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; | |
9bb48a70 | 142 | bool freerun; |
236d0f16 | 143 | bool dirty_exceeded; |
2bc00aef TH |
144 | }; |
145 | ||
eb608e3a JK |
146 | /* |
147 | * Length of period for aging writeout fractions of bdis. This is an | |
148 | * arbitrarily chosen number. The longer the period, the slower fractions will | |
149 | * reflect changes in current writeout rate. | |
150 | */ | |
151 | #define VM_COMPLETIONS_PERIOD_LEN (3*HZ) | |
04fbfdc1 | 152 | |
693108a8 TH |
153 | #ifdef CONFIG_CGROUP_WRITEBACK |
154 | ||
d60d1bdd TH |
155 | #define GDTC_INIT(__wb) .wb = (__wb), \ |
156 | .dom = &global_wb_domain, \ | |
157 | .wb_completions = &(__wb)->completions | |
158 | ||
9fc3a43e | 159 | #define GDTC_INIT_NO_WB .dom = &global_wb_domain |
d60d1bdd TH |
160 | |
161 | #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \ | |
162 | .dom = mem_cgroup_wb_domain(__wb), \ | |
163 | .wb_completions = &(__wb)->memcg_completions, \ | |
164 | .gdtc = __gdtc | |
c2aa723a TH |
165 | |
166 | static bool mdtc_valid(struct dirty_throttle_control *dtc) | |
167 | { | |
168 | return dtc->dom; | |
169 | } | |
e9f07dfd TH |
170 | |
171 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) | |
172 | { | |
173 | return dtc->dom; | |
174 | } | |
175 | ||
9fc3a43e TH |
176 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
177 | { | |
178 | return mdtc->gdtc; | |
179 | } | |
180 | ||
841710aa TH |
181 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
182 | { | |
183 | return &wb->memcg_completions; | |
184 | } | |
185 | ||
693108a8 TH |
186 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
187 | unsigned long *minp, unsigned long *maxp) | |
188 | { | |
20792ebf | 189 | unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth); |
693108a8 TH |
190 | unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); |
191 | unsigned long long min = wb->bdi->min_ratio; | |
192 | unsigned long long max = wb->bdi->max_ratio; | |
193 | ||
194 | /* | |
195 | * @wb may already be clean by the time control reaches here and | |
196 | * the total may not include its bw. | |
197 | */ | |
198 | if (this_bw < tot_bw) { | |
199 | if (min) { | |
200 | min *= this_bw; | |
6d9e8c65 | 201 | min = div64_ul(min, tot_bw); |
693108a8 | 202 | } |
ae82291e | 203 | if (max < 100 * BDI_RATIO_SCALE) { |
693108a8 | 204 | max *= this_bw; |
6d9e8c65 | 205 | max = div64_ul(max, tot_bw); |
693108a8 TH |
206 | } |
207 | } | |
208 | ||
209 | *minp = min; | |
210 | *maxp = max; | |
211 | } | |
212 | ||
213 | #else /* CONFIG_CGROUP_WRITEBACK */ | |
214 | ||
d60d1bdd TH |
215 | #define GDTC_INIT(__wb) .wb = (__wb), \ |
216 | .wb_completions = &(__wb)->completions | |
9fc3a43e | 217 | #define GDTC_INIT_NO_WB |
c2aa723a TH |
218 | #define MDTC_INIT(__wb, __gdtc) |
219 | ||
220 | static bool mdtc_valid(struct dirty_throttle_control *dtc) | |
221 | { | |
222 | return false; | |
223 | } | |
e9f07dfd TH |
224 | |
225 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) | |
226 | { | |
227 | return &global_wb_domain; | |
228 | } | |
229 | ||
9fc3a43e TH |
230 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
231 | { | |
232 | return NULL; | |
233 | } | |
234 | ||
841710aa TH |
235 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
236 | { | |
237 | return NULL; | |
238 | } | |
239 | ||
693108a8 TH |
240 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
241 | unsigned long *minp, unsigned long *maxp) | |
242 | { | |
243 | *minp = wb->bdi->min_ratio; | |
244 | *maxp = wb->bdi->max_ratio; | |
245 | } | |
246 | ||
247 | #endif /* CONFIG_CGROUP_WRITEBACK */ | |
248 | ||
a756cf59 JW |
249 | /* |
250 | * In a memory zone, there is a certain amount of pages we consider | |
251 | * available for the page cache, which is essentially the number of | |
252 | * free and reclaimable pages, minus some zone reserves to protect | |
253 | * lowmem and the ability to uphold the zone's watermarks without | |
254 | * requiring writeback. | |
255 | * | |
256 | * This number of dirtyable pages is the base value of which the | |
e0857cf5 | 257 | * user-configurable dirty ratio is the effective number of pages that |
a756cf59 JW |
258 | * are allowed to be actually dirtied. Per individual zone, or |
259 | * globally by using the sum of dirtyable pages over all zones. | |
260 | * | |
261 | * Because the user is allowed to specify the dirty limit globally as | |
262 | * absolute number of bytes, calculating the per-zone dirty limit can | |
263 | * require translating the configured limit into a percentage of | |
264 | * global dirtyable memory first. | |
265 | */ | |
266 | ||
a804552b | 267 | /** |
281e3726 MG |
268 | * node_dirtyable_memory - number of dirtyable pages in a node |
269 | * @pgdat: the node | |
a804552b | 270 | * |
a862f68a | 271 | * Return: the node's number of pages potentially available for dirty |
281e3726 | 272 | * page cache. This is the base value for the per-node dirty limits. |
a804552b | 273 | */ |
281e3726 | 274 | static unsigned long node_dirtyable_memory(struct pglist_data *pgdat) |
a804552b | 275 | { |
281e3726 MG |
276 | unsigned long nr_pages = 0; |
277 | int z; | |
278 | ||
279 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
280 | struct zone *zone = pgdat->node_zones + z; | |
281 | ||
282 | if (!populated_zone(zone)) | |
283 | continue; | |
284 | ||
285 | nr_pages += zone_page_state(zone, NR_FREE_PAGES); | |
286 | } | |
a804552b | 287 | |
a8d01437 JW |
288 | /* |
289 | * Pages reserved for the kernel should not be considered | |
290 | * dirtyable, to prevent a situation where reclaim has to | |
291 | * clean pages in order to balance the zones. | |
292 | */ | |
281e3726 | 293 | nr_pages -= min(nr_pages, pgdat->totalreserve_pages); |
a804552b | 294 | |
281e3726 MG |
295 | nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE); |
296 | nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE); | |
a804552b JW |
297 | |
298 | return nr_pages; | |
299 | } | |
300 | ||
1edf2234 JW |
301 | static unsigned long highmem_dirtyable_memory(unsigned long total) |
302 | { | |
303 | #ifdef CONFIG_HIGHMEM | |
304 | int node; | |
bb4cc2be | 305 | unsigned long x = 0; |
09b4ab3c | 306 | int i; |
1edf2234 JW |
307 | |
308 | for_each_node_state(node, N_HIGH_MEMORY) { | |
281e3726 MG |
309 | for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) { |
310 | struct zone *z; | |
9cb937e2 | 311 | unsigned long nr_pages; |
281e3726 MG |
312 | |
313 | if (!is_highmem_idx(i)) | |
314 | continue; | |
315 | ||
316 | z = &NODE_DATA(node)->node_zones[i]; | |
9cb937e2 MK |
317 | if (!populated_zone(z)) |
318 | continue; | |
1edf2234 | 319 | |
9cb937e2 | 320 | nr_pages = zone_page_state(z, NR_FREE_PAGES); |
281e3726 | 321 | /* watch for underflows */ |
9cb937e2 | 322 | nr_pages -= min(nr_pages, high_wmark_pages(z)); |
bb4cc2be MG |
323 | nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE); |
324 | nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE); | |
325 | x += nr_pages; | |
09b4ab3c | 326 | } |
1edf2234 | 327 | } |
281e3726 | 328 | |
1edf2234 JW |
329 | /* |
330 | * Make sure that the number of highmem pages is never larger | |
331 | * than the number of the total dirtyable memory. This can only | |
332 | * occur in very strange VM situations but we want to make sure | |
333 | * that this does not occur. | |
334 | */ | |
335 | return min(x, total); | |
336 | #else | |
337 | return 0; | |
338 | #endif | |
339 | } | |
340 | ||
341 | /** | |
ccafa287 | 342 | * global_dirtyable_memory - number of globally dirtyable pages |
1edf2234 | 343 | * |
a862f68a | 344 | * Return: the global number of pages potentially available for dirty |
ccafa287 | 345 | * page cache. This is the base value for the global dirty limits. |
1edf2234 | 346 | */ |
18cf8cf8 | 347 | static unsigned long global_dirtyable_memory(void) |
1edf2234 JW |
348 | { |
349 | unsigned long x; | |
350 | ||
c41f012a | 351 | x = global_zone_page_state(NR_FREE_PAGES); |
a8d01437 JW |
352 | /* |
353 | * Pages reserved for the kernel should not be considered | |
354 | * dirtyable, to prevent a situation where reclaim has to | |
355 | * clean pages in order to balance the zones. | |
356 | */ | |
357 | x -= min(x, totalreserve_pages); | |
1edf2234 | 358 | |
599d0c95 MG |
359 | x += global_node_page_state(NR_INACTIVE_FILE); |
360 | x += global_node_page_state(NR_ACTIVE_FILE); | |
a804552b | 361 | |
1edf2234 JW |
362 | if (!vm_highmem_is_dirtyable) |
363 | x -= highmem_dirtyable_memory(x); | |
364 | ||
365 | return x + 1; /* Ensure that we never return 0 */ | |
366 | } | |
367 | ||
9fc3a43e TH |
368 | /** |
369 | * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain | |
370 | * @dtc: dirty_throttle_control of interest | |
ccafa287 | 371 | * |
9fc3a43e TH |
372 | * Calculate @dtc->thresh and ->bg_thresh considering |
373 | * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller | |
374 | * must ensure that @dtc->avail is set before calling this function. The | |
a37b0715 | 375 | * dirty limits will be lifted by 1/4 for real-time tasks. |
ccafa287 | 376 | */ |
9fc3a43e | 377 | static void domain_dirty_limits(struct dirty_throttle_control *dtc) |
ccafa287 | 378 | { |
9fc3a43e TH |
379 | const unsigned long available_memory = dtc->avail; |
380 | struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc); | |
381 | unsigned long bytes = vm_dirty_bytes; | |
382 | unsigned long bg_bytes = dirty_background_bytes; | |
62a584fe TH |
383 | /* convert ratios to per-PAGE_SIZE for higher precision */ |
384 | unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100; | |
385 | unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100; | |
9fc3a43e TH |
386 | unsigned long thresh; |
387 | unsigned long bg_thresh; | |
ccafa287 JW |
388 | struct task_struct *tsk; |
389 | ||
9fc3a43e TH |
390 | /* gdtc is !NULL iff @dtc is for memcg domain */ |
391 | if (gdtc) { | |
392 | unsigned long global_avail = gdtc->avail; | |
393 | ||
394 | /* | |
395 | * The byte settings can't be applied directly to memcg | |
396 | * domains. Convert them to ratios by scaling against | |
62a584fe TH |
397 | * globally available memory. As the ratios are in |
398 | * per-PAGE_SIZE, they can be obtained by dividing bytes by | |
399 | * number of pages. | |
9fc3a43e TH |
400 | */ |
401 | if (bytes) | |
62a584fe TH |
402 | ratio = min(DIV_ROUND_UP(bytes, global_avail), |
403 | PAGE_SIZE); | |
9fc3a43e | 404 | if (bg_bytes) |
62a584fe TH |
405 | bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail), |
406 | PAGE_SIZE); | |
9fc3a43e TH |
407 | bytes = bg_bytes = 0; |
408 | } | |
409 | ||
410 | if (bytes) | |
411 | thresh = DIV_ROUND_UP(bytes, PAGE_SIZE); | |
ccafa287 | 412 | else |
62a584fe | 413 | thresh = (ratio * available_memory) / PAGE_SIZE; |
ccafa287 | 414 | |
9fc3a43e TH |
415 | if (bg_bytes) |
416 | bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE); | |
ccafa287 | 417 | else |
62a584fe | 418 | bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE; |
ccafa287 | 419 | |
ccafa287 | 420 | tsk = current; |
ae04f69d | 421 | if (rt_or_dl_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 | } |
68ed2a39 JK |
425 | /* |
426 | * Dirty throttling logic assumes the limits in page units fit into | |
427 | * 32-bits. This gives 16TB dirty limits max which is hopefully enough. | |
428 | */ | |
429 | if (thresh > UINT_MAX) | |
430 | thresh = UINT_MAX; | |
431 | /* This makes sure bg_thresh is within 32-bits as well */ | |
432 | if (bg_thresh >= thresh) | |
433 | bg_thresh = thresh / 2; | |
9fc3a43e TH |
434 | dtc->thresh = thresh; |
435 | dtc->bg_thresh = bg_thresh; | |
436 | ||
437 | /* we should eventually report the domain in the TP */ | |
438 | if (!gdtc) | |
439 | trace_global_dirty_state(bg_thresh, thresh); | |
440 | } | |
441 | ||
442 | /** | |
443 | * global_dirty_limits - background-writeback and dirty-throttling thresholds | |
444 | * @pbackground: out parameter for bg_thresh | |
445 | * @pdirty: out parameter for thresh | |
446 | * | |
447 | * Calculate bg_thresh and thresh for global_wb_domain. See | |
448 | * domain_dirty_limits() for details. | |
449 | */ | |
450 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) | |
451 | { | |
452 | struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB }; | |
453 | ||
454 | gdtc.avail = global_dirtyable_memory(); | |
455 | domain_dirty_limits(&gdtc); | |
456 | ||
457 | *pbackground = gdtc.bg_thresh; | |
458 | *pdirty = gdtc.thresh; | |
ccafa287 JW |
459 | } |
460 | ||
a756cf59 | 461 | /** |
281e3726 MG |
462 | * node_dirty_limit - maximum number of dirty pages allowed in a node |
463 | * @pgdat: the node | |
a756cf59 | 464 | * |
a862f68a | 465 | * Return: the maximum number of dirty pages allowed in a node, based |
281e3726 | 466 | * on the node's dirtyable memory. |
a756cf59 | 467 | */ |
281e3726 | 468 | static unsigned long node_dirty_limit(struct pglist_data *pgdat) |
a756cf59 | 469 | { |
281e3726 | 470 | unsigned long node_memory = node_dirtyable_memory(pgdat); |
a756cf59 JW |
471 | struct task_struct *tsk = current; |
472 | unsigned long dirty; | |
473 | ||
474 | if (vm_dirty_bytes) | |
475 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) * | |
281e3726 | 476 | node_memory / global_dirtyable_memory(); |
a756cf59 | 477 | else |
281e3726 | 478 | dirty = vm_dirty_ratio * node_memory / 100; |
a756cf59 | 479 | |
ae04f69d | 480 | if (rt_or_dl_task(tsk)) |
a756cf59 JW |
481 | dirty += dirty / 4; |
482 | ||
68ed2a39 JK |
483 | /* |
484 | * Dirty throttling logic assumes the limits in page units fit into | |
485 | * 32-bits. This gives 16TB dirty limits max which is hopefully enough. | |
486 | */ | |
487 | return min_t(unsigned long, dirty, UINT_MAX); | |
a756cf59 JW |
488 | } |
489 | ||
490 | /** | |
281e3726 MG |
491 | * node_dirty_ok - tells whether a node is within its dirty limits |
492 | * @pgdat: the node to check | |
a756cf59 | 493 | * |
a862f68a | 494 | * Return: %true when the dirty pages in @pgdat are within the node's |
a756cf59 JW |
495 | * dirty limit, %false if the limit is exceeded. |
496 | */ | |
281e3726 | 497 | bool node_dirty_ok(struct pglist_data *pgdat) |
a756cf59 | 498 | { |
281e3726 MG |
499 | unsigned long limit = node_dirty_limit(pgdat); |
500 | unsigned long nr_pages = 0; | |
501 | ||
11fb9989 | 502 | nr_pages += node_page_state(pgdat, NR_FILE_DIRTY); |
11fb9989 | 503 | nr_pages += node_page_state(pgdat, NR_WRITEBACK); |
a756cf59 | 504 | |
281e3726 | 505 | return nr_pages <= limit; |
a756cf59 JW |
506 | } |
507 | ||
aa779e51 | 508 | #ifdef CONFIG_SYSCTL |
78eb4ea2 | 509 | static int dirty_background_ratio_handler(const struct ctl_table *table, int write, |
32927393 | 510 | void *buffer, size_t *lenp, loff_t *ppos) |
2da02997 DR |
511 | { |
512 | int ret; | |
513 | ||
8d65af78 | 514 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 DR |
515 | if (ret == 0 && write) |
516 | dirty_background_bytes = 0; | |
517 | return ret; | |
518 | } | |
519 | ||
78eb4ea2 | 520 | static int dirty_background_bytes_handler(const struct ctl_table *table, int write, |
32927393 | 521 | void *buffer, size_t *lenp, loff_t *ppos) |
2da02997 DR |
522 | { |
523 | int ret; | |
68ed2a39 | 524 | unsigned long old_bytes = dirty_background_bytes; |
2da02997 | 525 | |
8d65af78 | 526 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
68ed2a39 JK |
527 | if (ret == 0 && write) { |
528 | if (DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE) > | |
529 | UINT_MAX) { | |
530 | dirty_background_bytes = old_bytes; | |
531 | return -ERANGE; | |
532 | } | |
2da02997 | 533 | dirty_background_ratio = 0; |
68ed2a39 | 534 | } |
2da02997 DR |
535 | return ret; |
536 | } | |
537 | ||
78eb4ea2 | 538 | static int dirty_ratio_handler(const struct ctl_table *table, int write, void *buffer, |
32927393 | 539 | size_t *lenp, loff_t *ppos) |
04fbfdc1 PZ |
540 | { |
541 | int old_ratio = vm_dirty_ratio; | |
2da02997 DR |
542 | int ret; |
543 | ||
8d65af78 | 544 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
04fbfdc1 | 545 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
eb608e3a | 546 | writeback_set_ratelimit(); |
2da02997 DR |
547 | vm_dirty_bytes = 0; |
548 | } | |
549 | return ret; | |
550 | } | |
551 | ||
78eb4ea2 | 552 | static int dirty_bytes_handler(const struct ctl_table *table, int write, |
32927393 | 553 | void *buffer, size_t *lenp, loff_t *ppos) |
2da02997 | 554 | { |
fc3501d4 | 555 | unsigned long old_bytes = vm_dirty_bytes; |
2da02997 DR |
556 | int ret; |
557 | ||
8d65af78 | 558 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
2da02997 | 559 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
68ed2a39 JK |
560 | if (DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) > UINT_MAX) { |
561 | vm_dirty_bytes = old_bytes; | |
