]>
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> | |
1bf27e98 | 16 | #include <linux/math64.h> |
b95f1b31 | 17 | #include <linux/export.h> |
1da177e4 LT |
18 | #include <linux/spinlock.h> |
19 | #include <linux/fs.h> | |
20 | #include <linux/mm.h> | |
21 | #include <linux/swap.h> | |
22 | #include <linux/slab.h> | |
23 | #include <linux/pagemap.h> | |
24 | #include <linux/writeback.h> | |
25 | #include <linux/init.h> | |
26 | #include <linux/backing-dev.h> | |
55e829af | 27 | #include <linux/task_io_accounting_ops.h> |
1da177e4 LT |
28 | #include <linux/blkdev.h> |
29 | #include <linux/mpage.h> | |
d08b3851 | 30 | #include <linux/rmap.h> |
1da177e4 | 31 | #include <linux/percpu.h> |
1da177e4 LT |
32 | #include <linux/smp.h> |
33 | #include <linux/sysctl.h> | |
34 | #include <linux/cpu.h> | |
35 | #include <linux/syscalls.h> | |
811d736f | 36 | #include <linux/pagevec.h> |
eb608e3a | 37 | #include <linux/timer.h> |
8bd75c77 | 38 | #include <linux/sched/rt.h> |
f361bf4a | 39 | #include <linux/sched/signal.h> |
6e543d57 | 40 | #include <linux/mm_inline.h> |
028c2dd1 | 41 | #include <trace/events/writeback.h> |
1da177e4 | 42 | |
6e543d57 LD |
43 | #include "internal.h" |
44 | ||
ffd1f609 WF |
45 | /* |
46 | * Sleep at most 200ms at a time in balance_dirty_pages(). | |
47 | */ | |
48 | #define MAX_PAUSE max(HZ/5, 1) | |
49 | ||
5b9b3574 WF |
50 | /* |
51 | * Try to keep balance_dirty_pages() call intervals higher than this many pages | |
52 | * by raising pause time to max_pause when falls below it. | |
53 | */ | |
54 | #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10)) | |
55 | ||
e98be2d5 WF |
56 | /* |
57 | * Estimate write bandwidth at 200ms intervals. | |
58 | */ | |
59 | #define BANDWIDTH_INTERVAL max(HZ/5, 1) | |
60 | ||
6c14ae1e WF |
61 | #define RATELIMIT_CALC_SHIFT 10 |
62 | ||
1da177e4 LT |
63 | /* |
64 | * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited | |
65 | * will look to see if it needs to force writeback or throttling. | |
66 | */ | |
67 | static long ratelimit_pages = 32; | |
68 | ||
1da177e4 LT |
69 | /* The following parameters are exported via /proc/sys/vm */ |
70 | ||
71 | /* | |
5b0830cb | 72 | * Start background writeback (via writeback threads) at this percentage |
1da177e4 | 73 | */ |
aa779e51 | 74 | static int dirty_background_ratio = 10; |
1da177e4 | 75 | |
2da02997 DR |
76 | /* |
77 | * dirty_background_bytes starts at 0 (disabled) so that it is a function of | |
78 | * dirty_background_ratio * the amount of dirtyable memory | |
79 | */ | |
aa779e51 | 80 | static unsigned long dirty_background_bytes; |
2da02997 | 81 | |
195cf453 BG |
82 | /* |
83 | * free highmem will not be subtracted from the total free memory | |
84 | * for calculating free ratios if vm_highmem_is_dirtyable is true | |
85 | */ | |
aa779e51 | 86 | static int vm_highmem_is_dirtyable; |
195cf453 | 87 | |
1da177e4 LT |
88 | /* |
89 | * The generator of dirty data starts writeback at this percentage | |
90 | */ | |
aa779e51 | 91 | static int vm_dirty_ratio = 20; |
1da177e4 | 92 | |
2da02997 DR |
93 | /* |
94 | * vm_dirty_bytes starts at 0 (disabled) so that it is a function of | |
95 | * vm_dirty_ratio * the amount of dirtyable memory | |
96 | */ | |
aa779e51 | 97 | static unsigned long vm_dirty_bytes; |
2da02997 | 98 | |
1da177e4 | 99 | /* |
704503d8 | 100 | * The interval between `kupdate'-style writebacks |
1da177e4 | 101 | */ |
22ef37ee | 102 | unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */ |
1da177e4 | 103 | |
91913a29 AB |
104 | EXPORT_SYMBOL_GPL(dirty_writeback_interval); |
105 | ||
1da177e4 | 106 | /* |
704503d8 | 107 | * The longest time for which data is allowed to remain dirty |
1da177e4 | 108 | */ |
22ef37ee | 109 | unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */ |
1da177e4 | 110 | |
1da177e4 | 111 | /* |
ed5b43f1 BS |
112 | * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: |
113 | * a full sync is triggered after this time elapses without any disk activity. | |
1da177e4 LT |
114 | */ |
115 | int laptop_mode; | |
116 | ||
117 | EXPORT_SYMBOL(laptop_mode); | |
118 | ||
119 | /* End of sysctl-exported parameters */ | |
120 | ||
dcc25ae7 | 121 | struct wb_domain global_wb_domain; |
1da177e4 | 122 | |
2bc00aef TH |
123 | /* consolidated parameters for balance_dirty_pages() and its subroutines */ |
124 | struct dirty_throttle_control { | |
e9f07dfd TH |
125 | #ifdef CONFIG_CGROUP_WRITEBACK |
126 | struct wb_domain *dom; | |
9fc3a43e | 127 | struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */ |
e9f07dfd | 128 | #endif |
2bc00aef | 129 | struct bdi_writeback *wb; |
e9770b34 | 130 | struct fprop_local_percpu *wb_completions; |
eb608e3a | 131 | |
9fc3a43e | 132 | unsigned long avail; /* dirtyable */ |
2bc00aef TH |
133 | unsigned long dirty; /* file_dirty + write + nfs */ |
134 | unsigned long thresh; /* dirty threshold */ | |
135 | unsigned long bg_thresh; /* dirty background threshold */ | |
136 | ||
137 | unsigned long wb_dirty; /* per-wb counterparts */ | |
138 | unsigned long wb_thresh; | |
970fb01a | 139 | unsigned long wb_bg_thresh; |
daddfa3c TH |
140 | |
141 | unsigned long pos_ratio; | |
2bc00aef TH |
142 | }; |
143 | ||
eb608e3a JK |
144 | /* |
145 | * Length of period for aging writeout fractions of bdis. This is an | |
146 | * arbitrarily chosen number. The longer the period, the slower fractions will | |
147 | * reflect changes in current writeout rate. | |
148 | */ | |
149 | #define VM_COMPLETIONS_PERIOD_LEN (3*HZ) | |
04fbfdc1 | 150 | |
693108a8 TH |
151 | #ifdef CONFIG_CGROUP_WRITEBACK |
152 | ||
d60d1bdd TH |
153 | #define GDTC_INIT(__wb) .wb = (__wb), \ |
154 | .dom = &global_wb_domain, \ | |
155 | .wb_completions = &(__wb)->completions | |
156 | ||
9fc3a43e | 157 | #define GDTC_INIT_NO_WB .dom = &global_wb_domain |
d60d1bdd TH |
158 | |
159 | #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \ | |
160 | .dom = mem_cgroup_wb_domain(__wb), \ | |
161 | .wb_completions = &(__wb)->memcg_completions, \ | |
162 | .gdtc = __gdtc | |
c2aa723a TH |
163 | |
164 | static bool mdtc_valid(struct dirty_throttle_control *dtc) | |
165 | { | |
166 | return dtc->dom; | |
167 | } | |
e9f07dfd TH |
168 | |
169 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) | |
170 | { | |
171 | return dtc->dom; | |
172 | } | |
173 | ||
9fc3a43e TH |
174 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
175 | { | |
176 | return mdtc->gdtc; | |
177 | } | |
178 | ||
841710aa TH |
179 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
180 | { | |
181 | return &wb->memcg_completions; | |
182 | } | |
183 | ||
693108a8 TH |
184 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
185 | unsigned long *minp, unsigned long *maxp) | |
186 | { | |
20792ebf | 187 | unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth); |
693108a8 TH |
188 | unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); |
189 | unsigned long long min = wb->bdi->min_ratio; | |
190 | unsigned long long max = wb->bdi->max_ratio; | |
191 | ||
192 | /* | |
193 | * @wb may already be clean by the time control reaches here and | |
194 | * the total may not include its bw. | |
195 | */ | |
196 | if (this_bw < tot_bw) { | |
197 | if (min) { | |
198 | min *= this_bw; | |
6d9e8c65 | 199 | min = div64_ul(min, tot_bw); |
693108a8 | 200 | } |
ae82291e | 201 | if (max < 100 * BDI_RATIO_SCALE) { |
693108a8 | 202 | max *= this_bw; |
6d9e8c65 | 203 | max = div64_ul(max, tot_bw); |
693108a8 TH |
204 | } |
205 | } | |
206 | ||
207 | *minp = min; | |
208 | *maxp = max; | |
209 | } | |
210 | ||
211 | #else /* CONFIG_CGROUP_WRITEBACK */ | |
212 | ||
d60d1bdd TH |
213 | #define GDTC_INIT(__wb) .wb = (__wb), \ |
214 | .wb_completions = &(__wb)->completions | |
9fc3a43e | 215 | #define GDTC_INIT_NO_WB |
c2aa723a TH |
216 | #define MDTC_INIT(__wb, __gdtc) |
217 | ||
218 | static bool mdtc_valid(struct dirty_throttle_control *dtc) | |
219 | { | |
220 | return false; | |
221 | } | |
e9f07dfd TH |
222 | |
223 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) | |
224 | { | |
225 | return &global_wb_domain; | |
226 | } | |
227 | ||
9fc3a43e TH |
228 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
229 | { | |
230 | return NULL; | |
231 | } | |
232 | ||
841710aa TH |
233 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
234 | { | |
235 | return NULL; | |
236 | } | |
237 | ||
693108a8 TH |
238 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
239 | unsigned long *minp, unsigned long *maxp) | |
240 | { | |
241 | *minp = wb->bdi->min_ratio; | |
242 | *maxp = wb->bdi->max_ratio; | |
243 | } | |
244 | ||
245 | #endif /* CONFIG_CGROUP_WRITEBACK */ | |
246 | ||
a756cf59 JW |
247 | /* |
248 | * In a memory zone, there is a certain amount of pages we consider | |
249 | * available for the page cache, which is essentially the number of | |
250 | * free and reclaimable pages, minus some zone reserves to protect | |
251 | * lowmem and the ability to uphold the zone's watermarks without | |
252 | * requiring writeback. | |
253 | * | |
254 | * This number of dirtyable pages is the base value of which the | |
e0857cf5 | 255 | * user-configurable dirty ratio is the effective number of pages that |
a756cf59 JW |
256 | * are allowed to be actually dirtied. Per individual zone, or |
257 | * globally by using the sum of dirtyable pages over all zones. | |
258 | * | |
259 | * Because the user is allowed to specify the dirty limit globally as | |
260 | * absolute number of bytes, calculating the per-zone dirty limit can | |
261 | * require translating the configured limit into a percentage of | |
262 | * global dirtyable memory first. | |
263 | */ | |
264 | ||
a804552b | 265 | /** |
281e3726 MG |
266 | * node_dirtyable_memory - number of dirtyable pages in a node |
267 | * @pgdat: the node | |
a804552b | 268 | * |
a862f68a | 269 | * Return: the node's number of pages potentially available for dirty |
281e3726 | 270 | * page cache. This is the base value for the per-node dirty limits. |
a804552b | 271 | */ |
281e3726 | 272 | static unsigned long node_dirtyable_memory(struct pglist_data *pgdat) |
a804552b | 273 | { |
281e3726 MG |
274 | unsigned long nr_pages = 0; |
275 | int z; | |
276 | ||
277 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
278 | struct zone *zone = pgdat->node_zones + z; | |
279 | ||
280 | if (!populated_zone(zone)) | |
281 | continue; | |
282 | ||
283 | nr_pages += zone_page_state(zone, NR_FREE_PAGES); | |
284 | } | |
a804552b | 285 | |
a8d01437 JW |
286 | /* |
287 | * Pages reserved for the kernel should not be considered | |
288 | * dirtyable, to prevent a situation where reclaim has to | |
289 | * clean pages in order to balance the zones. | |
290 | */ | |
281e3726 | 291 | nr_pages -= min(nr_pages, pgdat->totalreserve_pages); |
a804552b | 292 | |
281e3726 MG |
293 | nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE); |
294 | nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE); | |
a804552b JW |
295 | |
296 | return nr_pages; | |
297 | } | |
298 | ||
1edf2234 JW |
299 | static unsigned long highmem_dirtyable_memory(unsigned long total) |
300 | { | |
301 | #ifdef CONFIG_HIGHMEM | |
302 | int node; | |
bb4cc2be | 303 | unsigned long x = 0; |
09b4ab3c | 304 | int i; |
1edf2234 JW |
305 | |
306 | for_each_node_state(node, N_HIGH_MEMORY) { | |
281e3726 MG |
307 | for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) { |
308 | struct zone *z; | |
9cb937e2 | 309 | unsigned long nr_pages; |
281e3726 MG |
310 | |
311 | if (!is_highmem_idx(i)) | |
312 | continue; | |
313 | ||
314 | z = &NODE_DATA(node)->node_zones[i]; | |
9cb937e2 MK |
315 | if (!populated_zone(z)) |
316 | continue; | |
1edf2234 | 317 | |
9cb937e2 | 318 | nr_pages = zone_page_state(z, NR_FREE_PAGES); |
281e3726 | 319 | /* watch for underflows */ |
9cb937e2 | 320 | nr_pages -= min(nr_pages, high_wmark_pages(z)); |
bb4cc2be MG |
321 | nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE); |
322 | nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE); | |
323 | x += nr_pages; | |
09b4ab3c | 324 | } |
1edf2234 | 325 | } |
281e3726 | 326 | |
1edf2234 JW |
327 | /* |
328 | * Make sure that the number of highmem pages is never larger | |
329 | * than the number of the total dirtyable memory. This can only | |
330 | * occur in very strange VM situations but we want to make sure | |
331 | * that this does not occur. | |
332 | */ | |
333 | return min(x, total); | |
334 | #else | |
335 | return 0; | |
336 | #endif | |
337 | } | |
338 | ||
339 | /** | |
ccafa287 | 340 | * global_dirtyable_memory - number of globally dirtyable pages |
1edf2234 | 341 | * |
a862f68a | 342 | * Return: the global number of pages potentially available for dirty |
ccafa287 | 343 | * page cache. This is the base value for the global dirty limits. |
1edf2234 | 344 | */ |
18cf8cf8 | 345 | static unsigned long global_dirtyable_memory(void) |
1edf2234 JW |
346 | { |
347 | unsigned long x; | |
348 | ||
c41f012a | 349 | x = global_zone_page_state(NR_FREE_PAGES); |
a8d01437 JW |
350 | /* |
351 | * Pages reserved for the kernel should not be considered | |
352 | * dirtyable, to prevent a situation where reclaim has to | |
353 | * clean pages in order to balance the zones. | |
354 | */ | |
355 | x -= min(x, totalreserve_pages); | |
1edf2234 | 356 | |
599d0c95 MG |
357 | x += global_node_page_state(NR_INACTIVE_FILE); |
358 | x += global_node_page_state(NR_ACTIVE_FILE); | |
a804552b | 359 | |
1edf2234 JW |
360 | if (!vm_highmem_is_dirtyable) |
361 | x -= highmem_dirtyable_memory(x); | |
362 | ||
363 | return x + 1; /* Ensure that we never return 0 */ | |
364 | } | |
365 | ||
9fc3a43e TH |
366 | /** |
367 | * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain | |
368 | * @dtc: dirty_throttle_control of interest | |
ccafa287 | 369 | * |
9fc3a43e TH |
370 | * Calculate @dtc->thresh and ->bg_thresh considering |
371 | * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller | |
372 | * must ensure that @dtc->avail is set before calling this function. The | |
a37b0715 | 373 | * dirty limits will be lifted by 1/4 for real-time tasks. |
ccafa287 | 374 | */ |
9fc3a43e | 375 | static void domain_dirty_limits(struct dirty_throttle_control *dtc) |
ccafa287 | 376 | { |
9fc3a43e TH |
377 | const unsigned long available_memory = dtc->avail; |
378 | struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc); | |
379 | unsigned long bytes = vm_dirty_bytes; | |
380 | unsigned long bg_bytes = dirty_background_bytes; | |
62a584fe TH |
381 | /* convert ratios to per-PAGE_SIZE for higher precision */ |
382 | unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100; | |
383 | unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100; | |
9fc3a43e TH |
384 | unsigned long thresh; |
385 | unsigned long bg_thresh; | |
ccafa287 JW |
386 | struct task_struct *tsk; |
387 | ||
9fc3a43e TH |
388 | /* gdtc is !NULL iff @dtc is for memcg domain */ |
389 | if (gdtc) { | |
390 | unsigned long global_avail = gdtc->avail; | |
391 | ||
392 | /* | |
393 | * The byte settings can't be applied directly to memcg | |
394 | * domains. Convert them to ratios by scaling against | |
62a584fe TH |
395 | * globally available memory. As the ratios are in |
396 | * per-PAGE_SIZE, they can be obtained by dividing bytes by | |
397 | * number of pages. | |
9fc3a43e TH |
398 | */ |
399 | if (bytes) | |
62a584fe TH |
400 | ratio = min(DIV_ROUND_UP(bytes, global_avail), |
401 | PAGE_SIZE); | |
9fc3a43e | 402 | if (bg_bytes) |
62a584fe TH |
403 | bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail), |
404 | PAGE_SIZE); | |
9fc3a43e TH |
405 | bytes = bg_bytes = 0; |
406 | } | |
407 | ||
408 | if (bytes) | |
409 | thresh = DIV_ROUND_UP(bytes, PAGE_SIZE); | |
ccafa287 | 410 | else |
62a584fe | 411 | thresh = (ratio * available_memory) / PAGE_SIZE; |
ccafa287 | 412 | |
9fc3a43e TH |
413 | if (bg_bytes) |
414 | bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE); | |
ccafa287 | 415 | else |
62a584fe | 416 | bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE; |
ccafa287 | 417 | |
90daf306 | 418 | if (bg_thresh >= thresh) |
9fc3a43e | 419 | bg_thresh = thresh / 2; |
ccafa287 | 420 | tsk = current; |
a37b0715 | 421 | if (rt_task(tsk)) { |
a53eaff8 N |
422 | bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32; |
423 | thresh += thresh / 4 + global_wb_domain.dirty_limit / 32; | |
ccafa287 | 424 | } |
9fc3a43e TH |
425 | dtc->thresh = thresh; |
426 | dtc->bg_thresh = bg_thresh; | |
427 | ||
428 | /* we should eventually report the domain in the TP */ | |
429 | if (!gdtc) | |
430 | trace_global_dirty_state(bg_thresh, thresh); | |
431 | } | |
432 | ||
433 | /** | |
434 | * global_dirty_limits - background-writeback and dirty-throttling thresholds | |
435 | * @pbackground: out parameter for bg_thresh | |
436 | * @pdirty: out parameter for thresh | |
437 | * | |
438 | * Calculate bg_thresh and thresh for global_wb_domain. See | |
439 | * domain_dirty_limits() for details. | |
440 | */ | |
441 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) | |
442 | { | |
443 | struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB }; | |
444 | ||
445 | gdtc.avail = global_dirtyable_memory(); | |
