Version:  2.0.40 2.2.26 2.4.37 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 4.6 4.7 4.8

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_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 
285                 /* i_wb may have changed inbetween, can't use inode_to_wb() */
286                 if (likely(wb == inode->i_wb)) {
287                         wb_put(wb);     /* @inode already has ref */
288                         return wb;
289                 }
290 
291                 spin_unlock(&wb->list_lock);
292                 wb_put(wb);
293                 cpu_relax();
294                 spin_lock(&inode->i_lock);
295         }
296 }
297 
298 /**
299  * inode_to_wb_and_lock_list - determine an inode's wb and lock it
300  * @inode: inode of interest
301  *
302  * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
303  * on entry.
304  */
305 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
306         __acquires(&wb->list_lock)
307 {
308         spin_lock(&inode->i_lock);
309         return locked_inode_to_wb_and_lock_list(inode);
310 }
311 
312 struct inode_switch_wbs_context {
313         struct inode            *inode;
314         struct bdi_writeback    *new_wb;
315 
316         struct rcu_head         rcu_head;
317         struct work_struct      work;
318 };
319 
320 static void inode_switch_wbs_work_fn(struct work_struct *work)
321 {
322         struct inode_switch_wbs_context *isw =
323                 container_of(work, struct inode_switch_wbs_context, work);
324         struct inode *inode = isw->inode;
325         struct address_space *mapping = inode->i_mapping;
326         struct bdi_writeback *old_wb = inode->i_wb;
327         struct bdi_writeback *new_wb = isw->new_wb;
328         struct radix_tree_iter iter;
329         bool switched = false;
330         void **slot;
331 
332         /*
333          * By the time control reaches here, RCU grace period has passed
334          * since I_WB_SWITCH assertion and all wb stat update transactions
335          * between unlocked_inode_to_wb_begin/end() are guaranteed to be
336          * synchronizing against mapping->tree_lock.
337          *
338          * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock
339          * gives us exclusion against all wb related operations on @inode
340          * including IO list manipulations and stat updates.
341          */
342         if (old_wb < new_wb) {
343                 spin_lock(&old_wb->list_lock);
344                 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
345         } else {
346                 spin_lock(&new_wb->list_lock);
347                 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
348         }
349         spin_lock(&inode->i_lock);
350         spin_lock_irq(&mapping->tree_lock);
351 
352         /*
353          * Once I_FREEING is visible under i_lock, the eviction path owns
354          * the inode and we shouldn't modify ->i_io_list.
355          */
356         if (unlikely(inode->i_state & I_FREEING))
357                 goto skip_switch;
358 
359         /*
360          * Count and transfer stats.  Note that PAGECACHE_TAG_DIRTY points
361          * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
362          * pages actually under underwriteback.
363          */
364         radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
365                                    PAGECACHE_TAG_DIRTY) {
366                 struct page *page = radix_tree_deref_slot_protected(slot,
367                                                         &mapping->tree_lock);
368                 if (likely(page) && PageDirty(page)) {
369                         __dec_wb_stat(old_wb, WB_RECLAIMABLE);
370                         __inc_wb_stat(new_wb, WB_RECLAIMABLE);
371                 }
372         }
373 
374         radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0,
375                                    PAGECACHE_TAG_WRITEBACK) {
376                 struct page *page = radix_tree_deref_slot_protected(slot,
377                                                         &mapping->tree_lock);
378                 if (likely(page)) {
379                         WARN_ON_ONCE(!PageWriteback(page));
380                         __dec_wb_stat(old_wb, WB_WRITEBACK);
381                         __inc_wb_stat(new_wb, WB_WRITEBACK);
382                 }
383         }
384 
385         wb_get(new_wb);
386 
387         /*
388          * Transfer to @new_wb's IO list if necessary.  The specific list
389          * @inode was on is ignored and the inode is put on ->b_dirty which
390          * is always correct including from ->b_dirty_time.  The transfer
391          * preserves @inode->dirtied_when ordering.
392          */
393         if (!list_empty(&inode->i_io_list)) {
394                 struct inode *pos;
395 
396                 inode_io_list_del_locked(inode, old_wb);
397                 inode->i_wb = new_wb;
398                 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
399                         if (time_after_eq(inode->dirtied_when,
400                                           pos->dirtied_when))
401                                 break;
402                 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
403         } else {
404                 inode->i_wb = new_wb;
405         }
406 
407         /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
408         inode->i_wb_frn_winner = 0;
409         inode->i_wb_frn_avg_time = 0;
410         inode->i_wb_frn_history = 0;
411         switched = true;
412 skip_switch:
413         /*
414          * Paired with load_acquire in unlocked_inode_to_wb_begin() and
415          * ensures that the new wb is visible if they see !I_WB_SWITCH.
416          */
417         smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
418 
419         spin_unlock_irq(&mapping->tree_lock);
420         spin_unlock(&inode->i_lock);
421         spin_unlock(&new_wb->list_lock);
422         spin_unlock(&old_wb->list_lock);
423 
424         if (switched) {
425                 wb_wakeup(new_wb);
426                 wb_put(old_wb);
427         }
428         wb_put(new_wb);
429 
430         iput(inode);
431         kfree(isw);
432 
433         atomic_dec(&isw_nr_in_flight);
434 }
435 
436 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
437 {
438         struct inode_switch_wbs_context *isw = container_of(rcu_head,
439                                 struct inode_switch_wbs_context, rcu_head);
440 
441         /* needs to grab bh-unsafe locks, bounce to work item */
442         INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
443         queue_work(isw_wq, &isw->work);
444 }
445 
446 /**
447  * inode_switch_wbs - change the wb association of an inode
448  * @inode: target inode
449  * @new_wb_id: ID of the new wb
450  *
451  * Switch @inode's wb association to the wb identified by @new_wb_id.  The
452  * switching is performed asynchronously and may fail silently.
453  */
454 static void inode_switch_wbs(struct inode *inode, int new_wb_id)
455 {
456         struct backing_dev_info *bdi = inode_to_bdi(inode);
457         struct cgroup_subsys_state *memcg_css;
458         struct inode_switch_wbs_context *isw;
459 
460         /* noop if seems to be already in progress */
461         if (inode->i_state & I_WB_SWITCH)
462                 return;
463 
464         isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
465         if (!isw)
466                 return;
467 
468         /* find and pin the new wb */
469         rcu_read_lock();
470         memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
471         if (memcg_css)
472                 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
473         rcu_read_unlock();
474         if (!isw->new_wb)
475                 goto out_free;
476 
477         /* while holding I_WB_SWITCH, no one else can update the association */
478         spin_lock(&inode->i_lock);
479         if (!(inode->i_sb->s_flags & MS_ACTIVE) ||
480             inode->i_state & (I_WB_SWITCH | I_FREEING) ||
481             inode_to_wb(inode) == isw->new_wb) {
482                 spin_unlock(&inode->i_lock);
483                 goto out_free;
484         }
485         inode->i_state |= I_WB_SWITCH;
486         __iget(inode);
487         spin_unlock(&inode->i_lock);
488 
489         isw->inode = inode;
490 
491         atomic_inc(&isw_nr_in_flight);
492 
493         /*
494          * In addition to synchronizing among switchers, I_WB_SWITCH tells
495          * the RCU protected stat update paths to grab the mapping's
496          * tree_lock so that stat transfer can synchronize against them.
497          * Let's continue after I_WB_SWITCH is guaranteed to be visible.
498          */
499         call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
500         return;
501 
502 out_free:
503         if (isw->new_wb)
504                 wb_put(isw->new_wb);
505         kfree(isw);
506 }
507 
508 /**
509  * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
510  * @wbc: writeback_control of interest
511  * @inode: target inode
512  *
513  * @inode is locked and about to be written back under the control of @wbc.
514  * Record @inode's writeback context into @wbc and unlock the i_lock.  On
515  * writeback completion, wbc_detach_inode() should be called.  This is used
516  * to track the cgroup writeback context.
517  */
518 void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
519                                  struct inode *inode)
520 {
521         if (!inode_cgwb_enabled(inode)) {
522                 spin_unlock(&inode->i_lock);
523                 return;
524         }
525 
526         wbc->wb = inode_to_wb(inode);
527         wbc->inode = inode;
528 
529         wbc->wb_id = wbc->wb->memcg_css->id;
530         wbc->wb_lcand_id = inode->i_wb_frn_winner;
531         wbc->wb_tcand_id = 0;
532         wbc->wb_bytes = 0;
533         wbc->wb_lcand_bytes = 0;
534         wbc->wb_tcand_bytes = 0;
535 
536         wb_get(wbc->wb);
537         spin_unlock(&inode->i_lock);
538 
539         /*
540          * A dying wb indicates that the memcg-blkcg mapping has changed
541          * and a new wb is already serving the memcg.  Switch immediately.
542          */
543         if (unlikely(wb_dying(wbc->wb)))
544                 inode_switch_wbs(inode, wbc->wb_id);
545 }
546 
547 /**
548  * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
549  * @wbc: writeback_control of the just finished writeback
550  *
551  * To be called after a writeback attempt of an inode finishes and undoes
552  * wbc_attach_and_unlock_inode().  Can be called under any context.
553  *
554  * As concurrent write sharing of an inode is expected to be very rare and
555  * memcg only tracks page ownership on first-use basis severely confining
556  * the usefulness of such sharing, cgroup writeback tracks ownership
557  * per-inode.  While the support for concurrent write sharing of an inode
558  * is deemed unnecessary, an inode being written to by different cgroups at
559  * different points in time is a lot more common, and, more importantly,
560  * charging only by first-use can too readily lead to grossly incorrect
561  * behaviors (single foreign page can lead to gigabytes of writeback to be
562  * incorrectly attributed).
