538417c1fdbbcbb630cbcdb885c74ebe5f433844
[linux-2.6.git] / fs / jbd2 / journal.c
1 /*
2  * linux/fs/jbd2/journal.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- 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  * Generic filesystem journal-writing code; part of the ext2fs
13  * journaling system.
14  *
15  * This file manages journals: areas of disk reserved for logging
16  * transactional updates.  This includes the kernel journaling thread
17  * which is responsible for scheduling updates to the log.
18  *
19  * We do not actually manage the physical storage of the journal in this
20  * file: that is left to a per-journal policy function, which allows us
21  * to store the journal within a filesystem-specified area for ext2
22  * journaling (ext2 can use a reserved inode for storing the log).
23  */
24
25 #include <linux/module.h>
26 #include <linux/time.h>
27 #include <linux/fs.h>
28 #include <linux/jbd2.h>
29 #include <linux/errno.h>
30 #include <linux/slab.h>
31 #include <linux/init.h>
32 #include <linux/mm.h>
33 #include <linux/freezer.h>
34 #include <linux/pagemap.h>
35 #include <linux/kthread.h>
36 #include <linux/poison.h>
37 #include <linux/proc_fs.h>
38 #include <linux/debugfs.h>
39 #include <linux/seq_file.h>
40 #include <linux/math64.h>
41 #include <linux/hash.h>
42 #include <linux/log2.h>
43 #include <linux/vmalloc.h>
44 #include <linux/backing-dev.h>
45 #include <linux/bitops.h>
46
47 #define CREATE_TRACE_POINTS
48 #include <trace/events/jbd2.h>
49
50 #include <asm/uaccess.h>
51 #include <asm/page.h>
52 #include <asm/system.h>
53
54 EXPORT_SYMBOL(jbd2_journal_extend);
55 EXPORT_SYMBOL(jbd2_journal_stop);
56 EXPORT_SYMBOL(jbd2_journal_lock_updates);
57 EXPORT_SYMBOL(jbd2_journal_unlock_updates);
58 EXPORT_SYMBOL(jbd2_journal_get_write_access);
59 EXPORT_SYMBOL(jbd2_journal_get_create_access);
60 EXPORT_SYMBOL(jbd2_journal_get_undo_access);
61 EXPORT_SYMBOL(jbd2_journal_set_triggers);
62 EXPORT_SYMBOL(jbd2_journal_dirty_metadata);
63 EXPORT_SYMBOL(jbd2_journal_release_buffer);
64 EXPORT_SYMBOL(jbd2_journal_forget);
65 #if 0
66 EXPORT_SYMBOL(journal_sync_buffer);
67 #endif
68 EXPORT_SYMBOL(jbd2_journal_flush);
69 EXPORT_SYMBOL(jbd2_journal_revoke);
70
71 EXPORT_SYMBOL(jbd2_journal_init_dev);
72 EXPORT_SYMBOL(jbd2_journal_init_inode);
73 EXPORT_SYMBOL(jbd2_journal_update_format);
74 EXPORT_SYMBOL(jbd2_journal_check_used_features);
75 EXPORT_SYMBOL(jbd2_journal_check_available_features);
76 EXPORT_SYMBOL(jbd2_journal_set_features);
77 EXPORT_SYMBOL(jbd2_journal_load);
78 EXPORT_SYMBOL(jbd2_journal_destroy);
79 EXPORT_SYMBOL(jbd2_journal_abort);
80 EXPORT_SYMBOL(jbd2_journal_errno);
81 EXPORT_SYMBOL(jbd2_journal_ack_err);
82 EXPORT_SYMBOL(jbd2_journal_clear_err);
83 EXPORT_SYMBOL(jbd2_log_wait_commit);
84 EXPORT_SYMBOL(jbd2_log_start_commit);
85 EXPORT_SYMBOL(jbd2_journal_start_commit);
86 EXPORT_SYMBOL(jbd2_journal_force_commit_nested);
87 EXPORT_SYMBOL(jbd2_journal_wipe);
88 EXPORT_SYMBOL(jbd2_journal_blocks_per_page);
89 EXPORT_SYMBOL(jbd2_journal_invalidatepage);
90 EXPORT_SYMBOL(jbd2_journal_try_to_free_buffers);
91 EXPORT_SYMBOL(jbd2_journal_force_commit);
92 EXPORT_SYMBOL(jbd2_journal_file_inode);
93 EXPORT_SYMBOL(jbd2_journal_init_jbd_inode);
94 EXPORT_SYMBOL(jbd2_journal_release_jbd_inode);
95 EXPORT_SYMBOL(jbd2_journal_begin_ordered_truncate);
96
97 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
98 static void __journal_abort_soft (journal_t *journal, int errno);
99 static int jbd2_journal_create_slab(size_t slab_size);
100
101 /*
102  * Helper function used to manage commit timeouts
103  */
104
105 static void commit_timeout(unsigned long __data)
106 {
107         struct task_struct * p = (struct task_struct *) __data;
108
109         wake_up_process(p);
110 }
111
112 /*
113  * kjournald2: The main thread function used to manage a logging device
114  * journal.
115  *
116  * This kernel thread is responsible for two things:
117  *
118  * 1) COMMIT:  Every so often we need to commit the current state of the
119  *    filesystem to disk.  The journal thread is responsible for writing
120  *    all of the metadata buffers to disk.
121  *
122  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
123  *    of the data in that part of the log has been rewritten elsewhere on
124  *    the disk.  Flushing these old buffers to reclaim space in the log is
125  *    known as checkpointing, and this thread is responsible for that job.
126  */
127
128 static int kjournald2(void *arg)
129 {
130         journal_t *journal = arg;
131         transaction_t *transaction;
132
133         /*
134          * Set up an interval timer which can be used to trigger a commit wakeup
135          * after the commit interval expires
136          */
137         setup_timer(&journal->j_commit_timer, commit_timeout,
138                         (unsigned long)current);
139
140         /* Record that the journal thread is running */
141         journal->j_task = current;
142         wake_up(&journal->j_wait_done_commit);
143
144         /*
145          * And now, wait forever for commit wakeup events.
146          */
147         write_lock(&journal->j_state_lock);
148
149 loop:
150         if (journal->j_flags & JBD2_UNMOUNT)
151                 goto end_loop;
152
153         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
154                 journal->j_commit_sequence, journal->j_commit_request);
155
156         if (journal->j_commit_sequence != journal->j_commit_request) {
157                 jbd_debug(1, "OK, requests differ\n");
158                 write_unlock(&journal->j_state_lock);
159                 del_timer_sync(&journal->j_commit_timer);
160                 jbd2_journal_commit_transaction(journal);
161                 write_lock(&journal->j_state_lock);
162                 goto loop;
163         }
164
165         wake_up(&journal->j_wait_done_commit);
166         if (freezing(current)) {
167                 /*
168                  * The simpler the better. Flushing journal isn't a
169                  * good idea, because that depends on threads that may
170                  * be already stopped.
