page-writeback.c revision 1df647197c5b8aacaeb58592cba9a1df322c9000
1/* 2 * mm/page-writeback.c 3 * 4 * Copyright (C) 2002, Linus Torvalds. 5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> 6 * 7 * Contains functions related to writing back dirty pages at the 8 * address_space level. 9 * 10 * 10Apr2002 Andrew Morton 11 * Initial version 12 */ 13 14#include <linux/kernel.h> 15#include <linux/export.h> 16#include <linux/spinlock.h> 17#include <linux/fs.h> 18#include <linux/mm.h> 19#include <linux/swap.h> 20#include <linux/slab.h> 21#include <linux/pagemap.h> 22#include <linux/writeback.h> 23#include <linux/init.h> 24#include <linux/backing-dev.h> 25#include <linux/task_io_accounting_ops.h> 26#include <linux/blkdev.h> 27#include <linux/mpage.h> 28#include <linux/rmap.h> 29#include <linux/percpu.h> 30#include <linux/notifier.h> 31#include <linux/smp.h> 32#include <linux/sysctl.h> 33#include <linux/cpu.h> 34#include <linux/syscalls.h> 35#include <linux/buffer_head.h> 36#include <linux/pagevec.h> 37#include <trace/events/writeback.h> 38 39/* 40 * Sleep at most 200ms at a time in balance_dirty_pages(). 41 */ 42#define MAX_PAUSE max(HZ/5, 1) 43 44/* 45 * Estimate write bandwidth at 200ms intervals. 46 */ 47#define BANDWIDTH_INTERVAL max(HZ/5, 1) 48 49#define RATELIMIT_CALC_SHIFT 10 50 51/* 52 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited 53 * will look to see if it needs to force writeback or throttling. 54 */ 55static long ratelimit_pages = 32; 56 57/* The following parameters are exported via /proc/sys/vm */ 58 59/* 60 * Start background writeback (via writeback threads) at this percentage 61 */ 62int dirty_background_ratio = 10; 63 64/* 65 * dirty_background_bytes starts at 0 (disabled) so that it is a function of 66 * dirty_background_ratio * the amount of dirtyable memory 67 */ 68unsigned long dirty_background_bytes; 69 70/* 71 * free highmem will not be subtracted from the total free memory 72 * for calculating free ratios if vm_highmem_is_dirtyable is true 73 */ 74int vm_highmem_is_dirtyable; 75 76/* 77 * The generator of dirty data starts writeback at this percentage 78 */ 79int vm_dirty_ratio = 20; 80 81/* 82 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of 83 * vm_dirty_ratio * the amount of dirtyable memory 84 */ 85unsigned long vm_dirty_bytes; 86 87/* 88 * The interval between `kupdate'-style writebacks 89 */ 90unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */ 91 92/* 93 * The longest time for which data is allowed to remain dirty 94 */ 95unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */ 96 97/* 98 * Flag that makes the machine dump writes/reads and block dirtyings. 99 */ 100int block_dump; 101 102/* 103 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: 104 * a full sync is triggered after this time elapses without any disk activity. 105 */ 106int laptop_mode; 107 108EXPORT_SYMBOL(laptop_mode); 109 110/* End of sysctl-exported parameters */ 111 112unsigned long global_dirty_limit; 113 114/* 115 * Scale the writeback cache size proportional to the relative writeout speeds. 116 * 117 * We do this by keeping a floating proportion between BDIs, based on page 118 * writeback completions [end_page_writeback()]. Those devices that write out 119 * pages fastest will get the larger share, while the slower will get a smaller 120 * share. 121 * 122 * We use page writeout completions because we are interested in getting rid of 123 * dirty pages. Having them written out is the primary goal. 124 * 125 * We introduce a concept of time, a period over which we measure these events, 126 * because demand can/will vary over time. The length of this period itself is 127 * measured in page writeback completions. 128 * 129 */ 130static struct prop_descriptor vm_completions; 131static struct prop_descriptor vm_dirties; 132 133/* 134 * couple the period to the dirty_ratio: 135 * 136 * period/2 ~ roundup_pow_of_two(dirty limit) 137 */ 138static int calc_period_shift(void) 139{ 140 unsigned long dirty_total; 141 142 if (vm_dirty_bytes) 143 dirty_total = vm_dirty_bytes / PAGE_SIZE; 144 else 145 dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / 146 100; 147 return 2 + ilog2(dirty_total - 1); 148} 149 150/* 151 * update the period when the dirty threshold changes. 152 */ 153static void update_completion_period(void) 154{ 155 int shift = calc_period_shift(); 156 prop_change_shift(&vm_completions, shift); 157 prop_change_shift(&vm_dirties, shift); 158 159 writeback_set_ratelimit(); 160} 161 162int dirty_background_ratio_handler(struct ctl_table *table, int write, 163 void __user *buffer, size_t *lenp, 164 loff_t *ppos) 165{ 166 int ret; 167 168 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 169 if (ret == 0 && write) 170 dirty_background_bytes = 0; 171 return ret; 172} 173 174int dirty_background_bytes_handler(struct ctl_table *table, int write, 175 void __user *buffer, size_t *lenp, 176 loff_t *ppos) 177{ 178 int ret; 179 180 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); 181 if (ret == 0 && write) 182 dirty_background_ratio = 0; 183 return ret; 184} 185 186int dirty_ratio_handler(struct ctl_table *table, int write, 187 void __user *buffer, size_t *lenp, 188 loff_t *ppos) 189{ 190 int old_ratio = vm_dirty_ratio; 191 int ret; 192 193 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 194 if (ret == 0 && write && vm_dirty_ratio != old_ratio) { 195 update_completion_period(); 196 vm_dirty_bytes = 0; 197 } 198 return ret; 199} 200 201 202int dirty_bytes_handler(struct ctl_table *table, int write, 203 void __user *buffer, size_t *lenp, 204 loff_t *ppos) 205{ 206 unsigned long old_bytes = vm_dirty_bytes; 207 int ret; 208 209 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); 210 if (ret == 0 && write && vm_dirty_bytes != old_bytes) { 211 update_completion_period(); 212 vm_dirty_ratio = 0; 213 } 214 return ret; 215} 216 217/* 218 * Increment the BDI's writeout completion count and the global writeout 219 * completion count. Called from test_clear_page_writeback(). 220 */ 221static inline void __bdi_writeout_inc(struct backing_dev_info *bdi) 222{ 223 __inc_bdi_stat(bdi, BDI_WRITTEN); 224 __prop_inc_percpu_max(&vm_completions, &bdi->completions, 225 bdi->max_prop_frac); 226} 227 228void bdi_writeout_inc(struct backing_dev_info *bdi) 229{ 230 unsigned long flags; 231 232 local_irq_save(flags); 233 __bdi_writeout_inc(bdi); 234 local_irq_restore(flags); 235} 236EXPORT_SYMBOL_GPL(bdi_writeout_inc); 237 238void task_dirty_inc(struct task_struct *tsk) 239{ 240 prop_inc_single(&vm_dirties, &tsk->dirties); 241} 242 243/* 244 * Obtain an accurate fraction of the BDI's portion. 245 */ 246static void bdi_writeout_fraction(struct backing_dev_info *bdi, 247 long *numerator, long *denominator) 248{ 249 prop_fraction_percpu(&vm_completions, &bdi->completions, 250 numerator, denominator); 251} 252 253/* 254 * bdi_min_ratio keeps the sum of the minimum dirty shares of all 255 * registered backing devices, which, for obvious reasons, can not 256 * exceed 100%. 257 */ 258static unsigned int bdi_min_ratio; 259 260int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) 261{ 262 int ret = 0; 263 264 spin_lock_bh(&bdi_lock); 265 if (min_ratio > bdi->max_ratio) { 266 ret = -EINVAL; 267 } else { 268 min_ratio -= bdi->min_ratio; 269 if (bdi_min_ratio + min_ratio < 100) { 270 bdi_min_ratio += min_ratio; 271 bdi->min_ratio += min_ratio; 272 } else { 273 ret = -EINVAL; 274 } 275 } 276 spin_unlock_bh(&bdi_lock); 277 278 return ret; 279} 280 281int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio) 282{ 283 int ret = 0; 284 285 if (max_ratio > 100) 286 return -EINVAL; 287 288 spin_lock_bh(&bdi_lock); 289 if (bdi->min_ratio > max_ratio) { 290 ret = -EINVAL; 291 } else { 292 bdi->max_ratio = max_ratio; 293 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100; 294 } 295 spin_unlock_bh(&bdi_lock); 296 297 return ret; 298} 299EXPORT_SYMBOL(bdi_set_max_ratio); 300 301/* 302 * Work out the current dirty-memory clamping and background writeout 303 * thresholds. 