| 1 | |
| 2 | |
| 3 | * mm/page-writeback.c
|
| 4 | *
|
| 5 | * Copyright (C) 2002, Linus Torvalds.
|
| 6 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
|
| 7 | *
|
| 8 | * Contains functions related to writing back dirty pages at the
|
| 9 | * address_space level.
|
| 10 | *
|
| 11 | * 10Apr2002 Andrew Morton
|
| 12 | * Initial version
|
| 13 |
|
| 14 |
|
| 15 | #include <linux/kernel.h>
|
| 16 | #include <linux/math64.h>
|
| 17 | #include <linux/export.h>
|
| 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>
|
| 27 | #include <linux/task_io_accounting_ops.h>
|
| 28 | #include <linux/mpage.h>
|
| 29 | #include <linux/rmap.h>
|
| 30 | #include <linux/percpu.h>
|
| 31 | #include <linux/smp.h>
|
| 32 | #include <linux/sysctl.h>
|
| 33 | #include <linux/cpu.h>
|
| 34 | #include <linux/syscalls.h>
|
| 35 | #include <linux/folio_batch.h>
|
| 36 | #include <linux/timer.h>
|
| 37 | #include <linux/sched/rt.h>
|
| 38 | #include <linux/sched/signal.h>
|
| 39 | #include <linux/mm_inline.h>
|
| 40 | #include <linux/shmem_fs.h>
|
| 41 | #include <trace/events/writeback.h>
|
| 42 |
|
| 43 | #include "internal.h"
|
| 44 |
|
| 45 | |
| 46 | * Sleep at most 200ms at a time in balance_dirty_pages().
|
| 47 |
|
| 48 | #define MAX_PAUSE max(HZ/5, 1)
|
| 49 |
|
| 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 |
|
| 56 | |
| 57 | * Estimate write bandwidth or update dirty limit at 200ms intervals.
|
| 58 |
|
| 59 | #define BANDWIDTH_INTERVAL max(HZ/5, 1)
|
| 60 |
|
| 61 | #define RATELIMIT_CALC_SHIFT 10
|
| 62 |
|
| 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 |
|
| 69 |
|
| 70 |
|
| 71 | |
| 72 | * Start background writeback (via writeback threads) at this percentage
|
| 73 |
|
| 74 | static int dirty_background_ratio = 10;
|
| 75 |
|
| 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 |
|
| 80 | static unsigned long dirty_background_bytes;
|
| 81 |
|
| 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 |
|
| 86 | static int vm_highmem_is_dirtyable;
|
| 87 |
|
| 88 | |
| 89 | * The generator of dirty data starts writeback at this percentage
|
| 90 |
|
| 91 | static int vm_dirty_ratio = 20;
|
| 92 |
|
| 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 |
|
| 97 | static unsigned long vm_dirty_bytes;
|
| 98 |
|
| 99 | |
| 100 | * The interval between `kupdate'-style writebacks
|
| 101 |
|
| 102 | unsigned int dirty_writeback_interval = 5 * 100;
|
| 103 |
|
| 104 | EXPORT_SYMBOL_GPL(dirty_writeback_interval);
|
| 105 |
|
| 106 | |
| 107 | * The longest time for which data is allowed to remain dirty
|
| 108 |
|
| 109 | unsigned int dirty_expire_interval = 30 * 100;
|
| 110 |
|
| 111 |
|
| 112 |
|
| 113 | struct wb_domain global_wb_domain;
|
| 114 |
|
| 115 | |
| 116 | * Length of period for aging writeout fractions of bdis. This is an
|
| 117 | * arbitrarily chosen number. The longer the period, the slower fractions will
|
| 118 | * reflect changes in current writeout rate.
|
| 119 |
|
| 120 | #define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
|
| 121 |
|
| 122 | #ifdef CONFIG_CGROUP_WRITEBACK
|
| 123 |
|
| 124 | #define GDTC_INIT(__wb) .wb = (__wb), \
|
| 125 | .dom = &global_wb_domain, \
|
| 126 | .wb_completions = &(__wb)->completions
|
| 127 |
|
| 128 | #define GDTC_INIT_NO_WB .dom = &global_wb_domain
|
| 129 |
|
| 130 | #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
|
| 131 | .dom = mem_cgroup_wb_domain(__wb), \
|
| 132 | .wb_completions = &(__wb)->memcg_completions, \
|
| 133 | .gdtc = __gdtc
|
| 134 |
|
| 135 | static bool mdtc_valid(struct dirty_throttle_control *dtc)
|
| 136 | {
|
| 137 | return dtc->dom;
|
| 138 | }
|
| 139 |
|
| 140 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
|
| 141 | {
|
| 142 | return dtc->dom;
|
| 143 | }
|
| 144 |
|
| 145 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
|
| 146 | {
|
| 147 | return mdtc->gdtc;
|
| 148 | }
|
| 149 |
|
| 150 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
|
| 151 | {
|
| 152 | return &wb->memcg_completions;
|
| 153 | }
|
| 154 |
|
| 155 | static void wb_min_max_ratio(struct bdi_writeback *wb,
|
| 156 | unsigned long *minp, unsigned long *maxp)
|
| 157 | {
|
| 158 | unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth);
|
| 159 | unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
|
| 160 | unsigned long long min = wb->bdi->min_ratio;
|
| 161 | unsigned long long max = wb->bdi->max_ratio;
|
| 162 |
|
| 163 | |
| 164 | * @wb may already be clean by the time control reaches here and
|
| 165 | * the total may not include its bw.
|
| 166 |
|
| 167 | if (this_bw < tot_bw) {
|
| 168 | if (min) {
|
| 169 | min *= this_bw;
|
| 170 | min = div64_ul(min, tot_bw);
|
| 171 | }
|
| 172 | if (max < 100 * BDI_RATIO_SCALE) {
|
| 173 | max *= this_bw;
|
| 174 | max = div64_ul(max, tot_bw);
|
| 175 | }
|
| 176 | }
|
| 177 |
|
| 178 | *minp = min;
|
| 179 | *maxp = max;
|
| 180 | }
|
| 181 |
|
| 182 | #else
|
| 183 |
|
| 184 | #define GDTC_INIT(__wb) .wb = (__wb), \
|
| 185 | .wb_completions = &(__wb)->completions
|
| 186 | #define GDTC_INIT_NO_WB
|
| 187 | #define MDTC_INIT(__wb, __gdtc)
|
| 188 |
|
| 189 | static bool mdtc_valid(struct dirty_throttle_control *dtc)
|
| 190 | {
|
| 191 | return false;
|
| 192 | }
|
| 193 |
|
| 194 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
|
| 195 | {
|
| 196 | return &global_wb_domain;
|
| 197 | }
|
| 198 |
|
| 199 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
|
| 200 | {
|
| 201 | return NULL;
|
| 202 | }
|
| 203 |
|
| 204 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
|
| 205 | {
|
| 206 | return NULL;
|
| 207 | }
|
| 208 |
|
| 209 | static void wb_min_max_ratio(struct bdi_writeback *wb,
|
| 210 | unsigned long *minp, unsigned long *maxp)
|
| 211 | {
|
| 212 | *minp = wb->bdi->min_ratio;
|
| 213 | *maxp = wb->bdi->max_ratio;
|
| 214 | }
|
| 215 |
|
| 216 | #endif
|
| 217 |
|
| 218 | |
| 219 | * In a memory zone, there is a certain amount of pages we consider
|
| 220 | * available for the page cache, which is essentially the number of
|
| 221 | * free and reclaimable pages, minus some zone reserves to protect
|
| 222 | * lowmem and the ability to uphold the zone's watermarks without
|
| 223 | * requiring writeback.
|
| 224 | *
|
| 225 | * This number of dirtyable pages is the base value of which the
|
| 226 | * user-configurable dirty ratio is the effective number of pages that
|
| 227 | * are allowed to be actually dirtied. Per individual zone, or
|
| 228 | * globally by using the sum of dirtyable pages over all zones.
|
| 229 | *
|
| 230 | * Because the user is allowed to specify the dirty limit globally as
|
| 231 | * absolute number of bytes, calculating the per-zone dirty limit can
|
| 232 | * require translating the configured limit into a percentage of
|
| 233 | * global dirtyable memory first.
|
| 234 |
|
| 235 |
|
| 236 | |
| 237 | * node_dirtyable_memory - number of dirtyable pages in a node
|
| 238 | * @pgdat: the node
|
| 239 | *
|
| 240 | * Return: the node's number of pages potentially available for dirty
|
| 241 | * page cache. This is the base value for the per-node dirty limits.
|
| 242 |
|
| 243 | static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
|
| 244 | {
|
| 245 | unsigned long nr_pages = 0;
|
| 246 | int z;
|
| 247 |
|
| 248 | for (z = 0; z < MAX_NR_ZONES; z++) {
|
| 249 | struct zone *zone = pgdat->node_zones + z;
|
| 250 |
|
| 251 | if (!populated_zone(zone))
|
| 252 | continue;
|
| 253 |
|
| 254 | nr_pages += zone_page_state(zone, NR_FREE_PAGES);
|
| 255 | }
|
| 256 |
|
| 257 | |
| 258 | * Pages reserved for the kernel should not be considered
|
| 259 | * dirtyable, to prevent a situation where reclaim has to
|
| 260 | * clean pages in order to balance the zones.
|
| 261 |
|
| 262 | nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
|
| 263 |
|
| 264 | nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
|
| 265 | nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
|
| 266 |
|
| 267 | return nr_pages;
|
| 268 | }
|
| 269 |
|
| 270 | static unsigned long highmem_dirtyable_memory(unsigned long total)
|
| 271 | {
|
| 272 | #ifdef CONFIG_HIGHMEM
|
| 273 | int node;
|
| 274 | unsigned long x = 0;
|
| 275 | int i;
|
| 276 |
|
| 277 | for_each_node_state(node, N_HIGH_MEMORY) {
|
| 278 | for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
|
| 279 | struct zone *z;
|
| 280 | unsigned long nr_pages;
|
| 281 |
|
| 282 | if (!is_highmem_idx(i))
|
| 283 | continue;
|
| 284 |
|
| 285 | z = &NODE_DATA(node)->node_zones[i];
|
| 286 | if (!populated_zone(z))
|
| 287 | continue;
|
| 288 |
|
| 289 | nr_pages = zone_page_state(z, NR_FREE_PAGES);
|
| 290 |
|
| 291 | nr_pages -= min(nr_pages, high_wmark_pages(z));
|
| 292 | nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
|
| 293 | nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
|
| 294 | x += nr_pages;
|
| 295 | }
|
| 296 | }
|
| 297 |
|
| 298 | |
| 299 | * Make sure that the number of highmem pages is never larger
|
| 300 | * than the number of the total dirtyable memory. This can only
|
| 301 | * occur in very strange VM situations but we want to make sure
|
| 302 | * that this does not occur.
|
| 303 |
|
| 304 | return min(x, total);
|
| 305 | #else
|
| 306 | return 0;
|
| 307 | #endif
|
| 308 | }
|
| 309 |
|
| 310 | |
| 311 | * global_dirtyable_memory - number of globally dirtyable pages
|
| 312 | *
|
| 313 | * Return: the global number of pages potentially available for dirty
|
| 314 | * page cache. This is the base value for the global dirty limits.
|
| 315 |
|
| 316 | static unsigned long global_dirtyable_memory(void)
|
| 317 | {
|
| 318 | unsigned long x;
|
| 319 |
|
| 320 | x = global_zone_page_state(NR_FREE_PAGES);
|
| 321 | |
| 322 | * Pages reserved for the kernel should not be considered
|
| 323 | * dirtyable, to prevent a situation where reclaim has to
|
| 324 | * clean pages in order to balance the zones.
|
| 325 |
|
| 326 | x -= min(x, totalreserve_pages);
|
| 327 |
|
| 328 | x += global_node_page_state(NR_INACTIVE_FILE);
|
| 329 | x += global_node_page_state(NR_ACTIVE_FILE);
|
| 330 |
|
| 331 | if (!vm_highmem_is_dirtyable)
|
| 332 | x -= highmem_dirtyable_memory(x);
|
| 333 |
|
| 334 | return x + 1;
|
| 335 | }
|
| 336 |
|
| 337 | |
| 338 | * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
|
| 339 | * @dtc: dirty_throttle_control of interest
|
| 340 | *
|
| 341 | * Calculate @dtc->thresh and ->bg_thresh considering
|
| 342 | * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
|
| 343 | * must ensure that @dtc->avail is set before calling this function. The
|
| 344 | * dirty limits will be lifted by 1/4 for real-time tasks.
|
| 345 |
|
| 346 | static void domain_dirty_limits(struct dirty_throttle_control *dtc)
|
| 347 | {
|
| 348 | const unsigned long available_memory = dtc->avail;
|
| 349 | struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
|
| 350 | unsigned long bytes = vm_dirty_bytes;
|
| 351 | unsigned long bg_bytes = dirty_background_bytes;
|
| 352 |
|
| 353 | unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
|
| 354 | unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
|
| 355 | unsigned long thresh;
|
| 356 | unsigned long bg_thresh;
|
| 357 | struct task_struct *tsk;
|
| 358 |
|
| 359 |
|
| 360 | if (gdtc) {
|
| 361 | unsigned long global_avail = gdtc->avail;
|
| 362 |
|
| 363 | |
| 364 | * The byte settings can't be applied directly to memcg
|
| 365 | * domains. Convert them to ratios by scaling against
|
| 366 | * globally available memory. As the ratios are in
|
| 367 | * per-PAGE_SIZE, they can be obtained by dividing bytes by
|
| 368 | * number of pages.
|
| 369 |
|
| 370 | if (bytes)
|
| 371 | ratio = min(DIV_ROUND_UP(bytes, global_avail),
|
| 372 | PAGE_SIZE);
|
| 373 | if (bg_bytes)
|
| 374 | bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
|
| 375 | PAGE_SIZE);
|
| 376 | bytes = bg_bytes = 0;
|
| 377 | }
|
| 378 |
|
| 379 | if (bytes)
|
| 380 | thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
|
| 381 | else
|
| 382 | thresh = (ratio * available_memory) / PAGE_SIZE;
|
| 383 |
|
| 384 | if (bg_bytes)
|
| 385 | bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
|
| 386 | else
|
| 387 | bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
|
| 388 |
|
| 389 | tsk = current;
|
| 390 | if (rt_or_dl_task(tsk)) {
|
| 391 | bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
|
| 392 | thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
|
| 393 | }
|
| 394 | |
| 395 | * Dirty throttling logic assumes the limits in page units fit into
|
| 396 | * 32-bits. This gives 16TB dirty limits max which is hopefully enough.
|
| 397 |
|
| 398 | if (thresh > UINT_MAX)
|
| 399 | thresh = UINT_MAX;
|
| 400 |
|
| 401 | if (bg_thresh >= thresh)
|
| 402 | bg_thresh = thresh / 2;
|
| 403 | dtc->thresh = thresh;
|
| 404 | dtc->bg_thresh = bg_thresh;
|
| 405 |
|
| 406 |
|
| 407 | if (!gdtc)
|
| 408 | trace_global_dirty_state(bg_thresh, thresh);
|
| 409 | }
|
| 410 |
|
| 411 | |
| 412 | * global_dirty_limits - background-writeback and dirty-throttling thresholds
|
| 413 | * @pbackground: out parameter for bg_thresh
|
| 414 | * @pdirty: out parameter for thresh
|
| 415 | *
|
| 416 | * Calculate bg_thresh and thresh for global_wb_domain. See
|
| 417 | * domain_dirty_limits() for details.
|
| 418 |
|
| 419 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
|
| 420 | {
|
| 421 | struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
|
| 422 |
|
| 423 | gdtc.avail = global_dirtyable_memory();
|
| 424 | domain_dirty_limits(&gdtc);
|
| 425 |
|
| 426 | *pbackground = gdtc.bg_thresh;
|
| 427 | *pdirty = gdtc.thresh;
|
| 428 | }
|
| 429 |
|
| 430 | |
| 431 | * node_dirty_limit - maximum number of dirty pages allowed in a node
|
| 432 | * @pgdat: the node
|
| 433 | *
|
| 434 | * Return: the maximum number of dirty pages allowed in a node, based
|
| 435 | * on the node's dirtyable memory.
|
| 436 |
|
| 437 | static unsigned long node_dirty_limit(struct pglist_data *pgdat)
|
| 438 | {
|
| 439 | unsigned long node_memory = node_dirtyable_memory(pgdat);
|
| 440 | struct task_struct *tsk = current;
|
| 441 | unsigned long dirty;
|
| 442 |
|
| 443 | if (vm_dirty_bytes)
|
| 444 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
|
| 445 | node_memory / global_dirtyable_memory();
|
| 446 | else
|
| 447 | dirty = vm_dirty_ratio * node_memory / 100;
|
| 448 |
|
| 449 | if (rt_or_dl_task(tsk))
|
| 450 | dirty += dirty / 4;
|
| 451 |
|
| 452 | |
| 453 | * Dirty throttling logic assumes the limits in page units fit into
|
| 454 | * 32-bits. This gives 16TB dirty limits max which is hopefully enough.
|
| 455 |
|
| 456 | return min_t(unsigned long, dirty, UINT_MAX);
|
| 457 | }
|
| 458 |
|
| 459 | |
| 460 | * node_dirty_ok - tells whether a node is within its dirty limits
|
| 461 | * @pgdat: the node to check
|
| 462 | *
|
| 463 | * Return: %true when the dirty pages in @pgdat are within the node's
|
| 464 | * dirty limit, %false if the limit is exceeded.
