Version:  2.0.40 2.2.26 2.4.37 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5

Linux/fs/fs-writeback.c

  1 /*
  2  * fs/fs-writeback.c
  3  *
  4  * Copyright (C) 2002, Linus Torvalds.
  5  *
  6  * Contains all the functions related to writing back and waiting
  7  * upon dirty inodes against superblocks, and writing back dirty
  8  * pages against inodes.  ie: data writeback.  Writeout of the
  9  * inode itself is not handled here.
 10  *
 11  * 10Apr2002    Andrew Morton
 12  *              Split out of fs/inode.c
 13  *              Additions for address_space-based writeback
 14  */
 15 
 16 #include <linux/kernel.h>
 17 #include <linux/export.h>
 18 #include <linux/spinlock.h>
 19 #include <linux/slab.h>
 20 #include <linux/sched.h>
 21 #include <linux/fs.h>
 22 #include <linux/mm.h>
 23 #include <linux/pagemap.h>
 24 #include <linux/kthread.h>
 25 #include <linux/writeback.h>
 26 #include <linux/blkdev.h>
 27 #include <linux/backing-dev.h>
 28 #include <linux/tracepoint.h>
 29 #include <linux/device.h>
 30 #include <linux/memcontrol.h>
 31 #include "internal.h"
 32 
 33 /*
 34  * 4MB minimal write chunk size
 35  */
 36 #define MIN_WRITEBACK_PAGES     (4096UL >> (PAGE_CACHE_SHIFT - 10))
 37 
 38 struct wb_completion {
 39         atomic_t                cnt;
 40 };
 41 
 42 /*
 43  * Passed into wb_writeback(), essentially a subset of writeback_control
 44  */
 45 struct wb_writeback_work {
 46         long nr_pages;
 47         struct super_block *sb;
 48         unsigned long *older_than_this;
 49         enum writeback_sync_modes sync_mode;
 50         unsigned int tagged_writepages:1;
 51         unsigned int for_kupdate:1;
 52         unsigned int range_cyclic:1;
 53         unsigned int for_background:1;
 54         unsigned int for_sync:1;        /* sync(2) WB_SYNC_ALL writeback */
 55         unsigned int auto_free:1;       /* free on completion */
 56         enum wb_reason reason;          /* why was writeback initiated? */
 57 
 58         struct list_head list;          /* pending work list */
 59         struct wb_completion *done;     /* set if the caller waits */
 60 };
 61 
 62 /*
 63  * If one wants to wait for one or more wb_writeback_works, each work's
 64  * ->done should be set to a wb_completion defined using the following
 65  * macro.  Once all work items are issued with wb_queue_work(), the caller
 66  * can wait for the completion of all using wb_wait_for_completion().  Work
 67  * items which are waited upon aren't freed automatically on completion.
 68  */
 69 #define DEFINE_WB_COMPLETION_ONSTACK(cmpl)                              \
 70         struct wb_completion cmpl = {                                   \
 71                 .cnt            = ATOMIC_INIT(1),                       \
 72         }
 73 
 74 
 75 /*
 76  * If an inode is constantly having its pages dirtied, but then the
 77  * updates stop dirtytime_expire_interval seconds in the past, it's
 78  * possible for the worst case time between when an inode has its
 79  * timestamps updated and when they finally get written out to be two
 80  * dirtytime_expire_intervals.  We set the default to 12 hours (in
 81  * seconds), which means most of the time inodes will have their
 82  * timestamps written to disk after 12 hours, but in the worst case a
 83  * few inodes might not their timestamps updated for 24 hours.
 84  */
 85 unsigned int dirtytime_expire_interval = 12 * 60 * 60;
 86 
 87 static inline struct inode *wb_inode(struct list_head *head)
 88 {
 89         return list_entry(head, struct inode, i_io_list);
 90 }
 91 
 92 /*
 93  * Include the creation of the trace points after defining the
 94  * wb_writeback_work structure and inline functions so that the definition
 95  * remains local to this file.
 96  */
 97 #define CREATE_TRACE_POINTS
 98 #include <trace/events/writeback.h>
 99 
100 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
101 
102 static bool wb_io_lists_populated(struct bdi_writeback *wb)
103 {
104         if (wb_has_dirty_io(wb)) {
105                 return false;
106         } else {
107                 set_bit(WB_has_dirty_io, &wb->state);
108                 WARN_ON_ONCE(!wb->avg_write_bandwidth);
109                 atomic_long_add(wb->avg_write_bandwidth,
110                                 &wb->bdi->tot_write_bandwidth);
111                 return true;
112         }
113 }
114 
115 static void wb_io_lists_depopulated(struct bdi_writeback *wb)
116 {
117         if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
118             list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
119                 clear_bit(WB_has_dirty_io, &wb->state);
120                 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
121                                         &wb->bdi->tot_write_bandwidth) < 0);
122         }
123 }
124 
125 /**
126  * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
127  * @inode: inode to be moved
128  * @wb: target bdi_writeback
129  * @head: one of @wb->b_{dirty|io|more_io}
130  *
131  * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
132  * Returns %true if @inode is the first occupant of the !dirty_time IO
133  * lists; otherwise, %false.
134  */
135 static bool inode_io_list_move_locked(struct inode *inode,
136                                       struct bdi_writeback *wb,
137                                       struct list_head *head)
138 {
139         assert_spin_locked(&wb->list_lock);
140 
141         list_move(&inode->i_io_list, head);
142 
143         /* dirty_time doesn't count as dirty_io until expiration */
144         if (head != &wb->b_dirty_time)
145                 return wb_io_lists_populated(wb);
146 
147         wb_io_lists_depopulated(wb);
148         return false;
149 }
150 
151 /**
152  * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
153  * @inode: inode to be removed
154  * @wb: bdi_writeback @inode is being removed from
155  *
156  * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
157  * clear %WB_has_dirty_io if all are empty afterwards.
158  */
159 static void inode_io_list_del_locked(struct inode *inode,
160                                      struct bdi_writeback *wb)
161 {
162         assert_spin_locked(&wb->list_lock);
163 
164         list_del_init(&inode->i_io_list);
165         wb_io_lists_depopulated(wb);
166 }
167 
168 static void wb_wakeup(struct bdi_writeback *wb)
169 {
170         spin_lock_bh(&wb->work_lock);
171         if (test_bit(WB_registered, &wb->state))
172                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
173         spin_unlock_bh(&wb->work_lock);
174 }
175 
176 static void wb_queue_work(struct bdi_writeback *wb,
177                           struct wb_writeback_work *work)
178 {
179         trace_writeback_queue(wb, work);
180 
181         spin_lock_bh(&wb->work_lock);
182         if (!test_bit(WB_registered, &wb->state))
183                 goto out_unlock;
184         if (work->done)
185                 atomic_inc(&work->done->cnt);
186         list_add_tail(&work->list, &wb->work_list);
187         mod_delayed_work(bdi_wq, &wb->dwork, 0);
188 out_unlock:
189         spin_unlock_bh(&wb->work_lock);
190 }
191 
192 /**
193  * wb_wait_for_completion - wait for completion of bdi_writeback_works
194  * @bdi: bdi work items were issued to
195  * @done: target wb_completion
196  *
197  * Wait for one or more work items issued to @bdi with their ->done field
198  * set to @done, which should have been defined with
199  * DEFINE_WB_COMPLETION_ONSTACK().  This function returns after all such
200  * work items are completed.  Work items which are waited upon aren't freed
201  * automatically on completion.
202  */
203 static void wb_wait_for_completion(struct backing_dev_info *bdi,
204                                    struct wb_completion *done)
205 {
206         atomic_dec(&done->cnt);         /* put down the initial count */
207         wait_event(bdi->wb_waitq, !atomic_read(&done->cnt));
208 }
209 
210 #ifdef CONFIG_CGROUP_WRITEBACK
211 
212 /* parameters for foreign inode detection, see wb_detach_inode() */
213 #define WB_FRN_TIME_SHIFT       13      /* 1s = 2^13, upto 8 secs w/ 16bit */
214 #define WB_FRN_TIME_AVG_SHIFT   3       /* avg = avg * 7/8 + new * 1/8 */
215 #define WB_FRN_TIME_CUT_DIV     2       /* ignore rounds < avg / 2 */
216 #define WB_FRN_TIME_PERIOD      (2 * (1 << WB_FRN_TIME_SHIFT))  /* 2s */
217 
218 #define WB_FRN_HIST_SLOTS       16      /* inode->i_wb_frn_history is 16bit */
219 #define WB_FRN_HIST_UNIT        (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
220                                         /* each slot's duration is 2s / 16 */
221 #define WB_FRN_HIST_THR_SLOTS   (WB_FRN_HIST_SLOTS / 2)
222                                         /* if foreign slots >= 8, switch */
223 #define WB_FRN_HIST_MAX_SLOTS   (WB_FRN_HIST_THR_SLOTS / 2 + 1)
224                                         /* one round can affect upto 5 slots */
225 
226 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
227 static struct workqueue_struct *isw_wq;
228 
229 void __inode_attach_wb(struct inode *inode, struct page *page)
230 {
231         struct backing_dev_info *bdi = inode_to_bdi(inode);
232         struct bdi_writeback *wb = NULL;
233 
234         if (inode_cgwb_enabled(inode)) {
235                 struct cgroup_subsys_state *memcg_css;
236 
237                 if (page) {
238                         memcg_css = mem_cgroup_css_from_page(page);
239                         wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
240                 } else {
241                         /* must pin memcg_css, see wb_get_create() */
242                         memcg_css = task_get_css(current, memory_cgrp_id);
243                         wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
244                         css_put(memcg_css);
245                 }
246         }
247 
248         if (!wb)
249                 wb = &bdi->wb;
250 
251         /*
252          * There may be multiple instances of this function racing to
253          * update the same inode.  Use cmpxchg() to tell the winner.
254          */
255         if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
256                 wb_put(wb);
257 }
258 
259 /**
260  * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
261  * @inode: inode of interest with i_lock held
262  *
263  * Returns @inode's wb with its list_lock held.  @inode->i_lock must be
264  * held on entry and is released on return.  The returned wb is guaranteed
265  * to stay @inode's associated wb until its list_lock is released.
