rate <= (write_bw / 2) for bringing down wb_dirty. |
5a537485 | 1529 | */ |
b1cbc6d4 | 1530 | dtc->wb_thresh = __wb_calc_thresh(dtc); |
970fb01a TH |
1531 | dtc->wb_bg_thresh = dtc->thresh ? |
1532 | div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0; | |
5a537485 MP |
1533 | |
1534 | /* | |
1535 | * In order to avoid the stacked BDI deadlock we need | |
1536 | * to ensure we accurately count the 'dirty' pages when | |
1537 | * the threshold is low. | |
1538 | * | |
1539 | * Otherwise it would be possible to get thresh+n pages | |
1540 | * reported dirty, even though there are thresh-m pages | |
1541 | * actually dirty; with m+n sitting in the percpu | |
1542 | * deltas. | |
1543 | */ | |
2bce774e | 1544 | if (dtc->wb_thresh < 2 * wb_stat_error()) { |
93f78d88 | 1545 | wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); |
2bc00aef | 1546 | dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK); |
5a537485 | 1547 | } else { |
93f78d88 | 1548 | wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE); |
2bc00aef | 1549 | dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK); |
5a537485 MP |
1550 | } |
1551 | } | |
1552 | ||
1da177e4 LT |
1553 | /* |
1554 | * balance_dirty_pages() must be called by processes which are generating dirty | |
1555 | * data. It looks at the number of dirty pages in the machine and will force | |
143dfe86 | 1556 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. |
5b0830cb JA |
1557 | * If we're over `background_thresh' then the writeback threads are woken to |
1558 | * perform some writeout. | |
1da177e4 | 1559 | */ |
4c578dce | 1560 | static void balance_dirty_pages(struct bdi_writeback *wb, |
143dfe86 | 1561 | unsigned long pages_dirtied) |
1da177e4 | 1562 | { |
2bc00aef | 1563 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) }; |
c2aa723a | 1564 | struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) }; |
2bc00aef | 1565 | struct dirty_throttle_control * const gdtc = &gdtc_stor; |
c2aa723a TH |
1566 | struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ? |
1567 | &mdtc_stor : NULL; | |
1568 | struct dirty_throttle_control *sdtc; | |
8d92890b | 1569 | unsigned long nr_reclaimable; /* = file_dirty */ |
83712358 | 1570 | long period; |
7ccb9ad5 WF |
1571 | long pause; |
1572 | long max_pause; | |
1573 | long min_pause; | |
1574 | int nr_dirtied_pause; | |
e50e3720 | 1575 | bool dirty_exceeded = false; |
143dfe86 | 1576 | unsigned long task_ratelimit; |
7ccb9ad5 | 1577 | unsigned long dirty_ratelimit; |
dfb8ae56 | 1578 | struct backing_dev_info *bdi = wb->bdi; |
5a537485 | 1579 | bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT; |
e98be2d5 | 1580 | unsigned long start_time = jiffies; |
1da177e4 LT |
1581 | |
1582 | for (;;) { | |
83712358 | 1583 | unsigned long now = jiffies; |
2bc00aef | 1584 | unsigned long dirty, thresh, bg_thresh; |
50e55bf6 YS |
1585 | unsigned long m_dirty = 0; /* stop bogus uninit warnings */ |
1586 | unsigned long m_thresh = 0; | |
1587 | unsigned long m_bg_thresh = 0; | |
83712358 | 1588 | |
8d92890b | 1589 | nr_reclaimable = global_node_page_state(NR_FILE_DIRTY); |
9fc3a43e | 1590 | gdtc->avail = global_dirtyable_memory(); |
11fb9989 | 1591 | gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK); |
5fce25a9 | 1592 | |
9fc3a43e | 1593 | domain_dirty_limits(gdtc); |
16c4042f | 1594 | |
5a537485 | 1595 | if (unlikely(strictlimit)) { |
970fb01a | 1596 | wb_dirty_limits(gdtc); |
5a537485 | 1597 | |
2bc00aef TH |
1598 | dirty = gdtc->wb_dirty; |
1599 | thresh = gdtc->wb_thresh; | |
970fb01a | 1600 | bg_thresh = gdtc->wb_bg_thresh; |
5a537485 | 1601 | } else { |
2bc00aef TH |
1602 | dirty = gdtc->dirty; |
1603 | thresh = gdtc->thresh; | |
1604 | bg_thresh = gdtc->bg_thresh; | |
5a537485 MP |
1605 | } |
1606 | ||
c2aa723a | 1607 | if (mdtc) { |
c5edf9cd | 1608 | unsigned long filepages, headroom, writeback; |
c2aa723a TH |
1609 | |
1610 | /* | |
1611 | * If @wb belongs to !root memcg, repeat the same | |
1612 | * basic calculations for the memcg domain. | |
1613 | */ | |
c5edf9cd TH |
1614 | mem_cgroup_wb_stats(wb, &filepages, &headroom, |
1615 | &mdtc->dirty, &writeback); | |
c2aa723a | 1616 | mdtc->dirty += writeback; |
c5edf9cd | 1617 | mdtc_calc_avail(mdtc, filepages, headroom); |
c2aa723a TH |
1618 | |
1619 | domain_dirty_limits(mdtc); | |
1620 | ||
1621 | if (unlikely(strictlimit)) { | |
1622 | wb_dirty_limits(mdtc); | |
1623 | m_dirty = mdtc->wb_dirty; | |
1624 | m_thresh = mdtc->wb_thresh; | |
1625 | m_bg_thresh = mdtc->wb_bg_thresh; | |
1626 | } else { | |
1627 | m_dirty = mdtc->dirty; | |
1628 | m_thresh = mdtc->thresh; | |
1629 | m_bg_thresh = mdtc->bg_thresh; | |
1630 | } | |
5a537485 MP |
1631 | } |
1632 | ||
16c4042f WF |
1633 | /* |
1634 | * Throttle it only when the background writeback cannot | |
1635 | * catch-up. This avoids (excessively) small writeouts | |
de1fff37 | 1636 | * when the wb limits are ramping up in case of !strictlimit. |
5a537485 | 1637 | * |
de1fff37 TH |
1638 | * In strictlimit case make decision based on the wb counters |
1639 | * and limits. Small writeouts when the wb limits are ramping | |
5a537485 | 1640 | * up are the price we consciously pay for strictlimit-ing. |
c2aa723a TH |
1641 | * |
1642 | * If memcg domain is in effect, @dirty should be under | |
1643 | * both global and memcg freerun ceilings. | |
16c4042f | 1644 | */ |
c2aa723a TH |
1645 | if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) && |
1646 | (!mdtc || | |
1647 | m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) { | |
a37b0715 N |
1648 | unsigned long intv; |
1649 | unsigned long m_intv; | |
1650 | ||
1651 | free_running: | |
1652 | intv = dirty_poll_interval(dirty, thresh); | |
1653 | m_intv = ULONG_MAX; | |
c2aa723a | 1654 | |
83712358 WF |
1655 | current->dirty_paused_when = now; |
1656 | current->nr_dirtied = 0; | |
c2aa723a TH |
1657 | if (mdtc) |
1658 | m_intv = dirty_poll_interval(m_dirty, m_thresh); | |
1659 | current->nr_dirtied_pause = min(intv, m_intv); | |
16c4042f | 1660 | break; |
83712358 | 1661 | } |
16c4042f | 1662 | |
bc05873d | 1663 | if (unlikely(!writeback_in_progress(wb))) |
9ecf4866 | 1664 | wb_start_background_writeback(wb); |
143dfe86 | 1665 | |
97b27821 TH |
1666 | mem_cgroup_flush_foreign(wb); |
1667 | ||
c2aa723a TH |
1668 | /* |
1669 | * Calculate global domain's pos_ratio and select the | |
1670 | * global dtc by default. | |
1671 | */ | |
a37b0715 | 1672 | if (!strictlimit) { |
970fb01a | 1673 | wb_dirty_limits(gdtc); |
5fce25a9 | 1674 | |
a37b0715 N |
1675 | if ((current->flags & PF_LOCAL_THROTTLE) && |
1676 | gdtc->wb_dirty < | |
1677 | dirty_freerun_ceiling(gdtc->wb_thresh, | |
1678 | gdtc->wb_bg_thresh)) | |
1679 | /* | |
1680 | * LOCAL_THROTTLE tasks must not be throttled | |
1681 | * when below the per-wb freerun ceiling. | |
1682 | */ | |
1683 | goto free_running; | |
1684 | } | |
1685 | ||
2bc00aef TH |
1686 | dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) && |
1687 | ((gdtc->dirty > gdtc->thresh) || strictlimit); | |
daddfa3c TH |
1688 | |
1689 | wb_position_ratio(gdtc); | |
c2aa723a TH |
1690 | sdtc = gdtc; |
1691 | ||
1692 | if (mdtc) { | |
1693 | /* | |
1694 | * If memcg domain is in effect, calculate its | |
1695 | * pos_ratio. @wb should satisfy constraints from | |
1696 | * both global and memcg domains. Choose the one | |
1697 | * w/ lower pos_ratio. | |
1698 | */ | |
a37b0715 | 1699 | if (!strictlimit) { |
c2aa723a TH |
1700 | wb_dirty_limits(mdtc); |
1701 | ||
a37b0715 N |
1702 | if ((current->flags & PF_LOCAL_THROTTLE) && |
1703 | mdtc->wb_dirty < | |
1704 | dirty_freerun_ceiling(mdtc->wb_thresh, | |
1705 | mdtc->wb_bg_thresh)) | |
1706 | /* | |
1707 | * LOCAL_THROTTLE tasks must not be | |
1708 | * throttled when below the per-wb | |
1709 | * freerun ceiling. | |
1710 | */ | |
1711 | goto free_running; | |
1712 | } | |
c2aa723a TH |
1713 | dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) && |
1714 | ((mdtc->dirty > mdtc->thresh) || strictlimit); | |
1715 | ||
1716 | wb_position_ratio(mdtc); | |
