Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
checkpoint.c
Go to the documentation of this file.
1 /*
2  * linux/fs/jbd2/checkpoint.c
3  *
4  * Written by Stephen C. Tweedie <[email protected]>, 1999
5  *
6  * Copyright 1999 Red Hat Software --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Checkpoint routines for the generic filesystem journaling code.
13  * Part of the ext2fs journaling system.
14  *
15  * Checkpointing is the process of ensuring that a section of the log is
16  * committed fully to disk, so that that portion of the log can be
17  * reused.
18  */
19 
20 #include <linux/time.h>
21 #include <linux/fs.h>
22 #include <linux/jbd2.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/blkdev.h>
26 #include <trace/events/jbd2.h>
27 
28 /*
29  * Unlink a buffer from a transaction checkpoint list.
30  *
31  * Called with j_list_lock held.
32  */
33 static inline void __buffer_unlink_first(struct journal_head *jh)
34 {
36 
37  jh->b_cpnext->b_cpprev = jh->b_cpprev;
38  jh->b_cpprev->b_cpnext = jh->b_cpnext;
39  if (transaction->t_checkpoint_list == jh) {
40  transaction->t_checkpoint_list = jh->b_cpnext;
41  if (transaction->t_checkpoint_list == jh)
42  transaction->t_checkpoint_list = NULL;
43  }
44 }
45 
46 /*
47  * Unlink a buffer from a transaction checkpoint(io) list.
48  *
49  * Called with j_list_lock held.
50  */
51 static inline void __buffer_unlink(struct journal_head *jh)
52 {
53  transaction_t *transaction = jh->b_cp_transaction;
54 
55  __buffer_unlink_first(jh);
56  if (transaction->t_checkpoint_io_list == jh) {
57  transaction->t_checkpoint_io_list = jh->b_cpnext;
58  if (transaction->t_checkpoint_io_list == jh)
59  transaction->t_checkpoint_io_list = NULL;
60  }
61 }
62 
63 /*
64  * Move a buffer from the checkpoint list to the checkpoint io list
65  *
66  * Called with j_list_lock held
67  */
68 static inline void __buffer_relink_io(struct journal_head *jh)
69 {
70  transaction_t *transaction = jh->b_cp_transaction;
71 
72  __buffer_unlink_first(jh);
73 
74  if (!transaction->t_checkpoint_io_list) {
75  jh->b_cpnext = jh->b_cpprev = jh;
76  } else {
77  jh->b_cpnext = transaction->t_checkpoint_io_list;
78  jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
79  jh->b_cpprev->b_cpnext = jh;
80  jh->b_cpnext->b_cpprev = jh;
81  }
82  transaction->t_checkpoint_io_list = jh;
83 }
84 
85 /*
86  * Try to release a checkpointed buffer from its transaction.
87  * Returns 1 if we released it and 2 if we also released the
88  * whole transaction.
89  *
90  * Requires j_list_lock
91  */
92 static int __try_to_free_cp_buf(struct journal_head *jh)
93 {
94  int ret = 0;
95  struct buffer_head *bh = jh2bh(jh);
96 
97  if (jh->b_transaction == NULL && !buffer_locked(bh) &&
98  !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
99  /*
100  * Get our reference so that bh cannot be freed before
101  * we unlock it
102  */
103  get_bh(bh);
104  JBUFFER_TRACE(jh, "remove from checkpoint list");
105  ret = __jbd2_journal_remove_checkpoint(jh) + 1;
106  BUFFER_TRACE(bh, "release");
107  __brelse(bh);
108  }
109  return ret;
110 }
111 
112 /*
113  * __jbd2_log_wait_for_space: wait until there is space in the journal.
114  *
115  * Called under j-state_lock *only*. It will be unlocked if we have to wait
116  * for a checkpoint to free up some space in the log.