266  */
267 static struct bdi_writeback *
268 locked_inode_to_wb_and_lock_list(struct inode *inode)
269         __releases(&inode->i_lock)
270         __acquires(&wb->list_lock)
271 {
272         while (true) {
273                 struct bdi_writeback *wb = inode_to_wb(inode);
274 
275                 /*
276                  * inode_to_wb() association is protected by both
277                  * @inode->i_lock and @wb->list_lock but list_lock nests
278                  * outside i_lock.  Drop i_lock and verify that the
279                  * association hasn't changed after acquiring list_lock.
280                  */
281                 wb_get(wb);
282                 spin_unlock(&inode->i_lock);
283                 spin_lock(&wb->list_lock);
284                 wb_put(wb);             /* not gonna deref it anymore */
285 
286                 /* i_wb may have changed inbetween, can't use inode_to_wb() */
287                 if (likely(wb == inode->i_wb))
288                         return wb;      /* @inode already has ref */
289 
290                 spin_unlock(&wb->list_lock);
291                 cpu_relax();
292                 spin_lock(&inode->i_lock);
293         }
294 }
295 
296 /**
297  * inode_to_wb_and_lock_list - determine an inode's wb and lock it
298  * @inode: inode of interest
299  *
300  * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
301  * on entry.
302  */
303 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
304         __acquires(&wb->list_lock)
305 {
306         spin_lock(&inode->i_lock);
307         return locked_inode_to_wb_and_lock_list(inode);
308 }
309 
310 struct inode_switch_wbs_context {
311         struct inode            *inode;
312         struct bdi_writeback    *new_wb;
313 
314         struct rcu_head         rcu_head;
315         struct work_struct      work;
316 };
317 
318 static void inode_switch_wbs_work_fn(struct work_struct *work)
319 {
320         struct inode_switch_wbs_context *isw =
321                 container_of(work, struct inode_switch_wbs_context, work);
322         struct inode *inode = isw->inode;
323         struct address_space *mapping = inode->i_mapping;
324         struct bdi_writeback *old_wb = inode->i_wb;
325         struct bdi_writeback *new_wb = isw->new_wb;
326         struct radix_tree_iter iter;
327         bool switched = false;
328         void **slot;
329 
330         /*
331          * By the time control reaches here, RCU grace period has passed
332          * since I_WB_SWITCH assertion and all wb stat update transactions
333          * between unlocked_inode_to_wb_begin/end() are guaranteed to be
334          * synchronizing against mapping->tree_lock.
335          *
336          * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock
337          * gives us exclusion against all wb related operations on @inode
338          * including IO list manipulations and stat updates.
339          */
340         if (old_wb < new_wb) {
341                 spin_lock(&old_wb->list_lock);
342                 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
343         } else {
344                 spin_lock(&new_wb->list_lock);
345                 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
346         }
347         spin_lock(&inode->i_lock);
348         spin_lock_irq(&mapping->tree_lock);
349 
350         /*
351          * Once I_FREEING is visible under i_lock, the eviction path owns
352          * the inode and we shouldn't modify ->i_io_list.
353          */
354         if (unlikely(inode->i_state & I_FREEING))
355                 goto skip_switch;
356 
357         /*
358          * Count and transfer stats.  Note that PAGECACHE_TAG_DIRTY points
359          * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
360          * pages actually under underwriteback.
361          */
362         radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
363                                    PAGECACHE_TAG_DIRTY) {
364                 struct page *page = radix_tree_deref_slot_protected(slot,
365                                                         &mapping->tree_lock);
366                 if (likely(page) && PageDirty(page)) {
367                         __dec_wb_stat(old_wb, WB_RECLAIMABLE);
368                         __inc_wb_stat(new_wb, WB_RECLAIMABLE);
369                 }
370         }
371 
372         radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
373                                    PAGECACHE_TAG_WRITEBACK) {
374                 struct page *page = radix_tree_deref_slot_protected(slot,
375                                                         &mapping->tree_lock);
376                 if (likely(page)) {
377                         WARN_ON_ONCE(!PageWriteback(page));
378                         __dec_wb_stat(old_wb, WB_WRITEBACK);
379                         __inc_wb_stat(new_wb, WB_WRITEBACK);
380                 }
381         }
382 
383         wb_get(new_wb);
384 
385         /*
386          * Transfer to @new_wb's IO list if necessary.  The specific list
387          * @inode was on is ignored and the inode is put on ->b_dirty which
388          * is always correct including from ->b_dirty_time.  The transfer
389          * preserves @inode->dirtied_when ordering.
390          */
391         if (!list_empty(&inode->i_io_list)) {
392                 struct inode *pos;
393 
394                 inode_io_list_del_locked(inode, old_wb);
395                 inode->i_wb = new_wb;
396                 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
397                         if (time_after_eq(inode->dirtied_when,
398                                           pos->dirtied_when))
399                                 break;
400                 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
401         } else {
402                 inode->i_wb = new_wb;
403         }
404 
405         /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
406         inode->i_wb_frn_winner = 0;
407         inode->i_wb_frn_avg_time = 0;
408         inode->i_wb_frn_history = 0;
409         switched = true;
410 skip_switch:
411         /*
412          * Paired with load_acquire in unlocked_inode_to_wb_begin() and
413          * ensures that the new wb is visible if they see !I_WB_SWITCH.
414          */
415         smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
416 
417         spin_unlock_irq(&mapping->tree_lock);
418         spin_unlock(&inode->i_lock);
419         spin_unlock(&new_wb->list_lock);
420         spin_unlock(&old_wb->list_lock);
421 
422         if (switched) {
423                 wb_wakeup(new_wb);
424                 wb_put(old_wb);
425         }
426         wb_put(new_wb);
427 
428         iput(inode);
429         kfree(isw);
430 
431         atomic_dec(&isw_nr_in_flight);
432 }
433 
434 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
435 {
436         struct inode_switch_wbs_context *isw = container_of(rcu_head,
437                                 struct inode_switch_wbs_context, rcu_head);
438 
439         /* needs to grab bh-unsafe locks, bounce to work item */
440         INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
441         queue_work(isw_wq, &isw->work);
442 }
443 
444 /**
445  * inode_switch_wbs - change the wb association of an inode
446  * @inode: target inode
447  * @new_wb_id: ID of the new wb
448  *
449  * Switch @inode's wb association to the wb identified by @new_wb_id.  The
450  * switching is performed asynchronously and may fail silently.
451  */
452 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
453 {
454         struct backing_dev_info *bdi = inode_to_bdi(inode);
455         struct cgroup_subsys_state *memcg_css;
456         struct inode_switch_wbs_context *isw;
457 
458         /* noop if seems to be already in progress */
459         if (inode->i_state & I_WB_SWITCH)
460                 return;
461 
462         isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
463         if (!isw)
464                 return;
465 
466         /* find and pin the new wb */
467         rcu_read_lock();
468         memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
469         if (memcg_css)
470                 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
471         rcu_read_unlock();
472         if (!isw->new_wb)
473                 goto out_free;
474 
475         /* while holding I_WB_SWITCH, no one else can update the association */
476         spin_lock(&inode->i_lock);
477         if (!(inode->i_sb->s_flags & MS_ACTIVE) ||
478             inode->i_state & (I_WB_SWITCH | I_FREEING) ||
479             inode_to_wb(inode) == isw->new_wb) {
480                 spin_unlock(&inode->i_lock);
481                 goto out_free;
482         }
483         inode->i_state |= I_WB_SWITCH;
484         spin_unlock(&inode->i_lock);
485 
486         ihold(inode);
487         isw->inode = inode;
488 
489         atomic_inc(&isw_nr_in_flight);
490 
491         /*
492          * In addition to synchronizing among switchers, I_WB_SWITCH tells
493          * the RCU protected stat update paths to grab the mapping's
494          * tree_lock so that stat transfer can synchronize against them.
495          * Let's continue after I_WB_SWITCH is guaranteed to be visible.
496          */
497         call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
498         return;
499 
500 out_free:
501         if (isw->new_wb)
502                 wb_put(isw->new_wb);
503         kfree(isw);
504 }
505 
506 /**
507  * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
508  * @wbc: writeback_control of interest
509  * @inode: target inode
510  *
511  * @inode is locked and about to be written back under the control of @wbc.
512  * Record @inode's writeback context into @wbc and unlock the i_lock.  On
513  * writeback completion, wbc_detach_inode() should be called.  This is used
514  * to track the cgroup writeback context.
515  */
516 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
517                                  struct inode *inode)
518 {
519         if (!inode_cgwb_enabled(inode)) {
520                 spin_unlock(&inode->i_lock);
521                 return;
522         }
523 
524         wbc->wb = inode_to_wb(inode);
525         wbc->inode = inode;
526 
527         wbc->wb_id = wbc->wb->memcg_css->id;
528         wbc->wb_lcand_id = inode->i_wb_frn_winner;
529         wbc->wb_tcand_id = 0;
530         wbc->wb_bytes = 0;
531         wbc->wb_lcand_bytes = 0;
532         wbc->wb_tcand_bytes = 0;
533 
534         wb_get(wbc->wb);
535         spin_unlock(&inode->i_lock);
536 
537         /*
538          * A dying wb indicates that the memcg-blkcg mapping has changed
539          * and a new wb is already serving the memcg.  Switch immediately.
540          */
541         if (unlikely(wb_dying(wbc->wb)))
542                 inode_switch_wbs(inode, wbc->wb_id);
543 }
544 
545 /**
546  * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
547  * @wbc: writeback_control of the just finished writeback
548  *
549  * To be called after a writeback attempt of an inode finishes and undoes
550  * wbc_attach_and_unlock_inode().  Can be called under any context.
551  *
552  * As concurrent write sharing of an inode is expected to be very rare and
553  * memcg only tracks page ownership on first-use basis severely confining
554  * the usefulness of such sharing, cgroup writeback tracks ownership
555  * per-inode.  While the support for concurrent write sharing of an inode
556  * is deemed unnecessary, an inode being written to by different cgroups at
557  * different points in time is a lot more common, and, more importantly,
558  * charging only by first-use can too readily lead to grossly incorrect
559  * behaviors (single foreign page can lead to gigabytes of writeback to be
560  * incorrectly attributed).
561  *
562  * To resolve this issue, cgroup writeback detects the majority dirtier of
563  * an inode and transfers the ownership to it.  To avoid unnnecessary
564  * oscillation, the detection mechanism keeps track of history and gives
565  * out the switch verdict only if the foreign usage pattern is stable over
566  * a certain amount of time and/or writeback attempts.