1717 | if (mdtc->pos_ratio < gdtc->pos_ratio) | |
1718 | sdtc = mdtc; | |
1719 | } | |
daddfa3c | 1720 | |
a88a341a TH |
1721 | if (dirty_exceeded && !wb->dirty_exceeded) |
1722 | wb->dirty_exceeded = 1; | |
1da177e4 | 1723 | |
20792ebf | 1724 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) + |
45a2966f | 1725 | BANDWIDTH_INTERVAL)) |
fee468fd | 1726 | __wb_update_bandwidth(gdtc, mdtc, true); |
e98be2d5 | 1727 | |
c2aa723a | 1728 | /* throttle according to the chosen dtc */ |
20792ebf | 1729 | dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit); |
c2aa723a | 1730 | task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >> |
3a73dbbc | 1731 | RATELIMIT_CALC_SHIFT; |
c2aa723a | 1732 | max_pause = wb_max_pause(wb, sdtc->wb_dirty); |
a88a341a TH |
1733 | min_pause = wb_min_pause(wb, max_pause, |
1734 | task_ratelimit, dirty_ratelimit, | |
1735 | &nr_dirtied_pause); | |
7ccb9ad5 | 1736 | |
3a73dbbc | 1737 | if (unlikely(task_ratelimit == 0)) { |
83712358 | 1738 | period = max_pause; |
c8462cc9 | 1739 | pause = max_pause; |
143dfe86 | 1740 | goto pause; |
04fbfdc1 | 1741 | } |
83712358 WF |
1742 | period = HZ * pages_dirtied / task_ratelimit; |
1743 | pause = period; | |
1744 | if (current->dirty_paused_when) | |
1745 | pause -= now - current->dirty_paused_when; | |
1746 | /* | |
1747 | * For less than 1s think time (ext3/4 may block the dirtier | |
1748 | * for up to 800ms from time to time on 1-HDD; so does xfs, | |
1749 | * however at much less frequency), try to compensate it in | |
1750 | * future periods by updating the virtual time; otherwise just | |
1751 | * do a reset, as it may be a light dirtier. | |
1752 | */ | |
7ccb9ad5 | 1753 | if (pause < min_pause) { |
5634cc2a | 1754 | trace_balance_dirty_pages(wb, |
c2aa723a TH |
1755 | sdtc->thresh, |
1756 | sdtc->bg_thresh, | |
1757 | sdtc->dirty, | |
1758 | sdtc->wb_thresh, | |
1759 | sdtc->wb_dirty, | |
ece13ac3 WF |
1760 | dirty_ratelimit, |
1761 | task_ratelimit, | |
1762 | pages_dirtied, | |
83712358 | 1763 | period, |
7ccb9ad5 | 1764 | min(pause, 0L), |
ece13ac3 | 1765 | start_time); |
83712358 WF |
1766 | if (pause < -HZ) { |
1767 | current->dirty_paused_when = now; | |
1768 | current->nr_dirtied = 0; | |
1769 | } else if (period) { | |
1770 | current->dirty_paused_when += period; | |
1771 | current->nr_dirtied = 0; | |
7ccb9ad5 WF |
1772 | } else if (current->nr_dirtied_pause <= pages_dirtied) |
1773 | current->nr_dirtied_pause += pages_dirtied; | |
57fc978c | 1774 | break; |
04fbfdc1 | 1775 | } |
7ccb9ad5 WF |
1776 | if (unlikely(pause > max_pause)) { |
1777 | /* for occasional dropped task_ratelimit */ | |
1778 | now += min(pause - max_pause, max_pause); | |
1779 | pause = max_pause; | |
1780 | } | |
143dfe86 WF |
1781 | |
1782 | pause: | |
5634cc2a | 1783 | trace_balance_dirty_pages(wb, |
c2aa723a TH |
1784 | sdtc->thresh, |
1785 | sdtc->bg_thresh, | |
1786 | sdtc->dirty, | |
1787 | sdtc->wb_thresh, | |
1788 | sdtc->wb_dirty, | |
ece13ac3 WF |
1789 | dirty_ratelimit, |
1790 | task_ratelimit, | |
1791 | pages_dirtied, | |
83712358 | 1792 | period, |
ece13ac3 WF |
1793 | pause, |
1794 | start_time); | |
499d05ec | 1795 | __set_current_state(TASK_KILLABLE); |
b57d74af | 1796 | wb->dirty_sleep = now; |
d25105e8 | 1797 | io_schedule_timeout(pause); |
87c6a9b2 | 1798 | |
83712358 WF |
1799 | current->dirty_paused_when = now + pause; |
1800 | current->nr_dirtied = 0; | |
7ccb9ad5 | 1801 | current->nr_dirtied_pause = nr_dirtied_pause; |
83712358 | 1802 | |
ffd1f609 | 1803 | /* |
2bc00aef TH |
1804 | * This is typically equal to (dirty < thresh) and can also |
1805 | * keep "1000+ dd on a slow USB stick" under control. | |
ffd1f609 | 1806 | */ |
1df64719 | 1807 | if (task_ratelimit) |
ffd1f609 | 1808 | break; |
499d05ec | 1809 | |
c5c6343c | 1810 | /* |
f0953a1b | 1811 | * In the case of an unresponsive NFS server and the NFS dirty |
de1fff37 | 1812 | * pages exceeds dirty_thresh, give the other good wb's a pipe |
c5c6343c WF |
1813 | * to go through, so that tasks on them still remain responsive. |
1814 | * | |
3f8b6fb7 | 1815 | * In theory 1 page is enough to keep the consumer-producer |
c5c6343c | 1816 | * pipe going: the flusher cleans 1 page => the task dirties 1 |
de1fff37 | 1817 | * more page. However wb_dirty has accounting errors. So use |
93f78d88 | 1818 | * the larger and more IO friendly wb_stat_error. |
c5c6343c | 1819 | */ |
2bce774e | 1820 | if (sdtc->wb_dirty <= wb_stat_error()) |
c5c6343c WF |
1821 | break; |
1822 | ||
499d05ec JK |
1823 | if (fatal_signal_pending(current)) |
1824 | break; | |
1da177e4 LT |
1825 | } |
1826 | ||
a88a341a TH |
1827 | if (!dirty_exceeded && wb->dirty_exceeded) |
1828 | wb->dirty_exceeded = 0; | |
1da177e4 | 1829 | |
bc05873d | 1830 | if (writeback_in_progress(wb)) |
5b0830cb | 1831 | return; |
1da177e4 LT |
1832 | |
1833 | /* | |
1834 | * In laptop mode, we wait until hitting the higher threshold before | |
1835 | * starting background writeout, and then write out all the way down | |
1836 | * to the lower threshold. So slow writers cause minimal disk activity. | |
1837 | * | |
1838 | * In normal mode, we start background writeout at the lower | |
1839 | * background_thresh, to keep the amount of dirty memory low. | |
1840 | */ | |
143dfe86 WF |
1841 | if (laptop_mode) |
1842 | return; | |
1843 | ||
2bc00aef | 1844 | if (nr_reclaimable > gdtc->bg_thresh) |
9ecf4866 | 1845 | wb_start_background_writeback(wb); |
1da177e4 LT |
1846 | } |
1847 | ||
9d823e8f | 1848 | static DEFINE_PER_CPU(int, bdp_ratelimits); |
245b2e70 | 1849 | |
54848d73 WF |
1850 | /* |
1851 | * Normal tasks are throttled by | |
1852 | * loop { | |
1853 | * dirty tsk->nr_dirtied_pause pages; | |
1854 | * take a snap in balance_dirty_pages(); | |
1855 | * } | |
1856 | * However there is a worst case. If every task exit immediately when dirtied | |
1857 | * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be | |
1858 | * called to throttle the page dirties. The solution is to save the not yet | |
1859 | * throttled page dirties in dirty_throttle_leaks on task exit and charge them | |
1860 | * randomly into the running tasks. This works well for the above worst case, | |
1861 | * as the new task will pick up and accumulate the old task's leaked dirty | |
1862 | * count and eventually get throttled. | |
1863 | */ | |
1864 | DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0; | |
1865 | ||
1da177e4 | 1866 | /** |
d0e1d66b | 1867 | * balance_dirty_pages_ratelimited - balance dirty memory state |
67be2dd1 | 1868 | * @mapping: address_space which was dirtied |
1da177e4 LT |
1869 | * |
1870 | * Processes which are dirtying memory should call in here once for each page | |
1871 | * which was newly dirtied. The function will periodically check the system's | |
1872 | * dirty state and will initiate writeback if needed. | |
1873 | * | |
5defd497 KW |
1874 | * Once we're over the dirty memory limit we decrease the ratelimiting |
1875 | * by a lot, to prevent individual processes from overshooting the limit | |
1876 | * by (ratelimit_pages) each. | |
1da177e4 | 1877 | */ |
d0e1d66b | 1878 | void balance_dirty_pages_ratelimited(struct address_space *mapping) |
1da177e4 | 1879 | { |
dfb8ae56 TH |
1880 | struct inode *inode = mapping->host; |
1881 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1882 | struct bdi_writeback *wb = NULL; | |
9d823e8f WF |
1883 | int ratelimit; |
1884 | int *p; | |
1da177e4 | 1885 | |
f56753ac | 1886 | if (!(bdi->capabilities & BDI_CAP_WRITEBACK)) |
36715cef WF |
1887 | return; |
1888 | ||
dfb8ae56 TH |
1889 | if (inode_cgwb_enabled(inode)) |
1890 | wb = wb_get_create_current(bdi, GFP_KERNEL); | |
1891 | if (!wb) | |
1892 | wb = &bdi->wb; | |
1893 | ||
9d823e8f | 1894 | ratelimit = current->nr_dirtied_pause; |
a88a341a | 1895 | if (wb->dirty_exceeded) |
9d823e8f WF |
1896 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); |
1897 | ||
9d823e8f | 1898 | preempt_disable(); |
1da177e4 | 1899 | /* |
9d823e8f WF |
1900 | * This prevents one CPU to accumulate too many dirtied pages without |