562 | return -ERANGE; | |
563 | } | |
eb608e3a | 564 | writeback_set_ratelimit(); |
2da02997 | 565 | vm_dirty_ratio = 0; |
04fbfdc1 PZ |
566 | } |
567 | return ret; | |
568 | } | |
aa779e51 | 569 | #endif |
04fbfdc1 | 570 | |
eb608e3a JK |
571 | static unsigned long wp_next_time(unsigned long cur_time) |
572 | { | |
573 | cur_time += VM_COMPLETIONS_PERIOD_LEN; | |
574 | /* 0 has a special meaning... */ | |
575 | if (!cur_time) | |
576 | return 1; | |
577 | return cur_time; | |
578 | } | |
579 | ||
cc24df4c | 580 | static void wb_domain_writeout_add(struct wb_domain *dom, |
c7981433 | 581 | struct fprop_local_percpu *completions, |
cc24df4c | 582 | unsigned int max_prop_frac, long nr) |
04fbfdc1 | 583 | { |
be5f1797 | 584 | __fprop_add_percpu_max(&dom->completions, completions, |
cc24df4c | 585 | max_prop_frac, nr); |
eb608e3a | 586 | /* First event after period switching was turned off? */ |
517663ed | 587 | if (unlikely(!dom->period_time)) { |
eb608e3a JK |
588 | /* |
589 | * We can race with other __bdi_writeout_inc calls here but | |
590 | * it does not cause any harm since the resulting time when | |
591 | * timer will fire and what is in writeout_period_time will be | |
592 | * roughly the same. | |
593 | */ | |
380c27ca TH |
594 | dom->period_time = wp_next_time(jiffies); |
595 | mod_timer(&dom->period_timer, dom->period_time); | |
eb608e3a | 596 | } |
04fbfdc1 PZ |
597 | } |
598 | ||
c7981433 TH |
599 | /* |
600 | * Increment @wb's writeout completion count and the global writeout | |
269ccca3 | 601 | * completion count. Called from __folio_end_writeback(). |
c7981433 | 602 | */ |
cc24df4c | 603 | static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr) |
dd5656e5 | 604 | { |
841710aa | 605 | struct wb_domain *cgdom; |
dd5656e5 | 606 | |
cc24df4c MWO |
607 | wb_stat_mod(wb, WB_WRITTEN, nr); |
608 | wb_domain_writeout_add(&global_wb_domain, &wb->completions, | |
609 | wb->bdi->max_prop_frac, nr); | |
841710aa TH |
610 | |
611 | cgdom = mem_cgroup_wb_domain(wb); | |
612 | if (cgdom) | |
cc24df4c MWO |
613 | wb_domain_writeout_add(cgdom, wb_memcg_completions(wb), |
614 | wb->bdi->max_prop_frac, nr); | |
dd5656e5 | 615 | } |
dd5656e5 | 616 | |
93f78d88 | 617 | void wb_writeout_inc(struct bdi_writeback *wb) |
04fbfdc1 | 618 | { |
dd5656e5 MS |
619 | unsigned long flags; |
620 | ||
621 | local_irq_save(flags); | |
cc24df4c | 622 | __wb_writeout_add(wb, 1); |
dd5656e5 | 623 | local_irq_restore(flags); |
04fbfdc1 | 624 | } |
93f78d88 | 625 | EXPORT_SYMBOL_GPL(wb_writeout_inc); |
04fbfdc1 | 626 | |
eb608e3a JK |
627 | /* |
628 | * On idle system, we can be called long after we scheduled because we use | |
629 | * deferred timers so count with missed periods. | |
630 | */ | |
9823e51b | 631 | static void writeout_period(struct timer_list *t) |
eb608e3a | 632 | { |
9823e51b | 633 | struct wb_domain *dom = from_timer(dom, t, period_timer); |
380c27ca | 634 | int miss_periods = (jiffies - dom->period_time) / |
eb608e3a JK |
635 | VM_COMPLETIONS_PERIOD_LEN; |
636 | ||
380c27ca TH |
637 | if (fprop_new_period(&dom->completions, miss_periods + 1)) { |
638 | dom->period_time = wp_next_time(dom->period_time + | |
eb608e3a | 639 | miss_periods * VM_COMPLETIONS_PERIOD_LEN); |
380c27ca | 640 | mod_timer(&dom->period_timer, dom->period_time); |
eb608e3a JK |
641 | } else { |
642 | /* | |
643 | * Aging has zeroed all fractions. Stop wasting CPU on period | |
644 | * updates. | |
645 | */ | |
380c27ca | 646 | dom->period_time = 0; |
eb608e3a JK |
647 | } |
648 | } | |
649 | ||
380c27ca TH |
650 | int wb_domain_init(struct wb_domain *dom, gfp_t gfp) |
651 | { | |
652 | memset(dom, 0, sizeof(*dom)); | |
dcc25ae7 TH |
653 | |
654 | spin_lock_init(&dom->lock); | |
655 | ||
9823e51b | 656 | timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE); |
dcc25ae7 TH |
657 | |
658 | dom->dirty_limit_tstamp = jiffies; | |
659 | ||
380c27ca TH |
660 | return fprop_global_init(&dom->completions, gfp); |
661 | } | |
662 | ||
841710aa TH |
663 | #ifdef CONFIG_CGROUP_WRITEBACK |
664 | void wb_domain_exit(struct wb_domain *dom) | |
665 | { | |
666 | del_timer_sync(&dom->period_timer); | |
667 | fprop_global_destroy(&dom->completions); | |
668 | } | |
669 | #endif | |
670 | ||
189d3c4a | 671 | /* |
d08c429b JW |
672 | * bdi_min_ratio keeps the sum of the minimum dirty shares of all |
673 | * registered backing devices, which, for obvious reasons, can not | |
674 | * exceed 100%. | |
189d3c4a | 675 | */ |
189d3c4a PZ |
676 | static unsigned int bdi_min_ratio; |
677 | ||
1bf27e98 SR |
678 | static int bdi_check_pages_limit(unsigned long pages) |
679 | { | |
680 | unsigned long max_dirty_pages = global_dirtyable_memory(); | |
681 | ||
682 | if (pages > max_dirty_pages) | |
683 | return -EINVAL; | |
684 | ||
685 | return 0; | |
686 | } | |
687 | ||
688 | static unsigned long bdi_ratio_from_pages(unsigned long pages) | |
689 | { | |
690 | unsigned long background_thresh; | |
691 | unsigned long dirty_thresh; | |
692 | unsigned long ratio; | |
693 | ||
694 | global_dirty_limits(&background_thresh, &dirty_thresh); | |
695 | ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh); | |
696 | ||
697 | return ratio; | |
698 | } | |
699 | ||
00df7d51 SR |
700 | static u64 bdi_get_bytes(unsigned int ratio) |
701 | { | |
702 | unsigned long background_thresh; | |
703 | unsigned long dirty_thresh; | |
704 | u64 bytes; | |
705 | ||
706 | global_dirty_limits(&background_thresh, &dirty_thresh); | |
707 | bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100; | |
708 | ||
709 | return bytes; | |
710 | } | |
711 | ||
8021fb32 | 712 | static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) |
189d3c4a | 713 | { |
21f0dd88 | 714 | unsigned int delta; |
189d3c4a | 715 | int ret = 0; |
189d3c4a | 716 | |
2c44af4f SR |
717 | if (min_ratio > 100 * BDI_RATIO_SCALE) |
718 | return -EINVAL; | |
ae82291e | 719 | |
cfc4ba53 | 720 | spin_lock_bh(&bdi_lock); |
a42dde04 | 721 | if (min_ratio > bdi->max_ratio) { |
189d3c4a | 722 | ret = -EINVAL; |
a42dde04 | 723 | } else { |
21f0dd88 CW |
724 | if (min_ratio < bdi->min_ratio) { |
725 | delta = bdi->min_ratio - min_ratio; | |
726 | bdi_min_ratio -= delta; | |
727 | bdi->min_ratio = min_ratio; | |
a42dde04 | 728 | } else { |
21f0dd88 | 729 | delta = min_ratio - bdi->min_ratio; |
ae82291e | 730 | if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) { |
21f0dd88 CW |
731 | bdi_min_ratio += delta; |
732 | bdi->min_ratio = min_ratio; | |
733 | } else { | |
734 | ret = -EINVAL; | |
735 | } | |
a42dde04 PZ |
736 | } |
737 | } | |
cfc4ba53 | 738 | spin_unlock_bh(&bdi_lock); |
a42dde04 PZ |
739 | |
740 | return ret; | |
741 | } | |
742 | ||
efc3e6ad | 743 | static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio) |
a42dde04 | 744 | { |
a42dde04 PZ |
745 | int ret = 0; |
746 | ||
4e230b40 SR |
747 | if (max_ratio > 100 * BDI_RATIO_SCALE) |
748 | return -EINVAL; | |
749 | ||
cfc4ba53 | 750 | spin_lock_bh(&bdi_lock); |
a42dde04 PZ |
751 | if (bdi->min_ratio > max_ratio) { |
752 | ret = -EINVAL; | |
753 | } else { | |
754 | bdi->max_ratio = max_ratio; | |
fa151a39 JX |
755 | bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / |
756 | (100 * BDI_RATIO_SCALE); | |
a42dde04 | 757 | } |
cfc4ba53 | 758 | spin_unlock_bh(&bdi_lock); |
189d3c4a PZ |
759 | |
760 | return ret; | |
761 | } | |
efc3e6ad | 762 | |
2c44af4f SR |
763 | int bdi_set_min_ratio_no_scale(struct backing_dev_info *bdi, unsigned int min_ratio) |
764 | { | |
765 | return __bdi_set_min_ratio(bdi, min_ratio); | |
766 | } | |
767 | ||
4e230b40 SR |
768 | int bdi_set_max_ratio_no_scale(struct backing_dev_info *bdi, unsigned int max_ratio) |
769 | { | |
770 | return __bdi_set_max_ratio(bdi, max_ratio); | |
771 | } | |
772 | ||
8021fb32 SR |
773 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) |
774 | { | |
775 | return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE); | |
776 | } | |
777 | ||
efc3e6ad SR |
778 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio) |
779 | { | |
efc3e6ad SR |
780 | return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE); |
781 | } | |
a42dde04 | 782 | EXPORT_SYMBOL(bdi_set_max_ratio); |
189d3c4a | 783 | |
712c00d6 SR |
784 | u64 bdi_get_min_bytes(struct backing_dev_info *bdi) |
785 | { | |
786 | return bdi_get_bytes(bdi->min_ratio); | |
787 | } | |
788 | ||
803c9805 SR |
789 | int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes) |
790 | { | |
791 | int ret; | |
792 | unsigned long pages = min_bytes >> PAGE_SHIFT; | |
793 | unsigned long min_ratio; | |
794 | ||
795 | ret = bdi_check_pages_limit(pages); | |
796 | if (ret) | |
797 | return ret; | |
798 | ||
799 | min_ratio = bdi_ratio_from_pages(pages); | |
800 | return __bdi_set_min_ratio(bdi, min_ratio); | |
801 | } | |
802 | ||
00df7d51 SR |
803 | u64 bdi_get_max_bytes(struct backing_dev_info *bdi) |
804 | { | |
805 | return bdi_get_bytes(bdi->max_ratio); | |
806 | } | |
807 | ||
1bf27e98 SR |
808 | int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes) |
809 | { | |
810 | int ret; | |
811 | unsigned long pages = max_bytes >> PAGE_SHIFT; | |
812 | unsigned long max_ratio; | |
813 | ||
814 | ret = bdi_check_pages_limit(pages); | |
815 | if (ret) | |
816 | return ret; | |
817 | ||
818 | max_ratio = bdi_ratio_from_pages(pages); | |
819 | return __bdi_set_max_ratio(bdi, max_ratio); | |
820 | } | |
821 | ||
8e9d5ead SR |
822 | int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit) |
823 | { | |
824 | if (strict_limit > 1) | |
825 | return -EINVAL; | |
826 | ||
827 | spin_lock_bh(&bdi_lock); | |
828 | if (strict_limit) | |
829 | bdi->capabilities |= BDI_CAP_STRICTLIMIT; | |
830 | else | |
831 | bdi->capabilities &= ~BDI_CAP_STRICTLIMIT; | |
832 | spin_unlock_bh(&bdi_lock); | |
833 | ||
834 | return 0; | |
835 | } | |
836 | ||
6c14ae1e WF |
837 | static unsigned long dirty_freerun_ceiling(unsigned long thresh, |
838 | unsigned long bg_thresh) | |
839 | { | |
840 | return (thresh + bg_thresh) / 2; | |
841 | } | |
842 | ||
c7981433 TH |
843 | static unsigned long hard_dirty_limit(struct wb_domain *dom, |
844 | unsigned long thresh) | |
ffd1f609 | 845 | { |
dcc25ae7 | 846 | return max(thresh, dom->dirty_limit); |
ffd1f609 WF |
847 | } |
848 | ||
c5edf9cd TH |
849 | /* |
850 | * Memory which can be further allocated to a memcg domain is capped by | |
851 | * system-wide clean memory excluding the amount being used in the domain. | |
852 | */ | |
853 | static void mdtc_calc_avail(struct dirty_throttle_control *mdtc, | |
854 | unsigned long filepages, unsigned long headroom) | |
c2aa723a TH |
855 | { |
856 | struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc); | |
c5edf9cd TH |
857 | unsigned long clean = filepages - min(filepages, mdtc->dirty); |
858 | unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty); | |
859 | unsigned long other_clean = global_clean - min(global_clean, clean); | |
c2aa723a | 860 | |
c5edf9cd | 861 | mdtc->avail = filepages + min(headroom, other_clean); |
ffd1f609 WF |
862 | } |
863 | ||
ba62d5cf KS |
864 | static inline bool dtc_is_global(struct dirty_throttle_control *dtc) |
865 | { | |
866 | return mdtc_gdtc(dtc) == NULL; | |
867 | } | |
868 | ||
869 | /* | |
870 | * Dirty background will ignore pages being written as we're trying to | |
871 | * decide whether to put more under writeback. | |
872 | */ | |
873 | static void domain_dirty_avail(struct dirty_throttle_control *dtc, | |
874 | bool include_writeback) | |
875 | { | |
876 | if (dtc_is_global(dtc)) { | |
877 | dtc->avail = global_dirtyable_memory(); | |
878 | dtc->dirty = global_node_page_state(NR_FILE_DIRTY); | |
879 | if (include_writeback) | |
880 | dtc->dirty += global_node_page_state(NR_WRITEBACK); | |
881 | } else { | |
882 | unsigned long filepages = 0, headroom = 0, writeback = 0; | |
883 | ||
884 | mem_cgroup_wb_stats(dtc->wb, &filepages, &headroom, &dtc->dirty, | |
885 | &writeback); | |
886 | if (include_writeback) | |
887 | dtc->dirty += writeback; | |
888 | mdtc_calc_avail(dtc, filepages, headroom); | |
889 | } | |
890 | } | |
891 | ||
6f718656 | 892 | /** |
13fc4412 | 893 | * __wb_calc_thresh - @wb's share of dirty threshold |
b1cbc6d4 | 894 | * @dtc: dirty_throttle_context of interest |
13fc4412 | 895 | * @thresh: dirty throttling or dirty background threshold of wb_domain in @dtc |
1babe183 | 896 | * |
13fc4412 KS |
897 | * Note that balance_dirty_pages() will only seriously take dirty throttling |
898 | * threshold as a hard limit when sleeping max_pause per page is not enough | |
899 | * to keep the dirty pages under control. For example, when the device is | |
900 | * completely stalled due to some error conditions, or when there are 1000 | |
901 | * dd tasks writing to a slow 10MB/s USB key. | |
aed21ad2 | 902 | * In the other normal situations, it acts more gently by throttling the tasks |
a88a341a | 903 | * more (rather than completely block them) when the wb dirty pages go high. |
1babe183 | 904 | * |
6f718656 | 905 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
1babe183 WF |
906 | * - starving fast devices |
907 | * - piling up dirty pages (that will take long time to sync) on slow devices | |
908 | * | |
a88a341a | 909 | * The wb's share of dirty limit will be adapting to its throughput and |
1babe183 | 910 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. |
a862f68a | 911 | * |
13fc4412 KS |
912 | * Return: @wb's dirty limit in pages. For dirty throttling limit, the term |
913 | * "dirty" in the context of dirty balancing includes all PG_dirty and | |
914 | * PG_writeback pages. | |
1babe183 | 915 | */ |
13fc4412 KS |
916 | static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc, |
917 | unsigned long thresh) | |
16c4042f | 918 | { |
e9f07dfd | 919 | struct wb_domain *dom = dtc_dom(dtc); |
0d960a38 | 920 | u64 wb_thresh; |
d3ac946e | 921 | unsigned long numerator, denominator; |
693108a8 | 922 | unsigned long wb_min_ratio, wb_max_ratio; |
04fbfdc1 | 923 | |
16c4042f | 924 | /* |
13fc4412 | 925 | * Calculate this wb's share of the thresh ratio. |
16c4042f | 926 | */ |
e9770b34 | 927 | fprop_fraction_percpu(&dom->completions, dtc->wb_completions, |
380c27ca | 928 | &numerator, &denominator); |
04fbfdc1 | 929 | |
ae82291e | 930 | wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE); |
0d960a38 | 931 | wb_thresh *= numerator; |
d3ac946e | 932 | wb_thresh = div64_ul(wb_thresh, denominator); |
04fbfdc1 | 933 | |
b1cbc6d4 | 934 | wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio); |
04fbfdc1 | 935 | |
ae82291e SR |
936 | wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE); |
937 | if (wb_thresh > (thresh * wb_max_ratio) / (100 * BDI_RATIO_SCALE)) | |
938 | wb_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE); | |
16c4042f | 939 | |
0d960a38 | 940 | return wb_thresh; |
1da177e4 LT |
941 | } |
942 | ||
b1cbc6d4 TH |
943 | unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh) |
944 | { | |
13fc4412 KS |
945 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
946 | ||
947 | return __wb_calc_thresh(&gdtc, thresh); | |
1da177e4 LT |
948 | } |
949 | ||
4b5bbc39 KS |
950 | unsigned long cgwb_calc_thresh(struct bdi_writeback *wb) |
951 | { | |
952 | struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB }; | |
953 | struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) }; | |
4b5bbc39 | 954 | |
ba62d5cf KS |
955 | domain_dirty_avail(&gdtc, true); |
956 | domain_dirty_avail(&mdtc, true); | |
4b5bbc39 KS |
957 | domain_dirty_limits(&mdtc); |
958 | ||
13fc4412 | 959 | return __wb_calc_thresh(&mdtc, mdtc.thresh); |
1da177e4 LT |
960 | } |
961 | ||
5a537485 MP |
962 | /* |
963 | * setpoint - dirty 3 | |
964 | * f(dirty) := 1.0 + (----------------) | |
965 | * limit - setpoint | |
966 | * | |
967 | * it's a 3rd order polynomial that subjects to | |
968 | * | |
969 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast | |
970 | * (2) f(setpoint) = 1.0 => the balance point | |
971 | * (3) f(limit) = 0 => the hard limit | |
972 | * (4) df/dx <= 0 => negative feedback control | |
973 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) | |
974 | * => fast response on large errors; small oscillation near setpoint | |
975 | */ | |
d5c9fde3 | 976 | static long long pos_ratio_polynom(unsigned long setpoint, |
5a537485 MP |
977 | unsigned long dirty, |
978 | unsigned long limit) | |
979 | { | |
980 | long long pos_ratio; | |
981 | long x; | |
982 | ||
d5c9fde3 | 983 | x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT, |
464d1387 | 984 | (limit - setpoint) | 1); |
5a537485 MP |
985 | pos_ratio = x; |
986 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
987 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; | |
988 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT; | |