563  *
564  * To resolve this issue, cgroup writeback detects the majority dirtier of
565  * an inode and transfers the ownership to it.  To avoid unnnecessary
566  * oscillation, the detection mechanism keeps track of history and gives
567  * out the switch verdict only if the foreign usage pattern is stable over
568  * a certain amount of time and/or writeback attempts.
569  *
570  * On each writeback attempt, @wbc tries to detect the majority writer
571  * using Boyer-Moore majority vote algorithm.  In addition to the byte
572  * count from the majority voting, it also counts the bytes written for the
573  * current wb and the last round's winner wb (max of last round's current
574  * wb, the winner from two rounds ago, and the last round's majority
575  * candidate).  Keeping track of the historical winner helps the algorithm
576  * to semi-reliably detect the most active writer even when it's not the
577  * absolute majority.
578  *
579  * Once the winner of the round is determined, whether the winner is
580  * foreign or not and how much IO time the round consumed is recorded in
581  * inode->i_wb_frn_history.  If the amount of recorded foreign IO time is
582  * over a certain threshold, the switch verdict is given.
583  */
584 void wbc_detach_inode(struct writeback_control *wbc)
585 {
586         struct bdi_writeback *wb = wbc->wb;
587         struct inode *inode = wbc->inode;
588         unsigned long avg_time, max_bytes, max_time;
589         u16 history;
590         int max_id;
591 
592         if (!wb)
593                 return;
594 
595         history = inode->i_wb_frn_history;
596         avg_time = inode->i_wb_frn_avg_time;
597 
598         /* pick the winner of this round */
599         if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
600             wbc->wb_bytes >= wbc->wb_tcand_bytes) {
601                 max_id = wbc->wb_id;
602                 max_bytes = wbc->wb_bytes;
603         } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
604                 max_id = wbc->wb_lcand_id;
605                 max_bytes = wbc->wb_lcand_bytes;
606         } else {
607                 max_id = wbc->wb_tcand_id;
608                 max_bytes = wbc->wb_tcand_bytes;
609         }
610 
611         /*
612          * Calculate the amount of IO time the winner consumed and fold it
613          * into the running average kept per inode.  If the consumed IO
614          * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
615          * deciding whether to switch or not.  This is to prevent one-off
616          * small dirtiers from skewing the verdict.
617          */
618         max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
619                                 wb->avg_write_bandwidth);
620         if (avg_time)
621                 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
622                             (avg_time >> WB_FRN_TIME_AVG_SHIFT);
623         else
624                 avg_time = max_time;    /* immediate catch up on first run */
625 
626         if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
627                 int slots;
628 
629                 /*
630                  * The switch verdict is reached if foreign wb's consume
631                  * more than a certain proportion of IO time in a
632                  * WB_FRN_TIME_PERIOD.  This is loosely tracked by 16 slot
633                  * history mask where each bit represents one sixteenth of
634                  * the period.  Determine the number of slots to shift into
635                  * history from @max_time.
636                  */
637                 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
638                             (unsigned long)WB_FRN_HIST_MAX_SLOTS);
639                 history <<= slots;
640                 if (wbc->wb_id != max_id)
641                         history |= (1U << slots) - 1;
642 
643                 /*
644                  * Switch if the current wb isn't the consistent winner.
645                  * If there are multiple closely competing dirtiers, the
646                  * inode may switch across them repeatedly over time, which
647                  * is okay.  The main goal is avoiding keeping an inode on
648                  * the wrong wb for an extended period of time.
649                  */
650                 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
651                         inode_switch_wbs(inode, max_id);
652         }
653 
654         /*
655          * Multiple instances of this function may race to update the
656          * following fields but we don't mind occassional inaccuracies.
657          */
658         inode->i_wb_frn_winner = max_id;
659         inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
660         inode->i_wb_frn_history = history;
661 
662         wb_put(wbc->wb);
663         wbc->wb = NULL;
664 }
665 
666 /**
667  * wbc_account_io - account IO issued during writeback
668  * @wbc: writeback_control of the writeback in progress
669  * @page: page being written out
670  * @bytes: number of bytes being written out
671  *
672  * @bytes from @page are about to written out during the writeback
673  * controlled by @wbc.  Keep the book for foreign inode detection.  See
674  * wbc_detach_inode().
675  */
676 void wbc_account_io(struct writeback_control *wbc, struct page *page,
677                     size_t bytes)
678 {
679         int id;
680 
681         /*
682          * pageout() path doesn't attach @wbc to the inode being written
683          * out.  This is intentional as we don't want the function to block
684          * behind a slow cgroup.  Ultimately, we want pageout() to kick off
685          * regular writeback instead of writing things out itself.
686          */
687         if (!wbc->wb)
688                 return;
689 
690         id = mem_cgroup_css_from_page(page)->id;
691 
692         if (id == wbc->wb_id) {
693                 wbc->wb_bytes += bytes;
694                 return;
695         }
696 
697         if (id == wbc->wb_lcand_id)
698                 wbc->wb_lcand_bytes += bytes;
699 
700         /* Boyer-Moore majority vote algorithm */
701         if (!wbc->wb_tcand_bytes)
702                 wbc->wb_tcand_id = id;
703         if (id == wbc->wb_tcand_id)
704                 wbc->wb_tcand_bytes += bytes;
705         else
706                 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
707 }
708 EXPORT_SYMBOL_GPL(wbc_account_io);
709 
710 /**
711  * inode_congested - test whether an inode is congested
712  * @inode: inode to test for congestion (may be NULL)
713  * @cong_bits: mask of WB_[a]sync_congested bits to test
714  *
715  * Tests whether @inode is congested.  @cong_bits is the mask of congestion
716  * bits to test and the return value is the mask of set bits.
717  *
718  * If cgroup writeback is enabled for @inode, the congestion state is
719  * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
720  * associated with @inode is congested; otherwise, the root wb's congestion
721  * state is used.
722  *
723  * @inode is allowed to be NULL as this function is often called on
724  * mapping->host which is NULL for the swapper space.
725  */
726 int inode_congested(struct inode *inode, int cong_bits)
727 {
728         /*
729          * Once set, ->i_wb never becomes NULL while the inode is alive.
730          * Start transaction iff ->i_wb is visible.
731          */
732         if (inode && inode_to_wb_is_valid(inode)) {
733                 struct bdi_writeback *wb;
734                 bool locked, congested;
735 
736                 wb = unlocked_inode_to_wb_begin(inode, &locked);
737                 congested = wb_congested(wb, cong_bits);
738                 unlocked_inode_to_wb_end(inode, locked);
739                 return congested;
740         }
741 
742         return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
743 }
744 EXPORT_SYMBOL_GPL(inode_congested);
745 
746 /**
747  * wb_split_bdi_pages - split nr_pages to write according to bandwidth
748  * @wb: target bdi_writeback to split @nr_pages to
749  * @nr_pages: number of pages to write for the whole bdi
750  *
751  * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
752  * relation to the total write bandwidth of all wb's w/ dirty inodes on
753  * @wb->bdi.
754  */
755 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
756 {
757         unsigned long this_bw = wb->avg_write_bandwidth;
758         unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
759 
760         if (nr_pages == LONG_MAX)
761                 return LONG_MAX;
762 
763         /*
764          * This may be called on clean wb's and proportional distribution
765          * may not make sense, just use the original @nr_pages in those
766          * cases.  In general, we wanna err on the side of writing more.
767          */
768         if (!tot_bw || this_bw >= tot_bw)
769                 return nr_pages;
770         else
771                 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
772 }
773 
774 /**
775  * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
776  * @bdi: target backing_dev_info
777  * @base_work: wb_writeback_work to issue
778  * @skip_if_busy: skip wb's which already have writeback in progress
779  *
780  * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
781  * have dirty inodes.  If @base_work->nr_page isn't %LONG_MAX, it's
782  * distributed to the busy wbs according to each wb's proportion in the
783  * total active write bandwidth of @bdi.
784  */
785 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
786                                   struct wb_writeback_work *base_work,
787                                   bool skip_if_busy)
788 {
789         struct bdi_writeback *last_wb = NULL;
790         struct bdi_writeback *wb = list_entry(&bdi->wb_list,
791                                               struct bdi_writeback, bdi_node);
792 
793         might_sleep();
794 restart:
795         rcu_read_lock();
796         list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
797                 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done);
798                 struct wb_writeback_work fallback_work;
799                 struct wb_writeback_work *work;
800                 long nr_pages;
801 
802                 if (last_wb) {
803                         wb_put(last_wb);
804                         last_wb = NULL;
805                 }
806 
807                 /* SYNC_ALL writes out I_DIRTY_TIME too */
808                 if (!wb_has_dirty_io(wb) &&
809                     (base_work->sync_mode == WB_SYNC_NONE ||
810                      list_empty(&wb->b_dirty_time)))
811                         continue;
812                 if (skip_if_busy && writeback_in_progress(wb))
813                         continue;
814 
815                 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
816 
817                 work = kmalloc(sizeof(*work), GFP_ATOMIC);
818                 if (work) {
819                         *work = *base_work;
820                         work->nr_pages = nr_pages;
821                         work->auto_free = 1;
822                         wb_queue_work(wb, work);
823                         continue;
824                 }
825 
826                 /* alloc failed, execute synchronously using on-stack fallback */
827                 work = &fallback_work;
828                 *work = *base_work;
829                 work->nr_pages = nr_pages;
830                 work->auto_free = 0;
831                 work->done = &fallback_work_done;
832 
833                 wb_queue_work(wb, work);
834 
835                 /*
836                  * Pin @wb so that it stays on @bdi->wb_list.  This allows
837                  * continuing iteration from @wb after dropping and
838                  * regrabbing rcu read lock.
