Version:  2.0.40 2.2.26 2.4.37 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2

Linux/fs/fs-writeback.c

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

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