446 | domain_dirty_limits(&gdtc); | |
447 | ||
448 | *pbackground = gdtc.bg_thresh; | |
449 | *pdirty = gdtc.thresh; | |
ccafa287 JW |
450 | } |
451 | ||
a756cf59 | 452 | /** |
281e3726 MG |
453 | * node_dirty_limit - maximum number of dirty pages allowed in a node |
454 | * @pgdat: the node | |
a756cf59 | 455 | * |
a862f68a | 456 | * Return: the maximum number of dirty pages allowed in a node, based |
281e3726 | 457 | * on the node's dirtyable memory. |
a756cf59 | 458 | */ |
281e3726 | 459 | static unsigned long node_dirty_limit(struct pglist_data *pgdat) |
a756cf59 | 460 | { |
281e3726 | 461 | unsigned long node_memory = node_dirtyable_memory(pgdat); |
a756cf59 JW |
462 | struct task_struct *tsk = current; |
463 | unsigned long dirty; | |
464 | ||
465 | if (vm_dirty_bytes) | |
466 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) * | |
281e3726 | 467 | node_memory / global_dirtyable_memory(); |
a756cf59 | 468 | else |
281e3726 | 469 | dirty = vm_dirty_ratio * node_memory / 100; |
a756cf59 | 470 | |
a37b0715 | 471 | if (rt_task(tsk)) |
a756cf59 JW |
472 | dirty += dirty / 4; |
473 | ||
474 | return dirty; | |
475 | } | |
476 | ||
477 | /** | |
281e3726 MG |
478 | * node_dirty_ok - tells whether a node is within its dirty limits |
479 | * @pgdat: the node to check | |
a756cf59 | 480 | * |
a862f68a | 481 | * Return: %true when the dirty pages in @pgdat are within the node's |
a756cf59 JW |
482 | * dirty limit, %false if the limit is exceeded. |
483 | */ | |
281e3726 | 484 | bool node_dirty_ok(struct pglist_data *pgdat) |
a756cf59 | 485 | { |
281e3726 MG |
486 | unsigned long limit = node_dirty_limit(pgdat); |
487 | unsigned long nr_pages = 0; | |
488 | ||
11fb9989 | 489 | nr_pages += node_page_state(pgdat, NR_FILE_DIRTY); |
11fb9989 | 490 | nr_pages += node_page_state(pgdat, NR_WRITEBACK); |
a756cf59 | 491 | |
281e3726 | 492 | return nr_pages <= limit; |
a756cf59 JW |
493 | } |
494 | ||
aa779e51 | 495 | #ifdef CONFIG_SYSCTL |
496 | static int dirty_background_ratio_handler(struct ctl_table *table, int write, | |
32927393 | 497 | void *buffer, size_t *lenp, loff_t *ppos) |
2da02997 DR |
498 | { |
499 | int ret; | |
500 | ||
8d65af78 | 501 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
502 | if (ret == 0 && write) |
503 | dirty_background_bytes = 0; | |
504 | return ret; | |
505 | } | |
506 | ||
aa779e51 | 507 | static int dirty_background_bytes_handler(struct ctl_table *table, int write, |
32927393 | 508 | void *buffer, size_t *lenp, loff_t *ppos) |
2da02997 DR |
509 | { |
510 | int ret; | |
511 | ||
8d65af78 | 512 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
513 | if (ret == 0 && write) |
514 | dirty_background_ratio = 0; | |
515 | return ret; | |
516 | } | |
517 | ||
aa779e51 | 518 | static int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer, |
32927393 | 519 | size_t *lenp, loff_t *ppos) |
04fbfdc1 PZ |
520 | { |
521 | int old_ratio = vm_dirty_ratio; | |
2da02997 DR |
522 | int ret; |
523 | ||
8d65af78 | 524 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
04fbfdc1 | 525 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
eb608e3a | 526 | writeback_set_ratelimit(); |
2da02997 DR |
527 | vm_dirty_bytes = 0; |
528 | } | |
529 | return ret; | |
530 | } | |
531 | ||
aa779e51 | 532 | static int dirty_bytes_handler(struct ctl_table *table, int write, |
32927393 | 533 | void *buffer, size_t *lenp, loff_t *ppos) |
2da02997 | 534 | { |
fc3501d4 | 535 | unsigned long old_bytes = vm_dirty_bytes; |
2da02997 DR |
536 | int ret; |
537 | ||
8d65af78 | 538 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 | 539 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
eb608e3a | 540 | writeback_set_ratelimit(); |
2da02997 | 541 | vm_dirty_ratio = 0; |
04fbfdc1 PZ |
542 | } |
543 | return ret; | |
544 | } | |
aa779e51 | 545 | #endif |
04fbfdc1 | 546 | |
eb608e3a JK |
547 | static unsigned long wp_next_time(unsigned long cur_time) |
548 | { | |
549 | cur_time += VM_COMPLETIONS_PERIOD_LEN; | |
550 | /* 0 has a special meaning... */ | |
551 | if (!cur_time) | |
552 | return 1; | |
553 | return cur_time; | |
554 | } | |
555 | ||
cc24df4c | 556 | static void wb_domain_writeout_add(struct wb_domain *dom, |
c7981433 | 557 | struct fprop_local_percpu *completions, |
cc24df4c | 558 | unsigned int max_prop_frac, long nr) |
04fbfdc1 | 559 | { |
be5f1797 | 560 | __fprop_add_percpu_max(&dom->completions, completions, |
cc24df4c | 561 | max_prop_frac, nr); |
eb608e3a | 562 | /* First event after period switching was turned off? */ |
517663ed | 563 | if (unlikely(!dom->period_time)) { |
eb608e3a JK |
564 | /* |
565 | * We can race with other __bdi_writeout_inc calls here but | |
566 | * it does not cause any harm since the resulting time when | |
567 | * timer will fire and what is in writeout_period_time will be | |
568 | * roughly the same. | |
569 | */ | |
380c27ca TH |
570 | dom->period_time = wp_next_time(jiffies); |
571 | mod_timer(&dom->period_timer, dom->period_time); | |
eb608e3a | 572 | } |
04fbfdc1 PZ |
573 | } |
574 | ||
c7981433 TH |
575 | /* |
576 | * Increment @wb's writeout completion count and the global writeout | |
269ccca3 | 577 | * completion count. Called from __folio_end_writeback(). |
c7981433 | 578 | */ |
cc24df4c | 579 | static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr) |
dd5656e5 | 580 | { |
841710aa | 581 | struct wb_domain *cgdom; |
dd5656e5 | 582 | |
cc24df4c MWO |
583 | wb_stat_mod(wb, WB_WRITTEN, nr); |
584 | wb_domain_writeout_add(&global_wb_domain, &wb->completions, | |
585 | wb->bdi->max_prop_frac, nr); | |
841710aa TH |
586 | |
587 | cgdom = mem_cgroup_wb_domain(wb); | |
588 | if (cgdom) | |
cc24df4c MWO |
589 | wb_domain_writeout_add(cgdom, wb_memcg_completions(wb), |
590 | wb->bdi->max_prop_frac, nr); | |
dd5656e5 | 591 | } |
dd5656e5 | 592 | |
93f78d88 | 593 | void wb_writeout_inc(struct bdi_writeback *wb) |
04fbfdc1 | 594 | { |
dd5656e5 MS |
595 | unsigned long flags; |
596 | ||
597 | local_irq_save(flags); | |
cc24df4c | 598 | __wb_writeout_add(wb, 1); |
dd5656e5 | 599 | local_irq_restore(flags); |
04fbfdc1 | 600 | } |
93f78d88 | 601 | EXPORT_SYMBOL_GPL(wb_writeout_inc); |
04fbfdc1 | 602 | |
eb608e3a JK |
603 | /* |
604 | * On idle system, we can be called long after we scheduled because we use | |
605 | * deferred timers so count with missed periods. | |
606 | */ | |
9823e51b | 607 | static void writeout_period(struct timer_list *t) |
eb608e3a | 608 | { |
9823e51b | 609 | struct wb_domain *dom = from_timer(dom, t, period_timer); |
380c27ca | 610 | int miss_periods = (jiffies - dom->period_time) / |
eb608e3a JK |
611 | VM_COMPLETIONS_PERIOD_LEN; |
612 | ||
380c27ca TH |
613 | if (fprop_new_period(&dom->completions, miss_periods + 1)) { |
614 | dom->period_time = wp_next_time(dom->period_time + | |
eb608e3a | 615 | miss_periods * VM_COMPLETIONS_PERIOD_LEN); |
380c27ca | 616 | mod_timer(&dom->period_timer, dom->period_time); |
eb608e3a JK |
617 | } else { |
618 | /* | |
619 | * Aging has zeroed all fractions. Stop wasting CPU on period | |
620 | * updates. | |
621 | */ | |
380c27ca | 622 | dom->period_time = 0; |
eb608e3a JK |
623 | } |
624 | } | |
625 | ||
380c27ca TH |
626 | int wb_domain_init(struct wb_domain *dom, gfp_t gfp) |
627 | { | |
628 | memset(dom, 0, sizeof(*dom)); | |
dcc25ae7 TH |
629 | |
630 | spin_lock_init(&dom->lock); | |
631 | ||
9823e51b | 632 | timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE); |
dcc25ae7 TH |
633 | |
634 | dom->dirty_limit_tstamp = jiffies; | |
635 | ||
380c27ca TH |
636 | return fprop_global_init(&dom->completions, gfp); |
637 | } | |
638 | ||
841710aa TH |
639 | #ifdef CONFIG_CGROUP_WRITEBACK |
640 | void wb_domain_exit(struct wb_domain *dom) | |
641 | { | |
642 | del_timer_sync(&dom->period_timer); | |
643 | fprop_global_destroy(&dom->completions); | |
644 | } | |
645 | #endif | |
646 | ||
189d3c4a | 647 | /* |
d08c429b JW |
648 | * bdi_min_ratio keeps the sum of the minimum dirty shares of all |
649 | * registered backing devices, which, for obvious reasons, can not | |
650 | * exceed 100%. | |
189d3c4a | 651 | */ |
189d3c4a PZ |
652 | static unsigned int bdi_min_ratio; |
653 | ||
1bf27e98 SR |
654 | static int bdi_check_pages_limit(unsigned long pages) |
655 | { | |
656 | unsigned long max_dirty_pages = global_dirtyable_memory(); | |
657 | ||
658 | if (pages > max_dirty_pages) | |
659 | return -EINVAL; | |
660 | ||
661 | return 0; | |
662 | } | |
663 | ||
664 | static unsigned long bdi_ratio_from_pages(unsigned long pages) | |
665 | { | |
666 | unsigned long background_thresh; | |
667 | unsigned long dirty_thresh; | |
668 | unsigned long ratio; | |
669 | ||
670 | global_dirty_limits(&background_thresh, &dirty_thresh); | |
671 | ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh); | |
672 | ||
673 | return ratio; | |
674 | } | |
675 | ||
00df7d51 SR |
676 | static u64 bdi_get_bytes(unsigned int ratio) |
677 | { | |
678 | unsigned long background_thresh; | |
679 | unsigned long dirty_thresh; | |
680 | u64 bytes; | |
681 | ||
682 | global_dirty_limits(&background_thresh, &dirty_thresh); | |
683 | bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100; | |
684 | ||
685 | return bytes; | |
686 | } | |
687 | ||
8021fb32 | 688 | static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) |
189d3c4a | 689 | { |
21f0dd88 | 690 | unsigned int delta; |
189d3c4a | 691 | int ret = 0; |
189d3c4a | 692 | |
ae82291e SR |
693 | min_ratio *= BDI_RATIO_SCALE; |
694 | ||
cfc4ba53 | 695 | spin_lock_bh(&bdi_lock); |
a42dde04 | 696 | if (min_ratio > bdi->max_ratio) { |
189d3c4a | 697 | ret = -EINVAL; |
a42dde04 | 698 | } else { |
21f0dd88 CW |
699 | if (min_ratio < bdi->min_ratio) { |
700 | delta = bdi->min_ratio - min_ratio; | |
701 | bdi_min_ratio -= delta; | |
702 | bdi->min_ratio = min_ratio; | |
a42dde04 | 703 | } else { |
21f0dd88 | 704 | delta = min_ratio - bdi->min_ratio; |
ae82291e | 705 | if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) { |
21f0dd88 CW |
706 | bdi_min_ratio += delta; |
707 | bdi->min_ratio = min_ratio; | |
708 | } else { | |
709 | ret = -EINVAL; | |
710 | } | |
a42dde04 PZ |
711 | } |
712 | } | |
cfc4ba53 | 713 | spin_unlock_bh(&bdi_lock); |
a42dde04 PZ |
714 | |
715 | return ret; | |
716 | } | |
717 | ||
efc3e6ad | 718 | static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio) |
a42dde04 | 719 | { |
a42dde04 PZ |
720 | int ret = 0; |
721 | ||
cfc4ba53 | 722 | spin_lock_bh(&bdi_lock); |
a42dde04 PZ |
723 | if (bdi->min_ratio > max_ratio) { |
724 | ret = -EINVAL; | |
725 | } else { | |
726 | bdi->max_ratio = max_ratio; | |
eb608e3a | 727 | bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100; |
a42dde04 | 728 | } |
cfc4ba53 | 729 | spin_unlock_bh(&bdi_lock); |
189d3c4a PZ |
730 | |
731 | return ret; | |
732 | } | |
efc3e6ad | 733 | |
8021fb32 SR |
734 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) |
735 | { | |
736 | return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE); | |
737 | } | |
738 | ||
efc3e6ad SR |
739 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio) |
740 | { | |
741 | if (max_ratio > 100) | |
742 | return -EINVAL; | |
743 | ||
744 | return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE); | |
745 | } | |
a42dde04 | 746 | EXPORT_SYMBOL(bdi_set_max_ratio); |
189d3c4a | 747 | |
712c00d6 SR |
748 | u64 bdi_get_min_bytes(struct backing_dev_info *bdi) |
749 | { | |
750 | return bdi_get_bytes(bdi->min_ratio); | |
751 | } | |
752 | ||
803c9805 SR |
753 | int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes) |
754 | { | |
755 | int ret; | |
756 | unsigned long pages = min_bytes >> PAGE_SHIFT; | |
757 | unsigned long min_ratio; | |
758 | ||
759 | ret = bdi_check_pages_limit(pages); | |
760 | if (ret) | |
761 | return ret; | |
762 | ||
763 | min_ratio = bdi_ratio_from_pages(pages); | |
764 | return __bdi_set_min_ratio(bdi, min_ratio); | |
765 | } | |
766 | ||
00df7d51 SR |
767 | u64 bdi_get_max_bytes(struct backing_dev_info *bdi) |
768 | { | |
769 | return bdi_get_bytes(bdi->max_ratio); | |
770 | } | |
771 | ||
1bf27e98 SR |
772 | int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes) |
773 | { | |
774 | int ret; | |
775 | unsigned long pages = max_bytes >> PAGE_SHIFT; | |
776 | unsigned long max_ratio; | |
777 | ||
778 | ret = bdi_check_pages_limit(pages); | |
779 | if (ret) | |
780 | return ret; | |
781 | ||
782 | max_ratio = bdi_ratio_from_pages(pages); | |
783 | return __bdi_set_max_ratio(bdi, max_ratio); | |
784 | } | |
785 | ||
8e9d5ead SR |
786 | int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit) |
787 | { | |
788 | if (strict_limit > 1) | |
789 | return -EINVAL; | |
790 | ||
791 | spin_lock_bh(&bdi_lock); | |
792 | if (strict_limit) | |
793 | bdi->capabilities |= BDI_CAP_STRICTLIMIT; | |
794 | else | |
795 | bdi->capabilities &= ~BDI_CAP_STRICTLIMIT; | |
796 | spin_unlock_bh(&bdi_lock); | |
797 | ||
798 | return 0; | |
799 | } | |
800 | ||
6c14ae1e WF |
801 | static unsigned long dirty_freerun_ceiling(unsigned long thresh, |
802 | unsigned long bg_thresh) | |
803 | { | |
804 | return (thresh + bg_thresh) / 2; | |
805 | } | |
806 | ||
c7981433 TH |
807 | static unsigned long hard_dirty_limit(struct wb_domain *dom, |
808 | unsigned long thresh) | |
ffd1f609 | 809 | { |
dcc25ae7 | 810 | return max(thresh, dom->dirty_limit); |
ffd1f609 WF |
811 | } |
812 | ||
c5edf9cd TH |
813 | /* |
814 | * Memory which can be further allocated to a memcg domain is capped by | |
815 | * system-wide clean memory excluding the amount being used in the domain. | |
816 | */ | |
817 | static void mdtc_calc_avail(struct dirty_throttle_control *mdtc, | |
818 | unsigned long filepages, unsigned long headroom) | |
c2aa723a TH |
819 | { |
820 | struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc); | |
c5edf9cd TH |
821 | unsigned long clean = filepages - min(filepages, mdtc->dirty); |
822 | unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty); | |
823 | unsigned long other_clean = global_clean - min(global_clean, clean); | |
c2aa723a | 824 | |
c5edf9cd | 825 | mdtc->avail = filepages + min(headroom, other_clean); |
ffd1f609 WF |
826 | } |
827 | ||
6f718656 | 828 | /** |
b1cbc6d4 TH |
829 | * __wb_calc_thresh - @wb's share of dirty throttling threshold |
830 | * @dtc: dirty_throttle_context of interest | |
1babe183 | 831 | * |
aed21ad2 WF |
832 | * Note that balance_dirty_pages() will only seriously take it as a hard limit |
833 | * when sleeping max_pause per page is not enough to keep the dirty pages under | |
834 | * control. For example, when the device is completely stalled due to some error | |
835 | * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key. | |
836 | * In the other normal situations, it acts more gently by throttling the tasks | |
a88a341a | 837 | * more (rather than completely block them) when the wb dirty pages go high. |
1babe183 | 838 | * |
6f718656 | 839 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
1babe183 WF |
840 | * - starving fast devices |
841 | * - piling up dirty pages (that will take long time to sync) on slow devices | |
842 | * | |
a88a341a | 843 | * The wb's share of dirty limit will be adapting to its throughput and |
1babe183 | 844 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. |
a862f68a MR |
845 | * |
846 | * Return: @wb's dirty limit in pages. The term "dirty" in the context of | |
8d92890b | 847 | * dirty balancing includes all PG_dirty and PG_writeback pages. |
1babe183 | 848 | */ |
b1cbc6d4 | 849 | static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc) |
16c4042f | 850 | { |
e9f07dfd | 851 | struct wb_domain *dom = dtc_dom(dtc); |
b1cbc6d4 | 852 | unsigned long thresh = dtc->thresh; |
0d960a38 | 853 | u64 wb_thresh; |
d3ac946e | 854 | unsigned long numerator, denominator; |
693108a8 | 855 | unsigned long wb_min_ratio, wb_max_ratio; |
04fbfdc1 | 856 | |
16c4042f | 857 | /* |
0d960a38 | 858 | * Calculate this BDI's share of the thresh ratio. |
16c4042f | 859 | */ |
e9770b34 | 860 | fprop_fraction_percpu(&dom->completions, dtc->wb_completions, |
380c27ca | 861 | &numerator, &denominator); |
04fbfdc1 | 862 | |
ae82291e | 863 | wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE); |
0d960a38 | 864 | wb_thresh *= numerator; |
d3ac946e | 865 | wb_thresh = div64_ul(wb_thresh, denominator); |
04fbfdc1 | 866 | |
b1cbc6d4 | 867 | wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio); |
04fbfdc1 | 868 | |
ae82291e SR |
869 | wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE); |
870 | if (wb_thresh > (thresh * wb_max_ratio) / (100 * BDI_RATIO_SCALE)) | |
871 | wb_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE); | |
16c4042f | 872 | |
0d960a38 | 873 | return wb_thresh; |
1da177e4 LT |
874 | } |
875 | ||
b1cbc6d4 TH |
876 | unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh) |
877 | { | |
878 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb), | |
879 | .thresh = thresh }; | |
880 | return __wb_calc_thresh(&gdtc); | |
1da177e4 LT |
881 | } |
882 | ||
5a537485 MP |
883 | /* |
884 | * setpoint - dirty 3 | |
885 | * f(dirty) := 1.0 + (----------------) | |
886 | * limit - setpoint | |
887 | * | |