567  *
568  * On each writeback attempt, @wbc tries to detect the majority writer
569  * using Boyer-Moore majority vote algorithm.  In addition to the byte
570  * count from the majority voting, it also counts the bytes written for the
571  * current wb and the last round's winner wb (max of last round's current
572  * wb, the winner from two rounds ago, and the last round's majority
573  * candidate).  Keeping track of the historical winner helps the algorithm
574  * to semi-reliably detect the most active writer even when it's not the
575  * absolute majority.
576  *
577  * Once the winner of the round is determined, whether the winner is
578  * foreign or not and how much IO time the round consumed is recorded in
579  * inode->i_wb_frn_history.  If the amount of recorded foreign IO time is
580  * over a certain threshold, the switch verdict is given.
581  */
582 void wbc_detach_inode(struct writeback_control *wbc)
583 {
584         struct bdi_writeback *wb = wbc->wb;
585         struct inode *inode = wbc->inode;
586         unsigned long avg_time, max_bytes, max_time;
587         u16 history;
588         int max_id;
589 
590         if (!wb)
591                 return;
592 
593         history = inode->i_wb_frn_history;
594         avg_time = inode->i_wb_frn_avg_time;
595 
596         /* pick the winner of this round */
597         if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
598             wbc->wb_bytes >= wbc->wb_tcand_bytes) {
599                 max_id = wbc->wb_id;
600                 max_bytes = wbc->wb_bytes;
601         } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
602                 max_id = wbc->wb_lcand_id;
603                 max_bytes = wbc->wb_lcand_bytes;
604         } else {
605                 max_id = wbc->wb_tcand_id;
606                 max_bytes = wbc->wb_tcand_bytes;
607         }
608 
609         /*
610          * Calculate the amount of IO time the winner consumed and fold it
611          * into the running average kept per inode.  If the consumed IO
612          * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
613          * deciding whether to switch or not.  This is to prevent one-off
614          * small dirtiers from skewing the verdict.
615          */
616         max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
617                                 wb->avg_write_bandwidth);
618         if (avg_time)
619                 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
620                             (avg_time >> WB_FRN_TIME_AVG_SHIFT);
621         else
622                 avg_time = max_time;    /* immediate catch up on first run */
623 
624         if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
625                 int slots;
626 
627                 /*
628                  * The switch verdict is reached if foreign wb's consume
629                  * more than a certain proportion of IO time in a
630                  * WB_FRN_TIME_PERIOD.  This is loosely tracked by 16 slot
631                  * history mask where each bit represents one sixteenth of
632                  * the period.  Determine the number of slots to shift into
633                  * history from @max_time.
634                  */
635                 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
636                             (unsigned long)WB_FRN_HIST_MAX_SLOTS);
637                 history <<= slots;
638                 if (wbc->wb_id != max_id)
639                         history |= (1U << slots) - 1;
640 
641                 /*
642                  * Switch if the current wb isn't the consistent winner.
643                  * If there are multiple closely competing dirtiers, the
644                  * inode may switch across them repeatedly over time, which
645                  * is okay.  The main goal is avoiding keeping an inode on
646                  * the wrong wb for an extended period of time.
647                  */
648                 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
649                         inode_switch_wbs(inode, max_id);
650         }
651 
652         /*
653          * Multiple instances of this function may race to update the
654          * following fields but we don't mind occassional inaccuracies.
655          */
656         inode->i_wb_frn_winner = max_id;
657         inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
658         inode->i_wb_frn_history = history;
659 
660         wb_put(wbc->wb);
661         wbc->wb = NULL;
662 }
663 
664 /**
665  * wbc_account_io - account IO issued during writeback
666  * @wbc: writeback_control of the writeback in progress
667  * @page: page being written out
668  * @bytes: number of bytes being written out
669  *
670  * @bytes from @page are about to written out during the writeback
671  * controlled by @wbc.  Keep the book for foreign inode detection.  See
672  * wbc_detach_inode().
673  */
674 void wbc_account_io(struct writeback_control *wbc, struct page *page,
675                     size_t bytes)
676 {
677         int id;
678 
679         /*
680          * pageout() path doesn't attach @wbc to the inode being written
681          * out.  This is intentional as we don't want the function to block
682          * behind a slow cgroup.  Ultimately, we want pageout() to kick off
683          * regular writeback instead of writing things out itself.
684          */
685         if (!wbc->wb)
686                 return;
687 
688         id = mem_cgroup_css_from_page(page)->id;
689 
690         if (id == wbc->wb_id) {
691                 wbc->wb_bytes += bytes;
692                 return;
693         }
694 
695         if (id == wbc->wb_lcand_id)
696                 wbc->wb_lcand_bytes += bytes;
697 
698         /* Boyer-Moore majority vote algorithm */
699         if (!wbc->wb_tcand_bytes)
700                 wbc->wb_tcand_id = id;
701         if (id == wbc->wb_tcand_id)
702                 wbc->wb_tcand_bytes += bytes;
703         else
704                 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
705 }
706 EXPORT_SYMBOL_GPL(wbc_account_io);
707 
708 /**
709  * inode_congested - test whether an inode is congested
710  * @inode: inode to test for congestion (may be NULL)
711  * @cong_bits: mask of WB_[a]sync_congested bits to test
712  *
713  * Tests whether @inode is congested.  @cong_bits is the mask of congestion
714  * bits to test and the return value is the mask of set bits.
715  *
716  * If cgroup writeback is enabled for @inode, the congestion state is
717  * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
718  * associated with @inode is congested; otherwise, the root wb's congestion
719  * state is used.
720  *
721  * @inode is allowed to be NULL as this function is often called on
722  * mapping->host which is NULL for the swapper space.
723  */
724 int inode_congested(struct inode *inode, int cong_bits)
725 {
726         /*
727          * Once set, ->i_wb never becomes NULL while the inode is alive.
728          * Start transaction iff ->i_wb is visible.
729          */
730         if (inode && inode_to_wb_is_valid(inode)) {
731                 struct bdi_writeback *wb;
732                 bool locked, congested;
733 
734                 wb = unlocked_inode_to_wb_begin(inode, &locked);
735                 congested = wb_congested(wb, cong_bits);
736                 unlocked_inode_to_wb_end(inode, locked);
737                 return congested;
738         }
739 
740         return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
741 }
742 EXPORT_SYMBOL_GPL(inode_congested);
743 
744 /**
745  * wb_split_bdi_pages - split nr_pages to write according to bandwidth
746  * @wb: target bdi_writeback to split @nr_pages to
747  * @nr_pages: number of pages to write for the whole bdi
748  *
749  * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
750  * relation to the total write bandwidth of all wb's w/ dirty inodes on
751  * @wb->bdi.
752  */
753 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
754 {
755         unsigned long this_bw = wb->avg_write_bandwidth;
756         unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
757 
758         if (nr_pages == LONG_MAX)
759                 return LONG_MAX;
760 
761         /*
762          * This may be called on clean wb's and proportional distribution
763          * may not make sense, just use the original @nr_pages in those
764          * cases.  In general, we wanna err on the side of writing more.
765          */
766         if (!tot_bw || this_bw >= tot_bw)
767                 return nr_pages;
768         else
769                 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
770 }
771 
772 /**
773  * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
774  * @bdi: target backing_dev_info
775  * @base_work: wb_writeback_work to issue
776  * @skip_if_busy: skip wb's which already have writeback in progress
777  *
778  * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
779  * have dirty inodes.  If @base_work->nr_page isn't %LONG_MAX, it's
780  * distributed to the busy wbs according to each wb's proportion in the
781  * total active write bandwidth of @bdi.
782  */
783 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
784                                   struct wb_writeback_work *base_work,
785                                   bool skip_if_busy)
786 {
787         struct bdi_writeback *last_wb = NULL;
788         struct bdi_writeback *wb = list_entry(&bdi->wb_list,
789                                               struct bdi_writeback, bdi_node);
790 
791         might_sleep();
792 restart:
793         rcu_read_lock();
794         list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
795                 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done);
796                 struct wb_writeback_work fallback_work;
797                 struct wb_writeback_work *work;
798                 long nr_pages;
799 
800                 if (last_wb) {
801                         wb_put(last_wb);
802                         last_wb = NULL;
803                 }
804 
805                 /* SYNC_ALL writes out I_DIRTY_TIME too */
806                 if (!wb_has_dirty_io(wb) &&
807                     (base_work->sync_mode == WB_SYNC_NONE ||
808                      list_empty(&wb->b_dirty_time)))
809                         continue;
810                 if (skip_if_busy && writeback_in_progress(wb))
811                         continue;
812 
813                 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
814 
815                 work = kmalloc(sizeof(*work), GFP_ATOMIC);
816                 if (work) {
817                         *work = *base_work;
818                         work->nr_pages = nr_pages;
819                         work->auto_free = 1;
820                         wb_queue_work(wb, work);
821                         continue;
822                 }
823 
824                 /* alloc failed, execute synchronously using on-stack fallback */
825                 work = &fallback_work;
826                 *work = *base_work;
827                 work->nr_pages = nr_pages;
828                 work->auto_free = 0;
829                 work->done = &fallback_work_done;
830 
831                 wb_queue_work(wb, work);
832 
833                 /*
834                  * Pin @wb so that it stays on @bdi->wb_list.  This allows
835                  * continuing iteration from @wb after dropping and
836                  * regrabbing rcu read lock.
837                  */
838                 wb_get(wb);
839                 last_wb = wb;
840 
841                 rcu_read_unlock();
842                 wb_wait_for_completion(bdi, &fallback_work_done);
843                 goto restart;
844         }
845         rcu_read_unlock();
846 
847         if (last_wb)
848                 wb_put(last_wb);
849 }
850 
851 /**
852  * cgroup_writeback_umount - flush inode wb switches for umount
853  *
854  * This function is called when a super_block is about to be destroyed and
855  * flushes in-flight inode wb switches.  An inode wb switch goes through
856  * RCU and then workqueue, so the two need to be flushed in order to ensure
857  * that all previously scheduled switches are finished.  As wb switches are
858  * rare occurrences and synchronize_rcu() can take a while, perform
859  * flushing iff wb switches are in flight.