839                  */
840                 wb_get(wb);
841                 last_wb = wb;
842 
843                 rcu_read_unlock();
844                 wb_wait_for_completion(bdi, &fallback_work_done);
845                 goto restart;
846         }
847         rcu_read_unlock();
848 
849         if (last_wb)
850                 wb_put(last_wb);
851 }
852 
853 /**
854  * cgroup_writeback_umount - flush inode wb switches for umount
855  *
856  * This function is called when a super_block is about to be destroyed and
857  * flushes in-flight inode wb switches.  An inode wb switch goes through
858  * RCU and then workqueue, so the two need to be flushed in order to ensure
859  * that all previously scheduled switches are finished.  As wb switches are
860  * rare occurrences and synchronize_rcu() can take a while, perform
861  * flushing iff wb switches are in flight.
862  */
863 void cgroup_writeback_umount(void)
864 {
865         if (atomic_read(&isw_nr_in_flight)) {
866                 synchronize_rcu();
867                 flush_workqueue(isw_wq);
868         }
869 }
870 
871 static int __init cgroup_writeback_init(void)
872 {
873         isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
874         if (!isw_wq)
875                 return -ENOMEM;
876         return 0;
877 }
878 fs_initcall(cgroup_writeback_init);
879 
880 #else   /* CONFIG_CGROUP_WRITEBACK */
881 
882 static struct bdi_writeback *
883 locked_inode_to_wb_and_lock_list(struct inode *inode)
884         __releases(&inode->i_lock)
885         __acquires(&wb->list_lock)
886 {
887         struct bdi_writeback *wb = inode_to_wb(inode);
888 
889         spin_unlock(&inode->i_lock);
890         spin_lock(&wb->list_lock);
891         return wb;
892 }
893 
894 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
895         __acquires(&wb->list_lock)
896 {
897         struct bdi_writeback *wb = inode_to_wb(inode);
898 
899         spin_lock(&wb->list_lock);
900         return wb;
901 }
902 
903 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
904 {
905         return nr_pages;
906 }
907 
908 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
909                                   struct wb_writeback_work *base_work,
910                                   bool skip_if_busy)
911 {
912         might_sleep();
913 
914         if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
915                 base_work->auto_free = 0;
916                 wb_queue_work(&bdi->wb, base_work);
917         }
918 }
919 
920 #endif  /* CONFIG_CGROUP_WRITEBACK */
921 
922 void wb_start_writeback(struct bdi_writeback *wb, long nr_pages,
923                         bool range_cyclic, enum wb_reason reason)
924 {
925         struct wb_writeback_work *work;
926 
927         if (!wb_has_dirty_io(wb))
928                 return;
929 
930         /*
931          * This is WB_SYNC_NONE writeback, so if allocation fails just
932          * wakeup the thread for old dirty data writeback
933          */
934         work = kzalloc(sizeof(*work),
935                        GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN);
936         if (!work) {
937                 trace_writeback_nowork(wb);
938                 wb_wakeup(wb);
939                 return;
940         }
941 
942         work->sync_mode = WB_SYNC_NONE;
943         work->nr_pages  = nr_pages;
944         work->range_cyclic = range_cyclic;
945         work->reason    = reason;
946         work->auto_free = 1;
947 
948         wb_queue_work(wb, work);
949 }
950 
951 /**
952  * wb_start_background_writeback - start background writeback
953  * @wb: bdi_writback to write from
954  *
955  * Description:
956  *   This makes sure WB_SYNC_NONE background writeback happens. When
957  *   this function returns, it is only guaranteed that for given wb
958  *   some IO is happening if we are over background dirty threshold.
959  *   Caller need not hold sb s_umount semaphore.
960  */
961 void wb_start_background_writeback(struct bdi_writeback *wb)
962 {
963         /*
964          * We just wake up the flusher thread. It will perform background
965          * writeback as soon as there is no other work to do.
966          */
967         trace_writeback_wake_background(wb);
968         wb_wakeup(wb);
969 }
970 
971 /*
972  * Remove the inode from the writeback list it is on.
973  */
974 void inode_io_list_del(struct inode *inode)
975 {
976         struct bdi_writeback *wb;
977 
978         wb = inode_to_wb_and_lock_list(inode);
979         inode_io_list_del_locked(inode, wb);
980         spin_unlock(&wb->list_lock);
981 }
982 
983 /*
984  * mark an inode as under writeback on the sb
985  */
986 void sb_mark_inode_writeback(struct inode *inode)
987 {
988         struct super_block *sb = inode->i_sb;
989         unsigned long flags;
990 
991         if (list_empty(&inode->i_wb_list)) {
992                 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
993                 if (list_empty(&inode->i_wb_list)) {
994                         list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
995                         trace_sb_mark_inode_writeback(inode);
996                 }
997                 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
998         }
999 }
1000 
1001 /*
1002  * clear an inode as under writeback on the sb
1003  */
1004 void sb_clear_inode_writeback(struct inode *inode)
1005 {
1006         struct super_block *sb = inode->i_sb;
1007         unsigned long flags;
1008 
1009         if (!list_empty(&inode->i_wb_list)) {
1010                 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
1011                 if (!list_empty(&inode->i_wb_list)) {
1012                         list_del_init(&inode->i_wb_list);
1013                         trace_sb_clear_inode_writeback(inode);
1014                 }
1015                 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
1016         }
1017 }
1018 
1019 /*
1020  * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
1021  * furthest end of its superblock's dirty-inode list.
1022  *
1023  * Before stamping the inode's ->dirtied_when, we check to see whether it is
1024  * already the most-recently-dirtied inode on the b_dirty list.  If that is
1025  * the case then the inode must have been redirtied while it was being written
1026  * out and we don't reset its dirtied_when.
1027  */
1028 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
1029 {
1030         if (!list_empty(&wb->b_dirty)) {
1031                 struct inode *tail;
1032 
1033                 tail = wb_inode(wb->b_dirty.next);
1034                 if (time_before(inode->dirtied_when, tail->dirtied_when))
1035                         inode->dirtied_when = jiffies;
1036         }
1037         inode_io_list_move_locked(inode, wb, &wb->b_dirty);
1038 }
1039 
1040 /*
1041  * requeue inode for re-scanning after bdi->b_io list is exhausted.
1042  */
1043 static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
1044 {
1045         inode_io_list_move_locked(inode, wb, &wb->b_more_io);
1046 }
1047 
1048 static void inode_sync_complete(struct inode *inode)
1049 {
1050         inode->i_state &= ~I_SYNC;
1051         /* If inode is clean an unused, put it into LRU now... */
1052         inode_add_lru(inode);
1053         /* Waiters must see I_SYNC cleared before being woken up */
1054         smp_mb();
1055         wake_up_bit(&inode->i_state, __I_SYNC);
1056 }
1057 
1058 static bool inode_dirtied_after(struct inode *inode, unsigned long t)
1059 {
1060         bool ret = time_after(inode->dirtied_when, t);
1061 #ifndef CONFIG_64BIT
1062         /*
1063          * For inodes being constantly redirtied, dirtied_when can get stuck.
1064          * It _appears_ to be in the future, but is actually in distant past.
1065          * This test is necessary to prevent such wrapped-around relative times
1066          * from permanently stopping the whole bdi writeback.
1067          */
1068         ret = ret && time_before_eq(inode->dirtied_when, jiffies);
1069 #endif
1070         return ret;
1071 }
1072 
1073 #define EXPIRE_DIRTY_ATIME 0x0001
1074 
1075 /*
1076  * Move expired (dirtied before work->older_than_this) dirty inodes from
1077  * @delaying_queue to @dispatch_queue.
1078  */
1079 static int move_expired_inodes(struct list_head *delaying_queue,
1080                                struct list_head *dispatch_queue,
1081                                int flags,
1082                                struct wb_writeback_work *work)
1083 {
1084         unsigned long *older_than_this = NULL;
1085         unsigned long expire_time;
1086         LIST_HEAD(tmp);
1087         struct list_head *pos, *node;
1088         struct super_block *sb = NULL;
1089         struct inode *inode;
1090         int do_sb_sort = 0;
1091         int moved = 0;
1092 
1093         if ((flags & EXPIRE_DIRTY_ATIME) == 0)
1094                 older_than_this = work->older_than_this;
1095         else if (!work->for_sync) {
1096                 expire_time = jiffies - (dirtytime_expire_interval * HZ);
1097                 older_than_this = &expire_time;
1098         }
1099         while (!list_empty(delaying_queue)) {
1100                 inode = wb_inode(delaying_queue->prev);
1101                 if (older_than_this &&
1102                     inode_dirtied_after(inode, *older_than_this))
1103                         break;
1104                 list_move(&inode->i_io_list, &tmp);
1105                 moved++;
1106                 if (flags & EXPIRE_DIRTY_ATIME)
1107                         set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state);
1108                 if (sb_is_blkdev_sb(inode->i_sb))
1109                         continue;
1110                 if (sb && sb != inode->i_sb)
1111                         do_sb_sort = 1;
1112                 sb = inode->i_sb;
1113         }
1114 
1115         /* just one sb in list, splice to dispatch_queue and we're done */
1116         if (!do_sb_sort) {
1117                 list_splice(&tmp, dispatch_queue);
1118                 goto out;
1119         }
1120 
1121         /* Move inodes from one superblock together */
1122         while (!list_empty(&tmp)) {
1123                 sb = wb_inode(tmp.prev)->i_sb;
1124                 list_for_each_prev_safe(pos, node, &tmp) {
1125                         inode = wb_inode(pos);
1126                         if (inode->i_sb == sb)
1127                                 list_move(&inode->i_io_list, dispatch_queue);
1128                 }
1129         }
1130 out:
1131         return moved;
1132 }
1133 
1134 /*
1135  * Queue all expired dirty inodes for io, eldest first.