888 | * it's a 3rd order polynomial that subjects to | |
889 | * | |
890 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast | |
891 | * (2) f(setpoint) = 1.0 => the balance point | |
892 | * (3) f(limit) = 0 => the hard limit | |
893 | * (4) df/dx <= 0 => negative feedback control | |
894 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) | |
895 | * => fast response on large errors; small oscillation near setpoint | |
896 | */ | |
d5c9fde3 | 897 | static long long pos_ratio_polynom(unsigned long setpoint, |
5a537485 MP |
898 | unsigned long dirty, |
899 | unsigned long limit) | |
900 | { | |
901 | long long pos_ratio; | |
902 | long x; | |
903 | ||
d5c9fde3 | 904 | x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT, |
464d1387 | 905 | (limit - setpoint) | 1); |
5a537485 MP |
906 | pos_ratio = x; |
907 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
908 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
909 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT; | |
910 | ||
911 | return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT); | |
912 | } | |
913 | ||
6c14ae1e WF |
914 | /* |
915 | * Dirty position control. | |
916 | * | |
917 | * (o) global/bdi setpoints | |
918 | * | |
de1fff37 | 919 | * We want the dirty pages be balanced around the global/wb setpoints. |
6c14ae1e WF |
920 | * When the number of dirty pages is higher/lower than the setpoint, the |
921 | * dirty position control ratio (and hence task dirty ratelimit) will be | |
922 | * decreased/increased to bring the dirty pages back to the setpoint. | |
923 | * | |
924 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT | |
925 | * | |
926 | * if (dirty < setpoint) scale up pos_ratio | |
927 | * if (dirty > setpoint) scale down pos_ratio | |
928 | * | |
de1fff37 TH |
929 | * if (wb_dirty < wb_setpoint) scale up pos_ratio |
930 | * if (wb_dirty > wb_setpoint) scale down pos_ratio | |
6c14ae1e WF |
931 | * |
932 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT | |
933 | * | |
934 | * (o) global control line | |
935 | * | |
936 | * ^ pos_ratio | |
937 | * | | |
938 | * | |<===== global dirty control scope ======>| | |
03231554 | 939 | * 2.0 * * * * * * * |
6c14ae1e WF |
940 | * | .* |
941 | * | . * | |
942 | * | . * | |
943 | * | . * | |
944 | * | . * | |
945 | * | . * | |
946 | * 1.0 ................................* | |
947 | * | . . * | |
948 | * | . . * | |
949 | * | . . * | |
950 | * | . . * | |
951 | * | . . * | |
952 | * 0 +------------.------------------.----------------------*-------------> | |
953 | * freerun^ setpoint^ limit^ dirty pages | |
954 | * | |
de1fff37 | 955 | * (o) wb control line |
6c14ae1e WF |
956 | * |
957 | * ^ pos_ratio | |
958 | * | | |
959 | * | * | |
960 | * | * | |
961 | * | * | |
962 | * | * | |
963 | * | * |<=========== span ============>| | |
964 | * 1.0 .......................* | |
965 | * | . * | |
966 | * | . * | |
967 | * | . * | |
968 | * | . * | |
969 | * | . * | |
970 | * | . * | |
971 | * | . * | |
972 | * | . * | |
973 | * | . * | |
974 | * | . * | |
975 | * | . * | |
976 | * 1/4 ...............................................* * * * * * * * * * * * | |
977 | * | . . | |
978 | * | . . | |
979 | * | . . | |
980 | * 0 +----------------------.-------------------------------.-------------> | |
de1fff37 | 981 | * wb_setpoint^ x_intercept^ |
6c14ae1e | 982 | * |
de1fff37 | 983 | * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can |
6c14ae1e WF |
984 | * be smoothly throttled down to normal if it starts high in situations like |
985 | * - start writing to a slow SD card and a fast disk at the same time. The SD | |
de1fff37 TH |
986 | * card's wb_dirty may rush to many times higher than wb_setpoint. |
987 | * - the wb dirty thresh drops quickly due to change of JBOD workload | |
6c14ae1e | 988 | */ |
daddfa3c | 989 | static void wb_position_ratio(struct dirty_throttle_control *dtc) |
6c14ae1e | 990 | { |
2bc00aef | 991 | struct bdi_writeback *wb = dtc->wb; |
20792ebf | 992 | unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth); |
2bc00aef | 993 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); |
c7981433 | 994 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
2bc00aef | 995 | unsigned long wb_thresh = dtc->wb_thresh; |
6c14ae1e WF |
996 | unsigned long x_intercept; |
997 | unsigned long setpoint; /* dirty pages' target balance point */ | |
de1fff37 | 998 | unsigned long wb_setpoint; |
6c14ae1e WF |
999 | unsigned long span; |
1000 | long long pos_ratio; /* for scaling up/down the rate limit */ | |
1001 | long x; | |
1002 | ||
daddfa3c TH |
1003 | dtc->pos_ratio = 0; |
1004 | ||
2bc00aef | 1005 | if (unlikely(dtc->dirty >= limit)) |
daddfa3c | 1006 | return; |
6c14ae1e WF |
1007 | |
1008 | /* | |
1009 | * global setpoint | |
1010 | * | |
5a537485 MP |
1011 | * See comment for pos_ratio_polynom(). |
1012 | */ | |
1013 | setpoint = (freerun + limit) / 2; | |
2bc00aef | 1014 | pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit); |
5a537485 MP |
1015 | |
1016 | /* | |
1017 | * The strictlimit feature is a tool preventing mistrusted filesystems | |
1018 | * from growing a large number of dirty pages before throttling. For | |
de1fff37 TH |
1019 | * such filesystems balance_dirty_pages always checks wb counters |
1020 | * against wb limits. Even if global "nr_dirty" is under "freerun". | |
5a537485 MP |
1021 | * This is especially important for fuse which sets bdi->max_ratio to |
1022 | * 1% by default. Without strictlimit feature, fuse writeback may | |
1023 | * consume arbitrary amount of RAM because it is accounted in | |
1024 | * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty". | |
6c14ae1e | 1025 | * |
a88a341a | 1026 | * Here, in wb_position_ratio(), we calculate pos_ratio based on |
de1fff37 | 1027 | * two values: wb_dirty and wb_thresh. Let's consider an example: |
5a537485 MP |
1028 | * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global |
1029 | * limits are set by default to 10% and 20% (background and throttle). | |
de1fff37 | 1030 | * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages. |
0d960a38 | 1031 | * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is |
de1fff37 | 1032 | * about ~6K pages (as the average of background and throttle wb |
5a537485 | 1033 | * limits). The 3rd order polynomial will provide positive feedback if |
de1fff37 | 1034 | * wb_dirty is under wb_setpoint and vice versa. |
6c14ae1e | 1035 | * |
5a537485 | 1036 | * Note, that we cannot use global counters in these calculations |
de1fff37 | 1037 | * because we want to throttle process writing to a strictlimit wb |
5a537485 MP |
1038 | * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB |
1039 | * in the example above). | |
6c14ae1e | 1040 | */ |
a88a341a | 1041 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
de1fff37 | 1042 | long long wb_pos_ratio; |
5a537485 | 1043 | |
daddfa3c TH |
1044 | if (dtc->wb_dirty < 8) { |
1045 | dtc->pos_ratio = min_t(long long, pos_ratio * 2, | |
1046 | 2 << RATELIMIT_CALC_SHIFT); | |
1047 | return; | |
1048 | } | |
5a537485 | 1049 | |
2bc00aef | 1050 | if (dtc->wb_dirty >= wb_thresh) |
daddfa3c | 1051 | return; |
5a537485 | 1052 | |
970fb01a TH |
1053 | wb_setpoint = dirty_freerun_ceiling(wb_thresh, |
1054 | dtc->wb_bg_thresh); | |
5a537485 | 1055 | |
de1fff37 | 1056 | if (wb_setpoint == 0 || wb_setpoint == wb_thresh) |
daddfa3c | 1057 | return; |
5a537485 | 1058 | |
2bc00aef | 1059 | wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty, |
de1fff37 | 1060 | wb_thresh); |
5a537485 MP |
1061 | |
1062 | /* | |
de1fff37 TH |
1063 | * Typically, for strictlimit case, wb_setpoint << setpoint |
1064 | * and pos_ratio >> wb_pos_ratio. In the other words global | |
5a537485 | 1065 | * state ("dirty") is not limiting factor and we have to |
de1fff37 | 1066 | * make decision based on wb counters. But there is an |
5a537485 MP |
1067 | * important case when global pos_ratio should get precedence: |
1068 | * global limits are exceeded (e.g. due to activities on other | |
de1fff37 | 1069 | * wb's) while given strictlimit wb is below limit. |
5a537485 | 1070 | * |
de1fff37 | 1071 | * "pos_ratio * wb_pos_ratio" would work for the case above, |
5a537485 | 1072 | * but it would look too non-natural for the case of all |
de1fff37 | 1073 | * activity in the system coming from a single strictlimit wb |
5a537485 MP |
1074 | * with bdi->max_ratio == 100%. |
1075 | * | |
1076 | * Note that min() below somewhat changes the dynamics of the | |
1077 | * control system. Normally, pos_ratio value can be well over 3 | |
de1fff37 | 1078 | * (when globally we are at freerun and wb is well below wb |
5a537485 MP |
1079 | * setpoint). Now the maximum pos_ratio in the same situation |
1080 | * is 2. We might want to tweak this if we observe the control | |
1081 | * system is too slow to adapt. | |
1082 | */ | |
daddfa3c TH |
1083 | dtc->pos_ratio = min(pos_ratio, wb_pos_ratio); |
1084 | return; | |
5a537485 | 1085 | } |
6c14ae1e WF |
1086 | |
1087 | /* | |
1088 | * We have computed basic pos_ratio above based on global situation. If | |
de1fff37 | 1089 | * the wb is over/under its share of dirty pages, we want to scale |
6c14ae1e WF |
1090 | * pos_ratio further down/up. That is done by the following mechanism. |
1091 | */ | |
1092 | ||
1093 | /* | |
de1fff37 | 1094 | * wb setpoint |
6c14ae1e | 1095 | * |
de1fff37 | 1096 | * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint) |
6c14ae1e | 1097 | * |
de1fff37 | 1098 | * x_intercept - wb_dirty |
6c14ae1e | 1099 | * := -------------------------- |
de1fff37 | 1100 | * x_intercept - wb_setpoint |
6c14ae1e | 1101 | * |
de1fff37 | 1102 | * The main wb control line is a linear function that subjects to |
6c14ae1e | 1103 | * |
de1fff37 TH |
1104 | * (1) f(wb_setpoint) = 1.0 |
1105 | * (2) k = - 1 / (8 * write_bw) (in single wb case) | |
1106 | * or equally: x_intercept = wb_setpoint + 8 * write_bw | |
6c14ae1e | 1107 | * |
de1fff37 | 1108 | * For single wb case, the dirty pages are observed to fluctuate |
6c14ae1e | 1109 | * regularly within range |
de1fff37 | 1110 | * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2] |
6c14ae1e WF |
1111 | * for various filesystems, where (2) can yield in a reasonable 12.5% |
1112 | * fluctuation range for pos_ratio. | |
1113 | * | |
de1fff37 | 1114 | * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its |
6c14ae1e | 1115 | * own size, so move the slope over accordingly and choose a slope that |
de1fff37 | 1116 | * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh. |
6c14ae1e | 1117 | */ |
2bc00aef TH |
1118 | if (unlikely(wb_thresh > dtc->thresh)) |
1119 | wb_thresh = dtc->thresh; | |
aed21ad2 | 1120 | /* |
de1fff37 | 1121 | * It's very possible that wb_thresh is close to 0 not because the |
aed21ad2 WF |
1122 | * device is slow, but that it has remained inactive for long time. |
1123 | * Honour such devices a reasonable good (hopefully IO efficient) | |
1124 | * threshold, so that the occasional writes won't be blocked and active | |
1125 | * writes can rampup the threshold quickly. | |
1126 | */ | |
2bc00aef | 1127 | wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8); |
6c14ae1e | 1128 | /* |
de1fff37 TH |
1129 | * scale global setpoint to wb's: |
1130 | * wb_setpoint = setpoint * wb_thresh / thresh | |
6c14ae1e | 1131 | */ |
e4bc13ad | 1132 | x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1); |
de1fff37 | 1133 | wb_setpoint = setpoint * (u64)x >> 16; |
6c14ae1e | 1134 | /* |
de1fff37 TH |
1135 | * Use span=(8*write_bw) in single wb case as indicated by |
1136 | * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case. | |
6c14ae1e | 1137 | * |
de1fff37 TH |
1138 | * wb_thresh thresh - wb_thresh |
1139 | * span = --------- * (8 * write_bw) + ------------------ * wb_thresh | |
1140 | * thresh thresh | |
6c14ae1e | 1141 | */ |
2bc00aef | 1142 | span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16; |
de1fff37 | 1143 | x_intercept = wb_setpoint + span; |
6c14ae1e | 1144 | |
2bc00aef TH |
1145 | if (dtc->wb_dirty < x_intercept - span / 4) { |
1146 | pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty), | |
e4bc13ad | 1147 | (x_intercept - wb_setpoint) | 1); |
6c14ae1e WF |
1148 | } else |
1149 | pos_ratio /= 4; | |
1150 | ||
8927f66c | 1151 | /* |
de1fff37 | 1152 | * wb reserve area, safeguard against dirty pool underrun and disk idle |
8927f66c WF |
1153 | * It may push the desired control point of global dirty pages higher |
1154 | * than setpoint. | |
1155 | */ | |
de1fff37 | 1156 | x_intercept = wb_thresh / 2; |
2bc00aef TH |
1157 | if (dtc->wb_dirty < x_intercept) { |
1158 | if (dtc->wb_dirty > x_intercept / 8) | |
1159 | pos_ratio = div_u64(pos_ratio * x_intercept, | |
1160 | dtc->wb_dirty); | |
50657fc4 | 1161 | else |
8927f66c WF |
1162 | pos_ratio *= 8; |
1163 | } | |
1164 | ||
daddfa3c | 1165 | dtc->pos_ratio = pos_ratio; |
6c14ae1e WF |
1166 | } |
1167 | ||
a88a341a TH |
1168 | static void wb_update_write_bandwidth(struct bdi_writeback *wb, |
1169 | unsigned long elapsed, | |
1170 | unsigned long written) | |
e98be2d5 WF |
1171 | { |
1172 | const unsigned long period = roundup_pow_of_two(3 * HZ); | |
a88a341a TH |
1173 | unsigned long avg = wb->avg_write_bandwidth; |
1174 | unsigned long old = wb->write_bandwidth; | |
e98be2d5 WF |
1175 | u64 bw; |
1176 | ||
1177 | /* | |
1178 | * bw = written * HZ / elapsed | |
1179 | * | |
1180 | * bw * elapsed + write_bandwidth * (period - elapsed) | |
1181 | * write_bandwidth = --------------------------------------------------- | |
1182 | * period | |
c72efb65 | 1183 | * |
25ff8b15 | 1184 | * @written may have decreased due to folio_account_redirty(). |
c72efb65 | 1185 | * Avoid underflowing @bw calculation. |
e98be2d5 | 1186 | */ |
a88a341a | 1187 | bw = written - min(written, wb->written_stamp); |
e98be2d5 WF |
1188 | bw *= HZ; |
1189 | if (unlikely(elapsed > period)) { | |
0a5d1a7f | 1190 | bw = div64_ul(bw, elapsed); |
e98be2d5 WF |
1191 | avg = bw; |
1192 | goto out; | |
1193 | } | |
a88a341a | 1194 | bw += (u64)wb->write_bandwidth * (period - elapsed); |
e98be2d5 WF |
1195 | bw >>= ilog2(period); |
1196 | ||
1197 | /* | |
1198 | * one more level of smoothing, for filtering out sudden spikes | |
1199 | */ | |
1200 | if (avg > old && old >= (unsigned long)bw) | |
1201 | avg -= (avg - old) >> 3; | |
1202 | ||
1203 | if (avg < old && old <= (unsigned long)bw) | |
1204 | avg += (old - avg) >> 3; | |
1205 | ||
1206 | out: | |
95a46c65 TH |
1207 | /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */ |
1208 | avg = max(avg, 1LU); | |
1209 | if (wb_has_dirty_io(wb)) { | |
1210 | long delta = avg - wb->avg_write_bandwidth; | |
1211 | WARN_ON_ONCE(atomic_long_add_return(delta, | |
1212 | &wb->bdi->tot_write_bandwidth) <= 0); | |
1213 | } | |
a88a341a | 1214 | wb->write_bandwidth = bw; |
20792ebf | 1215 | WRITE_ONCE(wb->avg_write_bandwidth, avg); |
e98be2d5 WF |
1216 | } |
1217 | ||
2bc00aef | 1218 | static void update_dirty_limit(struct dirty_throttle_control *dtc) |
c42843f2 | 1219 | { |
e9f07dfd | 1220 | struct wb_domain *dom = dtc_dom(dtc); |
2bc00aef | 1221 | unsigned long thresh = dtc->thresh; |
dcc25ae7 | 1222 | unsigned long limit = dom->dirty_limit; |
c42843f2 WF |
1223 | |
1224 | /* | |
1225 | * Follow up in one step. | |
1226 | */ | |
1227 | if (limit < thresh) { | |
1228 | limit = thresh; | |
1229 | goto update; | |
1230 | } | |
1231 | ||
1232 | /* | |
1233 | * Follow down slowly. Use the higher one as the target, because thresh | |
1234 | * may drop below dirty. This is exactly the reason to introduce | |
dcc25ae7 | 1235 | * dom->dirty_limit which is guaranteed to lie above the dirty pages. |
c42843f2 | 1236 | */ |
2bc00aef | 1237 | thresh = max(thresh, dtc->dirty); |
c42843f2 WF |
1238 | if (limit > thresh) { |
1239 | limit -= (limit - thresh) >> 5; | |
1240 | goto update; | |
1241 | } | |
1242 | return; | |
1243 | update: | |
dcc25ae7 | 1244 | dom->dirty_limit = limit; |
c42843f2 WF |
1245 | } |
1246 | ||
42dd235c JK |
1247 | static void domain_update_dirty_limit(struct dirty_throttle_control *dtc, |
1248 | unsigned long now) | |
c42843f2 | 1249 | { |
e9f07dfd | 1250 | struct wb_domain *dom = dtc_dom(dtc); |
c42843f2 WF |
1251 | |
1252 | /* | |
1253 | * check locklessly first to optimize away locking for the most time | |
1254 | */ | |
dcc25ae7 | 1255 | if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) |
c42843f2 WF |
1256 | return; |
1257 | ||
dcc25ae7 TH |
1258 | spin_lock(&dom->lock); |
1259 | if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) { | |
2bc00aef | 1260 | update_dirty_limit(dtc); |
dcc25ae7 | 1261 | dom->dirty_limit_tstamp = now; |
c42843f2 | 1262 | } |
dcc25ae7 | 1263 | spin_unlock(&dom->lock); |
c42843f2 WF |
1264 | } |
1265 | ||
be3ffa27 | 1266 | /* |
de1fff37 | 1267 | * Maintain wb->dirty_ratelimit, the base dirty throttle rate. |
be3ffa27 | 1268 | * |
de1fff37 | 1269 | * Normal wb tasks will be curbed at or below it in long term. |
be3ffa27 WF |
1270 | * Obviously it should be around (write_bw / N) when there are N dd tasks. |
1271 | */ | |
2bc00aef | 1272 | static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc, |
