fs-writeback.c

   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * fs/fs-writeback.c
   4 *
   5 * Copyright (C) 2002, Linus Torvalds.
   6 *
   7 * Contains all the functions related to writing back and waiting
   8 * upon dirty inodes against superblocks, and writing back dirty
   9 * pages against inodes.  ie: data writeback.  Writeout of the
  10 * inode itself is not handled here.
  11 *
  12 * 10Apr2002	Andrew Morton
  13 *		Split out of fs/inode.c
  14 *		Additions for address_space-based writeback
  15 */
  16
  17#include <linux/sched/sysctl.h>
  18#include <linux/kernel.h>
  19#include <linux/export.h>
  20#include <linux/spinlock.h>
  21#include <linux/slab.h>
  22#include <linux/sched.h>
  23#include <linux/fs.h>
  24#include <linux/mm.h>
  25#include <linux/pagemap.h>
  26#include <linux/kthread.h>
  27#include <linux/writeback.h>
  28#include <linux/blkdev.h>
  29#include <linux/backing-dev.h>
  30#include <linux/tracepoint.h>
  31#include <linux/device.h>
  32#include <linux/memcontrol.h>
  33#include "internal.h"
  34
  35/*
  36 * Passed into wb_writeback(), essentially a subset of writeback_control
  37 */
  38struct wb_writeback_work {
  39	long nr_pages;
  40	struct super_block *sb;
  41	enum writeback_sync_modes sync_mode;
  42	unsigned int tagged_writepages:1;
  43	unsigned int for_kupdate:1;
  44	unsigned int range_cyclic:1;
  45	unsigned int for_background:1;
  46	unsigned int for_sync:1;	/* sync(2) WB_SYNC_ALL writeback */
  47	unsigned int auto_free:1;	/* free on completion */
  48	enum wb_reason reason;		/* why was writeback initiated? */
  49
  50	struct list_head list;		/* pending work list */
  51	struct wb_completion *done;	/* set if the caller waits */
  52};
  53
  54/*
  55 * If an inode is constantly having its pages dirtied, but then the
  56 * updates stop dirtytime_expire_interval seconds in the past, it's
  57 * possible for the worst case time between when an inode has its
  58 * timestamps updated and when they finally get written out to be two
  59 * dirtytime_expire_intervals.  We set the default to 12 hours (in
  60 * seconds), which means most of the time inodes will have their
  61 * timestamps written to disk after 12 hours, but in the worst case a
  62 * few inodes might not their timestamps updated for 24 hours.
  63 */
  64static unsigned int dirtytime_expire_interval = 12 * 60 * 60;
  65
  66static inline struct inode *wb_inode(struct list_head *head)
  67{
  68	return list_entry(head, struct inode, i_io_list);
  69}
  70
  71/*
  72 * Include the creation of the trace points after defining the
  73 * wb_writeback_work structure and inline functions so that the definition
  74 * remains local to this file.
  75 */
  76#define CREATE_TRACE_POINTS
  77#include <trace/events/writeback.h>
  78
  79EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
  80
  81static bool wb_io_lists_populated(struct bdi_writeback *wb)
  82{
  83	if (wb_has_dirty_io(wb)) {
  84		return false;
  85	} else {
  86		set_bit(WB_has_dirty_io, &wb->state);
  87		WARN_ON_ONCE(!wb->avg_write_bandwidth);
  88		atomic_long_add(wb->avg_write_bandwidth,
  89				&wb->bdi->tot_write_bandwidth);
  90		return true;
  91	}
  92}
  93
  94static void wb_io_lists_depopulated(struct bdi_writeback *wb)
  95{
  96	if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
  97	    list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
  98		clear_bit(WB_has_dirty_io, &wb->state);
  99		WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
 100					&wb->bdi->tot_write_bandwidth) < 0);
 101	}
 102}
 103
 104/**
 105 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
 106 * @inode: inode to be moved
 107 * @wb: target bdi_writeback
 108 * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
 109 *
 110 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
 111 * Returns %true if @inode is the first occupant of the !dirty_time IO
 112 * lists; otherwise, %false.
 113 */
 114static bool inode_io_list_move_locked(struct inode *inode,
 115				      struct bdi_writeback *wb,
 116				      struct list_head *head)
 117{
 118	assert_spin_locked(&wb->list_lock);
 119	assert_spin_locked(&inode->i_lock);
 120	WARN_ON_ONCE(inode_state_read(inode) & I_FREEING);
 121
 122	list_move(&inode->i_io_list, head);
 123
 124	/* dirty_time doesn't count as dirty_io until expiration */
 125	if (head != &wb->b_dirty_time)
 126		return wb_io_lists_populated(wb);
 127
 128	wb_io_lists_depopulated(wb);
 129	return false;
 130}
 131
 132static void wb_wakeup(struct bdi_writeback *wb)
 133{
 134	spin_lock_irq(&wb->work_lock);
 135	if (test_bit(WB_registered, &wb->state))
 136		mod_delayed_work(bdi_wq, &wb->dwork, 0);
 137	spin_unlock_irq(&wb->work_lock);
 138}
 139
 140/*
 141 * This function is used when the first inode for this wb is marked dirty. It
 142 * wakes-up the corresponding bdi thread which should then take care of the
 143 * periodic background write-out of dirty inodes. Since the write-out would
 144 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 145 * set up a timer which wakes the bdi thread up later.
 146 *
 147 * Note, we wouldn't bother setting up the timer, but this function is on the
 148 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 149 * by delaying the wake-up.
 150 *
 151 * We have to be careful not to postpone flush work if it is scheduled for
 152 * earlier. Thus we use queue_delayed_work().
 153 */
 154static void wb_wakeup_delayed(struct bdi_writeback *wb)
 155{
 156	unsigned long timeout;
 157
 158	timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 159	spin_lock_irq(&wb->work_lock);
 160	if (test_bit(WB_registered, &wb->state))
 161		queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 162	spin_unlock_irq(&wb->work_lock);
 163}
 164
 165static void finish_writeback_work(struct wb_writeback_work *work)
 166{
 167	struct wb_completion *done = work->done;
 168
 169	if (work->auto_free)
 170		kfree(work);
 171	if (done) {
 172		wait_queue_head_t *waitq = done->waitq;
 173
 174		/* @done can't be accessed after the following dec */
 175		if (atomic_dec_and_test(&done->cnt))
 176			wake_up_all(waitq);
 177	}
 178}
 179
 180static void wb_queue_work(struct bdi_writeback *wb,
 181			  struct wb_writeback_work *work)
 182{
 183	trace_writeback_queue(wb, work);
 184
 185	if (work->done)
 186		atomic_inc(&work->done->cnt);
 187
 188	spin_lock_irq(&wb->work_lock);
 189
 190	if (test_bit(WB_registered, &wb->state)) {
 191		list_add_tail(&work->list, &wb->work_list);
 192		mod_delayed_work(bdi_wq, &wb->dwork, 0);
 193	} else
 194		finish_writeback_work(work);
 195
 196	spin_unlock_irq(&wb->work_lock);
 197}
 198
 199static bool wb_wait_for_completion_cb(struct wb_completion *done)
 200{
 201	unsigned long timeout = sysctl_hung_task_timeout_secs;
 202	unsigned long waited_secs = (jiffies - done->wait_start) / HZ;
 203
 204	done->progress_stamp = jiffies;
 205	if (timeout && (waited_secs > timeout))
 206		pr_info("INFO: The task %s:%d has been waiting for writeback "
 207			"completion for more than %lu seconds.",
 208			current->comm, current->pid, waited_secs);
 209
 210	return !atomic_read(&done->cnt);
 211}
 212
 213/**
 214 * wb_wait_for_completion - wait for completion of bdi_writeback_works
 215 * @done: target wb_completion
 216 *
 217 * Wait for one or more work items issued to @bdi with their ->done field
 218 * set to @done, which should have been initialized with
 219 * DEFINE_WB_COMPLETION().  This function returns after all such work items
 220 * are completed.  Work items which are waited upon aren't freed
 221 * automatically on completion.
 222 */
 223void wb_wait_for_completion(struct wb_completion *done)
 224{
 225	done->wait_start = jiffies;
 226	atomic_dec(&done->cnt);		/* put down the initial count */
 227	wait_event(*done->waitq, wb_wait_for_completion_cb(done));
 228}
 229
 230#ifdef CONFIG_CGROUP_WRITEBACK
 231
 232/*
 233 * Parameters for foreign inode detection, see wbc_detach_inode() to see
 234 * how they're used.
 235 *
 236 * These paramters are inherently heuristical as the detection target
 237 * itself is fuzzy.  All we want to do is detaching an inode from the
 238 * current owner if it's being written to by some other cgroups too much.
 239 *
 240 * The current cgroup writeback is built on the assumption that multiple
 241 * cgroups writing to the same inode concurrently is very rare and a mode
 242 * of operation which isn't well supported.  As such, the goal is not
 243 * taking too long when a different cgroup takes over an inode while
 244 * avoiding too aggressive flip-flops from occasional foreign writes.
 245 *
 246 * We record, very roughly, 2s worth of IO time history and if more than
 247 * half of that is foreign, trigger the switch.  The recording is quantized
 248 * to 16 slots.  To avoid tiny writes from swinging the decision too much,
 249 * writes smaller than 1/8 of avg size are ignored.
 250 */
 251#define WB_FRN_TIME_SHIFT	13	/* 1s = 2^13, upto 8 secs w/ 16bit */
 252#define WB_FRN_TIME_AVG_SHIFT	3	/* avg = avg * 7/8 + new * 1/8 */
 253#define WB_FRN_TIME_CUT_DIV	8	/* ignore rounds < avg / 8 */
 254#define WB_FRN_TIME_PERIOD	(2 * (1 << WB_FRN_TIME_SHIFT))	/* 2s */
 255
 256#define WB_FRN_HIST_SLOTS	16	/* inode->i_wb_frn_history is 16bit */
 257#define WB_FRN_HIST_UNIT	(WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
 258					/* each slot's duration is 2s / 16 */
 259#define WB_FRN_HIST_THR_SLOTS	(WB_FRN_HIST_SLOTS / 2)
 260					/* if foreign slots >= 8, switch */
 261#define WB_FRN_HIST_MAX_SLOTS	(WB_FRN_HIST_THR_SLOTS / 2 + 1)
 262					/* one round can affect upto 5 slots */
 263#define WB_FRN_MAX_IN_FLIGHT	1024	/* don't queue too many concurrently */
 264
 265/*
 266 * Maximum inodes per isw.  A specific value has been chosen to make
 267 * struct inode_switch_wbs_context fit into 1024 bytes kmalloc.
 268 */
 269#define WB_MAX_INODES_PER_ISW  ((1024UL - sizeof(struct inode_switch_wbs_context)) \
 270                                / sizeof(struct inode *))
 271
 272static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
 273static struct workqueue_struct *isw_wq;
 274
 275void __inode_attach_wb(struct inode *inode, struct folio *folio)
 276{
 277	struct backing_dev_info *bdi = inode_to_bdi(inode);
 278	struct bdi_writeback *wb = NULL;
 279
 280	if (inode_cgwb_enabled(inode)) {
 281		struct cgroup_subsys_state *memcg_css;
 282
 283		if (folio) {
 284			memcg_css = mem_cgroup_css_from_folio(folio);
 285			wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
 286		} else {
 287			/* must pin memcg_css, see wb_get_create() */
 288			memcg_css = task_get_css(current, memory_cgrp_id);
 289			wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
 290			css_put(memcg_css);
 291		}
 292	}
 293
 294	if (!wb)
 295		wb = &bdi->wb;
 296
 297	/*
 298	 * There may be multiple instances of this function racing to
 299	 * update the same inode.  Use cmpxchg() to tell the winner.
 300	 */
 301	if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
 302		wb_put(wb);
 303}
 304
 305/**
 306 * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list
 307 * @inode: inode of interest with i_lock held
 308 * @wb: target bdi_writeback
 309 *
 310 * Remove the inode from wb's io lists and if necessarily put onto b_attached
 311 * list.  Only inodes attached to cgwb's are kept on this list.
 312 */
 313static void inode_cgwb_move_to_attached(struct inode *inode,
 314					struct bdi_writeback *wb)
 315{
 316	assert_spin_locked(&wb->list_lock);
 317	assert_spin_locked(&inode->i_lock);
 318	WARN_ON_ONCE(inode_state_read(inode) & I_FREEING);
 319
 320	inode_state_clear(inode, I_SYNC_QUEUED);
 321	if (wb != &wb->bdi->wb)
 322		list_move(&inode->i_io_list, &wb->b_attached);
 323	else
 324		list_del_init(&inode->i_io_list);
 325	wb_io_lists_depopulated(wb);
 326}
 327
 328/**
 329 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
 330 * @inode: inode of interest with i_lock held
 331 *
 332 * Returns @inode's wb with its list_lock held.  @inode->i_lock must be
 333 * held on entry and is released on return.  The returned wb is guaranteed
 334 * to stay @inode's associated wb until its list_lock is released.
 335 */
 336static struct bdi_writeback *
 337locked_inode_to_wb_and_lock_list(struct inode *inode)
 338	__releases(&inode->i_lock)
 339	__acquires(&wb->list_lock)
 340{
 341	while (true) {
 342		struct bdi_writeback *wb = inode_to_wb(inode);
 343
 344		/*
 345		 * inode_to_wb() association is protected by both
 346		 * @inode->i_lock and @wb->list_lock but list_lock nests
 347		 * outside i_lock.  Drop i_lock and verify that the
 348		 * association hasn't changed after acquiring list_lock.