989 | ||
990 | return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT); | |
991 | } | |
992 | ||
6c14ae1e WF |
993 | /* |
994 | * Dirty position control. | |
995 | * | |
996 | * (o) global/bdi setpoints | |
997 | * | |
de1fff37 | 998 | * We want the dirty pages be balanced around the global/wb setpoints. |
6c14ae1e WF |
999 | * When the number of dirty pages is higher/lower than the setpoint, the |
1000 | * dirty position control ratio (and hence task dirty ratelimit) will be | |
1001 | * decreased/increased to bring the dirty pages back to the setpoint. | |
1002 | * | |
1003 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT | |
1004 | * | |
1005 | * if (dirty < setpoint) scale up pos_ratio | |
1006 | * if (dirty > setpoint) scale down pos_ratio | |
1007 | * | |
de1fff37 TH |
1008 | * if (wb_dirty < wb_setpoint) scale up pos_ratio |
1009 | * if (wb_dirty > wb_setpoint) scale down pos_ratio | |
6c14ae1e WF |
1010 | * |
1011 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT | |
1012 | * | |
1013 | * (o) global control line | |
1014 | * | |
1015 | * ^ pos_ratio | |
1016 | * | | |
1017 | * | |<===== global dirty control scope ======>| | |
03231554 | 1018 | * 2.0 * * * * * * * |
6c14ae1e WF |
1019 | * | .* |
1020 | * | . * | |
1021 | * | . * | |
1022 | * | . * | |
1023 | * | . * | |
1024 | * | . * | |
1025 | * 1.0 ................................* | |
1026 | * | . . * | |
1027 | * | . . * | |
1028 | * | . . * | |
1029 | * | . . * | |
1030 | * | . . * | |
1031 | * 0 +------------.------------------.----------------------*-------------> | |
1032 | * freerun^ setpoint^ limit^ dirty pages | |
1033 | * | |
de1fff37 | 1034 | * (o) wb control line |
6c14ae1e WF |
1035 | * |
1036 | * ^ pos_ratio | |
1037 | * | | |
1038 | * | * | |
1039 | * | * | |
1040 | * | * | |
1041 | * | * | |
1042 | * | * |<=========== span ============>| | |
1043 | * 1.0 .......................* | |
1044 | * | . * | |
1045 | * | . * | |
1046 | * | . * | |
1047 | * | . * | |
1048 | * | . * | |
1049 | * | . * | |
1050 | * | . * | |
1051 | * | . * | |
1052 | * | . * | |
1053 | * | . * | |
1054 | * | . * | |
1055 | * 1/4 ...............................................* * * * * * * * * * * * | |
1056 | * | . . | |
1057 | * | . . | |
1058 | * | . . | |
1059 | * 0 +----------------------.-------------------------------.-------------> | |
de1fff37 | 1060 | * wb_setpoint^ x_intercept^ |
6c14ae1e | 1061 | * |
de1fff37 | 1062 | * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can |
6c14ae1e WF |
1063 | * be smoothly throttled down to normal if it starts high in situations like |
1064 | * - start writing to a slow SD card and a fast disk at the same time. The SD | |
de1fff37 TH |
1065 | * card's wb_dirty may rush to many times higher than wb_setpoint. |
1066 | * - the wb dirty thresh drops quickly due to change of JBOD workload | |
6c14ae1e | 1067 | */ |
daddfa3c | 1068 | static void wb_position_ratio(struct dirty_throttle_control *dtc) |
6c14ae1e | 1069 | { |
2bc00aef | 1070 | struct bdi_writeback *wb = dtc->wb; |
20792ebf | 1071 | unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth); |
2bc00aef | 1072 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); |
c7981433 | 1073 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
2bc00aef | 1074 | unsigned long wb_thresh = dtc->wb_thresh; |
6c14ae1e WF |
1075 | unsigned long x_intercept; |
1076 | unsigned long setpoint; /* dirty pages' target balance point */ | |
de1fff37 | 1077 | unsigned long wb_setpoint; |
6c14ae1e WF |
1078 | unsigned long span; |
1079 | long long pos_ratio; /* for scaling up/down the rate limit */ | |
1080 | long x; | |
1081 | ||
daddfa3c TH |
1082 | dtc->pos_ratio = 0; |
1083 | ||
2bc00aef | 1084 | if (unlikely(dtc->dirty >= limit)) |
daddfa3c | 1085 | return; |
6c14ae1e WF |
1086 | |
1087 | /* | |
1088 | * global setpoint | |
1089 | * | |
5a537485 MP |
1090 | * See comment for pos_ratio_polynom(). |
1091 | */ | |
1092 | setpoint = (freerun + limit) / 2; | |
2bc00aef | 1093 | pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit); |
5a537485 MP |
1094 | |
1095 | /* | |
1096 | * The strictlimit feature is a tool preventing mistrusted filesystems | |
1097 | * from growing a large number of dirty pages before throttling. For | |
de1fff37 TH |
1098 | * such filesystems balance_dirty_pages always checks wb counters |
1099 | * against wb limits. Even if global "nr_dirty" is under "freerun". | |
5a537485 MP |
1100 | * This is especially important for fuse which sets bdi->max_ratio to |
1101 | * 1% by default. Without strictlimit feature, fuse writeback may | |
1102 | * consume arbitrary amount of RAM because it is accounted in | |
1103 | * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty". | |
6c14ae1e | 1104 | * |
a88a341a | 1105 | * Here, in wb_position_ratio(), we calculate pos_ratio based on |
de1fff37 | 1106 | * two values: wb_dirty and wb_thresh. Let's consider an example: |
5a537485 MP |
1107 | * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global |
1108 | * limits are set by default to 10% and 20% (background and throttle). | |
de1fff37 | 1109 | * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages. |
0d960a38 | 1110 | * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is |
de1fff37 | 1111 | * about ~6K pages (as the average of background and throttle wb |
5a537485 | 1112 | * limits). The 3rd order polynomial will provide positive feedback if |
de1fff37 | 1113 | * wb_dirty is under wb_setpoint and vice versa. |
6c14ae1e | 1114 | * |
5a537485 | 1115 | * Note, that we cannot use global counters in these calculations |
de1fff37 | 1116 | * because we want to throttle process writing to a strictlimit wb |
5a537485 MP |
1117 | * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB |
1118 | * in the example above). | |
6c14ae1e | 1119 | */ |
a88a341a | 1120 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
de1fff37 | 1121 | long long wb_pos_ratio; |
5a537485 | 1122 | |
daddfa3c TH |
1123 | if (dtc->wb_dirty < 8) { |
1124 | dtc->pos_ratio = min_t(long long, pos_ratio * 2, | |
1125 | 2 << RATELIMIT_CALC_SHIFT); | |
1126 | return; | |
1127 | } | |
5a537485 | 1128 | |
2bc00aef | 1129 | if (dtc->wb_dirty >= wb_thresh) |
daddfa3c | 1130 | return; |
5a537485 | 1131 | |
970fb01a TH |
1132 | wb_setpoint = dirty_freerun_ceiling(wb_thresh, |
1133 | dtc->wb_bg_thresh); | |
5a537485 | 1134 | |
de1fff37 | 1135 | if (wb_setpoint == 0 || wb_setpoint == wb_thresh) |
daddfa3c | 1136 | return; |
5a537485 | 1137 | |
2bc00aef | 1138 | wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty, |
de1fff37 | 1139 | wb_thresh); |
5a537485 MP |
1140 | |
1141 | /* | |
de1fff37 TH |
1142 | * Typically, for strictlimit case, wb_setpoint << setpoint |
1143 | * and pos_ratio >> wb_pos_ratio. In the other words global | |
5a537485 | 1144 | * state ("dirty") is not limiting factor and we have to |
de1fff37 | 1145 | * make decision based on wb counters. But there is an |
5a537485 MP |
1146 | * important case when global pos_ratio should get precedence: |
1147 | * global limits are exceeded (e.g. due to activities on other | |
de1fff37 | 1148 | * wb's) while given strictlimit wb is below limit. |
5a537485 | 1149 | * |
de1fff37 | 1150 | * "pos_ratio * wb_pos_ratio" would work for the case above, |
5a537485 | 1151 | * but it would look too non-natural for the case of all |
de1fff37 | 1152 | * activity in the system coming from a single strictlimit wb |
5a537485 MP |
1153 | * with bdi->max_ratio == 100%. |
1154 | * | |
1155 | * Note that min() below somewhat changes the dynamics of the | |
1156 | * control system. Normally, pos_ratio value can be well over 3 | |
de1fff37 | 1157 | * (when globally we are at freerun and wb is well below wb |
5a537485 MP |
1158 | * setpoint). Now the maximum pos_ratio in the same situation |
1159 | * is 2. We might want to tweak this if we observe the control | |
1160 | * system is too slow to adapt. | |
1161 | */ | |
daddfa3c TH |
1162 | dtc->pos_ratio = min(pos_ratio, wb_pos_ratio); |
1163 | return; | |
5a537485 | 1164 | } |
6c14ae1e WF |
1165 | |
1166 | /* | |
1167 | * We have computed basic pos_ratio above based on global situation. If | |
de1fff37 | 1168 | * the wb is over/under its share of dirty pages, we want to scale |
6c14ae1e WF |
1169 | * pos_ratio further down/up. That is done by the following mechanism. |
1170 | */ | |
1171 | ||
1172 | /* | |
de1fff37 | 1173 | * wb setpoint |
6c14ae1e | 1174 | * |
de1fff37 | 1175 | * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint) |
6c14ae1e | 1176 | * |
de1fff37 | 1177 | * x_intercept - wb_dirty |
6c14ae1e | 1178 | * := -------------------------- |
de1fff37 | 1179 | * x_intercept - wb_setpoint |
6c14ae1e | 1180 | * |
de1fff37 | 1181 | * The main wb control line is a linear function that subjects to |
6c14ae1e | 1182 | * |
de1fff37 TH |
1183 | * (1) f(wb_setpoint) = 1.0 |
1184 | * (2) k = - 1 / (8 * write_bw) (in single wb case) | |
1185 | * or equally: x_intercept = wb_setpoint + 8 * write_bw | |
6c14ae1e | 1186 | * |
de1fff37 | 1187 | * For single wb case, the dirty pages are observed to fluctuate |
6c14ae1e | 1188 | * regularly within range |
de1fff37 | 1189 | * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2] |
6c14ae1e WF |
1190 | * for various filesystems, where (2) can yield in a reasonable 12.5% |
1191 | * fluctuation range for pos_ratio. | |
1192 | * | |
de1fff37 | 1193 | * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its |
6c14ae1e | 1194 | * own size, so move the slope over accordingly and choose a slope that |
de1fff37 | 1195 | * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh. |
6c14ae1e | 1196 | */ |
2bc00aef TH |
1197 | if (unlikely(wb_thresh > dtc->thresh)) |
1198 | wb_thresh = dtc->thresh; | |
aed21ad2 | 1199 | /* |
de1fff37 | 1200 | * It's very possible that wb_thresh is close to 0 not because the |
aed21ad2 WF |
1201 | * device is slow, but that it has remained inactive for long time. |
1202 | * Honour such devices a reasonable good (hopefully IO efficient) | |
1203 | * threshold, so that the occasional writes won't be blocked and active | |
1204 | * writes can rampup the threshold quickly. | |
1205 | */ | |
2bc00aef | 1206 | wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8); |
6c14ae1e | 1207 | /* |
de1fff37 TH |
1208 | * scale global setpoint to wb's: |
1209 | * wb_setpoint = setpoint * wb_thresh / thresh | |
6c14ae1e | 1210 | */ |
e4bc13ad | 1211 | x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1); |
de1fff37 | 1212 | wb_setpoint = setpoint * (u64)x >> 16; |
6c14ae1e | 1213 | /* |
de1fff37 TH |
1214 | * Use span=(8*write_bw) in single wb case as indicated by |
1215 | * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case. | |
6c14ae1e | 1216 | * |
de1fff37 TH |
1217 | * wb_thresh thresh - wb_thresh |
1218 | * span = --------- * (8 * write_bw) + ------------------ * wb_thresh | |
1219 | * thresh thresh | |
6c14ae1e | 1220 | */ |
2bc00aef | 1221 | span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16; |
de1fff37 | 1222 | x_intercept = wb_setpoint + span; |
6c14ae1e | 1223 | |
2bc00aef TH |
1224 | if (dtc->wb_dirty < x_intercept - span / 4) { |
1225 | pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty), | |
e4bc13ad | 1226 | (x_intercept - wb_setpoint) | 1); |
6c14ae1e WF |
1227 | } else |
1228 | pos_ratio /= 4; | |
1229 | ||
8927f66c | 1230 | /* |
de1fff37 | 1231 | * wb reserve area, safeguard against dirty pool underrun and disk idle |
8927f66c WF |
1232 | * It may push the desired control point of global dirty pages higher |
1233 | * than setpoint. | |
1234 | */ | |
de1fff37 | 1235 | x_intercept = wb_thresh / 2; |
2bc00aef TH |
1236 | if (dtc->wb_dirty < x_intercept) { |
1237 | if (dtc->wb_dirty > x_intercept / 8) | |
1238 | pos_ratio = div_u64(pos_ratio * x_intercept, | |
1239 | dtc->wb_dirty); | |
50657fc4 | 1240 | else |
8927f66c WF |
1241 | pos_ratio *= 8; |
1242 | } | |
1243 | ||
daddfa3c | 1244 | dtc->pos_ratio = pos_ratio; |
6c14ae1e WF |
1245 | } |
1246 | ||
a88a341a TH |
1247 | static void wb_update_write_bandwidth(struct bdi_writeback *wb, |
1248 | unsigned long elapsed, | |
1249 | unsigned long written) | |
e98be2d5 WF |
1250 | { |
1251 | const unsigned long period = roundup_pow_of_two(3 * HZ); | |
a88a341a TH |
1252 | unsigned long avg = wb->avg_write_bandwidth; |
1253 | unsigned long old = wb->write_bandwidth; | |
e98be2d5 WF |
1254 | u64 bw; |
1255 | ||
1256 | /* | |
1257 | * bw = written * HZ / elapsed | |
1258 | * | |
1259 | * bw * elapsed + write_bandwidth * (period - elapsed) | |
1260 | * write_bandwidth = --------------------------------------------------- | |
1261 | * period | |
c72efb65 | 1262 | * |
ed2da924 | 1263 | * @written may have decreased due to folio_redirty_for_writepage(). |
c72efb65 | 1264 | * Avoid underflowing @bw calculation. |
e98be2d5 | 1265 | */ |
a88a341a | 1266 | bw = written - min(written, wb->written_stamp); |
e98be2d5 WF |
1267 | bw *= HZ; |
1268 | if (unlikely(elapsed > period)) { | |
0a5d1a7f | 1269 | bw = div64_ul(bw, elapsed); |
e98be2d5 WF |
1270 | avg = bw; |
1271 | goto out; | |
1272 | } | |
a88a341a | 1273 | bw += (u64)wb->write_bandwidth * (period - elapsed); |
e98be2d5 WF |
1274 | bw >>= ilog2(period); |
1275 | ||
1276 | /* | |
1277 | * one more level of smoothing, for filtering out sudden spikes | |
1278 | */ | |
1279 | if (avg > old && old >= (unsigned long)bw) | |
1280 | avg -= (avg - old) >> 3; | |
1281 | ||
1282 | if (avg < old && old <= (unsigned long)bw) | |
1283 | avg += (old - avg) >> 3; | |
1284 | ||
1285 | out: | |
95a46c65 TH |
1286 | /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */ |
1287 | avg = max(avg, 1LU); | |
1288 | if (wb_has_dirty_io(wb)) { | |
1289 | long delta = avg - wb->avg_write_bandwidth; | |
1290 | WARN_ON_ONCE(atomic_long_add_return(delta, | |
1291 | &wb->bdi->tot_write_bandwidth) <= 0); | |
1292 | } | |
a88a341a | 1293 | wb->write_bandwidth = bw; |
20792ebf | 1294 | WRITE_ONCE(wb->avg_write_bandwidth, avg); |
e98be2d5 WF |
1295 | } |
1296 | ||
2bc00aef | 1297 | static void update_dirty_limit(struct dirty_throttle_control *dtc) |
c42843f2 | 1298 | { |
e9f07dfd | 1299 | struct wb_domain *dom = dtc_dom(dtc); |
2bc00aef | 1300 | unsigned long thresh = dtc->thresh; |
dcc25ae7 | 1301 | unsigned long limit = dom->dirty_limit; |
c42843f2 WF |
1302 | |
1303 | /* | |
1304 | * Follow up in one step. | |
1305 | */ | |
1306 | if (limit < thresh) { | |
1307 | limit = thresh; | |
1308 | goto update; | |
1309 | } | |
1310 | ||
1311 | /* | |
1312 | * Follow down slowly. Use the higher one as the target, because thresh | |
1313 | * may drop below dirty. This is exactly the reason to introduce | |
dcc25ae7 | 1314 | * dom->dirty_limit which is guaranteed to lie above the dirty pages. |
c42843f2 | 1315 | */ |
2bc00aef | 1316 | thresh = max(thresh, dtc->dirty); |
c42843f2 WF |
1317 | if (limit > thresh) { |
1318 | limit -= (limit - thresh) >> 5; | |
1319 | goto update; | |
1320 | } | |
1321 | return; | |
1322 | update: | |
dcc25ae7 | 1323 | dom->dirty_limit = limit; |
c42843f2 WF |
1324 | } |
1325 | ||
42dd235c JK |
1326 | static void domain_update_dirty_limit(struct dirty_throttle_control *dtc, |
1327 | unsigned long now) | |
c42843f2 | 1328 | { |
e9f07dfd | 1329 | struct wb_domain *dom = dtc_dom(dtc); |
c42843f2 WF |
1330 | |
1331 | /* | |
1332 | * check locklessly first to optimize away locking for the most time | |
1333 | */ | |
dcc25ae7 | 1334 | if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) |
c42843f2 WF |
1335 | return; |
1336 | ||
dcc25ae7 TH |
1337 | spin_lock(&dom->lock); |
1338 | if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) { | |
2bc00aef | 1339 | update_dirty_limit(dtc); |
dcc25ae7 | 1340 | dom->dirty_limit_tstamp = now; |
c42843f2 | 1341 | } |
dcc25ae7 | 1342 | spin_unlock(&dom->lock); |
c42843f2 WF |
1343 | } |
1344 | ||
be3ffa27 | 1345 | /* |
de1fff37 | 1346 | * Maintain wb->dirty_ratelimit, the base dirty throttle rate. |
be3ffa27 | 1347 | * |
de1fff37 | 1348 | * Normal wb tasks will be curbed at or below it in long term. |
be3ffa27 WF |
1349 | * Obviously it should be around (write_bw / N) when there are N dd tasks. |
1350 | */ | |
2bc00aef | 1351 | static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc, |
a88a341a TH |
1352 | unsigned long dirtied, |
1353 | unsigned long elapsed) | |
be3ffa27 | 1354 | { |
2bc00aef TH |
1355 | struct bdi_writeback *wb = dtc->wb; |
1356 | unsigned long dirty = dtc->dirty; | |
1357 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); | |
c7981433 | 1358 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
7381131c | 1359 | unsigned long setpoint = (freerun + limit) / 2; |
a88a341a TH |
1360 | unsigned long write_bw = wb->avg_write_bandwidth; |