117  */
118 void __jbd2_log_wait_for_space(journal_t *journal)
119 {
120  int nblocks, space_left;
121  /* assert_spin_locked(&journal->j_state_lock); */
122 
123  nblocks = jbd_space_needed(journal);
124  while (__jbd2_log_space_left(journal) < nblocks) {
125  if (journal->j_flags & JBD2_ABORT)
126  return;
127  write_unlock(&journal->j_state_lock);
128  mutex_lock(&journal->j_checkpoint_mutex);
129 
130  /*
131  * Test again, another process may have checkpointed while we
132  * were waiting for the checkpoint lock. If there are no
133  * transactions ready to be checkpointed, try to recover
134  * journal space by calling cleanup_journal_tail(), and if
135  * that doesn't work, by waiting for the currently committing
136  * transaction to complete. If there is absolutely no way
137  * to make progress, this is either a BUG or corrupted
138  * filesystem, so abort the journal and leave a stack
139  * trace for forensic evidence.
140  */
141  write_lock(&journal->j_state_lock);
142  spin_lock(&journal->j_list_lock);
143  nblocks = jbd_space_needed(journal);
144  space_left = __jbd2_log_space_left(journal);
145  if (space_left < nblocks) {
146  int chkpt = journal->j_checkpoint_transactions != NULL;
147  tid_t tid = 0;
148 
149  if (journal->j_committing_transaction)
150  tid = journal->j_committing_transaction->t_tid;
151  spin_unlock(&journal->j_list_lock);
152  write_unlock(&journal->j_state_lock);
153  if (chkpt) {
154  jbd2_log_do_checkpoint(journal);
155  } else if (jbd2_cleanup_journal_tail(journal) == 0) {
156  /* We were able to recover space; yay! */
157  ;
158  } else if (tid) {
159  jbd2_log_wait_commit(journal, tid);
160  } else {
161  printk(KERN_ERR "%s: needed %d blocks and "
162  "only had %d space available\n",
163  __func__, nblocks, space_left);
164  printk(KERN_ERR "%s: no way to get more "
165  "journal space in %s\n", __func__,
166  journal->j_devname);
167  WARN_ON(1);
168  jbd2_journal_abort(journal, 0);
169  }
170  write_lock(&journal->j_state_lock);
171  } else {
172  spin_unlock(&journal->j_list_lock);
173  }
174  mutex_unlock(&journal->j_checkpoint_mutex);
175  }
176 }
177 
178 /*
179  * Clean up transaction's list of buffers submitted for io.
180  * We wait for any pending IO to complete and remove any clean
181  * buffers. Note that we take the buffers in the opposite ordering
182  * from the one in which they were submitted for IO.
183  *
184  * Return 0 on success, and return <0 if some buffers have failed
185  * to be written out.
186  *
187  * Called with j_list_lock held.
188  */
189 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
190 {
191  struct journal_head *jh;
192  struct buffer_head *bh;
193  tid_t this_tid;
194  int released = 0;
195  int ret = 0;
196 
197  this_tid = transaction->t_tid;
198 restart:
199  /* Did somebody clean up the transaction in the meanwhile? */
200  if (journal->j_checkpoint_transactions != transaction ||
201  transaction->t_tid != this_tid)
202  return ret;
203  while (!released && transaction->t_checkpoint_io_list) {
204  jh = transaction->t_checkpoint_io_list;
205  bh = jh2bh(jh);
206  get_bh(bh);
207  if (buffer_locked(bh)) {
208  spin_unlock(&journal->j_list_lock);
209  wait_on_buffer(bh);
210  /* the journal_head may have gone by now */
211  BUFFER_TRACE(bh, "brelse");
212  __brelse(bh);
213  spin_lock(&journal->j_list_lock);
214  goto restart;
215  }
216  if (unlikely(buffer_write_io_error(bh)))
217  ret = -EIO;
218 
219  /*
220  * Now in whatever state the buffer currently is, we know that
221  * it has been written out and so we can drop it from the list
222  */
223  released = __jbd2_journal_remove_checkpoint(jh);
224  __brelse(bh);
225  }
226 
227  return ret;
228 }
229 
230 static void
231 __flush_batch(journal_t *journal, int *batch_count)
232 {
233  int i;
234  struct blk_plug plug;
235 
236  blk_start_plug(&plug);
237  for (i = 0; i < *batch_count; i++)
238  write_dirty_buffer(journal->j_chkpt_bhs[i], WRITE_SYNC);
239  blk_finish_plug(&plug);
240 
241  for (i = 0; i < *batch_count; i++) {
242  struct buffer_head *bh = journal->j_chkpt_bhs[i];
243  BUFFER_TRACE(bh, "brelse");
244  __brelse(bh);
245  }
246  *batch_count = 0;
247 }
248 
249 /*
250  * Try to flush one buffer from the checkpoint list to disk.