|
| 465 |
|
| 466 | bool node_dirty_ok(struct pglist_data *pgdat)
|
| 467 | {
|
| 468 | unsigned long limit = node_dirty_limit(pgdat);
|
| 469 | unsigned long nr_pages = 0;
|
| 470 |
|
| 471 | nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
|
| 472 | nr_pages += node_page_state(pgdat, NR_WRITEBACK);
|
| 473 |
|
| 474 | return nr_pages <= limit;
|
| 475 | }
|
| 476 |
|
| 477 | #ifdef CONFIG_SYSCTL
|
| 478 | static int dirty_background_ratio_handler(const struct ctl_table *table, int write,
|
| 479 | void *buffer, size_t *lenp, loff_t *ppos)
|
| 480 | {
|
| 481 | int ret;
|
| 482 |
|
| 483 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
|
| 484 | if (ret == 0 && write)
|
| 485 | dirty_background_bytes = 0;
|
| 486 | return ret;
|
| 487 | }
|
| 488 |
|
| 489 | static int dirty_background_bytes_handler(const struct ctl_table *table, int write,
|
| 490 | void *buffer, size_t *lenp, loff_t *ppos)
|
| 491 | {
|
| 492 | int ret;
|
| 493 | unsigned long old_bytes = dirty_background_bytes;
|
| 494 |
|
| 495 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
|
| 496 | if (ret == 0 && write) {
|
| 497 | if (DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE) >
|
| 498 | UINT_MAX) {
|
| 499 | dirty_background_bytes = old_bytes;
|
| 500 | return -ERANGE;
|
| 501 | }
|
| 502 | dirty_background_ratio = 0;
|
| 503 | }
|
| 504 | return ret;
|
| 505 | }
|
| 506 |
|
| 507 | static int dirty_ratio_handler(const struct ctl_table *table, int write, void *buffer,
|
| 508 | size_t *lenp, loff_t *ppos)
|
| 509 | {
|
| 510 | int old_ratio = vm_dirty_ratio;
|
| 511 | int ret;
|
| 512 |
|
| 513 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
|
| 514 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
|
| 515 | vm_dirty_bytes = 0;
|
| 516 | writeback_set_ratelimit();
|
| 517 | }
|
| 518 | return ret;
|
| 519 | }
|
| 520 |
|
| 521 | static int dirty_bytes_handler(const struct ctl_table *table, int write,
|
| 522 | void *buffer, size_t *lenp, loff_t *ppos)
|
| 523 | {
|
| 524 | unsigned long old_bytes = vm_dirty_bytes;
|
| 525 | int ret;
|
| 526 |
|
| 527 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
|
| 528 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
|
| 529 | if (DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) > UINT_MAX) {
|
| 530 | vm_dirty_bytes = old_bytes;
|
| 531 | return -ERANGE;
|
| 532 | }
|
| 533 | writeback_set_ratelimit();
|
| 534 | vm_dirty_ratio = 0;
|
| 535 | }
|
| 536 | return ret;
|
| 537 | }
|
| 538 | #endif
|
| 539 |
|
| 540 | static unsigned long wp_next_time(unsigned long cur_time)
|
| 541 | {
|
| 542 | cur_time += VM_COMPLETIONS_PERIOD_LEN;
|
| 543 |
|
| 544 | if (!cur_time)
|
| 545 | return 1;
|
| 546 | return cur_time;
|
| 547 | }
|
| 548 |
|
| 549 | static void wb_domain_writeout_add(struct wb_domain *dom,
|
| 550 | struct fprop_local_percpu *completions,
|
| 551 | unsigned int max_prop_frac, long nr)
|
| 552 | {
|
| 553 | __fprop_add_percpu_max(&dom->completions, completions,
|
| 554 | max_prop_frac, nr);
|
| 555 |
|
| 556 | if (unlikely(!dom->period_time)) {
|
| 557 | |
| 558 | * We can race with other wb_domain_writeout_add calls here but
|
| 559 | * it does not cause any harm since the resulting time when
|
| 560 | * timer will fire and what is in writeout_period_time will be
|
| 561 | * roughly the same.
|
| 562 |
|
| 563 | dom->period_time = wp_next_time(jiffies);
|
| 564 | mod_timer(&dom->period_timer, dom->period_time);
|
| 565 | }
|
| 566 | }
|
| 567 |
|
| 568 | |
| 569 | * Increment @wb's writeout completion count and the global writeout
|
| 570 | * completion count. Called from __folio_end_writeback().
|
| 571 |
|
| 572 | static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr)
|
| 573 | {
|
| 574 | struct wb_domain *cgdom;
|
| 575 |
|
| 576 | wb_stat_mod(wb, WB_WRITTEN, nr);
|
| 577 | wb_domain_writeout_add(&global_wb_domain, &wb->completions,
|
| 578 | wb->bdi->max_prop_frac, nr);
|
| 579 |
|
| 580 | cgdom = mem_cgroup_wb_domain(wb);
|
| 581 | if (cgdom)
|
| 582 | wb_domain_writeout_add(cgdom, wb_memcg_completions(wb),
|
| 583 | wb->bdi->max_prop_frac, nr);
|
| 584 | }
|
| 585 |
|
| 586 | |
| 587 | * On idle system, we can be called long after we scheduled because we use
|
| 588 | * deferred timers so count with missed periods.
|
| 589 |
|
| 590 | static void writeout_period(struct timer_list *t)
|
| 591 | {
|
| 592 | struct wb_domain *dom = timer_container_of(dom, t, period_timer);
|
| 593 | int miss_periods = (jiffies - dom->period_time) /
|
| 594 | VM_COMPLETIONS_PERIOD_LEN;
|
| 595 |
|
| 596 | if (fprop_new_period(&dom->completions, miss_periods + 1)) {
|
| 597 | dom->period_time = wp_next_time(dom->period_time +
|
| 598 | miss_periods * VM_COMPLETIONS_PERIOD_LEN);
|
| 599 | mod_timer(&dom->period_timer, dom->period_time);
|
| 600 | } else {
|
| 601 | |
| 602 | * Aging has zeroed all fractions. Stop wasting CPU on period
|
| 603 | * updates.
|
| 604 |
|
| 605 | dom->period_time = 0;
|
| 606 | }
|
| 607 | }
|
| 608 |
|
| 609 | int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
|
| 610 | {
|
| 611 | memset(dom, 0, sizeof(*dom));
|
| 612 |
|
| 613 | spin_lock_init(&dom->lock);
|
| 614 |
|
| 615 | timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
|
| 616 |
|
| 617 | dom->dirty_limit_tstamp = jiffies;
|
| 618 |
|
| 619 | return fprop_global_init(&dom->completions, gfp);
|
| 620 | }
|
| 621 |
|
| 622 | #ifdef CONFIG_CGROUP_WRITEBACK
|
| 623 | void wb_domain_exit(struct wb_domain *dom)
|
| 624 | {
|
| 625 | timer_delete_sync(&dom->period_timer);
|
| 626 | fprop_global_destroy(&dom->completions);
|
| 627 | }
|
| 628 | #endif
|
| 629 |
|
| 630 | |
| 631 | * bdi_min_ratio keeps the sum of the minimum dirty shares of all
|
| 632 | * registered backing devices, which, for obvious reasons, can not
|
| 633 | * exceed 100%.
|
| 634 |
|
| 635 | static unsigned int bdi_min_ratio;
|
| 636 |
|
| 637 | static int bdi_check_pages_limit(unsigned long pages)
|
| 638 | {
|
| 639 | unsigned long max_dirty_pages = global_dirtyable_memory();
|
| 640 |
|
| 641 | if (pages > max_dirty_pages)
|
| 642 | return -EINVAL;
|
| 643 |
|
| 644 | return 0;
|
| 645 | }
|
| 646 |
|
| 647 | static unsigned long bdi_ratio_from_pages(unsigned long pages)
|
| 648 | {
|
| 649 | unsigned long background_thresh;
|
| 650 | unsigned long dirty_thresh;
|
| 651 | unsigned long ratio;
|
| 652 |
|
| 653 | global_dirty_limits(&background_thresh, &dirty_thresh);
|
| 654 | if (!dirty_thresh)
|
| 655 | return -EINVAL;
|
| 656 | ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh);
|
| 657 |
|
| 658 | return ratio;
|
| 659 | }
|
| 660 |
|
| 661 | static u64 bdi_get_bytes(unsigned int ratio)
|
| 662 | {
|
| 663 | unsigned long background_thresh;
|
| 664 | unsigned long dirty_thresh;
|
| 665 | u64 bytes;
|
| 666 |
|
| 667 | global_dirty_limits(&background_thresh, &dirty_thresh);
|
| 668 | bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100;
|
| 669 |
|
| 670 | return bytes;
|
| 671 | }
|
| 672 |
|
| 673 | static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
|
| 674 | {
|
| 675 | unsigned int delta;
|
| 676 | int ret = 0;
|
| 677 |
|
| 678 | if (min_ratio > 100 * BDI_RATIO_SCALE)
|
| 679 | return -EINVAL;
|
| 680 |
|
| 681 | spin_lock_bh(&bdi_lock);
|
| 682 | if (min_ratio > bdi->max_ratio) {
|
| 683 | ret = -EINVAL;
|
| 684 | } else {
|
| 685 | if (min_ratio < bdi->min_ratio) {
|
| 686 | delta = bdi->min_ratio - min_ratio;
|
| 687 | bdi_min_ratio -= delta;
|
| 688 | bdi->min_ratio = min_ratio;
|
| 689 | } else {
|
| 690 | delta = min_ratio - bdi->min_ratio;
|
| 691 | if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) {
|
| 692 | bdi_min_ratio += delta;
|
| 693 | bdi->min_ratio = min_ratio;
|
| 694 | } else {
|
| 695 | ret = -EINVAL;
|
| 696 | }
|
| 697 | }
|
| 698 | }
|
| 699 | spin_unlock_bh(&bdi_lock);
|
| 700 |
|
| 701 | return ret;
|
| 702 | }
|
| 703 |
|
| 704 | static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
|
| 705 | {
|
| 706 | int ret = 0;
|
| 707 |
|
| 708 | if (max_ratio > 100 * BDI_RATIO_SCALE)
|
| 709 | return -EINVAL;
|
| 710 |
|
| 711 | spin_lock_bh(&bdi_lock);
|
| 712 | if (bdi->min_ratio > max_ratio) {
|
| 713 | ret = -EINVAL;
|
| 714 | } else {
|
| 715 | bdi->max_ratio = max_ratio;
|
| 716 | bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) /
|
| 717 | (100 * BDI_RATIO_SCALE);
|
| 718 | }
|
| 719 | spin_unlock_bh(&bdi_lock);
|
| 720 |
|
| 721 | return ret;
|
| 722 | }
|
| 723 |
|
| 724 | int bdi_set_min_ratio_no_scale(struct backing_dev_info *bdi, unsigned int min_ratio)
|
| 725 | {
|
| 726 | return __bdi_set_min_ratio(bdi, min_ratio);
|
| 727 | }
|
| 728 |
|
| 729 | int bdi_set_max_ratio_no_scale(struct backing_dev_info *bdi, unsigned int max_ratio)
|
| 730 | {
|
| 731 | return __bdi_set_max_ratio(bdi, max_ratio);
|
| 732 | }
|
| 733 |
|
| 734 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
|
| 735 | {
|
| 736 | return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE);
|
| 737 | }
|
| 738 |
|
| 739 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio)
|
| 740 | {
|
| 741 | return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE);
|
| 742 | }
|
| 743 | EXPORT_SYMBOL(bdi_set_max_ratio);
|
| 744 |
|
| 745 | u64 bdi_get_min_bytes(struct backing_dev_info *bdi)
|
| 746 | {
|
| 747 | return bdi_get_bytes(bdi->min_ratio);
|
| 748 | }
|
| 749 |
|
| 750 | int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes)
|
| 751 | {
|
| 752 | int ret;
|
| 753 | unsigned long pages = min_bytes >> PAGE_SHIFT;
|
| 754 | long min_ratio;
|
| 755 |
|
| 756 | ret = bdi_check_pages_limit(pages);
|
| 757 | if (ret)
|
| 758 | return ret;
|
| 759 |
|
| 760 | min_ratio = bdi_ratio_from_pages(pages);
|
| 761 | if (min_ratio < 0)
|
| 762 | return min_ratio;
|
| 763 | return __bdi_set_min_ratio(bdi, min_ratio);
|
| 764 | }
|
| 765 |
|
| 766 | u64 bdi_get_max_bytes(struct backing_dev_info *bdi)
|
| 767 | {
|
| 768 | return bdi_get_bytes(bdi->max_ratio);
|
| 769 | }
|
| 770 |
|
| 771 | int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes)
|
| 772 | {
|
| 773 | int ret;
|
| 774 | unsigned long pages = max_bytes >> PAGE_SHIFT;
|
| 775 | long max_ratio;
|
| 776 |
|
| 777 | ret = bdi_check_pages_limit(pages);
|
| 778 | if (ret)
|
| 779 | return ret;
|
| 780 |
|
| 781 | max_ratio = bdi_ratio_from_pages(pages);
|
| 782 | if (max_ratio < 0)
|
| 783 | return max_ratio;
|
| 784 | return __bdi_set_max_ratio(bdi, max_ratio);
|
| 785 | }
|
| 786 |
|
| 787 | int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit)
|
| 788 | {
|
| 789 | if (strict_limit > 1)
|
| 790 | return -EINVAL;
|
| 791 |
|
| 792 | spin_lock_bh(&bdi_lock);
|
| 793 | if (strict_limit)
|
| 794 | bdi->capabilities |= BDI_CAP_STRICTLIMIT;
|
| 795 | else
|
| 796 | bdi->capabilities &= ~BDI_CAP_STRICTLIMIT;
|
| 797 | spin_unlock_bh(&bdi_lock);
|
| 798 |
|
| 799 | return 0;
|
| 800 | }
|
| 801 |
|
| 802 | static unsigned long dirty_freerun_ceiling(unsigned long thresh,
|
| 803 | unsigned long bg_thresh)
|
| 804 | {
|
| 805 | return (thresh + bg_thresh) / 2;
|
| 806 | }
|
| 807 |
|
| 808 | static unsigned long hard_dirty_limit(struct wb_domain *dom,
|
| 809 | unsigned long thresh)
|
| 810 | {
|
| 811 | return max(thresh, dom->dirty_limit);
|
| 812 | }
|
| 813 |
|
| 814 | |
| 815 | * Memory which can be further allocated to a memcg domain is capped by
|
| 816 | * system-wide clean memory excluding the amount being used in the domain.
|
| 817 |
|
| 818 | static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
|
| 819 | unsigned long filepages, unsigned long headroom)
|
| 820 | {
|
| 821 | struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
|
| 822 | unsigned long clean = filepages - min(filepages, mdtc->dirty);
|
| 823 | unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
|
| 824 | unsigned long other_clean = global_clean - min(global_clean, clean);
|
| 825 |
|
| 826 | mdtc->avail = filepages + min(headroom, other_clean);
|
| 827 | }
|
| 828 |
|
| 829 | static inline bool dtc_is_global(struct dirty_throttle_control *dtc)
|
| 830 | {
|
| 831 | return mdtc_gdtc(dtc) == NULL;
|
| 832 | }
|
| 833 |
|
| 834 | |
| 835 | * Dirty background will ignore pages being written as we're trying to
|
| 836 | * decide whether to put more under writeback.
|
| 837 |
|
| 838 | static void domain_dirty_avail(struct dirty_throttle_control *dtc,
|
| 839 | bool include_writeback)
|
| 840 | {
|
| 841 | if (dtc_is_global(dtc)) {
|
| 842 | dtc->avail = global_dirtyable_memory();
|
| 843 | dtc->dirty = global_node_page_state(NR_FILE_DIRTY);
|
| 844 | if (include_writeback)
|
| 845 | dtc->dirty += global_node_page_state(NR_WRITEBACK);
|
| 846 | } else {
|
| 847 | unsigned long filepages = 0, headroom = 0, writeback = 0;
|
| 848 |
|
| 849 | mem_cgroup_wb_stats(dtc->wb, &filepages, &headroom, &dtc->dirty,
|
| 850 | &writeback);
|
| 851 | if (include_writeback)
|
| 852 | dtc->dirty += writeback;
|
| 853 | mdtc_calc_avail(dtc, filepages, headroom);
|
| 854 | }
|
| 855 | }
|
| 856 |
|
| 857 | |
| 858 | * __wb_calc_thresh - @wb's share of dirty threshold
|
| 859 | * @dtc: dirty_throttle_context of interest
|
| 860 | * @thresh: dirty throttling or dirty background threshold of wb_domain in @dtc
|
| 861 | *
|
| 862 | * Note that balance_dirty_pages() will only seriously take dirty throttling
|
| 863 | * threshold as a hard limit when sleeping max_pause per page is not enough
|
| 864 | * to keep the dirty pages under control. For example, when the device is
|
| 865 | * completely stalled due to some error conditions, or when there are 1000
|
| 866 | * dd tasks writing to a slow 10MB/s USB key.
|
| 867 | * In the other normal situations, it acts more gently by throttling the tasks
|
| 868 | * more (rather than completely block them) when the wb dirty pages go high.
|
| 869 | *
|
| 870 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent
|
| 871 | * - starving fast devices
|
| 872 | * - piling up dirty pages (that will take long time to sync) on slow devices
|
| 873 | *
|
| 874 | * The wb's share of dirty limit will be adapting to its throughput and
|
| 875 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
|
| 876 | *
|
| 877 | * Return: @wb's dirty limit in pages. For dirty throttling limit, the term
|
| 878 | * "dirty" in the context of dirty balancing includes all PG_dirty and
|
| 879 | * PG_writeback pages.