304 * 305 * The main aim here is to lower them aggressively if there is a lot of mapped 306 * memory around. To avoid stressing page reclaim with lots of unreclaimable 307 * pages. It is better to clamp down on writers than to start swapping, and 308 * performing lots of scanning. 309 * 310 * We only allow 1/2 of the currently-unmapped memory to be dirtied. 311 * 312 * We don't permit the clamping level to fall below 5% - that is getting rather 313 * excessive. 314 * 315 * We make sure that the background writeout level is below the adjusted 316 * clamping level. 317 */ 318 319static unsigned long highmem_dirtyable_memory(unsigned long total) 320{ 321#ifdef CONFIG_HIGHMEM 322 int node; 323 unsigned long x = 0; 324 325 for_each_node_state(node, N_HIGH_MEMORY) { 326 struct zone *z = 327 &NODE_DATA(node)->node_zones[ZONE_HIGHMEM]; 328 329 x += zone_page_state(z, NR_FREE_PAGES) + 330 zone_reclaimable_pages(z); 331 } 332 /* 333 * Make sure that the number of highmem pages is never larger 334 * than the number of the total dirtyable memory. This can only 335 * occur in very strange VM situations but we want to make sure 336 * that this does not occur. 337 */ 338 return min(x, total); 339#else 340 return 0; 341#endif 342} 343 344/** 345 * determine_dirtyable_memory - amount of memory that may be used 346 * 347 * Returns the numebr of pages that can currently be freed and used 348 * by the kernel for direct mappings. 349 */ 350unsigned long determine_dirtyable_memory(void) 351{ 352 unsigned long x; 353 354 x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages(); 355 356 if (!vm_highmem_is_dirtyable) 357 x -= highmem_dirtyable_memory(x); 358 359 return x + 1; /* Ensure that we never return 0 */ 360} 361 362static unsigned long dirty_freerun_ceiling(unsigned long thresh, 363 unsigned long bg_thresh) 364{ 365 return (thresh + bg_thresh) / 2; 366} 367 368static unsigned long hard_dirty_limit(unsigned long thresh) 369{ 370 return max(thresh, global_dirty_limit); 371} 372 373/* 374 * global_dirty_limits - background-writeback and dirty-throttling thresholds 375 * 376 * Calculate the dirty thresholds based on sysctl parameters 377 * - vm.dirty_background_ratio or vm.dirty_background_bytes 378 * - vm.dirty_ratio or vm.dirty_bytes 379 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and 380 * real-time tasks. 381 */ 382void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) 383{ 384 unsigned long background; 385 unsigned long dirty; 386 unsigned long uninitialized_var(available_memory); 387 struct task_struct *tsk; 388 389 if (!vm_dirty_bytes || !dirty_background_bytes) 390 available_memory = determine_dirtyable_memory(); 391 392 if (vm_dirty_bytes) 393 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE); 394 else 395 dirty = (vm_dirty_ratio * available_memory) / 100; 396 397 if (dirty_background_bytes) 398 background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE); 399 else 400 background = (dirty_background_ratio * available_memory) / 100; 401 402 if (background >= dirty) 403 background = dirty / 2; 404 tsk = current; 405 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) { 406 background += background / 4; 407 dirty += dirty / 4; 408 } 409 *pbackground = background; 410 *pdirty = dirty; 411 trace_global_dirty_state(background, dirty); 412} 413 414/** 415 * bdi_dirty_limit - @bdi's share of dirty throttling threshold 416 * @bdi: the backing_dev_info to query 417 * @dirty: global dirty limit in pages 418 * 419 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of 420 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages. 421 * And the "limit" in the name is not seriously taken as hard limit in 422 * balance_dirty_pages(). 423 * 424 * It allocates high/low dirty limits to fast/slow devices, in order to prevent 425 * - starving fast devices 426 * - piling up dirty pages (that will take long time to sync) on slow devices 427 * 428 * The bdi's share of dirty limit will be adapting to its throughput and 429 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. 430 */ 431unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty) 432{ 433 u64 bdi_dirty; 434 long numerator, denominator; 435 436 /* 437 * Calculate this BDI's share of the dirty ratio. 438 */ 439 bdi_writeout_fraction(bdi, &numerator, &denominator); 440 441 bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100; 442 bdi_dirty *= numerator; 443 do_div(bdi_dirty, denominator); 444 445 bdi_dirty += (dirty * bdi->min_ratio) / 100; 446 if (bdi_dirty > (dirty * bdi->max_ratio) / 100) 447 bdi_dirty = dirty * bdi->max_ratio / 100; 448 449 return bdi_dirty; 450} 451 452/* 453 * Dirty position control. 454 * 455 * (o) global/bdi setpoints 456 * 457 * We want the dirty pages be balanced around the global/bdi setpoints. 458 * When the number of dirty pages is higher/lower than the setpoint, the 459 * dirty position control ratio (and hence task dirty ratelimit) will be 460 * decreased/increased to bring the dirty pages back to the setpoint. 461 * 462 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT 463 * 464 * if (dirty < setpoint) scale up pos_ratio 465 * if (dirty > setpoint) scale down pos_ratio 466 * 467 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio 468 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio 469 * 470 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT 471 * 472 * (o) global control line 473 * 474 * ^ pos_ratio 475 * | 476 * | |<===== global dirty control scope ======>| 477 * 2.0 .............* 478 * | .* 479 * | . * 480 * | . * 481 * | . * 482 * | . * 483 * | . * 484 * 1.0 ................................* 485 * | . . * 486 * | . . * 487 * | . . * 488 * | . . * 489 * | . . * 490 * 0 +------------.------------------.----------------------*-------------> 491 * freerun^ setpoint^ limit^ dirty pages 492 * 493 * (o) bdi control line 494 * 495 * ^ pos_ratio 496 * | 497 * | * 498 * | * 499 * | * 500 * | * 501 * | * |<=========== span ============>| 502 * 1.0 .......................* 503 * | . * 504 * | . * 505 * | . * 506 * | . * 507 * | . * 508 * | . * 509 * | . * 510 * | . * 511 * | . * 512 * | . * 513 * | . * 514 * 1/4 ...............................................* * * * * * * * * * * * 515 * | . . 516 * | . . 517 * | . . 518 * 0 +----------------------.-------------------------------.-------------> 519 * bdi_setpoint^ x_intercept^ 520 * 521 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can 522 * be smoothly throttled down to normal if it starts high in situations like 523 * - start writing to a slow SD card and a fast disk at the same time. The SD 524 * card's bdi_dirty may rush to many times higher than bdi_setpoint. 525 * - the bdi dirty thresh drops quickly due to change of JBOD workload 526 */ 527static unsigned long bdi_position_ratio(struct backing_dev_info *bdi, 528 unsigned long thresh, 529 unsigned long bg_thresh, 530 unsigned long dirty, 531 unsigned long bdi_thresh, 532 unsigned long bdi_dirty) 533{ 534 unsigned long write_bw = bdi->avg_write_bandwidth; 535 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh); 536 unsigned long limit = hard_dirty_limit(thresh); 537 unsigned long x_intercept; 538 unsigned long setpoint; /* dirty pages' target balance point */ 539 unsigned long bdi_setpoint; 540 unsigned long span; 541 long long pos_ratio; /* for scaling up/down the rate limit */ 542 long x; 543 544 if (unlikely(dirty >= limit)) 545 return 0; 546 547 /* 548 * global setpoint 549 * 550 * setpoint - dirty 3 551 * f(dirty) := 1.0 + (----------------) 552 * limit - setpoint 553 * 554 * it's a 3rd order polynomial that subjects to 555 * 556 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast 557 * (2) f(setpoint) = 1.0 => the balance point 558 * (3) f(limit) = 0 => the hard limit 559 * (4) df/dx <= 0 => negative feedback control 560 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) 561 * => fast response on large errors; small oscillation near setpoint 562 */ 563 setpoint = (freerun + limit) / 2; 564 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT, 565 limit - setpoint + 1); 566 pos_ratio = x; 567 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; 568 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; 569 pos_ratio += 1 << RATELIMIT_CALC_SHIFT; 570 571 /* 572 * We have computed basic pos_ratio above based on global situation. If 573 * the bdi is over/under its share of dirty pages, we want to scale 574 * pos_ratio further down/up. That is done by the following mechanism. 