171                  */
172                 jbd_debug(1, "Now suspending kjournald2\n");
173                 write_unlock(&journal->j_state_lock);
174                 refrigerator();
175                 write_lock(&journal->j_state_lock);
176         } else {
177                 /*
178                  * We assume on resume that commits are already there,
179                  * so we don't sleep
180                  */
181                 DEFINE_WAIT(wait);
182                 int should_sleep = 1;
183
184                 prepare_to_wait(&journal->j_wait_commit, &wait,
185                                 TASK_INTERRUPTIBLE);
186                 if (journal->j_commit_sequence != journal->j_commit_request)
187                         should_sleep = 0;
188                 transaction = journal->j_running_transaction;
189                 if (transaction && time_after_eq(jiffies,
190                                                 transaction->t_expires))
191                         should_sleep = 0;
192                 if (journal->j_flags & JBD2_UNMOUNT)
193                         should_sleep = 0;
194                 if (should_sleep) {
195                         write_unlock(&journal->j_state_lock);
196                         schedule();
197                         write_lock(&journal->j_state_lock);
198                 }
199                 finish_wait(&journal->j_wait_commit, &wait);
200         }
201
202         jbd_debug(1, "kjournald2 wakes\n");
203
204         /*
205          * Were we woken up by a commit wakeup event?
206          */
207         transaction = journal->j_running_transaction;
208         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
209                 journal->j_commit_request = transaction->t_tid;
210                 jbd_debug(1, "woke because of timeout\n");
211         }
212         goto loop;
213
214 end_loop:
215         write_unlock(&journal->j_state_lock);
216         del_timer_sync(&journal->j_commit_timer);
217         journal->j_task = NULL;
218         wake_up(&journal->j_wait_done_commit);
219         jbd_debug(1, "Journal thread exiting.\n");
220         return 0;
221 }
222
223 static int jbd2_journal_start_thread(journal_t *journal)
224 {
225         struct task_struct *t;
226
227         t = kthread_run(kjournald2, journal, "jbd2/%s",
228                         journal->j_devname);
229         if (IS_ERR(t))
230                 return PTR_ERR(t);
231
232         wait_event(journal->j_wait_done_commit, journal->j_task != NULL);
233         return 0;
234 }
235
236 static void journal_kill_thread(journal_t *journal)
237 {
238         write_lock(&journal->j_state_lock);
239         journal->j_flags |= JBD2_UNMOUNT;
240
241         while (journal->j_task) {
242                 wake_up(&journal->j_wait_commit);
243                 write_unlock(&journal->j_state_lock);
244                 wait_event(journal->j_wait_done_commit, journal->j_task == NULL);
245                 write_lock(&journal->j_state_lock);
246         }
247         write_unlock(&journal->j_state_lock);
248 }
249
250 /*
251  * jbd2_journal_write_metadata_buffer: write a metadata buffer to the journal.
252  *
253  * Writes a metadata buffer to a given disk block.  The actual IO is not
254  * performed but a new buffer_head is constructed which labels the data
255  * to be written with the correct destination disk block.
256  *
257  * Any magic-number escaping which needs to be done will cause a
258  * copy-out here.  If the buffer happens to start with the
259  * JBD2_MAGIC_NUMBER, then we can't write it to the log directly: the
260  * magic number is only written to the log for descripter blocks.  In
261  * this case, we copy the data and replace the first word with 0, and we
262  * return a result code which indicates that this buffer needs to be
263  * marked as an escaped buffer in the corresponding log descriptor
264  * block.  The missing word can then be restored when the block is read
265  * during recovery.
266  *
267  * If the source buffer has already been modified by a new transaction
268  * since we took the last commit snapshot, we use the frozen copy of
269  * that data for IO.  If we end up using the existing buffer_head's data
270  * for the write, then we *have* to lock the buffer to prevent anyone
271  * else from using and possibly modifying it while the IO is in
272  * progress.
273  *
274  * The function returns a pointer to the buffer_heads to be used for IO.
275  *
276  * We assume that the journal has already been locked in this function.
277  *
278  * Return value:
279  *  <0: Error
280  * >=0: Finished OK
281  *
282  * On success:
283  * Bit 0 set == escape performed on the data
284  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
285  */
286
287 int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
288                                   struct journal_head  *jh_in,
289                                   struct journal_head **jh_out,
290                                   unsigned long long blocknr)
291 {
292         int need_copy_out = 0;
293         int done_copy_out = 0;
294         int do_escape = 0;
295         char *mapped_data;
296         struct buffer_head *new_bh;
297         struct journal_head *new_jh;
298         struct page *new_page;
299         unsigned int new_offset;
300         struct buffer_head *bh_in = jh2bh(jh_in);
301         journal_t *journal = transaction->t_journal;
302
303         /*
304          * The buffer really shouldn't be locked: only the current committing
305          * transaction is allowed to write it, so nobody else is allowed
306          * to do any IO.
307          *
308          * akpm: except if we're journalling data, and write() output is
309          * also part of a shared mapping, and another thread has
310          * decided to launch a writepage() against this buffer.
311          */
312         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
313
314 retry_alloc:
315         new_bh = alloc_buffer_head(GFP_NOFS);
316         if (!new_bh) {
317                 /*
318                  * Failure is not an option, but __GFP_NOFAIL is going
319                  * away; so we retry ourselves here.
320                  */
321                 congestion_wait(BLK_RW_ASYNC, HZ/50);
322                 goto retry_alloc;
323         }
324
325         /* keep subsequent assertions sane */
326         new_bh->b_state = 0;
327         init_buffer(new_bh, NULL, NULL);
328         atomic_set(&new_bh->b_count, 1);
329         new_jh = jbd2_journal_add_journal_head(new_bh); /* This sleeps */
330
331         /*
332          * If a new transaction has already done a buffer copy-out, then
333          * we use that version of the data for the commit.
334          */
335         jbd_lock_bh_state(bh_in);
336 repeat:
337         if (jh_in->b_frozen_data) {
338                 done_copy_out = 1;
339                 new_page = virt_to_page(jh_in->b_frozen_data);
340                 new_offset = offset_in_page(jh_in->b_frozen_data);
341         } else {
342                 new_page = jh2bh(jh_in)->b_page;
343                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
344         }
345
346         mapped_data = kmap_atomic(new_page, KM_USER0);
347         /*
348          * Fire data frozen trigger if data already wasn't frozen.  Do this
349          * before checking for escaping, as the trigger may modify the magic
350          * offset.  If a copy-out happens afterwards, it will have the correct
351          * data in the buffer.