251  *
252  * Return 1 if something happened which requires us to abort the current
253  * scan of the checkpoint list. Return <0 if the buffer has failed to
254  * be written out.
255  *
256  * Called with j_list_lock held and drops it if 1 is returned
257  */
258 static int __process_buffer(journal_t *journal, struct journal_head *jh,
259  int *batch_count, transaction_t *transaction)
260 {
261  struct buffer_head *bh = jh2bh(jh);
262  int ret = 0;
263 
264  if (buffer_locked(bh)) {
265  get_bh(bh);
266  spin_unlock(&journal->j_list_lock);
267  wait_on_buffer(bh);
268  /* the journal_head may have gone by now */
269  BUFFER_TRACE(bh, "brelse");
270  __brelse(bh);
271  ret = 1;
272  } else if (jh->b_transaction != NULL) {
274  tid_t tid = t->t_tid;
275 
276  transaction->t_chp_stats.cs_forced_to_close++;
277  spin_unlock(&journal->j_list_lock);
278  if (unlikely(journal->j_flags & JBD2_UNMOUNT))
279  /*
280  * The journal thread is dead; so starting and
281  * waiting for a commit to finish will cause
282  * us to wait for a _very_ long time.
283  */
284  printk(KERN_ERR "JBD2: %s: "
285  "Waiting for Godot: block %llu\n",
286  journal->j_devname,
287  (unsigned long long) bh->b_blocknr);
288  jbd2_log_start_commit(journal, tid);
289  jbd2_log_wait_commit(journal, tid);
290  ret = 1;
291  } else if (!buffer_dirty(bh)) {
292  ret = 1;
293  if (unlikely(buffer_write_io_error(bh)))
294  ret = -EIO;
295  get_bh(bh);
296  BUFFER_TRACE(bh, "remove from checkpoint");
298  spin_unlock(&journal->j_list_lock);
299  __brelse(bh);
300  } else {
301  /*
302  * Important: we are about to write the buffer, and
303  * possibly block, while still holding the journal lock.
304  * We cannot afford to let the transaction logic start
305  * messing around with this buffer before we write it to
306  * disk, as that would break recoverability.
307  */
308  BUFFER_TRACE(bh, "queue");
309  get_bh(bh);
310  J_ASSERT_BH(bh, !buffer_jwrite(bh));
311  journal->j_chkpt_bhs[*batch_count] = bh;
312  __buffer_relink_io(jh);
313  transaction->t_chp_stats.cs_written++;
314  (*batch_count)++;
315  if (*batch_count == JBD2_NR_BATCH) {
316  spin_unlock(&journal->j_list_lock);
317  __flush_batch(journal, batch_count);
318  ret = 1;
319  }
320  }
321  return ret;
322 }
323 
324 /*
325  * Perform an actual checkpoint. We take the first transaction on the
326  * list of transactions to be checkpointed and send all its buffers
327  * to disk. We submit larger chunks of data at once.
328  *
329  * The journal should be locked before calling this function.