a88a341a TH |
1273 | unsigned long dirtied, |
1274 | unsigned long elapsed) | |
be3ffa27 | 1275 | { |
2bc00aef TH |
1276 | struct bdi_writeback *wb = dtc->wb; |
1277 | unsigned long dirty = dtc->dirty; | |
1278 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); | |
c7981433 | 1279 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
7381131c | 1280 | unsigned long setpoint = (freerun + limit) / 2; |
a88a341a TH |
1281 | unsigned long write_bw = wb->avg_write_bandwidth; |
1282 | unsigned long dirty_ratelimit = wb->dirty_ratelimit; | |
be3ffa27 WF |
1283 | unsigned long dirty_rate; |
1284 | unsigned long task_ratelimit; | |
1285 | unsigned long balanced_dirty_ratelimit; | |
7381131c WF |
1286 | unsigned long step; |
1287 | unsigned long x; | |
d59b1087 | 1288 | unsigned long shift; |
be3ffa27 WF |
1289 | |
1290 | /* | |
1291 | * The dirty rate will match the writeout rate in long term, except | |
1292 | * when dirty pages are truncated by userspace or re-dirtied by FS. | |
1293 | */ | |
a88a341a | 1294 | dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed; |
be3ffa27 | 1295 | |
be3ffa27 WF |
1296 | /* |
1297 | * task_ratelimit reflects each dd's dirty rate for the past 200ms. | |
1298 | */ | |
1299 | task_ratelimit = (u64)dirty_ratelimit * | |
daddfa3c | 1300 | dtc->pos_ratio >> RATELIMIT_CALC_SHIFT; |
be3ffa27 WF |
1301 | task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ |
1302 | ||
1303 | /* | |
1304 | * A linear estimation of the "balanced" throttle rate. The theory is, | |
de1fff37 | 1305 | * if there are N dd tasks, each throttled at task_ratelimit, the wb's |
be3ffa27 WF |
1306 | * dirty_rate will be measured to be (N * task_ratelimit). So the below |
1307 | * formula will yield the balanced rate limit (write_bw / N). | |
1308 | * | |
1309 | * Note that the expanded form is not a pure rate feedback: | |
1310 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) | |
1311 | * but also takes pos_ratio into account: | |
1312 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) | |
1313 | * | |
1314 | * (1) is not realistic because pos_ratio also takes part in balancing | |
1315 | * the dirty rate. Consider the state | |
1316 | * pos_ratio = 0.5 (3) | |
1317 | * rate = 2 * (write_bw / N) (4) | |
1318 | * If (1) is used, it will stuck in that state! Because each dd will | |
1319 | * be throttled at | |
1320 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) | |
1321 | * yielding | |
1322 | * dirty_rate = N * task_ratelimit = write_bw (6) | |
1323 | * put (6) into (1) we get | |
1324 | * rate_(i+1) = rate_(i) (7) | |
1325 | * | |
1326 | * So we end up using (2) to always keep | |
1327 | * rate_(i+1) ~= (write_bw / N) (8) | |
1328 | * regardless of the value of pos_ratio. As long as (8) is satisfied, | |
1329 | * pos_ratio is able to drive itself to 1.0, which is not only where | |
1330 | * the dirty count meet the setpoint, but also where the slope of | |
1331 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated. | |
1332 | */ | |
1333 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, | |
1334 | dirty_rate | 1); | |
bdaac490 WF |
1335 | /* |
1336 | * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw | |
1337 | */ | |
1338 | if (unlikely(balanced_dirty_ratelimit > write_bw)) | |
1339 | balanced_dirty_ratelimit = write_bw; | |
be3ffa27 | 1340 | |
7381131c WF |
1341 | /* |
1342 | * We could safely do this and return immediately: | |
1343 | * | |
de1fff37 | 1344 | * wb->dirty_ratelimit = balanced_dirty_ratelimit; |
7381131c WF |
1345 | * |
1346 | * However to get a more stable dirty_ratelimit, the below elaborated | |
331cbdee | 1347 | * code makes use of task_ratelimit to filter out singular points and |
7381131c WF |
1348 | * limit the step size. |
1349 | * | |
1350 | * The below code essentially only uses the relative value of | |
1351 | * | |
1352 | * task_ratelimit - dirty_ratelimit | |
1353 | * = (pos_ratio - 1) * dirty_ratelimit | |
1354 | * | |
1355 | * which reflects the direction and size of dirty position error. | |
1356 | */ | |
1357 | ||
1358 | /* | |
1359 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff | |
1360 | * task_ratelimit is on the same side of dirty_ratelimit, too. | |
1361 | * For example, when | |
1362 | * - dirty_ratelimit > balanced_dirty_ratelimit | |
1363 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) | |
1364 | * lowering dirty_ratelimit will help meet both the position and rate | |
1365 | * control targets. Otherwise, don't update dirty_ratelimit if it will | |
1366 | * only help meet the rate target. After all, what the users ultimately | |
1367 | * feel and care are stable dirty rate and small position error. | |
1368 | * | |
1369 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size | |
331cbdee | 1370 | * and filter out the singular points of balanced_dirty_ratelimit. Which |
7381131c WF |
1371 | * keeps jumping around randomly and can even leap far away at times |
1372 | * due to the small 200ms estimation period of dirty_rate (we want to | |
1373 | * keep that period small to reduce time lags). | |
1374 | */ | |
1375 | step = 0; | |
5a537485 MP |
1376 | |
1377 | /* | |
de1fff37 | 1378 | * For strictlimit case, calculations above were based on wb counters |
a88a341a | 1379 | * and limits (starting from pos_ratio = wb_position_ratio() and up to |
5a537485 | 1380 | * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate). |
de1fff37 TH |
1381 | * Hence, to calculate "step" properly, we have to use wb_dirty as |
1382 | * "dirty" and wb_setpoint as "setpoint". | |
5a537485 | 1383 | * |
de1fff37 TH |
1384 | * We rampup dirty_ratelimit forcibly if wb_dirty is low because |
1385 | * it's possible that wb_thresh is close to zero due to inactivity | |
970fb01a | 1386 | * of backing device. |
5a537485 | 1387 | */ |
a88a341a | 1388 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
2bc00aef TH |
1389 | dirty = dtc->wb_dirty; |
1390 | if (dtc->wb_dirty < 8) | |
1391 | setpoint = dtc->wb_dirty + 1; | |
5a537485 | 1392 | else |
970fb01a | 1393 | setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2; |
5a537485 MP |
1394 | } |
1395 | ||
7381131c | 1396 | if (dirty < setpoint) { |
a88a341a | 1397 | x = min3(wb->balanced_dirty_ratelimit, |
7c809968 | 1398 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1399 | if (dirty_ratelimit < x) |
1400 | step = x - dirty_ratelimit; | |
1401 | } else { | |
a88a341a | 1402 | x = max3(wb->balanced_dirty_ratelimit, |
7c809968 | 1403 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1404 | if (dirty_ratelimit > x) |
1405 | step = dirty_ratelimit - x; | |
1406 | } | |
1407 | ||
1408 | /* | |
1409 | * Don't pursue 100% rate matching. It's impossible since the balanced | |
1410 | * rate itself is constantly fluctuating. So decrease the track speed | |
1411 | * when it gets close to the target. Helps eliminate pointless tremors. | |
1412 | */ | |
d59b1087 AR |
1413 | shift = dirty_ratelimit / (2 * step + 1); |
1414 | if (shift < BITS_PER_LONG) | |
1415 | step = DIV_ROUND_UP(step >> shift, 8); | |
1416 | else | |
1417 | step = 0; | |
7381131c WF |
1418 | |
1419 | if (dirty_ratelimit < balanced_dirty_ratelimit) | |
1420 | dirty_ratelimit += step; | |
1421 | else | |
1422 | dirty_ratelimit -= step; | |
1423 | ||
20792ebf | 1424 | WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL)); |
a88a341a | 1425 | wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit; |
b48c104d | 1426 | |
5634cc2a | 1427 | trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit); |
be3ffa27 WF |
1428 | } |
1429 | ||
c2aa723a TH |
1430 | static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc, |
1431 | struct dirty_throttle_control *mdtc, | |
8a731799 | 1432 | bool update_ratelimit) |
e98be2d5 | 1433 | { |
c2aa723a | 1434 | struct bdi_writeback *wb = gdtc->wb; |
e98be2d5 | 1435 | unsigned long now = jiffies; |
45a2966f | 1436 | unsigned long elapsed; |
be3ffa27 | 1437 | unsigned long dirtied; |
e98be2d5 WF |
1438 | unsigned long written; |
1439 | ||
45a2966f | 1440 | spin_lock(&wb->list_lock); |
8a731799 | 1441 | |
e98be2d5 | 1442 | /* |
45a2966f JK |
1443 | * Lockless checks for elapsed time are racy and delayed update after |
1444 | * IO completion doesn't do it at all (to make sure written pages are | |
1445 | * accounted reasonably quickly). Make sure elapsed >= 1 to avoid | |
1446 | * division errors. | |
e98be2d5 | 1447 | */ |
45a2966f | 1448 | elapsed = max(now - wb->bw_time_stamp, 1UL); |
a88a341a TH |
1449 | dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]); |
1450 | written = percpu_counter_read(&wb->stat[WB_WRITTEN]); | |
e98be2d5 | 1451 | |
8a731799 | 1452 | if (update_ratelimit) { |
42dd235c | 1453 | domain_update_dirty_limit(gdtc, now); |
c2aa723a TH |
1454 | wb_update_dirty_ratelimit(gdtc, dirtied, elapsed); |
1455 | ||
1456 | /* | |
1457 | * @mdtc is always NULL if !CGROUP_WRITEBACK but the | |
1458 | * compiler has no way to figure that out. Help it. | |
1459 | */ | |
1460 | if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) { | |
42dd235c | 1461 | domain_update_dirty_limit(mdtc, now); |
c2aa723a TH |
1462 | wb_update_dirty_ratelimit(mdtc, dirtied, elapsed); |
1463 | } | |
be3ffa27 | 1464 | } |
a88a341a | 1465 | wb_update_write_bandwidth(wb, elapsed, written); |
e98be2d5 | 1466 | |
a88a341a TH |
1467 | wb->dirtied_stamp = dirtied; |
1468 | wb->written_stamp = written; | |
20792ebf | 1469 | WRITE_ONCE(wb->bw_time_stamp, now); |
45a2966f | 1470 | spin_unlock(&wb->list_lock); |
e98be2d5 WF |
1471 | } |
1472 | ||
45a2966f | 1473 | void wb_update_bandwidth(struct bdi_writeback *wb) |
e98be2d5 | 1474 | { |
2bc00aef TH |
1475 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
1476 | ||
fee468fd | 1477 | __wb_update_bandwidth(&gdtc, NULL, false); |
fee468fd JK |
1478 | } |
1479 | ||
1480 | /* Interval after which we consider wb idle and don't estimate bandwidth */ | |
1481 | #define WB_BANDWIDTH_IDLE_JIF (HZ) | |
1482 | ||
1483 | static void wb_bandwidth_estimate_start(struct bdi_writeback *wb) | |
1484 | { | |
1485 | unsigned long now = jiffies; | |
1486 | unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp); | |
1487 | ||
1488 | if (elapsed > WB_BANDWIDTH_IDLE_JIF && | |
1489 | !atomic_read(&wb->writeback_inodes)) { | |
1490 | spin_lock(&wb->list_lock); | |
1491 | wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED); | |
1492 | wb->written_stamp = wb_stat(wb, WB_WRITTEN); | |
20792ebf | 1493 | WRITE_ONCE(wb->bw_time_stamp, now); |
fee468fd JK |
1494 | spin_unlock(&wb->list_lock); |
1495 | } | |
e98be2d5 WF |
1496 | } |
1497 | ||
9d823e8f | 1498 | /* |
d0e1d66b | 1499 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited() |
9d823e8f WF |
1500 | * will look to see if it needs to start dirty throttling. |
1501 | * | |
1502 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive | |
c41f012a | 1503 | * global_zone_page_state() too often. So scale it near-sqrt to the safety margin |
9d823e8f WF |
1504 | * (the number of pages we may dirty without exceeding the dirty limits). |
1505 | */ | |
1506 | static unsigned long dirty_poll_interval(unsigned long dirty, | |
1507 | unsigned long thresh) | |
1508 | { | |
1509 | if (thresh > dirty) | |
1510 | return 1UL << (ilog2(thresh - dirty) >> 1); | |
1511 | ||
1512 | return 1; | |
1513 | } | |
1514 | ||
a88a341a | 1515 | static unsigned long wb_max_pause(struct bdi_writeback *wb, |
de1fff37 | 1516 | unsigned long wb_dirty) |
c8462cc9 | 1517 | { |
20792ebf | 1518 | unsigned long bw = READ_ONCE(wb->avg_write_bandwidth); |
e3b6c655 | 1519 | unsigned long t; |
c8462cc9 | 1520 | |
7ccb9ad5 WF |
1521 | /* |
1522 | * Limit pause time for small memory systems. If sleeping for too long | |
1523 | * time, a small pool of dirty/writeback pages may go empty and disk go | |
1524 | * idle. | |
1525 | * | |
1526 | * 8 serves as the safety ratio. | |
1527 | */ | |
de1fff37 | 1528 | t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8)); |
7ccb9ad5 WF |
1529 | t++; |
1530 | ||
e3b6c655 | 1531 | return min_t(unsigned long, t, MAX_PAUSE); |
7ccb9ad5 WF |
1532 | } |
1533 | ||
a88a341a TH |
1534 | static long wb_min_pause(struct bdi_writeback *wb, |
1535 | long max_pause, | |
1536 | unsigned long task_ratelimit, | |
1537 | unsigned long dirty_ratelimit, | |
1538 | int *nr_dirtied_pause) | |
c8462cc9 | 1539 | { |
20792ebf JK |
1540 | long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth)); |
1541 | long lo = ilog2(READ_ONCE(wb->dirty_ratelimit)); | |
7ccb9ad5 WF |
1542 | long t; /* target pause */ |
1543 | long pause; /* estimated next pause */ | |
1544 | int pages; /* target nr_dirtied_pause */ | |
c8462cc9 | 1545 | |
7ccb9ad5 WF |
1546 | /* target for 10ms pause on 1-dd case */ |
1547 | t = max(1, HZ / 100); | |
c8462cc9 WF |
1548 | |
1549 | /* | |
1550 | * Scale up pause time for concurrent dirtiers in order to reduce CPU | |
1551 | * overheads. | |
1552 | * | |
7ccb9ad5 | 1553 | * (N * 10ms) on 2^N concurrent tasks. |
c8462cc9 WF |
1554 | */ |
1555 | if (hi > lo) | |
7ccb9ad5 | 1556 | t += (hi - lo) * (10 * HZ) / 1024; |
c8462cc9 WF |
1557 | |
1558 | /* | |
7ccb9ad5 WF |
1559 | * This is a bit convoluted. We try to base the next nr_dirtied_pause |
1560 | * on the much more stable dirty_ratelimit. However the next pause time | |
1561 | * will be computed based on task_ratelimit and the two rate limits may | |
1562 | * depart considerably at some time. Especially if task_ratelimit goes | |
1563 | * below dirty_ratelimit/2 and the target pause is max_pause, the next | |
1564 | * pause time will be max_pause*2 _trimmed down_ to max_pause. As a | |
1565 | * result task_ratelimit won't be executed faithfully, which could | |
1566 | * eventually bring down dirty_ratelimit. | |
c8462cc9 | 1567 | * |
7ccb9ad5 WF |
1568 | * We apply two rules to fix it up: |
1569 | * 1) try to estimate the next pause time and if necessary, use a lower | |
1570 | * nr_dirtied_pause so as not to exceed max_pause. When this happens, | |
1571 | * nr_dirtied_pause will be "dancing" with task_ratelimit. | |
1572 | * 2) limit the target pause time to max_pause/2, so that the normal | |
1573 | * small fluctuations of task_ratelimit won't trigger rule (1) and | |
1574 | * nr_dirtied_pause will remain as stable as dirty_ratelimit. | |
c8462cc9 | 1575 | */ |
7ccb9ad5 WF |
1576 | t = min(t, 1 + max_pause / 2); |
1577 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
c8462cc9 WF |
1578 | |
1579 | /* | |
5b9b3574 WF |
1580 | * Tiny nr_dirtied_pause is found to hurt I/O performance in the test |
1581 | * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}. | |
1582 | * When the 16 consecutive reads are often interrupted by some dirty | |
1583 | * throttling pause during the async writes, cfq will go into idles | |
1584 | * (deadline is fine). So push nr_dirtied_pause as high as possible | |
1585 | * until reaches DIRTY_POLL_THRESH=32 pages. | |
c8462cc9 | 1586 | */ |
5b9b3574 WF |
1587 | if (pages < DIRTY_POLL_THRESH) { |
1588 | t = max_pause; | |
1589 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
1590 | if (pages > DIRTY_POLL_THRESH) { | |
1591 | pages = DIRTY_POLL_THRESH; | |
1592 | t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit; | |
1593 | } | |
1594 | } | |
1595 | ||
7ccb9ad5 WF |
1596 | pause = HZ * pages / (task_ratelimit + 1); |
1597 | if (pause > max_pause) { | |
1598 | t = max_pause; | |
1599 | pages = task_ratelimit * t / roundup_pow_of_two(HZ); | |
1600 | } | |
c8462cc9 | 1601 | |
7ccb9ad5 | 1602 | *nr_dirtied_pause = pages; |
c8462cc9 | 1603 | /* |
7ccb9ad5 | 1604 | * The minimal pause time will normally be half the target pause time. |
c8462cc9 | 1605 | */ |
5b9b3574 | 1606 | return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t; |
c8462cc9 WF |
1607 | } |
1608 | ||
970fb01a | 1609 | static inline void wb_dirty_limits(struct dirty_throttle_control *dtc) |
5a537485 | 1610 | { |
2bc00aef | 1611 | struct bdi_writeback *wb = dtc->wb; |
93f78d88 | 1612 | unsigned long wb_reclaimable; |
5a537485 MP |
1613 | |
1614 | /* | |
de1fff37 | 1615 | * wb_thresh is not treated as some limiting factor as |
5a537485 | 1616 | * dirty_thresh, due to reasons |
de1fff37 | 1617 | * - in JBOD setup, wb_thresh can fluctuate a lot |
5a537485 | 1618 | * - in a system with HDD and USB key, the USB key may somehow |
de1fff37 TH |
1619 | * go into state (wb_dirty >> wb_thresh) either because |
1620 | * wb_dirty starts high, or because wb_thresh drops low. | |
5a537485 | 1621 | * In this case we don't want to hard throttle the USB key |
de1fff37 TH |
1622 | * dirtiers for 100 seconds until wb_dirty drops under |
1623 | * wb_thresh. Instead the auxiliary wb control line in | |
a88a341a | 1624 | * wb_position_ratio() will let the dirtier task progress |
de1fff37 | 1625 | * at some rate <= (write_bw / 2) for bringing down wb_dirty. |
5a537485 | 1626 | */ |
b1cbc6d4 | 1627 | dtc->wb_thresh = __wb_calc_thresh(dtc); |
970fb01a TH |
1628 | dtc->wb_bg_thresh = dtc->thresh ? |
1629 | div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0; | |
5a537485 MP |
1630 | |
1631 | /* | |
1632 | * In order to avoid the stacked BDI deadlock we need | |