352          */
353         if (!done_copy_out)
354                 jbd2_buffer_frozen_trigger(jh_in, mapped_data + new_offset,
355                                            jh_in->b_triggers);
356
357         /*
358          * Check for escaping
359          */
360         if (*((__be32 *)(mapped_data + new_offset)) ==
361                                 cpu_to_be32(JBD2_MAGIC_NUMBER)) {
362                 need_copy_out = 1;
363                 do_escape = 1;
364         }
365         kunmap_atomic(mapped_data, KM_USER0);
366
367         /*
368          * Do we need to do a data copy?
369          */
370         if (need_copy_out && !done_copy_out) {
371                 char *tmp;
372
373                 jbd_unlock_bh_state(bh_in);
374                 tmp = jbd2_alloc(bh_in->b_size, GFP_NOFS);
375                 if (!tmp) {
376                         jbd2_journal_put_journal_head(new_jh);
377                         return -ENOMEM;
378                 }
379                 jbd_lock_bh_state(bh_in);
380                 if (jh_in->b_frozen_data) {
381                         jbd2_free(tmp, bh_in->b_size);
382                         goto repeat;
383                 }
384
385                 jh_in->b_frozen_data = tmp;
386                 mapped_data = kmap_atomic(new_page, KM_USER0);
387                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
388                 kunmap_atomic(mapped_data, KM_USER0);
389
390                 new_page = virt_to_page(tmp);
391                 new_offset = offset_in_page(tmp);
392                 done_copy_out = 1;
393
394                 /*
395                  * This isn't strictly necessary, as we're using frozen
396                  * data for the escaping, but it keeps consistency with
397                  * b_frozen_data usage.
398                  */
399                 jh_in->b_frozen_triggers = jh_in->b_triggers;
400         }
401
402         /*
403          * Did we need to do an escaping?  Now we've done all the
404          * copying, we can finally do so.
405          */
406         if (do_escape) {
407                 mapped_data = kmap_atomic(new_page, KM_USER0);
408                 *((unsigned int *)(mapped_data + new_offset)) = 0;
409                 kunmap_atomic(mapped_data, KM_USER0);
410         }
411
412         set_bh_page(new_bh, new_page, new_offset);
413         new_jh->b_transaction = NULL;
414         new_bh->b_size = jh2bh(jh_in)->b_size;
415         new_bh->b_bdev = transaction->t_journal->j_dev;
416         new_bh->b_blocknr = blocknr;
417         set_buffer_mapped(new_bh);
418         set_buffer_dirty(new_bh);
419
420         *jh_out = new_jh;
421
422         /*
423          * The to-be-written buffer needs to get moved to the io queue,
424          * and the original buffer whose contents we are shadowing or
425          * copying is moved to the transaction's shadow queue.
426          */
427         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
428         spin_lock(&journal->j_list_lock);
429         __jbd2_journal_file_buffer(jh_in, transaction, BJ_Shadow);
430         spin_unlock(&journal->j_list_lock);
431         jbd_unlock_bh_state(bh_in);
432
433         JBUFFER_TRACE(new_jh, "file as BJ_IO");
434         jbd2_journal_file_buffer(new_jh, transaction, BJ_IO);
435
436         return do_escape | (done_copy_out << 1);
437 }
438
439 /*
440  * Allocation code for the journal file.  Manage the space left in the
441  * journal, so that we can begin checkpointing when appropriate.
442  */
443
444 /*
445  * __jbd2_log_space_left: Return the number of free blocks left in the journal.
446  *
447  * Called with the journal already locked.
448  *
449  * Called under j_state_lock
450  */
451
452 int __jbd2_log_space_left(journal_t *journal)
453 {
454         int left = journal->j_free;
455
456         /* assert_spin_locked(&journal->j_state_lock); */
457
458         /*
459          * Be pessimistic here about the number of those free blocks which
460          * might be required for log descriptor control blocks.
461          */
462
463 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
464
465         left -= MIN_LOG_RESERVED_BLOCKS;
466
467         if (left <= 0)
468                 return 0;
469         left -= (left >> 3);
470         return left;
471 }
472
473 /*
474  * Called under j_state_lock.  Returns true if a transaction commit was started.
475  */
476 int __jbd2_log_start_commit(journal_t *journal, tid_t target)
477 {
478         /*
479          * Are we already doing a recent enough commit?
480          */
481         if (!tid_geq(journal->j_commit_request, target)) {
482                 /*
483                  * We want a new commit: OK, mark the request and wakeup the
484                  * commit thread.  We do _not_ do the commit ourselves.
485                  */
486
487                 journal->j_commit_request = target;
488                 jbd_debug(1, "JBD: requesting commit %d/%d\n",
489                           journal->j_commit_request,
490                           journal->j_commit_sequence);
491                 wake_up(&journal->j_wait_commit);
492                 return 1;
493         }
494         return 0;
495 }
496
497 int jbd2_log_start_commit(journal_t *journal, tid_t tid)
498 {
499         int ret;
500
501         write_lock(&journal->j_state_lock);
502         ret = __jbd2_log_start_commit(journal, tid);
503         write_unlock(&journal->j_state_lock);
504         return ret;
505 }
506
507 /*
508  * Force and wait upon a commit if the calling process is not within
509  * transaction.  This is used for forcing out undo-protected data which contains
510  * bitmaps, when the fs is running out of space.
511  *
512  * We can only force the running transaction if we don't have an active handle;
513  * otherwise, we will deadlock.
514  *
515  * Returns true if a transaction was started.
516  */
517 int jbd2_journal_force_commit_nested(journal_t *journal)
518 {
519         transaction_t *transaction = NULL;
520         tid_t tid;
521
522         read_lock(&journal->j_state_lock);
523         if (journal->j_running_transaction && !current->journal_info) {
524                 transaction = journal->j_running_transaction;
525                 __jbd2_log_start_commit(journal, transaction->t_tid);
526         } else if (journal->j_committing_transaction)
527                 transaction = journal->j_committing_transaction;
528
529         if (!transaction) {
530                 read_unlock(&journal->j_state_lock);
531                 return 0;       /* Nothing to retry */
532         }
533
534         tid = transaction->t_tid;
535         read_unlock(&journal->j_state_lock);
536         jbd2_log_wait_commit(journal, tid);
537         return 1;
538 }
539
540 /*
541  * Start a commit of the current running transaction (if any).  Returns true
542  * if a transaction is going to be committed (or is currently already
543  * committing), and fills its tid in at *ptid
544  */
545 int jbd2_journal_start_commit(journal_t *journal, tid_t *ptid)
546 {
547         int ret = 0;
548
549         write_lock(&journal->j_state_lock);
550         if (journal->j_running_transaction) {
551                 tid_t tid = journal->j_running_transaction->t_tid;
552
553                 __jbd2_log_start_commit(journal, tid);
554                 /* There's a running transaction and we've just made sure
555                  * it's commit has been scheduled. */
556                 if (ptid)
557                         *ptid = tid;
558                 ret = 1;
559         } else if (journal->j_committing_transaction) {
560                 /*
561                  * If ext3_write_super() recently started a commit, then we
562                  * have to wait for completion of that transaction
563                  */
564                 if (ptid)
565                         *ptid = journal->j_committing_transaction->t_tid;
566                 ret = 1;
567         }
568         write_unlock(&journal->j_state_lock);
569         return ret;
570 }
571
572 /*
573  * Wait for a specified commit to complete.