1901 | * calling into balance_dirty_pages(), which can happen when there are | |
1902 | * 1000+ tasks, all of them start dirtying pages at exactly the same | |
1903 | * time, hence all honoured too large initial task->nr_dirtied_pause. | |
1da177e4 | 1904 | */ |
7c8e0181 | 1905 | p = this_cpu_ptr(&bdp_ratelimits); |
9d823e8f | 1906 | if (unlikely(current->nr_dirtied >= ratelimit)) |
fa5a734e | 1907 | *p = 0; |
d3bc1fef WF |
1908 | else if (unlikely(*p >= ratelimit_pages)) { |
1909 | *p = 0; | |
1910 | ratelimit = 0; | |
1da177e4 | 1911 | } |
54848d73 WF |
1912 | /* |
1913 | * Pick up the dirtied pages by the exited tasks. This avoids lots of | |
1914 | * short-lived tasks (eg. gcc invocations in a kernel build) escaping | |
1915 | * the dirty throttling and livelock other long-run dirtiers. | |
1916 | */ | |
7c8e0181 | 1917 | p = this_cpu_ptr(&dirty_throttle_leaks); |
54848d73 | 1918 | if (*p > 0 && current->nr_dirtied < ratelimit) { |
d0e1d66b | 1919 | unsigned long nr_pages_dirtied; |
54848d73 WF |
1920 | nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied); |
1921 | *p -= nr_pages_dirtied; | |
1922 | current->nr_dirtied += nr_pages_dirtied; | |
1da177e4 | 1923 | } |
fa5a734e | 1924 | preempt_enable(); |
9d823e8f WF |
1925 | |
1926 | if (unlikely(current->nr_dirtied >= ratelimit)) | |
4c578dce | 1927 | balance_dirty_pages(wb, current->nr_dirtied); |
dfb8ae56 TH |
1928 | |
1929 | wb_put(wb); | |
1da177e4 | 1930 | } |
d0e1d66b | 1931 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited); |
1da177e4 | 1932 | |
aa661bbe TH |
1933 | /** |
1934 | * wb_over_bg_thresh - does @wb need to be written back? | |
1935 | * @wb: bdi_writeback of interest | |
1936 | * | |
1937 | * Determines whether background writeback should keep writing @wb or it's | |
a862f68a MR |
1938 | * clean enough. |
1939 | * | |
1940 | * Return: %true if writeback should continue. | |
aa661bbe TH |
1941 | */ |
1942 | bool wb_over_bg_thresh(struct bdi_writeback *wb) | |
1943 | { | |
947e9762 | 1944 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) }; |
c2aa723a | 1945 | struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) }; |
947e9762 | 1946 | struct dirty_throttle_control * const gdtc = &gdtc_stor; |
c2aa723a TH |
1947 | struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ? |
1948 | &mdtc_stor : NULL; | |
ab19939a CW |
1949 | unsigned long reclaimable; |
1950 | unsigned long thresh; | |
aa661bbe | 1951 | |
947e9762 TH |
1952 | /* |
1953 | * Similar to balance_dirty_pages() but ignores pages being written | |
1954 | * as we're trying to decide whether to put more under writeback. | |
1955 | */ | |
1956 | gdtc->avail = global_dirtyable_memory(); | |
8d92890b | 1957 | gdtc->dirty = global_node_page_state(NR_FILE_DIRTY); |
947e9762 | 1958 | domain_dirty_limits(gdtc); |
aa661bbe | 1959 | |
947e9762 | 1960 | if (gdtc->dirty > gdtc->bg_thresh) |
aa661bbe TH |
1961 | return true; |
1962 | ||
ab19939a CW |
1963 | thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh); |
1964 | if (thresh < 2 * wb_stat_error()) | |
1965 | reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); | |
1966 | else | |
1967 | reclaimable = wb_stat(wb, WB_RECLAIMABLE); | |
1968 | ||
1969 | if (reclaimable > thresh) | |
aa661bbe TH |
1970 | return true; |
1971 | ||
c2aa723a | 1972 | if (mdtc) { |
c5edf9cd | 1973 | unsigned long filepages, headroom, writeback; |
c2aa723a | 1974 | |
c5edf9cd TH |
1975 | mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty, |
1976 | &writeback); | |
1977 | mdtc_calc_avail(mdtc, filepages, headroom); | |
c2aa723a TH |
1978 | domain_dirty_limits(mdtc); /* ditto, ignore writeback */ |
1979 | ||
1980 | if (mdtc->dirty > mdtc->bg_thresh) | |
1981 | return true; | |
1982 | ||
ab19939a CW |
1983 | thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh); |
1984 | if (thresh < 2 * wb_stat_error()) | |
1985 | reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); | |
1986 | else | |
1987 | reclaimable = wb_stat(wb, WB_RECLAIMABLE); | |
1988 | ||
1989 | if (reclaimable > thresh) | |
c2aa723a TH |
1990 | return true; |
1991 | } | |
1992 | ||
aa661bbe TH |
1993 | return false; |
1994 | } | |
1995 | ||
1da177e4 LT |
1996 | /* |
1997 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
1998 | */ | |
cccad5b9 | 1999 | int dirty_writeback_centisecs_handler(struct ctl_table *table, int write, |
32927393 | 2000 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 2001 | { |
94af5846 YS |
2002 | unsigned int old_interval = dirty_writeback_interval; |
2003 | int ret; | |
2004 | ||
2005 | ret = proc_dointvec(table, write, buffer, length, ppos); | |
515c24c1 YS |
2006 | |
2007 | /* | |
2008 | * Writing 0 to dirty_writeback_interval will disable periodic writeback | |
2009 | * and a different non-zero value will wakeup the writeback threads. | |
2010 | * wb_wakeup_delayed() would be more appropriate, but it's a pain to | |
2011 | * iterate over all bdis and wbs. | |
2012 | * The reason we do this is to make the change take effect immediately. | |
2013 | */ | |
2014 | if (!ret && write && dirty_writeback_interval && | |
2015 | dirty_writeback_interval != old_interval) | |
94af5846 YS |
2016 | wakeup_flusher_threads(WB_REASON_PERIODIC); |
2017 | ||
2018 | return ret; | |
1da177e4 LT |
2019 | } |
2020 | ||
bca237a5 | 2021 | void laptop_mode_timer_fn(struct timer_list *t) |
1da177e4 | 2022 | { |
bca237a5 KC |
2023 | struct backing_dev_info *backing_dev_info = |
2024 | from_timer(backing_dev_info, t, laptop_mode_wb_timer); | |
1da177e4 | 2025 | |
bca237a5 | 2026 | wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER); |
1da177e4 LT |
2027 | } |
2028 | ||
2029 | /* | |
2030 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
2031 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
2032 | * then push it back - the user is still using the disk. | |
2033 | */ | |
31373d09 | 2034 | void laptop_io_completion(struct backing_dev_info *info) |
1da177e4 | 2035 | { |
31373d09 | 2036 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
2037 | } |
2038 | ||
2039 | /* | |
2040 | * We're in laptop mode and we've just synced. The sync's writes will have | |
2041 | * caused another writeback to be scheduled by laptop_io_completion. | |
2042 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
2043 | */ | |
2044 | void laptop_sync_completion(void) | |
2045 | { | |
31373d09 MG |
2046 | struct backing_dev_info *bdi; |
2047 | ||
2048 | rcu_read_lock(); | |
2049 | ||
2050 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) | |
2051 | del_timer(&bdi->laptop_mode_wb_timer); | |
2052 | ||
2053 | rcu_read_unlock(); | |
1da177e4 LT |
2054 | } |
2055 | ||
2056 | /* | |
2057 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
2058 | * if a large number of processes all perform writes at the same time. | |
1da177e4 LT |
2059 | * |
2060 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
2061 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
9d823e8f | 2062 | * thresholds. |
1da177e4 LT |
2063 | */ |
2064 | ||
2d1d43f6 | 2065 | void writeback_set_ratelimit(void) |
1da177e4 | 2066 | { |
dcc25ae7 | 2067 | struct wb_domain *dom = &global_wb_domain; |
9d823e8f WF |
2068 | unsigned long background_thresh; |
2069 | unsigned long dirty_thresh; | |
dcc25ae7 | 2070 | |
9d823e8f | 2071 | global_dirty_limits(&background_thresh, &dirty_thresh); |
dcc25ae7 | 2072 | dom->dirty_limit = dirty_thresh; |
9d823e8f | 2073 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); |
1da177e4 LT |
2074 | if (ratelimit_pages < 16) |
2075 | ratelimit_pages = 16; | |
1da177e4 LT |
2076 | } |
2077 | ||
1d7ac6ae | 2078 | static int page_writeback_cpu_online(unsigned int cpu) |
1da177e4 | 2079 | { |
1d7ac6ae SAS |
2080 | writeback_set_ratelimit(); |
2081 | return 0; | |
1da177e4 LT |
2082 | } |
2083 | ||
1da177e4 | 2084 | /* |
dc6e29da LT |
2085 | * Called early on to tune the page writeback dirty limits. |
2086 | * | |
2087 | * We used to scale dirty pages according to how total memory | |
0a18e607 | 2088 | * related to pages that could be allocated for buffers. |