 349		 */
 350		wb_get(wb);
 351		spin_unlock(&inode->i_lock);
 352		spin_lock(&wb->list_lock);
 353
 354		/* i_wb may have changed inbetween, can't use inode_to_wb() */
 355		if (likely(wb == inode->i_wb)) {
 356			wb_put(wb);	/* @inode already has ref */
 357			return wb;
 358		}
 359
 360		spin_unlock(&wb->list_lock);
 361		wb_put(wb);
 362		cpu_relax();
 363		spin_lock(&inode->i_lock);
 364	}
 365}
 366
 367/**
 368 * inode_to_wb_and_lock_list - determine an inode's wb and lock it
 369 * @inode: inode of interest
 370 *
 371 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
 372 * on entry.
 373 */
 374static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
 375	__acquires(&wb->list_lock)
 376{
 377	spin_lock(&inode->i_lock);
 378	return locked_inode_to_wb_and_lock_list(inode);
 379}
 380
 381struct inode_switch_wbs_context {
 382	/* List of queued switching contexts for the wb */
 383	struct llist_node	list;
 384
 385	/*
 386	 * Multiple inodes can be switched at once.  The switching procedure
 387	 * consists of two parts, separated by a RCU grace period.  To make
 388	 * sure that the second part is executed for each inode gone through
 389	 * the first part, all inode pointers are placed into a NULL-terminated
 390	 * array embedded into struct inode_switch_wbs_context.  Otherwise
 391	 * an inode could be left in a non-consistent state.
 392	 */
 393	struct inode		*inodes[];
 394};
 395
 396static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
 397{
 398	down_write(&bdi->wb_switch_rwsem);
 399}
 400
 401static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
 402{
 403	up_write(&bdi->wb_switch_rwsem);
 404}
 405
 406static bool inode_do_switch_wbs(struct inode *inode,
 407				struct bdi_writeback *old_wb,
 408				struct bdi_writeback *new_wb)
 409{
 410	struct address_space *mapping = inode->i_mapping;
 411	XA_STATE(xas, &mapping->i_pages, 0);
 412	struct folio *folio;
 413	bool switched = false;
 414
 415	spin_lock(&inode->i_lock);
 416	xa_lock_irq(&mapping->i_pages);
 417
 418	/*
 419	 * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction
 420	 * path owns the inode and we shouldn't modify ->i_io_list.
 421	 */
 422	if (unlikely(inode_state_read(inode) & (I_FREEING | I_WILL_FREE)))
 423		goto skip_switch;
 424
 425	trace_inode_switch_wbs(inode, old_wb, new_wb);
 426
 427	/*
 428	 * Count and transfer stats.  Note that PAGECACHE_TAG_DIRTY points
 429	 * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to
 430	 * folios actually under writeback.
 431	 */
 432	xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
 433		if (folio_test_dirty(folio)) {
 434			long nr = folio_nr_pages(folio);
 435			wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr);
 436			wb_stat_mod(new_wb, WB_RECLAIMABLE, nr);
 437		}
 438	}
 439
 440	xas_set(&xas, 0);
 441	xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
 442		long nr = folio_nr_pages(folio);
 443		WARN_ON_ONCE(!folio_test_writeback(folio));
 444		wb_stat_mod(old_wb, WB_WRITEBACK, -nr);
 445		wb_stat_mod(new_wb, WB_WRITEBACK, nr);
 446	}
 447
 448	if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) {
 449		atomic_dec(&old_wb->writeback_inodes);
 450		atomic_inc(&new_wb->writeback_inodes);
 451	}
 452
 453	wb_get(new_wb);
 454
 455	/*
 456	 * Transfer to @new_wb's IO list if necessary.  If the @inode is dirty,
 457	 * the specific list @inode was on is ignored and the @inode is put on
 458	 * ->b_dirty which is always correct including from ->b_dirty_time.
 459	 * If the @inode was clean, it means it was on the b_attached list, so
 460	 * move it onto the b_attached list of @new_wb.
 461	 */
 462	if (!list_empty(&inode->i_io_list)) {
 463		inode->i_wb = new_wb;
 464
 465		if (inode_state_read(inode) & I_DIRTY_ALL) {
 466			/*
 467			 * We need to keep b_dirty list sorted by
 468			 * dirtied_time_when. However properly sorting the
 469			 * inode in the list gets too expensive when switching
 470			 * many inodes. So just attach inode at the end of the
 471			 * dirty list and clobber the dirtied_time_when.
 472			 */
 473			inode->dirtied_time_when = jiffies;
 474			inode_io_list_move_locked(inode, new_wb,
 475						  &new_wb->b_dirty);
 476		} else {
 477			inode_cgwb_move_to_attached(inode, new_wb);
 478		}
 479	} else {
 480		inode->i_wb = new_wb;
 481	}
 482
 483	/* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
 484	inode->i_wb_frn_winner = 0;
 485	inode->i_wb_frn_avg_time = 0;
 486	inode->i_wb_frn_history = 0;
 487	switched = true;
 488skip_switch:
 489	/*
 490	 * Paired with an acquire fence in unlocked_inode_to_wb_begin() and
 491	 * ensures that the new wb is visible if they see !I_WB_SWITCH.
 492	 */
 493	smp_wmb();
 494	inode_state_clear(inode, I_WB_SWITCH);
 495
 496	xa_unlock_irq(&mapping->i_pages);
 497	spin_unlock(&inode->i_lock);
 498
 499	return switched;
 500}
 501
 502static void process_inode_switch_wbs(struct bdi_writeback *new_wb,
 503				     struct inode_switch_wbs_context *isw)
 504{
 505	struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]);
 506	struct bdi_writeback *old_wb = isw->inodes[0]->i_wb;
 507	unsigned long nr_switched = 0;
 508	struct inode **inodep;
 509
 510	/*
 511	 * If @inode switches cgwb membership while sync_inodes_sb() is
 512	 * being issued, sync_inodes_sb() might miss it.  Synchronize.
 513	 */
 514	down_read(&bdi->wb_switch_rwsem);
 515
 516	inodep = isw->inodes;
 517	/*
 518	 * By the time control reaches here, RCU grace period has passed
 519	 * since I_WB_SWITCH assertion and all wb stat update transactions
 520	 * between unlocked_inode_to_wb_begin/end() are guaranteed to be
 521	 * synchronizing against the i_pages lock.
 522	 *
 523	 * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
 524	 * gives us exclusion against all wb related operations on @inode
 525	 * including IO list manipulations and stat updates.
 526	 */
 527relock:
 528	if (old_wb < new_wb) {
 529		spin_lock(&old_wb->list_lock);
 530		spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
 531	} else {
 532		spin_lock(&new_wb->list_lock);
 533		spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
 534	}
 535
 536	while (*inodep) {
 537		WARN_ON_ONCE((*inodep)->i_wb != old_wb);
 538		if (inode_do_switch_wbs(*inodep, old_wb, new_wb))
 539			nr_switched++;
 540		inodep++;
 541		if (*inodep && need_resched()) {
 542			spin_unlock(&new_wb->list_lock);
 543			spin_unlock(&old_wb->list_lock);
 544			cond_resched();
 545			goto relock;
 546		}
 547	}
 548
 549	spin_unlock(&new_wb->list_lock);
 550	spin_unlock(&old_wb->list_lock);
 551
 552	up_read(&bdi->wb_switch_rwsem);
 553
 554	if (nr_switched) {
 555		wb_wakeup(new_wb);
 556		wb_put_many(old_wb, nr_switched);
 557	}
 558
 559	for (inodep = isw->inodes; *inodep; inodep++)
 560		iput(*inodep);
 561	wb_put(new_wb);
 562	kfree(isw);
 563	atomic_dec(&isw_nr_in_flight);
 564}
 565
 566void inode_switch_wbs_work_fn(struct work_struct *work)
 567{
 568	struct bdi_writeback *new_wb = container_of(work, struct bdi_writeback,
 569						    switch_work);
 570	struct inode_switch_wbs_context *isw, *next_isw;
 571	struct llist_node *list;
 572
 573	/*
 574	 * Grab out reference to wb so that it cannot get freed under us
 575	 * after we process all the isw items.
 576	 */
 577	wb_get(new_wb);
 578	while (1) {
 579		list = llist_del_all(&new_wb->switch_wbs_ctxs);
 580		/* Nothing to do? */
 581		if (!list)
 582			break;
 583		/*
 584		 * In addition to synchronizing among switchers, I_WB_SWITCH
 585		 * tells the RCU protected stat update paths to grab the i_page
 586		 * lock so that stat transfer can synchronize against them.
 587		 * Let's continue after I_WB_SWITCH is guaranteed to be
 588		 * visible.
 589		 */
 590		synchronize_rcu();
 591
 592		llist_for_each_entry_safe(isw, next_isw, list, list)
 593			process_inode_switch_wbs(new_wb, isw);
 594	}
 595	wb_put(new_wb);
 596}
 597
 598static bool inode_prepare_wbs_switch(struct inode *inode,
 599				     struct bdi_writeback *new_wb)
 600{
 601	/*
 602	 * Paired with smp_mb() in cgroup_writeback_umount().
 603	 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
 604	 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
 605	 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
 606	 */
 607	smp_mb();
 608
 609	if (IS_DAX(inode))
 610		return false;
 611
 612	/* while holding I_WB_SWITCH, no one else can update the association */
 613	spin_lock(&inode->i_lock);
 614	if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
 615	    inode_state_read(inode) & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
 616	    inode_to_wb(inode) == new_wb) {
 617		spin_unlock(&inode->i_lock);
 618		return false;
 619	}
 620	inode_state_set(inode, I_WB_SWITCH);
 621	__iget(inode);
 622	spin_unlock(&inode->i_lock);
 623
 624	return true;
 625}
 626
 627static void wb_queue_isw(struct bdi_writeback *wb,
 628			 struct inode_switch_wbs_context *isw)
 629{
 630	if (llist_add(&isw->list, &wb->switch_wbs_ctxs))
 631		queue_work(isw_wq, &wb->switch_work);
 632}
 633
 634/**
 635 * inode_switch_wbs - change the wb association of an inode
 636 * @inode: target inode
 637 * @new_wb_id: ID of the new wb
 638 *
 639 * Switch @inode's wb association to the wb identified by @new_wb_id.  The
 640 * switching is performed asynchronously and may fail silently.
 641 */
 642static void inode_switch_wbs(struct inode *inode, int new_wb_id)
 643{
 644	struct backing_dev_info *bdi = inode_to_bdi(inode);
 645	struct cgroup_subsys_state *memcg_css;
 646	struct inode_switch_wbs_context *isw;
 647	struct bdi_writeback *new_wb = NULL;
 648
 649	/* noop if seems to be already in progress */
 650	if (inode_state_read_once(inode) & I_WB_SWITCH)
 651		return;
 652
 653	/* avoid queueing a new switch if too many are already in flight */
 654	if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
 655		return;
 656
 657	isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC);
 658	if (!isw)
 659		return;
 660
 661	atomic_inc(&isw_nr_in_flight);
 662
 663	/* find and pin the new wb */
 664	rcu_read_lock();
 665	memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
 666	if (memcg_css && !css_tryget(memcg_css))
 667		memcg_css = NULL;
 668	rcu_read_unlock();
 669	if (!memcg_css)
 670		goto out_free;
 671
 672	new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
 673	css_put(memcg_css);
 674	if (!new_wb)
 675		goto out_free;
 676
 677	if (!inode_prepare_wbs_switch(inode, new_wb))
 678		goto out_free;
 679
 680	isw->inodes[0] = inode;
 681
 682	trace_inode_switch_wbs_queue(inode->i_wb, new_wb, 1);
 683	wb_queue_isw(new_wb, isw);
 684	return;
 685
 686out_free:
 687	atomic_dec(&isw_nr_in_flight);
 688	if (new_wb)
 689		wb_put(new_wb);
 690	kfree(isw);
 691}
 692
 693static bool isw_prepare_wbs_switch(struct bdi_writeback *new_wb,
 694				   struct inode_switch_wbs_context *isw,
 695				   struct list_head *list, int *nr)
 696{
 697	struct inode *inode;
 698
 699	list_for_each_entry(inode, list, i_io_list) {
 700		if (!inode_prepare_wbs_switch(inode, new_wb))
 701			continue;
 702
 703		isw->inodes[*nr] = inode;
 704		(*nr)++;
 705
 706		if (*nr >= WB_MAX_INODES_PER_ISW - 1)
 707			return true;
 708	}
 709	return false;
 710}
 711
 712/**
 713 * cleanup_offline_cgwb - detach associated inodes
 714 * @wb: target wb
 715 *
 716 * Switch all inodes attached to @wb to a nearest living ancestor's wb in order
 717 * to eventually release the dying @wb.  Returns %true if not all inodes were
 718 * switched and the function has to be restarted.