1633 | * to ensure we accurately count the 'dirty' pages when | |
1634 | * the threshold is low. | |
1635 | * | |
1636 | * Otherwise it would be possible to get thresh+n pages | |
1637 | * reported dirty, even though there are thresh-m pages | |
1638 | * actually dirty; with m+n sitting in the percpu | |
1639 | * deltas. | |
1640 | */ | |
2bce774e | 1641 | if (dtc->wb_thresh < 2 * wb_stat_error()) { |
93f78d88 | 1642 | wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); |
2bc00aef | 1643 | dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK); |
5a537485 | 1644 | } else { |
93f78d88 | 1645 | wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE); |
2bc00aef | 1646 | dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK); |
5a537485 MP |
1647 | } |
1648 | } | |
1649 | ||
1da177e4 LT |
1650 | /* |
1651 | * balance_dirty_pages() must be called by processes which are generating dirty | |
1652 | * data. It looks at the number of dirty pages in the machine and will force | |
143dfe86 | 1653 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. |
5b0830cb JA |
1654 | * If we're over `background_thresh' then the writeback threads are woken to |
1655 | * perform some writeout. | |
1da177e4 | 1656 | */ |
fe6c9c6e JK |
1657 | static int balance_dirty_pages(struct bdi_writeback *wb, |
1658 | unsigned long pages_dirtied, unsigned int flags) | |
1da177e4 | 1659 | { |
2bc00aef | 1660 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) }; |
c2aa723a | 1661 | struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) }; |
2bc00aef | 1662 | struct dirty_throttle_control * const gdtc = &gdtc_stor; |
c2aa723a TH |
1663 | struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ? |
1664 | &mdtc_stor : NULL; | |
1665 | struct dirty_throttle_control *sdtc; | |
8d92890b | 1666 | unsigned long nr_reclaimable; /* = file_dirty */ |
83712358 | 1667 | long period; |
7ccb9ad5 WF |
1668 | long pause; |
1669 | long max_pause; | |
1670 | long min_pause; | |
1671 | int nr_dirtied_pause; | |
e50e3720 | 1672 | bool dirty_exceeded = false; |
143dfe86 | 1673 | unsigned long task_ratelimit; |
7ccb9ad5 | 1674 | unsigned long dirty_ratelimit; |
dfb8ae56 | 1675 | struct backing_dev_info *bdi = wb->bdi; |
5a537485 | 1676 | bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT; |
e98be2d5 | 1677 | unsigned long start_time = jiffies; |
fe6c9c6e | 1678 | int ret = 0; |
1da177e4 LT |
1679 | |
1680 | for (;;) { | |
83712358 | 1681 | unsigned long now = jiffies; |
2bc00aef | 1682 | unsigned long dirty, thresh, bg_thresh; |
50e55bf6 YS |
1683 | unsigned long m_dirty = 0; /* stop bogus uninit warnings */ |
1684 | unsigned long m_thresh = 0; | |
1685 | unsigned long m_bg_thresh = 0; | |
83712358 | 1686 | |
8d92890b | 1687 | nr_reclaimable = global_node_page_state(NR_FILE_DIRTY); |
9fc3a43e | 1688 | gdtc->avail = global_dirtyable_memory(); |
11fb9989 | 1689 | gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK); |
5fce25a9 | 1690 | |
9fc3a43e | 1691 | domain_dirty_limits(gdtc); |
16c4042f | 1692 | |
5a537485 | 1693 | if (unlikely(strictlimit)) { |
970fb01a | 1694 | wb_dirty_limits(gdtc); |
5a537485 | 1695 | |
2bc00aef TH |
1696 | dirty = gdtc->wb_dirty; |
1697 | thresh = gdtc->wb_thresh; | |
970fb01a | 1698 | bg_thresh = gdtc->wb_bg_thresh; |
5a537485 | 1699 | } else { |
2bc00aef TH |
1700 | dirty = gdtc->dirty; |
1701 | thresh = gdtc->thresh; | |
1702 | bg_thresh = gdtc->bg_thresh; | |
5a537485 MP |
1703 | } |
1704 | ||
c2aa723a | 1705 | if (mdtc) { |
c5edf9cd | 1706 | unsigned long filepages, headroom, writeback; |
c2aa723a TH |
1707 | |
1708 | /* | |
1709 | * If @wb belongs to !root memcg, repeat the same | |
1710 | * basic calculations for the memcg domain. | |
1711 | */ | |
c5edf9cd TH |
1712 | mem_cgroup_wb_stats(wb, &filepages, &headroom, |
1713 | &mdtc->dirty, &writeback); | |
c2aa723a | 1714 | mdtc->dirty += writeback; |
c5edf9cd | 1715 | mdtc_calc_avail(mdtc, filepages, headroom); |
c2aa723a TH |
1716 | |
1717 | domain_dirty_limits(mdtc); | |
1718 | ||
1719 | if (unlikely(strictlimit)) { | |
1720 | wb_dirty_limits(mdtc); | |
1721 | m_dirty = mdtc->wb_dirty; | |
1722 | m_thresh = mdtc->wb_thresh; | |
1723 | m_bg_thresh = mdtc->wb_bg_thresh; | |
1724 | } else { | |
1725 | m_dirty = mdtc->dirty; | |
1726 | m_thresh = mdtc->thresh; | |
1727 | m_bg_thresh = mdtc->bg_thresh; | |
1728 | } | |
5a537485 MP |
1729 | } |
1730 | ||
ea6813be JK |
1731 | /* |
1732 | * In laptop mode, we wait until hitting the higher threshold | |
1733 | * before starting background writeout, and then write out all | |
1734 | * the way down to the lower threshold. So slow writers cause | |
1735 | * minimal disk activity. | |
1736 | * | |
1737 | * In normal mode, we start background writeout at the lower | |
1738 | * background_thresh, to keep the amount of dirty memory low. | |
1739 | */ | |
1740 | if (!laptop_mode && nr_reclaimable > gdtc->bg_thresh && | |
1741 | !writeback_in_progress(wb)) | |
1742 | wb_start_background_writeback(wb); | |
1743 | ||
16c4042f WF |
1744 | /* |
1745 | * Throttle it only when the background writeback cannot | |
1746 | * catch-up. This avoids (excessively) small writeouts | |
de1fff37 | 1747 | * when the wb limits are ramping up in case of !strictlimit. |
5a537485 | 1748 | * |
de1fff37 TH |
1749 | * In strictlimit case make decision based on the wb counters |
1750 | * and limits. Small writeouts when the wb limits are ramping | |
5a537485 | 1751 | * up are the price we consciously pay for strictlimit-ing. |
c2aa723a TH |
1752 | * |
1753 | * If memcg domain is in effect, @dirty should be under | |
1754 | * both global and memcg freerun ceilings. | |
16c4042f | 1755 | */ |
c2aa723a TH |
1756 | if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) && |
1757 | (!mdtc || | |
1758 | m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) { | |
a37b0715 N |
1759 | unsigned long intv; |
1760 | unsigned long m_intv; | |
1761 | ||
1762 | free_running: | |
1763 | intv = dirty_poll_interval(dirty, thresh); | |
1764 | m_intv = ULONG_MAX; | |
c2aa723a | 1765 | |
83712358 WF |
1766 | current->dirty_paused_when = now; |
1767 | current->nr_dirtied = 0; | |
c2aa723a TH |
1768 | if (mdtc) |
1769 | m_intv = dirty_poll_interval(m_dirty, m_thresh); | |
1770 | current->nr_dirtied_pause = min(intv, m_intv); | |
16c4042f | 1771 | break; |
83712358 | 1772 | } |
16c4042f | 1773 | |
ea6813be | 1774 | /* Start writeback even when in laptop mode */ |
bc05873d | 1775 | if (unlikely(!writeback_in_progress(wb))) |
9ecf4866 | 1776 | wb_start_background_writeback(wb); |
143dfe86 | 1777 | |
97b27821 TH |
1778 | mem_cgroup_flush_foreign(wb); |
1779 | ||
c2aa723a TH |
1780 | /* |
1781 | * Calculate global domain's pos_ratio and select the | |
1782 | * global dtc by default. | |
1783 | */ | |
a37b0715 | 1784 | if (!strictlimit) { |
970fb01a | 1785 | wb_dirty_limits(gdtc); |
5fce25a9 | 1786 | |
a37b0715 N |
1787 | if ((current->flags & PF_LOCAL_THROTTLE) && |
1788 | gdtc->wb_dirty < | |
1789 | dirty_freerun_ceiling(gdtc->wb_thresh, | |
1790 | gdtc->wb_bg_thresh)) | |
1791 | /* | |
1792 | * LOCAL_THROTTLE tasks must not be throttled | |
1793 | * when below the per-wb freerun ceiling. | |
1794 | */ | |
1795 | goto free_running; | |
1796 | } | |
1797 | ||
2bc00aef TH |
1798 | dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) && |
1799 | ((gdtc->dirty > gdtc->thresh) || strictlimit); | |
daddfa3c TH |
1800 | |
1801 | wb_position_ratio(gdtc); | |
c2aa723a TH |
1802 | sdtc = gdtc; |
1803 | ||
1804 | if (mdtc) { | |
1805 | /* | |
1806 | * If memcg domain is in effect, calculate its | |
1807 | * pos_ratio. @wb should satisfy constraints from | |
1808 | * both global and memcg domains. Choose the one | |
1809 | * w/ lower pos_ratio. | |
1810 | */ | |
a37b0715 | 1811 | if (!strictlimit) { |
c2aa723a TH |
1812 | wb_dirty_limits(mdtc); |
1813 | ||
a37b0715 N |
1814 | if ((current->flags & PF_LOCAL_THROTTLE) && |
1815 | mdtc->wb_dirty < | |
1816 | dirty_freerun_ceiling(mdtc->wb_thresh, | |
1817 | mdtc->wb_bg_thresh)) | |
1818 | /* | |
1819 | * LOCAL_THROTTLE tasks must not be | |
1820 | * throttled when below the per-wb | |
1821 | * freerun ceiling. | |
1822 | */ | |
1823 | goto free_running; | |
1824 | } | |
c2aa723a TH |
1825 | dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) && |
1826 | ((mdtc->dirty > mdtc->thresh) || strictlimit); | |
1827 | ||
1828 | wb_position_ratio(mdtc); | |
1829 | if (mdtc->pos_ratio < gdtc->pos_ratio) | |
1830 | sdtc = mdtc; | |
1831 | } | |
daddfa3c | 1832 | |
e92eebbb JK |
1833 | if (dirty_exceeded != wb->dirty_exceeded) |
1834 | wb->dirty_exceeded = dirty_exceeded; | |
1da177e4 | 1835 | |
20792ebf | 1836 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) + |
45a2966f | 1837 | BANDWIDTH_INTERVAL)) |
fee468fd | 1838 | __wb_update_bandwidth(gdtc, mdtc, true); |
e98be2d5 | 1839 | |
c2aa723a | 1840 | /* throttle according to the chosen dtc */ |
20792ebf | 1841 | dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit); |
c2aa723a | 1842 | task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >> |
3a73dbbc | 1843 | RATELIMIT_CALC_SHIFT; |
c2aa723a | 1844 | max_pause = wb_max_pause(wb, sdtc->wb_dirty); |
a88a341a TH |
1845 | min_pause = wb_min_pause(wb, max_pause, |
1846 | task_ratelimit, dirty_ratelimit, | |
1847 | &nr_dirtied_pause); | |
7ccb9ad5 | 1848 | |
3a73dbbc | 1849 | if (unlikely(task_ratelimit == 0)) { |
83712358 | 1850 | period = max_pause; |
c8462cc9 | 1851 | pause = max_pause; |
143dfe86 | 1852 | goto pause; |
04fbfdc1 | 1853 | } |
83712358 WF |
1854 | period = HZ * pages_dirtied / task_ratelimit; |
1855 | pause = period; | |
1856 | if (current->dirty_paused_when) | |
1857 | pause -= now - current->dirty_paused_when; | |
1858 | /* | |
1859 | * For less than 1s think time (ext3/4 may block the dirtier | |
1860 | * for up to 800ms from time to time on 1-HDD; so does xfs, | |
1861 | * however at much less frequency), try to compensate it in | |
1862 | * future periods by updating the virtual time; otherwise just | |
1863 | * do a reset, as it may be a light dirtier. | |
1864 | */ | |
7ccb9ad5 | 1865 | if (pause < min_pause) { |
5634cc2a | 1866 | trace_balance_dirty_pages(wb, |
c2aa723a TH |
1867 | sdtc->thresh, |
1868 | sdtc->bg_thresh, | |
1869 | sdtc->dirty, | |
1870 | sdtc->wb_thresh, | |
1871 | sdtc->wb_dirty, | |
ece13ac3 WF |
1872 | dirty_ratelimit, |
1873 | task_ratelimit, | |
1874 | pages_dirtied, | |
83712358 | 1875 | period, |
7ccb9ad5 | 1876 | min(pause, 0L), |
ece13ac3 | 1877 | start_time); |
83712358 WF |
1878 | if (pause < -HZ) { |
1879 | current->dirty_paused_when = now; | |
1880 | current->nr_dirtied = 0; | |
1881 | } else if (period) { | |
1882 | current->dirty_paused_when += period; | |
1883 | current->nr_dirtied = 0; | |
7ccb9ad5 WF |
1884 | } else if (current->nr_dirtied_pause <= pages_dirtied) |
1885 | current->nr_dirtied_pause += pages_dirtied; | |
57fc978c | 1886 | break; |
04fbfdc1 | 1887 | } |
7ccb9ad5 WF |
1888 | if (unlikely(pause > max_pause)) { |
1889 | /* for occasional dropped task_ratelimit */ | |
1890 | now += min(pause - max_pause, max_pause); | |
1891 | pause = max_pause; | |
1892 | } | |
143dfe86 WF |
1893 | |
1894 | pause: | |
5634cc2a | 1895 | trace_balance_dirty_pages(wb, |
c2aa723a TH |
1896 | sdtc->thresh, |
1897 | sdtc->bg_thresh, | |
1898 | sdtc->dirty, | |
1899 | sdtc->wb_thresh, | |
1900 | sdtc->wb_dirty, | |
ece13ac3 WF |
1901 | dirty_ratelimit, |
1902 | task_ratelimit, | |
1903 | pages_dirtied, | |
83712358 | 1904 | period, |
ece13ac3 WF |
1905 | pause, |
1906 | start_time); | |
fe6c9c6e JK |
1907 | if (flags & BDP_ASYNC) { |
1908 | ret = -EAGAIN; | |
1909 | break; | |
1910 | } | |
499d05ec | 1911 | __set_current_state(TASK_KILLABLE); |
b57d74af | 1912 | wb->dirty_sleep = now; |
d25105e8 | 1913 | io_schedule_timeout(pause); |
87c6a9b2 | 1914 | |
83712358 WF |
1915 | current->dirty_paused_when = now + pause; |
1916 | current->nr_dirtied = 0; | |
7ccb9ad5 | 1917 | current->nr_dirtied_pause = nr_dirtied_pause; |
83712358 | 1918 | |
ffd1f609 | 1919 | /* |
2bc00aef TH |
1920 | * This is typically equal to (dirty < thresh) and can also |
1921 | * keep "1000+ dd on a slow USB stick" under control. | |
ffd1f609 | 1922 | */ |
1df64719 | 1923 | if (task_ratelimit) |
ffd1f609 | 1924 | break; |
499d05ec | 1925 | |
c5c6343c | 1926 | /* |
f0953a1b | 1927 | * In the case of an unresponsive NFS server and the NFS dirty |
de1fff37 | 1928 | * pages exceeds dirty_thresh, give the other good wb's a pipe |
c5c6343c WF |
1929 | * to go through, so that tasks on them still remain responsive. |
1930 | * | |
3f8b6fb7 | 1931 | * In theory 1 page is enough to keep the consumer-producer |
c5c6343c | 1932 | * pipe going: the flusher cleans 1 page => the task dirties 1 |
de1fff37 | 1933 | * more page. However wb_dirty has accounting errors. So use |
93f78d88 | 1934 | * the larger and more IO friendly wb_stat_error. |
c5c6343c | 1935 | */ |
2bce774e | 1936 | if (sdtc->wb_dirty <= wb_stat_error()) |
c5c6343c WF |
1937 | break; |
1938 | ||
499d05ec JK |
1939 | if (fatal_signal_pending(current)) |
1940 | break; | |
1da177e4 | 1941 | } |
fe6c9c6e | 1942 | return ret; |
1da177e4 LT |
1943 | } |
1944 | ||
9d823e8f | 1945 | static DEFINE_PER_CPU(int, bdp_ratelimits); |
245b2e70 | 1946 | |
54848d73 WF |
1947 | /* |
1948 | * Normal tasks are throttled by | |
1949 | * loop { | |
1950 | * dirty tsk->nr_dirtied_pause pages; | |
1951 | * take a snap in balance_dirty_pages(); | |
1952 | * } | |
1953 | * However there is a worst case. If every task exit immediately when dirtied | |
1954 | * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be | |
1955 | * called to throttle the page dirties. The solution is to save the not yet | |
1956 | * throttled page dirties in dirty_throttle_leaks on task exit and charge them | |
1957 | * randomly into the running tasks. This works well for the above worst case, | |
1958 | * as the new task will pick up and accumulate the old task's leaked dirty | |
1959 | * count and eventually get throttled. | |
1960 | */ | |
1961 | DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0; | |
1962 | ||
1da177e4 | 1963 | /** |
fe6c9c6e JK |
1964 | * balance_dirty_pages_ratelimited_flags - Balance dirty memory state. |
1965 | * @mapping: address_space which was dirtied. | |
1966 | * @flags: BDP flags. | |
1da177e4 LT |
1967 | * |
1968 | * Processes which are dirtying memory should call in here once for each page | |
1969 | * which was newly dirtied. The function will periodically check the system's | |
1970 | * dirty state and will initiate writeback if needed. | |
1971 | * | |
fe6c9c6e JK |
1972 | * See balance_dirty_pages_ratelimited() for details. |
1973 | * | |
1974 | * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to | |
1975 | * indicate that memory is out of balance and the caller must wait | |
1976 | * for I/O to complete. Otherwise, it will return 0 to indicate | |
1977 | * that either memory was already in balance, or it was able to sleep | |
1978 | * until the amount of dirty memory returned to balance. | |
1da177e4 | 1979 | */ |
fe6c9c6e JK |
1980 | int balance_dirty_pages_ratelimited_flags(struct address_space *mapping, |
1981 | unsigned int flags) | |
1da177e4 | 1982 | { |
dfb8ae56 TH |
1983 | struct inode *inode = mapping->host; |
1984 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1985 | struct bdi_writeback *wb = NULL; | |
9d823e8f | 1986 | int ratelimit; |
fe6c9c6e | 1987 | int ret = 0; |
9d823e8f | 1988 | int *p; |
1da177e4 | 1989 | |
f56753ac | 1990 | if (!(bdi->capabilities & BDI_CAP_WRITEBACK)) |
fe6c9c6e | 1991 | return ret; |
36715cef | 1992 | |
dfb8ae56 TH |
1993 | if (inode_cgwb_enabled(inode)) |
1994 | wb = wb_get_create_current(bdi, GFP_KERNEL); | |
1995 | if (!wb) | |
1996 | wb = &bdi->wb; | |
1997 | ||
9d823e8f | 1998 | ratelimit = current->nr_dirtied_pause; |
a88a341a | 1999 | if (wb->dirty_exceeded) |
9d823e8f WF |
2000 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); |
2001 | ||
9d823e8f | 2002 | preempt_disable(); |
1da177e4 | 2003 | /* |
9d823e8f WF |
2004 | * This prevents one CPU to accumulate too many dirtied pages without |
2005 | * calling into balance_dirty_pages(), which can happen when there are | |
2006 | * 1000+ tasks, all of them start dirtying pages at exactly the same | |
2007 | * time, hence all honoured too large initial task->nr_dirtied_pause. | |
1da177e4 | 2008 | */ |
7c8e0181 | 2009 | p = this_cpu_ptr(&bdp_ratelimits); |
9d823e8f | 2010 | if (unlikely(current->nr_dirtied >= ratelimit)) |
fa5a734e | 2011 | *p = 0; |
d3bc1fef WF |
2012 | else if (unlikely(*p >= ratelimit_pages)) { |
2013 | *p = 0; | |
2014 | ratelimit = 0; | |
1da177e4 | 2015 | } |
54848d73 WF |
2016 | /* |
2017 | * Pick up the dirtied pages by the exited tasks. This avoids lots of | |
2018 | * short-lived tasks (eg. gcc invocations in a kernel build) escaping | |
2019 | * the dirty throttling and livelock other long-run dirtiers. | |
2020 | */ | |
7c8e0181 | 2021 | p = this_cpu_ptr(&dirty_throttle_leaks); |
54848d73 | 2022 | if (*p > 0 && current->nr_dirtied < ratelimit) { |
d0e1d66b | 2023 | unsigned long nr_pages_dirtied; |
54848d73 WF |