|
| 880 |
|
| 881 | static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc,
|
| 882 | unsigned long thresh)
|
| 883 | {
|
| 884 | struct wb_domain *dom = dtc_dom(dtc);
|
| 885 | struct bdi_writeback *wb = dtc->wb;
|
| 886 | u64 wb_thresh;
|
| 887 | u64 wb_max_thresh;
|
| 888 | unsigned long numerator, denominator;
|
| 889 | unsigned long wb_min_ratio, wb_max_ratio;
|
| 890 |
|
| 891 | |
| 892 | * Calculate this wb's share of the thresh ratio.
|
| 893 |
|
| 894 | fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
|
| 895 | &numerator, &denominator);
|
| 896 |
|
| 897 | wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE);
|
| 898 | wb_thresh *= numerator;
|
| 899 | wb_thresh = div64_ul(wb_thresh, denominator);
|
| 900 |
|
| 901 | wb_min_max_ratio(wb, &wb_min_ratio, &wb_max_ratio);
|
| 902 |
|
| 903 | wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE);
|
| 904 |
|
| 905 | |
| 906 | * It's very possible that wb_thresh is close to 0 not because the
|
| 907 | * device is slow, but that it has remained inactive for long time.
|
| 908 | * Honour such devices a reasonable good (hopefully IO efficient)
|
| 909 | * threshold, so that the occasional writes won't be blocked and active
|
| 910 | * writes can rampup the threshold quickly.
|
| 911 |
|
| 912 | if (thresh > dtc->dirty) {
|
| 913 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT))
|
| 914 | wb_thresh = max(wb_thresh, (thresh - dtc->dirty) / 100);
|
| 915 | else
|
| 916 | wb_thresh = max(wb_thresh, (thresh - dtc->dirty) / 8);
|
| 917 | }
|
| 918 |
|
| 919 | wb_max_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE);
|
| 920 | if (wb_thresh > wb_max_thresh)
|
| 921 | wb_thresh = wb_max_thresh;
|
| 922 |
|
| 923 | return wb_thresh;
|
| 924 | }
|
| 925 |
|
| 926 | unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
|
| 927 | {
|
| 928 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
|
| 929 |
|
| 930 | domain_dirty_avail(&gdtc, true);
|
| 931 | return __wb_calc_thresh(&gdtc, thresh);
|
| 932 | }
|
| 933 |
|
| 934 | unsigned long cgwb_calc_thresh(struct bdi_writeback *wb)
|
| 935 | {
|
| 936 | struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
|
| 937 | struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
|
| 938 |
|
| 939 | domain_dirty_avail(&gdtc, true);
|
| 940 | domain_dirty_avail(&mdtc, true);
|
| 941 | domain_dirty_limits(&mdtc);
|
| 942 |
|
| 943 | return __wb_calc_thresh(&mdtc, mdtc.thresh);
|
| 944 | }
|
| 945 |
|
| 946 | |
| 947 | * setpoint - dirty 3
|
| 948 | * f(dirty) := 1.0 + (----------------)
|
| 949 | * limit - setpoint
|
| 950 | *
|
| 951 | * it's a 3rd order polynomial that subjects to
|
| 952 | *
|
| 953 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
|
| 954 | * (2) f(setpoint) = 1.0 => the balance point
|
| 955 | * (3) f(limit) = 0 => the hard limit
|
| 956 | * (4) df/dx <= 0 => negative feedback control
|
| 957 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
|
| 958 | * => fast response on large errors; small oscillation near setpoint
|
| 959 |
|
| 960 | static long long pos_ratio_polynom(unsigned long setpoint,
|
| 961 | unsigned long dirty,
|
| 962 | unsigned long limit)
|
| 963 | {
|
| 964 | long long pos_ratio;
|
| 965 | long x;
|
| 966 |
|
| 967 | x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
|
| 968 | (limit - setpoint) | 1);
|
| 969 | pos_ratio = x;
|
| 970 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
|
| 971 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
|
| 972 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
|
| 973 |
|
| 974 | return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
|
| 975 | }
|
| 976 |
|
| 977 | |
| 978 | * Dirty position control.
|
| 979 | *
|
| 980 | * (o) global/bdi setpoints
|
| 981 | *
|
| 982 | * We want the dirty pages be balanced around the global/wb setpoints.
|
| 983 | * When the number of dirty pages is higher/lower than the setpoint, the
|
| 984 | * dirty position control ratio (and hence task dirty ratelimit) will be
|
| 985 | * decreased/increased to bring the dirty pages back to the setpoint.
|
| 986 | *
|
| 987 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
|
| 988 | *
|
| 989 | * if (dirty < setpoint) scale up pos_ratio
|
| 990 | * if (dirty > setpoint) scale down pos_ratio
|
| 991 | *
|
| 992 | * if (wb_dirty < wb_setpoint) scale up pos_ratio
|
| 993 | * if (wb_dirty > wb_setpoint) scale down pos_ratio
|
| 994 | *
|
| 995 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
|
| 996 | *
|
| 997 | * (o) global control line
|
| 998 | *
|
| 999 | * ^ pos_ratio
|
| 1000 | * |
|
| 1001 | * | |<===== global dirty control scope ======>|
|
| 1002 | * 2.0 * * * * * * *
|
| 1003 | * | .*
|
| 1004 | * | . *
|
| 1005 | * | . *
|
| 1006 | * | . *
|
| 1007 | * | . *
|
| 1008 | * | . *
|
| 1009 | * 1.0 ................................*
|
| 1010 | * | . . *
|
| 1011 | * | . . *
|
| 1012 | * | . . *
|
| 1013 | * | . . *
|
| 1014 | * | . . *
|
| 1015 | * 0 +------------.------------------.----------------------*------------->
|
| 1016 | * freerun^ setpoint^ limit^ dirty pages
|
| 1017 | *
|
| 1018 | * (o) wb control line
|
| 1019 | *
|
| 1020 | * ^ pos_ratio
|
| 1021 | * |
|
| 1022 | * | *
|
| 1023 | * | *
|
| 1024 | * | *
|
| 1025 | * | *
|
| 1026 | * | * |<=========== span ============>|
|
| 1027 | * 1.0 .......................*
|
| 1028 | * | . *
|
| 1029 | * | . *
|
| 1030 | * | . *
|
| 1031 | * | . *
|
| 1032 | * | . *
|
| 1033 | * | . *
|
| 1034 | * | . *
|
| 1035 | * | . *
|
| 1036 | * | . *
|
| 1037 | * | . *
|
| 1038 | * | . *
|
| 1039 | * 1/4 ...............................................* * * * * * * * * * * *
|
| 1040 | * | . .
|
| 1041 | * | . .
|
| 1042 | * | . .
|
| 1043 | * 0 +----------------------.-------------------------------.------------->
|
| 1044 | * wb_setpoint^ x_intercept^
|
| 1045 | *
|
| 1046 | * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
|
| 1047 | * be smoothly throttled down to normal if it starts high in situations like
|
| 1048 | * - start writing to a slow SD card and a fast disk at the same time. The SD
|
| 1049 | * card's wb_dirty may rush to many times higher than wb_setpoint.
|
| 1050 | * - the wb dirty thresh drops quickly due to change of JBOD workload
|
| 1051 |
|
| 1052 | static void wb_position_ratio(struct dirty_throttle_control *dtc)
|
| 1053 | {
|
| 1054 | struct bdi_writeback *wb = dtc->wb;
|
| 1055 | unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth);
|
| 1056 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
|
| 1057 | unsigned long limit = dtc->limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
|
| 1058 | unsigned long wb_thresh = dtc->wb_thresh;
|
| 1059 | unsigned long x_intercept;
|
| 1060 | unsigned long setpoint;
|
| 1061 | unsigned long wb_setpoint;
|
| 1062 | unsigned long span;
|
| 1063 | long long pos_ratio;
|
| 1064 | long x;
|
| 1065 |
|
| 1066 | dtc->pos_ratio = 0;
|
| 1067 |
|
| 1068 | if (unlikely(dtc->dirty >= limit))
|
| 1069 | return;
|
| 1070 |
|
| 1071 | |
| 1072 | * global setpoint
|
| 1073 | *
|
| 1074 | * See comment for pos_ratio_polynom().
|
| 1075 |
|
| 1076 | setpoint = (freerun + limit) / 2;
|
| 1077 | pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
|
| 1078 |
|
| 1079 | |
| 1080 | * The strictlimit feature is a tool preventing mistrusted filesystems
|
| 1081 | * from growing a large number of dirty pages before throttling. For
|
| 1082 | * such filesystems balance_dirty_pages always checks wb counters
|
| 1083 | * against wb limits. Even if global "nr_dirty" is under "freerun".
|
| 1084 | * This is especially important for fuse which sets bdi->max_ratio to
|
| 1085 | * 1% by default.
|
| 1086 | *
|
| 1087 | * Here, in wb_position_ratio(), we calculate pos_ratio based on
|
| 1088 | * two values: wb_dirty and wb_thresh. Let's consider an example:
|
| 1089 | * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
|
| 1090 | * limits are set by default to 10% and 20% (background and throttle).
|
| 1091 | * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
|
| 1092 | * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
|
| 1093 | * about ~6K pages (as the average of background and throttle wb
|
| 1094 | * limits). The 3rd order polynomial will provide positive feedback if
|
| 1095 | * wb_dirty is under wb_setpoint and vice versa.
|
| 1096 | *
|
| 1097 | * Note, that we cannot use global counters in these calculations
|
| 1098 | * because we want to throttle process writing to a strictlimit wb
|
| 1099 | * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
|
| 1100 | * in the example above).
|
| 1101 |
|
| 1102 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
|
| 1103 | long long wb_pos_ratio;
|
| 1104 |
|
| 1105 | if (dtc->wb_dirty >= wb_thresh)
|
| 1106 | return;
|
| 1107 |
|
| 1108 | wb_setpoint = dirty_freerun_ceiling(wb_thresh,
|
| 1109 | dtc->wb_bg_thresh);
|
| 1110 |
|
| 1111 | if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
|
| 1112 | return;
|
| 1113 |
|
| 1114 | wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
|
| 1115 | wb_thresh);
|
| 1116 |
|
| 1117 | |
| 1118 | * Typically, for strictlimit case, wb_setpoint << setpoint
|
| 1119 | * and pos_ratio >> wb_pos_ratio. In the other words global
|
| 1120 | * state ("dirty") is not limiting factor and we have to
|
| 1121 | * make decision based on wb counters. But there is an
|
| 1122 | * important case when global pos_ratio should get precedence:
|
| 1123 | * global limits are exceeded (e.g. due to activities on other
|
| 1124 | * wb's) while given strictlimit wb is below limit.
|
| 1125 | *
|
| 1126 | * "pos_ratio * wb_pos_ratio" would work for the case above,
|
| 1127 | * but it would look too non-natural for the case of all
|
| 1128 | * activity in the system coming from a single strictlimit wb
|
| 1129 | * with bdi->max_ratio == 100%.
|
| 1130 | *
|
| 1131 | * Note that min() below somewhat changes the dynamics of the
|
| 1132 | * control system. Normally, pos_ratio value can be well over 3
|
| 1133 | * (when globally we are at freerun and wb is well below wb
|
| 1134 | * setpoint). Now the maximum pos_ratio in the same situation
|
| 1135 | * is 2. We might want to tweak this if we observe the control
|
| 1136 | * system is too slow to adapt.
|
| 1137 |
|
| 1138 | dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
|
| 1139 | return;
|
| 1140 | }
|
| 1141 |
|
| 1142 | |
| 1143 | * We have computed basic pos_ratio above based on global situation. If
|
| 1144 | * the wb is over/under its share of dirty pages, we want to scale
|
| 1145 | * pos_ratio further down/up. That is done by the following mechanism.
|
| 1146 |
|
| 1147 |
|
| 1148 | |
| 1149 | * wb setpoint
|
| 1150 | *
|
| 1151 | * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
|
| 1152 | *
|
| 1153 | * x_intercept - wb_dirty
|
| 1154 | * := --------------------------
|
| 1155 | * x_intercept - wb_setpoint
|
| 1156 | *
|
| 1157 | * The main wb control line is a linear function that subjects to
|
| 1158 | *
|
| 1159 | * (1) f(wb_setpoint) = 1.0
|
| 1160 | * (2) k = - 1 / (8 * write_bw) (in single wb case)
|
| 1161 | * or equally: x_intercept = wb_setpoint + 8 * write_bw
|
| 1162 | *
|
| 1163 | * For single wb case, the dirty pages are observed to fluctuate
|
| 1164 | * regularly within range
|
| 1165 | * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
|
| 1166 | * for various filesystems, where (2) can yield in a reasonable 12.5%
|
| 1167 | * fluctuation range for pos_ratio.
|
| 1168 | *
|
| 1169 | * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
|
| 1170 | * own size, so move the slope over accordingly and choose a slope that
|
| 1171 | * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
|
| 1172 |
|
| 1173 | if (unlikely(wb_thresh > dtc->thresh))
|
| 1174 | wb_thresh = dtc->thresh;
|
| 1175 | |
| 1176 | * scale global setpoint to wb's:
|
| 1177 | * wb_setpoint = setpoint * wb_thresh / thresh
|
| 1178 |
|
| 1179 | x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
|
| 1180 | wb_setpoint = setpoint * (u64)x >> 16;
|
| 1181 | |
| 1182 | * Use span=(8*write_bw) in single wb case as indicated by
|
| 1183 | * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
|
| 1184 | *
|
| 1185 | * wb_thresh thresh - wb_thresh
|
| 1186 | * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
|
| 1187 | * thresh thresh
|
| 1188 |
|
| 1189 | span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
|
| 1190 | x_intercept = wb_setpoint + span;
|
| 1191 |
|
| 1192 | if (dtc->wb_dirty < x_intercept - span / 4) {
|
| 1193 | pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
|
| 1194 | (x_intercept - wb_setpoint) | 1);
|
| 1195 | } else
|
| 1196 | pos_ratio /= 4;
|
| 1197 |
|
| 1198 | |
| 1199 | * wb reserve area, safeguard against dirty pool underrun and disk idle
|
| 1200 | * It may push the desired control point of global dirty pages higher
|
| 1201 | * than setpoint.
|
| 1202 |
|
| 1203 | x_intercept = wb_thresh / 2;
|
| 1204 | if (dtc->wb_dirty < x_intercept) {
|
| 1205 | if (dtc->wb_dirty > x_intercept / 8)
|
| 1206 | pos_ratio = div_u64(pos_ratio * x_intercept,
|
| 1207 | dtc->wb_dirty);
|
| 1208 | else
|
| 1209 | pos_ratio *= 8;
|
| 1210 | }
|
| 1211 |
|
| 1212 | dtc->pos_ratio = pos_ratio;
|
| 1213 | }
|
| 1214 |
|
| 1215 | static void wb_update_write_bandwidth(struct bdi_writeback *wb,
|
| 1216 | unsigned long elapsed,
|
| 1217 | unsigned long written)
|
| 1218 | {
|
| 1219 | const unsigned long period = roundup_pow_of_two(3 * HZ);
|
| 1220 | unsigned long avg = wb->avg_write_bandwidth;
|
| 1221 | unsigned long old = wb->write_bandwidth;
|
| 1222 | u64 bw;
|
| 1223 |
|
| 1224 | |
| 1225 | * bw = written * HZ / elapsed
|
| 1226 | *
|
| 1227 | * bw * elapsed + write_bandwidth * (period - elapsed)
|
| 1228 | * write_bandwidth = ---------------------------------------------------
|
| 1229 | * period
|
| 1230 | *
|
| 1231 | * @written may have decreased due to folio_redirty_for_writepage().
|
| 1232 | * Avoid underflowing @bw calculation.
|
| 1233 |
|
| 1234 | bw = written - min(written, wb->written_stamp);
|
| 1235 | bw *= HZ;
|
| 1236 | if (unlikely(elapsed > period)) {
|
| 1237 | bw = div64_ul(bw, elapsed);
|
| 1238 | avg = bw;
|
| 1239 | goto out;
|
| 1240 | }
|
| 1241 | bw += (u64)wb->write_bandwidth * (period - elapsed);
|
| 1242 | bw >>= ilog2(period);
|
| 1243 |
|
| 1244 | |
| 1245 | * one more level of smoothing, for filtering out sudden spikes
|
| 1246 |
|
| 1247 | if (avg > old && old >= (unsigned long)bw)
|
| 1248 | avg -= (avg - old) >> 3;
|
| 1249 |
|
| 1250 | if (avg < old && old <= (unsigned long)bw)
|
| 1251 | avg += (old - avg) >> 3;
|
| 1252 |
|
| 1253 | out:
|
| 1254 |
|
| 1255 | avg = max(avg, 1LU);
|
| 1256 | if (wb_has_dirty_io(wb)) {
|
| 1257 | long delta = avg - wb->avg_write_bandwidth;
|
| 1258 | WARN_ON_ONCE(atomic_long_add_return(delta,
|
| 1259 | &wb->bdi->tot_write_bandwidth) <= 0);
|
| 1260 | }
|
| 1261 | wb->write_bandwidth = bw;
|
| 1262 | WRITE_ONCE(wb->avg_write_bandwidth, avg);
|
| 1263 | }
|
| 1264 |
|
| 1265 | static void update_dirty_limit(struct dirty_throttle_control *dtc)
|
| 1266 | {
|
| 1267 | struct wb_domain *dom = dtc_dom(dtc);
|
| 1268 | unsigned long thresh = dtc->thresh;
|
| 1269 | unsigned long limit = dom->dirty_limit;
|
| 1270 |
|
| 1271 | |
| 1272 | * Follow up in one step.
|
| 1273 |
|
| 1274 | if (limit < thresh) {
|
| 1275 | limit = thresh;
|
| 1276 | goto update;
|
| 1277 | }
|
| 1278 |
|
| 1279 | |
| 1280 | * Follow down slowly. Use the higher one as the target, because thresh
|
| 1281 | * may drop below dirty. This is exactly the reason to introduce
|
| 1282 | * dom->dirty_limit which is guaranteed to lie above the dirty pages.