1361 | unsigned long dirty_ratelimit = wb->dirty_ratelimit; | |
be3ffa27 WF |
1362 | unsigned long dirty_rate; |
1363 | unsigned long task_ratelimit; | |
1364 | unsigned long balanced_dirty_ratelimit; | |
7381131c WF |
1365 | unsigned long step; |
1366 | unsigned long x; | |
d59b1087 | 1367 | unsigned long shift; |
be3ffa27 WF |
1368 | |
1369 | /* | |
1370 | * The dirty rate will match the writeout rate in long term, except | |
1371 | * when dirty pages are truncated by userspace or re-dirtied by FS. | |
1372 | */ | |
a88a341a | 1373 | dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed; |
be3ffa27 | 1374 | |
be3ffa27 WF |
1375 | /* |
1376 | * task_ratelimit reflects each dd's dirty rate for the past 200ms. | |
1377 | */ | |
1378 | task_ratelimit = (u64)dirty_ratelimit * | |
daddfa3c | 1379 | dtc->pos_ratio >> RATELIMIT_CALC_SHIFT; |
be3ffa27 WF |
1380 | task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ |
1381 | ||
1382 | /* | |
1383 | * A linear estimation of the "balanced" throttle rate. The theory is, | |
de1fff37 | 1384 | * if there are N dd tasks, each throttled at task_ratelimit, the wb's |
be3ffa27 WF |
1385 | * dirty_rate will be measured to be (N * task_ratelimit). So the below |
1386 | * formula will yield the balanced rate limit (write_bw / N). | |
1387 | * | |
1388 | * Note that the expanded form is not a pure rate feedback: | |
1389 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) | |
1390 | * but also takes pos_ratio into account: | |
1391 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) | |
1392 | * | |
1393 | * (1) is not realistic because pos_ratio also takes part in balancing | |
1394 | * the dirty rate. Consider the state | |
1395 | * pos_ratio = 0.5 (3) | |
1396 | * rate = 2 * (write_bw / N) (4) | |
1397 | * If (1) is used, it will stuck in that state! Because each dd will | |
1398 | * be throttled at | |
1399 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) | |
1400 | * yielding | |
1401 | * dirty_rate = N * task_ratelimit = write_bw (6) | |
1402 | * put (6) into (1) we get | |
1403 | * rate_(i+1) = rate_(i) (7) | |
1404 | * | |
1405 | * So we end up using (2) to always keep | |
1406 | * rate_(i+1) ~= (write_bw / N) (8) | |
1407 | * regardless of the value of pos_ratio. As long as (8) is satisfied, | |
1408 | * pos_ratio is able to drive itself to 1.0, which is not only where | |
1409 | * the dirty count meet the setpoint, but also where the slope of | |
1410 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated. | |
1411 | */ | |
1412 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, | |
1413 | dirty_rate | 1); | |
bdaac490 WF |
1414 | /* |
1415 | * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw | |
1416 | */ | |
1417 | if (unlikely(balanced_dirty_ratelimit > write_bw)) | |
1418 | balanced_dirty_ratelimit = write_bw; | |
be3ffa27 | 1419 | |
7381131c WF |
1420 | /* |
1421 | * We could safely do this and return immediately: | |
1422 | * | |
de1fff37 | 1423 | * wb->dirty_ratelimit = balanced_dirty_ratelimit; |
7381131c WF |
1424 | * |
1425 | * However to get a more stable dirty_ratelimit, the below elaborated | |
331cbdee | 1426 | * code makes use of task_ratelimit to filter out singular points and |
7381131c WF |
1427 | * limit the step size. |
1428 | * | |
1429 | * The below code essentially only uses the relative value of | |
1430 | * | |
1431 | * task_ratelimit - dirty_ratelimit | |
1432 | * = (pos_ratio - 1) * dirty_ratelimit | |
1433 | * | |
1434 | * which reflects the direction and size of dirty position error. | |
1435 | */ | |
1436 | ||
1437 | /* | |
1438 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff | |
1439 | * task_ratelimit is on the same side of dirty_ratelimit, too. | |
1440 | * For example, when | |
1441 | * - dirty_ratelimit > balanced_dirty_ratelimit | |
1442 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) | |
1443 | * lowering dirty_ratelimit will help meet both the position and rate | |
1444 | * control targets. Otherwise, don't update dirty_ratelimit if it will | |
1445 | * only help meet the rate target. After all, what the users ultimately | |
1446 | * feel and care are stable dirty rate and small position error. | |
1447 | * | |
1448 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size | |
331cbdee | 1449 | * and filter out the singular points of balanced_dirty_ratelimit. Which |
7381131c WF |
1450 | * keeps jumping around randomly and can even leap far away at times |
1451 | * due to the small 200ms estimation period of dirty_rate (we want to | |
1452 | * keep that period small to reduce time lags). | |
1453 | */ | |
1454 | step = 0; | |
5a537485 MP |
1455 | |
1456 | /* | |
de1fff37 | 1457 | * For strictlimit case, calculations above were based on wb counters |
a88a341a | 1458 | * and limits (starting from pos_ratio = wb_position_ratio() and up to |
5a537485 | 1459 | * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate). |
de1fff37 TH |
1460 | * Hence, to calculate "step" properly, we have to use wb_dirty as |
1461 | * "dirty" and wb_setpoint as "setpoint". | |
5a537485 | 1462 | * |
de1fff37 TH |
1463 | * We rampup dirty_ratelimit forcibly if wb_dirty is low because |
1464 | * it's possible that wb_thresh is close to zero due to inactivity | |
970fb01a | 1465 | * of backing device. |
5a537485 | 1466 | */ |
a88a341a | 1467 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
2bc00aef TH |
1468 | dirty = dtc->wb_dirty; |
1469 | if (dtc->wb_dirty < 8) | |
1470 | setpoint = dtc->wb_dirty + 1; | |
5a537485 | 1471 | else |
970fb01a | 1472 | setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2; |
5a537485 MP |
1473 | } |
1474 | ||
7381131c | 1475 | if (dirty < setpoint) { |
a88a341a | 1476 | x = min3(wb->balanced_dirty_ratelimit, |
7c809968 | 1477 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1478 | if (dirty_ratelimit < x) |
1479 | step = x - dirty_ratelimit; | |
1480 | } else { | |
a88a341a | 1481 | x = max3(wb->balanced_dirty_ratelimit, |
7c809968 | 1482 | balanced_dirty_ratelimit, task_ratelimit); |
7381131c WF |
1483 | if (dirty_ratelimit > x) |
1484 | step = dirty_ratelimit - x; | |
1485 | } | |
1486 | ||
1487 | /* | |
1488 | * Don't pursue 100% rate matching. It's impossible since the balanced | |
1489 | * rate itself is constantly fluctuating. So decrease the track speed | |
1490 | * when it gets close to the target. Helps eliminate pointless tremors. | |
1491 | */ | |
d59b1087 AR |
1492 | shift = dirty_ratelimit / (2 * step + 1); |
1493 | if (shift < BITS_PER_LONG) | |
1494 | step = DIV_ROUND_UP(step >> shift, 8); | |
1495 | else | |
1496 | step = 0; | |
7381131c WF |
1497 | |
1498 | if (dirty_ratelimit < balanced_dirty_ratelimit) | |
1499 | dirty_ratelimit += step; | |
1500 | else | |
1501 | dirty_ratelimit -= step; | |
1502 | ||
20792ebf | 1503 | WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL)); |
a88a341a | 1504 | wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit; |
b48c104d | 1505 | |
5634cc2a | 1506 | trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit); |
be3ffa27 WF |
1507 | } |
1508 | ||
c2aa723a TH |
1509 | static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc, |
1510 | struct dirty_throttle_control *mdtc, | |
8a731799 | 1511 | bool update_ratelimit) |
e98be2d5 | 1512 | { |
c2aa723a | 1513 | struct bdi_writeback *wb = gdtc->wb; |
e98be2d5 | 1514 | unsigned long now = jiffies; |
45a2966f | 1515 | unsigned long elapsed; |
be3ffa27 | 1516 | unsigned long dirtied; |
e98be2d5 WF |
1517 | unsigned long written; |
1518 | ||
45a2966f | 1519 | spin_lock(&wb->list_lock); |
8a731799 | 1520 | |
e98be2d5 | 1521 | /* |
45a2966f JK |
1522 | * Lockless checks for elapsed time are racy and delayed update after |
1523 | * IO completion doesn't do it at all (to make sure written pages are | |
1524 | * accounted reasonably quickly). Make sure elapsed >= 1 to avoid | |
1525 | * division errors. | |
e98be2d5 | 1526 | */ |
45a2966f | 1527 | elapsed = max(now - wb->bw_time_stamp, 1UL); |
a88a341a TH |
1528 | dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]); |
1529 | written = percpu_counter_read(&wb->stat[WB_WRITTEN]); | |
e98be2d5 | 1530 | |
8a731799 | 1531 | if (update_ratelimit) { |
42dd235c | 1532 | domain_update_dirty_limit(gdtc, now); |
c2aa723a TH |
1533 | wb_update_dirty_ratelimit(gdtc, dirtied, elapsed); |
1534 | ||
1535 | /* | |
1536 | * @mdtc is always NULL if !CGROUP_WRITEBACK but the | |
1537 | * compiler has no way to figure that out. Help it. | |
1538 | */ | |
1539 | if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) { | |
42dd235c | 1540 | domain_update_dirty_limit(mdtc, now); |
c2aa723a TH |
1541 | wb_update_dirty_ratelimit(mdtc, dirtied, elapsed); |
1542 | } | |
be3ffa27 | 1543 | } |
a88a341a | 1544 | wb_update_write_bandwidth(wb, elapsed, written); |
e98be2d5 | 1545 | |
a88a341a TH |
1546 | wb->dirtied_stamp = dirtied; |
1547 | wb->written_stamp = written; | |
20792ebf | 1548 | WRITE_ONCE(wb->bw_time_stamp, now); |
45a2966f | 1549 | spin_unlock(&wb->list_lock); |
e98be2d5 WF |
1550 | } |
1551 | ||
45a2966f | 1552 | void wb_update_bandwidth(struct bdi_writeback *wb) |
e98be2d5 | 1553 | { |
2bc00aef TH |
1554 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
1555 | ||
fee468fd | 1556 | __wb_update_bandwidth(&gdtc, NULL, false); |
fee468fd JK |
1557 | } |
1558 | ||
1559 | /* Interval after which we consider wb idle and don't estimate bandwidth */ | |
1560 | #define WB_BANDWIDTH_IDLE_JIF (HZ) | |
1561 | ||
1562 | static void wb_bandwidth_estimate_start(struct bdi_writeback *wb) | |
1563 | { | |
1564 | unsigned long now = jiffies; | |
1565 | unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp); | |
1566 | ||
1567 | if (elapsed > WB_BANDWIDTH_IDLE_JIF && | |
1568 | !atomic_read(&wb->writeback_inodes)) { | |
1569 | spin_lock(&wb->list_lock); | |
1570 | wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED); | |
1571 | wb->written_stamp = wb_stat(wb, WB_WRITTEN); | |
20792ebf | 1572 | WRITE_ONCE(wb->bw_time_stamp, now); |
fee468fd JK |
1573 | spin_unlock(&wb->list_lock); |
1574 | } | |
e98be2d5 WF |
1575 | } |
1576 | ||
9d823e8f | 1577 | /* |
d0e1d66b | 1578 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited() |
9d823e8f WF |
1579 | * will look to see if it needs to start dirty throttling. |
1580 | * | |
1581 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive | |
c41f012a | 1582 | * global_zone_page_state() too often. So scale it near-sqrt to the safety margin |
9d823e8f WF |
1583 | * (the number of pages we may dirty without exceeding the dirty limits). |
1584 | */ | |
1585 | static unsigned long dirty_poll_interval(unsigned long dirty, | |
1586 | unsigned long thresh) | |
1587 | { | |
1588 | if (thresh > dirty) | |
1589 | return 1UL << (ilog2(thresh - dirty) >> 1); | |
1590 | ||
1591 | return 1; | |
1592 | } | |
1593 | ||
a88a341a | 1594 | static unsigned long wb_max_pause(struct bdi_writeback *wb, |
de1fff37 | 1595 | unsigned long wb_dirty) |
c8462cc9 | 1596 | { |
20792ebf | 1597 | unsigned long bw = READ_ONCE(wb->avg_write_bandwidth); |
e3b6c655 | 1598 | unsigned long t; |
c8462cc9 | 1599 | |
7ccb9ad5 WF |
1600 | /* |
1601 | * Limit pause time for small memory systems. If sleeping for too long | |
1602 | * time, a small pool of dirty/writeback pages may go empty and disk go | |
1603 | * idle. | |
1604 | * | |
1605 | * 8 serves as the safety ratio. | |
1606 | */ | |
de1fff37 | 1607 | t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8)); |
7ccb9ad5 WF |
1608 | t++; |
1609 | ||
e3b6c655 | 1610 | return min_t(unsigned long, t, MAX_PAUSE); |
7ccb9ad5 WF |
1611 | } |
1612 | ||
a88a341a TH |
1613 | static long wb_min_pause(struct bdi_writeback *wb, |
1614 | long max_pause, | |
1615 | unsigned long task_ratelimit, | |
1616 | unsigned long dirty_ratelimit, | |
1617 | int *nr_dirtied_pause) | |
c8462cc9 | 1618 | { |
20792ebf JK |
1619 | long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth)); |
1620 | long lo = ilog2(READ_ONCE(wb->dirty_ratelimit)); | |
7ccb9ad5 WF |
1621 | long t; /* target pause */ |
1622 | long pause; /* estimated next pause */ | |
1623 | int pages; /* target nr_dirtied_pause */ | |
c8462cc9 | 1624 | |
7ccb9ad5 WF |
1625 | /* target for 10ms pause on 1-dd case */ |
1626 | t = max(1, HZ / 100); | |
c8462cc9 WF |
1627 | |
1628 | /* | |
1629 | * Scale up pause time for concurrent dirtiers in order to reduce CPU | |
1630 | * overheads. | |
1631 | * | |
7ccb9ad5 | 1632 | * (N * 10ms) on 2^N concurrent tasks. |
c8462cc9 WF |
1633 | */ |
1634 | if (hi > lo) | |
7ccb9ad5 | 1635 | t += (hi - lo) * (10 * HZ) / 1024; |
c8462cc9 WF |
1636 | |
1637 | /* | |
7ccb9ad5 WF |
1638 | * This is a bit convoluted. We try to base the next nr_dirtied_pause |
1639 | * on the much more stable dirty_ratelimit. However the next pause time | |
1640 | * will be computed based on task_ratelimit and the two rate limits may | |
1641 | * depart considerably at some time. Especially if task_ratelimit goes | |
1642 | * below dirty_ratelimit/2 and the target pause is max_pause, the next | |
1643 | * pause time will be max_pause*2 _trimmed down_ to max_pause. As a | |
1644 | * result task_ratelimit won't be executed faithfully, which could | |
1645 | * eventually bring down dirty_ratelimit. | |
c8462cc9 | 1646 | * |
7ccb9ad5 WF |
1647 | * We apply two rules to fix it up: |
1648 | * 1) try to estimate the next pause time and if necessary, use a lower | |
1649 | * nr_dirtied_pause so as not to exceed max_pause. When this happens, | |
1650 | * nr_dirtied_pause will be "dancing" with task_ratelimit. | |
1651 | * 2) limit the target pause time to max_pause/2, so that the normal | |
1652 | * small fluctuations of task_ratelimit won't trigger rule (1) and | |
1653 | * nr_dirtied_pause will remain as stable as dirty_ratelimit. | |
c8462cc9 | 1654 | */ |
7ccb9ad5 WF |
1655 | t = min(t, 1 + max_pause / 2); |
1656 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
c8462cc9 WF |
1657 | |
1658 | /* | |
5b9b3574 WF |
1659 | * Tiny nr_dirtied_pause is found to hurt I/O performance in the test |
1660 | * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}. | |
1661 | * When the 16 consecutive reads are often interrupted by some dirty | |
1662 | * throttling pause during the async writes, cfq will go into idles | |
1663 | * (deadline is fine). So push nr_dirtied_pause as high as possible | |
1664 | * until reaches DIRTY_POLL_THRESH=32 pages. | |
c8462cc9 | 1665 | */ |
5b9b3574 WF |
1666 | if (pages < DIRTY_POLL_THRESH) { |
1667 | t = max_pause; | |
1668 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); | |
1669 | if (pages > DIRTY_POLL_THRESH) { | |
1670 | pages = DIRTY_POLL_THRESH; | |
1671 | t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit; | |
1672 | } | |
1673 | } | |
1674 | ||
7ccb9ad5 WF |
1675 | pause = HZ * pages / (task_ratelimit + 1); |
1676 | if (pause > max_pause) { | |
1677 | t = max_pause; | |
1678 | pages = task_ratelimit * t / roundup_pow_of_two(HZ); | |
1679 | } | |
c8462cc9 | 1680 | |
7ccb9ad5 | 1681 | *nr_dirtied_pause = pages; |
c8462cc9 | 1682 | /* |
7ccb9ad5 | 1683 | * The minimal pause time will normally be half the target pause time. |
c8462cc9 | 1684 | */ |
5b9b3574 | 1685 | return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t; |
c8462cc9 WF |
1686 | } |
1687 | ||
970fb01a | 1688 | static inline void wb_dirty_limits(struct dirty_throttle_control *dtc) |
5a537485 | 1689 | { |
2bc00aef | 1690 | struct bdi_writeback *wb = dtc->wb; |
93f78d88 | 1691 | unsigned long wb_reclaimable; |
5a537485 MP |
1692 | |
1693 | /* | |
de1fff37 | 1694 | * wb_thresh is not treated as some limiting factor as |
5a537485 | 1695 | * dirty_thresh, due to reasons |
de1fff37 | 1696 | * - in JBOD setup, wb_thresh can fluctuate a lot |
5a537485 | 1697 | * - in a system with HDD and USB key, the USB key may somehow |
de1fff37 TH |
1698 | * go into state (wb_dirty >> wb_thresh) either because |
1699 | * wb_dirty starts high, or because wb_thresh drops low. | |
5a537485 | 1700 | * In this case we don't want to hard throttle the USB key |
de1fff37 TH |
1701 | * dirtiers for 100 seconds until wb_dirty drops under |
1702 | * wb_thresh. Instead the auxiliary wb control line in | |
a88a341a | 1703 | * wb_position_ratio() will let the dirtier task progress |
de1fff37 | 1704 | * at some rate <= (write_bw / 2) for bringing down wb_dirty. |
5a537485 | 1705 | */ |
13fc4412 | 1706 | dtc->wb_thresh = __wb_calc_thresh(dtc, dtc->thresh); |
970fb01a | 1707 | dtc->wb_bg_thresh = dtc->thresh ? |
8dfcffa3 | 1708 | div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0; |
5a537485 MP |
1709 | |
1710 | /* | |
1711 | * In order to avoid the stacked BDI deadlock we need | |