2024 | nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied); |
2025 | *p -= nr_pages_dirtied; | |
2026 | current->nr_dirtied += nr_pages_dirtied; | |
1da177e4 | 2027 | } |
fa5a734e | 2028 | preempt_enable(); |
9d823e8f WF |
2029 | |
2030 | if (unlikely(current->nr_dirtied >= ratelimit)) | |
fe6c9c6e | 2031 | ret = balance_dirty_pages(wb, current->nr_dirtied, flags); |
dfb8ae56 TH |
2032 | |
2033 | wb_put(wb); | |
fe6c9c6e JK |
2034 | return ret; |
2035 | } | |
611df5d6 | 2036 | EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags); |
fe6c9c6e JK |
2037 | |
2038 | /** | |
2039 | * balance_dirty_pages_ratelimited - balance dirty memory state. | |
2040 | * @mapping: address_space which was dirtied. | |
2041 | * | |
2042 | * Processes which are dirtying memory should call in here once for each page | |
2043 | * which was newly dirtied. The function will periodically check the system's | |
2044 | * dirty state and will initiate writeback if needed. | |
2045 | * | |
2046 | * Once we're over the dirty memory limit we decrease the ratelimiting | |
2047 | * by a lot, to prevent individual processes from overshooting the limit | |
2048 | * by (ratelimit_pages) each. | |
2049 | */ | |
2050 | void balance_dirty_pages_ratelimited(struct address_space *mapping) | |
2051 | { | |
2052 | balance_dirty_pages_ratelimited_flags(mapping, 0); | |
1da177e4 | 2053 | } |
d0e1d66b | 2054 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited); |
1da177e4 | 2055 | |
aa661bbe TH |
2056 | /** |
2057 | * wb_over_bg_thresh - does @wb need to be written back? | |
2058 | * @wb: bdi_writeback of interest | |
2059 | * | |
2060 | * Determines whether background writeback should keep writing @wb or it's | |
a862f68a MR |
2061 | * clean enough. |
2062 | * | |
2063 | * Return: %true if writeback should continue. | |
aa661bbe TH |
2064 | */ |
2065 | bool wb_over_bg_thresh(struct bdi_writeback *wb) | |
2066 | { | |
947e9762 | 2067 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) }; |
c2aa723a | 2068 | struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) }; |
947e9762 | 2069 | struct dirty_throttle_control * const gdtc = &gdtc_stor; |
c2aa723a TH |
2070 | struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ? |
2071 | &mdtc_stor : NULL; | |
ab19939a CW |
2072 | unsigned long reclaimable; |
2073 | unsigned long thresh; | |
aa661bbe | 2074 | |
947e9762 TH |
2075 | /* |
2076 | * Similar to balance_dirty_pages() but ignores pages being written | |
2077 | * as we're trying to decide whether to put more under writeback. | |
2078 | */ | |
2079 | gdtc->avail = global_dirtyable_memory(); | |
8d92890b | 2080 | gdtc->dirty = global_node_page_state(NR_FILE_DIRTY); |
947e9762 | 2081 | domain_dirty_limits(gdtc); |
aa661bbe | 2082 | |
947e9762 | 2083 | if (gdtc->dirty > gdtc->bg_thresh) |
aa661bbe TH |
2084 | return true; |
2085 | ||
ab19939a CW |
2086 | thresh = wb_calc_thresh(gdtc->wb, gdtc->bg_thresh); |
2087 | if (thresh < 2 * wb_stat_error()) | |
2088 | reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); | |
2089 | else | |
2090 | reclaimable = wb_stat(wb, WB_RECLAIMABLE); | |
2091 | ||
2092 | if (reclaimable > thresh) | |
aa661bbe TH |
2093 | return true; |
2094 | ||
c2aa723a | 2095 | if (mdtc) { |
c5edf9cd | 2096 | unsigned long filepages, headroom, writeback; |
c2aa723a | 2097 | |
c5edf9cd TH |
2098 | mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty, |
2099 | &writeback); | |
2100 | mdtc_calc_avail(mdtc, filepages, headroom); | |
c2aa723a TH |
2101 | domain_dirty_limits(mdtc); /* ditto, ignore writeback */ |
2102 | ||
2103 | if (mdtc->dirty > mdtc->bg_thresh) | |
2104 | return true; | |
2105 | ||
ab19939a CW |
2106 | thresh = wb_calc_thresh(mdtc->wb, mdtc->bg_thresh); |
2107 | if (thresh < 2 * wb_stat_error()) | |
2108 | reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); | |
2109 | else | |
2110 | reclaimable = wb_stat(wb, WB_RECLAIMABLE); | |
2111 | ||
2112 | if (reclaimable > thresh) | |
c2aa723a TH |
2113 | return true; |
2114 | } | |
2115 | ||
aa661bbe TH |
2116 | return false; |
2117 | } | |
2118 | ||
aa779e51 | 2119 | #ifdef CONFIG_SYSCTL |
1da177e4 LT |
2120 | /* |
2121 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
2122 | */ | |
aa779e51 | 2123 | static int dirty_writeback_centisecs_handler(struct ctl_table *table, int write, |
32927393 | 2124 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 2125 | { |
94af5846 YS |
2126 | unsigned int old_interval = dirty_writeback_interval; |
2127 | int ret; | |
2128 | ||
2129 | ret = proc_dointvec(table, write, buffer, length, ppos); | |
515c24c1 YS |
2130 | |
2131 | /* | |
2132 | * Writing 0 to dirty_writeback_interval will disable periodic writeback | |
2133 | * and a different non-zero value will wakeup the writeback threads. | |
2134 | * wb_wakeup_delayed() would be more appropriate, but it's a pain to | |
2135 | * iterate over all bdis and wbs. | |
2136 | * The reason we do this is to make the change take effect immediately. | |
2137 | */ | |
2138 | if (!ret && write && dirty_writeback_interval && | |
2139 | dirty_writeback_interval != old_interval) | |
94af5846 YS |
2140 | wakeup_flusher_threads(WB_REASON_PERIODIC); |
2141 | ||
2142 | return ret; | |
1da177e4 | 2143 | } |
aa779e51 | 2144 | #endif |
1da177e4 | 2145 | |
bca237a5 | 2146 | void laptop_mode_timer_fn(struct timer_list *t) |
1da177e4 | 2147 | { |
bca237a5 KC |
2148 | struct backing_dev_info *backing_dev_info = |
2149 | from_timer(backing_dev_info, t, laptop_mode_wb_timer); | |
1da177e4 | 2150 | |
bca237a5 | 2151 | wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER); |
1da177e4 LT |
2152 | } |
2153 | ||
2154 | /* | |
2155 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
2156 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
2157 | * then push it back - the user is still using the disk. | |
2158 | */ | |
31373d09 | 2159 | void laptop_io_completion(struct backing_dev_info *info) |
1da177e4 | 2160 | { |
31373d09 | 2161 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
2162 | } |
2163 | ||
2164 | /* | |
2165 | * We're in laptop mode and we've just synced. The sync's writes will have | |
2166 | * caused another writeback to be scheduled by laptop_io_completion. | |
2167 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
2168 | */ | |
2169 | void laptop_sync_completion(void) | |
2170 | { | |
31373d09 MG |
2171 | struct backing_dev_info *bdi; |
2172 | ||
2173 | rcu_read_lock(); | |
2174 | ||
2175 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) | |
2176 | del_timer(&bdi->laptop_mode_wb_timer); | |
2177 | ||
2178 | rcu_read_unlock(); | |
1da177e4 LT |
2179 | } |
2180 | ||
2181 | /* | |
2182 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
2183 | * if a large number of processes all perform writes at the same time. | |
1da177e4 LT |
2184 | * |
2185 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
2186 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
9d823e8f | 2187 | * thresholds. |
1da177e4 LT |
2188 | */ |
2189 | ||
2d1d43f6 | 2190 | void writeback_set_ratelimit(void) |
1da177e4 | 2191 | { |
dcc25ae7 | 2192 | struct wb_domain *dom = &global_wb_domain; |
9d823e8f WF |
2193 | unsigned long background_thresh; |
2194 | unsigned long dirty_thresh; | |
dcc25ae7 | 2195 | |
9d823e8f | 2196 | global_dirty_limits(&background_thresh, &dirty_thresh); |
dcc25ae7 | 2197 | dom->dirty_limit = dirty_thresh; |
9d823e8f | 2198 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); |
1da177e4 LT |
2199 | if (ratelimit_pages < 16) |
2200 | ratelimit_pages = 16; | |
1da177e4 LT |
2201 | } |
2202 | ||
1d7ac6ae | 2203 | static int page_writeback_cpu_online(unsigned int cpu) |
1da177e4 | 2204 | { |
1d7ac6ae SAS |
2205 | writeback_set_ratelimit(); |
2206 | return 0; | |
1da177e4 LT |
2207 | } |
2208 | ||
aa779e51 | 2209 | #ifdef CONFIG_SYSCTL |
3c6a4cba LC |
2210 | |
2211 | /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */ | |
2212 | static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE; | |
2213 | ||
aa779e51 | 2214 | static struct ctl_table vm_page_writeback_sysctls[] = { |
2215 | { | |
2216 | .procname = "dirty_background_ratio", | |
2217 | .data = &dirty_background_ratio, | |
2218 | .maxlen = sizeof(dirty_background_ratio), | |
2219 | .mode = 0644, | |
2220 | .proc_handler = dirty_background_ratio_handler, | |
2221 | .extra1 = SYSCTL_ZERO, | |
2222 | .extra2 = SYSCTL_ONE_HUNDRED, | |
2223 | }, | |
2224 | { | |
2225 | .procname = "dirty_background_bytes", | |
2226 | .data = &dirty_background_bytes, | |
2227 | .maxlen = sizeof(dirty_background_bytes), | |
2228 | .mode = 0644, | |
2229 | .proc_handler = dirty_background_bytes_handler, | |
2230 | .extra1 = SYSCTL_LONG_ONE, | |
2231 | }, | |
2232 | { | |
2233 | .procname = "dirty_ratio", | |
2234 | .data = &vm_dirty_ratio, | |
2235 | .maxlen = sizeof(vm_dirty_ratio), | |
2236 | .mode = 0644, | |
2237 | .proc_handler = dirty_ratio_handler, | |
2238 | .extra1 = SYSCTL_ZERO, | |
2239 | .extra2 = SYSCTL_ONE_HUNDRED, | |
2240 | }, | |
2241 | { | |
2242 | .procname = "dirty_bytes", | |
2243 | .data = &vm_dirty_bytes, | |
2244 | .maxlen = sizeof(vm_dirty_bytes), | |
2245 | .mode = 0644, | |
2246 | .proc_handler = dirty_bytes_handler, | |
2247 | .extra1 = (void *)&dirty_bytes_min, | |
2248 | }, | |
2249 | { | |
2250 | .procname = "dirty_writeback_centisecs", | |
2251 | .data = &dirty_writeback_interval, | |
2252 | .maxlen = sizeof(dirty_writeback_interval), | |
2253 | .mode = 0644, | |
2254 | .proc_handler = dirty_writeback_centisecs_handler, | |
2255 | }, | |
2256 | { | |
2257 | .procname = "dirty_expire_centisecs", | |
2258 | .data = &dirty_expire_interval, | |
2259 | .maxlen = sizeof(dirty_expire_interval), | |
2260 | .mode = 0644, | |
2261 | .proc_handler = proc_dointvec_minmax, | |
2262 | .extra1 = SYSCTL_ZERO, | |
2263 | }, | |
2264 | #ifdef CONFIG_HIGHMEM | |
2265 | { | |
2266 | .procname = "highmem_is_dirtyable", | |
2267 | .data = &vm_highmem_is_dirtyable, | |
2268 | .maxlen = sizeof(vm_highmem_is_dirtyable), | |
2269 | .mode = 0644, | |
2270 | .proc_handler = proc_dointvec_minmax, | |
2271 | .extra1 = SYSCTL_ZERO, | |
2272 | .extra2 = SYSCTL_ONE, | |
2273 | }, | |
2274 | #endif | |
2275 | { | |
2276 | .procname = "laptop_mode", | |
2277 | .data = &laptop_mode, | |
2278 | .maxlen = sizeof(laptop_mode), | |
2279 | .mode = 0644, | |
2280 | .proc_handler = proc_dointvec_jiffies, | |
2281 | }, | |
2282 | {} | |
2283 | }; | |
2284 | #endif | |
2285 | ||
1da177e4 | 2286 | /* |
dc6e29da LT |
2287 | * Called early on to tune the page writeback dirty limits. |
2288 | * | |
2289 | * We used to scale dirty pages according to how total memory | |
0a18e607 | 2290 | * related to pages that could be allocated for buffers. |
dc6e29da LT |
2291 | * |
2292 | * However, that was when we used "dirty_ratio" to scale with | |
2293 | * all memory, and we don't do that any more. "dirty_ratio" | |
0a18e607 | 2294 | * is now applied to total non-HIGHPAGE memory, and as such we can't |
dc6e29da LT |
2295 | * get into the old insane situation any more where we had |
2296 | * large amounts of dirty pages compared to a small amount of | |
2297 | * non-HIGHMEM memory. | |
2298 | * | |
2299 | * But we might still want to scale the dirty_ratio by how | |
2300 | * much memory the box has.. | |
1da177e4 LT |
2301 | */ |
2302 | void __init page_writeback_init(void) | |
2303 | { | |
a50fcb51 RV |
2304 | BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL)); |
2305 | ||
1d7ac6ae SAS |
2306 | cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online", |
2307 | page_writeback_cpu_online, NULL); | |
2308 | cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL, | |
2309 | page_writeback_cpu_online); | |
aa779e51 | 2310 | #ifdef CONFIG_SYSCTL |
2311 | register_sysctl_init("vm", vm_page_writeback_sysctls); | |
2312 | #endif | |
1da177e4 LT |
2313 | } |
2314 | ||
f446daae JK |
2315 | /** |
2316 | * tag_pages_for_writeback - tag pages to be written by write_cache_pages | |
2317 | * @mapping: address space structure to write | |
2318 | * @start: starting page index | |
2319 | * @end: ending page index (inclusive) | |
2320 | * | |
2321 | * This function scans the page range from @start to @end (inclusive) and tags | |
2322 | * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is | |
2323 | * that write_cache_pages (or whoever calls this function) will then use | |
2324 | * TOWRITE tag to identify pages eligible for writeback. This mechanism is | |
2325 | * used to avoid livelocking of writeback by a process steadily creating new | |
2326 | * dirty pages in the file (thus it is important for this function to be quick | |
2327 | * so that it can tag pages faster than a dirtying process can create them). | |
2328 | */ | |
f446daae JK |
2329 | void tag_pages_for_writeback(struct address_space *mapping, |
2330 | pgoff_t start, pgoff_t end) | |
2331 | { | |
ff9c745b MW |
2332 | XA_STATE(xas, &mapping->i_pages, start); |
2333 | unsigned int tagged = 0; | |
2334 | void *page; | |
268f42de | 2335 | |
ff9c745b MW |
2336 | xas_lock_irq(&xas); |
2337 | xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) { | |
2338 | xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE); | |
2339 | if (++tagged % XA_CHECK_SCHED) | |
268f42de | 2340 | continue; |
ff9c745b MW |
2341 | |
2342 | xas_pause(&xas); | |
2343 | xas_unlock_irq(&xas); | |
f446daae | 2344 | cond_resched(); |
ff9c745b | 2345 | xas_lock_irq(&xas); |
268f42de | 2346 | } |
ff9c745b | 2347 | xas_unlock_irq(&xas); |
f446daae JK |
2348 | } |
2349 | EXPORT_SYMBOL(tag_pages_for_writeback); | |
2350 | ||
811d736f | 2351 | /** |
0ea97180 | 2352 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. |
811d736f DH |
2353 | * @mapping: address space structure to write |
2354 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
0ea97180 MS |
2355 | * @writepage: function called for each page |
2356 | * @data: data passed to writepage function | |
811d736f | 2357 | * |
0ea97180 | 2358 | * If a page is already under I/O, write_cache_pages() skips it, even |
811d736f DH |
2359 | * if it's dirty. This is desirable behaviour for memory-cleaning writeback, |
2360 | * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() | |
2361 | * and msync() need to guarantee that all the data which was dirty at the time | |
2362 | * the call was made get new I/O started against them. If wbc->sync_mode is | |
2363 | * WB_SYNC_ALL then we were called for data integrity and we must wait for | |
2364 | * existing IO to complete. | |
f446daae JK |
2365 | * |
2366 | * To avoid livelocks (when other process dirties new pages), we first tag | |
2367 | * pages which should be written back with TOWRITE tag and only then start | |
2368 | * writing them. For data-integrity sync we have to be careful so that we do | |
2369 | * not miss some pages (e.g., because some other process has cleared TOWRITE | |
2370 | * tag we set). The rule we follow is that TOWRITE tag can be cleared only | |
2371 | * by the process clearing the DIRTY tag (and submitting the page for IO). | |
64081362 DC |
2372 | * |
2373 | * To avoid deadlocks between range_cyclic writeback and callers that hold | |
2374 | * pages in PageWriteback to aggregate IO until write_cache_pages() returns, | |
2375 | * we do not loop back to the start of the file. Doing so causes a page | |
2376 | * lock/page writeback access order inversion - we should only ever lock | |
2377 | * multiple pages in ascending page->index order, and looping back to the start | |
2378 | * of the file violates that rule and causes deadlocks. | |
a862f68a MR |
2379 | * |
2380 | * Return: %0 on success, negative error code otherwise | |
811d736f | 2381 | */ |
0ea97180 MS |
2382 | int write_cache_pages(struct address_space *mapping, |
2383 | struct writeback_control *wbc, writepage_t writepage, | |
2384 | void *data) | |
811d736f | 2385 | { |
811d736f DH |
2386 | int ret = 0; |
2387 | int done = 0; | |
3fa750dc | 2388 | int error; |
811d736f DH |
2389 | struct pagevec pvec; |
2390 | int nr_pages; | |
2391 | pgoff_t index; | |
2392 | pgoff_t end; /* Inclusive */ | |
bd19e012 | 2393 | pgoff_t done_index; |
811d736f | 2394 | int range_whole = 0; |
ff9c745b | 2395 | xa_mark_t tag; |
811d736f | 2396 | |
86679820 | 2397 | pagevec_init(&pvec); |
811d736f | 2398 | if (wbc->range_cyclic) { |
28659cc8 | 2399 | index = mapping->writeback_index; /* prev offset */ |
811d736f DH |
2400 | end = -1; |
2401 | } else { | |
09cbfeaf KS |
2402 | index = wbc->range_start >> PAGE_SHIFT; |
2403 | end = wbc->range_end >> PAGE_SHIFT; | |
811d736f DH |
2404 | if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) |
2405 | range_whole = 1; | |
811d736f | 2406 | } |
cc7b8f62 MFO |
2407 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) { |
2408 | tag_pages_for_writeback(mapping, index, end); | |
f446daae | 2409 | tag = PAGECACHE_TAG_TOWRITE; |
cc7b8f62 | 2410 | } else { |
f446daae | 2411 | tag = PAGECACHE_TAG_DIRTY; |
cc7b8f62 | 2412 | } |
bd19e012 | 2413 | done_index = index; |
5a3d5c98 NP |
2414 | while (!done && (index <= end)) { |
2415 | int i; | |
2416 | ||
2b9775ae | 2417 | nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end, |