860  */
861 void cgroup_writeback_umount(void)
862 {
863         if (atomic_read(&isw_nr_in_flight)) {
864                 synchronize_rcu();
865                 flush_workqueue(isw_wq);
866         }
867 }
868 
869 static int __init cgroup_writeback_init(void)
870 {
871         isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
872         if (!isw_wq)
873                 return -ENOMEM;
874         return 0;
875 }
876 fs_initcall(cgroup_writeback_init);
877 
878 #else   /* CONFIG_CGROUP_WRITEBACK */
879 
880 static struct bdi_writeback *
881 locked_inode_to_wb_and_lock_list(struct inode *inode)
882         __releases(&inode->i_lock)
883         __acquires(&wb->list_lock)
884 {
885         struct bdi_writeback *wb = inode_to_wb(inode);
886 
887         spin_unlock(&inode->i_lock);
888         spin_lock(&wb->list_lock);
889         return wb;
890 }
891 
892 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
893         __acquires(&wb->list_lock)
894 {
895         struct bdi_writeback *wb = inode_to_wb(inode);
896 
897         spin_lock(&wb->list_lock);
898         return wb;
899 }
900 
901 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
902 {
903         return nr_pages;
904 }
905 
906 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
907                                   struct wb_writeback_work *base_work,
908                                   bool skip_if_busy)
909 {
910         might_sleep();
911 
912         if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
913                 base_work->auto_free = 0;
914                 wb_queue_work(&bdi->wb, base_work);
915         }
916 }
917 
918 #endif  /* CONFIG_CGROUP_WRITEBACK */
919 
920 void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
921                         bool range_cyclic, enum wb_reason reason)
922 {
923         struct wb_writeback_work *work;
924 
925         if (!wb_has_dirty_io(wb))
926                 return;
927 
928         /*
929          * This is WB_SYNC_NONE writeback, so if allocation fails just
930          * wakeup the thread for old dirty data writeback
931          */
932         work = kzalloc(sizeof(*work), GFP_ATOMIC);
933         if (!work) {
934                 trace_writeback_nowork(wb);
935                 wb_wakeup(wb);
936                 return;
937         }
938 
939         work->sync_mode = WB_SYNC_NONE;
940         work->nr_pages  = nr_pages;
941         work->range_cyclic = range_cyclic;
942         work->reason    = reason;
943         work->auto_free = 1;
944 
945         wb_queue_work(wb, work);
946 }
947 
948 /**
949  * wb_start_background_writeback - start background writeback
950  * @wb: bdi_writback to write from
951  *
952  * Description:
953  *   This makes sure WB_SYNC_NONE background writeback happens. When
954  *   this function returns, it is only guaranteed that for given wb
955  *   some IO is happening if we are over background dirty threshold.
956  *   Caller need not hold sb s_umount semaphore.
957  */
958 void wb_start_background_writeback(struct bdi_writeback *wb)
959 {
960         /*
961          * We just wake up the flusher thread. It will perform background
962          * writeback as soon as there is no other work to do.
963          */
964         trace_writeback_wake_background(wb);
965         wb_wakeup(wb);
966 }
967 
968 /*
969  * Remove the inode from the writeback list it is on.
970  */
971 void inode_io_list_del(struct inode *inode)
972 {
973         struct bdi_writeback *wb;
974 
975         wb = inode_to_wb_and_lock_list(inode);
976         inode_io_list_del_locked(inode, wb);
977         spin_unlock(&wb->list_lock);
978 }
979 
980 /*
981  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
982  * furthest end of its superblock's dirty-inode list.
983  *
984  * Before stamping the inode's ->dirtied_when, we check to see whether it is
985  * already the most-recently-dirtied inode on the b_dirty list.  If that is
986  * the case then the inode must have been redirtied while it was being written
987  * out and we don't reset its dirtied_when.
988  */
989 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
990 {
991         if (!list_empty(&wb->b_dirty)) {
992                 struct inode *tail;
993 
994                 tail = wb_inode(wb->b_dirty.next);
995                 if (time_before(inode->dirtied_when, tail->dirtied_when))
996                         inode->dirtied_when = jiffies;
997         }
998         inode_io_list_move_locked(inode, wb, &wb->b_dirty);
999 }
1000 
1001 /*
1002  * requeue inode for re-scanning after bdi->b_io list is exhausted.
1003  */
1004 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1005 {
1006         inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1007 }
1008 
1009 static void inode_sync_complete(struct inode *inode)
1010 {
1011         inode->i_state &= ~I_SYNC;
1012         /* If inode is clean an unused, put it into LRU now... */
1013         inode_add_lru(inode);
1014         /* Waiters must see I_SYNC cleared before being woken up */
1015         smp_mb();
1016         wake_up_bit(&inode->i_state, __I_SYNC);
1017 }
1018 
1019 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1020 {
1021         bool ret = time_after(inode->dirtied_when, t);
1022 #ifndef CONFIG_64BIT
1023         /*
1024          * For inodes being constantly redirtied, dirtied_when can get stuck.
1025          * It _appears_ to be in the future, but is actually in distant past.
1026          * This test is necessary to prevent such wrapped-around relative times
1027          * from permanently stopping the whole bdi writeback.
1028          */
1029         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1030 #endif
1031         return ret;
1032 }
1033 
1034 #define EXPIRE_DIRTY_ATIME 0x0001
1035 
1036 /*
1037  * Move expired (dirtied before work->older_than_this) dirty inodes from
1038  * @delaying_queue to @dispatch_queue.
1039  */
1040 static int move_expired_inodes(struct list_head *delaying_queue,
1041                                struct list_head *dispatch_queue,
1042                                int flags,
1043                                struct wb_writeback_work *work)
1044 {
1045         unsigned long *older_than_this = NULL;
1046         unsigned long expire_time;
1047         LIST_HEAD(tmp);
1048         struct list_head *pos, *node;
1049         struct super_block *sb = NULL;
1050         struct inode *inode;
1051         int do_sb_sort = 0;
1052         int moved = 0;
1053 
1054         if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1055                 older_than_this = work->older_than_this;
1056         else if (!work->for_sync) {
1057                 expire_time = jiffies - (dirtytime_expire_interval * HZ);
1058                 older_than_this = &expire_time;
1059         }
1060         while (!list_empty(delaying_queue)) {
1061                 inode = wb_inode(delaying_queue->prev);
1062                 if (older_than_this &&
1063                     inode_dirtied_after(inode, *older_than_this))
1064                         break;
1065                 list_move(&inode->i_io_list, &tmp);
1066                 moved++;
1067                 if (flags & EXPIRE_DIRTY_ATIME)
1068                         set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
1069                 if (sb_is_blkdev_sb(inode->i_sb))
1070                         continue;
1071                 if (sb && sb != inode->i_sb)
1072                         do_sb_sort = 1;
1073                 sb = inode->i_sb;
1074         }
1075 
1076         /* just one sb in list, splice to dispatch_queue and we're done */
1077         if (!do_sb_sort) {
1078                 list_splice(&tmp, dispatch_queue);
1079                 goto out;
1080         }
1081 
1082         /* Move inodes from one superblock together */
1083         while (!list_empty(&tmp)) {
1084                 sb = wb_inode(tmp.prev)->i_sb;
1085                 list_for_each_prev_safe(pos, node, &tmp) {
1086                         inode = wb_inode(pos);
1087                         if (inode->i_sb == sb)
1088                                 list_move(&inode->i_io_list, dispatch_queue);
1089                 }
1090         }
1091 out:
1092         return moved;
1093 }
1094 
1095 /*
1096  * Queue all expired dirty inodes for io, eldest first.
1097  * Before
1098  *         newly dirtied     b_dirty    b_io    b_more_io
1099  *         =============>    gf         edc     BA
1100  * After
1101  *         newly dirtied     b_dirty    b_io    b_more_io
1102  *         =============>    g          fBAedc
1103  *                                           |
1104  *                                           +--> dequeue for IO
1105  */
1106 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
1107 {
1108         int moved;
1109 
1110         assert_spin_locked(&wb->list_lock);
1111         list_splice_init(&wb->b_more_io, &wb->b_io);
1112         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1113         moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1114                                      EXPIRE_DIRTY_ATIME, work);
1115         if (moved)
1116                 wb_io_lists_populated(wb);
1117         trace_writeback_queue_io(wb, work, moved);
1118 }
1119 
1120 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1121 {
1122         int ret;
1123 
1124         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1125                 trace_writeback_write_inode_start(inode, wbc);
1126                 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1127                 trace_writeback_write_inode(inode, wbc);
1128                 return ret;
1129         }
1130         return 0;
1131 }
1132 
1133 /*
1134  * Wait for writeback on an inode to complete. Called with i_lock held.
1135  * Caller must make sure inode cannot go away when we drop i_lock.
1136  */
1137 static void __inode_wait_for_writeback(struct inode *inode)
1138         __releases(inode->i_lock)
1139         __acquires(inode->i_lock)
1140 {
1141         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1142         wait_queue_head_t *wqh;
1143 
1144         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1145         while (inode->i_state & I_SYNC) {
1146                 spin_unlock(&inode->i_lock);
1147                 __wait_on_bit(wqh, &wq, bit_wait,
1148                               TASK_UNINTERRUPTIBLE);
1149                 spin_lock(&inode->i_lock);
1150         }
1151 }
1152 
1153 /*
1154  * Wait for writeback on an inode to complete. Caller must have inode pinned.
1155  */
1156 void inode_wait_for_writeback(struct inode *inode)
1157 {
1158         spin_lock(&inode->i_lock);
1159         __inode_wait_for_writeback(inode);
1160         spin_unlock(&inode->i_lock);
1161 }
1162 
1163 /*
1164  * Sleep until I_SYNC is cleared. This function must be called with i_lock
1165  * held and drops it. It is aimed for callers not holding any inode reference
1166  * so once i_lock is dropped, inode can go away.
1167  */
1168 static void inode_sleep_on_writeback(struct inode *inode)
1169         __releases(inode->i_lock)
1170 {
1171         DEFINE_WAIT(wait);
1172         wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1173         int sleep;
1174 
1175         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1176         sleep = inode->i_state & I_SYNC;
1177         spin_unlock(&inode->i_lock);
1178         if (sleep)
1179                 schedule();
1180         finish_wait(wqh, &wait);
1181 }
1182 
1183 /*
1184  * Find proper writeback list for the inode depending on its current state and
1185  * possibly also change of its state while we were doing writeback.  Here we
1186  * handle things such as livelock prevention or fairness of writeback among
1187  * inodes. This function can be called only by flusher thread - noone else
1188  * processes all inodes in writeback lists and requeueing inodes behind flusher
1189  * thread's back can have unexpected consequences.