 719 */
 720bool cleanup_offline_cgwb(struct bdi_writeback *wb)
 721{
 722	struct cgroup_subsys_state *memcg_css;
 723	struct inode_switch_wbs_context *isw;
 724	struct bdi_writeback *new_wb;
 725	int nr;
 726	bool restart = false;
 727
 728	isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW),
 729		      GFP_KERNEL);
 730	if (!isw)
 731		return restart;
 732
 733	atomic_inc(&isw_nr_in_flight);
 734
 735	for (memcg_css = wb->memcg_css->parent; memcg_css;
 736	     memcg_css = memcg_css->parent) {
 737		new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL);
 738		if (new_wb)
 739			break;
 740	}
 741	if (unlikely(!new_wb))
 742		new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
 743
 744	nr = 0;
 745	spin_lock(&wb->list_lock);
 746	/*
 747	 * In addition to the inodes that have completed writeback, also switch
 748	 * cgwbs for those inodes only with dirty timestamps. Otherwise, those
 749	 * inodes won't be written back for a long time when lazytime is
 750	 * enabled, and thus pinning the dying cgwbs. It won't break the
 751	 * bandwidth restrictions, as writeback of inode metadata is not
 752	 * accounted for.
 753	 */
 754	restart = isw_prepare_wbs_switch(new_wb, isw, &wb->b_attached, &nr);
 755	if (!restart)
 756		restart = isw_prepare_wbs_switch(new_wb, isw, &wb->b_dirty_time,
 757						 &nr);
 758	spin_unlock(&wb->list_lock);
 759
 760	/* no attached inodes? bail out */
 761	if (nr == 0) {
 762		atomic_dec(&isw_nr_in_flight);
 763		wb_put(new_wb);
 764		kfree(isw);
 765		return restart;
 766	}
 767
 768	trace_inode_switch_wbs_queue(wb, new_wb, nr);
 769	wb_queue_isw(new_wb, isw);
 770
 771	return restart;
 772}
 773
 774/**
 775 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
 776 * @wbc: writeback_control of interest
 777 * @inode: target inode
 778 *
 779 * @inode is locked and about to be written back under the control of @wbc.
 780 * Record @inode's writeback context into @wbc and unlock the i_lock.  On
 781 * writeback completion, wbc_detach_inode() should be called.  This is used
 782 * to track the cgroup writeback context.
 783 */
 784static void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
 785		struct inode *inode)
 786	__releases(&inode->i_lock)
 787{
 788	if (!inode_cgwb_enabled(inode)) {
 789		spin_unlock(&inode->i_lock);
 790		return;
 791	}
 792
 793	wbc->wb = inode_to_wb(inode);
 794	wbc->inode = inode;
 795
 796	wbc->wb_id = wbc->wb->memcg_css->id;
 797	wbc->wb_lcand_id = inode->i_wb_frn_winner;
 798	wbc->wb_tcand_id = 0;
 799	wbc->wb_bytes = 0;
 800	wbc->wb_lcand_bytes = 0;
 801	wbc->wb_tcand_bytes = 0;
 802
 803	wb_get(wbc->wb);
 804	spin_unlock(&inode->i_lock);
 805
 806	/*
 807	 * A dying wb indicates that either the blkcg associated with the
 808	 * memcg changed or the associated memcg is dying.  In the first
 809	 * case, a replacement wb should already be available and we should
 810	 * refresh the wb immediately.  In the second case, trying to
 811	 * refresh will keep failing.
 812	 */
 813	if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
 814		inode_switch_wbs(inode, wbc->wb_id);
 815}
 816
 817/**
 818 * wbc_attach_fdatawrite_inode - associate wbc and inode for fdatawrite
 819 * @wbc: writeback_control of interest
 820 * @inode: target inode
 821 *
 822 * This function is to be used by filemap_writeback(), which is an alternative
 823 * entry point into writeback code, and first ensures @inode is associated with
 824 * a bdi_writeback and attaches it to @wbc.
 825 */
 826void wbc_attach_fdatawrite_inode(struct writeback_control *wbc,
 827		struct inode *inode)
 828{
 829	spin_lock(&inode->i_lock);
 830	inode_attach_wb(inode, NULL);
 831	wbc_attach_and_unlock_inode(wbc, inode);
 832}
 833EXPORT_SYMBOL_GPL(wbc_attach_fdatawrite_inode);
 834
 835/**
 836 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
 837 * @wbc: writeback_control of the just finished writeback
 838 *
 839 * To be called after a writeback attempt of an inode finishes and undoes
 840 * wbc_attach_and_unlock_inode().  Can be called under any context.
 841 *
 842 * As concurrent write sharing of an inode is expected to be very rare and
 843 * memcg only tracks page ownership on first-use basis severely confining
 844 * the usefulness of such sharing, cgroup writeback tracks ownership
 845 * per-inode.  While the support for concurrent write sharing of an inode
 846 * is deemed unnecessary, an inode being written to by different cgroups at
 847 * different points in time is a lot more common, and, more importantly,
 848 * charging only by first-use can too readily lead to grossly incorrect
 849 * behaviors (single foreign page can lead to gigabytes of writeback to be
 850 * incorrectly attributed).
 851 *
 852 * To resolve this issue, cgroup writeback detects the majority dirtier of
 853 * an inode and transfers the ownership to it.  To avoid unnecessary
 854 * oscillation, the detection mechanism keeps track of history and gives
 855 * out the switch verdict only if the foreign usage pattern is stable over
 856 * a certain amount of time and/or writeback attempts.
 857 *
 858 * On each writeback attempt, @wbc tries to detect the majority writer
 859 * using Boyer-Moore majority vote algorithm.  In addition to the byte
 860 * count from the majority voting, it also counts the bytes written for the
 861 * current wb and the last round's winner wb (max of last round's current
 862 * wb, the winner from two rounds ago, and the last round's majority
 863 * candidate).  Keeping track of the historical winner helps the algorithm
 864 * to semi-reliably detect the most active writer even when it's not the
 865 * absolute majority.
 866 *
 867 * Once the winner of the round is determined, whether the winner is
 868 * foreign or not and how much IO time the round consumed is recorded in
 869 * inode->i_wb_frn_history.  If the amount of recorded foreign IO time is
 870 * over a certain threshold, the switch verdict is given.
 871 */
 872void wbc_detach_inode(struct writeback_control *wbc)
 873{
 874	struct bdi_writeback *wb = wbc->wb;
 875	struct inode *inode = wbc->inode;
 876	unsigned long avg_time, max_bytes, max_time;
 877	u16 history;
 878	int max_id;
 879
 880	if (!wb)
 881		return;
 882
 883	history = inode->i_wb_frn_history;
 884	avg_time = inode->i_wb_frn_avg_time;
 885
 886	/* pick the winner of this round */
 887	if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
 888	    wbc->wb_bytes >= wbc->wb_tcand_bytes) {
 889		max_id = wbc->wb_id;
 890		max_bytes = wbc->wb_bytes;
 891	} else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
 892		max_id = wbc->wb_lcand_id;
 893		max_bytes = wbc->wb_lcand_bytes;
 894	} else {
 895		max_id = wbc->wb_tcand_id;
 896		max_bytes = wbc->wb_tcand_bytes;
 897	}
 898
 899	/*
 900	 * Calculate the amount of IO time the winner consumed and fold it
 901	 * into the running average kept per inode.  If the consumed IO
 902	 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
 903	 * deciding whether to switch or not.  This is to prevent one-off
 904	 * small dirtiers from skewing the verdict.
 905	 */
 906	max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
 907				wb->avg_write_bandwidth);
 908	if (avg_time)
 909		avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
 910			    (avg_time >> WB_FRN_TIME_AVG_SHIFT);
 911	else
 912		avg_time = max_time;	/* immediate catch up on first run */
 913
 914	if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
 915		int slots;
 916
 917		/*
 918		 * The switch verdict is reached if foreign wb's consume
 919		 * more than a certain proportion of IO time in a
 920		 * WB_FRN_TIME_PERIOD.  This is loosely tracked by 16 slot
 921		 * history mask where each bit represents one sixteenth of
 922		 * the period.  Determine the number of slots to shift into
 923		 * history from @max_time.
 924		 */
 925		slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
 926			    (unsigned long)WB_FRN_HIST_MAX_SLOTS);
 927		history <<= slots;
 928		if (wbc->wb_id != max_id)
 929			history |= (1U << slots) - 1;
 930
 931		if (history)
 932			trace_inode_foreign_history(inode, wbc, history);
 933
 934		/*
 935		 * Switch if the current wb isn't the consistent winner.
 936		 * If there are multiple closely competing dirtiers, the
 937		 * inode may switch across them repeatedly over time, which
 938		 * is okay.  The main goal is avoiding keeping an inode on
 939		 * the wrong wb for an extended period of time.
 940		 */
 941		if (hweight16(history) > WB_FRN_HIST_THR_SLOTS)
 942			inode_switch_wbs(inode, max_id);
 943	}
 944
 945	/*
 946	 * Multiple instances of this function may race to update the
 947	 * following fields but we don't mind occassional inaccuracies.
 948	 */
 949	inode->i_wb_frn_winner = max_id;
 950	inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
 951	inode->i_wb_frn_history = history;
 952
 953	wb_put(wbc->wb);
 954	wbc->wb = NULL;
 955}
 956EXPORT_SYMBOL_GPL(wbc_detach_inode);
 957
 958/**
 959 * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
 960 * @wbc: writeback_control of the writeback in progress
 961 * @folio: folio being written out
 962 * @bytes: number of bytes being written out
 963 *
 964 * @bytes from @folio are about to written out during the writeback
 965 * controlled by @wbc.  Keep the book for foreign inode detection.  See
 966 * wbc_detach_inode().
 967 */
 968void wbc_account_cgroup_owner(struct writeback_control *wbc, struct folio *folio,
 969			      size_t bytes)
 970{
 971	struct cgroup_subsys_state *css;
 972	int id;
 973
 974	/*
 975	 * pageout() path doesn't attach @wbc to the inode being written
 976	 * out.  This is intentional as we don't want the function to block
 977	 * behind a slow cgroup.  Ultimately, we want pageout() to kick off
 978	 * regular writeback instead of writing things out itself.
 979	 */
 980	if (!wbc->wb || wbc->no_cgroup_owner)
 981		return;
 982
 983	css = mem_cgroup_css_from_folio(folio);
 984	/* dead cgroups shouldn't contribute to inode ownership arbitration */
 985	if (!(css->flags & CSS_ONLINE))
 986		return;
 987
 988	id = css->id;
 989
 990	if (id == wbc->wb_id) {
 991		wbc->wb_bytes += bytes;
 992		return;
 993	}
 994
 995	if (id == wbc->wb_lcand_id)
 996		wbc->wb_lcand_bytes += bytes;
 997
 998	/* Boyer-Moore majority vote algorithm */
 999	if (!wbc->wb_tcand_bytes)
1000		wbc->wb_tcand_id = id;
1001	if (id == wbc->wb_tcand_id)
1002		wbc->wb_tcand_bytes += bytes;
1003	else
1004		wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
1005}
1006EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
1007
1008/**
1009 * wb_split_bdi_pages - split nr_pages to write according to bandwidth
1010 * @wb: target bdi_writeback to split @nr_pages to
1011 * @nr_pages: number of pages to write for the whole bdi
1012 *
1013 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
1014 * relation to the total write bandwidth of all wb's w/ dirty inodes on
1015 * @wb->bdi.
1016 */
1017static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1018{
1019	unsigned long this_bw = wb->avg_write_bandwidth;
1020	unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
1021
1022	if (nr_pages == LONG_MAX)
1023		return LONG_MAX;
1024
1025	/*
1026	 * This may be called on clean wb's and proportional distribution
1027	 * may not make sense, just use the original @nr_pages in those
1028	 * cases.  In general, we wanna err on the side of writing more.
1029	 */
1030	if (!tot_bw || this_bw >= tot_bw)
1031		return nr_pages;
1032	else
1033		return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
1034}
1035
1036/**
1037 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
1038 * @bdi: target backing_dev_info
1039 * @base_work: wb_writeback_work to issue
1040 * @skip_if_busy: skip wb's which already have writeback in progress
1041 *
1042 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
1043 * have dirty inodes.  If @base_work->nr_page isn't %LONG_MAX, it's
1044 * distributed to the busy wbs according to each wb's proportion in the
1045 * total active write bandwidth of @bdi.
1046 */
1047static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1048				  struct wb_writeback_work *base_work,
1049				  bool skip_if_busy)
1050{
1051	struct bdi_writeback *last_wb = NULL;
1052	struct bdi_writeback *wb = list_entry(&bdi->wb_list,
1053					      struct bdi_writeback, bdi_node);
1054
1055	might_sleep();
1056restart:
1057	rcu_read_lock();
1058	list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
1059		DEFINE_WB_COMPLETION(fallback_work_done, bdi);
1060		struct wb_writeback_work fallback_work;
1061		struct wb_writeback_work *work;
1062		long nr_pages;
1063
1064		if (last_wb) {
1065			wb_put(last_wb);
1066			last_wb = NULL;
1067		}
1068
1069		/* SYNC_ALL writes out I_DIRTY_TIME too */
1070		if (!wb_has_dirty_io(wb) &&
1071		    (base_work->sync_mode == WB_SYNC_NONE ||
1072		     list_empty(&wb->b_dirty_time)))
1073			continue;
1074		if (skip_if_busy && writeback_in_progress(wb))
1075			continue;
1076
1077		nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
1078
1079		work = kmalloc(sizeof(*work), GFP_ATOMIC);
1080		if (work) {
1081			*work = *base_work;
1082			work->nr_pages = nr_pages;
1083			work->auto_free = 1;
1084			wb_queue_work(wb, work);
1085			continue;
1086		}
1087
1088		/*
1089		 * If wb_tryget fails, the wb has been shutdown, skip it.