575 */ 576 577 /* 578 * bdi setpoint 579 * 580 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint) 581 * 582 * x_intercept - bdi_dirty 583 * := -------------------------- 584 * x_intercept - bdi_setpoint 585 * 586 * The main bdi control line is a linear function that subjects to 587 * 588 * (1) f(bdi_setpoint) = 1.0 589 * (2) k = - 1 / (8 * write_bw) (in single bdi case) 590 * or equally: x_intercept = bdi_setpoint + 8 * write_bw 591 * 592 * For single bdi case, the dirty pages are observed to fluctuate 593 * regularly within range 594 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2] 595 * for various filesystems, where (2) can yield in a reasonable 12.5% 596 * fluctuation range for pos_ratio. 597 * 598 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its 599 * own size, so move the slope over accordingly and choose a slope that 600 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh. 601 */ 602 if (unlikely(bdi_thresh > thresh)) 603 bdi_thresh = thresh; 604 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8); 605 /* 606 * scale global setpoint to bdi's: 607 * bdi_setpoint = setpoint * bdi_thresh / thresh 608 */ 609 x = div_u64((u64)bdi_thresh << 16, thresh + 1); 610 bdi_setpoint = setpoint * (u64)x >> 16; 611 /* 612 * Use span=(8*write_bw) in single bdi case as indicated by 613 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case. 614 * 615 * bdi_thresh thresh - bdi_thresh 616 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh 617 * thresh thresh 618 */ 619 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16; 620 x_intercept = bdi_setpoint + span; 621 622 if (bdi_dirty < x_intercept - span / 4) { 623 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty), 624 x_intercept - bdi_setpoint + 1); 625 } else 626 pos_ratio /= 4; 627 628 /* 629 * bdi reserve area, safeguard against dirty pool underrun and disk idle 630 * It may push the desired control point of global dirty pages higher 631 * than setpoint. 632 */ 633 x_intercept = bdi_thresh / 2; 634 if (bdi_dirty < x_intercept) { 635 if (bdi_dirty > x_intercept / 8) 636 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty); 637 else 638 pos_ratio *= 8; 639 } 640 641 return pos_ratio; 642} 643 644static void bdi_update_write_bandwidth(struct backing_dev_info *bdi, 645 unsigned long elapsed, 646 unsigned long written) 647{ 648 const unsigned long period = roundup_pow_of_two(3 * HZ); 649 unsigned long avg = bdi->avg_write_bandwidth; 650 unsigned long old = bdi->write_bandwidth; 651 u64 bw; 652 653 /* 654 * bw = written * HZ / elapsed 655 * 656 * bw * elapsed + write_bandwidth * (period - elapsed) 657 * write_bandwidth = --------------------------------------------------- 658 * period 659 */ 660 bw = written - bdi->written_stamp; 661 bw *= HZ; 662 if (unlikely(elapsed > period)) { 663 do_div(bw, elapsed); 664 avg = bw; 665 goto out; 666 } 667 bw += (u64)bdi->write_bandwidth * (period - elapsed); 668 bw >>= ilog2(period); 669 670 /* 671 * one more level of smoothing, for filtering out sudden spikes 672 */ 673 if (avg > old && old >= (unsigned long)bw) 674 avg -= (avg - old) >> 3; 675 676 if (avg < old && old <= (unsigned long)bw) 677 avg += (old - avg) >> 3; 678 679out: 680 bdi->write_bandwidth = bw; 681 bdi->avg_write_bandwidth = avg; 682} 683 684/* 685 * The global dirtyable memory and dirty threshold could be suddenly knocked 686 * down by a large amount (eg. on the startup of KVM in a swapless system). 687 * This may throw the system into deep dirty exceeded state and throttle 688 * heavy/light dirtiers alike. To retain good responsiveness, maintain 689 * global_dirty_limit for tracking slowly down to the knocked down dirty 690 * threshold. 691 */ 692static void update_dirty_limit(unsigned long thresh, unsigned long dirty) 693{ 694 unsigned long limit = global_dirty_limit; 695 696 /* 697 * Follow up in one step. 698 */ 699 if (limit < thresh) { 700 limit = thresh; 701 goto update; 702 } 703 704 /* 705 * Follow down slowly. Use the higher one as the target, because thresh 706 * may drop below dirty. This is exactly the reason to introduce 707 * global_dirty_limit which is guaranteed to lie above the dirty pages. 708 */ 709 thresh = max(thresh, dirty); 710 if (limit > thresh) { 711 limit -= (limit - thresh) >> 5; 712 goto update; 713 } 714 return; 715update: 716 global_dirty_limit = limit; 717} 718 719static void global_update_bandwidth(unsigned long thresh, 720 unsigned long dirty, 721 unsigned long now) 722{ 723 static DEFINE_SPINLOCK(dirty_lock); 724 static unsigned long update_time; 725 726 /* 727 * check locklessly first to optimize away locking for the most time 728 */ 729 if (time_before(now, update_time + BANDWIDTH_INTERVAL)) 730 return; 731 732 spin_lock(&dirty_lock); 733 if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) { 734 update_dirty_limit(thresh, dirty); 735 update_time = now; 736 } 737 spin_unlock(&dirty_lock); 738} 739 740/* 741 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate. 742 * 743 * Normal bdi tasks will be curbed at or below it in long term. 744 * Obviously it should be around (write_bw / N) when there are N dd tasks. 745 */ 746static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi, 747 unsigned long thresh, 748 unsigned long bg_thresh, 749 unsigned long dirty, 750 unsigned long bdi_thresh, 751 unsigned long bdi_dirty, 752 unsigned long dirtied, 753 unsigned long elapsed) 754{ 755 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh); 756 unsigned long limit = hard_dirty_limit(thresh); 757 unsigned long setpoint = (freerun + limit) / 2; 758 unsigned long write_bw = bdi->avg_write_bandwidth; 759 unsigned long dirty_ratelimit = bdi->dirty_ratelimit; 760 unsigned long dirty_rate; 761 unsigned long task_ratelimit; 762 unsigned long balanced_dirty_ratelimit; 763 unsigned long pos_ratio; 764 unsigned long step; 765 unsigned long x; 766 767 /* 768 * The dirty rate will match the writeout rate in long term, except 769 * when dirty pages are truncated by userspace or re-dirtied by FS. 770 */ 771 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed; 772 773 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty, 774 bdi_thresh, bdi_dirty); 775 /* 776 * task_ratelimit reflects each dd's dirty rate for the past 200ms. 777 */ 778 task_ratelimit = (u64)dirty_ratelimit * 779 pos_ratio >> RATELIMIT_CALC_SHIFT; 780 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ 781 782 /* 783 * A linear estimation of the "balanced" throttle rate. The theory is, 784 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's 785 * dirty_rate will be measured to be (N * task_ratelimit). So the below 786 * formula will yield the balanced rate limit (write_bw / N). 