|
| 1283 |
|
| 1284 | thresh = max(thresh, dtc->dirty);
|
| 1285 | if (limit > thresh) {
|
| 1286 | limit -= (limit - thresh) >> 5;
|
| 1287 | goto update;
|
| 1288 | }
|
| 1289 | return;
|
| 1290 | update:
|
| 1291 | dom->dirty_limit = limit;
|
| 1292 | }
|
| 1293 |
|
| 1294 | static void domain_update_dirty_limit(struct dirty_throttle_control *dtc,
|
| 1295 | unsigned long now)
|
| 1296 | {
|
| 1297 | struct wb_domain *dom = dtc_dom(dtc);
|
| 1298 |
|
| 1299 | |
| 1300 | * check locklessly first to optimize away locking for the most time
|
| 1301 |
|
| 1302 | if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
|
| 1303 | return;
|
| 1304 |
|
| 1305 | spin_lock(&dom->lock);
|
| 1306 | if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
|
| 1307 | update_dirty_limit(dtc);
|
| 1308 | dom->dirty_limit_tstamp = now;
|
| 1309 | }
|
| 1310 | spin_unlock(&dom->lock);
|
| 1311 | }
|
| 1312 |
|
| 1313 | |
| 1314 | * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
|
| 1315 | *
|
| 1316 | * Normal wb tasks will be curbed at or below it in long term.
|
| 1317 | * Obviously it should be around (write_bw / N) when there are N dd tasks.
|
| 1318 |
|
| 1319 | static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
|
| 1320 | unsigned long dirtied,
|
| 1321 | unsigned long elapsed)
|
| 1322 | {
|
| 1323 | struct bdi_writeback *wb = dtc->wb;
|
| 1324 | unsigned long dirty = dtc->dirty;
|
| 1325 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
|
| 1326 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
|
| 1327 | unsigned long setpoint = (freerun + limit) / 2;
|
| 1328 | unsigned long write_bw = wb->avg_write_bandwidth;
|
| 1329 | unsigned long dirty_ratelimit = wb->dirty_ratelimit;
|
| 1330 | unsigned long dirty_rate;
|
| 1331 | unsigned long task_ratelimit;
|
| 1332 | unsigned long balanced_dirty_ratelimit;
|
| 1333 | unsigned long step;
|
| 1334 | unsigned long x;
|
| 1335 | unsigned long shift;
|
| 1336 |
|
| 1337 | |
| 1338 | * The dirty rate will match the writeout rate in long term, except
|
| 1339 | * when dirty pages are truncated by userspace or re-dirtied by FS.
|
| 1340 |
|
| 1341 | dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
|
| 1342 |
|
| 1343 | |
| 1344 | * task_ratelimit reflects each dd's dirty rate for the past 200ms.
|
| 1345 |
|
| 1346 | task_ratelimit = (u64)dirty_ratelimit *
|
| 1347 | dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
|
| 1348 | task_ratelimit++;
|
| 1349 |
|
| 1350 | |
| 1351 | * A linear estimation of the "balanced" throttle rate. The theory is,
|
| 1352 | * if there are N dd tasks, each throttled at task_ratelimit, the wb's
|
| 1353 | * dirty_rate will be measured to be (N * task_ratelimit). So the below
|
| 1354 | * formula will yield the balanced rate limit (write_bw / N).
|
| 1355 | *
|
| 1356 | * Note that the expanded form is not a pure rate feedback:
|
| 1357 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
|
| 1358 | * but also takes pos_ratio into account:
|
| 1359 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
|
| 1360 | *
|
| 1361 | * (1) is not realistic because pos_ratio also takes part in balancing
|
| 1362 | * the dirty rate. Consider the state
|
| 1363 | * pos_ratio = 0.5 (3)
|
| 1364 | * rate = 2 * (write_bw / N) (4)
|
| 1365 | * If (1) is used, it will stuck in that state! Because each dd will
|
| 1366 | * be throttled at
|
| 1367 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
|
| 1368 | * yielding
|
| 1369 | * dirty_rate = N * task_ratelimit = write_bw (6)
|
| 1370 | * put (6) into (1) we get
|
| 1371 | * rate_(i+1) = rate_(i) (7)
|
| 1372 | *
|
| 1373 | * So we end up using (2) to always keep
|
| 1374 | * rate_(i+1) ~= (write_bw / N) (8)
|
| 1375 | * regardless of the value of pos_ratio. As long as (8) is satisfied,
|
| 1376 | * pos_ratio is able to drive itself to 1.0, which is not only where
|
| 1377 | * the dirty count meet the setpoint, but also where the slope of
|
| 1378 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
|
| 1379 |
|
| 1380 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
|
| 1381 | dirty_rate | 1);
|
| 1382 | |
| 1383 | * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
|
| 1384 |
|
| 1385 | if (unlikely(balanced_dirty_ratelimit > write_bw))
|
| 1386 | balanced_dirty_ratelimit = write_bw;
|
| 1387 |
|
| 1388 | |
| 1389 | * We could safely do this and return immediately:
|
| 1390 | *
|
| 1391 | * wb->dirty_ratelimit = balanced_dirty_ratelimit;
|
| 1392 | *
|
| 1393 | * However to get a more stable dirty_ratelimit, the below elaborated
|
| 1394 | * code makes use of task_ratelimit to filter out singular points and
|
| 1395 | * limit the step size.
|
| 1396 | *
|
| 1397 | * The below code essentially only uses the relative value of
|
| 1398 | *
|
| 1399 | * task_ratelimit - dirty_ratelimit
|
| 1400 | * = (pos_ratio - 1) * dirty_ratelimit
|
| 1401 | *
|
| 1402 | * which reflects the direction and size of dirty position error.
|
| 1403 |
|
| 1404 |
|
| 1405 | |
| 1406 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff
|
| 1407 | * task_ratelimit is on the same side of dirty_ratelimit, too.
|
| 1408 | * For example, when
|
| 1409 | * - dirty_ratelimit > balanced_dirty_ratelimit
|
| 1410 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
|
| 1411 | * lowering dirty_ratelimit will help meet both the position and rate
|
| 1412 | * control targets. Otherwise, don't update dirty_ratelimit if it will
|
| 1413 | * only help meet the rate target. After all, what the users ultimately
|
| 1414 | * feel and care are stable dirty rate and small position error.
|
| 1415 | *
|
| 1416 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size
|
| 1417 | * and filter out the singular points of balanced_dirty_ratelimit. Which
|
| 1418 | * keeps jumping around randomly and can even leap far away at times
|
| 1419 | * due to the small 200ms estimation period of dirty_rate (we want to
|
| 1420 | * keep that period small to reduce time lags).
|
| 1421 |
|
| 1422 | step = 0;
|
| 1423 |
|
| 1424 | |
| 1425 | * For strictlimit case, calculations above were based on wb counters
|
| 1426 | * and limits (starting from pos_ratio = wb_position_ratio() and up to
|
| 1427 | * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
|
| 1428 | * Hence, to calculate "step" properly, we have to use wb_dirty as
|
| 1429 | * "dirty" and wb_setpoint as "setpoint".
|
| 1430 |
|
| 1431 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
|
| 1432 | dirty = dtc->wb_dirty;
|
| 1433 | setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
|
| 1434 | }
|
| 1435 |
|
| 1436 | if (dirty < setpoint) {
|
| 1437 | x = min3(wb->balanced_dirty_ratelimit,
|
| 1438 | balanced_dirty_ratelimit, task_ratelimit);
|
| 1439 | if (dirty_ratelimit < x)
|
| 1440 | step = x - dirty_ratelimit;
|
| 1441 | } else {
|
| 1442 | x = max3(wb->balanced_dirty_ratelimit,
|
| 1443 | balanced_dirty_ratelimit, task_ratelimit);
|
| 1444 | if (dirty_ratelimit > x)
|
| 1445 | step = dirty_ratelimit - x;
|
| 1446 | }
|
| 1447 |
|
| 1448 | |
| 1449 | * Don't pursue 100% rate matching. It's impossible since the balanced
|
| 1450 | * rate itself is constantly fluctuating. So decrease the track speed
|
| 1451 | * when it gets close to the target. Helps eliminate pointless tremors.
|
| 1452 |
|
| 1453 | shift = dirty_ratelimit / (2 * step + 1);
|
| 1454 | if (shift < BITS_PER_LONG)
|
| 1455 | step = DIV_ROUND_UP(step >> shift, 8);
|
| 1456 | else
|
| 1457 | step = 0;
|
| 1458 |
|
| 1459 | if (dirty_ratelimit < balanced_dirty_ratelimit)
|
| 1460 | dirty_ratelimit += step;
|
| 1461 | else
|
| 1462 | dirty_ratelimit -= step;
|
| 1463 |
|
| 1464 | WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL));
|
| 1465 | wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
|
| 1466 |
|
| 1467 | trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
|
| 1468 | }
|
| 1469 |
|
| 1470 | static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
|
| 1471 | struct dirty_throttle_control *mdtc,
|
| 1472 | bool update_ratelimit)
|
| 1473 | {
|
| 1474 | struct bdi_writeback *wb = gdtc->wb;
|
| 1475 | unsigned long now = jiffies;
|
| 1476 | unsigned long elapsed;
|
| 1477 | unsigned long dirtied;
|
| 1478 | unsigned long written;
|
| 1479 |
|
| 1480 | spin_lock(&wb->list_lock);
|
| 1481 |
|
| 1482 | |
| 1483 | * Lockless checks for elapsed time are racy and delayed update after
|
| 1484 | * IO completion doesn't do it at all (to make sure written pages are
|
| 1485 | * accounted reasonably quickly). Make sure elapsed >= 1 to avoid
|
| 1486 | * division errors.
|
| 1487 |
|
| 1488 | elapsed = max(now - wb->bw_time_stamp, 1UL);
|
| 1489 | dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
|
| 1490 | written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
|
| 1491 |
|
| 1492 | if (update_ratelimit) {
|
| 1493 | domain_update_dirty_limit(gdtc, now);
|
| 1494 | wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
|
| 1495 |
|
| 1496 | |
| 1497 | * @mdtc is always NULL if !CGROUP_WRITEBACK but the
|
| 1498 | * compiler has no way to figure that out. Help it.
|
| 1499 |
|
| 1500 | if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
|
| 1501 | domain_update_dirty_limit(mdtc, now);
|
| 1502 | wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
|
| 1503 | }
|
| 1504 | }
|
| 1505 | wb_update_write_bandwidth(wb, elapsed, written);
|
| 1506 |
|
| 1507 | wb->dirtied_stamp = dirtied;
|
| 1508 | wb->written_stamp = written;
|
| 1509 | WRITE_ONCE(wb->bw_time_stamp, now);
|
| 1510 | spin_unlock(&wb->list_lock);
|
| 1511 | }
|
| 1512 |
|
| 1513 | void wb_update_bandwidth(struct bdi_writeback *wb)
|
| 1514 | {
|
| 1515 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
|
| 1516 |
|
| 1517 | __wb_update_bandwidth(&gdtc, NULL, false);
|
| 1518 | }
|
| 1519 |
|
| 1520 |
|
| 1521 | #define WB_BANDWIDTH_IDLE_JIF (HZ)
|
| 1522 |
|
| 1523 | static void wb_bandwidth_estimate_start(struct bdi_writeback *wb)
|
| 1524 | {
|
| 1525 | unsigned long now = jiffies;
|
| 1526 | unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp);
|
| 1527 |
|
| 1528 | if (elapsed > WB_BANDWIDTH_IDLE_JIF &&
|
| 1529 | !atomic_read(&wb->writeback_inodes)) {
|
| 1530 | spin_lock(&wb->list_lock);
|
| 1531 | wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED);
|
| 1532 | wb->written_stamp = wb_stat(wb, WB_WRITTEN);
|
| 1533 | WRITE_ONCE(wb->bw_time_stamp, now);
|
| 1534 | spin_unlock(&wb->list_lock);
|
| 1535 | }
|
| 1536 | }
|
| 1537 |
|
| 1538 | |
| 1539 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
|
| 1540 | * will look to see if it needs to start dirty throttling.
|
| 1541 | *
|
| 1542 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive
|
| 1543 | * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
|
| 1544 | * (the number of pages we may dirty without exceeding the dirty limits).
|
| 1545 |
|
| 1546 | static unsigned long dirty_poll_interval(unsigned long dirty,
|
| 1547 | unsigned long thresh)
|
| 1548 | {
|
| 1549 | if (thresh > dirty)
|
| 1550 | return 1UL << (ilog2(thresh - dirty) >> 1);
|
| 1551 |
|
| 1552 | return 1;
|
| 1553 | }
|
| 1554 |
|
| 1555 | static unsigned long wb_max_pause(struct bdi_writeback *wb,
|
| 1556 | unsigned long wb_dirty)
|
| 1557 | {
|
| 1558 | unsigned long bw = READ_ONCE(wb->avg_write_bandwidth);
|
| 1559 | unsigned long t;
|
| 1560 |
|
| 1561 | |
| 1562 | * Limit pause time for small memory systems. If sleeping for too long
|
| 1563 | * time, a small pool of dirty/writeback pages may go empty and disk go
|
| 1564 | * idle.
|
| 1565 | *
|
| 1566 | * 8 serves as the safety ratio.
|
| 1567 |
|
| 1568 | t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
|
| 1569 | t++;
|
| 1570 |
|
| 1571 | return min_t(unsigned long, t, MAX_PAUSE);
|
| 1572 | }
|
| 1573 |
|
| 1574 | static long wb_min_pause(struct bdi_writeback *wb,
|
| 1575 | long max_pause,
|
| 1576 | unsigned long task_ratelimit,
|
| 1577 | unsigned long dirty_ratelimit,
|
| 1578 | int *nr_dirtied_pause)
|
| 1579 | {
|
| 1580 | long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth));
|
| 1581 | long lo = ilog2(READ_ONCE(wb->dirty_ratelimit));
|
| 1582 | long t;
|
| 1583 | long pause;
|
| 1584 | int pages;
|
| 1585 |
|
| 1586 |
|
| 1587 | t = max(1, HZ / 100);
|
| 1588 |
|
| 1589 | |
| 1590 | * Scale up pause time for concurrent dirtiers in order to reduce CPU
|
| 1591 | * overheads.
|
| 1592 | *
|
| 1593 | * (N * 10ms) on 2^N concurrent tasks.
|
| 1594 |
|
| 1595 | if (hi > lo)
|
| 1596 | t += (hi - lo) * (10 * HZ) / 1024;
|
| 1597 |
|
| 1598 | |
| 1599 | * This is a bit convoluted. We try to base the next nr_dirtied_pause
|
| 1600 | * on the much more stable dirty_ratelimit. However the next pause time
|
| 1601 | * will be computed based on task_ratelimit and the two rate limits may
|
| 1602 | * depart considerably at some time. Especially if task_ratelimit goes
|
| 1603 | * below dirty_ratelimit/2 and the target pause is max_pause, the next
|
| 1604 | * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
|
| 1605 | * result task_ratelimit won't be executed faithfully, which could
|
| 1606 | * eventually bring down dirty_ratelimit.
|
| 1607 | *
|
| 1608 | * We apply two rules to fix it up:
|
| 1609 | * 1) try to estimate the next pause time and if necessary, use a lower
|
| 1610 | * nr_dirtied_pause so as not to exceed max_pause. When this happens,
|
| 1611 | * nr_dirtied_pause will be "dancing" with task_ratelimit.
|
| 1612 | * 2) limit the target pause time to max_pause/2, so that the normal
|
| 1613 | * small fluctuations of task_ratelimit won't trigger rule (1) and
|
| 1614 | * nr_dirtied_pause will remain as stable as dirty_ratelimit.
|
| 1615 |
|
| 1616 | t = min(t, 1 + max_pause / 2);
|
| 1617 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
|
| 1618 |
|
| 1619 | |
| 1620 | * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
|
| 1621 | * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
|
| 1622 | * When the 16 consecutive reads are often interrupted by some dirty
|
| 1623 | * throttling pause during the async writes, cfq will go into idles
|
| 1624 | * (deadline is fine). So push nr_dirtied_pause as high as possible
|
| 1625 | * until reaches DIRTY_POLL_THRESH=32 pages.
|
| 1626 |
|
| 1627 | if (pages < DIRTY_POLL_THRESH) {
|
| 1628 | t = max_pause;
|
| 1629 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
|
| 1630 | if (pages > DIRTY_POLL_THRESH) {
|
| 1631 | pages = DIRTY_POLL_THRESH;
|
| 1632 | t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
|
| 1633 | }
|
| 1634 | }
|
| 1635 |
|
| 1636 | pause = HZ * pages / (task_ratelimit + 1);
|
| 1637 | if (pause > max_pause) {
|
| 1638 | t = max_pause;
|
| 1639 | pages = task_ratelimit * t / roundup_pow_of_two(HZ);
|
| 1640 | }
|
| 1641 |
|
| 1642 | *nr_dirtied_pause = pages;
|
| 1643 | |
| 1644 | * The minimal pause time will normally be half the target pause time.
|
| 1645 |
|
| 1646 | return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
|
| 1647 | }
|
| 1648 |
|
| 1649 | static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
|
| 1650 | {
|
| 1651 | struct bdi_writeback *wb = dtc->wb;
|
| 1652 | unsigned long wb_reclaimable;
|
| 1653 |
|
| 1654 | |
| 1655 | * wb_thresh is not treated as some limiting factor as
|
| 1656 | * dirty_thresh, due to reasons
|
| 1657 | * - in JBOD setup, wb_thresh can fluctuate a lot
|
| 1658 | * - in a system with HDD and USB key, the USB key may somehow
|
| 1659 | * go into state (wb_dirty >> wb_thresh) either because
|
| 1660 | * wb_dirty starts high, or because wb_thresh drops low.