330  * Called with j_checkpoint_mutex held.
331  */
332 int jbd2_log_do_checkpoint(journal_t *journal)
333 {
335  tid_t this_tid;
336  int result;
337 
338  jbd_debug(1, "Start checkpoint\n");
339 
340  /*
341  * First thing: if there are any transactions in the log which
342  * don't need checkpointing, just eliminate them from the
343  * journal straight away.
344  */
345  result = jbd2_cleanup_journal_tail(journal);
346  trace_jbd2_checkpoint(journal, result);
347  jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
348  if (result <= 0)
349  return result;
350 
351  /*
352  * OK, we need to start writing disk blocks. Take one transaction
353  * and write it.
354  */
355  result = 0;
356  spin_lock(&journal->j_list_lock);
357  if (!journal->j_checkpoint_transactions)
358  goto out;
359  transaction = journal->j_checkpoint_transactions;
360  if (transaction->t_chp_stats.cs_chp_time == 0)
361  transaction->t_chp_stats.cs_chp_time = jiffies;
362  this_tid = transaction->t_tid;
363 restart:
364  /*
365  * If someone cleaned up this transaction while we slept, we're
366  * done (maybe it's a new transaction, but it fell at the same
367  * address).
368  */
369  if (journal->j_checkpoint_transactions == transaction &&
370  transaction->t_tid == this_tid) {
371  int batch_count = 0;
372  struct journal_head *jh;
373  int retry = 0, err;
374 
375  while (!retry && transaction->t_checkpoint_list) {
376  jh = transaction->t_checkpoint_list;
377  retry = __process_buffer(journal, jh, &batch_count,
378  transaction);
379  if (retry < 0 && !result)
380  result = retry;
381  if (!retry && (need_resched() ||
382  spin_needbreak(&journal->j_list_lock))) {
383  spin_unlock(&journal->j_list_lock);
384  retry = 1;
385  break;
386  }
387  }
388 
389  if (batch_count) {
390  if (!retry) {
391  spin_unlock(&journal->j_list_lock);
392  retry = 1;
393  }
394  __flush_batch(journal, &batch_count);
395  }
396 
397  if (retry) {
398  spin_lock(&journal->j_list_lock);
399  goto restart;
400  }
401  /*
402  * Now we have cleaned up the first transaction's checkpoint
403  * list. Let's clean up the second one
404  */
405  err = __wait_cp_io(journal, transaction);
406  if (!result)
407  result = err;
408  }
409 out:
410  spin_unlock(&journal->j_list_lock);
411  if (result < 0)
412  jbd2_journal_abort(journal, result);
413  else
414  result = jbd2_cleanup_journal_tail(journal);
415 
416  return (result < 0) ? result : 0;
417 }
418 
419 /*
420  * Check the list of checkpoint transactions for the journal to see if
421  * we have already got rid of any since the last update of the log tail
422  * in the journal superblock. If so, we can instantly roll the
423  * superblock forward to remove those transactions from the log.
424  *
425  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
426  *
427  * Called with the journal lock held.
428  *
429  * This is the only part of the journaling code which really needs to be
430  * aware of transaction aborts. Checkpointing involves writing to the
431  * main filesystem area rather than to the journal, so it can proceed
432  * even in abort state, but we must not update the super block if
433  * checkpointing may have failed. Otherwise, we would lose some metadata
434  * buffers which should be written-back to the filesystem.
435  */
436 
437 int jbd2_cleanup_journal_tail(journal_t *journal)
438 {
439  tid_t first_tid;
440  unsigned long blocknr;
441 
442  if (is_journal_aborted(journal))
443  return 1;
444 
445  if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
446  return 1;
447  J_ASSERT(blocknr != 0);
448 
449  /*
450  * We need to make sure that any blocks that were recently written out
451  * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
452  * we drop the transactions from the journal. It's unlikely this will
453  * be necessary, especially with an appropriately sized journal, but we
454  * need this to guarantee correctness. Fortunately
455  * jbd2_cleanup_journal_tail() doesn't get called all that often.