1090		 *
1091		 * Pin @wb so that it stays on @bdi->wb_list.  This allows
1092		 * continuing iteration from @wb after dropping and
1093		 * regrabbing rcu read lock.
1094		 */
1095		if (!wb_tryget(wb))
1096			continue;
1097
1098		/* alloc failed, execute synchronously using on-stack fallback */
1099		work = &fallback_work;
1100		*work = *base_work;
1101		work->nr_pages = nr_pages;
1102		work->auto_free = 0;
1103		work->done = &fallback_work_done;
1104
1105		wb_queue_work(wb, work);
1106		last_wb = wb;
1107
1108		rcu_read_unlock();
1109		wb_wait_for_completion(&fallback_work_done);
1110		goto restart;
1111	}
1112	rcu_read_unlock();
1113
1114	if (last_wb)
1115		wb_put(last_wb);
1116}
1117
1118/**
1119 * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
1120 * @bdi_id: target bdi id
1121 * @memcg_id: target memcg css id
1122 * @reason: reason why some writeback work initiated
1123 * @done: target wb_completion
1124 *
1125 * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
1126 * with the specified parameters.
1127 */
1128int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
1129			   enum wb_reason reason, struct wb_completion *done)
1130{
1131	struct backing_dev_info *bdi;
1132	struct cgroup_subsys_state *memcg_css;
1133	struct bdi_writeback *wb;
1134	struct wb_writeback_work *work;
1135	unsigned long dirty;
1136	int ret;
1137
1138	/* lookup bdi and memcg */
1139	bdi = bdi_get_by_id(bdi_id);
1140	if (!bdi)
1141		return -ENOENT;
1142
1143	rcu_read_lock();
1144	memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
1145	if (memcg_css && !css_tryget(memcg_css))
1146		memcg_css = NULL;
1147	rcu_read_unlock();
1148	if (!memcg_css) {
1149		ret = -ENOENT;
1150		goto out_bdi_put;
1151	}
1152
1153	/*
1154	 * And find the associated wb.  If the wb isn't there already
1155	 * there's nothing to flush, don't create one.
1156	 */
1157	wb = wb_get_lookup(bdi, memcg_css);
1158	if (!wb) {
1159		ret = -ENOENT;
1160		goto out_css_put;
1161	}
1162
1163	/*
1164	 * The caller is attempting to write out most of
1165	 * the currently dirty pages.  Let's take the current dirty page
1166	 * count and inflate it by 25% which should be large enough to
1167	 * flush out most dirty pages while avoiding getting livelocked by
1168	 * concurrent dirtiers.
1169	 *
1170	 * BTW the memcg stats are flushed periodically and this is best-effort
1171	 * estimation, so some potential error is ok.
1172	 */
1173	dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY);
1174	dirty = dirty * 10 / 8;
1175
1176	/* issue the writeback work */
1177	work = kzalloc(sizeof(*work), GFP_NOWAIT);
1178	if (work) {
1179		work->nr_pages = dirty;
1180		work->sync_mode = WB_SYNC_NONE;
1181		work->range_cyclic = 1;
1182		work->reason = reason;
1183		work->done = done;
1184		work->auto_free = 1;
1185		wb_queue_work(wb, work);
1186		ret = 0;
1187	} else {
1188		ret = -ENOMEM;
1189	}
1190
1191	wb_put(wb);
1192out_css_put:
1193	css_put(memcg_css);
1194out_bdi_put:
1195	bdi_put(bdi);
1196	return ret;
1197}
1198
1199/**
1200 * cgroup_writeback_umount - flush inode wb switches for umount
1201 * @sb: target super_block
1202 *
1203 * This function is called when a super_block is about to be destroyed and
1204 * flushes in-flight inode wb switches.  An inode wb switch goes through
1205 * RCU and then workqueue, so the two need to be flushed in order to ensure
1206 * that all previously scheduled switches are finished.  As wb switches are
1207 * rare occurrences and synchronize_rcu() can take a while, perform
1208 * flushing iff wb switches are in flight.
1209 */
1210void cgroup_writeback_umount(struct super_block *sb)
1211{
1212
1213	if (!(sb->s_bdi->capabilities & BDI_CAP_WRITEBACK))
1214		return;
1215
1216	/*
1217	 * SB_ACTIVE should be reliably cleared before checking
1218	 * isw_nr_in_flight, see generic_shutdown_super().
1219	 */
1220	smp_mb();
1221
1222	if (atomic_read(&isw_nr_in_flight)) {
1223		/*
1224		 * Use rcu_barrier() to wait for all pending callbacks to
1225		 * ensure that all in-flight wb switches are in the workqueue.
1226		 */
1227		rcu_barrier();
1228		flush_workqueue(isw_wq);
1229	}
1230}
1231
1232static int __init cgroup_writeback_init(void)
1233{
1234	isw_wq = alloc_workqueue("inode_switch_wbs", WQ_PERCPU, 0);
1235	if (!isw_wq)
1236		return -ENOMEM;
1237	return 0;
1238}
1239fs_initcall(cgroup_writeback_init);
1240
1241#else	/* CONFIG_CGROUP_WRITEBACK */
1242
1243static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1244static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
1245
1246static void inode_cgwb_move_to_attached(struct inode *inode,
1247					struct bdi_writeback *wb)
1248{
1249	assert_spin_locked(&wb->list_lock);
1250	assert_spin_locked(&inode->i_lock);
1251	WARN_ON_ONCE(inode_state_read(inode) & I_FREEING);
1252
1253	inode_state_clear(inode, I_SYNC_QUEUED);
1254	list_del_init(&inode->i_io_list);
1255	wb_io_lists_depopulated(wb);
1256}
1257
1258static struct bdi_writeback *
1259locked_inode_to_wb_and_lock_list(struct inode *inode)
1260	__releases(&inode->i_lock)
1261	__acquires(&wb->list_lock)
1262{
1263	struct bdi_writeback *wb = inode_to_wb(inode);
1264
1265	spin_unlock(&inode->i_lock);
1266	spin_lock(&wb->list_lock);
1267	return wb;
1268}
1269
1270static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
1271	__acquires(&wb->list_lock)
1272{
1273	struct bdi_writeback *wb = inode_to_wb(inode);
1274
1275	spin_lock(&wb->list_lock);
1276	return wb;
1277}
1278
1279static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
1280{
1281	return nr_pages;
1282}
1283
1284static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
1285				  struct wb_writeback_work *base_work,
1286				  bool skip_if_busy)
1287{
1288	might_sleep();
1289
1290	if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
1291		base_work->auto_free = 0;
1292		wb_queue_work(&bdi->wb, base_work);
1293	}
1294}
1295
1296static inline void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
1297					       struct inode *inode)
1298	__releases(&inode->i_lock)
1299{
1300	spin_unlock(&inode->i_lock);
1301}
1302
1303#endif	/* CONFIG_CGROUP_WRITEBACK */
1304
1305/*
1306 * Add in the number of potentially dirty inodes, because each inode
1307 * write can dirty pagecache in the underlying blockdev.
1308 */
1309static unsigned long get_nr_dirty_pages(void)
1310{
1311	return global_node_page_state(NR_FILE_DIRTY) +
1312		get_nr_dirty_inodes();
1313}
1314
1315static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
1316{
1317	if (!wb_has_dirty_io(wb))
1318		return;
1319
1320	/*
1321	 * All callers of this function want to start writeback of all
1322	 * dirty pages. Places like vmscan can call this at a very
1323	 * high frequency, causing pointless allocations of tons of
1324	 * work items and keeping the flusher threads busy retrieving
1325	 * that work. Ensure that we only allow one of them pending and
1326	 * inflight at the time.
1327	 */
1328	if (test_bit(WB_start_all, &wb->state) ||
1329	    test_and_set_bit(WB_start_all, &wb->state))
1330		return;
1331
1332	wb->start_all_reason = reason;
1333	wb_wakeup(wb);
1334}
1335
1336/**
1337 * wb_start_background_writeback - start background writeback
1338 * @wb: bdi_writback to write from
1339 *
1340 * Description:
1341 *   This makes sure WB_SYNC_NONE background writeback happens. When
1342 *   this function returns, it is only guaranteed that for given wb
1343 *   some IO is happening if we are over background dirty threshold.
1344 *   Caller need not hold sb s_umount semaphore.
1345 */
1346void wb_start_background_writeback(struct bdi_writeback *wb)
1347{
1348	/*
1349	 * We just wake up the flusher thread. It will perform background
1350	 * writeback as soon as there is no other work to do.
1351	 */
1352	trace_writeback_wake_background(wb);
1353	wb_wakeup(wb);
1354}
1355
1356/*
1357 * Remove the inode from the writeback list it is on.
1358 */
1359void inode_io_list_del(struct inode *inode)
1360{
1361	struct bdi_writeback *wb;
1362
1363	/*
1364	 * FIXME: ext4 can call here from ext4_evict_inode() after evict() already
1365	 * unlinked the inode.
1366	 */
1367	if (list_empty_careful(&inode->i_io_list))
1368		return;
1369
1370	wb = inode_to_wb_and_lock_list(inode);
1371	spin_lock(&inode->i_lock);
1372
1373	inode_state_clear(inode, I_SYNC_QUEUED);
1374	list_del_init(&inode->i_io_list);
1375	wb_io_lists_depopulated(wb);
1376
1377	spin_unlock(&inode->i_lock);
1378	spin_unlock(&wb->list_lock);
1379}
1380EXPORT_SYMBOL(inode_io_list_del);
1381
1382/*
1383 * mark an inode as under writeback on the sb
1384 */
1385void sb_mark_inode_writeback(struct inode *inode)
1386{
1387	struct super_block *sb = inode->i_sb;
1388	unsigned long flags;
1389
1390	if (list_empty(&inode->i_wb_list)) {
1391		spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1392		if (list_empty(&inode->i_wb_list)) {
1393			list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
1394			trace_sb_mark_inode_writeback(inode);
1395		}
1396		spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1397	}
1398}
1399
1400/*
1401 * clear an inode as under writeback on the sb
1402 */
1403void sb_clear_inode_writeback(struct inode *inode)
1404{
1405	struct super_block *sb = inode->i_sb;
1406	unsigned long flags;
1407
1408	if (!list_empty(&inode->i_wb_list)) {
1409		spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1410		if (!list_empty(&inode->i_wb_list)) {
1411			list_del_init(&inode->i_wb_list);
1412			trace_sb_clear_inode_writeback(inode);
1413		}
1414		spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1415	}
1416}
1417
1418/*
1419 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1420 * furthest end of its superblock's dirty-inode list.
1421 *
1422 * Before stamping the inode's ->dirtied_when, we check to see whether it is
1423 * already the most-recently-dirtied inode on the b_dirty list.  If that is
1424 * the case then the inode must have been redirtied while it was being written
1425 * out and we don't reset its dirtied_when.
1426 */
1427static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
1428{
1429	assert_spin_locked(&inode->i_lock);
1430
1431	inode_state_clear(inode, I_SYNC_QUEUED);
1432	/*
1433	 * When the inode is being freed just don't bother with dirty list
1434	 * tracking. Flush worker will ignore this inode anyway and it will
1435	 * trigger assertions in inode_io_list_move_locked().
1436	 */
1437	if (inode_state_read(inode) & I_FREEING) {
1438		list_del_init(&inode->i_io_list);
1439		wb_io_lists_depopulated(wb);
1440		return;
1441	}
1442	if (!list_empty(&wb->b_dirty)) {
1443		struct inode *tail;
1444
1445		tail = wb_inode(wb->b_dirty.next);
1446		if (time_before(inode->dirtied_when, tail->dirtied_when))
1447			inode->dirtied_when = jiffies;
1448	}
1449	inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1450}
1451
1452static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1453{
1454	spin_lock(&inode->i_lock);
1455	redirty_tail_locked(inode, wb);
1456	spin_unlock(&inode->i_lock);
1457}
1458
1459/*
1460 * requeue inode for re-scanning after bdi->b_io list is exhausted.
1461 */
1462static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1463{
1464	inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1465}
1466
1467static void inode_sync_complete(struct inode *inode)
1468{
1469	assert_spin_locked(&inode->i_lock);
1470
1471	inode_state_clear(inode, I_SYNC);
1472	/* If inode is clean an unused, put it into LRU now... */
1473	inode_lru_list_add(inode);
1474	/* Called with inode->i_lock which ensures memory ordering. */
1475	inode_wake_up_bit(inode, __I_SYNC);
1476}
1477
1478static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1479{
1480	bool ret = time_after(inode->dirtied_when, t);
1481#ifndef CONFIG_64BIT
1482	/*
1483	 * For inodes being constantly redirtied, dirtied_when can get stuck.
1484	 * It _appears_ to be in the future, but is actually in distant past.
1485	 * This test is necessary to prevent such wrapped-around relative times
1486	 * from permanently stopping the whole bdi writeback.
1487	 */
1488	ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1489#endif
1490	return ret;
1491}
1492
1493/*
1494 * Move expired (dirtied before dirtied_before) dirty inodes from
1495 * @delaying_queue to @dispatch_queue.