dc6e29da LT |
2089 | * |
2090 | * However, that was when we used "dirty_ratio" to scale with | |
2091 | * all memory, and we don't do that any more. "dirty_ratio" | |
0a18e607 | 2092 | * is now applied to total non-HIGHPAGE memory, and as such we can't |
dc6e29da LT |
2093 | * get into the old insane situation any more where we had |
2094 | * large amounts of dirty pages compared to a small amount of | |
2095 | * non-HIGHMEM memory. | |
2096 | * | |
2097 | * But we might still want to scale the dirty_ratio by how | |
2098 | * much memory the box has.. | |
1da177e4 LT |
2099 | */ |
2100 | void __init page_writeback_init(void) | |
2101 | { | |
a50fcb51 RV |
2102 | BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL)); |
2103 | ||
1d7ac6ae SAS |
2104 | cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online", |
2105 | page_writeback_cpu_online, NULL); | |
2106 | cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL, | |
2107 | page_writeback_cpu_online); | |
1da177e4 LT |
2108 | } |
2109 | ||
f446daae JK |
2110 | /** |
2111 | * tag_pages_for_writeback - tag pages to be written by write_cache_pages | |
2112 | * @mapping: address space structure to write | |
2113 | * @start: starting page index | |
2114 | * @end: ending page index (inclusive) | |
2115 | * | |
2116 | * This function scans the page range from @start to @end (inclusive) and tags | |
2117 | * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is | |
2118 | * that write_cache_pages (or whoever calls this function) will then use | |
2119 | * TOWRITE tag to identify pages eligible for writeback. This mechanism is | |
2120 | * used to avoid livelocking of writeback by a process steadily creating new | |
2121 | * dirty pages in the file (thus it is important for this function to be quick | |
2122 | * so that it can tag pages faster than a dirtying process can create them). | |
2123 | */ | |
f446daae JK |
2124 | void tag_pages_for_writeback(struct address_space *mapping, |
2125 | pgoff_t start, pgoff_t end) | |
2126 | { | |
ff9c745b MW |
2127 | XA_STATE(xas, &mapping->i_pages, start); |
2128 | unsigned int tagged = 0; | |
2129 | void *page; | |
268f42de | 2130 | |
ff9c745b MW |
2131 | xas_lock_irq(&xas); |
2132 | xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) { | |
2133 | xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE); | |
2134 | if (++tagged % XA_CHECK_SCHED) | |
268f42de | 2135 | continue; |
ff9c745b MW |
2136 | |
2137 | xas_pause(&xas); | |
2138 | xas_unlock_irq(&xas); | |
f446daae | 2139 | cond_resched(); |
ff9c745b | 2140 | xas_lock_irq(&xas); |
268f42de | 2141 | } |
ff9c745b | 2142 | xas_unlock_irq(&xas); |
f446daae JK |
2143 | } |
2144 | EXPORT_SYMBOL(tag_pages_for_writeback); | |
2145 | ||
811d736f | 2146 | /** |
0ea97180 | 2147 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. |
811d736f DH |
2148 | * @mapping: address space structure to write |
2149 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
0ea97180 MS |
2150 | * @writepage: function called for each page |
2151 | * @data: data passed to writepage function | |
811d736f | 2152 | * |
0ea97180 | 2153 | * If a page is already under I/O, write_cache_pages() skips it, even |
811d736f DH |
2154 | * if it's dirty. This is desirable behaviour for memory-cleaning writeback, |
2155 | * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() | |
2156 | * and msync() need to guarantee that all the data which was dirty at the time | |
2157 | * the call was made get new I/O started against them. If wbc->sync_mode is | |
2158 | * WB_SYNC_ALL then we were called for data integrity and we must wait for | |
2159 | * existing IO to complete. | |
f446daae JK |
2160 | * |
2161 | * To avoid livelocks (when other process dirties new pages), we first tag | |
2162 | * pages which should be written back with TOWRITE tag and only then start | |
2163 | * writing them. For data-integrity sync we have to be careful so that we do | |
2164 | * not miss some pages (e.g., because some other process has cleared TOWRITE | |
2165 | * tag we set). The rule we follow is that TOWRITE tag can be cleared only | |
2166 | * by the process clearing the DIRTY tag (and submitting the page for IO). | |
64081362 DC |
2167 | * |
2168 | * To avoid deadlocks between range_cyclic writeback and callers that hold | |
2169 | * pages in PageWriteback to aggregate IO until write_cache_pages() returns, | |
2170 | * we do not loop back to the start of the file. Doing so causes a page | |
2171 | * lock/page writeback access order inversion - we should only ever lock | |
2172 | * multiple pages in ascending page->index order, and looping back to the start | |
2173 | * of the file violates that rule and causes deadlocks. | |
a862f68a MR |
2174 | * |
2175 | * Return: %0 on success, negative error code otherwise | |
811d736f | 2176 | */ |
0ea97180 MS |
2177 | int write_cache_pages(struct address_space *mapping, |
2178 | struct writeback_control *wbc, writepage_t writepage, | |
2179 | void *data) | |
811d736f | 2180 | { |
811d736f DH |
2181 | int ret = 0; |
2182 | int done = 0; | |
3fa750dc | 2183 | int error; |
811d736f DH |
2184 | struct pagevec pvec; |
2185 | int nr_pages; | |
2186 | pgoff_t index; | |
2187 | pgoff_t end; /* Inclusive */ | |
bd19e012 | 2188 | pgoff_t done_index; |
811d736f | 2189 | int range_whole = 0; |
ff9c745b | 2190 | xa_mark_t tag; |
811d736f | 2191 | |
86679820 | 2192 | pagevec_init(&pvec); |
811d736f | 2193 | if (wbc->range_cyclic) { |
28659cc8 | 2194 | index = mapping->writeback_index; /* prev offset */ |
811d736f DH |
2195 | end = -1; |
2196 | } else { | |
09cbfeaf KS |
2197 | index = wbc->range_start >> PAGE_SHIFT; |
2198 | end = wbc->range_end >> PAGE_SHIFT; | |
811d736f DH |
2199 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) |
2200 | range_whole = 1; | |
811d736f | 2201 | } |
cc7b8f62 MFO |
2202 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) { |
2203 | tag_pages_for_writeback(mapping, index, end); | |
f446daae | 2204 | tag = PAGECACHE_TAG_TOWRITE; |
cc7b8f62 | 2205 | } else { |
f446daae | 2206 | tag = PAGECACHE_TAG_DIRTY; |
cc7b8f62 | 2207 | } |
bd19e012 | 2208 | done_index = index; |
5a3d5c98 NP |
2209 | while (!done && (index <= end)) { |
2210 | int i; | |
2211 | ||
2b9775ae | 2212 | nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end, |
67fd707f | 2213 | tag); |
5a3d5c98 NP |
2214 | if (nr_pages == 0) |
2215 | break; | |
811d736f | 2216 | |
811d736f DH |
2217 | for (i = 0; i < nr_pages; i++) { |
2218 | struct page *page = pvec.pages[i]; | |
2219 | ||
cf15b07c | 2220 | done_index = page->index; |
d5482cdf | 2221 | |
811d736f DH |
2222 | lock_page(page); |
2223 | ||
5a3d5c98 NP |
2224 | /* |
2225 | * Page truncated or invalidated. We can freely skip it | |
2226 | * then, even for data integrity operations: the page | |
2227 | * has disappeared concurrently, so there could be no | |
f0953a1b | 2228 | * real expectation of this data integrity operation |
5a3d5c98 NP |
2229 | * even if there is now a new, dirty page at the same |
2230 | * pagecache address. | |
2231 | */ | |
811d736f | 2232 | if (unlikely(page->mapping != mapping)) { |
5a3d5c98 | 2233 | continue_unlock: |
811d736f DH |
2234 | unlock_page(page); |
2235 | continue; | |
2236 | } | |
2237 | ||
515f4a03 NP |
2238 | if (!PageDirty(page)) { |
2239 | /* someone wrote it for us */ | |
2240 | goto continue_unlock; | |
2241 | } | |
2242 | ||
2243 | if (PageWriteback(page)) { | |
2244 | if (wbc->sync_mode != WB_SYNC_NONE) | |
2245 | wait_on_page_writeback(page); | |
2246 | else | |
2247 | goto continue_unlock; | |
2248 | } | |
811d736f | 2249 | |
515f4a03 NP |
2250 | BUG_ON(PageWriteback(page)); |
2251 | if (!clear_page_dirty_for_io(page)) | |
5a3d5c98 | 2252 | goto continue_unlock; |
811d736f | 2253 | |
de1414a6 | 2254 | trace_wbc_writepage(wbc, inode_to_bdi(mapping->host)); |
3fa750dc BF |
2255 | error = (*writepage)(page, wbc, data); |
2256 | if (unlikely(error)) { | |
2257 | /* | |
2258 | * Handle errors according to the type of | |
2259 | * writeback. There's no need to continue for | |