574  * The caller may not hold the journal lock.
575  */
576 int jbd2_log_wait_commit(journal_t *journal, tid_t tid)
577 {
578         int err = 0;
579
580         read_lock(&journal->j_state_lock);
581 #ifdef CONFIG_JBD2_DEBUG
582         if (!tid_geq(journal->j_commit_request, tid)) {
583                 printk(KERN_EMERG
584                        "%s: error: j_commit_request=%d, tid=%d\n",
585                        __func__, journal->j_commit_request, tid);
586         }
587 #endif
588         while (tid_gt(tid, journal->j_commit_sequence)) {
589                 jbd_debug(1, "JBD: want %d, j_commit_sequence=%d\n",
590                                   tid, journal->j_commit_sequence);
591                 wake_up(&journal->j_wait_commit);
592                 read_unlock(&journal->j_state_lock);
593                 wait_event(journal->j_wait_done_commit,
594                                 !tid_gt(tid, journal->j_commit_sequence));
595                 read_lock(&journal->j_state_lock);
596         }
597         read_unlock(&journal->j_state_lock);
598
599         if (unlikely(is_journal_aborted(journal))) {
600                 printk(KERN_EMERG "journal commit I/O error\n");
601                 err = -EIO;
602         }
603         return err;
604 }
605
606 /*
607  * Log buffer allocation routines:
608  */
609
610 int jbd2_journal_next_log_block(journal_t *journal, unsigned long long *retp)
611 {
612         unsigned long blocknr;
613
614         write_lock(&journal->j_state_lock);
615         J_ASSERT(journal->j_free > 1);
616
617         blocknr = journal->j_head;
618         journal->j_head++;
619         journal->j_free--;
620         if (journal->j_head == journal->j_last)
621                 journal->j_head = journal->j_first;
622         write_unlock(&journal->j_state_lock);
623         return jbd2_journal_bmap(journal, blocknr, retp);
624 }
625
626 /*
627  * Conversion of logical to physical block numbers for the journal
628  *
629  * On external journals the journal blocks are identity-mapped, so
630  * this is a no-op.  If needed, we can use j_blk_offset - everything is
631  * ready.
632  */
633 int jbd2_journal_bmap(journal_t *journal, unsigned long blocknr,
634                  unsigned long long *retp)
635 {
636         int err = 0;
637         unsigned long long ret;
638
639         if (journal->j_inode) {
640                 ret = bmap(journal->j_inode, blocknr);
641                 if (ret)
642                         *retp = ret;
643                 else {
644                         printk(KERN_ALERT "%s: journal block not found "
645                                         "at offset %lu on %s\n",
646                                __func__, blocknr, journal->j_devname);
647                         err = -EIO;
648                         __journal_abort_soft(journal, err);
649                 }
650         } else {
651                 *retp = blocknr; /* +journal->j_blk_offset */
652         }
653         return err;
654 }
655
656 /*
657  * We play buffer_head aliasing tricks to write data/metadata blocks to
658  * the journal without copying their contents, but for journal
659  * descriptor blocks we do need to generate bona fide buffers.
660  *
661  * After the caller of jbd2_journal_get_descriptor_buffer() has finished modifying
662  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
663  * But we don't bother doing that, so there will be coherency problems with
664  * mmaps of blockdevs which hold live JBD-controlled filesystems.
665  */
666 struct journal_head *jbd2_journal_get_descriptor_buffer(journal_t *journal)
667 {
668         struct buffer_head *bh;
669         unsigned long long blocknr;
670         int err;
671
672         err = jbd2_journal_next_log_block(journal, &blocknr);
673
674         if (err)
675                 return NULL;
676
677         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
678         if (!bh)
679                 return NULL;
680         lock_buffer(bh);
681         memset(bh->b_data, 0, journal->j_blocksize);
682         set_buffer_uptodate(bh);
683         unlock_buffer(bh);
684         BUFFER_TRACE(bh, "return this buffer");
685         return jbd2_journal_add_journal_head(bh);
686 }
687
688 struct jbd2_stats_proc_session {
689         journal_t *journal;
690         struct transaction_stats_s *stats;
691         int start;
692         int max;
693 };
694
695 static void *jbd2_seq_info_start(struct seq_file *seq, loff_t *pos)
696 {
697         return *pos ? NULL : SEQ_START_TOKEN;
698 }
699
700 static void *jbd2_seq_info_next(struct seq_file *seq, void *v, loff_t *pos)
701 {
702         return NULL;
703 }
704
705 static int jbd2_seq_info_show(struct seq_file *seq, void *v)
706 {
707         struct jbd2_stats_proc_session *s = seq->private;
708
709         if (v != SEQ_START_TOKEN)
710                 return 0;
711         seq_printf(seq, "%lu transaction, each up to %u blocks\n",
712                         s->stats->ts_tid,
713                         s->journal->j_max_transaction_buffers);
714         if (s->stats->ts_tid == 0)
715                 return 0;
716         seq_printf(seq, "average: \n  %ums waiting for transaction\n",
717             jiffies_to_msecs(s->stats->run.rs_wait / s->stats->ts_tid));
718         seq_printf(seq, "  %ums running transaction\n",
719             jiffies_to_msecs(s->stats->run.rs_running / s->stats->ts_tid));
720         seq_printf(seq, "  %ums transaction was being locked\n",
721             jiffies_to_msecs(s->stats->run.rs_locked / s->stats->ts_tid));
722         seq_printf(seq, "  %ums flushing data (in ordered mode)\n",
723             jiffies_to_msecs(s->stats->run.rs_flushing / s->stats->ts_tid));
724         seq_printf(seq, "  %ums logging transaction\n",
725             jiffies_to_msecs(s->stats->run.rs_logging / s->stats->ts_tid));
726         seq_printf(seq, "  %lluus average transaction commit time\n",
727                    div_u64(s->journal->j_average_commit_time, 1000));
728         seq_printf(seq, "  %lu handles per transaction\n",
729             s->stats->run.rs_handle_count / s->stats->ts_tid);
730         seq_printf(seq, "  %lu blocks per transaction\n",
731             s->stats->run.rs_blocks / s->stats->ts_tid);
732         seq_printf(seq, "  %lu logged blocks per transaction\n",
733             s->stats->run.rs_blocks_logged / s->stats->ts_tid);
734         return 0;
735 }
736
737 static void jbd2_seq_info_stop(struct seq_file *seq, void *v)
738 {
739 }
740
741 static const struct seq_operations jbd2_seq_info_ops = {
742         .start  = jbd2_seq_info_start,
743         .next   = jbd2_seq_info_next,
744         .stop   = jbd2_seq_info_stop,
745         .show   = jbd2_seq_info_show,
746 };
747
748 static int jbd2_seq_info_open(struct inode *inode, struct file *file)
749 {
750         journal_t *journal = PDE(inode)->data;
751         struct jbd2_stats_proc_session *s;
752         int rc, size;