1190  */
1191 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1192                           struct writeback_control *wbc)
1193 {
1194         if (inode->i_state & I_FREEING)
1195                 return;
1196 
1197         /*
1198          * Sync livelock prevention. Each inode is tagged and synced in one
1199          * shot. If still dirty, it will be redirty_tail()'ed below.  Update
1200          * the dirty time to prevent enqueue and sync it again.
1201          */
1202         if ((inode->i_state & I_DIRTY) &&
1203             (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1204                 inode->dirtied_when = jiffies;
1205 
1206         if (wbc->pages_skipped) {
1207                 /*
1208                  * writeback is not making progress due to locked
1209                  * buffers. Skip this inode for now.
1210                  */
1211                 redirty_tail(inode, wb);
1212                 return;
1213         }
1214 
1215         if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1216                 /*
1217                  * We didn't write back all the pages.  nfs_writepages()
1218                  * sometimes bales out without doing anything.
1219                  */
1220                 if (wbc->nr_to_write <= 0) {
1221                         /* Slice used up. Queue for next turn. */
1222                         requeue_io(inode, wb);
1223                 } else {
1224                         /*
1225                          * Writeback blocked by something other than
1226                          * congestion. Delay the inode for some time to
1227                          * avoid spinning on the CPU (100% iowait)
1228                          * retrying writeback of the dirty page/inode
1229                          * that cannot be performed immediately.
1230                          */
1231                         redirty_tail(inode, wb);
1232                 }
1233         } else if (inode->i_state & I_DIRTY) {
1234                 /*
1235                  * Filesystems can dirty the inode during writeback operations,
1236                  * such as delayed allocation during submission or metadata
1237                  * updates after data IO completion.
1238                  */
1239                 redirty_tail(inode, wb);
1240         } else if (inode->i_state & I_DIRTY_TIME) {
1241                 inode->dirtied_when = jiffies;
1242                 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1243         } else {
1244                 /* The inode is clean. Remove from writeback lists. */
1245                 inode_io_list_del_locked(inode, wb);
1246         }
1247 }
1248 
1249 /*
1250  * Write out an inode and its dirty pages. Do not update the writeback list
1251  * linkage. That is left to the caller. The caller is also responsible for
1252  * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1253  */
1254 static int
1255 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1256 {
1257         struct address_space *mapping = inode->i_mapping;
1258         long nr_to_write = wbc->nr_to_write;
1259         unsigned dirty;
1260         int ret;
1261 
1262         WARN_ON(!(inode->i_state & I_SYNC));
1263 
1264         trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1265 
1266         ret = do_writepages(mapping, wbc);
1267 
1268         /*
1269          * Make sure to wait on the data before writing out the metadata.
1270          * This is important for filesystems that modify metadata on data
1271          * I/O completion. We don't do it for sync(2) writeback because it has a
1272          * separate, external IO completion path and ->sync_fs for guaranteeing
1273          * inode metadata is written back correctly.
1274          */
1275         if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1276                 int err = filemap_fdatawait(mapping);
1277                 if (ret == 0)
1278                         ret = err;
1279         }
1280 
1281         /*
1282          * Some filesystems may redirty the inode during the writeback
1283          * due to delalloc, clear dirty metadata flags right before
1284          * write_inode()
1285          */
1286         spin_lock(&inode->i_lock);
1287 
1288         dirty = inode->i_state & I_DIRTY;
1289         if (inode->i_state & I_DIRTY_TIME) {
1290                 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1291                     unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1292                     unlikely(time_after(jiffies,
1293                                         (inode->dirtied_time_when +
1294                                          dirtytime_expire_interval * HZ)))) {
1295                         dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1296                         trace_writeback_lazytime(inode);
1297                 }
1298         } else
1299                 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1300         inode->i_state &= ~dirty;
1301 
1302         /*
1303          * Paired with smp_mb() in __mark_inode_dirty().  This allows
1304          * __mark_inode_dirty() to test i_state without grabbing i_lock -
1305          * either they see the I_DIRTY bits cleared or we see the dirtied
1306          * inode.
1307          *
1308          * I_DIRTY_PAGES is always cleared together above even if @mapping
1309          * still has dirty pages.  The flag is reinstated after smp_mb() if
1310          * necessary.  This guarantees that either __mark_inode_dirty()
1311          * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1312          */
1313         smp_mb();
1314 
1315         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1316                 inode->i_state |= I_DIRTY_PAGES;
1317 
1318         spin_unlock(&inode->i_lock);
1319 
1320         if (dirty & I_DIRTY_TIME)
1321                 mark_inode_dirty_sync(inode);
1322         /* Don't write the inode if only I_DIRTY_PAGES was set */
1323         if (dirty & ~I_DIRTY_PAGES) {
1324                 int err = write_inode(inode, wbc);
1325                 if (ret == 0)
1326                         ret = err;
1327         }
1328         trace_writeback_single_inode(inode, wbc, nr_to_write);
1329         return ret;
1330 }
1331 
1332 /*
1333  * Write out an inode's dirty pages. Either the caller has an active reference
1334  * on the inode or the inode has I_WILL_FREE set.
1335  *
1336  * This function is designed to be called for writing back one inode which
1337  * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1338  * and does more profound writeback list handling in writeback_sb_inodes().
1339  */
1340 static int
1341 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb,
1342                        struct writeback_control *wbc)
1343 {
1344         int ret = 0;
1345 
1346         spin_lock(&inode->i_lock);
1347         if (!atomic_read(&inode->i_count))
1348                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1349         else
1350                 WARN_ON(inode->i_state & I_WILL_FREE);
1351 
1352         if (inode->i_state & I_SYNC) {
1353                 if (wbc->sync_mode != WB_SYNC_ALL)
1354                         goto out;
1355                 /*
1356                  * It's a data-integrity sync. We must wait. Since callers hold
1357                  * inode reference or inode has I_WILL_FREE set, it cannot go
1358                  * away under us.
1359                  */
1360                 __inode_wait_for_writeback(inode);
1361         }
1362         WARN_ON(inode->i_state & I_SYNC);
1363         /*
1364          * Skip inode if it is clean and we have no outstanding writeback in
1365          * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1366          * function since flusher thread may be doing for example sync in
1367          * parallel and if we move the inode, it could get skipped. So here we
1368          * make sure inode is on some writeback list and leave it there unless
1369          * we have completely cleaned the inode.
1370          */
1371         if (!(inode->i_state & I_DIRTY_ALL) &&
1372             (wbc->sync_mode != WB_SYNC_ALL ||
1373              !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1374                 goto out;
1375         inode->i_state |= I_SYNC;
1376         wbc_attach_and_unlock_inode(wbc, inode);
1377 
1378         ret = __writeback_single_inode(inode, wbc);
1379 
1380         wbc_detach_inode(wbc);
1381         spin_lock(&wb->list_lock);
1382         spin_lock(&inode->i_lock);
1383         /*
1384          * If inode is clean, remove it from writeback lists. Otherwise don't
1385          * touch it. See comment above for explanation.
1386          */
1387         if (!(inode->i_state & I_DIRTY_ALL))
1388                 inode_io_list_del_locked(inode, wb);
1389         spin_unlock(&wb->list_lock);
1390         inode_sync_complete(inode);
1391 out:
1392         spin_unlock(&inode->i_lock);
1393         return ret;
1394 }
1395 
1396 static long writeback_chunk_size(struct bdi_writeback *wb,
1397                                  struct wb_writeback_work *work)
1398 {
1399         long pages;
1400 
1401         /*
1402          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1403          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1404          * here avoids calling into writeback_inodes_wb() more than once.
1405          *
1406          * The intended call sequence for WB_SYNC_ALL writeback is:
1407          *
1408          *      wb_writeback()
1409          *          writeback_sb_inodes()       <== called only once
1410          *              write_cache_pages()     <== called once for each inode
1411          *                   (quickly) tag currently dirty pages
1412          *                   (maybe slowly) sync all tagged pages
1413          */
1414         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1415                 pages = LONG_MAX;
1416         else {
1417                 pages = min(wb->avg_write_bandwidth / 2,
1418                             global_wb_domain.dirty_limit / DIRTY_SCOPE);
1419                 pages = min(pages, work->nr_pages);
1420                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1421                                    MIN_WRITEBACK_PAGES);
1422         }
1423 
1424         return pages;
1425 }
1426 
1427 /*
1428  * Write a portion of b_io inodes which belong to @sb.
1429  *
1430  * Return the number of pages and/or inodes written.
1431  *
1432  * NOTE! This is called with wb->list_lock held, and will
1433  * unlock and relock that for each inode it ends up doing
1434  * IO for.
1435  */
1436 static long writeback_sb_inodes(struct super_block *sb,
1437                                 struct bdi_writeback *wb,
1438                                 struct wb_writeback_work *work)
1439 {
1440         struct writeback_control wbc = {
1441                 .sync_mode              = work->sync_mode,
1442                 .tagged_writepages      = work->tagged_writepages,
1443                 .for_kupdate            = work->for_kupdate,
1444                 .for_background         = work->for_background,
1445                 .for_sync               = work->for_sync,
1446                 .range_cyclic           = work->range_cyclic,
1447                 .range_start            = 0,
1448                 .range_end              = LLONG_MAX,
1449         };
1450         unsigned long start_time = jiffies;
1451         long write_chunk;
1452         long wrote = 0;  /* count both pages and inodes */
1453 
1454         while (!list_empty(&wb->b_io)) {
1455                 struct inode *inode = wb_inode(wb->b_io.prev);
1456 
1457                 if (inode->i_sb != sb) {
1458                         if (work->sb) {
1459                                 /*
1460                                  * We only want to write back data for this
1461                                  * superblock, move all inodes not belonging
1462                                  * to it back onto the dirty list.
1463                                  */
1464                                 redirty_tail(inode, wb);
1465                                 continue;
1466                         }
1467 
1468                         /*
1469                          * The inode belongs to a different superblock.
1470                          * Bounce back to the caller to unpin this and
1471                          * pin the next superblock.
1472                          */
1473                         break;
1474                 }
1475 
1476                 /*
1477                  * Don't bother with new inodes or inodes being freed, first
1478                  * kind does not need periodic writeout yet, and for the latter
1479                  * kind writeout is handled by the freer.