2260 | * background writeback. Just push done_index | |
2261 | * past this page so media errors won't choke | |
2262 | * writeout for the entire file. For integrity | |
2263 | * writeback, we must process the entire dirty | |
2264 | * set regardless of errors because the fs may | |
2265 | * still have state to clear for each page. In | |
2266 | * that case we continue processing and return | |
2267 | * the first error. | |
2268 | */ | |
2269 | if (error == AOP_WRITEPAGE_ACTIVATE) { | |
00266770 | 2270 | unlock_page(page); |
3fa750dc BF |
2271 | error = 0; |
2272 | } else if (wbc->sync_mode != WB_SYNC_ALL) { | |
2273 | ret = error; | |
cf15b07c | 2274 | done_index = page->index + 1; |
00266770 NP |
2275 | done = 1; |
2276 | break; | |
2277 | } | |
3fa750dc BF |
2278 | if (!ret) |
2279 | ret = error; | |
0b564927 | 2280 | } |
00266770 | 2281 | |
546a1924 DC |
2282 | /* |
2283 | * We stop writing back only if we are not doing | |
2284 | * integrity sync. In case of integrity sync we have to | |
2285 | * keep going until we have written all the pages | |
2286 | * we tagged for writeback prior to entering this loop. | |
2287 | */ | |
2288 | if (--wbc->nr_to_write <= 0 && | |
2289 | wbc->sync_mode == WB_SYNC_NONE) { | |
2290 | done = 1; | |
2291 | break; | |
05fe478d | 2292 | } |
811d736f DH |
2293 | } |
2294 | pagevec_release(&pvec); | |
2295 | cond_resched(); | |
2296 | } | |
64081362 DC |
2297 | |
2298 | /* | |
2299 | * If we hit the last page and there is more work to be done: wrap | |
2300 | * back the index back to the start of the file for the next | |
2301 | * time we are called. | |
2302 | */ | |
2303 | if (wbc->range_cyclic && !done) | |
2304 | done_index = 0; | |
0b564927 DC |
2305 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) |
2306 | mapping->writeback_index = done_index; | |
06d6cf69 | 2307 | |
811d736f DH |
2308 | return ret; |
2309 | } | |
0ea97180 MS |
2310 | EXPORT_SYMBOL(write_cache_pages); |
2311 | ||
2312 | /* | |
2313 | * Function used by generic_writepages to call the real writepage | |
2314 | * function and set the mapping flags on error | |
2315 | */ | |
2316 | static int __writepage(struct page *page, struct writeback_control *wbc, | |
2317 | void *data) | |
2318 | { | |
2319 | struct address_space *mapping = data; | |
2320 | int ret = mapping->a_ops->writepage(page, wbc); | |
2321 | mapping_set_error(mapping, ret); | |
2322 | return ret; | |
2323 | } | |
2324 | ||
2325 | /** | |
2326 | * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. | |
2327 | * @mapping: address space structure to write | |
2328 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
2329 | * | |
2330 | * This is a library function, which implements the writepages() | |
2331 | * address_space_operation. | |
a862f68a MR |
2332 | * |
2333 | * Return: %0 on success, negative error code otherwise | |
0ea97180 MS |
2334 | */ |
2335 | int generic_writepages(struct address_space *mapping, | |
2336 | struct writeback_control *wbc) | |
2337 | { | |
9b6096a6 SL |
2338 | struct blk_plug plug; |
2339 | int ret; | |
2340 | ||
0ea97180 MS |
2341 | /* deal with chardevs and other special file */ |
2342 | if (!mapping->a_ops->writepage) | |
2343 | return 0; | |
2344 | ||
9b6096a6 SL |
2345 | blk_start_plug(&plug); |
2346 | ret = write_cache_pages(mapping, wbc, __writepage, mapping); | |
2347 | blk_finish_plug(&plug); | |
2348 | return ret; | |
0ea97180 | 2349 | } |
811d736f DH |
2350 | |
2351 | EXPORT_SYMBOL(generic_writepages); | |
2352 | ||
1da177e4 LT |
2353 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
2354 | { | |
22905f77 | 2355 | int ret; |
fee468fd | 2356 | struct bdi_writeback *wb; |
22905f77 | 2357 | |
1da177e4 LT |
2358 | if (wbc->nr_to_write <= 0) |
2359 | return 0; | |
fee468fd JK |
2360 | wb = inode_to_wb_wbc(mapping->host, wbc); |
2361 | wb_bandwidth_estimate_start(wb); | |
80a2ea9f TT |
2362 | while (1) { |
2363 | if (mapping->a_ops->writepages) | |
2364 | ret = mapping->a_ops->writepages(mapping, wbc); | |
2365 | else | |
2366 | ret = generic_writepages(mapping, wbc); | |
2367 | if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL)) | |
2368 | break; | |
8d58802f MG |
2369 | |
2370 | /* | |
2371 | * Lacking an allocation context or the locality or writeback | |
2372 | * state of any of the inode's pages, throttle based on | |
2373 | * writeback activity on the local node. It's as good a | |
2374 | * guess as any. | |
2375 | */ | |
2376 | reclaim_throttle(NODE_DATA(numa_node_id()), | |
c3f4a9a2 | 2377 | VMSCAN_THROTTLE_WRITEBACK); |
80a2ea9f | 2378 | } |
45a2966f JK |
2379 | /* |
2380 | * Usually few pages are written by now from those we've just submitted | |
2381 | * but if there's constant writeback being submitted, this makes sure | |
2382 | * writeback bandwidth is updated once in a while. | |
2383 | */ | |
20792ebf JK |
2384 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) + |
2385 | BANDWIDTH_INTERVAL)) | |
45a2966f | 2386 | wb_update_bandwidth(wb); |
22905f77 | 2387 | return ret; |
1da177e4 LT |
2388 | } |
2389 | ||
2390 | /** | |
121703c1 MWO |
2391 | * folio_write_one - write out a single folio and wait on I/O. |
2392 | * @folio: The folio to write. | |
1da177e4 | 2393 | * |
121703c1 | 2394 | * The folio must be locked by the caller and will be unlocked upon return. |
1da177e4 | 2395 | * |
37e51a76 JL |
2396 | * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this |
2397 | * function returns. | |
a862f68a MR |
2398 | * |
2399 | * Return: %0 on success, negative error code otherwise | |
1da177e4 | 2400 | */ |
121703c1 | 2401 | int folio_write_one(struct folio *folio) |
1da177e4 | 2402 | { |
121703c1 | 2403 | struct address_space *mapping = folio->mapping; |
1da177e4 LT |
2404 | int ret = 0; |
2405 | struct writeback_control wbc = { | |
2406 | .sync_mode = WB_SYNC_ALL, | |
121703c1 | 2407 | .nr_to_write = folio_nr_pages(folio), |
1da177e4 LT |
2408 | }; |
2409 | ||
121703c1 | 2410 | BUG_ON(!folio_test_locked(folio)); |
1da177e4 | 2411 | |
121703c1 | 2412 | folio_wait_writeback(folio); |
1da177e4 | 2413 | |
121703c1 MWO |
2414 | if (folio_clear_dirty_for_io(folio)) { |
2415 | folio_get(folio); | |
2416 | ret = mapping->a_ops->writepage(&folio->page, &wbc); | |
37e51a76 | 2417 | if (ret == 0) |
121703c1 MWO |
2418 | folio_wait_writeback(folio); |
2419 | folio_put(folio); | |
1da177e4 | 2420 | } else { |
121703c1 | 2421 | folio_unlock(folio); |
1da177e4 | 2422 | } |
37e51a76 JL |
2423 | |
2424 | if (!ret) | |
2425 | ret = filemap_check_errors(mapping); | |
1da177e4 LT |
2426 | return ret; |
2427 | } | |
121703c1 | 2428 | EXPORT_SYMBOL(folio_write_one); |
1da177e4 | 2429 | |
76719325 KC |
2430 | /* |
2431 | * For address_spaces which do not use buffers nor write back. | |
2432 | */ | |
2433 | int __set_page_dirty_no_writeback(struct page *page) | |
2434 | { | |
2435 | if (!PageDirty(page)) | |
c3f0da63 | 2436 | return !TestSetPageDirty(page); |
76719325 KC |
2437 | return 0; |
2438 | } | |
b82a96c9 | 2439 | EXPORT_SYMBOL(__set_page_dirty_no_writeback); |
76719325 | 2440 | |
e3a7cca1 ES |
2441 | /* |
2442 | * Helper function for set_page_dirty family. | |
c4843a75 | 2443 | * |
81f8c3a4 | 2444 | * Caller must hold lock_page_memcg(). |
c4843a75 | 2445 | * |
e3a7cca1 ES |
2446 | * NOTE: This relies on being atomic wrt interrupts. |
2447 | */ | |
203a3151 | 2448 | static void folio_account_dirtied(struct folio *folio, |
6e1cae88 | 2449 | struct address_space *mapping) |
e3a7cca1 | 2450 | { |
52ebea74 TH |
2451 | struct inode *inode = mapping->host; |
2452 | ||
b9b0ff61 | 2453 | trace_writeback_dirty_folio(folio, mapping); |
9fb0a7da | 2454 | |
f56753ac | 2455 | if (mapping_can_writeback(mapping)) { |
52ebea74 | 2456 | struct bdi_writeback *wb; |
203a3151 | 2457 | long nr = folio_nr_pages(folio); |
de1414a6 | 2458 | |
203a3151 | 2459 | inode_attach_wb(inode, &folio->page); |
52ebea74 | 2460 | wb = inode_to_wb(inode); |
de1414a6 | 2461 | |
203a3151 MWO |
2462 | __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr); |
2463 | __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr); | |