753
754         s = kmalloc(sizeof(*s), GFP_KERNEL);
755         if (s == NULL)
756                 return -ENOMEM;
757         size = sizeof(struct transaction_stats_s);
758         s->stats = kmalloc(size, GFP_KERNEL);
759         if (s->stats == NULL) {
760                 kfree(s);
761                 return -ENOMEM;
762         }
763         spin_lock(&journal->j_history_lock);
764         memcpy(s->stats, &journal->j_stats, size);
765         s->journal = journal;
766         spin_unlock(&journal->j_history_lock);
767
768         rc = seq_open(file, &jbd2_seq_info_ops);
769         if (rc == 0) {
770                 struct seq_file *m = file->private_data;
771                 m->private = s;
772         } else {
773                 kfree(s->stats);
774                 kfree(s);
775         }
776         return rc;
777
778 }
779
780 static int jbd2_seq_info_release(struct inode *inode, struct file *file)
781 {
782         struct seq_file *seq = file->private_data;
783         struct jbd2_stats_proc_session *s = seq->private;
784         kfree(s->stats);
785         kfree(s);
786         return seq_release(inode, file);
787 }
788
789 static const struct file_operations jbd2_seq_info_fops = {
790         .owner          = THIS_MODULE,
791         .open           = jbd2_seq_info_open,
792         .read           = seq_read,
793         .llseek         = seq_lseek,
794         .release        = jbd2_seq_info_release,
795 };
796
797 static struct proc_dir_entry *proc_jbd2_stats;
798
799 static void jbd2_stats_proc_init(journal_t *journal)
800 {
801         journal->j_proc_entry = proc_mkdir(journal->j_devname, proc_jbd2_stats);
802         if (journal->j_proc_entry) {
803                 proc_create_data("info", S_IRUGO, journal->j_proc_entry,
804                                  &jbd2_seq_info_fops, journal);
805         }
806 }
807
808 static void jbd2_stats_proc_exit(journal_t *journal)
809 {
810         remove_proc_entry("info", journal->j_proc_entry);
811         remove_proc_entry(journal->j_devname, proc_jbd2_stats);
812 }
813
814 /*
815  * Management for journal control blocks: functions to create and
816  * destroy journal_t structures, and to initialise and read existing
817  * journal blocks from disk.  */
818
819 /* First: create and setup a journal_t object in memory.  We initialise
820  * very few fields yet: that has to wait until we have created the
821  * journal structures from from scratch, or loaded them from disk. */
822
823 static journal_t * journal_init_common (void)
824 {
825         journal_t *journal;
826         int err;
827
828         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
829         if (!journal)
830                 goto fail;
831
832         init_waitqueue_head(&journal->j_wait_transaction_locked);
833         init_waitqueue_head(&journal->j_wait_logspace);
834         init_waitqueue_head(&journal->j_wait_done_commit);
835         init_waitqueue_head(&journal->j_wait_checkpoint);
836         init_waitqueue_head(&journal->j_wait_commit);
837         init_waitqueue_head(&journal->j_wait_updates);
838         mutex_init(&journal->j_barrier);
839         mutex_init(&journal->j_checkpoint_mutex);
840         spin_lock_init(&journal->j_revoke_lock);
841         spin_lock_init(&journal->j_list_lock);
842         rwlock_init(&journal->j_state_lock);
843
844         journal->j_commit_interval = (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE);
845         journal->j_min_batch_time = 0;
846         journal->j_max_batch_time = 15000; /* 15ms */
847
848         /* The journal is marked for error until we succeed with recovery! */
849         journal->j_flags = JBD2_ABORT;
850
851         /* Set up a default-sized revoke table for the new mount. */
852         err = jbd2_journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
853         if (err) {
854                 kfree(journal);
855                 goto fail;
856         }
857
858         spin_lock_init(&journal->j_history_lock);
859
860         return journal;
861 fail:
862         return NULL;
863 }
864
865 /* jbd2_journal_init_dev and jbd2_journal_init_inode:
866  *
867  * Create a journal structure assigned some fixed set of disk blocks to
868  * the journal.  We don't actually touch those disk blocks yet, but we
869  * need to set up all of the mapping information to tell the journaling
870  * system where the journal blocks are.
871  *
872  */
873
874 /**
875  *  journal_t * jbd2_journal_init_dev() - creates and initialises a journal structure
876  *  @bdev: Block device on which to create the journal
877  *  @fs_dev: Device which hold journalled filesystem for this journal.
878  *  @start: Block nr Start of journal.
879  *  @len:  Length of the journal in blocks.
880  *  @blocksize: blocksize of journalling device
881  *
882  *  Returns: a newly created journal_t *
883  *
884  *  jbd2_journal_init_dev creates a journal which maps a fixed contiguous
885  *  range of blocks on an arbitrary block device.
886  *
887  */
888 journal_t * jbd2_journal_init_dev(struct block_device *bdev,
889                         struct block_device *fs_dev,
890                         unsigned long long start, int len, int blocksize)
891 {
892         journal_t *journal = journal_init_common();
893         struct buffer_head *bh;
894         char *p;
895         int n;
896
897         if (!journal)
898                 return NULL;
899
900         /* journal descriptor can store up to n blocks -bzzz */
901         journal->j_blocksize = blocksize;
902         jbd2_stats_proc_init(journal);
903         n = journal->j_blocksize / sizeof(journal_block_tag_t);
904         journal->j_wbufsize = n;
905         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
906         if (!journal->j_wbuf) {
907                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
908                         __func__);
909                 goto out_err;
910         }
911         journal->j_dev = bdev;
912         journal->j_fs_dev = fs_dev;
913         journal->j_blk_offset = start;
914         journal->j_maxlen = len;
915         bdevname(journal->j_dev, journal->j_devname);
916         p = journal->j_devname;
917         while ((p = strchr(p, '/')))
918                 *p = '!';
919
920         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
921         if (!bh) {
922                 printk(KERN_ERR
923                        "%s: Cannot get buffer for journal superblock\n",
924                        __func__);
925                 goto out_err;
926         }
927         journal->j_sb_buffer = bh;
928         journal->j_superblock = (journal_superblock_t *)bh->b_data;
929
930         return journal;
931 out_err:
932         kfree(journal->j_wbuf);
933         jbd2_stats_proc_exit(journal);
934         kfree(journal);
935         return NULL;
936 }
937
938 /**
939  *  journal_t * jbd2_journal_init_inode () - creates a journal which maps to a inode.