|
| 1661 | * In this case we don't want to hard throttle the USB key
|
| 1662 | * dirtiers for 100 seconds until wb_dirty drops under
|
| 1663 | * wb_thresh. Instead the auxiliary wb control line in
|
| 1664 | * wb_position_ratio() will let the dirtier task progress
|
| 1665 | * at some rate <= (write_bw / 2) for bringing down wb_dirty.
|
| 1666 |
|
| 1667 | dtc->wb_thresh = __wb_calc_thresh(dtc, dtc->thresh);
|
| 1668 | dtc->wb_bg_thresh = dtc->thresh ?
|
| 1669 | div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
|
| 1670 |
|
| 1671 | |
| 1672 | * In order to avoid the stacked BDI deadlock we need
|
| 1673 | * to ensure we accurately count the 'dirty' pages when
|
| 1674 | * the threshold is low.
|
| 1675 | *
|
| 1676 | * Otherwise it would be possible to get thresh+n pages
|
| 1677 | * reported dirty, even though there are thresh-m pages
|
| 1678 | * actually dirty; with m+n sitting in the percpu
|
| 1679 | * deltas.
|
| 1680 |
|
| 1681 | if (dtc->wb_thresh < 2 * wb_stat_error()) {
|
| 1682 | wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
|
| 1683 | dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
|
| 1684 | } else {
|
| 1685 | wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
|
| 1686 | dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
|
| 1687 | }
|
| 1688 | }
|
| 1689 |
|
| 1690 | static unsigned long domain_poll_intv(struct dirty_throttle_control *dtc,
|
| 1691 | bool strictlimit)
|
| 1692 | {
|
| 1693 | unsigned long dirty, thresh;
|
| 1694 |
|
| 1695 | if (strictlimit) {
|
| 1696 | dirty = dtc->wb_dirty;
|
| 1697 | thresh = dtc->wb_thresh;
|
| 1698 | } else {
|
| 1699 | dirty = dtc->dirty;
|
| 1700 | thresh = dtc->thresh;
|
| 1701 | }
|
| 1702 |
|
| 1703 | return dirty_poll_interval(dirty, thresh);
|
| 1704 | }
|
| 1705 |
|
| 1706 | |
| 1707 | * Throttle it only when the background writeback cannot catch-up. This avoids
|
| 1708 | * (excessively) small writeouts when the wb limits are ramping up in case of
|
| 1709 | * !strictlimit.
|
| 1710 | *
|
| 1711 | * In strictlimit case make decision based on the wb counters and limits. Small
|
| 1712 | * writeouts when the wb limits are ramping up are the price we consciously pay
|
| 1713 | * for strictlimit-ing.
|
| 1714 |
|
| 1715 | static void domain_dirty_freerun(struct dirty_throttle_control *dtc,
|
| 1716 | bool strictlimit)
|
| 1717 | {
|
| 1718 | unsigned long dirty, thresh, bg_thresh;
|
| 1719 |
|
| 1720 | if (unlikely(strictlimit)) {
|
| 1721 | wb_dirty_limits(dtc);
|
| 1722 | dirty = dtc->wb_dirty;
|
| 1723 | thresh = dtc->wb_thresh;
|
| 1724 | bg_thresh = dtc->wb_bg_thresh;
|
| 1725 | } else {
|
| 1726 | dirty = dtc->dirty;
|
| 1727 | thresh = dtc->thresh;
|
| 1728 | bg_thresh = dtc->bg_thresh;
|
| 1729 | }
|
| 1730 | dtc->freerun = dirty <= dirty_freerun_ceiling(thresh, bg_thresh);
|
| 1731 | }
|
| 1732 |
|
| 1733 | static void balance_domain_limits(struct dirty_throttle_control *dtc,
|
| 1734 | bool strictlimit)
|
| 1735 | {
|
| 1736 | domain_dirty_avail(dtc, true);
|
| 1737 | domain_dirty_limits(dtc);
|
| 1738 | domain_dirty_freerun(dtc, strictlimit);
|
| 1739 | }
|
| 1740 |
|
| 1741 | static void wb_dirty_freerun(struct dirty_throttle_control *dtc,
|
| 1742 | bool strictlimit)
|
| 1743 | {
|
| 1744 | dtc->freerun = false;
|
| 1745 |
|
| 1746 |
|
| 1747 | if (strictlimit)
|
| 1748 | return;
|
| 1749 |
|
| 1750 | wb_dirty_limits(dtc);
|
| 1751 | |
| 1752 | * LOCAL_THROTTLE tasks must not be throttled when below the per-wb
|
| 1753 | * freerun ceiling.
|
| 1754 |
|
| 1755 | if (!(current->flags & PF_LOCAL_THROTTLE))
|
| 1756 | return;
|
| 1757 |
|
| 1758 | dtc->freerun = dtc->wb_dirty <
|
| 1759 | dirty_freerun_ceiling(dtc->wb_thresh, dtc->wb_bg_thresh);
|
| 1760 | }
|
| 1761 |
|
| 1762 | static inline void wb_dirty_exceeded(struct dirty_throttle_control *dtc,
|
| 1763 | bool strictlimit)
|
| 1764 | {
|
| 1765 | dtc->dirty_exceeded = (dtc->wb_dirty > dtc->wb_thresh) &&
|
| 1766 | ((dtc->dirty > dtc->thresh) || strictlimit);
|
| 1767 | }
|
| 1768 |
|
| 1769 | |
| 1770 | * The limits fields dirty_exceeded and pos_ratio won't be updated if wb is
|
| 1771 | * in freerun state. Please don't use these invalid fields in freerun case.
|
| 1772 |
|
| 1773 | static void balance_wb_limits(struct dirty_throttle_control *dtc,
|
| 1774 | bool strictlimit)
|
| 1775 | {
|
| 1776 | wb_dirty_freerun(dtc, strictlimit);
|
| 1777 | if (dtc->freerun)
|
| 1778 | return;
|
| 1779 |
|
| 1780 | wb_dirty_exceeded(dtc, strictlimit);
|
| 1781 | wb_position_ratio(dtc);
|
| 1782 | }
|
| 1783 |
|
| 1784 | |
| 1785 | * balance_dirty_pages() must be called by processes which are generating dirty
|
| 1786 | * data. It looks at the number of dirty pages in the machine and will force
|
| 1787 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
|
| 1788 | * If we're over `background_thresh' then the writeback threads are woken to
|
| 1789 | * perform some writeout.
|
| 1790 |
|
| 1791 | static int balance_dirty_pages(struct bdi_writeback *wb,
|
| 1792 | unsigned long pages_dirtied, unsigned int flags)
|
| 1793 | {
|
| 1794 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
|
| 1795 | struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
|
| 1796 | struct dirty_throttle_control * const gdtc = &gdtc_stor;
|
| 1797 | struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
|
| 1798 | &mdtc_stor : NULL;
|
| 1799 | struct dirty_throttle_control *sdtc;
|
| 1800 | unsigned long nr_dirty;
|
| 1801 | long period;
|
| 1802 | long pause;
|
| 1803 | long max_pause;
|
| 1804 | long min_pause;
|
| 1805 | int nr_dirtied_pause;
|
| 1806 | unsigned long task_ratelimit;
|
| 1807 | unsigned long dirty_ratelimit;
|
| 1808 | struct backing_dev_info *bdi = wb->bdi;
|
| 1809 | bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
|
| 1810 | unsigned long start_time = jiffies;
|
| 1811 | int ret = 0;
|
| 1812 |
|
| 1813 | for (;;) {
|
| 1814 | unsigned long now = jiffies;
|
| 1815 |
|
| 1816 | nr_dirty = global_node_page_state(NR_FILE_DIRTY);
|
| 1817 |
|
| 1818 | balance_domain_limits(gdtc, strictlimit);
|
| 1819 | if (mdtc) {
|
| 1820 | |
| 1821 | * If @wb belongs to !root memcg, repeat the same
|
| 1822 | * basic calculations for the memcg domain.
|
| 1823 |
|
| 1824 | balance_domain_limits(mdtc, strictlimit);
|
| 1825 | }
|
| 1826 |
|
| 1827 | if (!writeback_in_progress(wb) &&
|
| 1828 | (nr_dirty > gdtc->bg_thresh ||
|
| 1829 | (strictlimit && gdtc->wb_dirty > gdtc->wb_bg_thresh)))
|
| 1830 | wb_start_background_writeback(wb);
|
| 1831 |
|
| 1832 | |
| 1833 | * If memcg domain is in effect, @dirty should be under
|
| 1834 | * both global and memcg freerun ceilings.
|
| 1835 |
|
| 1836 | if (gdtc->freerun && (!mdtc || mdtc->freerun)) {
|
| 1837 | unsigned long intv;
|
| 1838 | unsigned long m_intv;
|
| 1839 |
|
| 1840 | free_running:
|
| 1841 | intv = domain_poll_intv(gdtc, strictlimit);
|
| 1842 | m_intv = ULONG_MAX;
|
| 1843 |
|
| 1844 | current->dirty_paused_when = now;
|
| 1845 | current->nr_dirtied = 0;
|
| 1846 | if (mdtc)
|
| 1847 | m_intv = domain_poll_intv(mdtc, strictlimit);
|
| 1848 | current->nr_dirtied_pause = min(intv, m_intv);
|
| 1849 | break;
|
| 1850 | }
|
| 1851 |
|
| 1852 | |
| 1853 | * Unconditionally start background writeback if it's not
|
| 1854 | * already in progress. We need to do this because the global
|
| 1855 | * dirty threshold check above (nr_dirty > gdtc->bg_thresh)
|
| 1856 | * doesn't account for the memcg-based throttling case. memcg
|
| 1857 | * uses its own dirty count and thresholds and can trigger
|
| 1858 | * throttling even when global nr_dirty < gdtc->bg_thresh
|
| 1859 | *
|
| 1860 | * Writeback needs to be started else the writer stalls in the
|
| 1861 | * throttle loop waiting for dirty pages to be written back
|
| 1862 | * while no writeback is running.
|
| 1863 |
|
| 1864 | if (unlikely(!writeback_in_progress(wb)))
|
| 1865 | wb_start_background_writeback(wb);
|
| 1866 |
|
| 1867 | mem_cgroup_flush_foreign(wb);
|
| 1868 |
|
| 1869 | |
| 1870 | * Calculate global domain's pos_ratio and select the
|
| 1871 | * global dtc by default.
|
| 1872 |
|
| 1873 | balance_wb_limits(gdtc, strictlimit);
|
| 1874 | if (gdtc->freerun)
|
| 1875 | goto free_running;
|
| 1876 | sdtc = gdtc;
|
| 1877 |
|
| 1878 | if (mdtc) {
|
| 1879 | |
| 1880 | * If memcg domain is in effect, calculate its
|
| 1881 | * pos_ratio. @wb should satisfy constraints from
|
| 1882 | * both global and memcg domains. Choose the one
|
| 1883 | * w/ lower pos_ratio.
|
| 1884 |
|
| 1885 | balance_wb_limits(mdtc, strictlimit);
|
| 1886 | if (mdtc->freerun)
|
| 1887 | goto free_running;
|
| 1888 | if (mdtc->pos_ratio < gdtc->pos_ratio)
|
| 1889 | sdtc = mdtc;
|
| 1890 | }
|
| 1891 |
|
| 1892 | wb->dirty_exceeded = gdtc->dirty_exceeded ||
|
| 1893 | (mdtc && mdtc->dirty_exceeded);
|
| 1894 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
|
| 1895 | BANDWIDTH_INTERVAL))
|
| 1896 | __wb_update_bandwidth(gdtc, mdtc, true);
|
| 1897 |
|
| 1898 |
|
| 1899 | dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit);
|
| 1900 | task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
|
| 1901 | RATELIMIT_CALC_SHIFT;
|
| 1902 | max_pause = wb_max_pause(wb, sdtc->wb_dirty);
|
| 1903 | min_pause = wb_min_pause(wb, max_pause,
|
| 1904 | task_ratelimit, dirty_ratelimit,
|
| 1905 | &nr_dirtied_pause);
|
| 1906 |
|
| 1907 | if (unlikely(task_ratelimit == 0)) {
|
| 1908 | period = max_pause;
|
| 1909 | pause = max_pause;
|
| 1910 | goto pause;
|
| 1911 | }
|
| 1912 | period = HZ * pages_dirtied / task_ratelimit;
|
| 1913 | pause = period;
|
| 1914 | if (current->dirty_paused_when)
|
| 1915 | pause -= now - current->dirty_paused_when;
|
| 1916 | |
| 1917 | * For less than 1s think time (ext3/4 may block the dirtier
|
| 1918 | * for up to 800ms from time to time on 1-HDD; so does xfs,
|
| 1919 | * however at much less frequency), try to compensate it in
|
| 1920 | * future periods by updating the virtual time; otherwise just
|
| 1921 | * do a reset, as it may be a light dirtier.
|
| 1922 |
|
| 1923 | if (pause < min_pause) {
|
| 1924 | trace_balance_dirty_pages(wb,
|
| 1925 | sdtc,
|
| 1926 | dirty_ratelimit,
|
| 1927 | task_ratelimit,
|
| 1928 | pages_dirtied,
|
| 1929 | period,
|
| 1930 | min(pause, 0L),
|
| 1931 | start_time);
|
| 1932 | if (pause < -HZ) {
|
| 1933 | current->dirty_paused_when = now;
|
| 1934 | current->nr_dirtied = 0;
|
| 1935 | } else if (period) {
|
| 1936 | current->dirty_paused_when += period;
|
| 1937 | current->nr_dirtied = 0;
|
| 1938 | } else if (current->nr_dirtied_pause <= pages_dirtied)
|
| 1939 | current->nr_dirtied_pause += pages_dirtied;
|
| 1940 | break;
|
| 1941 | }
|
| 1942 | if (unlikely(pause > max_pause)) {
|
| 1943 |
|
| 1944 | now += min(pause - max_pause, max_pause);
|
| 1945 | pause = max_pause;
|
| 1946 | }
|
| 1947 |
|
| 1948 | pause:
|
| 1949 | trace_balance_dirty_pages(wb,
|
| 1950 | sdtc,
|
| 1951 | dirty_ratelimit,
|
| 1952 | task_ratelimit,
|
| 1953 | pages_dirtied,
|
| 1954 | period,
|
| 1955 | pause,
|
| 1956 | start_time);
|
| 1957 | if (flags & BDP_ASYNC) {
|
| 1958 | ret = -EAGAIN;
|
| 1959 | break;
|
| 1960 | }
|
| 1961 | __set_current_state(TASK_KILLABLE);
|
| 1962 | bdi->last_bdp_sleep = jiffies;
|
| 1963 | io_schedule_timeout(pause);
|
| 1964 |
|
| 1965 | current->dirty_paused_when = now + pause;
|
| 1966 | current->nr_dirtied = 0;
|
| 1967 | current->nr_dirtied_pause = nr_dirtied_pause;
|
| 1968 |
|
| 1969 | |
| 1970 | * This is typically equal to (dirty < thresh) and can also
|
| 1971 | * keep "1000+ dd on a slow USB stick" under control.
|
| 1972 |
|
| 1973 | if (task_ratelimit)
|
| 1974 | break;
|
| 1975 |
|
| 1976 | |
| 1977 | * In the case of an unresponsive NFS server and the NFS dirty
|
| 1978 | * pages exceeds dirty_thresh, give the other good wb's a pipe
|
| 1979 | * to go through, so that tasks on them still remain responsive.
|
| 1980 | *
|
| 1981 | * In theory 1 page is enough to keep the consumer-producer
|
| 1982 | * pipe going: the flusher cleans 1 page => the task dirties 1
|
| 1983 | * more page. However wb_dirty has accounting errors. So use
|
| 1984 | * the larger and more IO friendly wb_stat_error.
|
| 1985 |
|
| 1986 | if (sdtc->wb_dirty <= wb_stat_error())
|
| 1987 | break;
|
| 1988 |
|
| 1989 | if (fatal_signal_pending(current))
|
| 1990 | break;
|
| 1991 | }
|
| 1992 | return ret;
|
| 1993 | }
|
| 1994 |
|
| 1995 | static DEFINE_PER_CPU(int, bdp_ratelimits);
|
| 1996 |
|
| 1997 | |
| 1998 | * Normal tasks are throttled by
|
| 1999 | * loop {
|
| 2000 | * dirty tsk->nr_dirtied_pause pages;
|
| 2001 | * take a snap in balance_dirty_pages();
|
| 2002 | * }
|
| 2003 | * However there is a worst case. If every task exit immediately when dirtied
|
| 2004 | * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
|
| 2005 | * called to throttle the page dirties. The solution is to save the not yet
|
| 2006 | * throttled page dirties in dirty_throttle_leaks on task exit and charge them
|
| 2007 | * randomly into the running tasks. This works well for the above worst case,
|
| 2008 | * as the new task will pick up and accumulate the old task's leaked dirty
|
| 2009 | * count and eventually get throttled.
|
| 2010 |
|
| 2011 | DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
|
| 2012 |
|
| 2013 | |
| 2014 | * balance_dirty_pages_ratelimited_flags - Balance dirty memory state.
|
| 2015 | * @mapping: address_space which was dirtied.
|
| 2016 | * @flags: BDP flags.
|
| 2017 | *
|
| 2018 | * Processes which are dirtying memory should call in here once for each page
|
| 2019 | * which was newly dirtied. The function will periodically check the system's
|
| 2020 | * dirty state and will initiate writeback if needed.
|
| 2021 | *
|
| 2022 | * See balance_dirty_pages_ratelimited() for details.
|
| 2023 | *
|
| 2024 | * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to
|
| 2025 | * indicate that memory is out of balance and the caller must wait
|
| 2026 | * for I/O to complete. Otherwise, it will return 0 to indicate
|
| 2027 | * that either memory was already in balance, or it was able to sleep
|
| 2028 | * until the amount of dirty memory returned to balance.