67fd707f | 2418 | tag); |
5a3d5c98 NP |
2419 | if (nr_pages == 0) |
2420 | break; | |
811d736f | 2421 | |
811d736f DH |
2422 | for (i = 0; i < nr_pages; i++) { |
2423 | struct page *page = pvec.pages[i]; | |
2424 | ||
cf15b07c | 2425 | done_index = page->index; |
d5482cdf | 2426 | |
811d736f DH |
2427 | lock_page(page); |
2428 | ||
5a3d5c98 NP |
2429 | /* |
2430 | * Page truncated or invalidated. We can freely skip it | |
2431 | * then, even for data integrity operations: the page | |
2432 | * has disappeared concurrently, so there could be no | |
f0953a1b | 2433 | * real expectation of this data integrity operation |
5a3d5c98 NP |
2434 | * even if there is now a new, dirty page at the same |
2435 | * pagecache address. | |
2436 | */ | |
811d736f | 2437 | if (unlikely(page->mapping != mapping)) { |
5a3d5c98 | 2438 | continue_unlock: |
811d736f DH |
2439 | unlock_page(page); |
2440 | continue; | |
2441 | } | |
2442 | ||
515f4a03 NP |
2443 | if (!PageDirty(page)) { |
2444 | /* someone wrote it for us */ | |
2445 | goto continue_unlock; | |
2446 | } | |
2447 | ||
2448 | if (PageWriteback(page)) { | |
2449 | if (wbc->sync_mode != WB_SYNC_NONE) | |
2450 | wait_on_page_writeback(page); | |
2451 | else | |
2452 | goto continue_unlock; | |
2453 | } | |
811d736f | 2454 | |
515f4a03 NP |
2455 | BUG_ON(PageWriteback(page)); |
2456 | if (!clear_page_dirty_for_io(page)) | |
5a3d5c98 | 2457 | goto continue_unlock; |
811d736f | 2458 | |
de1414a6 | 2459 | trace_wbc_writepage(wbc, inode_to_bdi(mapping->host)); |
3fa750dc BF |
2460 | error = (*writepage)(page, wbc, data); |
2461 | if (unlikely(error)) { | |
2462 | /* | |
2463 | * Handle errors according to the type of | |
2464 | * writeback. There's no need to continue for | |
2465 | * background writeback. Just push done_index | |
2466 | * past this page so media errors won't choke | |
2467 | * writeout for the entire file. For integrity | |
2468 | * writeback, we must process the entire dirty | |
2469 | * set regardless of errors because the fs may | |
2470 | * still have state to clear for each page. In | |
2471 | * that case we continue processing and return | |
2472 | * the first error. | |
2473 | */ | |
2474 | if (error == AOP_WRITEPAGE_ACTIVATE) { | |
00266770 | 2475 | unlock_page(page); |
3fa750dc BF |
2476 | error = 0; |
2477 | } else if (wbc->sync_mode != WB_SYNC_ALL) { | |
2478 | ret = error; | |
cf15b07c | 2479 | done_index = page->index + 1; |
00266770 NP |
2480 | done = 1; |
2481 | break; | |
2482 | } | |
3fa750dc BF |
2483 | if (!ret) |
2484 | ret = error; | |
0b564927 | 2485 | } |
00266770 | 2486 | |
546a1924 DC |
2487 | /* |
2488 | * We stop writing back only if we are not doing | |
2489 | * integrity sync. In case of integrity sync we have to | |
2490 | * keep going until we have written all the pages | |
2491 | * we tagged for writeback prior to entering this loop. | |
2492 | */ | |
2493 | if (--wbc->nr_to_write <= 0 && | |
2494 | wbc->sync_mode == WB_SYNC_NONE) { | |
2495 | done = 1; | |
2496 | break; | |
05fe478d | 2497 | } |
811d736f DH |
2498 | } |
2499 | pagevec_release(&pvec); | |
2500 | cond_resched(); | |
2501 | } | |
64081362 DC |
2502 | |
2503 | /* | |
2504 | * If we hit the last page and there is more work to be done: wrap | |
2505 | * back the index back to the start of the file for the next | |
2506 | * time we are called. | |
2507 | */ | |
2508 | if (wbc->range_cyclic && !done) | |
2509 | done_index = 0; | |
0b564927 DC |
2510 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) |
2511 | mapping->writeback_index = done_index; | |
06d6cf69 | 2512 | |
811d736f DH |
2513 | return ret; |
2514 | } | |
0ea97180 MS |
2515 | EXPORT_SYMBOL(write_cache_pages); |
2516 | ||
2517 | /* | |
2518 | * Function used by generic_writepages to call the real writepage | |
2519 | * function and set the mapping flags on error | |
2520 | */ | |
2521 | static int __writepage(struct page *page, struct writeback_control *wbc, | |
2522 | void *data) | |
2523 | { | |
2524 | struct address_space *mapping = data; | |
2525 | int ret = mapping->a_ops->writepage(page, wbc); | |
2526 | mapping_set_error(mapping, ret); | |
2527 | return ret; | |
2528 | } | |
2529 | ||
2530 | /** | |
2531 | * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. | |
2532 | * @mapping: address space structure to write | |
2533 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
2534 | * | |
2535 | * This is a library function, which implements the writepages() | |
2536 | * address_space_operation. | |
a862f68a MR |
2537 | * |
2538 | * Return: %0 on success, negative error code otherwise | |
0ea97180 MS |
2539 | */ |
2540 | int generic_writepages(struct address_space *mapping, | |
2541 | struct writeback_control *wbc) | |
2542 | { | |
9b6096a6 SL |
2543 | struct blk_plug plug; |
2544 | int ret; | |
2545 | ||
0ea97180 MS |
2546 | /* deal with chardevs and other special file */ |
2547 | if (!mapping->a_ops->writepage) | |
2548 | return 0; | |
2549 | ||
9b6096a6 SL |
2550 | blk_start_plug(&plug); |
2551 | ret = write_cache_pages(mapping, wbc, __writepage, mapping); | |
2552 | blk_finish_plug(&plug); | |
2553 | return ret; | |
0ea97180 | 2554 | } |
811d736f DH |
2555 | |
2556 | EXPORT_SYMBOL(generic_writepages); | |
2557 | ||
1da177e4 LT |
2558 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
2559 | { | |
22905f77 | 2560 | int ret; |
fee468fd | 2561 | struct bdi_writeback *wb; |
22905f77 | 2562 | |
1da177e4 LT |
2563 | if (wbc->nr_to_write <= 0) |
2564 | return 0; | |
fee468fd JK |
2565 | wb = inode_to_wb_wbc(mapping->host, wbc); |
2566 | wb_bandwidth_estimate_start(wb); | |
80a2ea9f TT |
2567 | while (1) { |
2568 | if (mapping->a_ops->writepages) | |
2569 | ret = mapping->a_ops->writepages(mapping, wbc); | |
2570 | else | |
2571 | ret = generic_writepages(mapping, wbc); | |
2572 | if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL)) | |
2573 | break; | |
8d58802f MG |
2574 | |
2575 | /* | |
2576 | * Lacking an allocation context or the locality or writeback | |
2577 | * state of any of the inode's pages, throttle based on | |
2578 | * writeback activity on the local node. It's as good a | |
2579 | * guess as any. | |
2580 | */ | |
2581 | reclaim_throttle(NODE_DATA(numa_node_id()), | |
c3f4a9a2 | 2582 | VMSCAN_THROTTLE_WRITEBACK); |
80a2ea9f | 2583 | } |
45a2966f JK |
2584 | /* |
2585 | * Usually few pages are written by now from those we've just submitted | |
2586 | * but if there's constant writeback being submitted, this makes sure | |
2587 | * writeback bandwidth is updated once in a while. | |
2588 | */ | |
20792ebf JK |
2589 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) + |
2590 | BANDWIDTH_INTERVAL)) | |
45a2966f | 2591 | wb_update_bandwidth(wb); |
22905f77 | 2592 | return ret; |
1da177e4 LT |
2593 | } |
2594 | ||
2595 | /** | |
121703c1 MWO |
2596 | * folio_write_one - write out a single folio and wait on I/O. |
2597 | * @folio: The folio to write. | |
1da177e4 | 2598 | * |
121703c1 | 2599 | * The folio must be locked by the caller and will be unlocked upon return. |
1da177e4 | 2600 | * |
37e51a76 JL |
2601 | * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this |
2602 | * function returns. | |
a862f68a MR |
2603 | * |
2604 | * Return: %0 on success, negative error code otherwise | |
1da177e4 | 2605 | */ |
121703c1 | 2606 | int folio_write_one(struct folio *folio) |
1da177e4 | 2607 | { |
121703c1 | 2608 | struct address_space *mapping = folio->mapping; |
1da177e4 LT |
2609 | int ret = 0; |
2610 | struct writeback_control wbc = { | |
2611 | .sync_mode = WB_SYNC_ALL, | |
121703c1 | 2612 | .nr_to_write = folio_nr_pages(folio), |
1da177e4 LT |
2613 | }; |
2614 | ||
121703c1 | 2615 | BUG_ON(!folio_test_locked(folio)); |
1da177e4 | 2616 | |
121703c1 | 2617 | folio_wait_writeback(folio); |
1da177e4 | 2618 | |
121703c1 MWO |
2619 | if (folio_clear_dirty_for_io(folio)) { |
2620 | folio_get(folio); | |
2621 | ret = mapping->a_ops->writepage(&folio->page, &wbc); | |
37e51a76 | 2622 | if (ret == 0) |
121703c1 MWO |
2623 | folio_wait_writeback(folio); |
2624 | folio_put(folio); | |
1da177e4 | 2625 | } else { |
121703c1 | 2626 | folio_unlock(folio); |
1da177e4 | 2627 | } |
37e51a76 JL |
2628 | |
2629 | if (!ret) | |
2630 | ret = filemap_check_errors(mapping); | |
1da177e4 LT |
2631 | return ret; |
2632 | } | |
121703c1 | 2633 | EXPORT_SYMBOL(folio_write_one); |
1da177e4 | 2634 | |
76719325 KC |
2635 | /* |
2636 | * For address_spaces which do not use buffers nor write back. | |
2637 | */ | |
46de8b97 | 2638 | bool noop_dirty_folio(struct address_space *mapping, struct folio *folio) |
76719325 | 2639 | { |
46de8b97 MWO |
2640 | if (!folio_test_dirty(folio)) |
2641 | return !folio_test_set_dirty(folio); | |
2642 | return false; | |
76719325 | 2643 | } |
46de8b97 | 2644 | EXPORT_SYMBOL(noop_dirty_folio); |
76719325 | 2645 | |
e3a7cca1 ES |
2646 | /* |
2647 | * Helper function for set_page_dirty family. | |
c4843a75 | 2648 | * |
81f8c3a4 | 2649 | * Caller must hold lock_page_memcg(). |
c4843a75 | 2650 | * |
e3a7cca1 ES |
2651 | * NOTE: This relies on being atomic wrt interrupts. |
2652 | */ | |
203a3151 | 2653 | static void folio_account_dirtied(struct folio *folio, |
6e1cae88 | 2654 | struct address_space *mapping) |
e3a7cca1 | 2655 | { |
52ebea74 TH |
2656 | struct inode *inode = mapping->host; |
2657 | ||
b9b0ff61 | 2658 | trace_writeback_dirty_folio(folio, mapping); |
9fb0a7da | 2659 | |
f56753ac | 2660 | if (mapping_can_writeback(mapping)) { |
52ebea74 | 2661 | struct bdi_writeback *wb; |
203a3151 | 2662 | long nr = folio_nr_pages(folio); |
de1414a6 | 2663 | |
203a3151 | 2664 | inode_attach_wb(inode, &folio->page); |
52ebea74 | 2665 | wb = inode_to_wb(inode); |
de1414a6 | 2666 | |
203a3151 MWO |
2667 | __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr); |
2668 | __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr); | |
2669 | __node_stat_mod_folio(folio, NR_DIRTIED, nr); | |
2670 | wb_stat_mod(wb, WB_RECLAIMABLE, nr); | |
2671 | wb_stat_mod(wb, WB_DIRTIED, nr); | |
2672 | task_io_account_write(nr * PAGE_SIZE); | |
2673 | current->nr_dirtied += nr; | |
2674 | __this_cpu_add(bdp_ratelimits, nr); | |
97b27821 | 2675 | |
203a3151 | 2676 | mem_cgroup_track_foreign_dirty(folio, wb); |
e3a7cca1 ES |
2677 | } |
2678 | } | |
2679 | ||
b9ea2515 KK |
2680 | /* |
2681 | * Helper function for deaccounting dirty page without writeback. | |
2682 | * | |
81f8c3a4 | 2683 | * Caller must hold lock_page_memcg(). |
b9ea2515 | 2684 | */ |
566d3362 | 2685 | void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb) |
b9ea2515 | 2686 | { |
566d3362 HD |
2687 | long nr = folio_nr_pages(folio); |
2688 | ||
2689 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr); | |
2690 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); | |
2691 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr); | |
2692 | task_io_account_cancelled_write(nr * PAGE_SIZE); | |
b9ea2515 | 2693 | } |
b9ea2515 | 2694 | |
6e1cae88 | 2695 | /* |
203a3151 MWO |
2696 | * Mark the folio dirty, and set it dirty in the page cache, and mark |
2697 | * the inode dirty. | |
6e1cae88 | 2698 | * |
203a3151 | 2699 | * If warn is true, then emit a warning if the folio is not uptodate and has |
6e1cae88 MWO |
2700 | * not been truncated. |
2701 | * | |
a229a4f0 MWO |
2702 | * The caller must hold lock_page_memcg(). Most callers have the folio |
2703 | * locked. A few have the folio blocked from truncation through other | |
2704 | * means (eg zap_page_range() has it mapped and is holding the page table | |
2705 | * lock). This can also be called from mark_buffer_dirty(), which I | |
2706 | * cannot prove is always protected against truncate. | |
6e1cae88 | 2707 | */ |
203a3151 | 2708 | void __folio_mark_dirty(struct folio *folio, struct address_space *mapping, |
6e1cae88 MWO |
2709 | int warn) |
2710 | { | |
2711 | unsigned long flags; | |
2712 | ||
2713 | xa_lock_irqsave(&mapping->i_pages, flags); | |
203a3151 MWO |
2714 | if (folio->mapping) { /* Race with truncate? */ |
2715 | WARN_ON_ONCE(warn && !folio_test_uptodate(folio)); | |
2716 | folio_account_dirtied(folio, mapping); | |
2717 | __xa_set_mark(&mapping->i_pages, folio_index(folio), | |
6e1cae88 MWO |
2718 | PAGECACHE_TAG_DIRTY); |
2719 | } | |
2720 | xa_unlock_irqrestore(&mapping->i_pages, flags); | |
2721 | } | |
2722 | ||
85d4d2eb MWO |
2723 | /** |
2724 | * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads. | |
2725 | * @mapping: Address space this folio belongs to. | |
2726 | * @folio: Folio to be marked as dirty. | |
1da177e4 | 2727 | * |
85d4d2eb MWO |
2728 | * Filesystems which do not use buffer heads should call this function |
2729 | * from their set_page_dirty address space operation. It ignores the | |
2730 | * contents of folio_get_private(), so if the filesystem marks individual | |
2731 | * blocks as dirty, the filesystem should handle that itself. | |
1da177e4 | 2732 | * |
85d4d2eb MWO |
2733 | * This is also sometimes used by filesystems which use buffer_heads when |
2734 | * a single buffer is being dirtied: we want to set the folio dirty in | |
2735 | * that case, but not all the buffers. This is a "bottom-up" dirtying, | |
e621900a | 2736 | * whereas block_dirty_folio() is a "top-down" dirtying. |
85d4d2eb MWO |
2737 | * |
2738 | * The caller must ensure this doesn't race with truncation. Most will | |
2739 | * simply hold the folio lock, but e.g. zap_pte_range() calls with the | |
2740 | * folio mapped and the pte lock held, which also locks out truncation. | |
1da177e4 | 2741 | */ |
85d4d2eb | 2742 | bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio) |
1da177e4 | 2743 | { |
85d4d2eb MWO |
2744 | folio_memcg_lock(folio); |
2745 | if (folio_test_set_dirty(folio)) { | |
2746 | folio_memcg_unlock(folio); | |
2747 | return false; | |
2748 | } | |
1da177e4 | 2749 | |
85d4d2eb MWO |
2750 | __folio_mark_dirty(folio, mapping, !folio_test_private(folio)); |
2751 | folio_memcg_unlock(folio); | |
c4843a75 | 2752 | |
85d4d2eb MWO |
2753 | if (mapping->host) { |
2754 | /* !PageAnon && !swapper_space */ | |
2755 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 2756 | } |
85d4d2eb | 2757 | return true; |
1da177e4 | 2758 | } |
85d4d2eb | 2759 | EXPORT_SYMBOL(filemap_dirty_folio); |
1da177e4 | 2760 | |
25ff8b15 MWO |
2761 | /** |
2762 | * folio_account_redirty - Manually account for redirtying a page. | |
2763 | * @folio: The folio which is being redirtied. | |
2764 | * | |
2765 | * Most filesystems should call folio_redirty_for_writepage() instead | |
2766 | * of this fuction. If your filesystem is doing writeback outside the | |
2767 | * context of a writeback_control(), it can call this when redirtying | |
2768 | * a folio, to de-account the dirty counters (NR_DIRTIED, WB_DIRTIED, | |
2769 | * tsk->nr_dirtied), so that they match the written counters (NR_WRITTEN, | |
2770 | * WB_WRITTEN) in long term. The mismatches will lead to systematic errors | |
2771 | * in balanced_dirty_ratelimit and the dirty pages position control. | |
2f800fbd | 2772 | */ |
25ff8b15 | 2773 | void folio_account_redirty(struct folio *folio) |
2f800fbd | 2774 | { |
25ff8b15 | 2775 | struct address_space *mapping = folio->mapping; |
91018134 | 2776 | |
f56753ac | 2777 | if (mapping && mapping_can_writeback(mapping)) { |
682aa8e1 TH |
2778 | struct inode *inode = mapping->host; |
2779 | struct bdi_writeback *wb; | |
2e898e4c | 2780 | struct wb_lock_cookie cookie = {}; |
25ff8b15 | 2781 | long nr = folio_nr_pages(folio); |
91018134 | 2782 | |
2e898e4c | 2783 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
25ff8b15 MWO |
2784 | current->nr_dirtied -= nr; |
2785 | node_stat_mod_folio(folio, NR_DIRTIED, -nr); | |
2786 | wb_stat_mod(wb, WB_DIRTIED, -nr); | |
2e898e4c | 2787 | unlocked_inode_to_wb_end(inode, &cookie); |
2f800fbd WF |
2788 | } |
2789 | } | |
25ff8b15 | 2790 | EXPORT_SYMBOL(folio_account_redirty); |
2f800fbd | 2791 | |
cd78ab11 MWO |
2792 | /** |
2793 | * folio_redirty_for_writepage - Decline to write a dirty folio. | |
2794 | * @wbc: The writeback control. | |
2795 | * @folio: The folio. | |
2796 | * | |
2797 | * When a writepage implementation decides that it doesn't want to write | |
2798 | * @folio for some reason, it should call this function, unlock @folio and | |
2799 | * return 0. | |
2800 | * | |
2801 | * Return: True if we redirtied the folio. False if someone else dirtied | |
2802 | * it first. | |
1da177e4 | 2803 | */ |
cd78ab11 MWO |
2804 | bool folio_redirty_for_writepage(struct writeback_control *wbc, |
2805 | struct folio *folio) | |
1da177e4 | 2806 | { |
cd78ab11 MWO |
2807 | bool ret; |
2808 | long nr = folio_nr_pages(folio); | |
2809 | ||
2810 | wbc->pages_skipped += nr; | |
2811 | ret = filemap_dirty_folio(folio->mapping, folio); | |
2812 | folio_account_redirty(folio); | |
8d38633c | 2813 | |
8d38633c | 2814 | return ret; |
1da177e4 | 2815 | } |
cd78ab11 | 2816 | EXPORT_SYMBOL(folio_redirty_for_writepage); |
1da177e4 | 2817 | |
b5e84594 MWO |
2818 | /** |
2819 | * folio_mark_dirty - Mark a folio as being modified. | |
2820 | * @folio: The folio. | |
6746aff7 | 2821 | * |
2ca456c2 MWO |