940  *  @inode: An inode to create the journal in
941  *
942  * jbd2_journal_init_inode creates a journal which maps an on-disk inode as
943  * the journal.  The inode must exist already, must support bmap() and
944  * must have all data blocks preallocated.
945  */
946 journal_t * jbd2_journal_init_inode (struct inode *inode)
947 {
948         struct buffer_head *bh;
949         journal_t *journal = journal_init_common();
950         char *p;
951         int err;
952         int n;
953         unsigned long long blocknr;
954
955         if (!journal)
956                 return NULL;
957
958         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
959         journal->j_inode = inode;
960         bdevname(journal->j_dev, journal->j_devname);
961         p = journal->j_devname;
962         while ((p = strchr(p, '/')))
963                 *p = '!';
964         p = journal->j_devname + strlen(journal->j_devname);
965         sprintf(p, "-%lu", journal->j_inode->i_ino);
966         jbd_debug(1,
967                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
968                   journal, inode->i_sb->s_id, inode->i_ino,
969                   (long long) inode->i_size,
970                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
971
972         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
973         journal->j_blocksize = inode->i_sb->s_blocksize;
974         jbd2_stats_proc_init(journal);
975
976         /* journal descriptor can store up to n blocks -bzzz */
977         n = journal->j_blocksize / sizeof(journal_block_tag_t);
978         journal->j_wbufsize = n;
979         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
980         if (!journal->j_wbuf) {
981                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
982                         __func__);
983                 goto out_err;
984         }
985
986         err = jbd2_journal_bmap(journal, 0, &blocknr);
987         /* If that failed, give up */
988         if (err) {
989                 printk(KERN_ERR "%s: Cannnot locate journal superblock\n",
990                        __func__);
991                 goto out_err;
992         }
993
994         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
995         if (!bh) {
996                 printk(KERN_ERR
997                        "%s: Cannot get buffer for journal superblock\n",
998                        __func__);
999                 goto out_err;
1000         }
1001         journal->j_sb_buffer = bh;
1002         journal->j_superblock = (journal_superblock_t *)bh->b_data;
1003
1004         return journal;
1005 out_err:
1006         kfree(journal->j_wbuf);
1007         jbd2_stats_proc_exit(journal);
1008         kfree(journal);
1009         return NULL;
1010 }
1011
1012 /*
1013  * If the journal init or create aborts, we need to mark the journal
1014  * superblock as being NULL to prevent the journal destroy from writing
1015  * back a bogus superblock.
1016  */
1017 static void journal_fail_superblock (journal_t *journal)
1018 {
1019         struct buffer_head *bh = journal->j_sb_buffer;
1020         brelse(bh);
1021         journal->j_sb_buffer = NULL;
1022 }
1023
1024 /*
1025  * Given a journal_t structure, initialise the various fields for
1026  * startup of a new journaling session.  We use this both when creating
1027  * a journal, and after recovering an old journal to reset it for
1028  * subsequent use.
1029  */
1030
1031 static int journal_reset(journal_t *journal)
1032 {
1033         journal_superblock_t *sb = journal->j_superblock;
1034         unsigned long long first, last;
1035
1036         first = be32_to_cpu(sb->s_first);
1037         last = be32_to_cpu(sb->s_maxlen);
1038         if (first + JBD2_MIN_JOURNAL_BLOCKS > last + 1) {
1039                 printk(KERN_ERR "JBD: Journal too short (blocks %llu-%llu).\n",
1040                        first, last);
1041                 journal_fail_superblock(journal);
1042                 return -EINVAL;
1043         }
1044
1045         journal->j_first = first;
1046         journal->j_last = last;
1047
1048         journal->j_head = first;
1049         journal->j_tail = first;
1050         journal->j_free = last - first;
1051
1052         journal->j_tail_sequence = journal->j_transaction_sequence;
1053         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
1054         journal->j_commit_request = journal->j_commit_sequence;
1055
1056         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
1057
1058         /* Add the dynamic fields and write it to disk. */
1059         jbd2_journal_update_superblock(journal, 1);
1060         return jbd2_journal_start_thread(journal);
1061 }
1062
1063 /**
1064  * void jbd2_journal_update_superblock() - Update journal sb on disk.
1065  * @journal: The journal to update.
1066  * @wait: Set to '0' if you don't want to wait for IO completion.
1067  *
1068  * Update a journal's dynamic superblock fields and write it to disk,
1069  * optionally waiting for the IO to complete.
1070  */
1071 void jbd2_journal_update_superblock(journal_t *journal, int wait)
1072 {
1073         journal_superblock_t *sb = journal->j_superblock;
1074         struct buffer_head *bh = journal->j_sb_buffer;
1075
1076         /*
1077          * As a special case, if the on-disk copy is already marked as needing
1078          * no recovery (s_start == 0) and there are no outstanding transactions
1079          * in the filesystem, then we can safely defer the superblock update
1080          * until the next commit by setting JBD2_FLUSHED.  This avoids
1081          * attempting a write to a potential-readonly device.
1082          */
1083         if (sb->s_start == 0 && journal->j_tail_sequence ==
1084                                 journal->j_transaction_sequence) {
1085                 jbd_debug(1,"JBD: Skipping superblock update on recovered sb "
1086                         "(start %ld, seq %d, errno %d)\n",
1087                         journal->j_tail, journal->j_tail_sequence,
1088                         journal->j_errno);
1089                 goto out;
1090         }
1091
1092         if (buffer_write_io_error(bh)) {
1093                 /*
1094                  * Oh, dear.  A previous attempt to write the journal
1095                  * superblock failed.  This could happen because the
1096                  * USB device was yanked out.  Or it could happen to
1097                  * be a transient write error and maybe the block will
1098                  * be remapped.  Nothing we can do but to retry the
1099                  * write and hope for the best.