2464 | __node_stat_mod_folio(folio, NR_DIRTIED, nr); | |
2465 | wb_stat_mod(wb, WB_RECLAIMABLE, nr); | |
2466 | wb_stat_mod(wb, WB_DIRTIED, nr); | |
2467 | task_io_account_write(nr * PAGE_SIZE); | |
2468 | current->nr_dirtied += nr; | |
2469 | __this_cpu_add(bdp_ratelimits, nr); | |
97b27821 | 2470 | |
203a3151 | 2471 | mem_cgroup_track_foreign_dirty(folio, wb); |
e3a7cca1 ES |
2472 | } |
2473 | } | |
2474 | ||
b9ea2515 KK |
2475 | /* |
2476 | * Helper function for deaccounting dirty page without writeback. | |
2477 | * | |
81f8c3a4 | 2478 | * Caller must hold lock_page_memcg(). |
b9ea2515 | 2479 | */ |
fc9b6a53 | 2480 | void folio_account_cleaned(struct folio *folio, struct address_space *mapping, |
62cccb8c | 2481 | struct bdi_writeback *wb) |
b9ea2515 | 2482 | { |
f56753ac | 2483 | if (mapping_can_writeback(mapping)) { |
fc9b6a53 MWO |
2484 | long nr = folio_nr_pages(folio); |
2485 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr); | |
2486 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); | |
2487 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr); | |
2488 | task_io_account_cancelled_write(nr * PAGE_SIZE); | |
b9ea2515 KK |
2489 | } |
2490 | } | |
b9ea2515 | 2491 | |
6e1cae88 | 2492 | /* |
203a3151 MWO |
2493 | * Mark the folio dirty, and set it dirty in the page cache, and mark |
2494 | * the inode dirty. | |
6e1cae88 | 2495 | * |
203a3151 | 2496 | * If warn is true, then emit a warning if the folio is not uptodate and has |
6e1cae88 MWO |
2497 | * not been truncated. |
2498 | * | |
a229a4f0 MWO |
2499 | * The caller must hold lock_page_memcg(). Most callers have the folio |
2500 | * locked. A few have the folio blocked from truncation through other | |
2501 | * means (eg zap_page_range() has it mapped and is holding the page table | |
2502 | * lock). This can also be called from mark_buffer_dirty(), which I | |
2503 | * cannot prove is always protected against truncate. | |
6e1cae88 | 2504 | */ |
203a3151 | 2505 | void __folio_mark_dirty(struct folio *folio, struct address_space *mapping, |
6e1cae88 MWO |
2506 | int warn) |
2507 | { | |
2508 | unsigned long flags; | |
2509 | ||
2510 | xa_lock_irqsave(&mapping->i_pages, flags); | |
203a3151 MWO |
2511 | if (folio->mapping) { /* Race with truncate? */ |
2512 | WARN_ON_ONCE(warn && !folio_test_uptodate(folio)); | |
2513 | folio_account_dirtied(folio, mapping); | |
2514 | __xa_set_mark(&mapping->i_pages, folio_index(folio), | |
6e1cae88 MWO |
2515 | PAGECACHE_TAG_DIRTY); |
2516 | } | |
2517 | xa_unlock_irqrestore(&mapping->i_pages, flags); | |
2518 | } | |
2519 | ||
85d4d2eb MWO |
2520 | /** |
2521 | * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads. | |
2522 | * @mapping: Address space this folio belongs to. | |
2523 | * @folio: Folio to be marked as dirty. | |
1da177e4 | 2524 | * |
85d4d2eb MWO |
2525 | * Filesystems which do not use buffer heads should call this function |
2526 | * from their set_page_dirty address space operation. It ignores the | |
2527 | * contents of folio_get_private(), so if the filesystem marks individual | |
2528 | * blocks as dirty, the filesystem should handle that itself. | |
1da177e4 | 2529 | * |
85d4d2eb MWO |
2530 | * This is also sometimes used by filesystems which use buffer_heads when |
2531 | * a single buffer is being dirtied: we want to set the folio dirty in | |
2532 | * that case, but not all the buffers. This is a "bottom-up" dirtying, | |
2533 | * whereas __set_page_dirty_buffers() is a "top-down" dirtying. | |
2534 | * | |
2535 | * The caller must ensure this doesn't race with truncation. Most will | |
2536 | * simply hold the folio lock, but e.g. zap_pte_range() calls with the | |
2537 | * folio mapped and the pte lock held, which also locks out truncation. | |
1da177e4 | 2538 | */ |
85d4d2eb | 2539 | bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio) |
1da177e4 | 2540 | { |
85d4d2eb MWO |
2541 | folio_memcg_lock(folio); |
2542 | if (folio_test_set_dirty(folio)) { | |
2543 | folio_memcg_unlock(folio); | |
2544 | return false; | |
2545 | } | |
1da177e4 | 2546 | |
85d4d2eb MWO |
2547 | __folio_mark_dirty(folio, mapping, !folio_test_private(folio)); |
2548 | folio_memcg_unlock(folio); | |
c4843a75 | 2549 | |
85d4d2eb MWO |
2550 | if (mapping->host) { |
2551 | /* !PageAnon && !swapper_space */ | |
2552 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 2553 | } |
85d4d2eb | 2554 | return true; |
1da177e4 | 2555 | } |
85d4d2eb | 2556 | EXPORT_SYMBOL(filemap_dirty_folio); |
1da177e4 | 2557 | |
25ff8b15 MWO |
2558 | /** |
2559 | * folio_account_redirty - Manually account for redirtying a page. | |
2560 | * @folio: The folio which is being redirtied. | |
2561 | * | |
2562 | * Most filesystems should call folio_redirty_for_writepage() instead | |
2563 | * of this fuction. If your filesystem is doing writeback outside the | |
2564 | * context of a writeback_control(), it can call this when redirtying | |
2565 | * a folio, to de-account the dirty counters (NR_DIRTIED, WB_DIRTIED, | |
2566 | * tsk->nr_dirtied), so that they match the written counters (NR_WRITTEN, | |
2567 | * WB_WRITTEN) in long term. The mismatches will lead to systematic errors | |
2568 | * in balanced_dirty_ratelimit and the dirty pages position control. | |
2f800fbd | 2569 | */ |
25ff8b15 | 2570 | void folio_account_redirty(struct folio *folio) |
2f800fbd | 2571 | { |
25ff8b15 | 2572 | struct address_space *mapping = folio->mapping; |
91018134 | 2573 | |
f56753ac | 2574 | if (mapping && mapping_can_writeback(mapping)) { |
682aa8e1 TH |
2575 | struct inode *inode = mapping->host; |
2576 | struct bdi_writeback *wb; | |
2e898e4c | 2577 | struct wb_lock_cookie cookie = {}; |
25ff8b15 | 2578 | long nr = folio_nr_pages(folio); |
91018134 | 2579 | |
2e898e4c | 2580 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
25ff8b15 MWO |
2581 | current->nr_dirtied -= nr; |
2582 | node_stat_mod_folio(folio, NR_DIRTIED, -nr); | |
2583 | wb_stat_mod(wb, WB_DIRTIED, -nr); | |
2e898e4c | 2584 | unlocked_inode_to_wb_end(inode, &cookie); |
2f800fbd WF |
2585 | } |
2586 | } | |
25ff8b15 | 2587 | EXPORT_SYMBOL(folio_account_redirty); |
2f800fbd | 2588 | |
cd78ab11 MWO |
2589 | /** |
2590 | * folio_redirty_for_writepage - Decline to write a dirty folio. | |
2591 | * @wbc: The writeback control. | |
2592 | * @folio: The folio. | |
2593 | * | |
2594 | * When a writepage implementation decides that it doesn't want to write | |
2595 | * @folio for some reason, it should call this function, unlock @folio and | |
2596 | * return 0. | |
2597 | * | |
2598 | * Return: True if we redirtied the folio. False if someone else dirtied | |
2599 | * it first. | |
1da177e4 | 2600 | */ |
cd78ab11 MWO |
2601 | bool folio_redirty_for_writepage(struct writeback_control *wbc, |
2602 | struct folio *folio) | |
1da177e4 | 2603 | { |
cd78ab11 MWO |
2604 | bool ret; |
2605 | long nr = folio_nr_pages(folio); | |
2606 | ||
2607 | wbc->pages_skipped += nr; | |
2608 | ret = filemap_dirty_folio(folio->mapping, folio); | |
2609 | folio_account_redirty(folio); | |
8d38633c | 2610 | |
8d38633c | 2611 | return ret; |
1da177e4 | 2612 | } |
cd78ab11 | 2613 | EXPORT_SYMBOL(folio_redirty_for_writepage); |
1da177e4 | 2614 | |
b5e84594 MWO |
2615 | /** |
2616 | * folio_mark_dirty - Mark a folio as being modified. | |
2617 | * @folio: The folio. | |
6746aff7 | 2618 | * |
b5e84594 MWO |
2619 | * For folios with a mapping this should be done under the page lock |
2620 | * for the benefit of asynchronous memory errors who prefer a consistent | |
2621 | * dirty state. This rule can be broken in some special cases, | |
2622 | * but should be better not to. | |
2623 | * | |
2624 | * Return: True if the folio was newly dirtied, false if it was already dirty. | |
1da177e4 | 2625 | */ |
b5e84594 | 2626 | bool folio_mark_dirty(struct folio *folio) |
1da177e4 | 2627 | { |
b5e84594 | 2628 | struct address_space *mapping = folio_mapping(folio); |
1da177e4 LT |
2629 | |
2630 | if (likely(mapping)) { | |
278df9f4 MK |
2631 | /* |
2632 | * readahead/lru_deactivate_page could remain | |