2822 | * The folio may not be truncated while this function is running. |
2823 | * Holding the folio lock is sufficient to prevent truncation, but some | |
2824 | * callers cannot acquire a sleeping lock. These callers instead hold | |
2825 | * the page table lock for a page table which contains at least one page | |
2826 | * in this folio. Truncation will block on the page table lock as it | |
2827 | * unmaps pages before removing the folio from its mapping. | |
b5e84594 MWO |
2828 | * |
2829 | * Return: True if the folio was newly dirtied, false if it was already dirty. | |
1da177e4 | 2830 | */ |
b5e84594 | 2831 | bool folio_mark_dirty(struct folio *folio) |
1da177e4 | 2832 | { |
b5e84594 | 2833 | struct address_space *mapping = folio_mapping(folio); |
1da177e4 LT |
2834 | |
2835 | if (likely(mapping)) { | |
278df9f4 MK |
2836 | /* |
2837 | * readahead/lru_deactivate_page could remain | |
6f31a5a2 MWO |
2838 | * PG_readahead/PG_reclaim due to race with folio_end_writeback |
2839 | * About readahead, if the folio is written, the flags would be | |
278df9f4 | 2840 | * reset. So no problem. |
6f31a5a2 MWO |
2841 | * About lru_deactivate_page, if the folio is redirtied, |
2842 | * the flag will be reset. So no problem. but if the | |
2843 | * folio is used by readahead it will confuse readahead | |
2844 | * and make it restart the size rampup process. But it's | |
2845 | * a trivial problem. | |
278df9f4 | 2846 | */ |
b5e84594 MWO |
2847 | if (folio_test_reclaim(folio)) |
2848 | folio_clear_reclaim(folio); | |
3a3bae50 | 2849 | return mapping->a_ops->dirty_folio(mapping, folio); |
4741c9fd | 2850 | } |
3a3bae50 MWO |
2851 | |
2852 | return noop_dirty_folio(mapping, folio); | |
1da177e4 | 2853 | } |
b5e84594 | 2854 | EXPORT_SYMBOL(folio_mark_dirty); |
1da177e4 LT |
2855 | |
2856 | /* | |
2857 | * set_page_dirty() is racy if the caller has no reference against | |
2858 | * page->mapping->host, and if the page is unlocked. This is because another | |
2859 | * CPU could truncate the page off the mapping and then free the mapping. | |
2860 | * | |
2861 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
2862 | * holds a reference on the inode by having an open file. | |
2863 | * | |
2864 | * In other cases, the page should be locked before running set_page_dirty(). | |
2865 | */ | |
2866 | int set_page_dirty_lock(struct page *page) | |
2867 | { | |
2868 | int ret; | |
2869 | ||
7eaceacc | 2870 | lock_page(page); |
1da177e4 LT |
2871 | ret = set_page_dirty(page); |
2872 | unlock_page(page); | |
2873 | return ret; | |
2874 | } | |
2875 | EXPORT_SYMBOL(set_page_dirty_lock); | |
2876 | ||
11f81bec TH |
2877 | /* |
2878 | * This cancels just the dirty bit on the kernel page itself, it does NOT | |
2879 | * actually remove dirty bits on any mmap's that may be around. It also | |
2880 | * leaves the page tagged dirty, so any sync activity will still find it on | |
2881 | * the dirty lists, and in particular, clear_page_dirty_for_io() will still | |
2882 | * look at the dirty bits in the VM. | |
2883 | * | |
2884 | * Doing this should *normally* only ever be done when a page is truncated, | |
2885 | * and is not actually mapped anywhere at all. However, fs/buffer.c does | |
2886 | * this when it notices that somebody has cleaned out all the buffers on a | |
2887 | * page without actually doing it through the VM. Can you say "ext3 is | |
2888 | * horribly ugly"? Thought you could. | |
2889 | */ | |
fdaf532a | 2890 | void __folio_cancel_dirty(struct folio *folio) |
11f81bec | 2891 | { |
fdaf532a | 2892 | struct address_space *mapping = folio_mapping(folio); |
c4843a75 | 2893 | |
f56753ac | 2894 | if (mapping_can_writeback(mapping)) { |
682aa8e1 TH |
2895 | struct inode *inode = mapping->host; |
2896 | struct bdi_writeback *wb; | |
2e898e4c | 2897 | struct wb_lock_cookie cookie = {}; |
c4843a75 | 2898 | |
fdaf532a | 2899 | folio_memcg_lock(folio); |
2e898e4c | 2900 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
c4843a75 | 2901 | |
fdaf532a | 2902 | if (folio_test_clear_dirty(folio)) |
566d3362 | 2903 | folio_account_cleaned(folio, wb); |
c4843a75 | 2904 | |
2e898e4c | 2905 | unlocked_inode_to_wb_end(inode, &cookie); |
fdaf532a | 2906 | folio_memcg_unlock(folio); |
c4843a75 | 2907 | } else { |
fdaf532a | 2908 | folio_clear_dirty(folio); |
c4843a75 | 2909 | } |
11f81bec | 2910 | } |
fdaf532a | 2911 | EXPORT_SYMBOL(__folio_cancel_dirty); |
11f81bec | 2912 | |
1da177e4 | 2913 | /* |
9350f20a MWO |
2914 | * Clear a folio's dirty flag, while caring for dirty memory accounting. |
2915 | * Returns true if the folio was previously dirty. | |
1da177e4 | 2916 | * |
9350f20a MWO |
2917 | * This is for preparing to put the folio under writeout. We leave |
2918 | * the folio tagged as dirty in the xarray so that a concurrent | |
2919 | * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk. | |
2920 | * The ->writepage implementation will run either folio_start_writeback() | |
2921 | * or folio_mark_dirty(), at which stage we bring the folio's dirty flag | |
2922 | * and xarray dirty tag back into sync. | |
1da177e4 | 2923 | * |
9350f20a MWO |
2924 | * This incoherency between the folio's dirty flag and xarray tag is |
2925 | * unfortunate, but it only exists while the folio is locked. | |
1da177e4 | 2926 | */ |
9350f20a | 2927 | bool folio_clear_dirty_for_io(struct folio *folio) |
1da177e4 | 2928 | { |
9350f20a MWO |
2929 | struct address_space *mapping = folio_mapping(folio); |
2930 | bool ret = false; | |
1da177e4 | 2931 | |
9350f20a | 2932 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
79352894 | 2933 | |
f56753ac | 2934 | if (mapping && mapping_can_writeback(mapping)) { |
682aa8e1 TH |
2935 | struct inode *inode = mapping->host; |
2936 | struct bdi_writeback *wb; | |
2e898e4c | 2937 | struct wb_lock_cookie cookie = {}; |
682aa8e1 | 2938 | |
7658cc28 LT |
2939 | /* |
2940 | * Yes, Virginia, this is indeed insane. | |
2941 | * | |
2942 | * We use this sequence to make sure that | |
2943 | * (a) we account for dirty stats properly | |
2944 | * (b) we tell the low-level filesystem to | |
9350f20a | 2945 | * mark the whole folio dirty if it was |
7658cc28 | 2946 | * dirty in a pagetable. Only to then |
9350f20a | 2947 | * (c) clean the folio again and return 1 to |
7658cc28 LT |
2948 | * cause the writeback. |
2949 | * | |
2950 | * This way we avoid all nasty races with the | |
2951 | * dirty bit in multiple places and clearing | |
2952 | * them concurrently from different threads. | |
2953 | * | |
9350f20a | 2954 | * Note! Normally the "folio_mark_dirty(folio)" |
7658cc28 LT |
2955 | * has no effect on the actual dirty bit - since |
2956 | * that will already usually be set. But we | |
2957 | * need the side effects, and it can help us | |
2958 | * avoid races. | |
2959 | * | |
9350f20a | 2960 | * We basically use the folio "master dirty bit" |
7658cc28 LT |
2961 | * as a serialization point for all the different |
2962 | * threads doing their things. | |
7658cc28 | 2963 | */ |
9350f20a MWO |
2964 | if (folio_mkclean(folio)) |
2965 | folio_mark_dirty(folio); | |
79352894 NP |
2966 | /* |
2967 | * We carefully synchronise fault handlers against | |
9350f20a | 2968 | * installing a dirty pte and marking the folio dirty |
2d6d7f98 | 2969 | * at this point. We do this by having them hold the |
9350f20a | 2970 | * page lock while dirtying the folio, and folios are |
2d6d7f98 JW |
2971 | * always locked coming in here, so we get the desired |
2972 | * exclusion. | |
79352894 | 2973 | */ |
2e898e4c | 2974 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
9350f20a MWO |
2975 | if (folio_test_clear_dirty(folio)) { |
2976 | long nr = folio_nr_pages(folio); | |
2977 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr); | |
2978 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); | |
2979 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr); | |
2980 | ret = true; | |
1da177e4 | 2981 | } |
2e898e4c | 2982 | unlocked_inode_to_wb_end(inode, &cookie); |
c4843a75 | 2983 | return ret; |
1da177e4 | 2984 | } |
9350f20a | 2985 | return folio_test_clear_dirty(folio); |
1da177e4 | 2986 | } |
9350f20a | 2987 | EXPORT_SYMBOL(folio_clear_dirty_for_io); |
1da177e4 | 2988 | |
633a2abb JK |
2989 | static void wb_inode_writeback_start(struct bdi_writeback *wb) |
2990 | { | |
2991 | atomic_inc(&wb->writeback_inodes); | |
2992 | } | |
2993 | ||
2994 | static void wb_inode_writeback_end(struct bdi_writeback *wb) | |
2995 | { | |
f87904c0 | 2996 | unsigned long flags; |
633a2abb | 2997 | atomic_dec(&wb->writeback_inodes); |
45a2966f JK |
2998 | /* |
2999 | * Make sure estimate of writeback throughput gets updated after | |
3000 | * writeback completed. We delay the update by BANDWIDTH_INTERVAL | |
3001 | * (which is the interval other bandwidth updates use for batching) so | |
3002 | * that if multiple inodes end writeback at a similar time, they get | |
3003 | * batched into one bandwidth update. | |
3004 | */ | |
f87904c0 KK |
3005 | spin_lock_irqsave(&wb->work_lock, flags); |
3006 | if (test_bit(WB_registered, &wb->state)) | |
3007 | queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL); | |
3008 | spin_unlock_irqrestore(&wb->work_lock, flags); | |
633a2abb JK |
3009 | } |
3010 | ||
269ccca3 | 3011 | bool __folio_end_writeback(struct folio *folio) |
1da177e4 | 3012 | { |
269ccca3 MWO |
3013 | long nr = folio_nr_pages(folio); |
3014 | struct address_space *mapping = folio_mapping(folio); | |
3015 | bool ret; | |
1da177e4 | 3016 | |
269ccca3 | 3017 | folio_memcg_lock(folio); |
371a096e | 3018 | if (mapping && mapping_use_writeback_tags(mapping)) { |
91018134 TH |
3019 | struct inode *inode = mapping->host; |
3020 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 LT |
3021 | unsigned long flags; |
3022 | ||
b93b0163 | 3023 | xa_lock_irqsave(&mapping->i_pages, flags); |
269ccca3 | 3024 | ret = folio_test_clear_writeback(folio); |
69cb51d1 | 3025 | if (ret) { |
269ccca3 | 3026 | __xa_clear_mark(&mapping->i_pages, folio_index(folio), |
1da177e4 | 3027 | PAGECACHE_TAG_WRITEBACK); |
823423ef | 3028 | if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) { |
91018134 TH |
3029 | struct bdi_writeback *wb = inode_to_wb(inode); |
3030 | ||
269ccca3 MWO |
3031 | wb_stat_mod(wb, WB_WRITEBACK, -nr); |
3032 | __wb_writeout_add(wb, nr); | |
633a2abb JK |
3033 | if (!mapping_tagged(mapping, |
3034 | PAGECACHE_TAG_WRITEBACK)) | |
3035 | wb_inode_writeback_end(wb); | |
04fbfdc1 | 3036 | } |
69cb51d1 | 3037 | } |
6c60d2b5 DC |
3038 | |
3039 | if (mapping->host && !mapping_tagged(mapping, | |
3040 | PAGECACHE_TAG_WRITEBACK)) | |
3041 | sb_clear_inode_writeback(mapping->host); | |
3042 | ||
b93b0163 | 3043 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
1da177e4 | 3044 | } else { |
269ccca3 | 3045 | ret = folio_test_clear_writeback(folio); |
1da177e4 | 3046 | } |
99b12e3d | 3047 | if (ret) { |
269ccca3 MWO |
3048 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr); |
3049 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); | |
3050 | node_stat_mod_folio(folio, NR_WRITTEN, nr); | |
99b12e3d | 3051 | } |
269ccca3 | 3052 | folio_memcg_unlock(folio); |
1da177e4 LT |
3053 | return ret; |
3054 | } | |
3055 | ||
f143f1ea | 3056 | bool __folio_start_writeback(struct folio *folio, bool keep_write) |
1da177e4 | 3057 | { |
f143f1ea MWO |
3058 | long nr = folio_nr_pages(folio); |
3059 | struct address_space *mapping = folio_mapping(folio); | |
3060 | bool ret; | |
3061 | int access_ret; | |
1da177e4 | 3062 | |
f143f1ea | 3063 | folio_memcg_lock(folio); |
371a096e | 3064 | if (mapping && mapping_use_writeback_tags(mapping)) { |
f143f1ea | 3065 | XA_STATE(xas, &mapping->i_pages, folio_index(folio)); |
91018134 TH |
3066 | struct inode *inode = mapping->host; |
3067 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 LT |
3068 | unsigned long flags; |
3069 | ||
ff9c745b MW |
3070 | xas_lock_irqsave(&xas, flags); |
3071 | xas_load(&xas); | |
f143f1ea | 3072 | ret = folio_test_set_writeback(folio); |
69cb51d1 | 3073 | if (!ret) { |
6c60d2b5 DC |
3074 | bool on_wblist; |
3075 | ||
3076 | on_wblist = mapping_tagged(mapping, | |
3077 | PAGECACHE_TAG_WRITEBACK); | |
3078 | ||
ff9c745b | 3079 | xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK); |
633a2abb JK |
3080 | if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) { |
3081 | struct bdi_writeback *wb = inode_to_wb(inode); | |
3082 | ||
f143f1ea | 3083 | wb_stat_mod(wb, WB_WRITEBACK, nr); |
633a2abb JK |
3084 | if (!on_wblist) |
3085 | wb_inode_writeback_start(wb); | |
3086 | } | |
6c60d2b5 DC |
3087 | |
3088 | /* | |
f143f1ea MWO |
3089 | * We can come through here when swapping |
3090 | * anonymous folios, so we don't necessarily | |
3091 | * have an inode to track for sync. | |
6c60d2b5 DC |
3092 | */ |
3093 | if (mapping->host && !on_wblist) | |
3094 | sb_mark_inode_writeback(mapping->host); | |
69cb51d1 | 3095 | } |
f143f1ea | 3096 | if (!folio_test_dirty(folio)) |
ff9c745b | 3097 | xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY); |
1c8349a1 | 3098 | if (!keep_write) |
ff9c745b MW |
3099 | xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE); |
3100 | xas_unlock_irqrestore(&xas, flags); | |
1da177e4 | 3101 | } else { |
f143f1ea | 3102 | ret = folio_test_set_writeback(folio); |
1da177e4 | 3103 | } |
3a3c02ec | 3104 | if (!ret) { |
f143f1ea MWO |
3105 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr); |
3106 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr); | |
3a3c02ec | 3107 | } |
f143f1ea MWO |
3108 | folio_memcg_unlock(folio); |
3109 | access_ret = arch_make_folio_accessible(folio); | |
f28d4363 CI |
3110 | /* |
3111 | * If writeback has been triggered on a page that cannot be made | |
3112 | * accessible, it is too late to recover here. | |
3113 | */ | |
f143f1ea | 3114 | VM_BUG_ON_FOLIO(access_ret != 0, folio); |
f28d4363 | 3115 | |
1da177e4 | 3116 | return ret; |
1da177e4 | 3117 | } |
f143f1ea | 3118 | EXPORT_SYMBOL(__folio_start_writeback); |
1da177e4 | 3119 | |
490e016f MWO |
3120 | /** |
3121 | * folio_wait_writeback - Wait for a folio to finish writeback. | |
3122 | * @folio: The folio to wait for. | |
3123 | * | |
3124 | * If the folio is currently being written back to storage, wait for the | |
3125 | * I/O to complete. | |
3126 | * | |
3127 | * Context: Sleeps. Must be called in process context and with | |
3128 | * no spinlocks held. Caller should hold a reference on the folio. | |
3129 | * If the folio is not locked, writeback may start again after writeback | |
3130 | * has finished. | |
19343b5b | 3131 | */ |
490e016f | 3132 | void folio_wait_writeback(struct folio *folio) |
19343b5b | 3133 | { |
490e016f | 3134 | while (folio_test_writeback(folio)) { |
b9b0ff61 | 3135 | trace_folio_wait_writeback(folio, folio_mapping(folio)); |
101c0bf6 | 3136 | folio_wait_bit(folio, PG_writeback); |
19343b5b YS |
3137 | } |
3138 | } | |
490e016f | 3139 | EXPORT_SYMBOL_GPL(folio_wait_writeback); |
19343b5b | 3140 | |
490e016f MWO |
3141 | /** |
3142 | * folio_wait_writeback_killable - Wait for a folio to finish writeback. | |
3143 | * @folio: The folio to wait for. | |
3144 | * | |
3145 | * If the folio is currently being written back to storage, wait for the | |
3146 | * I/O to complete or a fatal signal to arrive. | |
3147 | * | |
3148 | * Context: Sleeps. Must be called in process context and with | |
3149 | * no spinlocks held. Caller should hold a reference on the folio. | |
3150 | * If the folio is not locked, writeback may start again after writeback | |
3151 | * has finished. | |
3152 | * Return: 0 on success, -EINTR if we get a fatal signal while waiting. | |
e5dbd332 | 3153 | */ |
490e016f | 3154 | int folio_wait_writeback_killable(struct folio *folio) |
e5dbd332 | 3155 | { |
490e016f | 3156 | while (folio_test_writeback(folio)) { |
b9b0ff61 | 3157 | trace_folio_wait_writeback(folio, folio_mapping(folio)); |
101c0bf6 | 3158 | if (folio_wait_bit_killable(folio, PG_writeback)) |
e5dbd332 MWO |
3159 | return -EINTR; |
3160 | } | |
3161 | ||
3162 | return 0; | |
3163 | } | |
490e016f | 3164 | EXPORT_SYMBOL_GPL(folio_wait_writeback_killable); |
e5dbd332 | 3165 | |
1d1d1a76 | 3166 | /** |
a49d0c50 MWO |
3167 | * folio_wait_stable() - wait for writeback to finish, if necessary. |
3168 | * @folio: The folio to wait on. | |
1d1d1a76 | 3169 | * |
a49d0c50 MWO |
3170 | * This function determines if the given folio is related to a backing |
3171 | * device that requires folio contents to be held stable during writeback. | |
3172 | * If so, then it will wait for any pending writeback to complete. | |
3173 | * | |
3174 | * Context: Sleeps. Must be called in process context and with | |
3175 | * no spinlocks held. Caller should hold a reference on the folio. | |
3176 | * If the folio is not locked, writeback may start again after writeback | |
3177 | * has finished. | |
1d1d1a76 | 3178 | */ |
a49d0c50 | 3179 | void folio_wait_stable(struct folio *folio) |
1d1d1a76 | 3180 | { |
452c472e | 3181 | if (folio_inode(folio)->i_sb->s_iflags & SB_I_STABLE_WRITES) |
a49d0c50 | 3182 | folio_wait_writeback(folio); |
1d1d1a76 | 3183 | } |
a49d0c50 | 3184 | EXPORT_SYMBOL_GPL(folio_wait_stable); |