1496 */
1497static int move_expired_inodes(struct list_head *delaying_queue,
1498			       struct list_head *dispatch_queue,
1499			       unsigned long dirtied_before)
1500{
1501	LIST_HEAD(tmp);
1502	struct list_head *pos, *node;
1503	struct super_block *sb = NULL;
1504	struct inode *inode;
1505	int do_sb_sort = 0;
1506	int moved = 0;
1507
1508	while (!list_empty(delaying_queue)) {
1509		inode = wb_inode(delaying_queue->prev);
1510		if (inode_dirtied_after(inode, dirtied_before))
1511			break;
1512		spin_lock(&inode->i_lock);
1513		list_move(&inode->i_io_list, &tmp);
1514		moved++;
1515		inode_state_set(inode, I_SYNC_QUEUED);
1516		spin_unlock(&inode->i_lock);
1517		if (sb_is_blkdev_sb(inode->i_sb))
1518			continue;
1519		if (sb && sb != inode->i_sb)
1520			do_sb_sort = 1;
1521		sb = inode->i_sb;
1522	}
1523
1524	/* just one sb in list, splice to dispatch_queue and we're done */
1525	if (!do_sb_sort) {
1526		list_splice(&tmp, dispatch_queue);
1527		goto out;
1528	}
1529
1530	/*
1531	 * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue',
1532	 * we don't take inode->i_lock here because it is just a pointless overhead.
1533	 * Inode is already marked as I_SYNC_QUEUED so writeback list handling is
1534	 * fully under our control.
1535	 */
1536	while (!list_empty(&tmp)) {
1537		sb = wb_inode(tmp.prev)->i_sb;
1538		list_for_each_prev_safe(pos, node, &tmp) {
1539			inode = wb_inode(pos);
1540			if (inode->i_sb == sb)
1541				list_move(&inode->i_io_list, dispatch_queue);
1542		}
1543	}
1544out:
1545	return moved;
1546}
1547
1548/*
1549 * Queue all expired dirty inodes for io, eldest first.
1550 * Before
1551 *         newly dirtied     b_dirty    b_io    b_more_io
1552 *         =============>    gf         edc     BA
1553 * After
1554 *         newly dirtied     b_dirty    b_io    b_more_io
1555 *         =============>    g          fBAedc
1556 *                                           |
1557 *                                           +--> dequeue for IO
1558 */
1559static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
1560		     unsigned long dirtied_before)
1561{
1562	int moved;
1563	unsigned long time_expire_jif = dirtied_before;
1564
1565	assert_spin_locked(&wb->list_lock);
1566	list_splice_init(&wb->b_more_io, &wb->b_io);
1567	moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
1568	if (!work->for_sync)
1569		time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
1570	moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1571				     time_expire_jif);
1572	if (moved)
1573		wb_io_lists_populated(wb);
1574	trace_writeback_queue_io(wb, work, dirtied_before, moved);
1575}
1576
1577static int write_inode(struct inode *inode, struct writeback_control *wbc)
1578{
1579	int ret;
1580
1581	if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1582		trace_writeback_write_inode_start(inode, wbc);
1583		ret = inode->i_sb->s_op->write_inode(inode, wbc);
1584		trace_writeback_write_inode(inode, wbc);
1585		return ret;
1586	}
1587	return 0;
1588}
1589
1590/*
1591 * Wait for writeback on an inode to complete. Called with i_lock held.
1592 * Caller must make sure inode cannot go away when we drop i_lock.
1593 */
1594void inode_wait_for_writeback(struct inode *inode)
1595{
1596	struct wait_bit_queue_entry wqe;
1597	struct wait_queue_head *wq_head;
1598
1599	assert_spin_locked(&inode->i_lock);
1600
1601	if (!(inode_state_read(inode) & I_SYNC))
1602		return;
1603
1604	wq_head = inode_bit_waitqueue(&wqe, inode, __I_SYNC);
1605	for (;;) {
1606		prepare_to_wait_event(wq_head, &wqe.wq_entry, TASK_UNINTERRUPTIBLE);
1607		/* Checking I_SYNC with inode->i_lock guarantees memory ordering. */
1608		if (!(inode_state_read(inode) & I_SYNC))
1609			break;
1610		spin_unlock(&inode->i_lock);
1611		schedule();
1612		spin_lock(&inode->i_lock);
1613	}
1614	finish_wait(wq_head, &wqe.wq_entry);
1615}
1616
1617/*
1618 * Sleep until I_SYNC is cleared. This function must be called with i_lock
1619 * held and drops it. It is aimed for callers not holding any inode reference
1620 * so once i_lock is dropped, inode can go away.
1621 */
1622static void inode_sleep_on_writeback(struct inode *inode)
1623	__releases(inode->i_lock)
1624{
1625	struct wait_bit_queue_entry wqe;
1626	struct wait_queue_head *wq_head;
1627	bool sleep;
1628
1629	assert_spin_locked(&inode->i_lock);
1630
1631	wq_head = inode_bit_waitqueue(&wqe, inode, __I_SYNC);
1632	prepare_to_wait_event(wq_head, &wqe.wq_entry, TASK_UNINTERRUPTIBLE);
1633	/* Checking I_SYNC with inode->i_lock guarantees memory ordering. */
1634	sleep = !!(inode_state_read(inode) & I_SYNC);
1635	spin_unlock(&inode->i_lock);
1636	if (sleep)
1637		schedule();
1638	finish_wait(wq_head, &wqe.wq_entry);
1639}
1640
1641/*
1642 * Find proper writeback list for the inode depending on its current state and
1643 * possibly also change of its state while we were doing writeback.  Here we
1644 * handle things such as livelock prevention or fairness of writeback among
1645 * inodes. This function can be called only by flusher thread - noone else
1646 * processes all inodes in writeback lists and requeueing inodes behind flusher
1647 * thread's back can have unexpected consequences.
1648 */
1649static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1650			  struct writeback_control *wbc,
1651			  unsigned long dirtied_before)
1652{
1653	if (inode_state_read(inode) & I_FREEING)
1654		return;
1655
1656	/*
1657	 * Sync livelock prevention. Each inode is tagged and synced in one
1658	 * shot. If still dirty, it will be redirty_tail()'ed below.  Update
1659	 * the dirty time to prevent enqueue and sync it again.
1660	 */
1661	if ((inode_state_read(inode) & I_DIRTY) &&
1662	    (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1663		inode->dirtied_when = jiffies;
1664
1665	if (wbc->pages_skipped) {
1666		/*
1667		 * Writeback is not making progress due to locked buffers.
1668		 * Skip this inode for now. Although having skipped pages
1669		 * is odd for clean inodes, it can happen for some
1670		 * filesystems so handle that gracefully.
1671		 */
1672		if (inode_state_read(inode) & I_DIRTY_ALL)
1673			redirty_tail_locked(inode, wb);
1674		else
1675			inode_cgwb_move_to_attached(inode, wb);
1676		return;
1677	}
1678
1679	if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1680		/*
1681		 * We didn't write back all the pages.  nfs_writepages()
1682		 * sometimes bales out without doing anything.
1683		 */
1684		if (wbc->nr_to_write <= 0 &&
1685		    !inode_dirtied_after(inode, dirtied_before)) {
1686			/* Slice used up. Queue for next turn. */
1687			requeue_io(inode, wb);
1688		} else {
1689			/*
1690			 * Writeback blocked by something other than
1691			 * congestion. Delay the inode for some time to
1692			 * avoid spinning on the CPU (100% iowait)
1693			 * retrying writeback of the dirty page/inode
1694			 * that cannot be performed immediately.
1695			 */
1696			redirty_tail_locked(inode, wb);
1697		}
1698	} else if (inode_state_read(inode) & I_DIRTY) {
1699		/*
1700		 * Filesystems can dirty the inode during writeback operations,
1701		 * such as delayed allocation during submission or metadata
1702		 * updates after data IO completion.
1703		 */
1704		redirty_tail_locked(inode, wb);
1705	} else if (inode_state_read(inode) & I_DIRTY_TIME) {
1706		inode->dirtied_when = jiffies;
1707		inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1708		inode_state_clear(inode, I_SYNC_QUEUED);
1709	} else {
1710		/* The inode is clean. Remove from writeback lists. */
1711		inode_cgwb_move_to_attached(inode, wb);
1712	}
1713}
1714
1715/*
1716 * Write out an inode and its dirty pages (or some of its dirty pages, depending
1717 * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state.
1718 *
1719 * This doesn't remove the inode from the writeback list it is on, except
1720 * potentially to move it from b_dirty_time to b_dirty due to timestamp
1721 * expiration.  The caller is otherwise responsible for writeback list handling.
1722 *
1723 * The caller is also responsible for setting the I_SYNC flag beforehand and
1724 * calling inode_sync_complete() to clear it afterwards.
1725 */
1726static int
1727__writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1728{
1729	struct address_space *mapping = inode->i_mapping;
1730	long nr_to_write = wbc->nr_to_write;
1731	unsigned dirty;
1732	int ret;
1733
1734	WARN_ON(!(inode_state_read_once(inode) & I_SYNC));
1735
1736	trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1737
1738	ret = do_writepages(mapping, wbc);
1739
1740	/*
1741	 * Make sure to wait on the data before writing out the metadata.
1742	 * This is important for filesystems that modify metadata on data
1743	 * I/O completion. We don't do it for sync(2) writeback because it has a
1744	 * separate, external IO completion path and ->sync_fs for guaranteeing
1745	 * inode metadata is written back correctly.
1746	 */
1747	if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1748		int err = filemap_fdatawait(mapping);
1749		if (ret == 0)
1750			ret = err;
1751	}
1752
1753	/*
1754	 * If the inode has dirty timestamps and we need to write them, call
1755	 * mark_inode_dirty_sync() to notify the filesystem about it and to
1756	 * change I_DIRTY_TIME into I_DIRTY_SYNC.
1757	 */
1758	if ((inode_state_read_once(inode) & I_DIRTY_TIME) &&
1759	    (wbc->sync_mode == WB_SYNC_ALL ||
1760	     time_after(jiffies, inode->dirtied_time_when +
1761			dirtytime_expire_interval * HZ))) {
1762		trace_writeback_lazytime(inode);
1763		mark_inode_dirty_sync(inode);
1764	}
1765
1766	/*
1767	 * Get and clear the dirty flags from i_state.  This needs to be done
1768	 * after calling writepages because some filesystems may redirty the
1769	 * inode during writepages due to delalloc.  It also needs to be done
1770	 * after handling timestamp expiration, as that may dirty the inode too.
1771	 */
1772	spin_lock(&inode->i_lock);
1773	dirty = inode_state_read(inode) & I_DIRTY;
1774	inode_state_clear(inode, dirty);
1775
1776	/*
1777	 * Paired with smp_mb() in __mark_inode_dirty().  This allows
1778	 * __mark_inode_dirty() to test i_state without grabbing i_lock -
1779	 * either they see the I_DIRTY bits cleared or we see the dirtied
1780	 * inode.
1781	 *
1782	 * I_DIRTY_PAGES is always cleared together above even if @mapping
1783	 * still has dirty pages.  The flag is reinstated after smp_mb() if
1784	 * necessary.  This guarantees that either __mark_inode_dirty()
1785	 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1786	 */
1787	smp_mb();
1788
1789	if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1790		inode_state_set(inode, I_DIRTY_PAGES);
1791	else if (unlikely(inode_state_read(inode) & I_PINNING_NETFS_WB)) {
1792		if (!(inode_state_read(inode) & I_DIRTY_PAGES)) {
1793			inode_state_clear(inode, I_PINNING_NETFS_WB);
1794			wbc->unpinned_netfs_wb = true;
1795			dirty |= I_PINNING_NETFS_WB; /* Cause write_inode */
1796		}
1797	}
1798
1799	spin_unlock(&inode->i_lock);
1800
1801	/* Don't write the inode if only I_DIRTY_PAGES was set */
1802	if (dirty & ~I_DIRTY_PAGES) {
1803		int err = write_inode(inode, wbc);
1804		if (ret == 0)
1805			ret = err;
1806	}
1807	wbc->unpinned_netfs_wb = false;
1808	trace_writeback_single_inode(inode, wbc, nr_to_write);
1809	return ret;
1810}
1811
1812/*
1813 * Write out an inode's dirty data and metadata on-demand, i.e. separately from
1814 * the regular batched writeback done by the flusher threads in
1815 * writeback_sb_inodes().  @wbc controls various aspects of the write, such as
1816 * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE).
1817 *
1818 * To prevent the inode from going away, either the caller must have a reference
1819 * to the inode, or the inode must have I_WILL_FREE or I_FREEING set.
1820 */
1821static int writeback_single_inode(struct inode *inode,
1822				  struct writeback_control *wbc)
1823{
1824	struct bdi_writeback *wb;
1825	int ret = 0;
1826
1827	spin_lock(&inode->i_lock);
1828	if (!icount_read(inode))
1829		WARN_ON(!(inode_state_read(inode) & (I_WILL_FREE | I_FREEING)));
1830	else
1831		WARN_ON(inode_state_read(inode) & I_WILL_FREE);
1832
1833	if (inode_state_read(inode) & I_SYNC) {
1834		/*
1835		 * Writeback is already running on the inode.  For WB_SYNC_NONE,
1836		 * that's enough and we can just return.  For WB_SYNC_ALL, we
1837		 * must wait for the existing writeback to complete, then do
1838		 * writeback again if there's anything left.