787 * 788 * Note that the expanded form is not a pure rate feedback: 789 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) 790 * but also takes pos_ratio into account: 791 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) 792 * 793 * (1) is not realistic because pos_ratio also takes part in balancing 794 * the dirty rate. Consider the state 795 * pos_ratio = 0.5 (3) 796 * rate = 2 * (write_bw / N) (4) 797 * If (1) is used, it will stuck in that state! Because each dd will 798 * be throttled at 799 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) 800 * yielding 801 * dirty_rate = N * task_ratelimit = write_bw (6) 802 * put (6) into (1) we get 803 * rate_(i+1) = rate_(i) (7) 804 * 805 * So we end up using (2) to always keep 806 * rate_(i+1) ~= (write_bw / N) (8) 807 * regardless of the value of pos_ratio. As long as (8) is satisfied, 808 * pos_ratio is able to drive itself to 1.0, which is not only where 809 * the dirty count meet the setpoint, but also where the slope of 810 * pos_ratio is most flat and hence task_ratelimit is least fluctuated. 811 */ 812 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, 813 dirty_rate | 1); 814 815 /* 816 * We could safely do this and return immediately: 817 * 818 * bdi->dirty_ratelimit = balanced_dirty_ratelimit; 819 * 820 * However to get a more stable dirty_ratelimit, the below elaborated 821 * code makes use of task_ratelimit to filter out sigular points and 822 * limit the step size. 823 * 824 * The below code essentially only uses the relative value of 825 * 826 * task_ratelimit - dirty_ratelimit 827 * = (pos_ratio - 1) * dirty_ratelimit 828 * 829 * which reflects the direction and size of dirty position error. 830 */ 831 832 /* 833 * dirty_ratelimit will follow balanced_dirty_ratelimit iff 834 * task_ratelimit is on the same side of dirty_ratelimit, too. 835 * For example, when 836 * - dirty_ratelimit > balanced_dirty_ratelimit 837 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) 838 * lowering dirty_ratelimit will help meet both the position and rate 839 * control targets. Otherwise, don't update dirty_ratelimit if it will 840 * only help meet the rate target. After all, what the users ultimately 841 * feel and care are stable dirty rate and small position error. 842 * 843 * |task_ratelimit - dirty_ratelimit| is used to limit the step size 844 * and filter out the sigular points of balanced_dirty_ratelimit. Which 845 * keeps jumping around randomly and can even leap far away at times 846 * due to the small 200ms estimation period of dirty_rate (we want to 847 * keep that period small to reduce time lags). 848 */ 849 step = 0; 850 if (dirty < setpoint) { 851 x = min(bdi->balanced_dirty_ratelimit, 852 min(balanced_dirty_ratelimit, task_ratelimit)); 853 if (dirty_ratelimit < x) 854 step = x - dirty_ratelimit; 855 } else { 856 x = max(bdi->balanced_dirty_ratelimit, 857 max(balanced_dirty_ratelimit, task_ratelimit)); 858 if (dirty_ratelimit > x) 859 step = dirty_ratelimit - x; 860 } 861 862 /* 863 * Don't pursue 100% rate matching. It's impossible since the balanced 864 * rate itself is constantly fluctuating. So decrease the track speed 865 * when it gets close to the target. Helps eliminate pointless tremors. 866 */ 867 step >>= dirty_ratelimit / (2 * step + 1); 868 /* 869 * Limit the tracking speed to avoid overshooting. 870 */ 871 step = (step + 7) / 8; 872 873 if (dirty_ratelimit < balanced_dirty_ratelimit) 874 dirty_ratelimit += step; 875 else 876 dirty_ratelimit -= step; 877 878 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL); 879 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit; 880 881 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit); 882} 883 884void __bdi_update_bandwidth(struct backing_dev_info *bdi, 885 unsigned long thresh, 886 unsigned long bg_thresh, 887 unsigned long dirty, 888 unsigned long bdi_thresh, 889 unsigned long bdi_dirty, 890 unsigned long start_time) 891{ 892 unsigned long now = jiffies; 893 unsigned long elapsed = now - bdi->bw_time_stamp; 894 unsigned long dirtied; 895 unsigned long written; 896 897 /* 898 * rate-limit, only update once every 200ms. 899 */ 900 if (elapsed < BANDWIDTH_INTERVAL) 901 return; 902 903 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]); 904 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]); 905 906 /* 907 * Skip quiet periods when disk bandwidth is under-utilized. 908 * (at least 1s idle time between two flusher runs) 909 */ 910 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time)) 911 goto snapshot; 912 913 if (thresh) { 914 global_update_bandwidth(thresh, dirty, now); 915 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty, 916 bdi_thresh, bdi_dirty, 917 dirtied, elapsed); 918 } 919 bdi_update_write_bandwidth(bdi, elapsed, written); 920 921snapshot: 922 bdi->dirtied_stamp = dirtied; 923 bdi->written_stamp = written; 924 bdi->bw_time_stamp = now; 925} 926 927static void bdi_update_bandwidth(struct backing_dev_info *bdi, 928 unsigned long thresh, 929 unsigned long bg_thresh, 930 unsigned long dirty, 931 unsigned long bdi_thresh, 932 unsigned long bdi_dirty, 933 unsigned long start_time) 934{ 935 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL)) 936 return; 937 spin_lock(&bdi->wb.list_lock); 938 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty, 939 bdi_thresh, bdi_dirty, start_time); 940 spin_unlock(&bdi->wb.list_lock); 941} 942 943/* 944 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr() 945 * will look to see if it needs to start dirty throttling. 946 * 947 * If dirty_poll_interval is too low, big NUMA machines will call the expensive 948 * global_page_state() too often. So scale it near-sqrt to the safety margin 949 * (the number of pages we may dirty without exceeding the dirty limits). 950 */ 951static unsigned long dirty_poll_interval(unsigned long dirty, 952 unsigned long thresh) 953{ 954 if (thresh > dirty) 955 return 1UL << (ilog2(thresh - dirty) >> 1); 956 957 return 1; 958} 959 960static unsigned long bdi_max_pause(struct backing_dev_info *bdi, 961 unsigned long bdi_dirty) 962{ 963 unsigned long bw = bdi->avg_write_bandwidth; 964 unsigned long hi = ilog2(bw); 965 unsigned long lo = ilog2(bdi->dirty_ratelimit); 966 unsigned long t; 967 968 /* target for 20ms max pause on 1-dd case */ 969 t = HZ / 50; 970 971 /* 972 * Scale up pause time for concurrent dirtiers in order to reduce CPU 973 * overheads. 974 * 975 * (N * 20ms) on 2^N concurrent tasks. 976 */ 977 if (hi > lo) 978 t += (hi - lo) * (20 * HZ) / 1024; 979 980 /* 981 * Limit pause time for small memory systems. If sleeping for too long 982 * time, a small pool of dirty/writeback pages may go empty and disk go 983 * idle. 984 * 985 * 8 serves as the safety ratio. 986 */ 987 if (bdi_dirty) 988 t = min(t, bdi_dirty * HZ / (8 * bw + 1)); 989 990 /* 991 * The pause time will be settled within range (max_pause/4, max_pause). 992 * Apply a minimal value of 4 to get a non-zero max_pause/4. 993 */ 994 return clamp_val(t, 4, MAX_PAUSE); 995} 996 997/* 998 * balance_dirty_pages() must be called by processes which are generating dirty 999 * data. It looks at the number of dirty pages in the machine and will force 1000 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. 1001 * If we're over `background_thresh' then the writeback threads are woken to 1002 * perform some writeout. 1003 */ 1004static void balance_dirty_pages(struct address_space *mapping, 1005 unsigned long pages_dirtied) 1006{ 1007 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */ 1008 unsigned long bdi_reclaimable; 1009 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */ 1010 unsigned long bdi_dirty; 1011 unsigned long freerun; 1012 unsigned long background_thresh; 1013 unsigned long dirty_thresh; 1014 unsigned long bdi_thresh; 1015 long pause = 0; 1016 long uninitialized_var(max_pause); 1017 bool dirty_exceeded = false; 1018 unsigned long task_ratelimit; 1019 unsigned long uninitialized_var(dirty_ratelimit); 1020 unsigned long pos_ratio; 1021 struct backing_dev_info *bdi = mapping->backing_dev_info; 1022 unsigned long start_time = jiffies; 1023 1024 for (;;) { 1025 /* 1026 * Unstable writes are a feature of certain networked 1027 * filesystems (i.e. NFS) in which data may have been 1028 * written to the server's write cache, but has not yet 1029 * been flushed to permanent storage. 