1136  * Before
1137  *         newly dirtied     b_dirty    b_io    b_more_io
1138  *         =============>    gf         edc     BA
1139  * After
1140  *         newly dirtied     b_dirty    b_io    b_more_io
1141  *         =============>    g          fBAedc
1142  *                                           |
1143  *                                           +--> dequeue for IO
1144  */
1145 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
1146 {
1147         int moved;
1148 
1149         assert_spin_locked(&wb->list_lock);
1150         list_splice_init(&wb->b_more_io, &wb->b_io);
1151         moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work);
1152         moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
1153                                      EXPIRE_DIRTY_ATIME, work);
1154         if (moved)
1155                 wb_io_lists_populated(wb);
1156         trace_writeback_queue_io(wb, work, moved);
1157 }
1158 
1159 static int write_inode(struct inode *inode, struct writeback_control *wbc)
1160 {
1161         int ret;
1162 
1163         if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
1164                 trace_writeback_write_inode_start(inode, wbc);
1165                 ret = inode->i_sb->s_op->write_inode(inode, wbc);
1166                 trace_writeback_write_inode(inode, wbc);
1167                 return ret;
1168         }
1169         return 0;
1170 }
1171 
1172 /*
1173  * Wait for writeback on an inode to complete. Called with i_lock held.
1174  * Caller must make sure inode cannot go away when we drop i_lock.
1175  */
1176 static void __inode_wait_for_writeback(struct inode *inode)
1177         __releases(inode->i_lock)
1178         __acquires(inode->i_lock)
1179 {
1180         DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
1181         wait_queue_head_t *wqh;
1182 
1183         wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1184         while (inode->i_state & I_SYNC) {
1185                 spin_unlock(&inode->i_lock);
1186                 __wait_on_bit(wqh, &wq, bit_wait,
1187                               TASK_UNINTERRUPTIBLE);
1188                 spin_lock(&inode->i_lock);
1189         }
1190 }
1191 
1192 /*
1193  * Wait for writeback on an inode to complete. Caller must have inode pinned.
1194  */
1195 void inode_wait_for_writeback(struct inode *inode)
1196 {
1197         spin_lock(&inode->i_lock);
1198         __inode_wait_for_writeback(inode);
1199         spin_unlock(&inode->i_lock);
1200 }
1201 
1202 /*
1203  * Sleep until I_SYNC is cleared. This function must be called with i_lock
1204  * held and drops it. It is aimed for callers not holding any inode reference
1205  * so once i_lock is dropped, inode can go away.
1206  */
1207 static void inode_sleep_on_writeback(struct inode *inode)
1208         __releases(inode->i_lock)
1209 {
1210         DEFINE_WAIT(wait);
1211         wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
1212         int sleep;
1213 
1214         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1215         sleep = inode->i_state & I_SYNC;
1216         spin_unlock(&inode->i_lock);
1217         if (sleep)
1218                 schedule();
1219         finish_wait(wqh, &wait);
1220 }
1221 
1222 /*
1223  * Find proper writeback list for the inode depending on its current state and
1224  * possibly also change of its state while we were doing writeback.  Here we
1225  * handle things such as livelock prevention or fairness of writeback among
1226  * inodes. This function can be called only by flusher thread - noone else
1227  * processes all inodes in writeback lists and requeueing inodes behind flusher
1228  * thread's back can have unexpected consequences.
1229  */
1230 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
1231                           struct writeback_control *wbc)
1232 {
1233         if (inode->i_state & I_FREEING)
1234                 return;
1235 
1236         /*
1237          * Sync livelock prevention. Each inode is tagged and synced in one
1238          * shot. If still dirty, it will be redirty_tail()'ed below.  Update
1239          * the dirty time to prevent enqueue and sync it again.
1240          */
1241         if ((inode->i_state & I_DIRTY) &&
1242             (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
1243                 inode->dirtied_when = jiffies;
1244 
1245         if (wbc->pages_skipped) {
1246                 /*
1247                  * writeback is not making progress due to locked
1248                  * buffers. Skip this inode for now.
1249                  */
1250                 redirty_tail(inode, wb);
1251                 return;
1252         }
1253 
1254         if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
1255                 /*
1256                  * We didn't write back all the pages.  nfs_writepages()
1257                  * sometimes bales out without doing anything.
1258                  */
1259                 if (wbc->nr_to_write <= 0) {
1260                         /* Slice used up. Queue for next turn. */
1261                         requeue_io(inode, wb);
1262                 } else {
1263                         /*
1264                          * Writeback blocked by something other than
1265                          * congestion. Delay the inode for some time to
1266                          * avoid spinning on the CPU (100% iowait)
1267                          * retrying writeback of the dirty page/inode
1268                          * that cannot be performed immediately.
1269                          */
1270                         redirty_tail(inode, wb);
1271                 }
1272         } else if (inode->i_state & I_DIRTY) {
1273                 /*
1274                  * Filesystems can dirty the inode during writeback operations,
1275                  * such as delayed allocation during submission or metadata
1276                  * updates after data IO completion.
1277                  */
1278                 redirty_tail(inode, wb);
1279         } else if (inode->i_state & I_DIRTY_TIME) {
1280                 inode->dirtied_when = jiffies;
1281                 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
1282         } else {
1283                 /* The inode is clean. Remove from writeback lists. */
1284                 inode_io_list_del_locked(inode, wb);
1285         }
1286 }
1287 
1288 /*
1289  * Write out an inode and its dirty pages. Do not update the writeback list
1290  * linkage. That is left to the caller. The caller is also responsible for
1291  * setting I_SYNC flag and calling inode_sync_complete() to clear it.
1292  */
1293 static int
1294 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
1295 {
1296         struct address_space *mapping = inode->i_mapping;
1297         long nr_to_write = wbc->nr_to_write;
1298         unsigned dirty;
1299         int ret;
1300 
1301         WARN_ON(!(inode->i_state & I_SYNC));
1302 
1303         trace_writeback_single_inode_start(inode, wbc, nr_to_write);
1304 
1305         ret = do_writepages(mapping, wbc);
1306 
1307         /*
1308          * Make sure to wait on the data before writing out the metadata.
1309          * This is important for filesystems that modify metadata on data
1310          * I/O completion. We don't do it for sync(2) writeback because it has a
1311          * separate, external IO completion path and ->sync_fs for guaranteeing
1312          * inode metadata is written back correctly.
1313          */
1314         if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
1315                 int err = filemap_fdatawait(mapping);
1316                 if (ret == 0)
1317                         ret = err;
1318         }
1319 
1320         /*
1321          * Some filesystems may redirty the inode during the writeback
1322          * due to delalloc, clear dirty metadata flags right before
1323          * write_inode()
1324          */
1325         spin_lock(&inode->i_lock);
1326 
1327         dirty = inode->i_state & I_DIRTY;
1328         if (inode->i_state & I_DIRTY_TIME) {
1329                 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
1330                     wbc->sync_mode == WB_SYNC_ALL ||
1331                     unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) ||
1332                     unlikely(time_after(jiffies,
1333                                         (inode->dirtied_time_when +
1334                                          dirtytime_expire_interval * HZ)))) {
1335                         dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED;
1336                         trace_writeback_lazytime(inode);
1337                 }
1338         } else
1339                 inode->i_state &= ~I_DIRTY_TIME_EXPIRED;
1340         inode->i_state &= ~dirty;
1341 
1342         /*
1343          * Paired with smp_mb() in __mark_inode_dirty().  This allows
1344          * __mark_inode_dirty() to test i_state without grabbing i_lock -
1345          * either they see the I_DIRTY bits cleared or we see the dirtied
1346          * inode.
1347          *
1348          * I_DIRTY_PAGES is always cleared together above even if @mapping
1349          * still has dirty pages.  The flag is reinstated after smp_mb() if
1350          * necessary.  This guarantees that either __mark_inode_dirty()
1351          * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
1352          */
1353         smp_mb();
1354 
1355         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
1356                 inode->i_state |= I_DIRTY_PAGES;
1357 
1358         spin_unlock(&inode->i_lock);
1359 
1360         if (dirty & I_DIRTY_TIME)
1361                 mark_inode_dirty_sync(inode);
1362         /* Don't write the inode if only I_DIRTY_PAGES was set */
1363         if (dirty & ~I_DIRTY_PAGES) {
1364                 int err = write_inode(inode, wbc);
1365                 if (ret == 0)
1366                         ret = err;
1367         }
1368         trace_writeback_single_inode(inode, wbc, nr_to_write);
1369         return ret;
1370 }
1371 
1372 /*
1373  * Write out an inode's dirty pages. Either the caller has an active reference
1374  * on the inode or the inode has I_WILL_FREE set.
1375  *
1376  * This function is designed to be called for writing back one inode which
1377  * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
1378  * and does more profound writeback list handling in writeback_sb_inodes().
1379  */
1380 static int writeback_single_inode(struct inode *inode,
1381                                   struct writeback_control *wbc)
1382 {
1383         struct bdi_writeback *wb;
1384         int ret = 0;
1385 
1386         spin_lock(&inode->i_lock);
1387         if (!atomic_read(&inode->i_count))
1388                 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
1389         else
1390                 WARN_ON(inode->i_state & I_WILL_FREE);
1391 
1392         if (inode->i_state & I_SYNC) {
1393                 if (wbc->sync_mode != WB_SYNC_ALL)
1394                         goto out;
1395                 /*
1396                  * It's a data-integrity sync. We must wait. Since callers hold
1397                  * inode reference or inode has I_WILL_FREE set, it cannot go
1398                  * away under us.