456  */
457  if (journal->j_flags & JBD2_BARRIER)
458  blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
459 
460  __jbd2_update_log_tail(journal, first_tid, blocknr);
461  return 0;
462 }
463 
464 
465 /* Checkpoint list management */
466 
467 /*
468  * journal_clean_one_cp_list
469  *
470  * Find all the written-back checkpoint buffers in the given list and
471  * release them.
472  *
473  * Called with the journal locked.
474  * Called with j_list_lock held.
475  * Returns number of buffers reaped (for debug)
476  */
477 
478 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
479 {
480  struct journal_head *last_jh;
481  struct journal_head *next_jh = jh;
482  int ret, freed = 0;
483 
484  *released = 0;
485  if (!jh)
486  return 0;
487 
488  last_jh = jh->b_cpprev;
489  do {
490  jh = next_jh;
491  next_jh = jh->b_cpnext;
492  ret = __try_to_free_cp_buf(jh);
493  if (ret) {
494  freed++;
495  if (ret == 2) {
496  *released = 1;
497  return freed;
498  }
499  }
500  /*
501  * This function only frees up some memory
502  * if possible so we dont have an obligation
503  * to finish processing. Bail out if preemption
504  * requested:
505  */
506  if (need_resched())
507  return freed;
508  } while (jh != last_jh);
509 
510  return freed;
511 }
512 
513 /*
514  * journal_clean_checkpoint_list
515  *
516  * Find all the written-back checkpoint buffers in the journal and release them.
517  *
518  * Called with the journal locked.
519  * Called with j_list_lock held.
520  * Returns number of buffers reaped (for debug)
521  */
522 
524 {
525  transaction_t *transaction, *last_transaction, *next_transaction;
526  int ret = 0;
527  int released;
528 
529  transaction = journal->j_checkpoint_transactions;
530  if (!transaction)
531  goto out;
532 
533  last_transaction = transaction->t_cpprev;
534  next_transaction = transaction;
535  do {
536  transaction = next_transaction;
537  next_transaction = transaction->t_cpnext;
538  ret += journal_clean_one_cp_list(transaction->
539  t_checkpoint_list, &released);
540  /*
541  * This function only frees up some memory if possible so we
542  * dont have an obligation to finish processing. Bail out if
543  * preemption requested:
544  */
545  if (need_resched())
546  goto out;
547  if (released)
548  continue;
549  /*
550  * It is essential that we are as careful as in the case of
551  * t_checkpoint_list with removing the buffer from the list as
552  * we can possibly see not yet submitted buffers on io_list
553  */
554  ret += journal_clean_one_cp_list(transaction->
555  t_checkpoint_io_list, &released);
556  if (need_resched())
557  goto out;
558  } while (transaction != last_transaction);
559 out:
560  return ret;
561 }
562 
563 /*
564  * journal_remove_checkpoint: called after a buffer has been committed
565  * to disk (either by being write-back flushed to disk, or being
566  * committed to the log).
567  *
568  * We cannot safely clean a transaction out of the log until all of the
569  * buffer updates committed in that transaction have safely been stored
570  * elsewhere on disk. To achieve this, all of the buffers in a
571  * transaction need to be maintained on the transaction's checkpoint
572  * lists until they have been rewritten, at which point this function is
573  * called to remove the buffer from the existing transaction's
574  * checkpoint lists.
575  *
576  * The function returns 1 if it frees the transaction, 0 otherwise.
577  * The function can free jh and bh.
578  *
579  * This function is called with j_list_lock held.