1712 | * to ensure we accurately count the 'dirty' pages when | |
1713 | * the threshold is low. | |
1714 | * | |
1715 | * Otherwise it would be possible to get thresh+n pages | |
1716 | * reported dirty, even though there are thresh-m pages | |
1717 | * actually dirty; with m+n sitting in the percpu | |
1718 | * deltas. | |
1719 | */ | |
2bce774e | 1720 | if (dtc->wb_thresh < 2 * wb_stat_error()) { |
93f78d88 | 1721 | wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); |
2bc00aef | 1722 | dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK); |
5a537485 | 1723 | } else { |
93f78d88 | 1724 | wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE); |
2bc00aef | 1725 | dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK); |
5a537485 MP |
1726 | } |
1727 | } | |
1728 | ||
9bb48a70 KS |
1729 | static unsigned long domain_poll_intv(struct dirty_throttle_control *dtc, |
1730 | bool strictlimit) | |
1731 | { | |
1732 | unsigned long dirty, thresh; | |
1733 | ||
1734 | if (strictlimit) { | |
1735 | dirty = dtc->wb_dirty; | |
1736 | thresh = dtc->wb_thresh; | |
1737 | } else { | |
1738 | dirty = dtc->dirty; | |
1739 | thresh = dtc->thresh; | |
1740 | } | |
1741 | ||
1742 | return dirty_poll_interval(dirty, thresh); | |
1743 | } | |
1744 | ||
1745 | /* | |
1746 | * Throttle it only when the background writeback cannot catch-up. This avoids | |
1747 | * (excessively) small writeouts when the wb limits are ramping up in case of | |
1748 | * !strictlimit. | |
1749 | * | |
1750 | * In strictlimit case make decision based on the wb counters and limits. Small | |
1751 | * writeouts when the wb limits are ramping up are the price we consciously pay | |
1752 | * for strictlimit-ing. | |
1753 | */ | |
1754 | static void domain_dirty_freerun(struct dirty_throttle_control *dtc, | |
1755 | bool strictlimit) | |
1756 | { | |
1757 | unsigned long dirty, thresh, bg_thresh; | |
1758 | ||
1759 | if (unlikely(strictlimit)) { | |
1760 | wb_dirty_limits(dtc); | |
1761 | dirty = dtc->wb_dirty; | |
1762 | thresh = dtc->wb_thresh; | |
1763 | bg_thresh = dtc->wb_bg_thresh; | |
1764 | } else { | |
1765 | dirty = dtc->dirty; | |
1766 | thresh = dtc->thresh; | |
1767 | bg_thresh = dtc->bg_thresh; | |
1768 | } | |
1769 | dtc->freerun = dirty <= dirty_freerun_ceiling(thresh, bg_thresh); | |
1770 | } | |
1771 | ||
8c9918de KS |
1772 | static void balance_domain_limits(struct dirty_throttle_control *dtc, |
1773 | bool strictlimit) | |
1774 | { | |
1775 | domain_dirty_avail(dtc, true); | |
1776 | domain_dirty_limits(dtc); | |
1777 | domain_dirty_freerun(dtc, strictlimit); | |
1778 | } | |
1779 | ||
2530e239 KS |
1780 | static void wb_dirty_freerun(struct dirty_throttle_control *dtc, |
1781 | bool strictlimit) | |
1782 | { | |
1783 | dtc->freerun = false; | |
1784 | ||
1785 | /* was already handled in domain_dirty_freerun */ | |
1786 | if (strictlimit) | |
1787 | return; | |
1788 | ||
1789 | wb_dirty_limits(dtc); | |
1790 | /* | |
1791 | * LOCAL_THROTTLE tasks must not be throttled when below the per-wb | |
1792 | * freerun ceiling. | |
1793 | */ | |
1794 | if (!(current->flags & PF_LOCAL_THROTTLE)) | |
1795 | return; | |
1796 | ||
1797 | dtc->freerun = dtc->wb_dirty < | |
1798 | dirty_freerun_ceiling(dtc->wb_thresh, dtc->wb_bg_thresh); | |
1799 | } | |
1800 | ||
236d0f16 KS |
1801 | static inline void wb_dirty_exceeded(struct dirty_throttle_control *dtc, |
1802 | bool strictlimit) | |
1803 | { | |
1804 | dtc->dirty_exceeded = (dtc->wb_dirty > dtc->wb_thresh) && | |
1805 | ((dtc->dirty > dtc->thresh) || strictlimit); | |
1806 | } | |
1807 | ||
8246291e KS |
1808 | /* |
1809 | * The limits fields dirty_exceeded and pos_ratio won't be updated if wb is | |
1810 | * in freerun state. Please don't use these invalid fields in freerun case. | |
1811 | */ | |
1812 | static void balance_wb_limits(struct dirty_throttle_control *dtc, | |
1813 | bool strictlimit) | |
1814 | { | |
1815 | wb_dirty_freerun(dtc, strictlimit); | |
1816 | if (dtc->freerun) | |
1817 | return; | |
1818 | ||
1819 | wb_dirty_exceeded(dtc, strictlimit); | |
1820 | wb_position_ratio(dtc); | |
1821 | } | |
1822 | ||
1da177e4 LT |
1823 | /* |
1824 | * balance_dirty_pages() must be called by processes which are generating dirty | |
1825 | * data. It looks at the number of dirty pages in the machine and will force | |
143dfe86 | 1826 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. |
5b0830cb JA |
1827 | * If we're over `background_thresh' then the writeback threads are woken to |
1828 | * perform some writeout. | |
1da177e4 | 1829 | */ |
fe6c9c6e JK |
1830 | static int balance_dirty_pages(struct bdi_writeback *wb, |
1831 | unsigned long pages_dirtied, unsigned int flags) | |
1da177e4 | 1832 | { |
2bc00aef | 1833 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) }; |
c2aa723a | 1834 | struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) }; |
2bc00aef | 1835 | struct dirty_throttle_control * const gdtc = &gdtc_stor; |
c2aa723a TH |
1836 | struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ? |
1837 | &mdtc_stor : NULL; | |
1838 | struct dirty_throttle_control *sdtc; | |
826881a7 | 1839 | unsigned long nr_dirty; |
83712358 | 1840 | long period; |
7ccb9ad5 WF |
1841 | long pause; |
1842 | long max_pause; | |
1843 | long min_pause; | |
1844 | int nr_dirtied_pause; | |
143dfe86 | 1845 | unsigned long task_ratelimit; |
7ccb9ad5 | 1846 | unsigned long dirty_ratelimit; |
dfb8ae56 | 1847 | struct backing_dev_info *bdi = wb->bdi; |
5a537485 | 1848 | bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT; |
e98be2d5 | 1849 | unsigned long start_time = jiffies; |
fe6c9c6e | 1850 | int ret = 0; |
1da177e4 LT |
1851 | |
1852 | for (;;) { | |
83712358 WF |
1853 | unsigned long now = jiffies; |
1854 | ||
826881a7 | 1855 | nr_dirty = global_node_page_state(NR_FILE_DIRTY); |
5fce25a9 | 1856 | |
8c9918de | 1857 | balance_domain_limits(gdtc, strictlimit); |
c2aa723a | 1858 | if (mdtc) { |
c2aa723a TH |
1859 | /* |
1860 | * If @wb belongs to !root memcg, repeat the same | |
1861 | * basic calculations for the memcg domain. | |
1862 | */ | |
8c9918de | 1863 | balance_domain_limits(mdtc, strictlimit); |
5a537485 MP |
1864 | } |
1865 | ||
ea6813be JK |
1866 | /* |
1867 | * In laptop mode, we wait until hitting the higher threshold | |
1868 | * before starting background writeout, and then write out all | |
1869 | * the way down to the lower threshold. So slow writers cause | |
1870 | * minimal disk activity. | |
1871 | * | |
1872 | * In normal mode, we start background writeout at the lower | |
1873 | * background_thresh, to keep the amount of dirty memory low. | |
1874 | */ | |
826881a7 | 1875 | if (!laptop_mode && nr_dirty > gdtc->bg_thresh && |
ea6813be JK |
1876 | !writeback_in_progress(wb)) |
1877 | wb_start_background_writeback(wb); | |
1878 | ||
16c4042f | 1879 | /* |
c2aa723a TH |
1880 | * If memcg domain is in effect, @dirty should be under |
1881 | * both global and memcg freerun ceilings. | |
16c4042f | 1882 | */ |
9bb48a70 | 1883 | if (gdtc->freerun && (!mdtc || mdtc->freerun)) { |
a37b0715 N |
1884 | unsigned long intv; |
1885 | unsigned long m_intv; | |
1886 | ||
1887 | free_running: | |
9bb48a70 | 1888 | intv = domain_poll_intv(gdtc, strictlimit); |
a37b0715 | 1889 | m_intv = ULONG_MAX; |
c2aa723a | 1890 | |
83712358 WF |
1891 | current->dirty_paused_when = now; |
1892 | current->nr_dirtied = 0; | |
c2aa723a | 1893 | if (mdtc) |
9bb48a70 | 1894 | m_intv = domain_poll_intv(mdtc, strictlimit); |
c2aa723a | 1895 | current->nr_dirtied_pause = min(intv, m_intv); |
16c4042f | 1896 | break; |
83712358 | 1897 | } |
16c4042f | 1898 | |
ea6813be | 1899 | /* Start writeback even when in laptop mode */ |
bc05873d | 1900 | if (unlikely(!writeback_in_progress(wb))) |
9ecf4866 | 1901 | wb_start_background_writeback(wb); |
143dfe86 | 1902 | |
97b27821 TH |
1903 | mem_cgroup_flush_foreign(wb); |
1904 | ||
c2aa723a TH |
1905 | /* |
1906 | * Calculate global domain's pos_ratio and select the | |
1907 | * global dtc by default. | |
1908 | */ | |
8246291e | 1909 | balance_wb_limits(gdtc, strictlimit); |
2530e239 KS |
1910 | if (gdtc->freerun) |
1911 | goto free_running; | |
c2aa723a TH |
1912 | sdtc = gdtc; |
1913 | ||
1914 | if (mdtc) { | |
1915 | /* | |
1916 | * If memcg domain is in effect, calculate its | |
1917 | * pos_ratio. @wb should satisfy constraints from | |
1918 | * both global and memcg domains. Choose the one | |
1919 | * w/ lower pos_ratio. | |
1920 | */ | |
8246291e | 1921 | balance_wb_limits(mdtc, strictlimit); |
2530e239 KS |
1922 | if (mdtc->freerun) |
1923 | goto free_running; | |
c2aa723a TH |
1924 | if (mdtc->pos_ratio < gdtc->pos_ratio) |
1925 | sdtc = mdtc; | |
1926 | } | |
daddfa3c | 1927 | |
236d0f16 KS |
1928 | wb->dirty_exceeded = gdtc->dirty_exceeded || |
1929 | (mdtc && mdtc->dirty_exceeded); | |
20792ebf | 1930 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) + |
45a2966f | 1931 | BANDWIDTH_INTERVAL)) |
fee468fd | 1932 | __wb_update_bandwidth(gdtc, mdtc, true); |
e98be2d5 | 1933 | |
c2aa723a | 1934 | /* throttle according to the chosen dtc */ |
20792ebf | 1935 | dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit); |
c2aa723a | 1936 | task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >> |
3a73dbbc | 1937 | RATELIMIT_CALC_SHIFT; |
c2aa723a | 1938 | max_pause = wb_max_pause(wb, sdtc->wb_dirty); |
a88a341a TH |
1939 | min_pause = wb_min_pause(wb, max_pause, |
1940 | task_ratelimit, dirty_ratelimit, | |
1941 | &nr_dirtied_pause); | |
7ccb9ad5 | 1942 | |
3a73dbbc | 1943 | if (unlikely(task_ratelimit == 0)) { |
83712358 | 1944 | period = max_pause; |
c8462cc9 | 1945 | pause = max_pause; |
143dfe86 | 1946 | goto pause; |
04fbfdc1 | 1947 | } |
83712358 WF |
1948 | period = HZ * pages_dirtied / task_ratelimit; |
1949 | pause = period; | |
1950 | if (current->dirty_paused_when) | |
1951 | pause -= now - current->dirty_paused_when; | |
1952 | /* | |
1953 | * For less than 1s think time (ext3/4 may block the dirtier | |
1954 | * for up to 800ms from time to time on 1-HDD; so does xfs, | |
1955 | * however at much less frequency), try to compensate it in | |
1956 | * future periods by updating the virtual time; otherwise just | |
1957 | * do a reset, as it may be a light dirtier. | |
1958 | */ | |
7ccb9ad5 | 1959 | if (pause < min_pause) { |
5634cc2a | 1960 | trace_balance_dirty_pages(wb, |
c2aa723a TH |
1961 | sdtc->thresh, |
1962 | sdtc->bg_thresh, | |
1963 | sdtc->dirty, | |
1964 | sdtc->wb_thresh, | |
1965 | sdtc->wb_dirty, | |
ece13ac3 WF |
1966 | dirty_ratelimit, |
1967 | task_ratelimit, | |
1968 | pages_dirtied, | |
83712358 | 1969 | period, |
7ccb9ad5 | 1970 | min(pause, 0L), |
ece13ac3 | 1971 | start_time); |
83712358 WF |
1972 | if (pause < -HZ) { |
1973 | current->dirty_paused_when = now; | |
1974 | current->nr_dirtied = 0; | |
1975 | } else if (period) { | |
1976 | current->dirty_paused_when += period; | |
1977 | current->nr_dirtied = 0; | |
7ccb9ad5 WF |
1978 | } else if (current->nr_dirtied_pause <= pages_dirtied) |
1979 | current->nr_dirtied_pause += pages_dirtied; | |
57fc978c | 1980 | break; |
04fbfdc1 | 1981 | } |
7ccb9ad5 WF |
1982 | if (unlikely(pause > max_pause)) { |
1983 | /* for occasional dropped task_ratelimit */ | |
1984 | now += min(pause - max_pause, max_pause); | |
1985 | pause = max_pause; | |
1986 | } | |
143dfe86 WF |
1987 | |
1988 | pause: | |
5634cc2a | 1989 | trace_balance_dirty_pages(wb, |
c2aa723a TH |
1990 | sdtc->thresh, |
1991 | sdtc->bg_thresh, | |
1992 | sdtc->dirty, | |
1993 | sdtc->wb_thresh, | |
1994 | sdtc->wb_dirty, | |
ece13ac3 WF |
1995 | dirty_ratelimit, |
1996 | task_ratelimit, | |
1997 | pages_dirtied, | |
83712358 | 1998 | period, |
ece13ac3 WF |
1999 | pause, |
2000 | start_time); | |
fe6c9c6e JK |
2001 | if (flags & BDP_ASYNC) { |
2002 | ret = -EAGAIN; | |
2003 | break; | |
2004 | } | |
499d05ec | 2005 | __set_current_state(TASK_KILLABLE); |
f814bdda | 2006 | bdi->last_bdp_sleep = jiffies; |
d25105e8 | 2007 | io_schedule_timeout(pause); |
87c6a9b2 | 2008 | |
83712358 WF |
2009 | current->dirty_paused_when = now + pause; |
2010 | current->nr_dirtied = 0; | |
7ccb9ad5 | 2011 | current->nr_dirtied_pause = nr_dirtied_pause; |
83712358 | 2012 | |
ffd1f609 | 2013 | /* |
2bc00aef TH |
2014 | * This is typically equal to (dirty < thresh) and can also |
2015 | * keep "1000+ dd on a slow USB stick" under control. | |
ffd1f609 | 2016 | */ |
1df64719 | 2017 | if (task_ratelimit) |
ffd1f609 | 2018 | break; |
499d05ec | 2019 | |
c5c6343c | 2020 | /* |
f0953a1b | 2021 | * In the case of an unresponsive NFS server and the NFS dirty |
de1fff37 | 2022 | * pages exceeds dirty_thresh, give the other good wb's a pipe |
c5c6343c WF |
2023 | * to go through, so that tasks on them still remain responsive. |
2024 | * | |
3f8b6fb7 | 2025 | * In theory 1 page is enough to keep the consumer-producer |
c5c6343c | 2026 | * pipe going: the flusher cleans 1 page => the task dirties 1 |
de1fff37 | 2027 | * more page. However wb_dirty has accounting errors. So use |
93f78d88 | 2028 | * the larger and more IO friendly wb_stat_error. |
c5c6343c | 2029 | */ |
2bce774e | 2030 | if (sdtc->wb_dirty <= wb_stat_error()) |
c5c6343c WF |
2031 | break; |
2032 | ||
499d05ec JK |
2033 | if (fatal_signal_pending(current)) |
2034 | break; | |
1da177e4 | 2035 | } |
fe6c9c6e | 2036 | return ret; |
1da177e4 LT |
2037 | } |
2038 | ||
9d823e8f | 2039 | static DEFINE_PER_CPU(int, bdp_ratelimits); |
245b2e70 | 2040 | |
54848d73 WF |
2041 | /* |
2042 | * Normal tasks are throttled by | |
2043 | * loop { | |
2044 | * dirty tsk->nr_dirtied_pause pages; | |
2045 | * take a snap in balance_dirty_pages(); | |
2046 | * } | |
2047 | * However there is a worst case. If every task exit immediately when dirtied | |
2048 | * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be | |
2049 | * called to throttle the page dirties. The solution is to save the not yet | |
2050 | * throttled page dirties in dirty_throttle_leaks on task exit and charge them | |
2051 | * randomly into the running tasks. This works well for the above worst case, | |
2052 | * as the new task will pick up and accumulate the old task's leaked dirty | |
2053 | * count and eventually get throttled. | |
2054 | */ | |
2055 | DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0; | |
2056 | ||
1da177e4 | 2057 | /** |
fe6c9c6e JK |
2058 | * balance_dirty_pages_ratelimited_flags - Balance dirty memory state. |
2059 | * @mapping: address_space which was dirtied. | |
2060 | * @flags: BDP flags. | |
1da177e4 LT |
2061 | * |
2062 | * Processes which are dirtying memory should call in here once for each page | |
2063 | * which was newly dirtied. The function will periodically check the system's | |
2064 | * dirty state and will initiate writeback if needed. | |
2065 | * | |
fe6c9c6e JK |
2066 | * See balance_dirty_pages_ratelimited() for details. |
2067 | * | |
2068 | * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to | |
2069 | * indicate that memory is out of balance and the caller must wait | |
2070 | * for I/O to complete. Otherwise, it will return 0 to indicate | |
2071 | * that either memory was already in balance, or it was able to sleep | |
2072 | * until the amount of dirty memory returned to balance. | |
1da177e4 | 2073 | */ |
fe6c9c6e JK |
2074 | int balance_dirty_pages_ratelimited_flags(struct address_space *mapping, |
2075 | unsigned int flags) | |
1da177e4 | 2076 | { |
dfb8ae56 TH |
2077 | struct inode *inode = mapping->host; |
2078 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
2079 | struct bdi_writeback *wb = NULL; | |
9d823e8f | 2080 | int ratelimit; |
fe6c9c6e | 2081 | int ret = 0; |
9d823e8f | 2082 | int *p; |
1da177e4 | 2083 | |
f56753ac | 2084 | if (!(bdi->capabilities & BDI_CAP_WRITEBACK)) |
fe6c9c6e | 2085 | return ret; |
36715cef | 2086 | |
dfb8ae56 TH |
2087 | if (inode_cgwb_enabled(inode)) |
2088 | wb = wb_get_create_current(bdi, GFP_KERNEL); | |
2089 | if (!wb) | |
2090 | wb = &bdi->wb; | |
2091 | ||
9d823e8f | 2092 | ratelimit = current->nr_dirtied_pause; |
a88a341a | 2093 | if (wb->dirty_exceeded) |
9d823e8f WF |
2094 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); |
2095 | ||
9d823e8f | 2096 | preempt_disable(); |
1da177e4 | 2097 | /* |
9d823e8f WF |
2098 | * This prevents one CPU to accumulate too many dirtied pages without |
2099 | * calling into balance_dirty_pages(), which can happen when there are | |
2100 | * 1000+ tasks, all of them start dirtying pages at exactly the same | |
2101 | * time, hence all honoured too large initial task->nr_dirtied_pause. | |
1da177e4 | 2102 | */ |
7c8e0181 | 2103 | p = this_cpu_ptr(&bdp_ratelimits); |
9d823e8f | 2104 | if (unlikely(current->nr_dirtied >= ratelimit)) |
fa5a734e | 2105 | *p = 0; |
d3bc1fef WF |