1399                  */
1400                 __inode_wait_for_writeback(inode);
1401         }
1402         WARN_ON(inode->i_state & I_SYNC);
1403         /*
1404          * Skip inode if it is clean and we have no outstanding writeback in
1405          * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
1406          * function since flusher thread may be doing for example sync in
1407          * parallel and if we move the inode, it could get skipped. So here we
1408          * make sure inode is on some writeback list and leave it there unless
1409          * we have completely cleaned the inode.
1410          */
1411         if (!(inode->i_state & I_DIRTY_ALL) &&
1412             (wbc->sync_mode != WB_SYNC_ALL ||
1413              !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
1414                 goto out;
1415         inode->i_state |= I_SYNC;
1416         wbc_attach_and_unlock_inode(wbc, inode);
1417 
1418         ret = __writeback_single_inode(inode, wbc);
1419 
1420         wbc_detach_inode(wbc);
1421 
1422         wb = inode_to_wb_and_lock_list(inode);
1423         spin_lock(&inode->i_lock);
1424         /*
1425          * If inode is clean, remove it from writeback lists. Otherwise don't
1426          * touch it. See comment above for explanation.
1427          */
1428         if (!(inode->i_state & I_DIRTY_ALL))
1429                 inode_io_list_del_locked(inode, wb);
1430         spin_unlock(&wb->list_lock);
1431         inode_sync_complete(inode);
1432 out:
1433         spin_unlock(&inode->i_lock);
1434         return ret;
1435 }
1436 
1437 static long writeback_chunk_size(struct bdi_writeback *wb,
1438                                  struct wb_writeback_work *work)
1439 {
1440         long pages;
1441 
1442         /*
1443          * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
1444          * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
1445          * here avoids calling into writeback_inodes_wb() more than once.
1446          *
1447          * The intended call sequence for WB_SYNC_ALL writeback is:
1448          *
1449          *      wb_writeback()
1450          *          writeback_sb_inodes()       <== called only once
1451          *              write_cache_pages()     <== called once for each inode
1452          *                   (quickly) tag currently dirty pages
1453          *                   (maybe slowly) sync all tagged pages
1454          */
1455         if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
1456                 pages = LONG_MAX;
1457         else {
1458                 pages = min(wb->avg_write_bandwidth / 2,
1459                             global_wb_domain.dirty_limit / DIRTY_SCOPE);
1460                 pages = min(pages, work->nr_pages);
1461                 pages = round_down(pages + MIN_WRITEBACK_PAGES,
1462                                    MIN_WRITEBACK_PAGES);
1463         }
1464 
1465         return pages;
1466 }
1467 
1468 /*
1469  * Write a portion of b_io inodes which belong to @sb.
1470  *
1471  * Return the number of pages and/or inodes written.
1472  *
1473  * NOTE! This is called with wb->list_lock held, and will
1474  * unlock and relock that for each inode it ends up doing
1475  * IO for.
1476  */
1477 static long writeback_sb_inodes(struct super_block *sb,
1478                                 struct bdi_writeback *wb,
1479                                 struct wb_writeback_work *work)
1480 {
1481         struct writeback_control wbc = {
1482                 .sync_mode              = work->sync_mode,
1483                 .tagged_writepages      = work->tagged_writepages,
1484                 .for_kupdate            = work->for_kupdate,
1485                 .for_background         = work->for_background,
1486                 .for_sync               = work->for_sync,
1487                 .range_cyclic           = work->range_cyclic,
1488                 .range_start            = 0,
1489                 .range_end              = LLONG_MAX,
1490         };
1491         unsigned long start_time = jiffies;
1492         long write_chunk;
1493         long wrote = 0;  /* count both pages and inodes */
1494 
1495         while (!list_empty(&wb->b_io)) {
1496                 struct inode *inode = wb_inode(wb->b_io.prev);
1497                 struct bdi_writeback *tmp_wb;
1498 
1499                 if (inode->i_sb != sb) {
1500                         if (work->sb) {
1501                                 /*
1502                                  * We only want to write back data for this
1503                                  * superblock, move all inodes not belonging
1504                                  * to it back onto the dirty list.
1505                                  */
1506                                 redirty_tail(inode, wb);
1507                                 continue;
1508                         }
1509 
1510                         /*
1511                          * The inode belongs to a different superblock.
1512                          * Bounce back to the caller to unpin this and
1513                          * pin the next superblock.
1514                          */
1515                         break;
1516                 }
1517 
1518                 /*
1519                  * Don't bother with new inodes or inodes being freed, first
1520                  * kind does not need periodic writeout yet, and for the latter
1521                  * kind writeout is handled by the freer.
1522                  */
1523                 spin_lock(&inode->i_lock);
1524                 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
1525                         spin_unlock(&inode->i_lock);
1526                         redirty_tail(inode, wb);
1527                         continue;
1528                 }
1529                 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
1530                         /*
1531                          * If this inode is locked for writeback and we are not
1532                          * doing writeback-for-data-integrity, move it to
1533                          * b_more_io so that writeback can proceed with the
1534                          * other inodes on s_io.
1535                          *
1536                          * We'll have another go at writing back this inode
1537                          * when we completed a full scan of b_io.
1538                          */
1539                         spin_unlock(&inode->i_lock);
1540                         requeue_io(inode, wb);
1541                         trace_writeback_sb_inodes_requeue(inode);
1542                         continue;
1543                 }
1544                 spin_unlock(&wb->list_lock);
1545 
1546                 /*
1547                  * We already requeued the inode if it had I_SYNC set and we
1548                  * are doing WB_SYNC_NONE writeback. So this catches only the
1549                  * WB_SYNC_ALL case.
1550                  */
1551                 if (inode->i_state & I_SYNC) {
1552                         /* Wait for I_SYNC. This function drops i_lock... */
1553                         inode_sleep_on_writeback(inode);
1554                         /* Inode may be gone, start again */
1555                         spin_lock(&wb->list_lock);
1556                         continue;
1557                 }
1558                 inode->i_state |= I_SYNC;
1559                 wbc_attach_and_unlock_inode(&wbc, inode);
1560 
1561                 write_chunk = writeback_chunk_size(wb, work);
1562                 wbc.nr_to_write = write_chunk;
1563                 wbc.pages_skipped = 0;
1564 
1565                 /*
1566                  * We use I_SYNC to pin the inode in memory. While it is set
1567                  * evict_inode() will wait so the inode cannot be freed.
1568                  */
1569                 __writeback_single_inode(inode, &wbc);
1570 
1571                 wbc_detach_inode(&wbc);
1572                 work->nr_pages -= write_chunk - wbc.nr_to_write;
1573                 wrote += write_chunk - wbc.nr_to_write;
1574 
1575                 if (need_resched()) {
1576                         /*
1577                          * We're trying to balance between building up a nice
1578                          * long list of IOs to improve our merge rate, and
1579                          * getting those IOs out quickly for anyone throttling
1580                          * in balance_dirty_pages().  cond_resched() doesn't
1581                          * unplug, so get our IOs out the door before we
1582                          * give up the CPU.
1583                          */
1584                         blk_flush_plug(current);
1585                         cond_resched();
1586                 }
1587 
1588                 /*
1589                  * Requeue @inode if still dirty.  Be careful as @inode may
1590                  * have been switched to another wb in the meantime.
1591                  */
1592                 tmp_wb = inode_to_wb_and_lock_list(inode);
1593                 spin_lock(&inode->i_lock);
1594                 if (!(inode->i_state & I_DIRTY_ALL))
1595                         wrote++;
1596                 requeue_inode(inode, tmp_wb, &wbc);
1597                 inode_sync_complete(inode);
1598                 spin_unlock(&inode->i_lock);
1599 
1600                 if (unlikely(tmp_wb != wb)) {
1601                         spin_unlock(&tmp_wb->list_lock);
1602                         spin_lock(&wb->list_lock);
1603                 }
1604 
1605                 /*
1606                  * bail out to wb_writeback() often enough to check
1607                  * background threshold and other termination conditions.
1608                  */
1609                 if (wrote) {
1610                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1611                                 break;
1612                         if (work->nr_pages <= 0)
1613                                 break;
1614                 }
1615         }
1616         return wrote;
1617 }
1618 
1619 static long __writeback_inodes_wb(struct bdi_writeback *wb,
1620                                   struct wb_writeback_work *work)
1621 {
1622         unsigned long start_time = jiffies;
1623         long wrote = 0;
1624 
1625         while (!list_empty(&wb->b_io)) {
1626                 struct inode *inode = wb_inode(wb->b_io.prev);
1627                 struct super_block *sb = inode->i_sb;
1628 
1629                 if (!trylock_super(sb)) {
1630                         /*
1631                          * trylock_super() may fail consistently due to
1632                          * s_umount being grabbed by someone else. Don't use
1633                          * requeue_io() to avoid busy retrying the inode/sb.
1634                          */
1635                         redirty_tail(inode, wb);
1636                         continue;
1637                 }
1638                 wrote += writeback_sb_inodes(sb, wb, work);
1639                 up_read(&sb->s_umount);
1640 
1641                 /* refer to the same tests at the end of writeback_sb_inodes */
1642                 if (wrote) {
1643                         if (time_is_before_jiffies(start_time + HZ / 10UL))
1644                                 break;
1645                         if (work->nr_pages <= 0)
1646                                 break;
1647                 }
1648         }
1649         /* Leave any unwritten inodes on b_io */
1650         return wrote;
1651 }
1652 
1653 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
1654                                 enum wb_reason reason)
1655 {
1656         struct wb_writeback_work work = {
1657                 .nr_pages       = nr_pages,
1658                 .sync_mode      = WB_SYNC_NONE,
1659                 .range_cyclic   = 1,
1660                 .reason         = reason,
1661         };
1662         struct blk_plug plug;
1663 
1664         blk_start_plug(&plug);
1665         spin_lock(&wb->list_lock);
1666         if (list_empty(&wb->b_io))
1667                 queue_io(wb, &work);
1668         __writeback_inodes_wb(wb, &work);
1669         spin_unlock(&wb->list_lock);
1670         blk_finish_plug(&plug);
1671 
1672         return nr_pages - work.nr_pages;
1673 }
1674 
1675 /*
1676  * Explicit flushing or periodic writeback of "old" data.