|
| 2029 |
|
| 2030 | int balance_dirty_pages_ratelimited_flags(struct address_space *mapping,
|
| 2031 | unsigned int flags)
|
| 2032 | {
|
| 2033 | struct inode *inode = mapping->host;
|
| 2034 | struct backing_dev_info *bdi = inode_to_bdi(inode);
|
| 2035 | struct bdi_writeback *wb = NULL;
|
| 2036 | int ratelimit;
|
| 2037 | int ret = 0;
|
| 2038 | int *p;
|
| 2039 |
|
| 2040 | if (!(bdi->capabilities & BDI_CAP_WRITEBACK))
|
| 2041 | return ret;
|
| 2042 |
|
| 2043 | if (inode_cgwb_enabled(inode))
|
| 2044 | wb = wb_get_create_current(bdi, GFP_KERNEL);
|
| 2045 | if (!wb)
|
| 2046 | wb = &bdi->wb;
|
| 2047 |
|
| 2048 | ratelimit = current->nr_dirtied_pause;
|
| 2049 | if (wb->dirty_exceeded)
|
| 2050 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
|
| 2051 |
|
| 2052 | preempt_disable();
|
| 2053 | |
| 2054 | * This prevents one CPU to accumulate too many dirtied pages without
|
| 2055 | * calling into balance_dirty_pages(), which can happen when there are
|
| 2056 | * 1000+ tasks, all of them start dirtying pages at exactly the same
|
| 2057 | * time, hence all honoured too large initial task->nr_dirtied_pause.
|
| 2058 |
|
| 2059 | p = this_cpu_ptr(&bdp_ratelimits);
|
| 2060 | if (unlikely(current->nr_dirtied >= ratelimit))
|
| 2061 | *p = 0;
|
| 2062 | else if (unlikely(*p >= ratelimit_pages)) {
|
| 2063 | *p = 0;
|
| 2064 | ratelimit = 0;
|
| 2065 | }
|
| 2066 | |
| 2067 | * Pick up the dirtied pages by the exited tasks. This avoids lots of
|
| 2068 | * short-lived tasks (eg. gcc invocations in a kernel build) escaping
|
| 2069 | * the dirty throttling and livelock other long-run dirtiers.
|
| 2070 |
|
| 2071 | p = this_cpu_ptr(&dirty_throttle_leaks);
|
| 2072 | if (*p > 0 && current->nr_dirtied < ratelimit) {
|
| 2073 | unsigned long nr_pages_dirtied;
|
| 2074 | nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
|
| 2075 | *p -= nr_pages_dirtied;
|
| 2076 | current->nr_dirtied += nr_pages_dirtied;
|
| 2077 | }
|
| 2078 | preempt_enable();
|
| 2079 |
|
| 2080 | if (unlikely(current->nr_dirtied >= ratelimit))
|
| 2081 | ret = balance_dirty_pages(wb, current->nr_dirtied, flags);
|
| 2082 |
|
| 2083 | wb_put(wb);
|
| 2084 | return ret;
|
| 2085 | }
|
| 2086 | EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags);
|
| 2087 |
|
| 2088 | |
| 2089 | * balance_dirty_pages_ratelimited - balance dirty memory state.
|
| 2090 | * @mapping: address_space which was dirtied.
|
| 2091 | *
|
| 2092 | * Processes which are dirtying memory should call in here once for each page
|
| 2093 | * which was newly dirtied. The function will periodically check the system's
|
| 2094 | * dirty state and will initiate writeback if needed.
|
| 2095 | *
|
| 2096 | * Once we're over the dirty memory limit we decrease the ratelimiting
|
| 2097 | * by a lot, to prevent individual processes from overshooting the limit
|
| 2098 | * by (ratelimit_pages) each.
|
| 2099 |
|
| 2100 | void balance_dirty_pages_ratelimited(struct address_space *mapping)
|
| 2101 | {
|
| 2102 | balance_dirty_pages_ratelimited_flags(mapping, 0);
|
| 2103 | }
|
| 2104 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
|
| 2105 |
|
| 2106 | |
| 2107 | * Similar to wb_dirty_limits, wb_bg_dirty_limits also calculates dirty
|
| 2108 | * and thresh, but it's for background writeback.
|
| 2109 |
|
| 2110 | static void wb_bg_dirty_limits(struct dirty_throttle_control *dtc)
|
| 2111 | {
|
| 2112 | struct bdi_writeback *wb = dtc->wb;
|
| 2113 |
|
| 2114 | dtc->wb_bg_thresh = __wb_calc_thresh(dtc, dtc->bg_thresh);
|
| 2115 | if (dtc->wb_bg_thresh < 2 * wb_stat_error())
|
| 2116 | dtc->wb_dirty = wb_stat_sum(wb, WB_RECLAIMABLE);
|
| 2117 | else
|
| 2118 | dtc->wb_dirty = wb_stat(wb, WB_RECLAIMABLE);
|
| 2119 | }
|
| 2120 |
|
| 2121 | static bool domain_over_bg_thresh(struct dirty_throttle_control *dtc)
|
| 2122 | {
|
| 2123 | domain_dirty_avail(dtc, false);
|
| 2124 | domain_dirty_limits(dtc);
|
| 2125 | if (dtc->dirty > dtc->bg_thresh)
|
| 2126 | return true;
|
| 2127 |
|
| 2128 | wb_bg_dirty_limits(dtc);
|
| 2129 | if (dtc->wb_dirty > dtc->wb_bg_thresh)
|
| 2130 | return true;
|
| 2131 |
|
| 2132 | return false;
|
| 2133 | }
|
| 2134 |
|
| 2135 | |
| 2136 | * wb_over_bg_thresh - does @wb need to be written back?
|
| 2137 | * @wb: bdi_writeback of interest
|
| 2138 | *
|
| 2139 | * Determines whether background writeback should keep writing @wb or it's
|
| 2140 | * clean enough.
|
| 2141 | *
|
| 2142 | * Return: %true if writeback should continue.
|
| 2143 |
|
| 2144 | bool wb_over_bg_thresh(struct bdi_writeback *wb)
|
| 2145 | {
|
| 2146 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
|
| 2147 | struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) };
|
| 2148 |
|
| 2149 | if (domain_over_bg_thresh(&gdtc))
|
| 2150 | return true;
|
| 2151 |
|
| 2152 | if (mdtc_valid(&mdtc))
|
| 2153 | return domain_over_bg_thresh(&mdtc);
|
| 2154 |
|
| 2155 | return false;
|
| 2156 | }
|
| 2157 |
|
| 2158 | #ifdef CONFIG_SYSCTL
|
| 2159 | |
| 2160 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
|
| 2161 |
|
| 2162 | static int dirty_writeback_centisecs_handler(const struct ctl_table *table, int write,
|
| 2163 | void *buffer, size_t *length, loff_t *ppos)
|
| 2164 | {
|
| 2165 | unsigned int old_interval = dirty_writeback_interval;
|
| 2166 | int ret;
|
| 2167 |
|
| 2168 | ret = proc_dointvec(table, write, buffer, length, ppos);
|
| 2169 |
|
| 2170 | |
| 2171 | * Writing 0 to dirty_writeback_interval will disable periodic writeback
|
| 2172 | * and a different non-zero value will wakeup the writeback threads.
|
| 2173 | * wb_wakeup_delayed() would be more appropriate, but it's a pain to
|
| 2174 | * iterate over all bdis and wbs.
|
| 2175 | * The reason we do this is to make the change take effect immediately.
|
| 2176 |
|
| 2177 | if (!ret && write && dirty_writeback_interval &&
|
| 2178 | dirty_writeback_interval != old_interval)
|
| 2179 | wakeup_flusher_threads(WB_REASON_PERIODIC);
|
| 2180 |
|
| 2181 | return ret;
|
| 2182 | }
|
| 2183 | #endif
|
| 2184 |
|
| 2185 | |
| 2186 | * If ratelimit_pages is too high then we can get into dirty-data overload
|
| 2187 | * if a large number of processes all perform writes at the same time.
|
| 2188 | *
|
| 2189 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are
|
| 2190 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
|
| 2191 | * thresholds.
|
| 2192 |
|
| 2193 |
|
| 2194 | void writeback_set_ratelimit(void)
|
| 2195 | {
|
| 2196 | struct wb_domain *dom = &global_wb_domain;
|
| 2197 | unsigned long background_thresh;
|
| 2198 | unsigned long dirty_thresh;
|
| 2199 |
|
| 2200 | global_dirty_limits(&background_thresh, &dirty_thresh);
|
| 2201 | dom->dirty_limit = dirty_thresh;
|
| 2202 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
|
| 2203 | if (ratelimit_pages < 16)
|
| 2204 | ratelimit_pages = 16;
|
| 2205 | }
|
| 2206 |
|
| 2207 | static int page_writeback_cpu_online(unsigned int cpu)
|
| 2208 | {
|
| 2209 | writeback_set_ratelimit();
|
| 2210 | return 0;
|
| 2211 | }
|
| 2212 |
|
| 2213 | #ifdef CONFIG_SYSCTL
|
| 2214 |
|
| 2215 | static int laptop_mode;
|
| 2216 | static int laptop_mode_handler(const struct ctl_table *table, int write,
|
| 2217 | void *buffer, size_t *lenp, loff_t *ppos)
|
| 2218 | {
|
| 2219 | int ret = proc_dointvec_jiffies(table, write, buffer, lenp, ppos);
|
| 2220 |
|
| 2221 | if (!ret && write)
|
| 2222 | pr_warn("%s: vm.laptop_mode is deprecated. Ignoring setting.\n",
|
| 2223 | current->comm);
|
| 2224 |
|
| 2225 | return ret;
|
| 2226 | }
|
| 2227 |
|
| 2228 |
|
| 2229 | static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
|
| 2230 |
|
| 2231 | static const struct ctl_table vm_page_writeback_sysctls[] = {
|
| 2232 | {
|
| 2233 | .procname = "dirty_background_ratio",
|
| 2234 | .data = &dirty_background_ratio,
|
| 2235 | .maxlen = sizeof(dirty_background_ratio),
|
| 2236 | .mode = 0644,
|
| 2237 | .proc_handler = dirty_background_ratio_handler,
|
| 2238 | .extra1 = SYSCTL_ZERO,
|
| 2239 | .extra2 = SYSCTL_ONE_HUNDRED,
|
| 2240 | },
|
| 2241 | {
|
| 2242 | .procname = "dirty_background_bytes",
|
| 2243 | .data = &dirty_background_bytes,
|
| 2244 | .maxlen = sizeof(dirty_background_bytes),
|
| 2245 | .mode = 0644,
|
| 2246 | .proc_handler = dirty_background_bytes_handler,
|
| 2247 | .extra1 = SYSCTL_LONG_ONE,
|
| 2248 | },
|
| 2249 | {
|
| 2250 | .procname = "dirty_ratio",
|
| 2251 | .data = &vm_dirty_ratio,
|
| 2252 | .maxlen = sizeof(vm_dirty_ratio),
|
| 2253 | .mode = 0644,
|
| 2254 | .proc_handler = dirty_ratio_handler,
|
| 2255 | .extra1 = SYSCTL_ZERO,
|
| 2256 | .extra2 = SYSCTL_ONE_HUNDRED,
|
| 2257 | },
|
| 2258 | {
|
| 2259 | .procname = "dirty_bytes",
|
| 2260 | .data = &vm_dirty_bytes,
|
| 2261 | .maxlen = sizeof(vm_dirty_bytes),
|
| 2262 | .mode = 0644,
|
| 2263 | .proc_handler = dirty_bytes_handler,
|
| 2264 | .extra1 = (void *)&dirty_bytes_min,
|
| 2265 | },
|
| 2266 | {
|
| 2267 | .procname = "dirty_writeback_centisecs",
|
| 2268 | .data = &dirty_writeback_interval,
|
| 2269 | .maxlen = sizeof(dirty_writeback_interval),
|
| 2270 | .mode = 0644,
|
| 2271 | .proc_handler = dirty_writeback_centisecs_handler,
|
| 2272 | },
|
| 2273 | {
|
| 2274 | .procname = "dirty_expire_centisecs",
|
| 2275 | .data = &dirty_expire_interval,
|
| 2276 | .maxlen = sizeof(dirty_expire_interval),
|
| 2277 | .mode = 0644,
|
| 2278 | .proc_handler = proc_dointvec_minmax,
|
| 2279 | .extra1 = SYSCTL_ZERO,
|
| 2280 | },
|
| 2281 | #ifdef CONFIG_HIGHMEM
|
| 2282 | {
|
| 2283 | .procname = "highmem_is_dirtyable",
|
| 2284 | .data = &vm_highmem_is_dirtyable,
|
| 2285 | .maxlen = sizeof(vm_highmem_is_dirtyable),
|
| 2286 | .mode = 0644,
|
| 2287 | .proc_handler = proc_dointvec_minmax,
|
| 2288 | .extra1 = SYSCTL_ZERO,
|
| 2289 | .extra2 = SYSCTL_ONE,
|
| 2290 | },
|
| 2291 | #endif
|
| 2292 | {
|
| 2293 | .procname = "laptop_mode",
|
| 2294 | .data = &laptop_mode,
|
| 2295 | .maxlen = sizeof(laptop_mode),
|
| 2296 | .mode = 0644,
|
| 2297 | .proc_handler = laptop_mode_handler,
|
| 2298 | },
|
| 2299 | };
|
| 2300 | #endif
|
| 2301 |
|
| 2302 | |
| 2303 | * Called early on to tune the page writeback dirty limits.
|
| 2304 | *
|
| 2305 | * We used to scale dirty pages according to how total memory
|
| 2306 | * related to pages that could be allocated for buffers.
|
| 2307 | *
|
| 2308 | * However, that was when we used "dirty_ratio" to scale with
|
| 2309 | * all memory, and we don't do that any more. "dirty_ratio"
|
| 2310 | * is now applied to total non-HIGHPAGE memory, and as such we can't
|
| 2311 | * get into the old insane situation any more where we had
|
| 2312 | * large amounts of dirty pages compared to a small amount of
|
| 2313 | * non-HIGHMEM memory.
|
| 2314 | *
|
| 2315 | * But we might still want to scale the dirty_ratio by how
|
| 2316 | * much memory the box has..
|
| 2317 |
|
| 2318 | void __init page_writeback_init(void)
|
| 2319 | {
|
| 2320 | BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
|
| 2321 |
|
| 2322 | cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
|
| 2323 | page_writeback_cpu_online, NULL);
|
| 2324 | cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
|
| 2325 | page_writeback_cpu_online);
|
| 2326 | #ifdef CONFIG_SYSCTL
|
| 2327 | register_sysctl_init("vm", vm_page_writeback_sysctls);
|
| 2328 | #endif
|
| 2329 | }
|
| 2330 |
|
| 2331 | |
| 2332 | * tag_pages_for_writeback - tag pages to be written by writeback
|
| 2333 | * @mapping: address space structure to write
|
| 2334 | * @start: starting page index
|
| 2335 | * @end: ending page index (inclusive)
|
| 2336 | *
|
| 2337 | * This function scans the page range from @start to @end (inclusive) and tags
|
| 2338 | * all pages that have DIRTY tag set with a special TOWRITE tag. The caller
|
| 2339 | * can then use the TOWRITE tag to identify pages eligible for writeback.
|
| 2340 | * This mechanism is used to avoid livelocking of writeback by a process
|
| 2341 | * steadily creating new dirty pages in the file (thus it is important for this
|
| 2342 | * function to be quick so that it can tag pages faster than a dirtying process
|
| 2343 | * can create them).
|
| 2344 |
|
| 2345 | void tag_pages_for_writeback(struct address_space *mapping,
|
| 2346 | pgoff_t start, pgoff_t end)
|
| 2347 | {
|
| 2348 | XA_STATE(xas, &mapping->i_pages, start);
|
| 2349 | unsigned int tagged = 0;
|
| 2350 | void *page;
|
| 2351 |
|
| 2352 | xas_lock_irq(&xas);
|
| 2353 | xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) {
|
| 2354 | xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE);
|
| 2355 | if (++tagged % XA_CHECK_SCHED)
|
| 2356 | continue;
|
| 2357 |
|
| 2358 | xas_pause(&xas);
|
| 2359 | xas_unlock_irq(&xas);
|
| 2360 | cond_resched();
|
| 2361 | xas_lock_irq(&xas);
|
| 2362 | }
|
| 2363 | xas_unlock_irq(&xas);
|
| 2364 | }
|
| 2365 | EXPORT_SYMBOL(tag_pages_for_writeback);
|
| 2366 |
|
| 2367 | static bool folio_prepare_writeback(struct address_space *mapping,
|
| 2368 | struct writeback_control *wbc, struct folio *folio)
|
| 2369 | {
|
| 2370 | |
| 2371 | * Folio truncated or invalidated. We can freely skip it then,
|
| 2372 | * even for data integrity operations: the folio has disappeared
|
| 2373 | * concurrently, so there could be no real expectation of this
|
| 2374 | * data integrity operation even if there is now a new, dirty
|
| 2375 | * folio at the same pagecache index.
|
| 2376 |
|
| 2377 | if (unlikely(folio->mapping != mapping))
|
| 2378 | return false;
|
| 2379 |
|
| 2380 | |
| 2381 | * Did somebody else write it for us?