1030 */ 1031 nr_reclaimable = global_page_state(NR_FILE_DIRTY) + 1032 global_page_state(NR_UNSTABLE_NFS); 1033 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK); 1034 1035 global_dirty_limits(&background_thresh, &dirty_thresh); 1036 1037 /* 1038 * Throttle it only when the background writeback cannot 1039 * catch-up. This avoids (excessively) small writeouts 1040 * when the bdi limits are ramping up. 1041 */ 1042 freerun = dirty_freerun_ceiling(dirty_thresh, 1043 background_thresh); 1044 if (nr_dirty <= freerun) 1045 break; 1046 1047 if (unlikely(!writeback_in_progress(bdi))) 1048 bdi_start_background_writeback(bdi); 1049 1050 /* 1051 * bdi_thresh is not treated as some limiting factor as 1052 * dirty_thresh, due to reasons 1053 * - in JBOD setup, bdi_thresh can fluctuate a lot 1054 * - in a system with HDD and USB key, the USB key may somehow 1055 * go into state (bdi_dirty >> bdi_thresh) either because 1056 * bdi_dirty starts high, or because bdi_thresh drops low. 1057 * In this case we don't want to hard throttle the USB key 1058 * dirtiers for 100 seconds until bdi_dirty drops under 1059 * bdi_thresh. Instead the auxiliary bdi control line in 1060 * bdi_position_ratio() will let the dirtier task progress 1061 * at some rate <= (write_bw / 2) for bringing down bdi_dirty. 1062 */ 1063 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh); 1064 1065 /* 1066 * In order to avoid the stacked BDI deadlock we need 1067 * to ensure we accurately count the 'dirty' pages when 1068 * the threshold is low. 1069 * 1070 * Otherwise it would be possible to get thresh+n pages 1071 * reported dirty, even though there are thresh-m pages 1072 * actually dirty; with m+n sitting in the percpu 1073 * deltas. 1074 */ 1075 if (bdi_thresh < 2 * bdi_stat_error(bdi)) { 1076 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE); 1077 bdi_dirty = bdi_reclaimable + 1078 bdi_stat_sum(bdi, BDI_WRITEBACK); 1079 } else { 1080 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE); 1081 bdi_dirty = bdi_reclaimable + 1082 bdi_stat(bdi, BDI_WRITEBACK); 1083 } 1084 1085 dirty_exceeded = (bdi_dirty > bdi_thresh) || 1086 (nr_dirty > dirty_thresh); 1087 if (dirty_exceeded && !bdi->dirty_exceeded) 1088 bdi->dirty_exceeded = 1; 1089 1090 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh, 1091 nr_dirty, bdi_thresh, bdi_dirty, 1092 start_time); 1093 1094 max_pause = bdi_max_pause(bdi, bdi_dirty); 1095 1096 dirty_ratelimit = bdi->dirty_ratelimit; 1097 pos_ratio = bdi_position_ratio(bdi, dirty_thresh, 1098 background_thresh, nr_dirty, 1099 bdi_thresh, bdi_dirty); 1100 task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >> 1101 RATELIMIT_CALC_SHIFT; 1102 if (unlikely(task_ratelimit == 0)) { 1103 pause = max_pause; 1104 goto pause; 1105 } 1106 pause = HZ * pages_dirtied / task_ratelimit; 1107 if (unlikely(pause <= 0)) { 1108 trace_balance_dirty_pages(bdi, 1109 dirty_thresh, 1110 background_thresh, 1111 nr_dirty, 1112 bdi_thresh, 1113 bdi_dirty, 1114 dirty_ratelimit, 1115 task_ratelimit, 1116 pages_dirtied, 1117 pause, 1118 start_time); 1119 pause = 1; /* avoid resetting nr_dirtied_pause below */ 1120 break; 1121 } 1122 pause = min(pause, max_pause); 1123 1124pause: 1125 trace_balance_dirty_pages(bdi, 1126 dirty_thresh, 1127 background_thresh, 1128 nr_dirty, 1129 bdi_thresh, 1130 bdi_dirty, 1131 dirty_ratelimit, 1132 task_ratelimit, 1133 pages_dirtied, 1134 pause, 1135 start_time); 1136 __set_current_state(TASK_KILLABLE); 1137 io_schedule_timeout(pause); 1138 1139 /* 1140 * This is typically equal to (nr_dirty < dirty_thresh) and can 1141 * also keep "1000+ dd on a slow USB stick" under control. 1142 */ 1143 if (task_ratelimit) 1144 break; 1145 1146 if (fatal_signal_pending(current)) 1147 break; 1148 } 1149 1150 if (!dirty_exceeded && bdi->dirty_exceeded) 1151 bdi->dirty_exceeded = 0; 1152 1153 current->nr_dirtied = 0; 1154 if (pause == 0) { /* in freerun area */ 1155 current->nr_dirtied_pause = 1156 dirty_poll_interval(nr_dirty, dirty_thresh); 1157 } else if (pause <= max_pause / 4 && 1158 pages_dirtied >= current->nr_dirtied_pause) { 1159 current->nr_dirtied_pause = clamp_val( 1160 dirty_ratelimit * (max_pause / 2) / HZ, 1161 pages_dirtied + pages_dirtied / 8, 1162 pages_dirtied * 4); 1163 } else if (pause >= max_pause) { 1164 current->nr_dirtied_pause = 1 | clamp_val( 1165 dirty_ratelimit * (max_pause / 2) / HZ, 1166 pages_dirtied / 4, 1167 pages_dirtied - pages_dirtied / 8); 1168 } 1169 1170 if (writeback_in_progress(bdi)) 1171 return; 1172 1173 /* 1174 * In laptop mode, we wait until hitting the higher threshold before 1175 * starting background writeout, and then write out all the way down 1176 * to the lower threshold. So slow writers cause minimal disk activity. 1177 * 1178 * In normal mode, we start background writeout at the lower 1179 * background_thresh, to keep the amount of dirty memory low. 1180 */ 1181 if (laptop_mode) 1182 return; 1183 1184 if (nr_reclaimable > background_thresh) 1185 bdi_start_background_writeback(bdi); 1186} 1187 1188void set_page_dirty_balance(struct page *page, int page_mkwrite) 1189{ 1190 if (set_page_dirty(page) || page_mkwrite) { 1191 struct address_space *mapping = page_mapping(page); 1192 1193 if (mapping) 1194 balance_dirty_pages_ratelimited(mapping); 1195 } 1196} 1197 1198static DEFINE_PER_CPU(int, bdp_ratelimits); 1199 1200/** 1201 * balance_dirty_pages_ratelimited_nr - balance dirty memory state 1202 * @mapping: address_space which was dirtied 1203 * @nr_pages_dirtied: number of pages which the caller has just dirtied 1204 * 1205 * Processes which are dirtying memory should call in here once for each page 1206 * which was newly dirtied. The function will periodically check the system's 1207 * dirty state and will initiate writeback if needed. 1208 * 1209 * On really big machines, get_writeback_state is expensive, so try to avoid 1210 * calling it too often (ratelimiting). But once we're over the dirty memory 1211 * limit we decrease the ratelimiting by a lot, to prevent individual processes 1212 * from overshooting the limit by (ratelimit_pages) each. 1213 */ 1214void balance_dirty_pages_ratelimited_nr(struct address_space *mapping, 1215 unsigned long nr_pages_dirtied) 1216{ 1217 struct backing_dev_info *bdi = mapping->backing_dev_info; 1218 int ratelimit; 1219 int *p; 1220 1221 if (!bdi_cap_account_dirty(bdi)) 1222 return; 1223 1224 ratelimit = current->nr_dirtied_pause; 1225 if (bdi->dirty_exceeded) 1226 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); 1227 1228 current->nr_dirtied += nr_pages_dirtied; 1229 1230 preempt_disable(); 1231 /* 1232 * This prevents one CPU to accumulate too many dirtied pages without 1233 * calling into balance_dirty_pages(), which can happen when there are 1234 * 1000+ tasks, all of them start dirtying pages at exactly the same 1235 * time, hence all honoured too large initial task->nr_dirtied_pause. 1236 */ 1237 p = &__get_cpu_var(bdp_ratelimits); 1238 if (unlikely(current->nr_dirtied >= ratelimit)) 1239 *p = 0; 1240 else { 1241 *p += nr_pages_dirtied; 1242 if (unlikely(*p >= ratelimit_pages)) { 1243 *p = 0; 1244 ratelimit = 0; 1245 } 1246 } 1247 preempt_enable(); 1248 1249 if (unlikely(current->nr_dirtied >= ratelimit)) 1250 balance_dirty_pages(mapping, current->nr_dirtied); 1251} 1252EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr); 1253 1254void throttle_vm_writeout(gfp_t gfp_mask) 1255{ 1256 unsigned long background_thresh; 1257 unsigned long dirty_thresh; 1258 1259 for ( ; ; ) { 1260 global_dirty_limits(&background_thresh, &dirty_thresh); 1261 1262 /* 1263 * Boost the allowable dirty threshold a bit for page 1264 * allocators so they don't get DoS'ed by heavy writers 1265 */ 1266 dirty_thresh += dirty_thresh / 10; /* wheeee... */ 1267 1268 if (global_page_state(NR_UNSTABLE_NFS) + 1269 global_page_state(NR_WRITEBACK) <= dirty_thresh) 1270 break; 1271 congestion_wait(BLK_RW_ASYNC, HZ/10); 1272 1273 /* 1274 * The caller might hold locks which can prevent IO completion 1275 * or progress in the filesystem. So we cannot just sit here 1276 * waiting for IO to complete. 