2106 | else if (unlikely(*p >= ratelimit_pages)) { |
2107 | *p = 0; | |
2108 | ratelimit = 0; | |
1da177e4 | 2109 | } |
54848d73 WF |
2110 | /* |
2111 | * Pick up the dirtied pages by the exited tasks. This avoids lots of | |
2112 | * short-lived tasks (eg. gcc invocations in a kernel build) escaping | |
2113 | * the dirty throttling and livelock other long-run dirtiers. | |
2114 | */ | |
7c8e0181 | 2115 | p = this_cpu_ptr(&dirty_throttle_leaks); |
54848d73 | 2116 | if (*p > 0 && current->nr_dirtied < ratelimit) { |
d0e1d66b | 2117 | unsigned long nr_pages_dirtied; |
54848d73 WF |
2118 | nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied); |
2119 | *p -= nr_pages_dirtied; | |
2120 | current->nr_dirtied += nr_pages_dirtied; | |
1da177e4 | 2121 | } |
fa5a734e | 2122 | preempt_enable(); |
9d823e8f WF |
2123 | |
2124 | if (unlikely(current->nr_dirtied >= ratelimit)) | |
fe6c9c6e | 2125 | ret = balance_dirty_pages(wb, current->nr_dirtied, flags); |
dfb8ae56 TH |
2126 | |
2127 | wb_put(wb); | |
fe6c9c6e JK |
2128 | return ret; |
2129 | } | |
611df5d6 | 2130 | EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags); |
fe6c9c6e JK |
2131 | |
2132 | /** | |
2133 | * balance_dirty_pages_ratelimited - balance dirty memory state. | |
2134 | * @mapping: address_space which was dirtied. | |
2135 | * | |
2136 | * Processes which are dirtying memory should call in here once for each page | |
2137 | * which was newly dirtied. The function will periodically check the system's | |
2138 | * dirty state and will initiate writeback if needed. | |
2139 | * | |
2140 | * Once we're over the dirty memory limit we decrease the ratelimiting | |
2141 | * by a lot, to prevent individual processes from overshooting the limit | |
2142 | * by (ratelimit_pages) each. | |
2143 | */ | |
2144 | void balance_dirty_pages_ratelimited(struct address_space *mapping) | |
2145 | { | |
2146 | balance_dirty_pages_ratelimited_flags(mapping, 0); | |
1da177e4 | 2147 | } |
d0e1d66b | 2148 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited); |
1da177e4 | 2149 | |
7c0c629b KS |
2150 | /* |
2151 | * Similar to wb_dirty_limits, wb_bg_dirty_limits also calculates dirty | |
2152 | * and thresh, but it's for background writeback. | |
2153 | */ | |
2154 | static void wb_bg_dirty_limits(struct dirty_throttle_control *dtc) | |
2155 | { | |
2156 | struct bdi_writeback *wb = dtc->wb; | |
2157 | ||
2158 | dtc->wb_bg_thresh = __wb_calc_thresh(dtc, dtc->bg_thresh); | |
2159 | if (dtc->wb_bg_thresh < 2 * wb_stat_error()) | |
2160 | dtc->wb_dirty = wb_stat_sum(wb, WB_RECLAIMABLE); | |
2161 | else | |
2162 | dtc->wb_dirty = wb_stat(wb, WB_RECLAIMABLE); | |
2163 | } | |
2164 | ||
6e208329 KS |
2165 | static bool domain_over_bg_thresh(struct dirty_throttle_control *dtc) |
2166 | { | |
2167 | domain_dirty_avail(dtc, false); | |
2168 | domain_dirty_limits(dtc); | |
2169 | if (dtc->dirty > dtc->bg_thresh) | |
2170 | return true; | |
2171 | ||
2172 | wb_bg_dirty_limits(dtc); | |
2173 | if (dtc->wb_dirty > dtc->wb_bg_thresh) | |
2174 | return true; | |
2175 | ||
2176 | return false; | |
2177 | } | |
2178 | ||
aa661bbe TH |
2179 | /** |
2180 | * wb_over_bg_thresh - does @wb need to be written back? | |
2181 | * @wb: bdi_writeback of interest | |
2182 | * | |
2183 | * Determines whether background writeback should keep writing @wb or it's | |
a862f68a MR |
2184 | * clean enough. |
2185 | * | |
2186 | * Return: %true if writeback should continue. | |
aa661bbe TH |
2187 | */ |
2188 | bool wb_over_bg_thresh(struct bdi_writeback *wb) | |
2189 | { | |
6e208329 KS |
2190 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
2191 | struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) }; | |
aa661bbe | 2192 | |
6e208329 | 2193 | if (domain_over_bg_thresh(&gdtc)) |
aa661bbe TH |
2194 | return true; |
2195 | ||
6e208329 KS |
2196 | if (mdtc_valid(&mdtc)) |
2197 | return domain_over_bg_thresh(&mdtc); | |
c2aa723a | 2198 | |
aa661bbe TH |
2199 | return false; |
2200 | } | |
2201 | ||
aa779e51 | 2202 | #ifdef CONFIG_SYSCTL |
1da177e4 LT |
2203 | /* |
2204 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs | |
2205 | */ | |
78eb4ea2 | 2206 | static int dirty_writeback_centisecs_handler(const struct ctl_table *table, int write, |
32927393 | 2207 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 2208 | { |
94af5846 YS |
2209 | unsigned int old_interval = dirty_writeback_interval; |
2210 | int ret; | |
2211 | ||
2212 | ret = proc_dointvec(table, write, buffer, length, ppos); | |
515c24c1 YS |
2213 | |
2214 | /* | |
2215 | * Writing 0 to dirty_writeback_interval will disable periodic writeback | |
2216 | * and a different non-zero value will wakeup the writeback threads. | |
2217 | * wb_wakeup_delayed() would be more appropriate, but it's a pain to | |
2218 | * iterate over all bdis and wbs. | |
2219 | * The reason we do this is to make the change take effect immediately. | |
2220 | */ | |
2221 | if (!ret && write && dirty_writeback_interval && | |
2222 | dirty_writeback_interval != old_interval) | |
94af5846 YS |
2223 | wakeup_flusher_threads(WB_REASON_PERIODIC); |
2224 | ||
2225 | return ret; | |
1da177e4 | 2226 | } |
aa779e51 | 2227 | #endif |
1da177e4 | 2228 | |
bca237a5 | 2229 | void laptop_mode_timer_fn(struct timer_list *t) |
1da177e4 | 2230 | { |
bca237a5 KC |
2231 | struct backing_dev_info *backing_dev_info = |
2232 | from_timer(backing_dev_info, t, laptop_mode_wb_timer); | |
1da177e4 | 2233 | |
bca237a5 | 2234 | wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER); |
1da177e4 LT |
2235 | } |
2236 | ||
2237 | /* | |
2238 | * We've spun up the disk and we're in laptop mode: schedule writeback | |
2239 | * of all dirty data a few seconds from now. If the flush is already scheduled | |
2240 | * then push it back - the user is still using the disk. | |
2241 | */ | |
31373d09 | 2242 | void laptop_io_completion(struct backing_dev_info *info) |
1da177e4 | 2243 | { |
31373d09 | 2244 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
1da177e4 LT |
2245 | } |
2246 | ||
2247 | /* | |
2248 | * We're in laptop mode and we've just synced. The sync's writes will have | |
2249 | * caused another writeback to be scheduled by laptop_io_completion. | |
2250 | * Nothing needs to be written back anymore, so we unschedule the writeback. | |
2251 | */ | |
2252 | void laptop_sync_completion(void) | |
2253 | { | |
31373d09 MG |
2254 | struct backing_dev_info *bdi; |
2255 | ||
2256 | rcu_read_lock(); | |
2257 | ||
2258 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) | |
2259 | del_timer(&bdi->laptop_mode_wb_timer); | |
2260 | ||
2261 | rcu_read_unlock(); | |
1da177e4 LT |
2262 | } |
2263 | ||
2264 | /* | |
2265 | * If ratelimit_pages is too high then we can get into dirty-data overload | |
2266 | * if a large number of processes all perform writes at the same time. | |
1da177e4 LT |
2267 | * |
2268 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are | |
2269 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory | |
9d823e8f | 2270 | * thresholds. |
1da177e4 LT |
2271 | */ |
2272 | ||
2d1d43f6 | 2273 | void writeback_set_ratelimit(void) |
1da177e4 | 2274 | { |
dcc25ae7 | 2275 | struct wb_domain *dom = &global_wb_domain; |
9d823e8f WF |
2276 | unsigned long background_thresh; |
2277 | unsigned long dirty_thresh; | |
dcc25ae7 | 2278 | |
9d823e8f | 2279 | global_dirty_limits(&background_thresh, &dirty_thresh); |
dcc25ae7 | 2280 | dom->dirty_limit = dirty_thresh; |
9d823e8f | 2281 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); |
1da177e4 LT |
2282 | if (ratelimit_pages < 16) |
2283 | ratelimit_pages = 16; | |
1da177e4 LT |
2284 | } |
2285 | ||
1d7ac6ae | 2286 | static int page_writeback_cpu_online(unsigned int cpu) |
1da177e4 | 2287 | { |
1d7ac6ae SAS |
2288 | writeback_set_ratelimit(); |
2289 | return 0; | |
1da177e4 LT |
2290 | } |
2291 | ||
aa779e51 | 2292 | #ifdef CONFIG_SYSCTL |
3c6a4cba LC |
2293 | |
2294 | /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */ | |
2295 | static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE; | |
2296 | ||
aa779e51 | 2297 | static struct ctl_table vm_page_writeback_sysctls[] = { |
2298 | { | |
2299 | .procname = "dirty_background_ratio", | |
2300 | .data = &dirty_background_ratio, | |
2301 | .maxlen = sizeof(dirty_background_ratio), | |
2302 | .mode = 0644, | |
2303 | .proc_handler = dirty_background_ratio_handler, | |
2304 | .extra1 = SYSCTL_ZERO, | |
2305 | .extra2 = SYSCTL_ONE_HUNDRED, | |
2306 | }, | |
2307 | { | |
2308 | .procname = "dirty_background_bytes", | |
2309 | .data = &dirty_background_bytes, | |
2310 | .maxlen = sizeof(dirty_background_bytes), | |
2311 | .mode = 0644, | |
2312 | .proc_handler = dirty_background_bytes_handler, | |
2313 | .extra1 = SYSCTL_LONG_ONE, | |
2314 | }, | |
2315 | { | |
2316 | .procname = "dirty_ratio", | |
2317 | .data = &vm_dirty_ratio, | |
2318 | .maxlen = sizeof(vm_dirty_ratio), | |
2319 | .mode = 0644, | |
2320 | .proc_handler = dirty_ratio_handler, | |
2321 | .extra1 = SYSCTL_ZERO, | |
2322 | .extra2 = SYSCTL_ONE_HUNDRED, | |
2323 | }, | |
2324 | { | |
2325 | .procname = "dirty_bytes", | |
2326 | .data = &vm_dirty_bytes, | |
2327 | .maxlen = sizeof(vm_dirty_bytes), | |
2328 | .mode = 0644, | |
2329 | .proc_handler = dirty_bytes_handler, | |
2330 | .extra1 = (void *)&dirty_bytes_min, | |
2331 | }, | |
2332 | { | |
2333 | .procname = "dirty_writeback_centisecs", | |
2334 | .data = &dirty_writeback_interval, | |
2335 | .maxlen = sizeof(dirty_writeback_interval), | |
2336 | .mode = 0644, | |
2337 | .proc_handler = dirty_writeback_centisecs_handler, | |
2338 | }, | |
2339 | { | |
2340 | .procname = "dirty_expire_centisecs", | |
2341 | .data = &dirty_expire_interval, | |
2342 | .maxlen = sizeof(dirty_expire_interval), | |
2343 | .mode = 0644, | |
2344 | .proc_handler = proc_dointvec_minmax, | |
2345 | .extra1 = SYSCTL_ZERO, | |
2346 | }, | |
2347 | #ifdef CONFIG_HIGHMEM | |
2348 | { | |
2349 | .procname = "highmem_is_dirtyable", | |
2350 | .data = &vm_highmem_is_dirtyable, | |
2351 | .maxlen = sizeof(vm_highmem_is_dirtyable), | |
2352 | .mode = 0644, | |
2353 | .proc_handler = proc_dointvec_minmax, | |
2354 | .extra1 = SYSCTL_ZERO, | |
2355 | .extra2 = SYSCTL_ONE, | |
2356 | }, | |
2357 | #endif | |
2358 | { | |
2359 | .procname = "laptop_mode", | |
2360 | .data = &laptop_mode, | |
2361 | .maxlen = sizeof(laptop_mode), | |
2362 | .mode = 0644, | |
2363 | .proc_handler = proc_dointvec_jiffies, | |
2364 | }, | |
aa779e51 | 2365 | }; |
2366 | #endif | |
2367 | ||
1da177e4 | 2368 | /* |
dc6e29da LT |
2369 | * Called early on to tune the page writeback dirty limits. |
2370 | * | |
2371 | * We used to scale dirty pages according to how total memory | |
0a18e607 | 2372 | * related to pages that could be allocated for buffers. |
dc6e29da LT |
2373 | * |
2374 | * However, that was when we used "dirty_ratio" to scale with | |
2375 | * all memory, and we don't do that any more. "dirty_ratio" | |
0a18e607 | 2376 | * is now applied to total non-HIGHPAGE memory, and as such we can't |
dc6e29da LT |
2377 | * get into the old insane situation any more where we had |
2378 | * large amounts of dirty pages compared to a small amount of | |
2379 | * non-HIGHMEM memory. | |
2380 | * | |
2381 | * But we might still want to scale the dirty_ratio by how | |
2382 | * much memory the box has.. | |
1da177e4 LT |
2383 | */ |
2384 | void __init page_writeback_init(void) | |
2385 | { | |
a50fcb51 RV |
2386 | BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL)); |
2387 | ||
1d7ac6ae SAS |
2388 | cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online", |
2389 | page_writeback_cpu_online, NULL); | |
2390 | cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL, | |
2391 | page_writeback_cpu_online); | |
aa779e51 | 2392 | #ifdef CONFIG_SYSCTL |
2393 | register_sysctl_init("vm", vm_page_writeback_sysctls); | |
2394 | #endif | |
1da177e4 LT |
2395 | } |
2396 | ||
f446daae | 2397 | /** |
cdc150b5 | 2398 | * tag_pages_for_writeback - tag pages to be written by writeback |
f446daae JK |
2399 | * @mapping: address space structure to write |
2400 | * @start: starting page index | |
2401 | * @end: ending page index (inclusive) | |
2402 | * | |
2403 | * This function scans the page range from @start to @end (inclusive) and tags | |
cdc150b5 CH |
2404 | * all pages that have DIRTY tag set with a special TOWRITE tag. The caller |
2405 | * can then use the TOWRITE tag to identify pages eligible for writeback. | |
2406 | * This mechanism is used to avoid livelocking of writeback by a process | |
2407 | * steadily creating new dirty pages in the file (thus it is important for this | |
2408 | * function to be quick so that it can tag pages faster than a dirtying process | |
2409 | * can create them). | |
f446daae | 2410 | */ |
f446daae JK |
2411 | void tag_pages_for_writeback(struct address_space *mapping, |
2412 | pgoff_t start, pgoff_t end) | |
2413 | { | |
ff9c745b MW |
2414 | XA_STATE(xas, &mapping->i_pages, start); |
2415 | unsigned int tagged = 0; | |
2416 | void *page; | |
268f42de | 2417 | |
ff9c745b MW |
2418 | xas_lock_irq(&xas); |
2419 | xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) { | |
2420 | xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE); | |
2421 | if (++tagged % XA_CHECK_SCHED) | |
268f42de | 2422 | continue; |
ff9c745b MW |
2423 | |
2424 | xas_pause(&xas); | |
2425 | xas_unlock_irq(&xas); | |
f446daae | 2426 | cond_resched(); |
ff9c745b | 2427 | xas_lock_irq(&xas); |
268f42de | 2428 | } |
ff9c745b | 2429 | xas_unlock_irq(&xas); |
f446daae JK |
2430 | } |
2431 | EXPORT_SYMBOL(tag_pages_for_writeback); | |
2432 | ||
b1793929 MWO |
2433 | static bool folio_prepare_writeback(struct address_space *mapping, |
2434 | struct writeback_control *wbc, struct folio *folio) | |
2435 | { | |
2436 | /* | |
2437 | * Folio truncated or invalidated. We can freely skip it then, | |
2438 | * even for data integrity operations: the folio has disappeared | |
2439 | * concurrently, so there could be no real expectation of this | |
2440 | * data integrity operation even if there is now a new, dirty | |
2441 | * folio at the same pagecache index. | |
2442 | */ | |
2443 | if (unlikely(folio->mapping != mapping)) | |
2444 | return false; | |
2445 | ||
2446 | /* | |
2447 | * Did somebody else write it for us? | |
2448 | */ | |
2449 | if (!folio_test_dirty(folio)) | |
2450 | return false; | |
2451 | ||
2452 | if (folio_test_writeback(folio)) { | |
2453 | if (wbc->sync_mode == WB_SYNC_NONE) | |
2454 | return false; | |
2455 | folio_wait_writeback(folio); | |
2456 | } | |
2457 | BUG_ON(folio_test_writeback(folio)); | |
2458 | ||
2459 | if (!folio_clear_dirty_for_io(folio)) | |
2460 | return false; | |
2461 | ||
2462 | return true; | |
2463 | } | |
2464 | ||
751e0d55 MWO |
2465 | static xa_mark_t wbc_to_tag(struct writeback_control *wbc) |
2466 | { | |
2467 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) | |
2468 | return PAGECACHE_TAG_TOWRITE; | |
2469 | return PAGECACHE_TAG_DIRTY; | |
2470 | } | |
2471 | ||
2472 | static pgoff_t wbc_end(struct writeback_control *wbc) | |
2473 | { | |
2474 | if (wbc->range_cyclic) | |
2475 | return -1; | |
2476 | return wbc->range_end >> PAGE_SHIFT; | |
2477 | } | |
2478 | ||
e6d0ab87 | 2479 | static struct folio *writeback_get_folio(struct address_space *mapping, |
751e0d55 MWO |
2480 | struct writeback_control *wbc) |
2481 | { | |
e6d0ab87 MWO |
2482 | struct folio *folio; |
2483 | ||
a2cbc136 | 2484 | retry: |
e6d0ab87 MWO |
2485 | folio = folio_batch_next(&wbc->fbatch); |
2486 | if (!folio) { | |
2487 | folio_batch_release(&wbc->fbatch); | |
2488 | cond_resched(); | |
2489 | filemap_get_folios_tag(mapping, &wbc->index, wbc_end(wbc), | |
2490 | wbc_to_tag(wbc), &wbc->fbatch); | |
2491 | folio = folio_batch_next(&wbc->fbatch); | |
a2cbc136 MWO |
2492 | if (!folio) |
2493 | return NULL; | |
2494 | } | |
2495 | ||
2496 | folio_lock(folio); | |
2497 | if (unlikely(!folio_prepare_writeback(mapping, wbc, folio))) { | |
2498 | folio_unlock(folio); | |
2499 | goto retry; | |
e6d0ab87 MWO |
2500 | } |
2501 | ||
a2cbc136 | 2502 | trace_wbc_writepage(wbc, inode_to_bdi(mapping->host)); |
e6d0ab87 | 2503 | return folio; |
751e0d55 MWO |
2504 | } |
2505 | ||
811d736f | 2506 | /** |
cdc150b5 | 2507 | * writeback_iter - iterate folio of a mapping for writeback |
811d736f | 2508 | * @mapping: address space structure to write |
cdc150b5 CH |
2509 | * @wbc: writeback context |
2510 | * @folio: previously iterated folio (%NULL to start) | |
2511 | * @error: in-out pointer for writeback errors (see below) | |
811d736f | 2512 | * |
cdc150b5 CH |
2513 | * This function returns the next folio for the writeback operation described by |