|
| 2382 |
|
| 2383 | if (!folio_test_dirty(folio))
|
| 2384 | return false;
|
| 2385 |
|
| 2386 | if (folio_test_writeback(folio)) {
|
| 2387 | if (wbc->sync_mode == WB_SYNC_NONE)
|
| 2388 | return false;
|
| 2389 | folio_wait_writeback(folio);
|
| 2390 | }
|
| 2391 | BUG_ON(folio_test_writeback(folio));
|
| 2392 |
|
| 2393 | if (!folio_clear_dirty_for_io(folio))
|
| 2394 | return false;
|
| 2395 |
|
| 2396 | return true;
|
| 2397 | }
|
| 2398 |
|
| 2399 |
|
| 2400 | static pgoff_t wbc_end(struct writeback_control *wbc)
|
| 2401 | {
|
| 2402 | if (wbc->range_cyclic)
|
| 2403 | return -1;
|
| 2404 | return wbc->range_end >> PAGE_SHIFT;
|
| 2405 | }
|
| 2406 |
|
| 2407 | static struct folio *writeback_get_folio(struct address_space *mapping,
|
| 2408 | struct writeback_control *wbc)
|
| 2409 | {
|
| 2410 | struct folio *folio;
|
| 2411 |
|
| 2412 | retry:
|
| 2413 | folio = folio_batch_next(&wbc->fbatch);
|
| 2414 | if (!folio) {
|
| 2415 | folio_batch_release(&wbc->fbatch);
|
| 2416 | cond_resched();
|
| 2417 | filemap_get_folios_tag(mapping, &wbc->index, wbc_end(wbc),
|
| 2418 | wbc_to_tag(wbc), &wbc->fbatch);
|
| 2419 | folio = folio_batch_next(&wbc->fbatch);
|
| 2420 | if (!folio)
|
| 2421 | return NULL;
|
| 2422 | }
|
| 2423 |
|
| 2424 | folio_lock(folio);
|
| 2425 | if (unlikely(!folio_prepare_writeback(mapping, wbc, folio))) {
|
| 2426 | folio_unlock(folio);
|
| 2427 | goto retry;
|
| 2428 | }
|
| 2429 |
|
| 2430 | trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
|
| 2431 | return folio;
|
| 2432 | }
|
| 2433 |
|
| 2434 | |
| 2435 | * writeback_iter - iterate folio of a mapping for writeback
|
| 2436 | * @mapping: address space structure to write
|
| 2437 | * @wbc: writeback context
|
| 2438 | * @folio: previously iterated folio (%NULL to start)
|
| 2439 | * @error: in-out pointer for writeback errors (see below)
|
| 2440 | *
|
| 2441 | * This function returns the next folio for the writeback operation described by
|
| 2442 | * @wbc on @mapping and should be called in a while loop in the ->writepages
|
| 2443 | * implementation.
|
| 2444 | *
|
| 2445 | * To start the writeback operation, %NULL is passed in the @folio argument, and
|
| 2446 | * for every subsequent iteration the folio returned previously should be passed
|
| 2447 | * back in.
|
| 2448 | *
|
| 2449 | * If there was an error in the per-folio writeback inside the writeback_iter()
|
| 2450 | * loop, @error should be set to the error value.
|
| 2451 | *
|
| 2452 | * Once the writeback described in @wbc has finished, this function will return
|
| 2453 | * %NULL and if there was an error in any iteration restore it to @error.
|
| 2454 | *
|
| 2455 | * Note: callers should not manually break out of the loop using break or goto
|
| 2456 | * but must keep calling writeback_iter() until it returns %NULL.
|
| 2457 | *
|
| 2458 | * Return: the folio to write or %NULL if the loop is done.
|
| 2459 |
|
| 2460 | struct folio *writeback_iter(struct address_space *mapping,
|
| 2461 | struct writeback_control *wbc, struct folio *folio, int *error)
|
| 2462 | {
|
| 2463 | if (!folio) {
|
| 2464 | folio_batch_init(&wbc->fbatch);
|
| 2465 | wbc->saved_err = *error = 0;
|
| 2466 |
|
| 2467 | |
| 2468 | * For range cyclic writeback we remember where we stopped so
|
| 2469 | * that we can continue where we stopped.
|
| 2470 | *
|
| 2471 | * For non-cyclic writeback we always start at the beginning of
|
| 2472 | * the passed in range.
|
| 2473 |
|
| 2474 | if (wbc->range_cyclic)
|
| 2475 | wbc->index = mapping->writeback_index;
|
| 2476 | else
|
| 2477 | wbc->index = wbc->range_start >> PAGE_SHIFT;
|
| 2478 |
|
| 2479 | |
| 2480 | * To avoid livelocks when other processes dirty new pages, we
|
| 2481 | * first tag pages which should be written back and only then
|
| 2482 | * start writing them.
|
| 2483 | *
|
| 2484 | * For data-integrity writeback we have to be careful so that we
|
| 2485 | * do not miss some pages (e.g., because some other process has
|
| 2486 | * cleared the TOWRITE tag we set). The rule we follow is that
|
| 2487 | * TOWRITE tag can be cleared only by the process clearing the
|
| 2488 | * DIRTY tag (and submitting the page for I/O).
|
| 2489 |
|
| 2490 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
|
| 2491 | tag_pages_for_writeback(mapping, wbc->index,
|
| 2492 | wbc_end(wbc));
|
| 2493 | } else {
|
| 2494 | wbc->nr_to_write -= folio_nr_pages(folio);
|
| 2495 |
|
| 2496 | WARN_ON_ONCE(*error > 0);
|
| 2497 |
|
| 2498 | |
| 2499 | * For integrity writeback we have to keep going until we have
|
| 2500 | * written all the folios we tagged for writeback above, even if
|
| 2501 | * we run past wbc->nr_to_write or encounter errors.
|
| 2502 | * We stash away the first error we encounter in wbc->saved_err
|
| 2503 | * so that it can be retrieved when we're done. This is because
|
| 2504 | * the file system may still have state to clear for each folio.
|
| 2505 | *
|
| 2506 | * For background writeback we exit as soon as we run past
|
| 2507 | * wbc->nr_to_write or encounter the first error.
|
| 2508 |
|
| 2509 | if (wbc->sync_mode == WB_SYNC_ALL) {
|
| 2510 | if (*error && !wbc->saved_err)
|
| 2511 | wbc->saved_err = *error;
|
| 2512 | } else {
|
| 2513 | if (*error || wbc->nr_to_write <= 0)
|
| 2514 | goto done;
|
| 2515 | }
|
| 2516 | }
|
| 2517 |
|
| 2518 | folio = writeback_get_folio(mapping, wbc);
|
| 2519 | if (!folio) {
|
| 2520 | |
| 2521 | * To avoid deadlocks between range_cyclic writeback and callers
|
| 2522 | * that hold folios in writeback to aggregate I/O until
|
| 2523 | * the writeback iteration finishes, we do not loop back to the
|
| 2524 | * start of the file. Doing so causes a folio lock/folio
|
| 2525 | * writeback access order inversion - we should only ever lock
|
| 2526 | * multiple folios in ascending folio->index order, and looping
|
| 2527 | * back to the start of the file violates that rule and causes
|
| 2528 | * deadlocks.
|
| 2529 |
|
| 2530 | if (wbc->range_cyclic)
|
| 2531 | mapping->writeback_index = 0;
|
| 2532 |
|
| 2533 | |
| 2534 | * Return the first error we encountered (if there was any) to
|
| 2535 | * the caller.
|
| 2536 |
|
| 2537 | *error = wbc->saved_err;
|
| 2538 | }
|
| 2539 | return folio;
|
| 2540 |
|
| 2541 | done:
|
| 2542 | if (wbc->range_cyclic)
|
| 2543 | mapping->writeback_index = folio_next_index(folio);
|
| 2544 | folio_batch_release(&wbc->fbatch);
|
| 2545 | return NULL;
|
| 2546 | }
|
| 2547 | EXPORT_SYMBOL_GPL(writeback_iter);
|
| 2548 |
|
| 2549 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
|
| 2550 | {
|
| 2551 | int ret;
|
| 2552 | struct bdi_writeback *wb;
|
| 2553 |
|
| 2554 | if (wbc->nr_to_write <= 0)
|
| 2555 | return 0;
|
| 2556 | wb = inode_to_wb_wbc(mapping->host, wbc);
|
| 2557 | wb_bandwidth_estimate_start(wb);
|
| 2558 | while (1) {
|
| 2559 | if (mapping->a_ops->writepages)
|
| 2560 | ret = mapping->a_ops->writepages(mapping, wbc);
|
| 2561 | else
|
| 2562 |
|
| 2563 | ret = 0;
|
| 2564 | if (ret != -ENOMEM || wbc->sync_mode != WB_SYNC_ALL)
|
| 2565 | break;
|
| 2566 |
|
| 2567 | |
| 2568 | * Lacking an allocation context or the locality or writeback
|
| 2569 | * state of any of the inode's pages, throttle based on
|
| 2570 | * writeback activity on the local node. It's as good a
|
| 2571 | * guess as any.
|
| 2572 |
|
| 2573 | reclaim_throttle(NODE_DATA(numa_node_id()),
|
| 2574 | VMSCAN_THROTTLE_WRITEBACK);
|
| 2575 | }
|
| 2576 | |
| 2577 | * Usually few pages are written by now from those we've just submitted
|
| 2578 | * but if there's constant writeback being submitted, this makes sure
|
| 2579 | * writeback bandwidth is updated once in a while.
|
| 2580 |
|
| 2581 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) +
|
| 2582 | BANDWIDTH_INTERVAL))
|
| 2583 | wb_update_bandwidth(wb);
|
| 2584 | return ret;
|
| 2585 | }
|
| 2586 |
|
| 2587 | |
| 2588 | * For address_spaces which do not use buffers nor write back.
|
| 2589 |
|
| 2590 | bool noop_dirty_folio(struct address_space *mapping, struct folio *folio)
|
| 2591 | {
|
| 2592 | if (!folio_test_dirty(folio))
|
| 2593 | return !folio_test_set_dirty(folio);
|
| 2594 | return false;
|
| 2595 | }
|
| 2596 | EXPORT_SYMBOL(noop_dirty_folio);
|
| 2597 |
|
| 2598 | |
| 2599 | * Helper function for set_page_dirty family.
|
| 2600 | *
|
| 2601 | * NOTE: This relies on being atomic wrt interrupts.
|
| 2602 |
|
| 2603 | static void folio_account_dirtied(struct folio *folio,
|
| 2604 | struct address_space *mapping)
|
| 2605 | {
|
| 2606 | struct inode *inode = mapping->host;
|
| 2607 |
|
| 2608 | trace_writeback_dirty_folio(folio, mapping);
|
| 2609 |
|
| 2610 | if (mapping_can_writeback(mapping)) {
|
| 2611 | struct bdi_writeback *wb;
|
| 2612 | long nr = folio_nr_pages(folio);
|
| 2613 |
|
| 2614 | inode_attach_wb(inode, folio);
|
| 2615 | wb = inode_to_wb(inode);
|
| 2616 |
|
| 2617 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr);
|
| 2618 | if (folio_test_dropbehind(folio))
|
| 2619 | wb_stat_mod(wb, WB_DONTCACHE_DIRTY, nr);
|
| 2620 | __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
|
| 2621 | __node_stat_mod_folio(folio, NR_DIRTIED, nr);
|
| 2622 | wb_stat_mod(wb, WB_RECLAIMABLE, nr);
|
| 2623 | wb_stat_mod(wb, WB_DIRTIED, nr);
|
| 2624 | task_io_account_write(nr * PAGE_SIZE);
|
| 2625 | current->nr_dirtied += nr;
|
| 2626 | __this_cpu_add(bdp_ratelimits, nr);
|
| 2627 |
|
| 2628 | mem_cgroup_track_foreign_dirty(folio, wb);
|
| 2629 | }
|
| 2630 | }
|
| 2631 |
|
| 2632 | |
| 2633 | * Helper function for deaccounting dirty page without writeback.
|
| 2634 | *
|
| 2635 |
|
| 2636 | void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb)
|
| 2637 | {
|
| 2638 | long nr = folio_nr_pages(folio);
|
| 2639 |
|
| 2640 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
|
| 2641 | if (folio_test_dropbehind(folio))
|
| 2642 | wb_stat_mod(wb, WB_DONTCACHE_DIRTY, -nr);
|
| 2643 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
|
| 2644 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
|
| 2645 | task_io_account_cancelled_write(nr * PAGE_SIZE);
|
| 2646 | }
|
| 2647 |
|
| 2648 | |
| 2649 | * Mark the folio dirty, and set it dirty in the page cache.
|
| 2650 | *
|
| 2651 | * If warn is true, then emit a warning if the folio is not uptodate and has
|
| 2652 | * not been truncated.
|
| 2653 | *
|
| 2654 | * It is the caller's responsibility to prevent the folio from being truncated
|
| 2655 | * while this function is in progress, although it may have been truncated
|
| 2656 | * before this function is called. Most callers have the folio locked.
|
| 2657 | * A few have the folio blocked from truncation through other means (e.g.
|
| 2658 | * zap_vma() has it mapped and is holding the page table lock).
|
| 2659 | * When called from mark_buffer_dirty(), the filesystem should hold a
|
| 2660 | * reference to the buffer_head that is being marked dirty, which causes
|
| 2661 | * try_to_free_buffers() to fail.
|
| 2662 |
|
| 2663 | void __folio_mark_dirty(struct folio *folio, struct address_space *mapping,
|
| 2664 | int warn)
|
| 2665 | {
|
| 2666 | unsigned long flags;
|
| 2667 |
|
| 2668 | |
| 2669 | * Shmem writeback relies on swap, and swap writeback is LRU based,
|
| 2670 | * not using the dirty mark.
|
| 2671 |
|
| 2672 | VM_WARN_ON_ONCE(folio_test_swapcache(folio) || shmem_mapping(mapping));
|
| 2673 |
|
| 2674 | xa_lock_irqsave(&mapping->i_pages, flags);
|
| 2675 | if (folio->mapping) {
|
| 2676 | WARN_ON_ONCE(warn && !folio_test_uptodate(folio));
|
| 2677 | folio_account_dirtied(folio, mapping);
|
| 2678 | __xa_set_mark(&mapping->i_pages, folio->index,
|
| 2679 | PAGECACHE_TAG_DIRTY);
|
| 2680 | }
|
| 2681 | xa_unlock_irqrestore(&mapping->i_pages, flags);
|
| 2682 | }
|
| 2683 |
|
| 2684 | |
| 2685 | * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads.
|
| 2686 | * @mapping: Address space this folio belongs to.
|
| 2687 | * @folio: Folio to be marked as dirty.
|
| 2688 | *
|
| 2689 | * Filesystems which do not use buffer heads should call this function
|
| 2690 | * from their dirty_folio address space operation. It ignores the
|
| 2691 | * contents of folio_get_private(), so if the filesystem marks individual
|
| 2692 | * blocks as dirty, the filesystem should handle that itself.
|
| 2693 | *
|
| 2694 | * This is also sometimes used by filesystems which use buffer_heads when
|
| 2695 | * a single buffer is being dirtied: we want to set the folio dirty in
|
| 2696 | * that case, but not all the buffers. This is a "bottom-up" dirtying,
|
| 2697 | * whereas block_dirty_folio() is a "top-down" dirtying.
|
| 2698 | *
|
| 2699 | * The caller must ensure this doesn't race with truncation. Most will
|
| 2700 | * simply hold the folio lock, but e.g. zap_pte_range() calls with the
|
| 2701 | * folio mapped and the pte lock held, which also locks out truncation.
|
| 2702 |
|
| 2703 | bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio)
|
| 2704 | {
|
| 2705 | if (folio_test_set_dirty(folio))
|
| 2706 | return false;
|
| 2707 |
|
| 2708 | __folio_mark_dirty(folio, mapping, !folio_test_private(folio));
|
| 2709 |
|
| 2710 | if (mapping->host) {
|
| 2711 |
|
| 2712 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
|
| 2713 | }
|
| 2714 | return true;
|
| 2715 | }
|
| 2716 | EXPORT_SYMBOL(filemap_dirty_folio);
|
| 2717 |
|
| 2718 | |
| 2719 | * folio_redirty_for_writepage - Decline to write a dirty folio.
|
| 2720 | * @wbc: The writeback control.
|
| 2721 | * @folio: The folio.
|
| 2722 | *
|
| 2723 | * When a writepage implementation decides that it doesn't want to write
|
| 2724 | * @folio for some reason, it should call this function, unlock @folio and
|
| 2725 | * return 0.
|
| 2726 | *
|
| 2727 | * Return: True if we redirtied the folio. False if someone else dirtied
|
| 2728 | * it first.
|
| 2729 |
|
| 2730 | bool folio_redirty_for_writepage(struct writeback_control *wbc,
|
| 2731 | struct folio *folio)
|
| 2732 | {
|
| 2733 | struct address_space *mapping = folio->mapping;
|
| 2734 | long nr = folio_nr_pages(folio);
|
| 2735 | bool ret;
|
| 2736 |
|
| 2737 | wbc->pages_skipped += nr;
|
| 2738 | ret = filemap_dirty_folio(mapping, folio);
|
| 2739 | if (mapping && mapping_can_writeback(mapping)) {
|
| 2740 | struct inode *inode = mapping->host;
|
| 2741 | struct bdi_writeback *wb;
|
| 2742 | struct wb_lock_cookie cookie = {};
|
| 2743 |
|
| 2744 | wb = unlocked_inode_to_wb_begin(inode, &cookie);
|
| 2745 | current->nr_dirtied -= nr;
|
| 2746 | node_stat_mod_folio(folio, NR_DIRTIED, -nr);
|
| 2747 | wb_stat_mod(wb, WB_DIRTIED, -nr);
|
| 2748 | unlocked_inode_to_wb_end(inode, &cookie);
|
| 2749 | }
|
| 2750 | return ret;
|
| 2751 | }
|
| 2752 | EXPORT_SYMBOL(folio_redirty_for_writepage);
|
| 2753 |
|
| 2754 | |
| 2755 | * folio_mark_dirty - Mark a folio as being modified.