1480                  */
1481                 spin_lock(&inode->i_lock);
1482                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1483                         spin_unlock(&inode->i_lock);
1484                         redirty_tail(inode, wb);
1485                         continue;
1486                 }
1487                 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1488                         /*
1489                          * If this inode is locked for writeback and we are not
1490                          * doing writeback-for-data-integrity, move it to
1491                          * b_more_io so that writeback can proceed with the
1492                          * other inodes on s_io.
1493                          *
1494                          * We'll have another go at writing back this inode
1495                          * when we completed a full scan of b_io.
1496                          */
1497                         spin_unlock(&inode->i_lock);
1498                         requeue_io(inode, wb);
1499                         trace_writeback_sb_inodes_requeue(inode);
1500                         continue;
1501                 }
1502                 spin_unlock(&wb->list_lock);
1503 
1504                 /*
1505                  * We already requeued the inode if it had I_SYNC set and we
1506                  * are doing WB_SYNC_NONE writeback. So this catches only the
1507                  * WB_SYNC_ALL case.
1508                  */
1509                 if (inode->i_state & I_SYNC) {
1510                         /* Wait for I_SYNC. This function drops i_lock... */
1511                         inode_sleep_on_writeback(inode);
1512                         /* Inode may be gone, start again */
1513                         spin_lock(&wb->list_lock);
1514                         continue;
1515                 }
1516                 inode->i_state |= I_SYNC;
1517                 wbc_attach_and_unlock_inode(&wbc, inode);
1518 
1519                 write_chunk = writeback_chunk_size(wb, work);
1520                 wbc.nr_to_write = write_chunk;
1521                 wbc.pages_skipped = 0;
1522 
1523                 /*
1524                  * We use I_SYNC to pin the inode in memory. While it is set
1525                  * evict_inode() will wait so the inode cannot be freed.
1526                  */
1527                 __writeback_single_inode(inode, &wbc);
1528 
1529                 wbc_detach_inode(&wbc);
1530                 work->nr_pages -= write_chunk - wbc.nr_to_write;
1531                 wrote += write_chunk - wbc.nr_to_write;
1532 
1533                 if (need_resched()) {
1534                         /*
1535                          * We're trying to balance between building up a nice
1536                          * long list of IOs to improve our merge rate, and
1537                          * getting those IOs out quickly for anyone throttling
1538                          * in balance_dirty_pages().  cond_resched() doesn't
1539                          * unplug, so get our IOs out the door before we
1540                          * give up the CPU.
1541                          */
1542                         blk_flush_plug(current);
1543                         cond_resched();
1544                 }
1545 
1546 
1547                 spin_lock(&wb->list_lock);
1548                 spin_lock(&inode->i_lock);
1549                 if (!(inode->i_state & I_DIRTY_ALL))
1550                         wrote++;
1551                 requeue_inode(inode, wb, &wbc);
1552                 inode_sync_complete(inode);
1553                 spin_unlock(&inode->i_lock);
1554 
1555                 /*
1556                  * bail out to wb_writeback() often enough to check
1557                  * background threshold and other termination conditions.
1558                  */
1559                 if (wrote) {
1560                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1561                                 break;
1562                         if (work->nr_pages <= 0)
1563                                 break;
1564                 }
1565         }
1566         return wrote;
1567 }
1568 
1569 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1570                                   struct wb_writeback_work *work)
1571 {
1572         unsigned long start_time = jiffies;
1573         long wrote = 0;
1574 
1575         while (!list_empty(&wb->b_io)) {
1576                 struct inode *inode = wb_inode(wb->b_io.prev);
1577                 struct super_block *sb = inode->i_sb;
1578 
1579                 if (!trylock_super(sb)) {
1580                         /*
1581                          * trylock_super() may fail consistently due to
1582                          * s_umount being grabbed by someone else. Don't use
1583                          * requeue_io() to avoid busy retrying the inode/sb.
1584                          */
1585                         redirty_tail(inode, wb);
1586                         continue;
1587                 }
1588                 wrote += writeback_sb_inodes(sb, wb, work);
1589                 up_read(&sb->s_umount);
1590 
1591                 /* refer to the same tests at the end of writeback_sb_inodes */
1592                 if (wrote) {
1593                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1594                                 break;
1595                         if (work->nr_pages <= 0)
1596                                 break;
1597                 }
1598         }
1599         /* Leave any unwritten inodes on b_io */
1600         return wrote;
1601 }
1602 
1603 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1604                                 enum wb_reason reason)
1605 {
1606         struct wb_writeback_work work = {
1607                 .nr_pages       = nr_pages,
1608                 .sync_mode      = WB_SYNC_NONE,
1609                 .range_cyclic   = 1,
1610                 .reason         = reason,
1611         };
1612         struct blk_plug plug;
1613 
1614         blk_start_plug(&plug);
1615         spin_lock(&wb->list_lock);
1616         if (list_empty(&wb->b_io))
1617                 queue_io(wb, &work);
1618         __writeback_inodes_wb(wb, &work);
1619         spin_unlock(&wb->list_lock);
1620         blk_finish_plug(&plug);
1621 
1622         return nr_pages - work.nr_pages;
1623 }
1624 
1625 /*
1626  * Explicit flushing or periodic writeback of "old" data.
1627  *
1628  * Define "old": the first time one of an inode's pages is dirtied, we mark the
1629  * dirtying-time in the inode's address_space.  So this periodic writeback code
1630  * just walks the superblock inode list, writing back any inodes which are
1631  * older than a specific point in time.
1632  *
1633  * Try to run once per dirty_writeback_interval.  But if a writeback event
1634  * takes longer than a dirty_writeback_interval interval, then leave a
1635  * one-second gap.
1636  *
1637  * older_than_this takes precedence over nr_to_write.  So we'll only write back
1638  * all dirty pages if they are all attached to "old" mappings.
1639  */
1640 static long wb_writeback(struct bdi_writeback *wb,
1641                          struct wb_writeback_work *work)
1642 {
1643         unsigned long wb_start = jiffies;
1644         long nr_pages = work->nr_pages;
1645         unsigned long oldest_jif;
1646         struct inode *inode;
1647         long progress;
1648         struct blk_plug plug;
1649 
1650         oldest_jif = jiffies;
1651         work->older_than_this = &oldest_jif;
1652 
1653         blk_start_plug(&plug);
1654         spin_lock(&wb->list_lock);
1655         for (;;) {
1656                 /*
1657                  * Stop writeback when nr_pages has been consumed
1658                  */
1659                 if (work->nr_pages <= 0)
1660                         break;
1661 
1662                 /*
1663                  * Background writeout and kupdate-style writeback may
1664                  * run forever. Stop them if there is other work to do
1665                  * so that e.g. sync can proceed. They'll be restarted
1666                  * after the other works are all done.
1667                  */
1668                 if ((work->for_background || work->for_kupdate) &&
1669                     !list_empty(&wb->work_list))
1670                         break;
1671 
1672                 /*
1673                  * For background writeout, stop when we are below the
1674                  * background dirty threshold
1675                  */
1676                 if (work->for_background && !wb_over_bg_thresh(wb))
1677                         break;
1678 
1679                 /*
1680                  * Kupdate and background works are special and we want to
1681                  * include all inodes that need writing. Livelock avoidance is
1682                  * handled by these works yielding to any other work so we are
1683                  * safe.
1684                  */
1685                 if (work->for_kupdate) {
1686                         oldest_jif = jiffies -
1687                                 msecs_to_jiffies(dirty_expire_interval * 10);
1688                 } else if (work->for_background)
1689                         oldest_jif = jiffies;
1690 
1691                 trace_writeback_start(wb, work);
1692                 if (list_empty(&wb->b_io))
1693                         queue_io(wb, work);
1694                 if (work->sb)
1695                         progress = writeback_sb_inodes(work->sb, wb, work);
1696                 else
1697                         progress = __writeback_inodes_wb(wb, work);
1698                 trace_writeback_written(wb, work);
1699 
1700                 wb_update_bandwidth(wb, wb_start);
1701 
1702                 /*
1703                  * Did we write something? Try for more
1704                  *
1705                  * Dirty inodes are moved to b_io for writeback in batches.
1706                  * The completion of the current batch does not necessarily
1707                  * mean the overall work is done. So we keep looping as long
1708                  * as made some progress on cleaning pages or inodes.
1709                  */
1710                 if (progress)
1711                         continue;
1712                 /*
1713                  * No more inodes for IO, bail
1714                  */
1715                 if (list_empty(&wb->b_more_io))
1716                         break;
1717                 /*
1718                  * Nothing written. Wait for some inode to
1719                  * become available for writeback. Otherwise
1720                  * we'll just busyloop.
1721                  */
1722                 if (!list_empty(&wb->b_more_io))  {
1723                         trace_writeback_wait(wb, work);
1724                         inode = wb_inode(wb->b_more_io.prev);
1725                         spin_lock(&inode->i_lock);
1726                         spin_unlock(&wb->list_lock);
1727                         /* This function drops i_lock... */
1728                         inode_sleep_on_writeback(inode);
1729                         spin_lock(&wb->list_lock);
1730                 }
1731         }
1732         spin_unlock(&wb->list_lock);
1733         blk_finish_plug(&plug);
1734 
1735         return nr_pages - work->nr_pages;
1736 }
1737 
1738 /*
1739  * Return the next wb_writeback_work struct that hasn't been processed yet.
1740  */
1741 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1742 {
1743         struct wb_writeback_work *work = NULL;
1744 
1745         spin_lock_bh(&wb->work_lock);
1746         if (!list_empty(&wb->work_list)) {
1747                 work = list_entry(wb->work_list.next,
1748                                   struct wb_writeback_work, list);
1749                 list_del_init(&work->list);
1750         }
1751         spin_unlock_bh(&wb->work_lock);
1752         return work;
1753 }
1754 
1755 /*
1756  * Add in the number of potentially dirty inodes, because each inode
1757  * write can dirty pagecache in the underlying blockdev.