2514 | * @wbc on @mapping and should be called in a while loop in the ->writepages | |
2515 | * implementation. | |
f446daae | 2516 | * |
cdc150b5 CH |
2517 | * To start the writeback operation, %NULL is passed in the @folio argument, and |
2518 | * for every subsequent iteration the folio returned previously should be passed | |
2519 | * back in. | |
64081362 | 2520 | * |
cdc150b5 CH |
2521 | * If there was an error in the per-folio writeback inside the writeback_iter() |
2522 | * loop, @error should be set to the error value. | |
a862f68a | 2523 | * |
cdc150b5 CH |
2524 | * Once the writeback described in @wbc has finished, this function will return |
2525 | * %NULL and if there was an error in any iteration restore it to @error. | |
2526 | * | |
2527 | * Note: callers should not manually break out of the loop using break or goto | |
2528 | * but must keep calling writeback_iter() until it returns %NULL. | |
2529 | * | |
2530 | * Return: the folio to write or %NULL if the loop is done. | |
811d736f | 2531 | */ |
cdc150b5 CH |
2532 | struct folio *writeback_iter(struct address_space *mapping, |
2533 | struct writeback_control *wbc, struct folio *folio, int *error) | |
811d736f | 2534 | { |
cdc150b5 CH |
2535 | if (!folio) { |
2536 | folio_batch_init(&wbc->fbatch); | |
2537 | wbc->saved_err = *error = 0; | |
751e0d55 | 2538 | |
cdc150b5 CH |
2539 | /* |
2540 | * For range cyclic writeback we remember where we stopped so | |
2541 | * that we can continue where we stopped. | |
2542 | * | |
2543 | * For non-cyclic writeback we always start at the beginning of | |
2544 | * the passed in range. | |
2545 | */ | |
2546 | if (wbc->range_cyclic) | |
2547 | wbc->index = mapping->writeback_index; | |
2548 | else | |
2549 | wbc->index = wbc->range_start >> PAGE_SHIFT; | |
811d736f | 2550 | |
cdc150b5 CH |
2551 | /* |
2552 | * To avoid livelocks when other processes dirty new pages, we | |
2553 | * first tag pages which should be written back and only then | |
2554 | * start writing them. | |
2555 | * | |
2556 | * For data-integrity writeback we have to be careful so that we | |
2557 | * do not miss some pages (e.g., because some other process has | |
2558 | * cleared the TOWRITE tag we set). The rule we follow is that | |
2559 | * TOWRITE tag can be cleared only by the process clearing the | |
2560 | * DIRTY tag (and submitting the page for I/O). | |
2561 | */ | |
2562 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) | |
2563 | tag_pages_for_writeback(mapping, wbc->index, | |
2564 | wbc_end(wbc)); | |
2565 | } else { | |
807d1fe3 | 2566 | wbc->nr_to_write -= folio_nr_pages(folio); |
f946e0d2 | 2567 | |
cdc150b5 | 2568 | WARN_ON_ONCE(*error > 0); |
00266770 | 2569 | |
807d1fe3 MWO |
2570 | /* |
2571 | * For integrity writeback we have to keep going until we have | |
2572 | * written all the folios we tagged for writeback above, even if | |
2573 | * we run past wbc->nr_to_write or encounter errors. | |
2574 | * We stash away the first error we encounter in wbc->saved_err | |
2575 | * so that it can be retrieved when we're done. This is because | |
2576 | * the file system may still have state to clear for each folio. | |
2577 | * | |
2578 | * For background writeback we exit as soon as we run past | |
2579 | * wbc->nr_to_write or encounter the first error. | |
2580 | */ | |
2581 | if (wbc->sync_mode == WB_SYNC_ALL) { | |
cdc150b5 CH |
2582 | if (*error && !wbc->saved_err) |
2583 | wbc->saved_err = *error; | |
807d1fe3 | 2584 | } else { |
cdc150b5 | 2585 | if (*error || wbc->nr_to_write <= 0) |
807d1fe3 | 2586 | goto done; |
811d736f | 2587 | } |
811d736f | 2588 | } |
64081362 | 2589 | |
cdc150b5 CH |
2590 | folio = writeback_get_folio(mapping, wbc); |
2591 | if (!folio) { | |
2592 | /* | |
2593 | * To avoid deadlocks between range_cyclic writeback and callers | |
2594 | * that hold pages in PageWriteback to aggregate I/O until | |
2595 | * the writeback iteration finishes, we do not loop back to the | |
2596 | * start of the file. Doing so causes a page lock/page | |
2597 | * writeback access order inversion - we should only ever lock | |
2598 | * multiple pages in ascending page->index order, and looping | |
2599 | * back to the start of the file violates that rule and causes | |
2600 | * deadlocks. | |
2601 | */ | |
2602 | if (wbc->range_cyclic) | |
2603 | mapping->writeback_index = 0; | |
2604 | ||
2605 | /* | |
2606 | * Return the first error we encountered (if there was any) to | |
2607 | * the caller. | |
2608 | */ | |
2609 | *error = wbc->saved_err; | |
2610 | } | |
2611 | return folio; | |
f946e0d2 CH |
2612 | |
2613 | done: | |
2614 | if (wbc->range_cyclic) | |
0692fad5 | 2615 | mapping->writeback_index = folio_next_index(folio); |
751e0d55 | 2616 | folio_batch_release(&wbc->fbatch); |
cdc150b5 CH |
2617 | return NULL; |
2618 | } | |
5a550a0c | 2619 | EXPORT_SYMBOL_GPL(writeback_iter); |
cdc150b5 CH |
2620 | |
2621 | /** | |
2622 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. | |
2623 | * @mapping: address space structure to write | |
2624 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write | |
2625 | * @writepage: function called for each page | |
2626 | * @data: data passed to writepage function | |
2627 | * | |
2628 | * Return: %0 on success, negative error code otherwise | |
2629 | * | |
2630 | * Note: please use writeback_iter() instead. | |
2631 | */ | |
2632 | int write_cache_pages(struct address_space *mapping, | |
2633 | struct writeback_control *wbc, writepage_t writepage, | |
2634 | void *data) | |
2635 | { | |
2636 | struct folio *folio = NULL; | |
2637 | int error; | |
2638 | ||
2639 | while ((folio = writeback_iter(mapping, wbc, folio, &error))) { | |
2640 | error = writepage(folio, wbc, data); | |
2641 | if (error == AOP_WRITEPAGE_ACTIVATE) { | |
2642 | folio_unlock(folio); | |
2643 | error = 0; | |
2644 | } | |
2645 | } | |
2646 | ||
f946e0d2 | 2647 | return error; |
811d736f | 2648 | } |
0ea97180 MS |
2649 | EXPORT_SYMBOL(write_cache_pages); |
2650 | ||
c44ed5b7 MWO |
2651 | static int writeback_use_writepage(struct address_space *mapping, |
2652 | struct writeback_control *wbc) | |
0ea97180 | 2653 | { |
c44ed5b7 MWO |
2654 | struct folio *folio = NULL; |
2655 | struct blk_plug plug; | |
2656 | int err; | |
6768907e | 2657 | |
c44ed5b7 MWO |
2658 | blk_start_plug(&plug); |
2659 | while ((folio = writeback_iter(mapping, wbc, folio, &err))) { | |
2660 | err = mapping->a_ops->writepage(&folio->page, wbc); | |
2661 | if (err == AOP_WRITEPAGE_ACTIVATE) { | |
2662 | folio_unlock(folio); | |
2663 | err = 0; | |
2664 | } | |
2665 | mapping_set_error(mapping, err); | |
2666 | } | |
2667 | blk_finish_plug(&plug); | |
2668 | ||
2669 | return err; | |
0ea97180 MS |
2670 | } |
2671 | ||
1da177e4 LT |
2672 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
2673 | { | |
22905f77 | 2674 | int ret; |
fee468fd | 2675 | struct bdi_writeback *wb; |
22905f77 | 2676 | |
1da177e4 LT |
2677 | if (wbc->nr_to_write <= 0) |
2678 | return 0; | |
fee468fd JK |
2679 | wb = inode_to_wb_wbc(mapping->host, wbc); |
2680 | wb_bandwidth_estimate_start(wb); | |
80a2ea9f | 2681 | while (1) { |
c2ca7a59 | 2682 | if (mapping->a_ops->writepages) { |
80a2ea9f | 2683 | ret = mapping->a_ops->writepages(mapping, wbc); |
c2ca7a59 | 2684 | } else if (mapping->a_ops->writepage) { |
c44ed5b7 | 2685 | ret = writeback_use_writepage(mapping, wbc); |
c2ca7a59 CH |
2686 | } else { |
2687 | /* deal with chardevs and other special files */ | |
2688 | ret = 0; | |
2689 | } | |
2690 | if (ret != -ENOMEM || wbc->sync_mode != WB_SYNC_ALL) | |
80a2ea9f | 2691 | break; |
8d58802f MG |
2692 | |
2693 | /* | |
2694 | * Lacking an allocation context or the locality or writeback | |
2695 | * state of any of the inode's pages, throttle based on | |
2696 | * writeback activity on the local node. It's as good a | |
2697 | * guess as any. | |
2698 | */ | |
2699 | reclaim_throttle(NODE_DATA(numa_node_id()), | |
c3f4a9a2 | 2700 | VMSCAN_THROTTLE_WRITEBACK); |
80a2ea9f | 2701 | } |
45a2966f JK |
2702 | /* |
2703 | * Usually few pages are written by now from those we've just submitted | |
2704 | * but if there's constant writeback being submitted, this makes sure | |
2705 | * writeback bandwidth is updated once in a while. | |
2706 | */ | |
20792ebf JK |
2707 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) + |
2708 | BANDWIDTH_INTERVAL)) | |
45a2966f | 2709 | wb_update_bandwidth(wb); |
22905f77 | 2710 | return ret; |
1da177e4 LT |
2711 | } |
2712 | ||
76719325 KC |
2713 | /* |
2714 | * For address_spaces which do not use buffers nor write back. | |
2715 | */ | |
46de8b97 | 2716 | bool noop_dirty_folio(struct address_space *mapping, struct folio *folio) |
76719325 | 2717 | { |
46de8b97 MWO |
2718 | if (!folio_test_dirty(folio)) |
2719 | return !folio_test_set_dirty(folio); | |
2720 | return false; | |
76719325 | 2721 | } |
46de8b97 | 2722 | EXPORT_SYMBOL(noop_dirty_folio); |
76719325 | 2723 | |
e3a7cca1 ES |
2724 | /* |
2725 | * Helper function for set_page_dirty family. | |
c4843a75 | 2726 | * |
6c77b607 | 2727 | * Caller must hold folio_memcg_lock(). |
c4843a75 | 2728 | * |
e3a7cca1 ES |
2729 | * NOTE: This relies on being atomic wrt interrupts. |
2730 | */ | |
203a3151 | 2731 | static void folio_account_dirtied(struct folio *folio, |
6e1cae88 | 2732 | struct address_space *mapping) |
e3a7cca1 | 2733 | { |
52ebea74 TH |
2734 | struct inode *inode = mapping->host; |
2735 | ||
b9b0ff61 | 2736 | trace_writeback_dirty_folio(folio, mapping); |
9fb0a7da | 2737 | |
f56753ac | 2738 | if (mapping_can_writeback(mapping)) { |
52ebea74 | 2739 | struct bdi_writeback *wb; |
203a3151 | 2740 | long nr = folio_nr_pages(folio); |
de1414a6 | 2741 | |
9cfb816b | 2742 | inode_attach_wb(inode, folio); |
52ebea74 | 2743 | wb = inode_to_wb(inode); |
de1414a6 | 2744 | |
203a3151 MWO |
2745 | __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr); |
2746 | __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr); | |
2747 | __node_stat_mod_folio(folio, NR_DIRTIED, nr); | |
2748 | wb_stat_mod(wb, WB_RECLAIMABLE, nr); | |
2749 | wb_stat_mod(wb, WB_DIRTIED, nr); | |
2750 | task_io_account_write(nr * PAGE_SIZE); | |
2751 | current->nr_dirtied += nr; | |
2752 | __this_cpu_add(bdp_ratelimits, nr); | |
97b27821 | 2753 | |
203a3151 | 2754 | mem_cgroup_track_foreign_dirty(folio, wb); |
e3a7cca1 ES |
2755 | } |
2756 | } | |
2757 | ||
b9ea2515 KK |
2758 | /* |
2759 | * Helper function for deaccounting dirty page without writeback. | |
2760 | * | |
6c77b607 | 2761 | * Caller must hold folio_memcg_lock(). |
b9ea2515 | 2762 | */ |
566d3362 | 2763 | void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb) |
b9ea2515 | 2764 | { |
566d3362 HD |
2765 | long nr = folio_nr_pages(folio); |
2766 | ||
2767 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr); | |
2768 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); | |
2769 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr); | |
2770 | task_io_account_cancelled_write(nr * PAGE_SIZE); | |
b9ea2515 | 2771 | } |
b9ea2515 | 2772 | |
6e1cae88 | 2773 | /* |
dc6e0ae5 | 2774 | * Mark the folio dirty, and set it dirty in the page cache. |
6e1cae88 | 2775 | * |
203a3151 | 2776 | * If warn is true, then emit a warning if the folio is not uptodate and has |
6e1cae88 MWO |
2777 | * not been truncated. |
2778 | * | |
3d84d897 MWO |
2779 | * The caller must hold folio_memcg_lock(). It is the caller's |
2780 | * responsibility to prevent the folio from being truncated while | |
2781 | * this function is in progress, although it may have been truncated | |
2782 | * before this function is called. Most callers have the folio locked. | |
2783 | * A few have the folio blocked from truncation through other means (e.g. | |
2784 | * zap_vma_pages() has it mapped and is holding the page table lock). | |
2785 | * When called from mark_buffer_dirty(), the filesystem should hold a | |
2786 | * reference to the buffer_head that is being marked dirty, which causes | |
2787 | * try_to_free_buffers() to fail. | |
6e1cae88 | 2788 | */ |
203a3151 | 2789 | void __folio_mark_dirty(struct folio *folio, struct address_space *mapping, |
6e1cae88 MWO |
2790 | int warn) |
2791 | { | |
2792 | unsigned long flags; | |
2793 | ||
2794 | xa_lock_irqsave(&mapping->i_pages, flags); | |
203a3151 MWO |
2795 | if (folio->mapping) { /* Race with truncate? */ |
2796 | WARN_ON_ONCE(warn && !folio_test_uptodate(folio)); | |
2797 | folio_account_dirtied(folio, mapping); | |
2798 | __xa_set_mark(&mapping->i_pages, folio_index(folio), | |
6e1cae88 MWO |
2799 | PAGECACHE_TAG_DIRTY); |
2800 | } | |
2801 | xa_unlock_irqrestore(&mapping->i_pages, flags); | |
2802 | } | |
2803 | ||
85d4d2eb MWO |
2804 | /** |
2805 | * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads. | |
2806 | * @mapping: Address space this folio belongs to. | |
2807 | * @folio: Folio to be marked as dirty. | |
1da177e4 | 2808 | * |
85d4d2eb | 2809 | * Filesystems which do not use buffer heads should call this function |
ab428b4c | 2810 | * from their dirty_folio address space operation. It ignores the |
85d4d2eb MWO |
2811 | * contents of folio_get_private(), so if the filesystem marks individual |
2812 | * blocks as dirty, the filesystem should handle that itself. | |
1da177e4 | 2813 | * |
85d4d2eb MWO |
2814 | * This is also sometimes used by filesystems which use buffer_heads when |
2815 | * a single buffer is being dirtied: we want to set the folio dirty in | |
2816 | * that case, but not all the buffers. This is a "bottom-up" dirtying, | |
e621900a | 2817 | * whereas block_dirty_folio() is a "top-down" dirtying. |
85d4d2eb MWO |
2818 | * |
2819 | * The caller must ensure this doesn't race with truncation. Most will | |
2820 | * simply hold the folio lock, but e.g. zap_pte_range() calls with the | |
2821 | * folio mapped and the pte lock held, which also locks out truncation. | |
1da177e4 | 2822 | */ |
85d4d2eb | 2823 | bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio) |
1da177e4 | 2824 | { |
85d4d2eb MWO |
2825 | folio_memcg_lock(folio); |
2826 | if (folio_test_set_dirty(folio)) { | |
2827 | folio_memcg_unlock(folio); | |
2828 | return false; | |
2829 | } | |
1da177e4 | 2830 | |
85d4d2eb MWO |
2831 | __folio_mark_dirty(folio, mapping, !folio_test_private(folio)); |
2832 | folio_memcg_unlock(folio); | |
c4843a75 | 2833 | |
85d4d2eb MWO |
2834 | if (mapping->host) { |
2835 | /* !PageAnon && !swapper_space */ | |
2836 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); | |
1da177e4 | 2837 | } |
85d4d2eb | 2838 | return true; |
1da177e4 | 2839 | } |
85d4d2eb | 2840 | EXPORT_SYMBOL(filemap_dirty_folio); |
1da177e4 | 2841 | |
cd78ab11 MWO |
2842 | /** |
2843 | * folio_redirty_for_writepage - Decline to write a dirty folio. | |
2844 | * @wbc: The writeback control. | |
2845 | * @folio: The folio. | |
2846 | * | |
2847 | * When a writepage implementation decides that it doesn't want to write | |
2848 | * @folio for some reason, it should call this function, unlock @folio and | |
2849 | * return 0. | |
2850 | * | |
2851 | * Return: True if we redirtied the folio. False if someone else dirtied | |
2852 | * it first. | |
1da177e4 | 2853 | */ |
cd78ab11 MWO |
2854 | bool folio_redirty_for_writepage(struct writeback_control *wbc, |
2855 | struct folio *folio) | |
1da177e4 | 2856 | { |
ed2da924 | 2857 | struct address_space *mapping = folio->mapping; |
cd78ab11 | 2858 | long nr = folio_nr_pages(folio); |
ed2da924 | 2859 | bool ret; |
cd78ab11 MWO |
2860 | |
2861 | wbc->pages_skipped += nr; | |
ed2da924 CH |
2862 | ret = filemap_dirty_folio(mapping, folio); |
2863 | if (mapping && mapping_can_writeback(mapping)) { | |
2864 | struct inode *inode = mapping->host; | |
2865 | struct bdi_writeback *wb; | |
2866 | struct wb_lock_cookie cookie = {}; | |
8d38633c | 2867 | |
ed2da924 CH |
2868 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
2869 | current->nr_dirtied -= nr; | |
2870 | node_stat_mod_folio(folio, NR_DIRTIED, -nr); | |
2871 | wb_stat_mod(wb, WB_DIRTIED, -nr); | |
2872 | unlocked_inode_to_wb_end(inode, &cookie); | |
2873 | } | |
8d38633c | 2874 | return ret; |
1da177e4 | 2875 | } |
cd78ab11 | 2876 | EXPORT_SYMBOL(folio_redirty_for_writepage); |
1da177e4 | 2877 | |
b5e84594 MWO |
2878 | /** |
2879 | * folio_mark_dirty - Mark a folio as being modified. | |
2880 | * @folio: The folio. | |
6746aff7 | 2881 | * |
2ca456c2 MWO |
2882 | * The folio may not be truncated while this function is running. |
2883 | * Holding the folio lock is sufficient to prevent truncation, but some | |
2884 | * callers cannot acquire a sleeping lock. These callers instead hold | |
2885 | * the page table lock for a page table which contains at least one page | |
2886 | * in this folio. Truncation will block on the page table lock as it | |