1677  *
1678  * Define "old": the first time one of an inode's pages is dirtied, we mark the
1679  * dirtying-time in the inode's address_space.  So this periodic writeback code
1680  * just walks the superblock inode list, writing back any inodes which are
1681  * older than a specific point in time.
1682  *
1683  * Try to run once per dirty_writeback_interval.  But if a writeback event
1684  * takes longer than a dirty_writeback_interval interval, then leave a
1685  * one-second gap.
1686  *
1687  * older_than_this takes precedence over nr_to_write.  So we'll only write back
1688  * all dirty pages if they are all attached to "old" mappings.
1689  */
1690 static long wb_writeback(struct bdi_writeback *wb,
1691                          struct wb_writeback_work *work)
1692 {
1693         unsigned long wb_start = jiffies;
1694         long nr_pages = work->nr_pages;
1695         unsigned long oldest_jif;
1696         struct inode *inode;
1697         long progress;
1698         struct blk_plug plug;
1699 
1700         oldest_jif = jiffies;
1701         work->older_than_this = &oldest_jif;
1702 
1703         blk_start_plug(&plug);
1704         spin_lock(&wb->list_lock);
1705         for (;;) {
1706                 /*
1707                  * Stop writeback when nr_pages has been consumed
1708                  */
1709                 if (work->nr_pages <= 0)
1710                         break;
1711 
1712                 /*
1713                  * Background writeout and kupdate-style writeback may
1714                  * run forever. Stop them if there is other work to do
1715                  * so that e.g. sync can proceed. They'll be restarted
1716                  * after the other works are all done.
1717                  */
1718                 if ((work->for_background || work->for_kupdate) &&
1719                     !list_empty(&wb->work_list))
1720                         break;
1721 
1722                 /*
1723                  * For background writeout, stop when we are below the
1724                  * background dirty threshold
1725                  */
1726                 if (work->for_background && !wb_over_bg_thresh(wb))
1727                         break;
1728 
1729                 /*
1730                  * Kupdate and background works are special and we want to
1731                  * include all inodes that need writing. Livelock avoidance is
1732                  * handled by these works yielding to any other work so we are
1733                  * safe.
1734                  */
1735                 if (work->for_kupdate) {
1736                         oldest_jif = jiffies -
1737                                 msecs_to_jiffies(dirty_expire_interval * 10);
1738                 } else if (work->for_background)
1739                         oldest_jif = jiffies;
1740 
1741                 trace_writeback_start(wb, work);
1742                 if (list_empty(&wb->b_io))
1743                         queue_io(wb, work);
1744                 if (work->sb)
1745                         progress = writeback_sb_inodes(work->sb, wb, work);
1746                 else
1747                         progress = __writeback_inodes_wb(wb, work);
1748                 trace_writeback_written(wb, work);
1749 
1750                 wb_update_bandwidth(wb, wb_start);
1751 
1752                 /*
1753                  * Did we write something? Try for more
1754                  *
1755                  * Dirty inodes are moved to b_io for writeback in batches.
1756                  * The completion of the current batch does not necessarily
1757                  * mean the overall work is done. So we keep looping as long
1758                  * as made some progress on cleaning pages or inodes.
1759                  */
1760                 if (progress)
1761                         continue;
1762                 /*
1763                  * No more inodes for IO, bail
1764                  */
1765                 if (list_empty(&wb->b_more_io))
1766                         break;
1767                 /*
1768                  * Nothing written. Wait for some inode to
1769                  * become available for writeback. Otherwise
1770                  * we'll just busyloop.
1771                  */
1772                 if (!list_empty(&wb->b_more_io))  {
1773                         trace_writeback_wait(wb, work);
1774                         inode = wb_inode(wb->b_more_io.prev);
1775                         spin_lock(&inode->i_lock);
1776                         spin_unlock(&wb->list_lock);
1777                         /* This function drops i_lock... */
1778                         inode_sleep_on_writeback(inode);
1779                         spin_lock(&wb->list_lock);
1780                 }
1781         }
1782         spin_unlock(&wb->list_lock);
1783         blk_finish_plug(&plug);
1784 
1785         return nr_pages - work->nr_pages;
1786 }
1787 
1788 /*
1789  * Return the next wb_writeback_work struct that hasn't been processed yet.
1790  */
1791 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
1792 {
1793         struct wb_writeback_work *work = NULL;
1794 
1795         spin_lock_bh(&wb->work_lock);
1796         if (!list_empty(&wb->work_list)) {
1797                 work = list_entry(wb->work_list.next,
1798                                   struct wb_writeback_work, list);
1799                 list_del_init(&work->list);
1800         }
1801         spin_unlock_bh(&wb->work_lock);
1802         return work;
1803 }
1804 
1805 /*
1806  * Add in the number of potentially dirty inodes, because each inode
1807  * write can dirty pagecache in the underlying blockdev.
1808  */
1809 static unsigned long get_nr_dirty_pages(void)
1810 {
1811         return global_node_page_state(NR_FILE_DIRTY) +
1812                 global_node_page_state(NR_UNSTABLE_NFS) +
1813                 get_nr_dirty_inodes();
1814 }
1815 
1816 static long wb_check_background_flush(struct bdi_writeback *wb)
1817 {
1818         if (wb_over_bg_thresh(wb)) {
1819 
1820                 struct wb_writeback_work work = {
1821                         .nr_pages       = LONG_MAX,
1822                         .sync_mode      = WB_SYNC_NONE,
1823                         .for_background = 1,
1824                         .range_cyclic   = 1,
1825                         .reason         = WB_REASON_BACKGROUND,
1826                 };
1827 
1828                 return wb_writeback(wb, &work);
1829         }
1830 
1831         return 0;
1832 }
1833 
1834 static long wb_check_old_data_flush(struct bdi_writeback *wb)
1835 {
1836         unsigned long expired;
1837         long nr_pages;
1838 
1839         /*
1840          * When set to zero, disable periodic writeback
1841          */
1842         if (!dirty_writeback_interval)
1843                 return 0;
1844 
1845         expired = wb->last_old_flush +
1846                         msecs_to_jiffies(dirty_writeback_interval * 10);
1847         if (time_before(jiffies, expired))
1848                 return 0;
1849 
1850         wb->last_old_flush = jiffies;
1851         nr_pages = get_nr_dirty_pages();
1852 
1853         if (nr_pages) {
1854                 struct wb_writeback_work work = {
1855                         .nr_pages       = nr_pages,
1856                         .sync_mode      = WB_SYNC_NONE,
1857                         .for_kupdate    = 1,
1858                         .range_cyclic   = 1,
1859                         .reason         = WB_REASON_PERIODIC,
1860                 };
1861 
1862                 return wb_writeback(wb, &work);
1863         }
1864 
1865         return 0;
1866 }
1867 
1868 /*
1869  * Retrieve work items and do the writeback they describe
1870  */
1871 static long wb_do_writeback(struct bdi_writeback *wb)
1872 {
1873         struct wb_writeback_work *work;
1874         long wrote = 0;
1875 
1876         set_bit(WB_writeback_running, &wb->state);
1877         while ((work = get_next_work_item(wb)) != NULL) {
1878                 struct wb_completion *done = work->done;
1879 
1880                 trace_writeback_exec(wb, work);
1881 
1882                 wrote += wb_writeback(wb, work);
1883 
1884                 if (work->auto_free)
1885                         kfree(work);
1886                 if (done && atomic_dec_and_test(&done->cnt))
1887                         wake_up_all(&wb->bdi->wb_waitq);
1888         }
1889 
1890         /*
1891          * Check for periodic writeback, kupdated() style
1892          */
1893         wrote += wb_check_old_data_flush(wb);
1894         wrote += wb_check_background_flush(wb);
1895         clear_bit(WB_writeback_running, &wb->state);
1896 
1897         return wrote;
1898 }
1899 
1900 /*
1901  * Handle writeback of dirty data for the device backed by this bdi. Also
1902  * reschedules periodically and does kupdated style flushing.
1903  */
1904 void wb_workfn(struct work_struct *work)
1905 {
1906         struct bdi_writeback *wb = container_of(to_delayed_work(work),
1907                                                 struct bdi_writeback, dwork);
1908         long pages_written;
1909 
1910         set_worker_desc("flush-%s", dev_name(wb->bdi->dev));
1911         current->flags |= PF_SWAPWRITE;
1912 
1913         if (likely(!current_is_workqueue_rescuer() ||
1914                    !test_bit(WB_registered, &wb->state))) {
1915                 /*
1916                  * The normal path.  Keep writing back @wb until its
1917                  * work_list is empty.  Note that this path is also taken
1918                  * if @wb is shutting down even when we're running off the
1919                  * rescuer as work_list needs to be drained.
1920                  */
1921                 do {
1922                         pages_written = wb_do_writeback(wb);
1923                         trace_writeback_pages_written(pages_written);
1924                 } while (!list_empty(&wb->work_list));
1925         } else {
1926                 /*
1927                  * bdi_wq can't get enough workers and we're running off
1928                  * the emergency worker.  Don't hog it.  Hopefully, 1024 is
1929                  * enough for efficient IO.