1277 */ 1278 if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO)) 1279 break; 1280 } 1281} 1282 1283/* 1284 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs 1285 */ 1286int dirty_writeback_centisecs_handler(ctl_table *table, int write, 1287 void __user *buffer, size_t *length, loff_t *ppos) 1288{ 1289 proc_dointvec(table, write, buffer, length, ppos); 1290 bdi_arm_supers_timer(); 1291 return 0; 1292} 1293 1294#ifdef CONFIG_BLOCK 1295void laptop_mode_timer_fn(unsigned long data) 1296{ 1297 struct request_queue *q = (struct request_queue *)data; 1298 int nr_pages = global_page_state(NR_FILE_DIRTY) + 1299 global_page_state(NR_UNSTABLE_NFS); 1300 1301 /* 1302 * We want to write everything out, not just down to the dirty 1303 * threshold 1304 */ 1305 if (bdi_has_dirty_io(&q->backing_dev_info)) 1306 bdi_start_writeback(&q->backing_dev_info, nr_pages, 1307 WB_REASON_LAPTOP_TIMER); 1308} 1309 1310/* 1311 * We've spun up the disk and we're in laptop mode: schedule writeback 1312 * of all dirty data a few seconds from now. If the flush is already scheduled 1313 * then push it back - the user is still using the disk. 1314 */ 1315void laptop_io_completion(struct backing_dev_info *info) 1316{ 1317 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); 1318} 1319 1320/* 1321 * We're in laptop mode and we've just synced. The sync's writes will have 1322 * caused another writeback to be scheduled by laptop_io_completion. 1323 * Nothing needs to be written back anymore, so we unschedule the writeback. 1324 */ 1325void laptop_sync_completion(void) 1326{ 1327 struct backing_dev_info *bdi; 1328 1329 rcu_read_lock(); 1330 1331 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) 1332 del_timer(&bdi->laptop_mode_wb_timer); 1333 1334 rcu_read_unlock(); 1335} 1336#endif 1337 1338/* 1339 * If ratelimit_pages is too high then we can get into dirty-data overload 1340 * if a large number of processes all perform writes at the same time. 1341 * If it is too low then SMP machines will call the (expensive) 1342 * get_writeback_state too often. 1343 * 1344 * Here we set ratelimit_pages to a level which ensures that when all CPUs are 1345 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory 1346 * thresholds. 1347 */ 1348 1349void writeback_set_ratelimit(void) 1350{ 1351 unsigned long background_thresh; 1352 unsigned long dirty_thresh; 1353 global_dirty_limits(&background_thresh, &dirty_thresh); 1354 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); 1355 if (ratelimit_pages < 16) 1356 ratelimit_pages = 16; 1357} 1358 1359static int __cpuinit 1360ratelimit_handler(struct notifier_block *self, unsigned long u, void *v) 1361{ 1362 writeback_set_ratelimit(); 1363 return NOTIFY_DONE; 1364} 1365 1366static struct notifier_block __cpuinitdata ratelimit_nb = { 1367 .notifier_call = ratelimit_handler, 1368 .next = NULL, 1369}; 1370 1371/* 1372 * Called early on to tune the page writeback dirty limits. 1373 * 1374 * We used to scale dirty pages according to how total memory 1375 * related to pages that could be allocated for buffers (by 1376 * comparing nr_free_buffer_pages() to vm_total_pages. 1377 * 1378 * However, that was when we used "dirty_ratio" to scale with 1379 * all memory, and we don't do that any more. "dirty_ratio" 1380 * is now applied to total non-HIGHPAGE memory (by subtracting 1381 * totalhigh_pages from vm_total_pages), and as such we can't 1382 * get into the old insane situation any more where we had 1383 * large amounts of dirty pages compared to a small amount of 1384 * non-HIGHMEM memory. 1385 * 1386 * But we might still want to scale the dirty_ratio by how 1387 * much memory the box has.. 1388 */ 1389void __init page_writeback_init(void) 1390{ 1391 int shift; 1392 1393 writeback_set_ratelimit(); 1394 register_cpu_notifier(&ratelimit_nb); 1395 1396 shift = calc_period_shift(); 1397 prop_descriptor_init(&vm_completions, shift); 1398 prop_descriptor_init(&vm_dirties, shift); 1399} 1400 1401/** 1402 * tag_pages_for_writeback - tag pages to be written by write_cache_pages 1403 * @mapping: address space structure to write 1404 * @start: starting page index 1405 * @end: ending page index (inclusive) 1406 * 1407 * This function scans the page range from @start to @end (inclusive) and tags 1408 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is 1409 * that write_cache_pages (or whoever calls this function) will then use 1410 * TOWRITE tag to identify pages eligible for writeback. This mechanism is 1411 * used to avoid livelocking of writeback by a process steadily creating new 1412 * dirty pages in the file (thus it is important for this function to be quick 1413 * so that it can tag pages faster than a dirtying process can create them). 1414 */ 1415/* 1416 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency. 1417 */ 1418void tag_pages_for_writeback(struct address_space *mapping, 1419 pgoff_t start, pgoff_t end) 1420{ 1421#define WRITEBACK_TAG_BATCH 4096 1422 unsigned long tagged; 1423 1424 do { 1425 spin_lock_irq(&mapping->tree_lock); 1426 tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree, 1427 &start, end, WRITEBACK_TAG_BATCH, 1428 PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE); 1429 spin_unlock_irq(&mapping->tree_lock); 1430 WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH); 1431 cond_resched(); 1432 /* We check 'start' to handle wrapping when end == ~0UL */ 1433 } while (tagged >= WRITEBACK_TAG_BATCH && start); 1434} 1435EXPORT_SYMBOL(tag_pages_for_writeback); 1436 1437/** 1438 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. 1439 * @mapping: address space structure to write 1440 * @wbc: subtract the number of written pages from *@wbc->nr_to_write 1441 * @writepage: function called for each page 1442 * @data: data passed to writepage function 1443 * 1444 * If a page is already under I/O, write_cache_pages() skips it, even 1445 * if it's dirty. This is desirable behaviour for memory-cleaning writeback, 1446 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync() 1447 * and msync() need to guarantee that all the data which was dirty at the time 1448 * the call was made get new I/O started against them. If wbc->sync_mode is 1449 * WB_SYNC_ALL then we were called for data integrity and we must wait for 1450 * existing IO to complete. 1451 * 1452 * To avoid livelocks (when other process dirties new pages), we first tag 1453 * pages which should be written back with TOWRITE tag and only then start 1454 * writing them. For data-integrity sync we have to be careful so that we do 1455 * not miss some pages (e.g., because some other process has cleared TOWRITE 1456 * tag we set). The rule we follow is that TOWRITE tag can be cleared only 1457 * by the process clearing the DIRTY tag (and submitting the page for IO). 