1758  */
1759 static unsigned long get_nr_dirty_pages(void)
1760 {
1761         return global_page_state(NR_FILE_DIRTY) +
1762                 global_page_state(NR_UNSTABLE_NFS) +
1763                 get_nr_dirty_inodes();
1764 }
1765 
1766 static long wb_check_background_flush(struct bdi_writeback *wb)
1767 {
1768         if (wb_over_bg_thresh(wb)) {
1769 
1770                 struct wb_writeback_work work = {
1771                         .nr_pages       = LONG_MAX,
1772                         .sync_mode      = WB_SYNC_NONE,
1773                         .for_background = 1,
1774                         .range_cyclic   = 1,
1775                         .reason         = WB_REASON_BACKGROUND,
1776                 };
1777 
1778                 return wb_writeback(wb, &work);
1779         }
1780 
1781         return 0;
1782 }
1783 
1784 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1785 {
1786         unsigned long expired;
1787         long nr_pages;
1788 
1789         /*
1790          * When set to zero, disable periodic writeback
1791          */
1792         if (!dirty_writeback_interval)
1793                 return 0;
1794 
1795         expired = wb->last_old_flush +
1796                         msecs_to_jiffies(dirty_writeback_interval * 10);
1797         if (time_before(jiffies, expired))
1798                 return 0;
1799 
1800         wb->last_old_flush = jiffies;
1801         nr_pages = get_nr_dirty_pages();
1802 
1803         if (nr_pages) {
1804                 struct wb_writeback_work work = {
1805                         .nr_pages       = nr_pages,
1806                         .sync_mode      = WB_SYNC_NONE,
1807                         .for_kupdate    = 1,
1808                         .range_cyclic   = 1,
1809                         .reason         = WB_REASON_PERIODIC,
1810                 };
1811 
1812                 return wb_writeback(wb, &work);
1813         }
1814 
1815         return 0;
1816 }
1817 
1818 /*
1819  * Retrieve work items and do the writeback they describe
1820  */
1821 static long wb_do_writeback(struct bdi_writeback *wb)
1822 {
1823         struct wb_writeback_work *work;
1824         long wrote = 0;
1825 
1826         set_bit(WB_writeback_running, &wb->state);
1827         while ((work = get_next_work_item(wb)) != NULL) {
1828                 struct wb_completion *done = work->done;
1829 
1830                 trace_writeback_exec(wb, work);
1831 
1832                 wrote += wb_writeback(wb, work);
1833 
1834                 if (work->auto_free)
1835                         kfree(work);
1836                 if (done && atomic_dec_and_test(&done->cnt))
1837                         wake_up_all(&wb->bdi->wb_waitq);
1838         }
1839 
1840         /*
1841          * Check for periodic writeback, kupdated() style
1842          */
1843         wrote += wb_check_old_data_flush(wb);
1844         wrote += wb_check_background_flush(wb);
1845         clear_bit(WB_writeback_running, &wb->state);
1846 
1847         return wrote;
1848 }
1849 
1850 /*
1851  * Handle writeback of dirty data for the device backed by this bdi. Also
1852  * reschedules periodically and does kupdated style flushing.
1853  */
1854 void wb_workfn(struct work_struct *work)
1855 {
1856         struct bdi_writeback *wb = container_of(to_delayed_work(work),
1857                                                 struct bdi_writeback, dwork);
1858         long pages_written;
1859 
1860         set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
1861         current->flags |= PF_SWAPWRITE;
1862 
1863         if (likely(!current_is_workqueue_rescuer() ||
1864                    !test_bit(WB_registered, &wb->state))) {
1865                 /*
1866                  * The normal path.  Keep writing back @wb until its
1867                  * work_list is empty.  Note that this path is also taken
1868                  * if @wb is shutting down even when we're running off the
1869                  * rescuer as work_list needs to be drained.
1870                  */
1871                 do {
1872                         pages_written = wb_do_writeback(wb);
1873                         trace_writeback_pages_written(pages_written);
1874                 } while (!list_empty(&wb->work_list));
1875         } else {
1876                 /*
1877                  * bdi_wq can't get enough workers and we're running off
1878                  * the emergency worker.  Don't hog it.  Hopefully, 1024 is
1879                  * enough for efficient IO.
1880                  */
1881                 pages_written = writeback_inodes_wb(wb, 1024,
1882                                                     WB_REASON_FORKER_THREAD);
1883                 trace_writeback_pages_written(pages_written);
1884         }
1885 
1886         if (!list_empty(&wb->work_list))
1887                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1888         else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1889                 wb_wakeup_delayed(wb);
1890 
1891         current->flags &= ~PF_SWAPWRITE;
1892 }
1893 
1894 /*
1895  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1896  * the whole world.
1897  */
1898 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1899 {
1900         struct backing_dev_info *bdi;
1901 
1902         if (!nr_pages)
1903                 nr_pages = get_nr_dirty_pages();
1904 
1905         rcu_read_lock();
1906         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1907                 struct bdi_writeback *wb;
1908 
1909                 if (!bdi_has_dirty_io(bdi))
1910                         continue;
1911 
1912                 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
1913                         wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
1914                                            false, reason);
1915         }
1916         rcu_read_unlock();
1917 }
1918 
1919 /*
1920  * Wake up bdi's periodically to make sure dirtytime inodes gets
1921  * written back periodically.  We deliberately do *not* check the
1922  * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1923  * kernel to be constantly waking up once there are any dirtytime
1924  * inodes on the system.  So instead we define a separate delayed work
1925  * function which gets called much more rarely.  (By default, only
1926  * once every 12 hours.)
1927  *
1928  * If there is any other write activity going on in the file system,
1929  * this function won't be necessary.  But if the only thing that has
1930  * happened on the file system is a dirtytime inode caused by an atime
1931  * update, we need this infrastructure below to make sure that inode
1932  * eventually gets pushed out to disk.
1933  */
1934 static void wakeup_dirtytime_writeback(struct work_struct *w);
1935 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1936 
1937 static void wakeup_dirtytime_writeback(struct work_struct *w)
1938 {
1939         struct backing_dev_info *bdi;
1940 
1941         rcu_read_lock();
1942         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1943                 struct bdi_writeback *wb;
1944 
1945                 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
1946                         if (!list_empty(&wb->b_dirty_time))
1947                                 wb_wakeup(wb);
1948         }
1949         rcu_read_unlock();
1950         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1951 }
1952 
1953 static int __init start_dirtytime_writeback(void)
1954 {
1955         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
1956         return 0;
1957 }
1958 __initcall(start_dirtytime_writeback);
1959 
1960 int dirtytime_interval_handler(struct ctl_table *table, int write,
1961                                void __user *buffer, size_t *lenp, loff_t *ppos)
1962 {
1963         int ret;
1964 
1965         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1966         if (ret == 0 && write)
1967                 mod_delayed_work(system_wq, &dirtytime_work, 0);
1968         return ret;
1969 }
1970 
1971 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
1972 {
1973         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
1974                 struct dentry *dentry;
1975                 const char *name = "?";
1976 
1977                 dentry = d_find_alias(inode);
1978                 if (dentry) {
1979                         spin_lock(&dentry->d_lock);
1980                         name = (const char *) dentry->d_name.name;
1981                 }
1982                 printk(KERN_DEBUG
1983                        "%s(%d): dirtied inode %lu (%s) on %s\n",
1984                        current->comm, task_pid_nr(current), inode->i_ino,
1985                        name, inode->i_sb->s_id);
1986                 if (dentry) {
1987                         spin_unlock(&dentry->d_lock);
1988                         dput(dentry);
1989                 }
1990         }
1991 }
1992 
1993 /**
1994  *      __mark_inode_dirty -    internal function
1995  *      @inode: inode to mark
1996  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
1997  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
1998  *      mark_inode_dirty_sync.
1999  *
2000  * Put the inode on the super block's dirty list.
2001  *
2002  * CAREFUL! We mark it dirty unconditionally, but move it onto the
2003  * dirty list only if it is hashed or if it refers to a blockdev.
2004  * If it was not hashed, it will never be added to the dirty list
2005  * even if it is later hashed, as it will have been marked dirty already.
2006  *
2007  * In short, make sure you hash any inodes _before_ you start marking
2008  * them dirty.
2009  *
2010  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2011  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
2012  * the kernel-internal blockdev inode represents the dirtying time of the
2013  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
2014  * page->mapping->host, so the page-dirtying time is recorded in the internal
2015  * blockdev inode.
2016  */
2017 void __mark_inode_dirty(struct inode *inode, int flags)
2018 {
2019 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
2020         struct super_block *sb = inode->i_sb;
2021         int dirtytime;
2022 
2023         trace_writeback_mark_inode_dirty(inode, flags);
2024 
2025         /*
2026          * Don't do this for I_DIRTY_PAGES - that doesn't actually
2027          * dirty the inode itself
2028          */
2029         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
2030                 trace_writeback_dirty_inode_start(inode, flags);
2031 
2032                 if (sb->s_op->dirty_inode)
2033                         sb->s_op->dirty_inode(inode, flags);
2034 
2035                 trace_writeback_dirty_inode(inode, flags);
2036         }
2037         if (flags & I_DIRTY_INODE)
2038                 flags &= ~I_DIRTY_TIME;
2039         dirtytime = flags & I_DIRTY_TIME;
2040 
2041         /*
2042          * Paired with smp_mb() in __writeback_single_inode() for the
2043          * following lockless i_state test.  See there for details.
2044          */
2045         smp_mb();
2046 
2047         if (((inode->i_state & flags) == flags) ||
2048             (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2049                 return;
2050 
2051         if (unlikely(block_dump))
2052                 block_dump___mark_inode_dirty(inode);
2053 
2054         spin_lock(&inode->i_lock);
2055         if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2056                 goto out_unlock_inode;
2057         if ((inode->i_state & flags) != flags) {
2058                 const int was_dirty = inode->i_state & I_DIRTY;
2059 
2060                 inode_attach_wb(inode, NULL);
2061 
2062                 if (flags & I_DIRTY_INODE)
2063                         inode->i_state &= ~I_DIRTY_TIME;
2064                 inode->i_state |= flags;
2065 
2066                 /*
2067                  * If the inode is being synced, just update its dirty state.
2068                  * The unlocker will place the inode on the appropriate
2069                  * superblock list, based upon its state.
2070                  */
2071                 if (inode->i_state & I_SYNC)
2072                         goto out_unlock_inode;
2073 
2074                 /*
2075                  * Only add valid (hashed) inodes to the superblock's
2076                  * dirty list.  Add blockdev inodes as well.
2077                  */
2078                 if (!S_ISBLK(inode->i_mode)) {
2079                         if (inode_unhashed(inode))
2080                                 goto out_unlock_inode;
2081                 }
2082                 if (inode->i_state & I_FREEING)
2083                         goto out_unlock_inode;
2084 
2085                 /*
2086                  * If the inode was already on b_dirty/b_io/b_more_io, don't
2087                  * reposition it (that would break b_dirty time-ordering).