1839		 */
1840		if (wbc->sync_mode != WB_SYNC_ALL)
1841			goto out;
1842		inode_wait_for_writeback(inode);
1843	}
1844	WARN_ON(inode_state_read(inode) & I_SYNC);
1845	/*
1846	 * If the inode is already fully clean, then there's nothing to do.
1847	 *
1848	 * For data-integrity syncs we also need to check whether any pages are
1849	 * still under writeback, e.g. due to prior WB_SYNC_NONE writeback.  If
1850	 * there are any such pages, we'll need to wait for them.
1851	 */
1852	if (!(inode_state_read(inode) & I_DIRTY_ALL) &&
1853	    (wbc->sync_mode != WB_SYNC_ALL ||
1854	     !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1855		goto out;
1856	inode_state_set(inode, I_SYNC);
1857	wbc_attach_and_unlock_inode(wbc, inode);
1858
1859	ret = __writeback_single_inode(inode, wbc);
1860
1861	wbc_detach_inode(wbc);
1862
1863	wb = inode_to_wb_and_lock_list(inode);
1864	spin_lock(&inode->i_lock);
1865	/*
1866	 * If the inode is freeing, its i_io_list shoudn't be updated
1867	 * as it can be finally deleted at this moment.
1868	 */
1869	if (!(inode_state_read(inode) & I_FREEING)) {
1870		/*
1871		 * If the inode is now fully clean, then it can be safely
1872		 * removed from its writeback list (if any). Otherwise the
1873		 * flusher threads are responsible for the writeback lists.
1874		 */
1875		if (!(inode_state_read(inode) & I_DIRTY_ALL))
1876			inode_cgwb_move_to_attached(inode, wb);
1877		else if (!(inode_state_read(inode) & I_SYNC_QUEUED)) {
1878			if ((inode_state_read(inode) & I_DIRTY))
1879				redirty_tail_locked(inode, wb);
1880			else if (inode_state_read(inode) & I_DIRTY_TIME) {
1881				inode->dirtied_when = jiffies;
1882				inode_io_list_move_locked(inode,
1883							  wb,
1884							  &wb->b_dirty_time);
1885			}
1886		}
1887	}
1888
1889	spin_unlock(&wb->list_lock);
1890	inode_sync_complete(inode);
1891out:
1892	spin_unlock(&inode->i_lock);
1893	return ret;
1894}
1895
1896static long writeback_chunk_size(struct super_block *sb,
1897		struct bdi_writeback *wb, struct wb_writeback_work *work)
1898{
1899	long pages;
1900
1901	/*
1902	 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1903	 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1904	 * here avoids calling into writeback_inodes_wb() more than once.
1905	 *
1906	 * The intended call sequence for WB_SYNC_ALL writeback is:
1907	 *
1908	 *      wb_writeback()
1909	 *          writeback_sb_inodes()       <== called only once
1910	 *              write_cache_pages()     <== called once for each inode
1911	 *                   (quickly) tag currently dirty pages
1912	 *                   (maybe slowly) sync all tagged pages
1913	 */
1914	if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1915		return LONG_MAX;
1916
1917	pages = min(wb->avg_write_bandwidth / 2,
1918		    global_wb_domain.dirty_limit / DIRTY_SCOPE);
1919	pages = min(pages, work->nr_pages);
1920	return round_down(pages + sb->s_min_writeback_pages,
1921			sb->s_min_writeback_pages);
1922}
1923
1924/*
1925 * Write a portion of b_io inodes which belong to @sb.
1926 *
1927 * Return the number of pages and/or inodes written.
1928 *
1929 * NOTE! This is called with wb->list_lock held, and will
1930 * unlock and relock that for each inode it ends up doing
1931 * IO for.
1932 */
1933static long writeback_sb_inodes(struct super_block *sb,
1934				struct bdi_writeback *wb,
1935				struct wb_writeback_work *work)
1936{
1937	struct writeback_control wbc = {
1938		.sync_mode		= work->sync_mode,
1939		.tagged_writepages	= work->tagged_writepages,
1940		.for_kupdate		= work->for_kupdate,
1941		.for_background		= work->for_background,
1942		.for_sync		= work->for_sync,
1943		.range_cyclic		= work->range_cyclic,
1944		.range_start		= 0,
1945		.range_end		= LLONG_MAX,
1946	};
1947	unsigned long start_time = jiffies;
1948	unsigned long timeout = sysctl_hung_task_timeout_secs;
1949	long write_chunk;
1950	long total_wrote = 0;  /* count both pages and inodes */
1951	unsigned long dirtied_before = jiffies;
1952
1953	if (work->for_kupdate)
1954		dirtied_before = jiffies -
1955			msecs_to_jiffies(dirty_expire_interval * 10);
1956
1957	while (!list_empty(&wb->b_io)) {
1958		struct inode *inode = wb_inode(wb->b_io.prev);
1959		struct bdi_writeback *tmp_wb;
1960		long wrote;
1961
1962		if (inode->i_sb != sb) {
1963			if (work->sb) {
1964				/*
1965				 * We only want to write back data for this
1966				 * superblock, move all inodes not belonging
1967				 * to it back onto the dirty list.
1968				 */
1969				redirty_tail(inode, wb);
1970				continue;
1971			}
1972
1973			/*
1974			 * The inode belongs to a different superblock.
1975			 * Bounce back to the caller to unpin this and
1976			 * pin the next superblock.
1977			 */
1978			break;
1979		}
1980
1981		/*
1982		 * Don't bother with new inodes or inodes being freed, first
1983		 * kind does not need periodic writeout yet, and for the latter
1984		 * kind writeout is handled by the freer.
1985		 */
1986		spin_lock(&inode->i_lock);
1987		if (inode_state_read(inode) & (I_NEW | I_FREEING | I_WILL_FREE)) {
1988			redirty_tail_locked(inode, wb);
1989			spin_unlock(&inode->i_lock);
1990			continue;
1991		}
1992		if ((inode_state_read(inode) & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1993			/*
1994			 * If this inode is locked for writeback and we are not
1995			 * doing writeback-for-data-integrity, move it to
1996			 * b_more_io so that writeback can proceed with the
1997			 * other inodes on s_io.
1998			 *
1999			 * We'll have another go at writing back this inode
2000			 * when we completed a full scan of b_io.
2001			 */
2002			requeue_io(inode, wb);
2003			spin_unlock(&inode->i_lock);
2004			trace_writeback_sb_inodes_requeue(inode);
2005			continue;
2006		}
2007		spin_unlock(&wb->list_lock);
2008
2009		/*
2010		 * We already requeued the inode if it had I_SYNC set and we
2011		 * are doing WB_SYNC_NONE writeback. So this catches only the
2012		 * WB_SYNC_ALL case.
2013		 */
2014		if (inode_state_read(inode) & I_SYNC) {
2015			/* Wait for I_SYNC. This function drops i_lock... */
2016			inode_sleep_on_writeback(inode);
2017			/* Inode may be gone, start again */
2018			spin_lock(&wb->list_lock);
2019			continue;
2020		}
2021		inode_state_set(inode, I_SYNC);
2022		wbc_attach_and_unlock_inode(&wbc, inode);
2023
2024		write_chunk = writeback_chunk_size(inode->i_sb, wb, work);
2025		wbc.nr_to_write = write_chunk;
2026		wbc.pages_skipped = 0;
2027
2028		/*
2029		 * We use I_SYNC to pin the inode in memory. While it is set
2030		 * evict_inode() will wait so the inode cannot be freed.
2031		 */
2032		__writeback_single_inode(inode, &wbc);
2033
2034		/* Report progress to inform the hung task detector of the progress. */
2035		if (work->done && work->done->progress_stamp && timeout &&
2036		   (jiffies - work->done->progress_stamp) > HZ * timeout / 2)
2037			wake_up_all(work->done->waitq);
2038
2039		wbc_detach_inode(&wbc);
2040		work->nr_pages -= write_chunk - wbc.nr_to_write;
2041		wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped;
2042		wrote = wrote < 0 ? 0 : wrote;
2043		total_wrote += wrote;
2044
2045		if (need_resched()) {
2046			/*
2047			 * We're trying to balance between building up a nice
2048			 * long list of IOs to improve our merge rate, and
2049			 * getting those IOs out quickly for anyone throttling
2050			 * in balance_dirty_pages().  cond_resched() doesn't
2051			 * unplug, so get our IOs out the door before we
2052			 * give up the CPU.
2053			 */
2054			blk_flush_plug(current->plug, false);
2055			cond_resched();
2056		}
2057
2058		/*
2059		 * Requeue @inode if still dirty.  Be careful as @inode may
2060		 * have been switched to another wb in the meantime.
2061		 */
2062		tmp_wb = inode_to_wb_and_lock_list(inode);
2063		spin_lock(&inode->i_lock);
2064		if (!(inode_state_read(inode) & I_DIRTY_ALL))
2065			total_wrote++;
2066		requeue_inode(inode, tmp_wb, &wbc, dirtied_before);
2067		inode_sync_complete(inode);
2068		spin_unlock(&inode->i_lock);
2069
2070		if (unlikely(tmp_wb != wb)) {
2071			spin_unlock(&tmp_wb->list_lock);
2072			spin_lock(&wb->list_lock);
2073		}
2074
2075		/*
2076		 * bail out to wb_writeback() often enough to check
2077		 * background threshold and other termination conditions.
2078		 */
2079		if (total_wrote) {
2080			if (time_is_before_jiffies(start_time + HZ / 10UL))
2081				break;
2082			if (work->nr_pages <= 0)
2083				break;
2084		}
2085	}
2086	return total_wrote;
2087}
2088
2089static long __writeback_inodes_wb(struct bdi_writeback *wb,
2090				  struct wb_writeback_work *work)
2091{
2092	unsigned long start_time = jiffies;
2093	long wrote = 0;
2094
2095	while (!list_empty(&wb->b_io)) {
2096		struct inode *inode = wb_inode(wb->b_io.prev);
2097		struct super_block *sb = inode->i_sb;
2098
2099		if (!super_trylock_shared(sb)) {
2100			/*
2101			 * super_trylock_shared() may fail consistently due to
2102			 * s_umount being grabbed by someone else. Don't use
2103			 * requeue_io() to avoid busy retrying the inode/sb.
2104			 */
2105			redirty_tail(inode, wb);
2106			continue;
2107		}
2108		wrote += writeback_sb_inodes(sb, wb, work);
2109		up_read(&sb->s_umount);
2110
2111		/* refer to the same tests at the end of writeback_sb_inodes */
2112		if (wrote) {
2113			if (time_is_before_jiffies(start_time + HZ / 10UL))
2114				break;
2115			if (work->nr_pages <= 0)
2116				break;
2117		}
2118	}
2119	/* Leave any unwritten inodes on b_io */
2120	return wrote;
2121}
2122
2123static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
2124				enum wb_reason reason)
2125{
2126	struct wb_writeback_work work = {
2127		.nr_pages	= nr_pages,
2128		.sync_mode	= WB_SYNC_NONE,
2129		.range_cyclic	= 1,
2130		.reason		= reason,
2131	};
2132	struct blk_plug plug;
2133
2134	blk_start_plug(&plug);
2135	spin_lock(&wb->list_lock);
2136	if (list_empty(&wb->b_io))
2137		queue_io(wb, &work, jiffies);
2138	__writeback_inodes_wb(wb, &work);
2139	spin_unlock(&wb->list_lock);
2140	blk_finish_plug(&plug);
2141
2142	return nr_pages - work.nr_pages;
2143}
2144
2145/*
2146 * Explicit flushing or periodic writeback of "old" data.
2147 *
2148 * Define "old": the first time one of an inode's pages is dirtied, we mark the
2149 * dirtying-time in the inode's address_space.  So this periodic writeback code
2150 * just walks the superblock inode list, writing back any inodes which are
2151 * older than a specific point in time.
2152 *
2153 * Try to run once per dirty_writeback_interval.  But if a writeback event
2154 * takes longer than a dirty_writeback_interval interval, then leave a
2155 * one-second gap.
2156 *
2157 * dirtied_before takes precedence over nr_to_write.  So we'll only write back
2158 * all dirty pages if they are all attached to "old" mappings.
2159 */
2160static long wb_writeback(struct bdi_writeback *wb,
2161			 struct wb_writeback_work *work)
2162{
2163	long nr_pages = work->nr_pages;
2164	unsigned long dirtied_before = jiffies;
2165	struct inode *inode;
2166	long progress;
2167	struct blk_plug plug;
2168	bool queued = false;
2169
2170	blk_start_plug(&plug);
2171	for (;;) {
2172		/*
2173		 * Stop writeback when nr_pages has been consumed
2174		 */
2175		if (work->nr_pages <= 0)
2176			break;
2177
2178		/*
2179		 * Background writeout and kupdate-style writeback may
2180		 * run forever. Stop them if there is other work to do
2181		 * so that e.g. sync can proceed. They'll be restarted
2182		 * after the other works are all done.
2183		 */
2184		if ((work->for_background || work->for_kupdate) &&
2185		    !list_empty(&wb->work_list))
2186			break;
2187
2188		/*
2189		 * For background writeout, stop when we are below the
2190		 * background dirty threshold
2191		 */
2192		if (work->for_background && !wb_over_bg_thresh(wb))
2193			break;
2194
2195
2196		spin_lock(&wb->list_lock);
2197
2198		trace_writeback_start(wb, work);
2199		if (list_empty(&wb->b_io)) {
2200			/*
2201			 * Kupdate and background works are special and we want
2202			 * to include all inodes that need writing. Livelock
2203			 * avoidance is handled by these works yielding to any
2204			 * other work so we are safe.