2633 | * PG_readahead/PG_reclaim due to race with end_page_writeback | |
2634 | * About readahead, if the page is written, the flags would be | |
2635 | * reset. So no problem. | |
2636 | * About lru_deactivate_page, if the page is redirty, the flag | |
2637 | * will be reset. So no problem. but if the page is used by readahead | |
2638 | * it will confuse readahead and make it restart the size rampup | |
2639 | * process. But it's a trivial problem. | |
2640 | */ | |
b5e84594 MWO |
2641 | if (folio_test_reclaim(folio)) |
2642 | folio_clear_reclaim(folio); | |
2643 | return mapping->a_ops->set_page_dirty(&folio->page); | |
1da177e4 | 2644 | } |
b5e84594 MWO |
2645 | if (!folio_test_dirty(folio)) { |
2646 | if (!folio_test_set_dirty(folio)) | |
2647 | return true; | |
4741c9fd | 2648 | } |
b5e84594 | 2649 | return false; |
1da177e4 | 2650 | } |
b5e84594 | 2651 | EXPORT_SYMBOL(folio_mark_dirty); |
1da177e4 LT |
2652 | |
2653 | /* | |
2654 | * set_page_dirty() is racy if the caller has no reference against | |
2655 | * page->mapping->host, and if the page is unlocked. This is because another | |
2656 | * CPU could truncate the page off the mapping and then free the mapping. | |
2657 | * | |
2658 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
2659 | * holds a reference on the inode by having an open file. | |
2660 | * | |
2661 | * In other cases, the page should be locked before running set_page_dirty(). | |
2662 | */ | |
2663 | int set_page_dirty_lock(struct page *page) | |
2664 | { | |
2665 | int ret; | |
2666 | ||
7eaceacc | 2667 | lock_page(page); |
1da177e4 LT |
2668 | ret = set_page_dirty(page); |
2669 | unlock_page(page); | |
2670 | return ret; | |
2671 | } | |
2672 | EXPORT_SYMBOL(set_page_dirty_lock); | |
2673 | ||
11f81bec TH |
2674 | /* |
2675 | * This cancels just the dirty bit on the kernel page itself, it does NOT | |
2676 | * actually remove dirty bits on any mmap's that may be around. It also | |
2677 | * leaves the page tagged dirty, so any sync activity will still find it on | |
2678 | * the dirty lists, and in particular, clear_page_dirty_for_io() will still | |
2679 | * look at the dirty bits in the VM. | |
2680 | * | |
2681 | * Doing this should *normally* only ever be done when a page is truncated, | |
2682 | * and is not actually mapped anywhere at all. However, fs/buffer.c does | |
2683 | * this when it notices that somebody has cleaned out all the buffers on a | |
2684 | * page without actually doing it through the VM. Can you say "ext3 is | |
2685 | * horribly ugly"? Thought you could. | |
2686 | */ | |
fdaf532a | 2687 | void __folio_cancel_dirty(struct folio *folio) |
11f81bec | 2688 | { |
fdaf532a | 2689 | struct address_space *mapping = folio_mapping(folio); |
c4843a75 | 2690 | |
f56753ac | 2691 | if (mapping_can_writeback(mapping)) { |
682aa8e1 TH |
2692 | struct inode *inode = mapping->host; |
2693 | struct bdi_writeback *wb; | |
2e898e4c | 2694 | struct wb_lock_cookie cookie = {}; |
c4843a75 | 2695 | |
fdaf532a | 2696 | folio_memcg_lock(folio); |
2e898e4c | 2697 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
c4843a75 | 2698 | |
fdaf532a MWO |
2699 | if (folio_test_clear_dirty(folio)) |
2700 | folio_account_cleaned(folio, mapping, wb); | |
c4843a75 | 2701 | |
2e898e4c | 2702 | unlocked_inode_to_wb_end(inode, &cookie); |
fdaf532a | 2703 | folio_memcg_unlock(folio); |
c4843a75 | 2704 | } else { |
fdaf532a | 2705 | folio_clear_dirty(folio); |
c4843a75 | 2706 | } |
11f81bec | 2707 | } |
fdaf532a | 2708 | EXPORT_SYMBOL(__folio_cancel_dirty); |
11f81bec | 2709 | |
1da177e4 | 2710 | /* |
9350f20a MWO |
2711 | * Clear a folio's dirty flag, while caring for dirty memory accounting. |
2712 | * Returns true if the folio was previously dirty. | |
1da177e4 | 2713 | * |
9350f20a MWO |
2714 | * This is for preparing to put the folio under writeout. We leave |
2715 | * the folio tagged as dirty in the xarray so that a concurrent | |
2716 | * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk. | |
2717 | * The ->writepage implementation will run either folio_start_writeback() | |
2718 | * or folio_mark_dirty(), at which stage we bring the folio's dirty flag | |
2719 | * and xarray dirty tag back into sync. | |
1da177e4 | 2720 | * |
9350f20a MWO |
2721 | * This incoherency between the folio's dirty flag and xarray tag is |
2722 | * unfortunate, but it only exists while the folio is locked. | |
1da177e4 | 2723 | */ |
9350f20a | 2724 | bool folio_clear_dirty_for_io(struct folio *folio) |
1da177e4 | 2725 | { |
9350f20a MWO |
2726 | struct address_space *mapping = folio_mapping(folio); |
2727 | bool ret = false; | |
1da177e4 | 2728 | |
9350f20a | 2729 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
79352894 | 2730 | |
f56753ac | 2731 | if (mapping && mapping_can_writeback(mapping)) { |
682aa8e1 TH |
2732 | struct inode *inode = mapping->host; |
2733 | struct bdi_writeback *wb; | |
2e898e4c | 2734 | struct wb_lock_cookie cookie = {}; |
682aa8e1 | 2735 | |
7658cc28 LT |
2736 | /* |
2737 | * Yes, Virginia, this is indeed insane. | |
2738 | * | |
2739 | * We use this sequence to make sure that | |
2740 | * (a) we account for dirty stats properly | |
2741 | * (b) we tell the low-level filesystem to | |
9350f20a | 2742 | * mark the whole folio dirty if it was |
7658cc28 | 2743 | * dirty in a pagetable. Only to then |
9350f20a | 2744 | * (c) clean the folio again and return 1 to |
7658cc28 LT |
2745 | * cause the writeback. |
2746 | * | |
2747 | * This way we avoid all nasty races with the | |
2748 | * dirty bit in multiple places and clearing | |
2749 | * them concurrently from different threads. | |
2750 | * | |
9350f20a | 2751 | * Note! Normally the "folio_mark_dirty(folio)" |
7658cc28 LT |
2752 | * has no effect on the actual dirty bit - since |
2753 | * that will already usually be set. But we | |
2754 | * need the side effects, and it can help us | |
2755 | * avoid races. | |
2756 | * | |
9350f20a | 2757 | * We basically use the folio "master dirty bit" |
7658cc28 LT |
2758 | * as a serialization point for all the different |
2759 | * threads doing their things. | |
7658cc28 | 2760 | */ |
9350f20a MWO |
2761 | if (folio_mkclean(folio)) |
2762 | folio_mark_dirty(folio); | |
79352894 NP |
2763 | /* |
2764 | * We carefully synchronise fault handlers against | |
9350f20a | 2765 | * installing a dirty pte and marking the folio dirty |
2d6d7f98 | 2766 | * at this point. We do this by having them hold the |
9350f20a | 2767 | * page lock while dirtying the folio, and folios are |
2d6d7f98 JW |
2768 | * always locked coming in here, so we get the desired |
2769 | * exclusion. | |
79352894 | 2770 | */ |
2e898e4c | 2771 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
9350f20a MWO |
2772 | if (folio_test_clear_dirty(folio)) { |
2773 | long nr = folio_nr_pages(folio); | |
2774 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr); | |
2775 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); | |
2776 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr); | |
2777 | ret = true; | |
1da177e4 | 2778 | } |
2e898e4c | 2779 | unlocked_inode_to_wb_end(inode, &cookie); |
c4843a75 | 2780 | return ret; |
1da177e4 | 2781 | } |
9350f20a | 2782 | return folio_test_clear_dirty(folio); |
1da177e4 | 2783 | } |
9350f20a | 2784 | EXPORT_SYMBOL(folio_clear_dirty_for_io); |
1da177e4 | 2785 | |
633a2abb JK |
2786 | static void wb_inode_writeback_start(struct bdi_writeback *wb) |
2787 | { | |
2788 | atomic_inc(&wb->writeback_inodes); | |
2789 | } | |
2790 | ||
2791 | static void wb_inode_writeback_end(struct bdi_writeback *wb) | |
2792 | { | |
2793 | atomic_dec(&wb->writeback_inodes); | |
45a2966f JK |
2794 | /* |
2795 | * Make sure estimate of writeback throughput gets updated after | |
2796 | * writeback completed. We delay the update by BANDWIDTH_INTERVAL | |
2797 | * (which is the interval other bandwidth updates use for batching) so | |
2798 | * that if multiple inodes end writeback at a similar time, they get | |
2799 | * batched into one bandwidth update. | |