1930                  */
1931                 pages_written = writeback_inodes_wb(wb, 1024,
1932                                                     WB_REASON_FORKER_THREAD);
1933                 trace_writeback_pages_written(pages_written);
1934         }
1935 
1936         if (!list_empty(&wb->work_list))
1937                 mod_delayed_work(bdi_wq, &wb->dwork, 0);
1938         else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
1939                 wb_wakeup_delayed(wb);
1940 
1941         current->flags &= ~PF_SWAPWRITE;
1942 }
1943 
1944 /*
1945  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
1946  * the whole world.
1947  */
1948 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
1949 {
1950         struct backing_dev_info *bdi;
1951 
1952         /*
1953          * If we are expecting writeback progress we must submit plugged IO.
1954          */
1955         if (blk_needs_flush_plug(current))
1956                 blk_schedule_flush_plug(current);
1957 
1958         if (!nr_pages)
1959                 nr_pages = get_nr_dirty_pages();
1960 
1961         rcu_read_lock();
1962         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1963                 struct bdi_writeback *wb;
1964 
1965                 if (!bdi_has_dirty_io(bdi))
1966                         continue;
1967 
1968                 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
1969                         wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages),
1970                                            false, reason);
1971         }
1972         rcu_read_unlock();
1973 }
1974 
1975 /*
1976  * Wake up bdi's periodically to make sure dirtytime inodes gets
1977  * written back periodically.  We deliberately do *not* check the
1978  * b_dirtytime list in wb_has_dirty_io(), since this would cause the
1979  * kernel to be constantly waking up once there are any dirtytime
1980  * inodes on the system.  So instead we define a separate delayed work
1981  * function which gets called much more rarely.  (By default, only
1982  * once every 12 hours.)
1983  *
1984  * If there is any other write activity going on in the file system,
1985  * this function won't be necessary.  But if the only thing that has
1986  * happened on the file system is a dirtytime inode caused by an atime
1987  * update, we need this infrastructure below to make sure that inode
1988  * eventually gets pushed out to disk.
1989  */
1990 static void wakeup_dirtytime_writeback(struct work_struct *w);
1991 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
1992 
1993 static void wakeup_dirtytime_writeback(struct work_struct *w)
1994 {
1995         struct backing_dev_info *bdi;
1996 
1997         rcu_read_lock();
1998         list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
1999                 struct bdi_writeback *wb;
2000 
2001                 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
2002                         if (!list_empty(&wb->b_dirty_time))
2003                                 wb_wakeup(wb);
2004         }
2005         rcu_read_unlock();
2006         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2007 }
2008 
2009 static int __init start_dirtytime_writeback(void)
2010 {
2011         schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
2012         return 0;
2013 }
2014 __initcall(start_dirtytime_writeback);
2015 
2016 int dirtytime_interval_handler(struct ctl_table *table, int write,
2017                                void __user *buffer, size_t *lenp, loff_t *ppos)
2018 {
2019         int ret;
2020 
2021         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2022         if (ret == 0 && write)
2023                 mod_delayed_work(system_wq, &dirtytime_work, 0);
2024         return ret;
2025 }
2026 
2027 static noinline void block_dump___mark_inode_dirty(struct inode *inode)
2028 {
2029         if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
2030                 struct dentry *dentry;
2031                 const char *name = "?";
2032 
2033                 dentry = d_find_alias(inode);
2034                 if (dentry) {
2035                         spin_lock(&dentry->d_lock);
2036                         name = (const char *) dentry->d_name.name;
2037                 }
2038                 printk(KERN_DEBUG
2039                        "%s(%d): dirtied inode %lu (%s) on %s\n",
2040                        current->comm, task_pid_nr(current), inode->i_ino,
2041                        name, inode->i_sb->s_id);
2042                 if (dentry) {
2043                         spin_unlock(&dentry->d_lock);
2044                         dput(dentry);
2045                 }
2046         }
2047 }
2048 
2049 /**
2050  *      __mark_inode_dirty -    internal function
2051  *      @inode: inode to mark
2052  *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
2053  *      Mark an inode as dirty. Callers should use mark_inode_dirty or
2054  *      mark_inode_dirty_sync.
2055  *
2056  * Put the inode on the super block's dirty list.
2057  *
2058  * CAREFUL! We mark it dirty unconditionally, but move it onto the
2059  * dirty list only if it is hashed or if it refers to a blockdev.
2060  * If it was not hashed, it will never be added to the dirty list
2061  * even if it is later hashed, as it will have been marked dirty already.
2062  *
2063  * In short, make sure you hash any inodes _before_ you start marking
2064  * them dirty.
2065  *
2066  * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
2067  * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
2068  * the kernel-internal blockdev inode represents the dirtying time of the
2069  * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
2070  * page->mapping->host, so the page-dirtying time is recorded in the internal
2071  * blockdev inode.
2072  */
2073 void __mark_inode_dirty(struct inode *inode, int flags)
2074 {
2075 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC)
2076         struct super_block *sb = inode->i_sb;
2077         int dirtytime;
2078 
2079         trace_writeback_mark_inode_dirty(inode, flags);
2080 
2081         /*
2082          * Don't do this for I_DIRTY_PAGES - that doesn't actually
2083          * dirty the inode itself
2084          */
2085         if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) {
2086                 trace_writeback_dirty_inode_start(inode, flags);
2087 
2088                 if (sb->s_op->dirty_inode)
2089                         sb->s_op->dirty_inode(inode, flags);
2090 
2091                 trace_writeback_dirty_inode(inode, flags);
2092         }
2093         if (flags & I_DIRTY_INODE)
2094                 flags &= ~I_DIRTY_TIME;
2095         dirtytime = flags & I_DIRTY_TIME;
2096 
2097         /*
2098          * Paired with smp_mb() in __writeback_single_inode() for the
2099          * following lockless i_state test.  See there for details.
2100          */
2101         smp_mb();
2102 
2103         if (((inode->i_state & flags) == flags) ||
2104             (dirtytime && (inode->i_state & I_DIRTY_INODE)))
2105                 return;
2106 
2107         if (unlikely(block_dump))
2108                 block_dump___mark_inode_dirty(inode);
2109 
2110         spin_lock(&inode->i_lock);
2111         if (dirtytime && (inode->i_state & I_DIRTY_INODE))
2112                 goto out_unlock_inode;
2113         if ((inode->i_state & flags) != flags) {
2114                 const int was_dirty = inode->i_state & I_DIRTY;
2115 
2116                 inode_attach_wb(inode, NULL);
2117 
2118                 if (flags & I_DIRTY_INODE)
2119                         inode->i_state &= ~I_DIRTY_TIME;
2120                 inode->i_state |= flags;
2121 
2122                 /*
2123                  * If the inode is being synced, just update its dirty state.
2124                  * The unlocker will place the inode on the appropriate
2125                  * superblock list, based upon its state.
2126                  */
2127                 if (inode->i_state & I_SYNC)
2128                         goto out_unlock_inode;
2129 
2130                 /*
2131                  * Only add valid (hashed) inodes to the superblock's
2132                  * dirty list.  Add blockdev inodes as well.
2133                  */
2134                 if (!S_ISBLK(inode->i_mode)) {
2135                         if (inode_unhashed(inode))
2136                                 goto out_unlock_inode;
2137                 }
2138                 if (inode->i_state & I_FREEING)
2139                         goto out_unlock_inode;
2140 
2141                 /*
2142                  * If the inode was already on b_dirty/b_io/b_more_io, don't
2143                  * reposition it (that would break b_dirty time-ordering).
2144                  */
2145                 if (!was_dirty) {
2146                         struct bdi_writeback *wb;
2147                         struct list_head *dirty_list;
2148                         bool wakeup_bdi = false;
2149 
2150                         wb = locked_inode_to_wb_and_lock_list(inode);
2151 
2152                         WARN(bdi_cap_writeback_dirty(wb->bdi) &&
2153                              !test_bit(WB_registered, &wb->state),
2154                              "bdi-%s not registered\n", wb->bdi->name);
2155 
2156                         inode->dirtied_when = jiffies;
2157                         if (dirtytime)
2158                                 inode->dirtied_time_when = jiffies;
2159 
2160                         if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES))
2161                                 dirty_list = &wb->b_dirty;
2162                         else
2163                                 dirty_list = &wb->b_dirty_time;
2164 
2165                         wakeup_bdi = inode_io_list_move_locked(inode, wb,
2166                                                                dirty_list);
2167 
2168                         spin_unlock(&wb->list_lock);
2169                         trace_writeback_dirty_inode_enqueue(inode);
2170 
2171                         /*
2172                          * If this is the first dirty inode for this bdi,
2173                          * we have to wake-up the corresponding bdi thread
2174                          * to make sure background write-back happens
2175                          * later.
2176                          */
2177                         if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi)
2178                                 wb_wakeup_delayed(wb);
2179                         return;
2180                 }
2181         }
2182 out_unlock_inode:
2183         spin_unlock(&inode->i_lock);
2184 
2185 #undef I_DIRTY_INODE
2186 }
2187 EXPORT_SYMBOL(__mark_inode_dirty);
2188 
2189 /*
2190  * The @s_sync_lock is used to serialise concurrent sync operations
2191  * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
2192  * Concurrent callers will block on the s_sync_lock rather than doing contending
2193  * walks. The queueing maintains sync(2) required behaviour as all the IO that
2194  * has been issued up to the time this function is enter is guaranteed to be
2195  * completed by the time we have gained the lock and waited for all IO that is
2196  * in progress regardless of the order callers are granted the lock.