2887 | * unmaps pages before removing the folio from its mapping. | |
b5e84594 MWO |
2888 | * |
2889 | * Return: True if the folio was newly dirtied, false if it was already dirty. | |
1da177e4 | 2890 | */ |
b5e84594 | 2891 | bool folio_mark_dirty(struct folio *folio) |
1da177e4 | 2892 | { |
b5e84594 | 2893 | struct address_space *mapping = folio_mapping(folio); |
1da177e4 LT |
2894 | |
2895 | if (likely(mapping)) { | |
278df9f4 | 2896 | /* |
5a9e3474 | 2897 | * readahead/folio_deactivate could remain |
6f31a5a2 MWO |
2898 | * PG_readahead/PG_reclaim due to race with folio_end_writeback |
2899 | * About readahead, if the folio is written, the flags would be | |
278df9f4 | 2900 | * reset. So no problem. |
5a9e3474 | 2901 | * About folio_deactivate, if the folio is redirtied, |
6f31a5a2 MWO |
2902 | * the flag will be reset. So no problem. but if the |
2903 | * folio is used by readahead it will confuse readahead | |
2904 | * and make it restart the size rampup process. But it's | |
2905 | * a trivial problem. | |
278df9f4 | 2906 | */ |
b5e84594 MWO |
2907 | if (folio_test_reclaim(folio)) |
2908 | folio_clear_reclaim(folio); | |
3a3bae50 | 2909 | return mapping->a_ops->dirty_folio(mapping, folio); |
4741c9fd | 2910 | } |
3a3bae50 MWO |
2911 | |
2912 | return noop_dirty_folio(mapping, folio); | |
1da177e4 | 2913 | } |
b5e84594 | 2914 | EXPORT_SYMBOL(folio_mark_dirty); |
1da177e4 LT |
2915 | |
2916 | /* | |
2917 | * set_page_dirty() is racy if the caller has no reference against | |
2918 | * page->mapping->host, and if the page is unlocked. This is because another | |
2919 | * CPU could truncate the page off the mapping and then free the mapping. | |
2920 | * | |
2921 | * Usually, the page _is_ locked, or the caller is a user-space process which | |
2922 | * holds a reference on the inode by having an open file. | |
2923 | * | |
2924 | * In other cases, the page should be locked before running set_page_dirty(). | |
2925 | */ | |
2926 | int set_page_dirty_lock(struct page *page) | |
2927 | { | |
2928 | int ret; | |
2929 | ||
7eaceacc | 2930 | lock_page(page); |
1da177e4 LT |
2931 | ret = set_page_dirty(page); |
2932 | unlock_page(page); | |
2933 | return ret; | |
2934 | } | |
2935 | EXPORT_SYMBOL(set_page_dirty_lock); | |
2936 | ||
11f81bec TH |
2937 | /* |
2938 | * This cancels just the dirty bit on the kernel page itself, it does NOT | |
2939 | * actually remove dirty bits on any mmap's that may be around. It also | |
2940 | * leaves the page tagged dirty, so any sync activity will still find it on | |
2941 | * the dirty lists, and in particular, clear_page_dirty_for_io() will still | |
2942 | * look at the dirty bits in the VM. | |
2943 | * | |
2944 | * Doing this should *normally* only ever be done when a page is truncated, | |
2945 | * and is not actually mapped anywhere at all. However, fs/buffer.c does | |
2946 | * this when it notices that somebody has cleaned out all the buffers on a | |
2947 | * page without actually doing it through the VM. Can you say "ext3 is | |
2948 | * horribly ugly"? Thought you could. | |
2949 | */ | |
fdaf532a | 2950 | void __folio_cancel_dirty(struct folio *folio) |
11f81bec | 2951 | { |
fdaf532a | 2952 | struct address_space *mapping = folio_mapping(folio); |
c4843a75 | 2953 | |
f56753ac | 2954 | if (mapping_can_writeback(mapping)) { |
682aa8e1 TH |
2955 | struct inode *inode = mapping->host; |
2956 | struct bdi_writeback *wb; | |
2e898e4c | 2957 | struct wb_lock_cookie cookie = {}; |
c4843a75 | 2958 | |
fdaf532a | 2959 | folio_memcg_lock(folio); |
2e898e4c | 2960 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
c4843a75 | 2961 | |
fdaf532a | 2962 | if (folio_test_clear_dirty(folio)) |
566d3362 | 2963 | folio_account_cleaned(folio, wb); |
c4843a75 | 2964 | |
2e898e4c | 2965 | unlocked_inode_to_wb_end(inode, &cookie); |
fdaf532a | 2966 | folio_memcg_unlock(folio); |
c4843a75 | 2967 | } else { |
fdaf532a | 2968 | folio_clear_dirty(folio); |
c4843a75 | 2969 | } |
11f81bec | 2970 | } |
fdaf532a | 2971 | EXPORT_SYMBOL(__folio_cancel_dirty); |
11f81bec | 2972 | |
1da177e4 | 2973 | /* |
9350f20a MWO |
2974 | * Clear a folio's dirty flag, while caring for dirty memory accounting. |
2975 | * Returns true if the folio was previously dirty. | |
1da177e4 | 2976 | * |
9350f20a MWO |
2977 | * This is for preparing to put the folio under writeout. We leave |
2978 | * the folio tagged as dirty in the xarray so that a concurrent | |
2979 | * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk. | |
2980 | * The ->writepage implementation will run either folio_start_writeback() | |
2981 | * or folio_mark_dirty(), at which stage we bring the folio's dirty flag | |
2982 | * and xarray dirty tag back into sync. | |
1da177e4 | 2983 | * |
9350f20a MWO |
2984 | * This incoherency between the folio's dirty flag and xarray tag is |
2985 | * unfortunate, but it only exists while the folio is locked. | |
1da177e4 | 2986 | */ |
9350f20a | 2987 | bool folio_clear_dirty_for_io(struct folio *folio) |
1da177e4 | 2988 | { |
9350f20a MWO |
2989 | struct address_space *mapping = folio_mapping(folio); |
2990 | bool ret = false; | |
1da177e4 | 2991 | |
9350f20a | 2992 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
79352894 | 2993 | |
f56753ac | 2994 | if (mapping && mapping_can_writeback(mapping)) { |
682aa8e1 TH |
2995 | struct inode *inode = mapping->host; |
2996 | struct bdi_writeback *wb; | |
2e898e4c | 2997 | struct wb_lock_cookie cookie = {}; |
682aa8e1 | 2998 | |
7658cc28 LT |
2999 | /* |
3000 | * Yes, Virginia, this is indeed insane. | |
3001 | * | |
3002 | * We use this sequence to make sure that | |
3003 | * (a) we account for dirty stats properly | |
3004 | * (b) we tell the low-level filesystem to | |
9350f20a | 3005 | * mark the whole folio dirty if it was |
7658cc28 | 3006 | * dirty in a pagetable. Only to then |
9350f20a | 3007 | * (c) clean the folio again and return 1 to |
7658cc28 LT |
3008 | * cause the writeback. |
3009 | * | |
3010 | * This way we avoid all nasty races with the | |
3011 | * dirty bit in multiple places and clearing | |
3012 | * them concurrently from different threads. | |
3013 | * | |
9350f20a | 3014 | * Note! Normally the "folio_mark_dirty(folio)" |
7658cc28 LT |
3015 | * has no effect on the actual dirty bit - since |
3016 | * that will already usually be set. But we | |
3017 | * need the side effects, and it can help us | |
3018 | * avoid races. | |
3019 | * | |
9350f20a | 3020 | * We basically use the folio "master dirty bit" |
7658cc28 LT |
3021 | * as a serialization point for all the different |
3022 | * threads doing their things. | |
7658cc28 | 3023 | */ |
9350f20a MWO |
3024 | if (folio_mkclean(folio)) |
3025 | folio_mark_dirty(folio); | |
79352894 NP |
3026 | /* |
3027 | * We carefully synchronise fault handlers against | |
9350f20a | 3028 | * installing a dirty pte and marking the folio dirty |
2d6d7f98 | 3029 | * at this point. We do this by having them hold the |
9350f20a | 3030 | * page lock while dirtying the folio, and folios are |
2d6d7f98 JW |
3031 | * always locked coming in here, so we get the desired |
3032 | * exclusion. | |
79352894 | 3033 | */ |
2e898e4c | 3034 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
9350f20a MWO |
3035 | if (folio_test_clear_dirty(folio)) { |
3036 | long nr = folio_nr_pages(folio); | |
3037 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr); | |
3038 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); | |
3039 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr); | |
3040 | ret = true; | |
1da177e4 | 3041 | } |
2e898e4c | 3042 | unlocked_inode_to_wb_end(inode, &cookie); |
c4843a75 | 3043 | return ret; |
1da177e4 | 3044 | } |
9350f20a | 3045 | return folio_test_clear_dirty(folio); |
1da177e4 | 3046 | } |
9350f20a | 3047 | EXPORT_SYMBOL(folio_clear_dirty_for_io); |
1da177e4 | 3048 | |
633a2abb JK |
3049 | static void wb_inode_writeback_start(struct bdi_writeback *wb) |
3050 | { | |
3051 | atomic_inc(&wb->writeback_inodes); | |
3052 | } | |
3053 | ||
3054 | static void wb_inode_writeback_end(struct bdi_writeback *wb) | |
3055 | { | |
f87904c0 | 3056 | unsigned long flags; |
633a2abb | 3057 | atomic_dec(&wb->writeback_inodes); |
45a2966f JK |
3058 | /* |
3059 | * Make sure estimate of writeback throughput gets updated after | |
3060 | * writeback completed. We delay the update by BANDWIDTH_INTERVAL | |
3061 | * (which is the interval other bandwidth updates use for batching) so | |
3062 | * that if multiple inodes end writeback at a similar time, they get | |
3063 | * batched into one bandwidth update. | |
3064 | */ | |
f87904c0 KK |
3065 | spin_lock_irqsave(&wb->work_lock, flags); |
3066 | if (test_bit(WB_registered, &wb->state)) | |
3067 | queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL); | |
3068 | spin_unlock_irqrestore(&wb->work_lock, flags); | |
633a2abb JK |
3069 | } |
3070 | ||
2580d554 | 3071 | bool __folio_end_writeback(struct folio *folio) |
1da177e4 | 3072 | { |
269ccca3 MWO |
3073 | long nr = folio_nr_pages(folio); |
3074 | struct address_space *mapping = folio_mapping(folio); | |
2580d554 | 3075 | bool ret; |
1da177e4 | 3076 | |
269ccca3 | 3077 | folio_memcg_lock(folio); |
371a096e | 3078 | if (mapping && mapping_use_writeback_tags(mapping)) { |
91018134 TH |
3079 | struct inode *inode = mapping->host; |
3080 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 LT |
3081 | unsigned long flags; |
3082 | ||
b93b0163 | 3083 | xa_lock_irqsave(&mapping->i_pages, flags); |
2580d554 | 3084 | ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback); |
7d0795d0 MWO |
3085 | __xa_clear_mark(&mapping->i_pages, folio_index(folio), |
3086 | PAGECACHE_TAG_WRITEBACK); | |
3087 | if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) { | |
3088 | struct bdi_writeback *wb = inode_to_wb(inode); | |
3089 | ||
3090 | wb_stat_mod(wb, WB_WRITEBACK, -nr); | |
3091 | __wb_writeout_add(wb, nr); | |
3092 | if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) | |
3093 | wb_inode_writeback_end(wb); | |
69cb51d1 | 3094 | } |
6c60d2b5 DC |
3095 | |
3096 | if (mapping->host && !mapping_tagged(mapping, | |
3097 | PAGECACHE_TAG_WRITEBACK)) | |
3098 | sb_clear_inode_writeback(mapping->host); | |
3099 | ||
b93b0163 | 3100 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
1da177e4 | 3101 | } else { |
2580d554 | 3102 | ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback); |
99b12e3d | 3103 | } |
7d0795d0 MWO |
3104 | |
3105 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr); | |
3106 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); | |
3107 | node_stat_mod_folio(folio, NR_WRITTEN, nr); | |
269ccca3 | 3108 | folio_memcg_unlock(folio); |
2580d554 MWO |
3109 | |
3110 | return ret; | |
1da177e4 LT |
3111 | } |
3112 | ||
b5612c36 | 3113 | void __folio_start_writeback(struct folio *folio, bool keep_write) |
1da177e4 | 3114 | { |
f143f1ea MWO |
3115 | long nr = folio_nr_pages(folio); |
3116 | struct address_space *mapping = folio_mapping(folio); | |
f143f1ea | 3117 | int access_ret; |
1da177e4 | 3118 | |
b5612c36 MWO |
3119 | VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio); |
3120 | ||
f143f1ea | 3121 | folio_memcg_lock(folio); |
371a096e | 3122 | if (mapping && mapping_use_writeback_tags(mapping)) { |
f143f1ea | 3123 | XA_STATE(xas, &mapping->i_pages, folio_index(folio)); |
91018134 TH |
3124 | struct inode *inode = mapping->host; |
3125 | struct backing_dev_info *bdi = inode_to_bdi(inode); | |
1da177e4 | 3126 | unsigned long flags; |
b5612c36 | 3127 | bool on_wblist; |
1da177e4 | 3128 | |
ff9c745b MW |
3129 | xas_lock_irqsave(&xas, flags); |
3130 | xas_load(&xas); | |
b5612c36 | 3131 | folio_test_set_writeback(folio); |
6c60d2b5 | 3132 | |
b5612c36 | 3133 | on_wblist = mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK); |
6c60d2b5 | 3134 | |
b5612c36 MWO |
3135 | xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK); |
3136 | if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) { | |
3137 | struct bdi_writeback *wb = inode_to_wb(inode); | |
6c60d2b5 | 3138 | |
b5612c36 MWO |
3139 | wb_stat_mod(wb, WB_WRITEBACK, nr); |
3140 | if (!on_wblist) | |
3141 | wb_inode_writeback_start(wb); | |
69cb51d1 | 3142 | } |
b5612c36 MWO |
3143 | |
3144 | /* | |
3145 | * We can come through here when swapping anonymous | |
3146 | * folios, so we don't necessarily have an inode to | |
3147 | * track for sync. | |
3148 | */ | |
3149 | if (mapping->host && !on_wblist) | |
3150 | sb_mark_inode_writeback(mapping->host); | |
f143f1ea | 3151 | if (!folio_test_dirty(folio)) |
ff9c745b | 3152 | xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY); |
1c8349a1 | 3153 | if (!keep_write) |
ff9c745b MW |
3154 | xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE); |
3155 | xas_unlock_irqrestore(&xas, flags); | |
1da177e4 | 3156 | } else { |
b5612c36 | 3157 | folio_test_set_writeback(folio); |
3a3c02ec | 3158 | } |
b5612c36 MWO |
3159 | |
3160 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr); | |
3161 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr); | |
f143f1ea | 3162 | folio_memcg_unlock(folio); |
b5612c36 | 3163 | |
f143f1ea | 3164 | access_ret = arch_make_folio_accessible(folio); |
f28d4363 CI |
3165 | /* |
3166 | * If writeback has been triggered on a page that cannot be made | |
3167 | * accessible, it is too late to recover here. | |
3168 | */ | |
f143f1ea | 3169 | VM_BUG_ON_FOLIO(access_ret != 0, folio); |
1da177e4 | 3170 | } |
f143f1ea | 3171 | EXPORT_SYMBOL(__folio_start_writeback); |
1da177e4 | 3172 | |
490e016f MWO |
3173 | /** |
3174 | * folio_wait_writeback - Wait for a folio to finish writeback. | |
3175 | * @folio: The folio to wait for. | |
3176 | * | |
3177 | * If the folio is currently being written back to storage, wait for the | |
3178 | * I/O to complete. | |
3179 | * | |
3180 | * Context: Sleeps. Must be called in process context and with | |
3181 | * no spinlocks held. Caller should hold a reference on the folio. | |
3182 | * If the folio is not locked, writeback may start again after writeback | |
3183 | * has finished. | |
19343b5b | 3184 | */ |
490e016f | 3185 | void folio_wait_writeback(struct folio *folio) |
19343b5b | 3186 | { |
490e016f | 3187 | while (folio_test_writeback(folio)) { |
b9b0ff61 | 3188 | trace_folio_wait_writeback(folio, folio_mapping(folio)); |
101c0bf6 | 3189 | folio_wait_bit(folio, PG_writeback); |
19343b5b YS |
3190 | } |
3191 | } | |
490e016f | 3192 | EXPORT_SYMBOL_GPL(folio_wait_writeback); |
19343b5b | 3193 | |
490e016f MWO |
3194 | /** |
3195 | * folio_wait_writeback_killable - Wait for a folio to finish writeback. | |
3196 | * @folio: The folio to wait for. | |
3197 | * | |
3198 | * If the folio is currently being written back to storage, wait for the | |
3199 | * I/O to complete or a fatal signal to arrive. | |
3200 | * | |
3201 | * Context: Sleeps. Must be called in process context and with | |
3202 | * no spinlocks held. Caller should hold a reference on the folio. | |
3203 | * If the folio is not locked, writeback may start again after writeback | |
3204 | * has finished. | |
3205 | * Return: 0 on success, -EINTR if we get a fatal signal while waiting. | |
e5dbd332 | 3206 | */ |
490e016f | 3207 | int folio_wait_writeback_killable(struct folio *folio) |
e5dbd332 | 3208 | { |
490e016f | 3209 | while (folio_test_writeback(folio)) { |
b9b0ff61 | 3210 | trace_folio_wait_writeback(folio, folio_mapping(folio)); |
101c0bf6 | 3211 | if (folio_wait_bit_killable(folio, PG_writeback)) |
e5dbd332 MWO |
3212 | return -EINTR; |
3213 | } | |
3214 | ||
3215 | return 0; | |
3216 | } | |
490e016f | 3217 | EXPORT_SYMBOL_GPL(folio_wait_writeback_killable); |
e5dbd332 | 3218 | |
1d1d1a76 | 3219 | /** |
a49d0c50 MWO |
3220 | * folio_wait_stable() - wait for writeback to finish, if necessary. |
3221 | * @folio: The folio to wait on. | |
1d1d1a76 | 3222 | * |
a49d0c50 MWO |
3223 | * This function determines if the given folio is related to a backing |
3224 | * device that requires folio contents to be held stable during writeback. | |
3225 | * If so, then it will wait for any pending writeback to complete. | |
3226 | * | |
3227 | * Context: Sleeps. Must be called in process context and with | |
3228 | * no spinlocks held. Caller should hold a reference on the folio. | |
3229 | * If the folio is not locked, writeback may start again after writeback | |
3230 | * has finished. | |
1d1d1a76 | 3231 | */ |
a49d0c50 | 3232 | void folio_wait_stable(struct folio *folio) |
1d1d1a76 | 3233 | { |
762321da | 3234 | if (mapping_stable_writes(folio_mapping(folio))) |
a49d0c50 | 3235 | folio_wait_writeback(folio); |
1d1d1a76 | 3236 | } |
a49d0c50 | 3237 | EXPORT_SYMBOL_GPL(folio_wait_stable); |