1100                  */
1101                 printk(KERN_ERR "JBD2: previous I/O error detected "
1102                        "for journal superblock update for %s.\n",
1103                        journal->j_devname);
1104                 clear_buffer_write_io_error(bh);
1105                 set_buffer_uptodate(bh);
1106         }
1107
1108         read_lock(&journal->j_state_lock);
1109         jbd_debug(1,"JBD: updating superblock (start %ld, seq %d, errno %d)\n",
1110                   journal->j_tail, journal->j_tail_sequence, journal->j_errno);
1111
1112         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
1113         sb->s_start    = cpu_to_be32(journal->j_tail);
1114         sb->s_errno    = cpu_to_be32(journal->j_errno);
1115         read_unlock(&journal->j_state_lock);
1116
1117         BUFFER_TRACE(bh, "marking dirty");
1118         mark_buffer_dirty(bh);
1119         if (wait) {
1120                 sync_dirty_buffer(bh);
1121                 if (buffer_write_io_error(bh)) {
1122                         printk(KERN_ERR "JBD2: I/O error detected "
1123                                "when updating journal superblock for %s.\n",
1124                                journal->j_devname);
1125                         clear_buffer_write_io_error(bh);
1126                         set_buffer_uptodate(bh);
1127                 }
1128         } else
1129                 write_dirty_buffer(bh, WRITE);
1130
1131 out:
1132         /* If we have just flushed the log (by marking s_start==0), then
1133          * any future commit will have to be careful to update the
1134          * superblock again to re-record the true start of the log. */
1135
1136         write_lock(&journal->j_state_lock);
1137         if (sb->s_start)
1138                 journal->j_flags &= ~JBD2_FLUSHED;
1139         else
1140                 journal->j_flags |= JBD2_FLUSHED;
1141         write_unlock(&journal->j_state_lock);
1142 }
1143
1144 /*
1145  * Read the superblock for a given journal, performing initial
1146  * validation of the format.
1147  */
1148
1149 static int journal_get_superblock(journal_t *journal)
1150 {
1151         struct buffer_head *bh;
1152         journal_superblock_t *sb;
1153         int err = -EIO;
1154
1155         bh = journal->j_sb_buffer;
1156
1157         J_ASSERT(bh != NULL);
1158         if (!buffer_uptodate(bh)) {
1159                 ll_rw_block(READ, 1, &bh);
1160                 wait_on_buffer(bh);
1161                 if (!buffer_uptodate(bh)) {
1162                         printk (KERN_ERR
1163                                 "JBD: IO error reading journal superblock\n");
1164                         goto out;
1165                 }
1166         }
1167
1168         sb = journal->j_superblock;
1169
1170         err = -EINVAL;
1171
1172         if (sb->s_header.h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER) ||
1173             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1174                 printk(KERN_WARNING "JBD: no valid journal superblock found\n");
1175                 goto out;
1176         }
1177
1178         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1179         case JBD2_SUPERBLOCK_V1:
1180                 journal->j_format_version = 1;
1181                 break;
1182         case JBD2_SUPERBLOCK_V2:
1183                 journal->j_format_version = 2;
1184                 break;
1185         default:
1186                 printk(KERN_WARNING "JBD: unrecognised superblock format ID\n");
1187                 goto out;
1188         }
1189
1190         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1191                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1192         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1193                 printk (KERN_WARNING "JBD: journal file too short\n");
1194                 goto out;
1195         }
1196
1197         return 0;
1198
1199 out:
1200         journal_fail_superblock(journal);
1201         return err;
1202 }
1203
1204 /*
1205  * Load the on-disk journal superblock and read the key fields into the
1206  * journal_t.
1207  */
1208
1209 static int load_superblock(journal_t *journal)
1210 {
1211         int err;
1212         journal_superblock_t *sb;
1213
1214         err = journal_get_superblock(journal);
1215         if (err)
1216                 return err;
1217
1218         sb = journal->j_superblock;
1219
1220         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1221         journal->j_tail = be32_to_cpu(sb->s_start);
1222         journal->j_first = be32_to_cpu(sb->s_first);
1223         journal->j_last = be32_to_cpu(sb->s_maxlen);
1224         journal->j_errno = be32_to_cpu(sb->s_errno);
1225
1226         return 0;
1227 }
1228
1229
1230 /**
1231  * int jbd2_journal_load() - Read journal from disk.
1232  * @journal: Journal to act on.
1233  *
1234  * Given a journal_t structure which tells us which disk blocks contain
1235  * a journal, read the journal from disk to initialise the in-memory
1236  * structures.
1237  */
1238 int jbd2_journal_load(journal_t *journal)
1239 {
1240         int err;
1241         journal_superblock_t *sb;
1242
1243         err = load_superblock(journal);
1244         if (err)
1245                 return err;
1246
1247         sb = journal->j_superblock;
1248         /* If this is a V2 superblock, then we have to check the
1249          * features flags on it. */
1250
1251         if (journal->j_format_version >= 2) {
1252                 if ((sb->s_feature_ro_compat &
1253                      ~cpu_to_be32(JBD2_KNOWN_ROCOMPAT_FEATURES)) ||
1254                     (sb->s_feature_incompat &
1255                      ~cpu_to_be32(JBD2_KNOWN_INCOMPAT_FEATURES))) {
1256                         printk (KERN_WARNING
1257                                 "JBD: Unrecognised features on journal\n");
1258                         return -EINVAL;
1259                 }
1260         }
1261
1262         /*
1263          * Create a slab for this blocksize
1264          */
1265         err = jbd2_journal_create_slab(be32_to_cpu(sb->s_blocksize));
1266         if (err)
1267                 return err;
1268
1269         /* Let the recovery code check whether it needs to recover any
1270          * data from the journal. */
1271         if (jbd2_journal_recover(journal))
1272                 goto recovery_error;
1273
1274         if (journal->j_failed_commit) {
1275                 printk(KERN_ERR "JBD2: journal transaction %u on %s "
1276                        "is corrupt.\n", journal->j_failed_commit,
1277                        journal->j_devname);
1278                 return -EIO;
1279         }
1280
1281         /* OK, we've finished with the dynamic journal bits:
1282          * reinitialise the dynamic contents of the superblock in memory
1283          * and reset them on disk. */
1284         if (journal_reset(journal))
1285                 goto recovery_error;
1286
1287         journal->j_flags &= ~JBD2_ABORT;
1288         journal->j_flags |= JBD2_LOADED;
1289         return 0;
1290
1291 recovery_error:
1292         printk (KERN_WARNING "JBD: recovery failed\n");
1293         return -EIO;
1294 }
1295
1296 /**
1297  * void jbd2_journal_destroy() - Release a journal_t structure.
1298  * @journal: Journal to act on.
1299  *
1300  * Release a journal_t structure once it is no longer in use by the
1301  * journaled object.