2088                  */
2089                 if (!was_dirty) {
2090                         struct bdi_writeback *wb;
2091                         struct list_head *dirty_list;
2092                         bool wakeup_bdi = false;
2093 
2094                         wb = locked_inode_to_wb_and_lock_list(inode);
2095 
2096                         WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2097                              !test_bit(WB_registered, &wb->state),
2098                              "bdi-%s not registered\n", wb->bdi->name);
2099 
2100                         inode->dirtied_when = jiffies;
2101                         if (dirtytime)
2102                                 inode->dirtied_time_when = jiffies;
2103 
2104                         if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
2105                                 dirty_list = &wb->b_dirty;
2106                         else
2107                                 dirty_list = &wb->b_dirty_time;
2108 
2109                         wakeup_bdi = inode_io_list_move_locked(inode, wb,
2110                                                                dirty_list);
2111 
2112                         spin_unlock(&wb->list_lock);
2113                         trace_writeback_dirty_inode_enqueue(inode);
2114 
2115                         /*
2116                          * If this is the first dirty inode for this bdi,
2117                          * we have to wake-up the corresponding bdi thread
2118                          * to make sure background write-back happens
2119                          * later.
2120                          */
2121                         if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2122                                 wb_wakeup_delayed(wb);
2123                         return;
2124                 }
2125         }
2126 out_unlock_inode:
2127         spin_unlock(&inode->i_lock);
2128 
2129 #undef I_DIRTY_INODE
2130 }
2131 EXPORT_SYMBOL(__mark_inode_dirty);
2132 
2133 /*
2134  * The @s_sync_lock is used to serialise concurrent sync operations
2135  * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2136  * Concurrent callers will block on the s_sync_lock rather than doing contending
2137  * walks. The queueing maintains sync(2) required behaviour as all the IO that
2138  * has been issued up to the time this function is enter is guaranteed to be
2139  * completed by the time we have gained the lock and waited for all IO that is
2140  * in progress regardless of the order callers are granted the lock.
2141  */
2142 static void wait_sb_inodes(struct super_block *sb)
2143 {
2144         struct inode *inode, *old_inode = NULL;
2145 
2146         /*
2147          * We need to be protected against the filesystem going from
2148          * r/o to r/w or vice versa.
2149          */
2150         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2151 
2152         mutex_lock(&sb->s_sync_lock);
2153         spin_lock(&sb->s_inode_list_lock);
2154 
2155         /*
2156          * Data integrity sync. Must wait for all pages under writeback,
2157          * because there may have been pages dirtied before our sync
2158          * call, but which had writeout started before we write it out.
2159          * In which case, the inode may not be on the dirty list, but
2160          * we still have to wait for that writeout.
2161          */
2162         list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
2163                 struct address_space *mapping = inode->i_mapping;
2164 
2165                 spin_lock(&inode->i_lock);
2166                 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
2167                     (mapping->nrpages == 0)) {
2168                         spin_unlock(&inode->i_lock);
2169                         continue;
2170                 }
2171                 __iget(inode);
2172                 spin_unlock(&inode->i_lock);
2173                 spin_unlock(&sb->s_inode_list_lock);
2174 
2175                 /*
2176                  * We hold a reference to 'inode' so it couldn't have been
2177                  * removed from s_inodes list while we dropped the
2178                  * s_inode_list_lock.  We cannot iput the inode now as we can
2179                  * be holding the last reference and we cannot iput it under
2180                  * s_inode_list_lock. So we keep the reference and iput it
2181                  * later.
2182                  */
2183                 iput(old_inode);
2184                 old_inode = inode;
2185 
2186                 /*
2187                  * We keep the error status of individual mapping so that
2188                  * applications can catch the writeback error using fsync(2).
2189                  * See filemap_fdatawait_keep_errors() for details.
2190                  */
2191                 filemap_fdatawait_keep_errors(mapping);
2192 
2193                 cond_resched();
2194 
2195                 spin_lock(&sb->s_inode_list_lock);
2196         }
2197         spin_unlock(&sb->s_inode_list_lock);
2198         iput(old_inode);
2199         mutex_unlock(&sb->s_sync_lock);
2200 }
2201 
2202 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2203                                      enum wb_reason reason, bool skip_if_busy)
2204 {
2205         DEFINE_WB_COMPLETION_ONSTACK(done);
2206         struct wb_writeback_work work = {
2207                 .sb                     = sb,
2208                 .sync_mode              = WB_SYNC_NONE,
2209                 .tagged_writepages      = 1,
2210                 .done                   = &done,
2211                 .nr_pages               = nr,
2212                 .reason                 = reason,
2213         };
2214         struct backing_dev_info *bdi = sb->s_bdi;
2215 
2216         if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2217                 return;
2218         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2219 
2220         bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2221         wb_wait_for_completion(bdi, &done);
2222 }
2223 
2224 /**
2225  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
2226  * @sb: the superblock
2227  * @nr: the number of pages to write
2228  * @reason: reason why some writeback work initiated
2229  *
2230  * Start writeback on some inodes on this super_block. No guarantees are made
2231  * on how many (if any) will be written, and this function does not wait
2232  * for IO completion of submitted IO.
2233  */
2234 void writeback_inodes_sb_nr(struct super_block *sb,
2235                             unsigned long nr,
2236                             enum wb_reason reason)
2237 {
2238         __writeback_inodes_sb_nr(sb, nr, reason, false);
2239 }
2240 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2241 
2242 /**
2243  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
2244  * @sb: the superblock
2245  * @reason: reason why some writeback work was initiated
2246  *
2247  * Start writeback on some inodes on this super_block. No guarantees are made
2248  * on how many (if any) will be written, and this function does not wait
2249  * for IO completion of submitted IO.
2250  */
2251 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2252 {
2253         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2254 }
2255 EXPORT_SYMBOL(writeback_inodes_sb);
2256 
2257 /**
2258  * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
2259  * @sb: the superblock
2260  * @nr: the number of pages to write
2261  * @reason: the reason of writeback
2262  *
2263  * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
2264  * Returns 1 if writeback was started, 0 if not.
2265  */
2266 bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2267                                    enum wb_reason reason)
2268 {
2269         if (!down_read_trylock(&sb->s_umount))
2270                 return false;
2271 
2272         __writeback_inodes_sb_nr(sb, nr, reason, true);
2273         up_read(&sb->s_umount);
2274         return true;
2275 }
2276 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2277 
2278 /**
2279  * try_to_writeback_inodes_sb - try to start writeback if none underway
2280  * @sb: the superblock
2281  * @reason: reason why some writeback work was initiated
2282  *
2283  * Implement by try_to_writeback_inodes_sb_nr()
2284  * Returns 1 if writeback was started, 0 if not.
2285  */
2286 bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2287 {
2288         return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2289 }
2290 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2291 
2292 /**
2293  * sync_inodes_sb       -       sync sb inode pages
2294  * @sb: the superblock
2295  *
2296  * This function writes and waits on any dirty inode belonging to this
2297  * super_block.
2298  */
2299 void sync_inodes_sb(struct super_block *sb)
2300 {
2301         DEFINE_WB_COMPLETION_ONSTACK(done);
2302         struct wb_writeback_work work = {
2303                 .sb             = sb,
2304                 .sync_mode      = WB_SYNC_ALL,
2305                 .nr_pages       = LONG_MAX,
2306                 .range_cyclic   = 0,
2307                 .done           = &done,
2308                 .reason         = WB_REASON_SYNC,
2309                 .for_sync       = 1,
2310         };
2311         struct backing_dev_info *bdi = sb->s_bdi;
2312 
2313         /*
2314          * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2315          * inodes under writeback and I_DIRTY_TIME inodes ignored by
2316          * bdi_has_dirty() need to be written out too.
2317          */
2318         if (bdi == &noop_backing_dev_info)
2319                 return;
2320         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2321 
2322         bdi_split_work_to_wbs(bdi, &work, false);
2323         wb_wait_for_completion(bdi, &done);
2324 
2325         wait_sb_inodes(sb);
2326 }
2327 EXPORT_SYMBOL(sync_inodes_sb);
2328 
2329 /**
2330  * write_inode_now      -       write an inode to disk
2331  * @inode: inode to write to disk
2332  * @sync: whether the write should be synchronous or not
2333  *
2334  * This function commits an inode to disk immediately if it is dirty. This is
2335  * primarily needed by knfsd.
2336  *
2337  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2338  */
2339 int write_inode_now(struct inode *inode, int sync)
2340 {
2341         struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
2342         struct writeback_control wbc = {
2343                 .nr_to_write = LONG_MAX,
2344                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2345                 .range_start = 0,
2346                 .range_end = LLONG_MAX,
2347         };
2348 
2349         if (!mapping_cap_writeback_dirty(inode->i_mapping))
2350                 wbc.nr_to_write = 0;
2351 
2352         might_sleep();
2353         return writeback_single_inode(inode, wb, &wbc);
2354 }
2355 EXPORT_SYMBOL(write_inode_now);
2356 
2357 /**
2358  * sync_inode - write an inode and its pages to disk.
2359  * @inode: the inode to sync
2360  * @wbc: controls the writeback mode
2361  *
2362  * sync_inode() will write an inode and its pages to disk.  It will also
2363  * correctly update the inode on its superblock's dirty inode lists and will
2364  * update inode->i_state.
2365  *
2366  * The caller must have a ref on the inode.
2367  */
2368 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2369 {
2370         return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc);
2371 }
2372 EXPORT_SYMBOL(sync_inode);
2373 
2374 /**
2375  * sync_inode_metadata - write an inode to disk
2376  * @inode: the inode to sync
2377  * @wait: wait for I/O to complete.
2378  *
2379  * Write an inode to disk and adjust its dirty state after completion.
2380  *
2381  * Note: only writes the actual inode, no associated data or other metadata.
2382  */
2383 int sync_inode_metadata(struct inode *inode, int wait)
2384 {
2385         struct writeback_control wbc = {
2386                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2387                 .nr_to_write = 0, /* metadata-only */
2388         };
2389 
2390         return sync_inode(inode, &wbc);
2391 }
2392 EXPORT_SYMBOL(sync_inode_metadata);
2393 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us