2205			 */
2206			if (work->for_kupdate) {
2207				dirtied_before = jiffies -
2208					msecs_to_jiffies(dirty_expire_interval *
2209							 10);
2210			} else if (work->for_background)
2211				dirtied_before = jiffies;
2212
2213			queue_io(wb, work, dirtied_before);
2214			queued = true;
2215		}
2216		if (work->sb)
2217			progress = writeback_sb_inodes(work->sb, wb, work);
2218		else
2219			progress = __writeback_inodes_wb(wb, work);
2220		trace_writeback_written(wb, work);
2221
2222		/*
2223		 * Did we write something? Try for more
2224		 *
2225		 * Dirty inodes are moved to b_io for writeback in batches.
2226		 * The completion of the current batch does not necessarily
2227		 * mean the overall work is done. So we keep looping as long
2228		 * as made some progress on cleaning pages or inodes.
2229		 */
2230		if (progress || !queued) {
2231			spin_unlock(&wb->list_lock);
2232			continue;
2233		}
2234
2235		/*
2236		 * No more inodes for IO, bail
2237		 */
2238		if (list_empty(&wb->b_more_io)) {
2239			spin_unlock(&wb->list_lock);
2240			break;
2241		}
2242
2243		/*
2244		 * Nothing written. Wait for some inode to
2245		 * become available for writeback. Otherwise
2246		 * we'll just busyloop.
2247		 */
2248		trace_writeback_wait(wb, work);
2249		inode = wb_inode(wb->b_more_io.prev);
2250		spin_lock(&inode->i_lock);
2251		spin_unlock(&wb->list_lock);
2252		/* This function drops i_lock... */
2253		inode_sleep_on_writeback(inode);
2254	}
2255	blk_finish_plug(&plug);
2256
2257	return nr_pages - work->nr_pages;
2258}
2259
2260/*
2261 * Return the next wb_writeback_work struct that hasn't been processed yet.
2262 */
2263static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
2264{
2265	struct wb_writeback_work *work = NULL;
2266
2267	spin_lock_irq(&wb->work_lock);
2268	if (!list_empty(&wb->work_list)) {
2269		work = list_entry(wb->work_list.next,
2270				  struct wb_writeback_work, list);
2271		list_del_init(&work->list);
2272	}
2273	spin_unlock_irq(&wb->work_lock);
2274	return work;
2275}
2276
2277static long wb_check_background_flush(struct bdi_writeback *wb)
2278{
2279	if (wb_over_bg_thresh(wb)) {
2280
2281		struct wb_writeback_work work = {
2282			.nr_pages	= LONG_MAX,
2283			.sync_mode	= WB_SYNC_NONE,
2284			.for_background	= 1,
2285			.range_cyclic	= 1,
2286			.reason		= WB_REASON_BACKGROUND,
2287		};
2288
2289		return wb_writeback(wb, &work);
2290	}
2291
2292	return 0;
2293}
2294
2295static long wb_check_old_data_flush(struct bdi_writeback *wb)
2296{
2297	unsigned long expired;
2298	long nr_pages;
2299
2300	/*
2301	 * When set to zero, disable periodic writeback
2302	 */
2303	if (!dirty_writeback_interval)
2304		return 0;
2305
2306	expired = wb->last_old_flush +
2307			msecs_to_jiffies(dirty_writeback_interval * 10);
2308	if (time_before(jiffies, expired))
2309		return 0;
2310
2311	wb->last_old_flush = jiffies;
2312	nr_pages = get_nr_dirty_pages();
2313
2314	if (nr_pages) {
2315		struct wb_writeback_work work = {
2316			.nr_pages	= nr_pages,
2317			.sync_mode	= WB_SYNC_NONE,
2318			.for_kupdate	= 1,
2319			.range_cyclic	= 1,
2320			.reason		= WB_REASON_PERIODIC,
2321		};
2322
2323		return wb_writeback(wb, &work);
2324	}
2325
2326	return 0;
2327}
2328
2329static long wb_check_start_all(struct bdi_writeback *wb)
2330{
2331	long nr_pages;
2332
2333	if (!test_bit(WB_start_all, &wb->state))
2334		return 0;
2335
2336	nr_pages = get_nr_dirty_pages();
2337	if (nr_pages) {
2338		struct wb_writeback_work work = {
2339			.nr_pages	= wb_split_bdi_pages(wb, nr_pages),
2340			.sync_mode	= WB_SYNC_NONE,
2341			.range_cyclic	= 1,
2342			.reason		= wb->start_all_reason,
2343		};
2344
2345		nr_pages = wb_writeback(wb, &work);
2346	}
2347
2348	clear_bit(WB_start_all, &wb->state);
2349	return nr_pages;
2350}
2351
2352
2353/*
2354 * Retrieve work items and do the writeback they describe
2355 */
2356static long wb_do_writeback(struct bdi_writeback *wb)
2357{
2358	struct wb_writeback_work *work;
2359	long wrote = 0;
2360
2361	set_bit(WB_writeback_running, &wb->state);
2362	while ((work = get_next_work_item(wb)) != NULL) {
2363		trace_writeback_exec(wb, work);
2364		wrote += wb_writeback(wb, work);
2365		finish_writeback_work(work);
2366	}
2367
2368	/*
2369	 * Check for a flush-everything request
2370	 */
2371	wrote += wb_check_start_all(wb);
2372
2373	/*
2374	 * Check for periodic writeback, kupdated() style
2375	 */
2376	wrote += wb_check_old_data_flush(wb);
2377	wrote += wb_check_background_flush(wb);
2378	clear_bit(WB_writeback_running, &wb->state);
2379
2380	return wrote;
2381}
2382
2383/*
2384 * Handle writeback of dirty data for the device backed by this bdi. Also
2385 * reschedules periodically and does kupdated style flushing.
2386 */
2387void wb_workfn(struct work_struct *work)
2388{
2389	struct bdi_writeback *wb = container_of(to_delayed_work(work),
2390						struct bdi_writeback, dwork);
2391	long pages_written;
2392
2393	set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
2394
2395	if (likely(!current_is_workqueue_rescuer() ||
2396		   !test_bit(WB_registered, &wb->state))) {
2397		/*
2398		 * The normal path.  Keep writing back @wb until its
2399		 * work_list is empty.  Note that this path is also taken
2400		 * if @wb is shutting down even when we're running off the
2401		 * rescuer as work_list needs to be drained.
2402		 */
2403		do {
2404			pages_written = wb_do_writeback(wb);
2405			trace_writeback_pages_written(pages_written);
2406		} while (!list_empty(&wb->work_list));
2407	} else {
2408		/*
2409		 * bdi_wq can't get enough workers and we're running off
2410		 * the emergency worker.  Don't hog it.  Hopefully, 1024 is
2411		 * enough for efficient IO.
2412		 */
2413		pages_written = writeback_inodes_wb(wb, 1024,
2414						    WB_REASON_FORKER_THREAD);
2415		trace_writeback_pages_written(pages_written);
2416	}
2417
2418	if (!list_empty(&wb->work_list))
2419		wb_wakeup(wb);
2420	else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
2421		wb_wakeup_delayed(wb);
2422}
2423
2424/*
2425 * Start writeback of all dirty pages on this bdi.
2426 */
2427static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2428					 enum wb_reason reason)
2429{
2430	struct bdi_writeback *wb;
2431
2432	if (!bdi_has_dirty_io(bdi))
2433		return;
2434
2435	list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2436		wb_start_writeback(wb, reason);
2437}
2438
2439void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
2440				enum wb_reason reason)
2441{
2442	rcu_read_lock();
2443	__wakeup_flusher_threads_bdi(bdi, reason);
2444	rcu_read_unlock();
2445}
2446
2447/*
2448 * Wakeup the flusher threads to start writeback of all currently dirty pages
2449 */
2450void wakeup_flusher_threads(enum wb_reason reason)
2451{
2452	struct backing_dev_info *bdi;
2453
2454	/*
2455	 * If we are expecting writeback progress we must submit plugged IO.
2456	 */
2457	blk_flush_plug(current->plug, true);
2458
2459	rcu_read_lock();
2460	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2461		__wakeup_flusher_threads_bdi(bdi, reason);
2462	rcu_read_unlock();
2463}
2464
2465/*
2466 * Wake up bdi's periodically to make sure dirtytime inodes gets
2467 * written back periodically.  We deliberately do *not* check the
2468 * b_dirtytime list in wb_has_dirty_io(), since this would cause the
2469 * kernel to be constantly waking up once there are any dirtytime
2470 * inodes on the system.  So instead we define a separate delayed work
2471 * function which gets called much more rarely.  (By default, only
2472 * once every 12 hours.)
2473 *
2474 * If there is any other write activity going on in the file system,
2475 * this function won't be necessary.  But if the only thing that has
2476 * happened on the file system is a dirtytime inode caused by an atime
2477 * update, we need this infrastructure below to make sure that inode
2478 * eventually gets pushed out to disk.
2479 */
2480static void wakeup_dirtytime_writeback(struct work_struct *w);
2481static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
2482
2483static void wakeup_dirtytime_writeback(struct work_struct *w)
2484{
2485	struct backing_dev_info *bdi;
2486
2487	rcu_read_lock();
2488	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
2489		struct bdi_writeback *wb;
2490
2491		list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2492			if (!list_empty(&wb->b_dirty_time))
2493				wb_wakeup(wb);
2494	}
2495	rcu_read_unlock();
2496	if (dirtytime_expire_interval)
2497		schedule_delayed_work(&dirtytime_work,
2498				      round_jiffies_relative(dirtytime_expire_interval * HZ));
2499}
2500
2501static int dirtytime_interval_handler(const struct ctl_table *table, int write,
2502			       void *buffer, size_t *lenp, loff_t *ppos)
2503{
2504	int ret;
2505
2506	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2507	if (ret == 0 && write) {
2508		if (dirtytime_expire_interval)
2509			mod_delayed_work(system_percpu_wq, &dirtytime_work, 0);
2510		else
2511			cancel_delayed_work_sync(&dirtytime_work);
2512	}
2513	return ret;
2514}
2515
2516static const struct ctl_table vm_fs_writeback_table[] = {
2517	{
2518		.procname	= "dirtytime_expire_seconds",
2519		.data		= &dirtytime_expire_interval,
2520		.maxlen		= sizeof(dirtytime_expire_interval),
2521		.mode		= 0644,
2522		.proc_handler	= dirtytime_interval_handler,
2523		.extra1		= SYSCTL_ZERO,
2524	},
2525};
2526
2527static int __init start_dirtytime_writeback(void)
2528{
2529	if (dirtytime_expire_interval)
2530		schedule_delayed_work(&dirtytime_work,
2531				      round_jiffies_relative(dirtytime_expire_interval * HZ));
2532	register_sysctl_init("vm", vm_fs_writeback_table);
2533	return 0;
2534}
2535__initcall(start_dirtytime_writeback);
2536
2537/**
2538 * __mark_inode_dirty -	internal function to mark an inode dirty
2539 *
2540 * @inode: inode to mark
2541 * @flags: what kind of dirty, e.g. I_DIRTY_SYNC.  This can be a combination of
2542 *	   multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined
2543 *	   with I_DIRTY_PAGES.
2544 *
2545 * Mark an inode as dirty.  We notify the filesystem, then update the inode's
2546 * dirty flags.  Then, if needed we add the inode to the appropriate dirty list.
2547 *
2548 * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync()
2549 * instead of calling this directly.
2550 *
2551 * CAREFUL!  We only add the inode to the dirty list if it is hashed or if it
2552 * refers to a blockdev.  Unhashed inodes will never be added to the dirty list
2553 * even if they are later hashed, as they will have been marked dirty already.
2554 *
2555 * In short, ensure you hash any inodes _before_ you start marking them dirty.
2556 *
2557 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2558 * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
2559 * the kernel-internal blockdev inode represents the dirtying time of the
2560 * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
2561 * page->mapping->host, so the page-dirtying time is recorded in the internal
2562 * blockdev inode.
2563 */
2564void __mark_inode_dirty(struct inode *inode, int flags)
2565{
2566	struct super_block *sb = inode->i_sb;
2567	int dirtytime = 0;
2568	struct bdi_writeback *wb = NULL;
2569
2570	trace_writeback_mark_inode_dirty(inode, flags);
2571
2572	if (flags & I_DIRTY_INODE) {
2573		/*
2574		 * Inode timestamp update will piggback on this dirtying.
2575		 * We tell ->dirty_inode callback that timestamps need to
2576		 * be updated by setting I_DIRTY_TIME in flags.
2577		 */
2578		if (inode_state_read_once(inode) & I_DIRTY_TIME) {
2579			spin_lock(&inode->i_lock);
2580			if (inode_state_read(inode) & I_DIRTY_TIME) {
2581				inode_state_clear(inode, I_DIRTY_TIME);
2582				flags |= I_DIRTY_TIME;
2583			}
2584			spin_unlock(&inode->i_lock);
2585		}
2586
2587		/*
2588		 * Notify the filesystem about the inode being dirtied, so that
2589		 * (if needed) it can update on-disk fields and journal the
2590		 * inode.  This is only needed when the inode itself is being
2591		 * dirtied now.  I.e. it's only needed for I_DIRTY_INODE, not
2592		 * for just I_DIRTY_PAGES or I_DIRTY_TIME.