|
| 2756 | * @folio: The folio.
|
| 2757 | *
|
| 2758 | * The folio may not be truncated while this function is running.
|
| 2759 | * Holding the folio lock is sufficient to prevent truncation, but some
|
| 2760 | * callers cannot acquire a sleeping lock. These callers instead hold
|
| 2761 | * the page table lock for a page table which contains at least one page
|
| 2762 | * in this folio. Truncation will block on the page table lock as it
|
| 2763 | * unmaps pages before removing the folio from its mapping.
|
| 2764 | *
|
| 2765 | * .. DANGER::
|
| 2766 | * Do not use this on a folio obtained from a function like
|
| 2767 | * get_user_pages_fast() without holding appropriate locks; you might want to
|
| 2768 | * use set_page_dirty_lock() or folio_mark_dirty_lock() instead.
|
| 2769 | *
|
| 2770 | * Return: True if the folio was newly dirtied, false if it was already dirty.
|
| 2771 |
|
| 2772 | bool folio_mark_dirty(struct folio *folio)
|
| 2773 | {
|
| 2774 | struct address_space *mapping = folio_mapping(folio);
|
| 2775 |
|
| 2776 | if (likely(mapping)) {
|
| 2777 | |
| 2778 | * readahead/folio_deactivate could remain
|
| 2779 | * PG_readahead/PG_reclaim due to race with folio_end_writeback
|
| 2780 | * About readahead, if the folio is written, the flags would be
|
| 2781 | * reset. So no problem.
|
| 2782 | * About folio_deactivate, if the folio is redirtied,
|
| 2783 | * the flag will be reset. So no problem. but if the
|
| 2784 | * folio is used by readahead it will confuse readahead
|
| 2785 | * and make it restart the size rampup process. But it's
|
| 2786 | * a trivial problem.
|
| 2787 |
|
| 2788 | if (folio_test_reclaim(folio))
|
| 2789 | folio_clear_reclaim(folio);
|
| 2790 | return mapping->a_ops->dirty_folio(mapping, folio);
|
| 2791 | }
|
| 2792 |
|
| 2793 | return noop_dirty_folio(mapping, folio);
|
| 2794 | }
|
| 2795 | EXPORT_SYMBOL(folio_mark_dirty);
|
| 2796 |
|
| 2797 | |
| 2798 | * folio_mark_dirty() is racy if the caller has no reference against
|
| 2799 | * folio->mapping->host, and if the folio is unlocked. This is because another
|
| 2800 | * CPU could truncate the folio off the mapping and then free the mapping.
|
| 2801 | *
|
| 2802 | * Usually, the folio _is_ locked, or the caller is a user-space process which
|
| 2803 | * holds a reference on the inode by having an open file.
|
| 2804 | *
|
| 2805 | * In other cases, the folio should be locked before running folio_mark_dirty().
|
| 2806 |
|
| 2807 | bool folio_mark_dirty_lock(struct folio *folio)
|
| 2808 | {
|
| 2809 | bool ret;
|
| 2810 |
|
| 2811 | folio_lock(folio);
|
| 2812 | ret = folio_mark_dirty(folio);
|
| 2813 | folio_unlock(folio);
|
| 2814 | return ret;
|
| 2815 | }
|
| 2816 | EXPORT_SYMBOL(folio_mark_dirty_lock);
|
| 2817 |
|
| 2818 | |
| 2819 | * This cancels just the dirty bit on the kernel page itself, it does NOT
|
| 2820 | * actually remove dirty bits on any mmap's that may be around. It also
|
| 2821 | * leaves the page tagged dirty, so any sync activity will still find it on
|
| 2822 | * the dirty lists, and in particular, clear_page_dirty_for_io() will still
|
| 2823 | * look at the dirty bits in the VM.
|
| 2824 | *
|
| 2825 | * Doing this should *normally* only ever be done when a page is truncated,
|
| 2826 | * and is not actually mapped anywhere at all. However, fs/buffer.c does
|
| 2827 | * this when it notices that somebody has cleaned out all the buffers on a
|
| 2828 | * page without actually doing it through the VM. Can you say "ext3 is
|
| 2829 | * horribly ugly"? Thought you could.
|
| 2830 |
|
| 2831 | void __folio_cancel_dirty(struct folio *folio)
|
| 2832 | {
|
| 2833 | struct address_space *mapping = folio_mapping(folio);
|
| 2834 |
|
| 2835 | if (mapping_can_writeback(mapping)) {
|
| 2836 | struct inode *inode = mapping->host;
|
| 2837 | struct bdi_writeback *wb;
|
| 2838 | struct wb_lock_cookie cookie = {};
|
| 2839 |
|
| 2840 | wb = unlocked_inode_to_wb_begin(inode, &cookie);
|
| 2841 |
|
| 2842 | if (folio_test_clear_dirty(folio))
|
| 2843 | folio_account_cleaned(folio, wb);
|
| 2844 |
|
| 2845 | unlocked_inode_to_wb_end(inode, &cookie);
|
| 2846 | } else {
|
| 2847 | folio_clear_dirty(folio);
|
| 2848 | }
|
| 2849 | }
|
| 2850 | EXPORT_SYMBOL(__folio_cancel_dirty);
|
| 2851 |
|
| 2852 | |
| 2853 | * Clear a folio's dirty flag, while caring for dirty memory accounting.
|
| 2854 | * Returns true if the folio was previously dirty.
|
| 2855 | *
|
| 2856 | * This is for preparing to put the folio under writeout. We leave
|
| 2857 | * the folio tagged as dirty in the xarray so that a concurrent
|
| 2858 | * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk.
|
| 2859 | * The ->writepage implementation will run either folio_start_writeback()
|
| 2860 | * or folio_mark_dirty(), at which stage we bring the folio's dirty flag
|
| 2861 | * and xarray dirty tag back into sync.
|
| 2862 | *
|
| 2863 | * This incoherency between the folio's dirty flag and xarray tag is
|
| 2864 | * unfortunate, but it only exists while the folio is locked.
|
| 2865 |
|
| 2866 | bool folio_clear_dirty_for_io(struct folio *folio)
|
| 2867 | {
|
| 2868 | struct address_space *mapping = folio_mapping(folio);
|
| 2869 | bool ret = false;
|
| 2870 |
|
| 2871 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
|
| 2872 |
|
| 2873 | if (mapping && mapping_can_writeback(mapping)) {
|
| 2874 | struct inode *inode = mapping->host;
|
| 2875 | struct bdi_writeback *wb;
|
| 2876 | struct wb_lock_cookie cookie = {};
|
| 2877 |
|
| 2878 | |
| 2879 | * Yes, Virginia, this is indeed insane.
|
| 2880 | *
|
| 2881 | * We use this sequence to make sure that
|
| 2882 | * (a) we account for dirty stats properly
|
| 2883 | * (b) we tell the low-level filesystem to
|
| 2884 | * mark the whole folio dirty if it was
|
| 2885 | * dirty in a pagetable. Only to then
|
| 2886 | * (c) clean the folio again and return 1 to
|
| 2887 | * cause the writeback.
|
| 2888 | *
|
| 2889 | * This way we avoid all nasty races with the
|
| 2890 | * dirty bit in multiple places and clearing
|
| 2891 | * them concurrently from different threads.
|
| 2892 | *
|
| 2893 | * Note! Normally the "folio_mark_dirty(folio)"
|
| 2894 | * has no effect on the actual dirty bit - since
|
| 2895 | * that will already usually be set. But we
|
| 2896 | * need the side effects, and it can help us
|
| 2897 | * avoid races.
|
| 2898 | *
|
| 2899 | * We basically use the folio "master dirty bit"
|
| 2900 | * as a serialization point for all the different
|
| 2901 | * threads doing their things.
|
| 2902 |
|
| 2903 | if (folio_mkclean(folio))
|
| 2904 | folio_mark_dirty(folio);
|
| 2905 | |
| 2906 | * We carefully synchronise fault handlers against
|
| 2907 | * installing a dirty pte and marking the folio dirty
|
| 2908 | * at this point. We do this by having them hold the
|
| 2909 | * page lock while dirtying the folio, and folios are
|
| 2910 | * always locked coming in here, so we get the desired
|
| 2911 | * exclusion.
|
| 2912 |
|
| 2913 | wb = unlocked_inode_to_wb_begin(inode, &cookie);
|
| 2914 | if (folio_test_clear_dirty(folio)) {
|
| 2915 | long nr = folio_nr_pages(folio);
|
| 2916 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr);
|
| 2917 | if (folio_test_dropbehind(folio))
|
| 2918 | wb_stat_mod(wb, WB_DONTCACHE_DIRTY, -nr);
|
| 2919 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
|
| 2920 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr);
|
| 2921 | ret = true;
|
| 2922 | }
|
| 2923 | unlocked_inode_to_wb_end(inode, &cookie);
|
| 2924 | return ret;
|
| 2925 | }
|
| 2926 | return folio_test_clear_dirty(folio);
|
| 2927 | }
|
| 2928 | EXPORT_SYMBOL(folio_clear_dirty_for_io);
|
| 2929 |
|
| 2930 | static void wb_inode_writeback_start(struct bdi_writeback *wb)
|
| 2931 | {
|
| 2932 | atomic_inc(&wb->writeback_inodes);
|
| 2933 | }
|
| 2934 |
|
| 2935 | static void wb_inode_writeback_end(struct bdi_writeback *wb)
|
| 2936 | {
|
| 2937 | unsigned long flags;
|
| 2938 | atomic_dec(&wb->writeback_inodes);
|
| 2939 | |
| 2940 | * Make sure estimate of writeback throughput gets updated after
|
| 2941 | * writeback completed. We delay the update by BANDWIDTH_INTERVAL
|
| 2942 | * (which is the interval other bandwidth updates use for batching) so
|
| 2943 | * that if multiple inodes end writeback at a similar time, they get
|
| 2944 | * batched into one bandwidth update.
|
| 2945 |
|
| 2946 | spin_lock_irqsave(&wb->work_lock, flags);
|
| 2947 | if (test_bit(WB_registered, &wb->state))
|
| 2948 | queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL);
|
| 2949 | spin_unlock_irqrestore(&wb->work_lock, flags);
|
| 2950 | }
|
| 2951 |
|
| 2952 | bool __folio_end_writeback(struct folio *folio)
|
| 2953 | {
|
| 2954 | long nr = folio_nr_pages(folio);
|
| 2955 | struct address_space *mapping = folio_mapping(folio);
|
| 2956 | bool ret;
|
| 2957 |
|
| 2958 | if (mapping && mapping_use_writeback_tags(mapping)) {
|
| 2959 | struct inode *inode = mapping->host;
|
| 2960 | struct bdi_writeback *wb;
|
| 2961 | unsigned long flags;
|
| 2962 |
|
| 2963 | xa_lock_irqsave(&mapping->i_pages, flags);
|
| 2964 | ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
|
| 2965 | __xa_clear_mark(&mapping->i_pages, folio->index,
|
| 2966 | PAGECACHE_TAG_WRITEBACK);
|
| 2967 |
|
| 2968 | wb = inode_to_wb(inode);
|
| 2969 | wb_stat_mod(wb, WB_WRITEBACK, -nr);
|
| 2970 | __wb_writeout_add(wb, nr);
|
| 2971 | if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
|
| 2972 | wb_inode_writeback_end(wb);
|
| 2973 | if (mapping->host)
|
| 2974 | sb_clear_inode_writeback(mapping->host);
|
| 2975 | }
|
| 2976 |
|
| 2977 | xa_unlock_irqrestore(&mapping->i_pages, flags);
|
| 2978 | } else {
|
| 2979 | ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback);
|
| 2980 | }
|
| 2981 |
|
| 2982 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr);
|
| 2983 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr);
|
| 2984 | node_stat_mod_folio(folio, NR_WRITTEN, nr);
|
| 2985 |
|
| 2986 | return ret;
|
| 2987 | }
|
| 2988 |
|
| 2989 | void __folio_start_writeback(struct folio *folio, bool keep_write)
|
| 2990 | {
|
| 2991 | long nr = folio_nr_pages(folio);
|
| 2992 | struct address_space *mapping = folio_mapping(folio);
|
| 2993 | int access_ret;
|
| 2994 |
|
| 2995 | VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
|
| 2996 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
|
| 2997 |
|
| 2998 | if (mapping && mapping_use_writeback_tags(mapping)) {
|
| 2999 | XA_STATE(xas, &mapping->i_pages, folio->index);
|
| 3000 | struct inode *inode = mapping->host;
|
| 3001 | struct bdi_writeback *wb;
|
| 3002 | unsigned long flags;
|
| 3003 | bool on_wblist;
|
| 3004 |
|
| 3005 | xas_lock_irqsave(&xas, flags);
|
| 3006 | xas_load(&xas);
|
| 3007 | folio_test_set_writeback(folio);
|
| 3008 |
|
| 3009 | on_wblist = mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK);
|
| 3010 |
|
| 3011 | xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK);
|
| 3012 | wb = inode_to_wb(inode);
|
| 3013 | wb_stat_mod(wb, WB_WRITEBACK, nr);
|
| 3014 | if (!on_wblist) {
|
| 3015 | wb_inode_writeback_start(wb);
|
| 3016 | |
| 3017 | * We can come through here when swapping anonymous
|
| 3018 | * folios, so we don't necessarily have an inode to
|
| 3019 | * track for sync.
|
| 3020 |
|
| 3021 | if (mapping->host)
|
| 3022 | sb_mark_inode_writeback(mapping->host);
|
| 3023 | }
|
| 3024 |
|
| 3025 | if (!folio_test_dirty(folio))
|
| 3026 | xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY);
|
| 3027 | if (!keep_write)
|
| 3028 | xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE);
|
| 3029 | xas_unlock_irqrestore(&xas, flags);
|
| 3030 | } else {
|
| 3031 | folio_test_set_writeback(folio);
|
| 3032 | }
|
| 3033 |
|
| 3034 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr);
|
| 3035 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr);
|
| 3036 |
|
| 3037 | access_ret = arch_make_folio_accessible(folio);
|
| 3038 | |
| 3039 | * If writeback has been triggered on a page that cannot be made
|
| 3040 | * accessible, it is too late to recover here.
|
| 3041 |
|
| 3042 | VM_BUG_ON_FOLIO(access_ret != 0, folio);
|
| 3043 | }
|
| 3044 | EXPORT_SYMBOL(__folio_start_writeback);
|
| 3045 |
|
| 3046 | |
| 3047 | * folio_wait_writeback - Wait for a folio to finish writeback.
|
| 3048 | * @folio: The folio to wait for.
|
| 3049 | *
|
| 3050 | * If the folio is currently being written back to storage, wait for the
|
| 3051 | * I/O to complete.
|
| 3052 | *
|
| 3053 | * Context: Sleeps. Must be called in process context and with
|
| 3054 | * no spinlocks held. Caller should hold a reference on the folio.
|
| 3055 | * If the folio is not locked, writeback may start again after writeback
|
| 3056 | * has finished.
|
| 3057 |
|
| 3058 | void folio_wait_writeback(struct folio *folio)
|
| 3059 | {
|
| 3060 | while (folio_test_writeback(folio)) {
|
| 3061 | trace_folio_wait_writeback(folio, folio_mapping(folio));
|
| 3062 | folio_wait_bit(folio, PG_writeback);
|
| 3063 | }
|
| 3064 | }
|
| 3065 | EXPORT_SYMBOL_GPL(folio_wait_writeback);
|
| 3066 |
|
| 3067 | |
| 3068 | * folio_wait_writeback_killable - Wait for a folio to finish writeback.
|
| 3069 | * @folio: The folio to wait for.
|
| 3070 | *
|
| 3071 | * If the folio is currently being written back to storage, wait for the
|
| 3072 | * I/O to complete or a fatal signal to arrive.
|
| 3073 | *
|
| 3074 | * Context: Sleeps. Must be called in process context and with
|
| 3075 | * no spinlocks held. Caller should hold a reference on the folio.
|
| 3076 | * If the folio is not locked, writeback may start again after writeback
|
| 3077 | * has finished.
|
| 3078 | * Return: 0 on success, -EINTR if we get a fatal signal while waiting.
|
| 3079 |
|
| 3080 | int folio_wait_writeback_killable(struct folio *folio)
|
| 3081 | {
|
| 3082 | while (folio_test_writeback(folio)) {
|
| 3083 | trace_folio_wait_writeback(folio, folio_mapping(folio));
|
| 3084 | if (folio_wait_bit_killable(folio, PG_writeback))
|
| 3085 | return -EINTR;
|
| 3086 | }
|
| 3087 |
|
| 3088 | return 0;
|
| 3089 | }
|
| 3090 | EXPORT_SYMBOL_GPL(folio_wait_writeback_killable);
|
| 3091 |
|
| 3092 | |
| 3093 | * folio_wait_stable() - wait for writeback to finish, if necessary.
|
| 3094 | * @folio: The folio to wait on.
|
| 3095 | *
|
| 3096 | * This function determines if the given folio is related to a backing
|
| 3097 | * device that requires folio contents to be held stable during writeback.
|
| 3098 | * If so, then it will wait for any pending writeback to complete.
|
| 3099 | *
|
| 3100 | * Context: Sleeps. Must be called in process context and with
|
| 3101 | * no spinlocks held. Caller should hold a reference on the folio.
|
| 3102 | * If the folio is not locked, writeback may start again after writeback
|
| 3103 | * has finished.
|
| 3104 |
|
| 3105 | void folio_wait_stable(struct folio *folio)
|
| 3106 | {
|
| 3107 | if (mapping_stable_writes(folio_mapping(folio)))
|
| 3108 | folio_wait_writeback(folio);
|
| 3109 | }
|
| 3110 | EXPORT_SYMBOL_GPL(folio_wait_stable);
|