2197  */
2198 static void wait_sb_inodes(struct super_block *sb)
2199 {
2200         LIST_HEAD(sync_list);
2201 
2202         /*
2203          * We need to be protected against the filesystem going from
2204          * r/o to r/w or vice versa.
2205          */
2206         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2207 
2208         mutex_lock(&sb->s_sync_lock);
2209 
2210         /*
2211          * Splice the writeback list onto a temporary list to avoid waiting on
2212          * inodes that have started writeback after this point.
2213          *
2214          * Use rcu_read_lock() to keep the inodes around until we have a
2215          * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
2216          * the local list because inodes can be dropped from either by writeback
2217          * completion.
2218          */
2219         rcu_read_lock();
2220         spin_lock_irq(&sb->s_inode_wblist_lock);
2221         list_splice_init(&sb->s_inodes_wb, &sync_list);
2222 
2223         /*
2224          * Data integrity sync. Must wait for all pages under writeback, because
2225          * there may have been pages dirtied before our sync call, but which had
2226          * writeout started before we write it out.  In which case, the inode
2227          * may not be on the dirty list, but we still have to wait for that
2228          * writeout.
2229          */
2230         while (!list_empty(&sync_list)) {
2231                 struct inode *inode = list_first_entry(&sync_list, struct inode,
2232                                                        i_wb_list);
2233                 struct address_space *mapping = inode->i_mapping;
2234 
2235                 /*
2236                  * Move each inode back to the wb list before we drop the lock
2237                  * to preserve consistency between i_wb_list and the mapping
2238                  * writeback tag. Writeback completion is responsible to remove
2239                  * the inode from either list once the writeback tag is cleared.
2240                  */
2241                 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
2242 
2243                 /*
2244                  * The mapping can appear untagged while still on-list since we
2245                  * do not have the mapping lock. Skip it here, wb completion
2246                  * will remove it.
2247                  */
2248                 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
2249                         continue;
2250 
2251                 spin_unlock_irq(&sb->s_inode_wblist_lock);
2252 
2253                 spin_lock(&inode->i_lock);
2254                 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
2255                         spin_unlock(&inode->i_lock);
2256 
2257                         spin_lock_irq(&sb->s_inode_wblist_lock);
2258                         continue;
2259                 }
2260                 __iget(inode);
2261                 spin_unlock(&inode->i_lock);
2262                 rcu_read_unlock();
2263 
2264                 /*
2265                  * We keep the error status of individual mapping so that
2266                  * applications can catch the writeback error using fsync(2).
2267                  * See filemap_fdatawait_keep_errors() for details.
2268                  */
2269                 filemap_fdatawait_keep_errors(mapping);
2270 
2271                 cond_resched();
2272 
2273                 iput(inode);
2274 
2275                 rcu_read_lock();
2276                 spin_lock_irq(&sb->s_inode_wblist_lock);
2277         }
2278         spin_unlock_irq(&sb->s_inode_wblist_lock);
2279         rcu_read_unlock();
2280         mutex_unlock(&sb->s_sync_lock);
2281 }
2282 
2283 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2284                                      enum wb_reason reason, bool skip_if_busy)
2285 {
2286         DEFINE_WB_COMPLETION_ONSTACK(done);
2287         struct wb_writeback_work work = {
2288                 .sb                     = sb,
2289                 .sync_mode              = WB_SYNC_NONE,
2290                 .tagged_writepages      = 1,
2291                 .done                   = &done,
2292                 .nr_pages               = nr,
2293                 .reason                 = reason,
2294         };
2295         struct backing_dev_info *bdi = sb->s_bdi;
2296 
2297         if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
2298                 return;
2299         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2300 
2301         bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
2302         wb_wait_for_completion(bdi, &done);
2303 }
2304 
2305 /**
2306  * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
2307  * @sb: the superblock
2308  * @nr: the number of pages to write
2309  * @reason: reason why some writeback work initiated
2310  *
2311  * Start writeback on some inodes on this super_block. No guarantees are made
2312  * on how many (if any) will be written, and this function does not wait
2313  * for IO completion of submitted IO.
2314  */
2315 void writeback_inodes_sb_nr(struct super_block *sb,
2316                             unsigned long nr,
2317                             enum wb_reason reason)
2318 {
2319         __writeback_inodes_sb_nr(sb, nr, reason, false);
2320 }
2321 EXPORT_SYMBOL(writeback_inodes_sb_nr);
2322 
2323 /**
2324  * writeback_inodes_sb  -       writeback dirty inodes from given super_block
2325  * @sb: the superblock
2326  * @reason: reason why some writeback work was initiated
2327  *
2328  * Start writeback on some inodes on this super_block. No guarantees are made
2329  * on how many (if any) will be written, and this function does not wait
2330  * for IO completion of submitted IO.
2331  */
2332 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2333 {
2334         return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2335 }
2336 EXPORT_SYMBOL(writeback_inodes_sb);
2337 
2338 /**
2339  * try_to_writeback_inodes_sb_nr - try to start writeback if none underway
2340  * @sb: the superblock
2341  * @nr: the number of pages to write
2342  * @reason: the reason of writeback
2343  *
2344  * Invoke writeback_inodes_sb_nr if no writeback is currently underway.
2345  * Returns 1 if writeback was started, 0 if not.
2346  */
2347 bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
2348                                    enum wb_reason reason)
2349 {
2350         if (!down_read_trylock(&sb->s_umount))
2351                 return false;
2352 
2353         __writeback_inodes_sb_nr(sb, nr, reason, true);
2354         up_read(&sb->s_umount);
2355         return true;
2356 }
2357 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr);
2358 
2359 /**
2360  * try_to_writeback_inodes_sb - try to start writeback if none underway
2361  * @sb: the superblock
2362  * @reason: reason why some writeback work was initiated
2363  *
2364  * Implement by try_to_writeback_inodes_sb_nr()
2365  * Returns 1 if writeback was started, 0 if not.
2366  */
2367 bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
2368 {
2369         return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
2370 }
2371 EXPORT_SYMBOL(try_to_writeback_inodes_sb);
2372 
2373 /**
2374  * sync_inodes_sb       -       sync sb inode pages
2375  * @sb: the superblock
2376  *
2377  * This function writes and waits on any dirty inode belonging to this
2378  * super_block.
2379  */
2380 void sync_inodes_sb(struct super_block *sb)
2381 {
2382         DEFINE_WB_COMPLETION_ONSTACK(done);
2383         struct wb_writeback_work work = {
2384                 .sb             = sb,
2385                 .sync_mode      = WB_SYNC_ALL,
2386                 .nr_pages       = LONG_MAX,
2387                 .range_cyclic   = 0,
2388                 .done           = &done,
2389                 .reason         = WB_REASON_SYNC,
2390                 .for_sync       = 1,
2391         };
2392         struct backing_dev_info *bdi = sb->s_bdi;
2393 
2394         /*
2395          * Can't skip on !bdi_has_dirty() because we should wait for !dirty
2396          * inodes under writeback and I_DIRTY_TIME inodes ignored by
2397          * bdi_has_dirty() need to be written out too.
2398          */
2399         if (bdi == &noop_backing_dev_info)
2400                 return;
2401         WARN_ON(!rwsem_is_locked(&sb->s_umount));
2402 
2403         bdi_split_work_to_wbs(bdi, &work, false);
2404         wb_wait_for_completion(bdi, &done);
2405 
2406         wait_sb_inodes(sb);
2407 }
2408 EXPORT_SYMBOL(sync_inodes_sb);
2409 
2410 /**
2411  * write_inode_now      -       write an inode to disk
2412  * @inode: inode to write to disk
2413  * @sync: whether the write should be synchronous or not
2414  *
2415  * This function commits an inode to disk immediately if it is dirty. This is
2416  * primarily needed by knfsd.
2417  *
2418  * The caller must either have a ref on the inode or must have set I_WILL_FREE.
2419  */
2420 int write_inode_now(struct inode *inode, int sync)
2421 {
2422         struct writeback_control wbc = {
2423                 .nr_to_write = LONG_MAX,
2424                 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
2425                 .range_start = 0,
2426                 .range_end = LLONG_MAX,
2427         };
2428 
2429         if (!mapping_cap_writeback_dirty(inode->i_mapping))
2430                 wbc.nr_to_write = 0;
2431 
2432         might_sleep();
2433         return writeback_single_inode(inode, &wbc);
2434 }
2435 EXPORT_SYMBOL(write_inode_now);
2436 
2437 /**
2438  * sync_inode - write an inode and its pages to disk.
2439  * @inode: the inode to sync
2440  * @wbc: controls the writeback mode
2441  *
2442  * sync_inode() will write an inode and its pages to disk.  It will also
2443  * correctly update the inode on its superblock's dirty inode lists and will
2444  * update inode->i_state.
2445  *
2446  * The caller must have a ref on the inode.
2447  */
2448 int sync_inode(struct inode *inode, struct writeback_control *wbc)
2449 {
2450         return writeback_single_inode(inode, wbc);
2451 }
2452 EXPORT_SYMBOL(sync_inode);
2453 
2454 /**
2455  * sync_inode_metadata - write an inode to disk
2456  * @inode: the inode to sync
2457  * @wait: wait for I/O to complete.
2458  *
2459  * Write an inode to disk and adjust its dirty state after completion.
2460  *
2461  * Note: only writes the actual inode, no associated data or other metadata.
2462  */
2463 int sync_inode_metadata(struct inode *inode, int wait)
2464 {
2465         struct writeback_control wbc = {
2466                 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
2467                 .nr_to_write = 0, /* metadata-only */
2468         };
2469 
2470         return sync_inode(inode, &wbc);
2471 }
2472 EXPORT_SYMBOL(sync_inode_metadata);
2473 

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