1458 */ 1459int write_cache_pages(struct address_space *mapping, 1460 struct writeback_control *wbc, writepage_t writepage, 1461 void *data) 1462{ 1463 int ret = 0; 1464 int done = 0; 1465 struct pagevec pvec; 1466 int nr_pages; 1467 pgoff_t uninitialized_var(writeback_index); 1468 pgoff_t index; 1469 pgoff_t end; /* Inclusive */ 1470 pgoff_t done_index; 1471 int cycled; 1472 int range_whole = 0; 1473 int tag; 1474 1475 pagevec_init(&pvec, 0); 1476 if (wbc->range_cyclic) { 1477 writeback_index = mapping->writeback_index; /* prev offset */ 1478 index = writeback_index; 1479 if (index == 0) 1480 cycled = 1; 1481 else 1482 cycled = 0; 1483 end = -1; 1484 } else { 1485 index = wbc->range_start >> PAGE_CACHE_SHIFT; 1486 end = wbc->range_end >> PAGE_CACHE_SHIFT; 1487 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX) 1488 range_whole = 1; 1489 cycled = 1; /* ignore range_cyclic tests */ 1490 } 1491 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) 1492 tag = PAGECACHE_TAG_TOWRITE; 1493 else 1494 tag = PAGECACHE_TAG_DIRTY; 1495retry: 1496 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) 1497 tag_pages_for_writeback(mapping, index, end); 1498 done_index = index; 1499 while (!done && (index <= end)) { 1500 int i; 1501 1502 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag, 1503 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1); 1504 if (nr_pages == 0) 1505 break; 1506 1507 for (i = 0; i < nr_pages; i++) { 1508 struct page *page = pvec.pages[i]; 1509 1510 /* 1511 * At this point, the page may be truncated or 1512 * invalidated (changing page->mapping to NULL), or 1513 * even swizzled back from swapper_space to tmpfs file 1514 * mapping. However, page->index will not change 1515 * because we have a reference on the page. 1516 */ 1517 if (page->index > end) { 1518 /* 1519 * can't be range_cyclic (1st pass) because 1520 * end == -1 in that case. 1521 */ 1522 done = 1; 1523 break; 1524 } 1525 1526 done_index = page->index; 1527 1528 lock_page(page); 1529 1530 /* 1531 * Page truncated or invalidated. We can freely skip it 1532 * then, even for data integrity operations: the page 1533 * has disappeared concurrently, so there could be no 1534 * real expectation of this data interity operation 1535 * even if there is now a new, dirty page at the same 1536 * pagecache address. 1537 */ 1538 if (unlikely(page->mapping != mapping)) { 1539continue_unlock: 1540 unlock_page(page); 1541 continue; 1542 } 1543 1544 if (!PageDirty(page)) { 1545 /* someone wrote it for us */ 1546 goto continue_unlock; 1547 } 1548 1549 if (PageWriteback(page)) { 1550 if (wbc->sync_mode != WB_SYNC_NONE) 1551 wait_on_page_writeback(page); 1552 else 1553 goto continue_unlock; 1554 } 1555 1556 BUG_ON(PageWriteback(page)); 1557 if (!clear_page_dirty_for_io(page)) 1558 goto continue_unlock; 1559 1560 trace_wbc_writepage(wbc, mapping->backing_dev_info); 1561 ret = (*writepage)(page, wbc, data); 1562 if (unlikely(ret)) { 1563 if (ret == AOP_WRITEPAGE_ACTIVATE) { 1564 unlock_page(page); 1565 ret = 0; 1566 } else { 1567 /* 1568 * done_index is set past this page, 1569 * so media errors will not choke 1570 * background writeout for the entire 1571 * file. This has consequences for 1572 * range_cyclic semantics (ie. it may 1573 * not be suitable for data integrity 1574 * writeout). 1575 */ 1576 done_index = page->index + 1; 1577 done = 1; 1578 break; 1579 } 1580 } 1581 1582 /* 1583 * We stop writing back only if we are not doing 1584 * integrity sync. In case of integrity sync we have to 1585 * keep going until we have written all the pages 1586 * we tagged for writeback prior to entering this loop. 1587 */ 1588 if (--wbc->nr_to_write <= 0 && 1589 wbc->sync_mode == WB_SYNC_NONE) { 1590 done = 1; 1591 break; 1592 } 1593 } 1594 pagevec_release(&pvec); 1595 cond_resched(); 1596 } 1597 if (!cycled && !done) { 1598 /* 1599 * range_cyclic: 1600 * We hit the last page and there is more work to be done: wrap 1601 * back to the start of the file 1602 */ 1603 cycled = 1; 1604 index = 0; 1605 end = writeback_index - 1; 1606 goto retry; 1607 } 1608 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) 1609 mapping->writeback_index = done_index; 1610 1611 return ret; 1612} 1613EXPORT_SYMBOL(write_cache_pages); 1614 1615/* 1616 * Function used by generic_writepages to call the real writepage 1617 * function and set the mapping flags on error 1618 */ 1619static int __writepage(struct page *page, struct writeback_control *wbc, 1620 void *data) 1621{ 1622 struct address_space *mapping = data; 1623 int ret = mapping->a_ops->writepage(page, wbc); 1624 mapping_set_error(mapping, ret); 1625 return ret; 1626} 1627 1628/** 1629 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them. 1630 * @mapping: address space structure to write 1631 * @wbc: subtract the number of written pages from *@wbc->nr_to_write 1632 * 1633 * This is a library function, which implements the writepages() 1634 * address_space_operation. 1635 */ 1636int generic_writepages(struct address_space *mapping, 1637 struct writeback_control *wbc) 1638{ 1639 struct blk_plug plug; 1640 int ret; 1641 1642 /* deal with chardevs and other special file */ 1643 if (!mapping->a_ops->writepage) 1644 return 0; 1645 1646 blk_start_plug(&plug); 1647 ret = write_cache_pages(mapping, wbc, __writepage, mapping); 1648 blk_finish_plug(&plug); 1649 return ret; 1650} 1651 1652EXPORT_SYMBOL(generic_writepages); 1653 1654int do_writepages(struct address_space *mapping, struct writeback_control *wbc) 1655{ 1656 int ret; 1657 1658 if (wbc->nr_to_write <= 0) 1659 return 0; 1660 if (mapping->a_ops->writepages) 1661 ret = mapping->a_ops->writepages(mapping, wbc); 1662 else 1663 ret = generic_writepages(mapping, wbc); 1664 return ret; 1665} 1666 1667/** 1668 * write_one_page - write out a single page and optionally wait on I/O 1669 * @page: the page to write 1670 * @wait: if true, wait on writeout 1671 * 1672 * The page must be locked by the caller and will be unlocked upon return. 1673 * 1674 * write_one_page() returns a negative error code if I/O failed. 1675 */ 1676int write_one_page(struct page *page, int wait) 1677{ 1678 struct address_space *mapping = page->mapping; 1679 int ret = 0; 1680 struct writeback_control wbc = { 1681 .sync_mode = WB_SYNC_ALL, 1682 .nr_to_write = 1, 1683 }; 1684 1685 BUG_ON(!PageLocked(page)); 1686 1687 if (wait) 1688 wait_on_page_writeback(page); 1689 1690 if (clear_page_dirty_for_io(page)) { 1691 page_cache_get(page); 1692 ret = mapping->a_ops->writepage(page, &wbc); 1693 if (ret == 0 && wait) { 1694 wait_on_page_writeback(page); 1695 if (PageError(page)) 1696 ret = -EIO; 1697 } 1698 page_cache_release(page); 1699 } else { 1700 unlock_page(page); 1701 } 1702 return ret; 1703} 1704EXPORT_SYMBOL(write_one_page); 1705 1706/* 1707 * For address_spaces which do not use buffers nor write back. 1708 */ 1709int __set_page_dirty_no_writeback(struct page *page) 1710{ 1711 if (!PageDirty(page)) 1712 return !TestSetPageDirty(page); 1713 return 0; 1714} 1715 1716/* 1717 * Helper function for set_page_dirty family. 1718 * NOTE: This relies on being atomic wrt interrupts. 1719 */ 1720void account_page_dirtied(struct page *page, struct address_space *mapping) 1721{ 1722 if (mapping_cap_account_dirty(mapping)) { 1723 __inc_zone_page_state(page, NR_FILE_DIRTY); 1724 __inc_zone_page_state(page, NR_DIRTIED); 1725 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE); 1726 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED); 1727 task_dirty_inc(current); 1728 task_io_account_write(PAGE_CACHE_SIZE); 1729 } 1730} 1731EXPORT_SYMBOL(account_page_dirtied); 1732 1733/* 1734 * Helper function for set_page_writeback family. 1735 * NOTE: Unlike account_page_dirtied this does not rely on being atomic 1736 * wrt interrupts. 1737 */ 1738void account_page_writeback(struct page *page) 1739{ 1740 inc_zone_page_state(page, NR_WRITEBACK); 1741} 1742EXPORT_SYMBOL(account_page_writeback); 1743 1744/* 1745 * For address_spaces which do not use buffers. Just tag the page as dirty in 1746 * its radix tree. 1747 * 1748 * This is also used when a single buffer is being dirtied: we want to set the 1749 * page dirty in that case, but not all the buffers. This is a "bottom-up" 1750 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying. 1751 * 1752 * Most callers have locked the page, which pins the address_space in memory. 1753 * But zap_pte_range() does not lock the page, however in that case the 1754 * mapping is pinned by the vma's ->vm_file reference. 