2800 | */ | |
2801 | queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL); | |
633a2abb JK |
2802 | } |
2803 | ||
269ccca3 | 2804 | bool __folio_end_writeback(struct folio *folio) |
1da177e4 | 2805 | { |
269ccca3 MWO |
2806 | long nr = folio_nr_pages(folio); |
2807 | struct address_space *mapping = folio_mapping(folio); | |
2808 | bool ret; | |
1da177e4 | 2809 | |
269ccca3 | 2810 | folio_memcg_lock(folio); |
371a096e | 2811 | if (mapping && mapping_use_writeback_tags(mapping)) { |
91018134 TH |
2812 | struct inode *inode = mapping->host; |
2813 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 LT |
2814 | unsigned long flags; |
2815 | ||
b93b0163 | 2816 | xa_lock_irqsave(&mapping->i_pages, flags); |
269ccca3 | 2817 | ret = folio_test_clear_writeback(folio); |
69cb51d1 | 2818 | if (ret) { |
269ccca3 | 2819 | __xa_clear_mark(&mapping->i_pages, folio_index(folio), |
1da177e4 | 2820 | PAGECACHE_TAG_WRITEBACK); |
823423ef | 2821 | if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) { |
91018134 TH |
2822 | struct bdi_writeback *wb = inode_to_wb(inode); |
2823 | ||
269ccca3 MWO |
2824 | wb_stat_mod(wb, WB_WRITEBACK, -nr); |
2825 | __wb_writeout_add(wb, nr); | |
633a2abb JK |
2826 | if (!mapping_tagged(mapping, |
2827 | PAGECACHE_TAG_WRITEBACK)) | |
2828 | wb_inode_writeback_end(wb); | |
04fbfdc1 | 2829 | } |
69cb51d1 | 2830 | } |
6c60d2b5 DC |
2831 | |
2832 | if (mapping->host && !mapping_tagged(mapping, | |
2833 | PAGECACHE_TAG_WRITEBACK)) | |
2834 | sb_clear_inode_writeback(mapping->host); | |
2835 | ||
b93b0163 | 2836 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
1da177e4 | 2837 | } else { |
269ccca3 | 2838 | ret = folio_test_clear_writeback(folio); |
1da177e4 | 2839 | } |
99b12e3d | 2840 | if (ret) { |
269ccca3 MWO |
2841 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr); |
2842 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); | |
2843 | node_stat_mod_folio(folio, NR_WRITTEN, nr); | |
99b12e3d | 2844 | } |
269ccca3 | 2845 | folio_memcg_unlock(folio); |
1da177e4 LT |
2846 | return ret; |
2847 | } | |
2848 | ||
f143f1ea | 2849 | bool __folio_start_writeback(struct folio *folio, bool keep_write) |
1da177e4 | 2850 | { |
f143f1ea MWO |
2851 | long nr = folio_nr_pages(folio); |
2852 | struct address_space *mapping = folio_mapping(folio); | |
2853 | bool ret; | |
2854 | int access_ret; | |
1da177e4 | 2855 | |
f143f1ea | 2856 | folio_memcg_lock(folio); |
371a096e | 2857 | if (mapping && mapping_use_writeback_tags(mapping)) { |
f143f1ea | 2858 | XA_STATE(xas, &mapping->i_pages, folio_index(folio)); |
91018134 TH |
2859 | struct inode *inode = mapping->host; |
2860 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 LT |
2861 | unsigned long flags; |
2862 | ||
ff9c745b MW |
2863 | xas_lock_irqsave(&xas, flags); |
2864 | xas_load(&xas); | |
f143f1ea | 2865 | ret = folio_test_set_writeback(folio); |
69cb51d1 | 2866 | if (!ret) { |
6c60d2b5 DC |
2867 | bool on_wblist; |
2868 | ||
2869 | on_wblist = mapping_tagged(mapping, | |
2870 | PAGECACHE_TAG_WRITEBACK); | |
2871 | ||
ff9c745b | 2872 | xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK); |
633a2abb JK |
2873 | if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) { |
2874 | struct bdi_writeback *wb = inode_to_wb(inode); | |
2875 | ||
f143f1ea | 2876 | wb_stat_mod(wb, WB_WRITEBACK, nr); |
633a2abb JK |
2877 | if (!on_wblist) |
2878 | wb_inode_writeback_start(wb); | |
2879 | } | |
6c60d2b5 DC |
2880 | |
2881 | /* | |
f143f1ea MWO |
2882 | * We can come through here when swapping |
2883 | * anonymous folios, so we don't necessarily | |
2884 | * have an inode to track for sync. | |
6c60d2b5 DC |
2885 | */ |
2886 | if (mapping->host && !on_wblist) | |
2887 | sb_mark_inode_writeback(mapping->host); | |
69cb51d1 | 2888 | } |
f143f1ea | 2889 | if (!folio_test_dirty(folio)) |
ff9c745b | 2890 | xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY); |
1c8349a1 | 2891 | if (!keep_write) |
ff9c745b MW |
2892 | xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE); |
2893 | xas_unlock_irqrestore(&xas, flags); | |
1da177e4 | 2894 | } else { |
f143f1ea | 2895 | ret = folio_test_set_writeback(folio); |
1da177e4 | 2896 | } |
3a3c02ec | 2897 | if (!ret) { |
f143f1ea MWO |
2898 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr); |
2899 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr); | |
3a3c02ec | 2900 | } |
f143f1ea MWO |
2901 | folio_memcg_unlock(folio); |
2902 | access_ret = arch_make_folio_accessible(folio); | |
f28d4363 CI |
2903 | /* |
2904 | * If writeback has been triggered on a page that cannot be made | |
2905 | * accessible, it is too late to recover here. | |
2906 | */ | |
f143f1ea | 2907 | VM_BUG_ON_FOLIO(access_ret != 0, folio); |
f28d4363 | 2908 | |
1da177e4 | 2909 | return ret; |
1da177e4 | 2910 | } |
f143f1ea | 2911 | EXPORT_SYMBOL(__folio_start_writeback); |
1da177e4 | 2912 | |
490e016f MWO |
2913 | /** |
2914 | * folio_wait_writeback - Wait for a folio to finish writeback. | |
2915 | * @folio: The folio to wait for. | |
2916 | * | |
2917 | * If the folio is currently being written back to storage, wait for the | |
2918 | * I/O to complete. | |
2919 | * | |
2920 | * Context: Sleeps. Must be called in process context and with | |
2921 | * no spinlocks held. Caller should hold a reference on the folio. | |
2922 | * If the folio is not locked, writeback may start again after writeback | |
2923 | * has finished. | |
19343b5b | 2924 | */ |
490e016f | 2925 | void folio_wait_writeback(struct folio *folio) |
19343b5b | 2926 | { |
490e016f | 2927 | while (folio_test_writeback(folio)) { |
b9b0ff61 | 2928 | trace_folio_wait_writeback(folio, folio_mapping(folio)); |
101c0bf6 | 2929 | folio_wait_bit(folio, PG_writeback); |
19343b5b YS |
2930 | } |
2931 | } | |
490e016f | 2932 | EXPORT_SYMBOL_GPL(folio_wait_writeback); |
19343b5b | 2933 | |
490e016f MWO |
2934 | /** |
2935 | * folio_wait_writeback_killable - Wait for a folio to finish writeback. | |
2936 | * @folio: The folio to wait for. | |
2937 | * | |
2938 | * If the folio is currently being written back to storage, wait for the | |
2939 | * I/O to complete or a fatal signal to arrive. | |
2940 | * | |
2941 | * Context: Sleeps. Must be called in process context and with | |
2942 | * no spinlocks held. Caller should hold a reference on the folio. | |
2943 | * If the folio is not locked, writeback may start again after writeback | |
2944 | * has finished. | |
2945 | * Return: 0 on success, -EINTR if we get a fatal signal while waiting. | |
e5dbd332 | 2946 | */ |
490e016f | 2947 | int folio_wait_writeback_killable(struct folio *folio) |
e5dbd332 | 2948 | { |
490e016f | 2949 | while (folio_test_writeback(folio)) { |
b9b0ff61 | 2950 | trace_folio_wait_writeback(folio, folio_mapping(folio)); |
101c0bf6 | 2951 | if (folio_wait_bit_killable(folio, PG_writeback)) |
e5dbd332 MWO |
2952 | return -EINTR; |
2953 | } | |
2954 | ||
2955 | return 0; | |
2956 | } | |
490e016f | 2957 | EXPORT_SYMBOL_GPL(folio_wait_writeback_killable); |
e5dbd332 | 2958 | |
1d1d1a76 | 2959 | /** |
a49d0c50 MWO |
2960 | * folio_wait_stable() - wait for writeback to finish, if necessary. |
2961 | * @folio: The folio to wait on. | |
1d1d1a76 | 2962 | * |
a49d0c50 MWO |
2963 | * This function determines if the given folio is related to a backing |
2964 | * device that requires folio contents to be held stable during writeback. | |
2965 | * If so, then it will wait for any pending writeback to complete. | |
2966 | * | |
2967 | * Context: Sleeps. Must be called in process context and with | |
2968 | * no spinlocks held. Caller should hold a reference on the folio. | |
2969 | * If the folio is not locked, writeback may start again after writeback | |
2970 | * has finished. | |
1d1d1a76 | 2971 | */ |
a49d0c50 | 2972 | void folio_wait_stable(struct folio *folio) |
1d1d1a76 | 2973 | { |
452c472e | 2974 | if (folio_inode(folio)->i_sb->s_iflags & SB_I_STABLE_WRITES) |
a49d0c50 | 2975 | folio_wait_writeback(folio); |
1d1d1a76 | 2976 | } |
a49d0c50 | 2977 | EXPORT_SYMBOL_GPL(folio_wait_stable); |