2593		 */
2594		trace_writeback_dirty_inode_start(inode, flags);
2595		if (sb->s_op->dirty_inode)
2596			sb->s_op->dirty_inode(inode,
2597				flags & (I_DIRTY_INODE | I_DIRTY_TIME));
2598		trace_writeback_dirty_inode(inode, flags);
2599
2600		/* I_DIRTY_INODE supersedes I_DIRTY_TIME. */
2601		flags &= ~I_DIRTY_TIME;
2602	} else {
2603		/*
2604		 * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing.
2605		 * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME
2606		 * in one call to __mark_inode_dirty().)
2607		 */
2608		dirtytime = flags & I_DIRTY_TIME;
2609		WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME);
2610	}
2611
2612	/*
2613	 * Paired with smp_mb() in __writeback_single_inode() for the
2614	 * following lockless i_state test.  See there for details.
2615	 */
2616	smp_mb();
2617
2618	if ((inode_state_read_once(inode) & flags) == flags)
2619		return;
2620
2621	spin_lock(&inode->i_lock);
2622	if ((inode_state_read(inode) & flags) != flags) {
2623		const int was_dirty = inode_state_read(inode) & I_DIRTY;
2624
2625		inode_attach_wb(inode, NULL);
2626
2627		inode_state_set(inode, flags);
2628
2629		/*
2630		 * Grab inode's wb early because it requires dropping i_lock and we
2631		 * need to make sure following checks happen atomically with dirty
2632		 * list handling so that we don't move inodes under flush worker's
2633		 * hands.
2634		 */
2635		if (!was_dirty) {
2636			wb = locked_inode_to_wb_and_lock_list(inode);
2637			spin_lock(&inode->i_lock);
2638		}
2639
2640		/*
2641		 * If the inode is queued for writeback by flush worker, just
2642		 * update its dirty state. Once the flush worker is done with
2643		 * the inode it will place it on the appropriate superblock
2644		 * list, based upon its state.
2645		 */
2646		if (inode_state_read(inode) & I_SYNC_QUEUED)
2647			goto out_unlock;
2648
2649		/*
2650		 * Only add valid (hashed) inodes to the superblock's
2651		 * dirty list.  Add blockdev inodes as well.
2652		 */
2653		if (!S_ISBLK(inode->i_mode)) {
2654			if (inode_unhashed(inode))
2655				goto out_unlock;
2656		}
2657		if (inode_state_read(inode) & I_FREEING)
2658			goto out_unlock;
2659
2660		/*
2661		 * If the inode was already on b_dirty/b_io/b_more_io, don't
2662		 * reposition it (that would break b_dirty time-ordering).
2663		 */
2664		if (!was_dirty) {
2665			struct list_head *dirty_list;
2666			bool wakeup_bdi = false;
2667
2668			inode->dirtied_when = jiffies;
2669			if (dirtytime)
2670				inode->dirtied_time_when = jiffies;
2671
2672			if (inode_state_read(inode) & I_DIRTY)
2673				dirty_list = &wb->b_dirty;
2674			else
2675				dirty_list = &wb->b_dirty_time;
2676
2677			wakeup_bdi = inode_io_list_move_locked(inode, wb,
2678							       dirty_list);
2679
2680			/*
2681			 * If this is the first dirty inode for this bdi,
2682			 * we have to wake-up the corresponding bdi thread
2683			 * to make sure background write-back happens
2684			 * later.
2685			 */
2686			if (wakeup_bdi &&
2687			    (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
2688				wb_wakeup_delayed(wb);
2689
2690			spin_unlock(&wb->list_lock);
2691			spin_unlock(&inode->i_lock);
2692			trace_writeback_dirty_inode_enqueue(inode);
2693
2694			return;
2695		}
2696	}
2697out_unlock:
2698	if (wb)
2699		spin_unlock(&wb->list_lock);
2700	spin_unlock(&inode->i_lock);
2701}
2702EXPORT_SYMBOL(__mark_inode_dirty);
2703
2704/*
2705 * The @s_sync_lock is used to serialise concurrent sync operations
2706 * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2707 * Concurrent callers will block on the s_sync_lock rather than doing contending
2708 * walks. The queueing maintains sync(2) required behaviour as all the IO that
2709 * has been issued up to the time this function is enter is guaranteed to be
2710 * completed by the time we have gained the lock and waited for all IO that is
2711 * in progress regardless of the order callers are granted the lock.
2712 */
2713static void wait_sb_inodes(struct super_block *sb)
2714{
2715	LIST_HEAD(sync_list);
2716
2717	/*
2718	 * We need to be protected against the filesystem going from
2719	 * r/o to r/w or vice versa.
2720	 */
2721	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2722
2723	mutex_lock(&sb->s_sync_lock);
2724
2725	/*
2726	 * Splice the writeback list onto a temporary list to avoid waiting on
2727	 * inodes that have started writeback after this point.
2728	 *
2729	 * Use rcu_read_lock() to keep the inodes around until we have a
2730	 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2731	 * the local list because inodes can be dropped from either by writeback
2732	 * completion.
2733	 */
2734	rcu_read_lock();
2735	spin_lock_irq(&sb->s_inode_wblist_lock);
2736	list_splice_init(&sb->s_inodes_wb, &sync_list);
2737
2738	/*
2739	 * Data integrity sync. Must wait for all pages under writeback, because
2740	 * there may have been pages dirtied before our sync call, but which had
2741	 * writeout started before we write it out.  In which case, the inode
2742	 * may not be on the dirty list, but we still have to wait for that
2743	 * writeout.
2744	 */
2745	while (!list_empty(&sync_list)) {
2746		struct inode *inode = list_first_entry(&sync_list, struct inode,
2747						       i_wb_list);
2748		struct address_space *mapping = inode->i_mapping;
2749
2750		/*
2751		 * Move each inode back to the wb list before we drop the lock
2752		 * to preserve consistency between i_wb_list and the mapping
2753		 * writeback tag. Writeback completion is responsible to remove
2754		 * the inode from either list once the writeback tag is cleared.
2755		 */
2756		list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2757
2758		/*
2759		 * The mapping can appear untagged while still on-list since we
2760		 * do not have the mapping lock. Skip it here, wb completion
2761		 * will remove it.
2762		 */
2763		if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2764			continue;
2765
2766		spin_unlock_irq(&sb->s_inode_wblist_lock);
2767
2768		spin_lock(&inode->i_lock);
2769		if (inode_state_read(inode) & (I_FREEING | I_WILL_FREE | I_NEW)) {
2770			spin_unlock(&inode->i_lock);
2771
2772			spin_lock_irq(&sb->s_inode_wblist_lock);
2773			continue;
2774		}
2775		__iget(inode);
2776		spin_unlock(&inode->i_lock);
2777		rcu_read_unlock();
2778
2779		/*
2780		 * We keep the error status of individual mapping so that
2781		 * applications can catch the writeback error using fsync(2).
2782		 * See filemap_fdatawait_keep_errors() for details.
2783		 */
2784		filemap_fdatawait_keep_errors(mapping);
2785
2786		cond_resched();
2787
2788		iput(inode);
2789
2790		rcu_read_lock();
2791		spin_lock_irq(&sb->s_inode_wblist_lock);
2792	}
2793	spin_unlock_irq(&sb->s_inode_wblist_lock);
2794	rcu_read_unlock();
2795	mutex_unlock(&sb->s_sync_lock);
2796}
2797
2798static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2799				     enum wb_reason reason, bool skip_if_busy)
2800{
2801	struct backing_dev_info *bdi = sb->s_bdi;
2802	DEFINE_WB_COMPLETION(done, bdi);
2803	struct wb_writeback_work work = {
2804		.sb			= sb,
2805		.sync_mode		= WB_SYNC_NONE,
2806		.tagged_writepages	= 1,
2807		.done			= &done,
2808		.nr_pages		= nr,
2809		.reason			= reason,
2810	};
2811
2812	if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2813		return;
2814	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2815
2816	bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2817	wb_wait_for_completion(&done);
2818}
2819
2820/**
2821 * writeback_inodes_sb_nr -	writeback dirty inodes from given super_block
2822 * @sb: the superblock
2823 * @nr: the number of pages to write
2824 * @reason: reason why some writeback work initiated
2825 *
2826 * Start writeback on some inodes on this super_block. No guarantees are made
2827 * on how many (if any) will be written, and this function does not wait
2828 * for IO completion of submitted IO.
2829 */
2830void writeback_inodes_sb_nr(struct super_block *sb,
2831			    unsigned long nr,
2832			    enum wb_reason reason)
2833{
2834	__writeback_inodes_sb_nr(sb, nr, reason, false);
2835}
2836EXPORT_SYMBOL(writeback_inodes_sb_nr);
2837
2838/**
2839 * writeback_inodes_sb	-	writeback dirty inodes from given super_block
2840 * @sb: the superblock
2841 * @reason: reason why some writeback work was initiated
2842 *
2843 * Start writeback on some inodes on this super_block. No guarantees are made
2844 * on how many (if any) will be written, and this function does not wait
2845 * for IO completion of submitted IO.
2846 */
2847void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2848{
2849	writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2850}
2851EXPORT_SYMBOL(writeback_inodes_sb);
2852
2853/**
2854 * try_to_writeback_inodes_sb - try to start writeback if none underway
2855 * @sb: the superblock
2856 * @reason: reason why some writeback work was initiated
2857 *
2858 * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
2859 */
2860void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2861{
2862	if (!down_read_trylock(&sb->s_umount))
2863		return;
2864
2865	__writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
2866	up_read(&sb->s_umount);
2867}
2868EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2869
2870/**
2871 * sync_inodes_sb	-	sync sb inode pages
2872 * @sb: the superblock
2873 *
2874 * This function writes and waits on any dirty inode belonging to this
2875 * super_block.
2876 */
2877void sync_inodes_sb(struct super_block *sb)
2878{
2879	struct backing_dev_info *bdi = sb->s_bdi;
2880	DEFINE_WB_COMPLETION(done, bdi);
2881	struct wb_writeback_work work = {
2882		.sb		= sb,
2883		.sync_mode	= WB_SYNC_ALL,
2884		.nr_pages	= LONG_MAX,
2885		.range_cyclic	= 0,
2886		.done		= &done,
2887		.reason		= WB_REASON_SYNC,
2888		.for_sync	= 1,
2889	};
2890
2891	/*
2892	 * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2893	 * inodes under writeback and I_DIRTY_TIME inodes ignored by
2894	 * bdi_has_dirty() need to be written out too.
2895	 */
2896	if (bdi == &noop_backing_dev_info)
2897		return;
2898
2899	/*
2900	 * If the superblock has SB_I_NO_DATA_INTEGRITY set, there's no need to
2901	 * wait for the writeout to complete, as the filesystem cannot guarantee
2902	 * data persistence on sync. Just kick off writeback and return.
2903	 */
2904	if (sb->s_iflags & SB_I_NO_DATA_INTEGRITY) {
2905		wakeup_flusher_threads_bdi(bdi, WB_REASON_SYNC);
2906		return;
2907	}
2908
2909	WARN_ON(!rwsem_is_locked(&sb->s_umount));
2910
2911	/* protect against inode wb switch, see inode_switch_wbs_work_fn() */
2912	bdi_down_write_wb_switch_rwsem(bdi);
2913	bdi_split_work_to_wbs(bdi, &work, false);
2914	wb_wait_for_completion(&done);
2915	bdi_up_write_wb_switch_rwsem(bdi);
2916
2917	wait_sb_inodes(sb);
2918}
2919EXPORT_SYMBOL(sync_inodes_sb);
2920
2921/**
2922 * write_inode_now	-	write an inode to disk
2923 * @inode: inode to write to disk
2924 * @sync: whether the write should be synchronous or not
2925 *
2926 * This function commits an inode to disk immediately if it is dirty. This is
2927 * primarily needed by knfsd.
2928 *
2929 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2930 */
2931int write_inode_now(struct inode *inode, int sync)
2932{
2933	struct writeback_control wbc = {
2934		.nr_to_write = LONG_MAX,
2935		.sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2936		.range_start = 0,
2937		.range_end = LLONG_MAX,
2938	};
2939
2940	if (!mapping_can_writeback(inode->i_mapping))
2941		wbc.nr_to_write = 0;
2942
2943	might_sleep();
2944	return writeback_single_inode(inode, &wbc);
2945}
2946EXPORT_SYMBOL(write_inode_now);
2947
2948/**
2949 * sync_inode_metadata - write an inode to disk
2950 * @inode: the inode to sync
2951 * @wait: wait for I/O to complete.
2952 *
2953 * Write an inode to disk and adjust its dirty state after completion.
2954 *
2955 * Note: only writes the actual inode, no associated data or other metadata.
2956 */
2957int sync_inode_metadata(struct inode *inode, int wait)
2958{
2959	struct writeback_control wbc = {
2960		.sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2961		.nr_to_write = 0, /* metadata-only */
2962	};
2963
2964	return writeback_single_inode(inode, &wbc);
2965}
2966EXPORT_SYMBOL(sync_inode_metadata);