1755 * 1756 * We take care to handle the case where the page was truncated from the 1757 * mapping by re-checking page_mapping() inside tree_lock. 1758 */ 1759int __set_page_dirty_nobuffers(struct page *page) 1760{ 1761 if (!TestSetPageDirty(page)) { 1762 struct address_space *mapping = page_mapping(page); 1763 struct address_space *mapping2; 1764 1765 if (!mapping) 1766 return 1; 1767 1768 spin_lock_irq(&mapping->tree_lock); 1769 mapping2 = page_mapping(page); 1770 if (mapping2) { /* Race with truncate? */ 1771 BUG_ON(mapping2 != mapping); 1772 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page)); 1773 account_page_dirtied(page, mapping); 1774 radix_tree_tag_set(&mapping->page_tree, 1775 page_index(page), PAGECACHE_TAG_DIRTY); 1776 } 1777 spin_unlock_irq(&mapping->tree_lock); 1778 if (mapping->host) { 1779 /* !PageAnon && !swapper_space */ 1780 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 1781 } 1782 return 1; 1783 } 1784 return 0; 1785} 1786EXPORT_SYMBOL(__set_page_dirty_nobuffers); 1787 1788/* 1789 * When a writepage implementation decides that it doesn't want to write this 1790 * page for some reason, it should redirty the locked page via 1791 * redirty_page_for_writepage() and it should then unlock the page and return 0 1792 */ 1793int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page) 1794{ 1795 wbc->pages_skipped++; 1796 return __set_page_dirty_nobuffers(page); 1797} 1798EXPORT_SYMBOL(redirty_page_for_writepage); 1799 1800/* 1801 * Dirty a page. 1802 * 1803 * For pages with a mapping this should be done under the page lock 1804 * for the benefit of asynchronous memory errors who prefer a consistent 1805 * dirty state. This rule can be broken in some special cases, 1806 * but should be better not to. 1807 * 1808 * If the mapping doesn't provide a set_page_dirty a_op, then 1809 * just fall through and assume that it wants buffer_heads. 1810 */ 1811int set_page_dirty(struct page *page) 1812{ 1813 struct address_space *mapping = page_mapping(page); 1814 1815 if (likely(mapping)) { 1816 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty; 1817 /* 1818 * readahead/lru_deactivate_page could remain 1819 * PG_readahead/PG_reclaim due to race with end_page_writeback 1820 * About readahead, if the page is written, the flags would be 1821 * reset. So no problem. 1822 * About lru_deactivate_page, if the page is redirty, the flag 1823 * will be reset. So no problem. but if the page is used by readahead 1824 * it will confuse readahead and make it restart the size rampup 1825 * process. But it's a trivial problem. 1826 */ 1827 ClearPageReclaim(page); 1828#ifdef CONFIG_BLOCK 1829 if (!spd) 1830 spd = __set_page_dirty_buffers; 1831#endif 1832 return (*spd)(page); 1833 } 1834 if (!PageDirty(page)) { 1835 if (!TestSetPageDirty(page)) 1836 return 1; 1837 } 1838 return 0; 1839} 1840EXPORT_SYMBOL(set_page_dirty); 1841 1842/* 1843 * set_page_dirty() is racy if the caller has no reference against 1844 * page->mapping->host, and if the page is unlocked. This is because another 1845 * CPU could truncate the page off the mapping and then free the mapping. 1846 * 1847 * Usually, the page _is_ locked, or the caller is a user-space process which 1848 * holds a reference on the inode by having an open file. 1849 * 1850 * In other cases, the page should be locked before running set_page_dirty(). 1851 */ 1852int set_page_dirty_lock(struct page *page) 1853{ 1854 int ret; 1855 1856 lock_page(page); 1857 ret = set_page_dirty(page); 1858 unlock_page(page); 1859 return ret; 1860} 1861EXPORT_SYMBOL(set_page_dirty_lock); 1862 1863/* 1864 * Clear a page's dirty flag, while caring for dirty memory accounting. 1865 * Returns true if the page was previously dirty. 1866 * 1867 * This is for preparing to put the page under writeout. We leave the page 1868 * tagged as dirty in the radix tree so that a concurrent write-for-sync 1869 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage 1870 * implementation will run either set_page_writeback() or set_page_dirty(), 1871 * at which stage we bring the page's dirty flag and radix-tree dirty tag 1872 * back into sync. 1873 * 1874 * This incoherency between the page's dirty flag and radix-tree tag is 1875 * unfortunate, but it only exists while the page is locked. 1876 */ 1877int clear_page_dirty_for_io(struct page *page) 1878{ 1879 struct address_space *mapping = page_mapping(page); 1880 1881 BUG_ON(!PageLocked(page)); 1882 1883 if (mapping && mapping_cap_account_dirty(mapping)) { 1884 /* 1885 * Yes, Virginia, this is indeed insane. 1886 * 1887 * We use this sequence to make sure that 1888 * (a) we account for dirty stats properly 1889 * (b) we tell the low-level filesystem to 1890 * mark the whole page dirty if it was 1891 * dirty in a pagetable. Only to then 1892 * (c) clean the page again and return 1 to 1893 * cause the writeback. 1894 * 1895 * This way we avoid all nasty races with the 1896 * dirty bit in multiple places and clearing 1897 * them concurrently from different threads. 1898 * 1899 * Note! Normally the "set_page_dirty(page)" 1900 * has no effect on the actual dirty bit - since 1901 * that will already usually be set. But we 1902 * need the side effects, and it can help us 1903 * avoid races. 1904 * 1905 * We basically use the page "master dirty bit" 1906 * as a serialization point for all the different 1907 * threads doing their things. 1908 */ 1909 if (page_mkclean(page)) 1910 set_page_dirty(page); 1911 /* 1912 * We carefully synchronise fault handlers against 1913 * installing a dirty pte and marking the page dirty 1914 * at this point. We do this by having them hold the 1915 * page lock at some point after installing their 1916 * pte, but before marking the page dirty. 1917 * Pages are always locked coming in here, so we get 1918 * the desired exclusion. See mm/memory.c:do_wp_page() 1919 * for more comments. 1920 */ 1921 if (TestClearPageDirty(page)) { 1922 dec_zone_page_state(page, NR_FILE_DIRTY); 1923 dec_bdi_stat(mapping->backing_dev_info, 1924 BDI_RECLAIMABLE); 1925 return 1; 1926 } 1927 return 0; 1928 } 1929 return TestClearPageDirty(page); 1930} 1931EXPORT_SYMBOL(clear_page_dirty_for_io); 1932 1933int test_clear_page_writeback(struct page *page) 1934{ 1935 struct address_space *mapping = page_mapping(page); 1936 int ret; 1937 1938 if (mapping) { 1939 struct backing_dev_info *bdi = mapping->backing_dev_info; 1940 unsigned long flags; 1941 1942 spin_lock_irqsave(&mapping->tree_lock, flags); 1943 ret = TestClearPageWriteback(page); 1944 if (ret) { 1945 radix_tree_tag_clear(&mapping->page_tree, 1946 page_index(page), 1947 PAGECACHE_TAG_WRITEBACK); 1948 if (bdi_cap_account_writeback(bdi)) { 1949 __dec_bdi_stat(bdi, BDI_WRITEBACK); 1950 __bdi_writeout_inc(bdi); 1951 } 1952 } 1953 spin_unlock_irqrestore(&mapping->tree_lock, flags); 1954 } else { 1955 ret = TestClearPageWriteback(page); 1956 } 1957 if (ret) { 1958 dec_zone_page_state(page, NR_WRITEBACK); 1959 inc_zone_page_state(page, NR_WRITTEN); 1960 } 1961 return ret; 1962} 1963 1964int test_set_page_writeback(struct page *page) 1965{ 1966 struct address_space *mapping = page_mapping(page); 1967 int ret; 1968 1969 if (mapping) { 1970 struct backing_dev_info *bdi = mapping->backing_dev_info; 1971 unsigned long flags; 1972 1973 spin_lock_irqsave(&mapping->tree_lock, flags); 1974 ret = TestSetPageWriteback(page); 1975 if (!ret) { 1976 radix_tree_tag_set(&mapping->page_tree, 1977 page_index(page), 1978 PAGECACHE_TAG_WRITEBACK); 1979 if (bdi_cap_account_writeback(bdi)) 1980 __inc_bdi_stat(bdi, BDI_WRITEBACK); 1981 } 1982 if (!PageDirty(page)) 1983 radix_tree_tag_clear(&mapping->page_tree, 1984 page_index(page), 1985 PAGECACHE_TAG_DIRTY); 1986 radix_tree_tag_clear(&mapping->page_tree, 1987 page_index(page), 1988 PAGECACHE_TAG_TOWRITE); 1989 spin_unlock_irqrestore(&mapping->tree_lock, flags); 1990 } else { 1991 ret = TestSetPageWriteback(page); 1992 } 1993 if (!ret) 1994 account_page_writeback(page); 1995 return ret; 1996 1997} 1998EXPORT_SYMBOL(test_set_page_writeback); 1999 2000/* 2001 * Return true if any of the pages in the mapping are marked with the 2002 * passed tag. 2003 */ 2004int mapping_tagged(struct address_space *mapping, int tag) 2005{ 2006 return radix_tree_tagged(&mapping->page_tree, tag); 2007} 2008EXPORT_SYMBOL(mapping_tagged); 2009