580  */
582 {
583  struct transaction_chp_stats_s *stats;
585  journal_t *journal;
586  int ret = 0;
587 
588  JBUFFER_TRACE(jh, "entry");
589 
590  if ((transaction = jh->b_cp_transaction) == NULL) {
591  JBUFFER_TRACE(jh, "not on transaction");
592  goto out;
593  }
594  journal = transaction->t_journal;
595 
596  JBUFFER_TRACE(jh, "removing from transaction");
597  __buffer_unlink(jh);
598  jh->b_cp_transaction = NULL;
600 
601  if (transaction->t_checkpoint_list != NULL ||
602  transaction->t_checkpoint_io_list != NULL)
603  goto out;
604 
605  /*
606  * There is one special case to worry about: if we have just pulled the
607  * buffer off a running or committing transaction's checkpoing list,
608  * then even if the checkpoint list is empty, the transaction obviously
609  * cannot be dropped!
610  *
611  * The locking here around t_state is a bit sleazy.
612  * See the comment at the end of jbd2_journal_commit_transaction().
613  */
614  if (transaction->t_state != T_FINISHED)
615  goto out;
616 
617  /* OK, that was the last buffer for the transaction: we can now
618  safely remove this transaction from the log */
619  stats = &transaction->t_chp_stats;
620  if (stats->cs_chp_time)
621  stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
622  jiffies);
623  trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
624  transaction->t_tid, stats);
625 
626  __jbd2_journal_drop_transaction(journal, transaction);
627  jbd2_journal_free_transaction(transaction);
628 
629  /* Just in case anybody was waiting for more transactions to be
630  checkpointed... */
631  wake_up(&journal->j_wait_logspace);
632  ret = 1;
633 out:
634  return ret;
635 }
636 
637 /*
638  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
639  * list so that we know when it is safe to clean the transaction out of
640  * the log.
641  *
642  * Called with the journal locked.
643  * Called with j_list_lock held.
644  */
646  transaction_t *transaction)
647 {
648  JBUFFER_TRACE(jh, "entry");
649  J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
650  J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
651 
652  /* Get reference for checkpointing transaction */
655 
656  if (!transaction->t_checkpoint_list) {
657  jh->b_cpnext = jh->b_cpprev = jh;
658  } else {
659  jh->b_cpnext = transaction->t_checkpoint_list;
660  jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
661  jh->b_cpprev->b_cpnext = jh;
662  jh->b_cpnext->b_cpprev = jh;
663  }
664  transaction->t_checkpoint_list = jh;
665 }
666 
667 /*
668  * We've finished with this transaction structure: adios...
669  *
670  * The transaction must have no links except for the checkpoint by this
671  * point.
672  *
673  * Called with the journal locked.
674  * Called with j_list_lock held.
675  */
676 
677 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
678 {
679  assert_spin_locked(&journal->j_list_lock);
680  if (transaction->t_cpnext) {
681  transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
682  transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
683  if (journal->j_checkpoint_transactions == transaction)
684  journal->j_checkpoint_transactions =
685  transaction->t_cpnext;
686  if (journal->j_checkpoint_transactions == transaction)
687  journal->j_checkpoint_transactions = NULL;
688  }
689 
690  J_ASSERT(transaction->t_state == T_FINISHED);
691  J_ASSERT(transaction->t_buffers == NULL);
692  J_ASSERT(transaction->t_forget == NULL);
693  J_ASSERT(transaction->t_iobuf_list == NULL);
694  J_ASSERT(transaction->t_shadow_list == NULL);
695  J_ASSERT(transaction->t_log_list == NULL);
696  J_ASSERT(transaction->t_checkpoint_list == NULL);
697  J_ASSERT(transaction->t_checkpoint_io_list == NULL);
698  J_ASSERT(atomic_read(&transaction->t_updates) == 0);
699  J_ASSERT(journal->j_committing_transaction != transaction);
700  J_ASSERT(journal->j_running_transaction != transaction);
701 
702  trace_jbd2_drop_transaction(journal, transaction);
703 
704  jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
705 }