1302  * Return <0 if we couldn't clean up the journal.
1303  */
1304 int jbd2_journal_destroy(journal_t *journal)
1305 {
1306         int err = 0;
1307
1308         /* Wait for the commit thread to wake up and die. */
1309         journal_kill_thread(journal);
1310
1311         /* Force a final log commit */
1312         if (journal->j_running_transaction)
1313                 jbd2_journal_commit_transaction(journal);
1314
1315         /* Force any old transactions to disk */
1316
1317         /* Totally anal locking here... */
1318         spin_lock(&journal->j_list_lock);
1319         while (journal->j_checkpoint_transactions != NULL) {
1320                 spin_unlock(&journal->j_list_lock);
1321                 mutex_lock(&journal->j_checkpoint_mutex);
1322                 jbd2_log_do_checkpoint(journal);
1323                 mutex_unlock(&journal->j_checkpoint_mutex);
1324                 spin_lock(&journal->j_list_lock);
1325         }
1326
1327         J_ASSERT(journal->j_running_transaction == NULL);
1328         J_ASSERT(journal->j_committing_transaction == NULL);
1329         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1330         spin_unlock(&journal->j_list_lock);
1331
1332         if (journal->j_sb_buffer) {
1333                 if (!is_journal_aborted(journal)) {
1334                         /* We can now mark the journal as empty. */
1335                         journal->j_tail = 0;
1336                         journal->j_tail_sequence =
1337                                 ++journal->j_transaction_sequence;
1338                         jbd2_journal_update_superblock(journal, 1);
1339                 } else {
1340                         err = -EIO;
1341                 }
1342                 brelse(journal->j_sb_buffer);
1343         }
1344
1345         if (journal->j_proc_entry)
1346                 jbd2_stats_proc_exit(journal);
1347         if (journal->j_inode)
1348                 iput(journal->j_inode);
1349         if (journal->j_revoke)
1350                 jbd2_journal_destroy_revoke(journal);
1351         kfree(journal->j_wbuf);
1352         kfree(journal);
1353
1354         return err;
1355 }
1356
1357
1358 /**
1359  *int jbd2_journal_check_used_features () - Check if features specified are used.
1360  * @journal: Journal to check.
1361  * @compat: bitmask of compatible features
1362  * @ro: bitmask of features that force read-only mount
1363  * @incompat: bitmask of incompatible features
1364  *
1365  * Check whether the journal uses all of a given set of
1366  * features.  Return true (non-zero) if it does.
1367  **/
1368
1369 int jbd2_journal_check_used_features (journal_t *journal, unsigned long compat,
1370                                  unsigned long ro, unsigned long incompat)
1371 {
1372         journal_superblock_t *sb;
1373
1374         if (!compat && !ro && !incompat)
1375                 return 1;
1376         /* Load journal superblock if it is not loaded yet. */
1377         if (journal->j_format_version == 0 &&
1378             journal_get_superblock(journal) != 0)
1379                 return 0;
1380         if (journal->j_format_version == 1)
1381                 return 0;
1382
1383         sb = journal->j_superblock;
1384
1385         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1386             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1387             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1388                 return 1;
1389
1390         return 0;
1391 }
1392
1393 /**
1394  * int jbd2_journal_check_available_features() - Check feature set in journalling layer
1395  * @journal: Journal to check.
1396  * @compat: bitmask of compatible features
1397  * @ro: bitmask of features that force read-only mount
1398  * @incompat: bitmask of incompatible features
1399  *
1400  * Check whether the journaling code supports the use of
1401  * all of a given set of features on this journal.  Return true
1402  * (non-zero) if it can. */
1403
1404 int jbd2_journal_check_available_features (journal_t *journal, unsigned long compat,
1405                                       unsigned long ro, unsigned long incompat)
1406 {
1407         if (!compat && !ro && !incompat)
1408                 return 1;
1409
1410         /* We can support any known requested features iff the
1411          * superblock is in version 2.  Otherwise we fail to support any
1412          * extended sb features. */
1413
1414         if (journal->j_format_version != 2)
1415                 return 0;
1416
1417         if ((compat   & JBD2_KNOWN_COMPAT_FEATURES) == compat &&
1418             (ro       & JBD2_KNOWN_ROCOMPAT_FEATURES) == ro &&
1419             (incompat & JBD2_KNOWN_INCOMPAT_FEATURES) == incompat)
1420                 return 1;
1421
1422         return 0;
1423 }
1424
1425 /**
1426  * int jbd2_journal_set_features () - Mark a given journal feature in the superblock
1427  * @journal: Journal to act on.
1428  * @compat: bitmask of compatible features
1429  * @ro: bitmask of features that force read-only mount
1430  * @incompat: bitmask of incompatible features
1431  *
1432  * Mark a given journal feature as present on the
1433  * superblock.  Returns true if the requested features could be set.
1434  *
1435  */
1436
1437 int jbd2_journal_set_features (journal_t *journal, unsigned long compat,
1438                           unsigned long ro, unsigned long incompat)
1439 {
1440         journal_superblock_t *sb;
1441
1442         if (jbd2_journal_check_used_features(journal, compat, ro, incompat))
1443                 return 1;
1444
1445         if (!jbd2_journal_check_available_features(journal, compat, ro, incompat))
1446                 return 0;
1447
1448         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1449                   compat, ro, incompat);
1450
1451         sb = journal->j_superblock;
1452
1453         sb->s_feature_compat    |= cpu_to_be32(compat);
1454         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1455         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1456
1457         return 1;
1458 }
1459
1460 /*
1461  * jbd2_journal_clear_features () - Clear a given journal feature in the
1462  *                                  superblock
1463  * @journal: Journal to act on.
1464  * @compat: bitmask of compatible features
1465  * @ro: bitmask of features that force read-only mount
1466  * @incompat: bitmask of incompatible features
1467  *
1468  * Clear a given journal feature as present on the
1469  * superblock.
1470  */
1471 void jbd2_journal_clear_features(journal_t *journal, unsigned long compat,
1472                                 unsigned long ro, unsigned long incompat)
1473 {
1474         journal_superblock_t *sb;
1475
1476         jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
1477                   compat, ro, incompat);
1478
1479         sb = journal->j_superblock;
1480
1481         sb->s_feature_compat    &= ~cpu_to_be32(compat);
1482         sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
1483         sb->s_feature_incompat  &= ~cpu_to_be32(incompat);
1484 }
1485 EXPORT_SYMBOL(jbd2_journal_clear_features);
1486
1487 /**
1488  * int jbd2_journal_update_format () - Update on-disk journal structure.
1489  * @journal: Journal to act on.
1490  *
1491  * Given an initialised but unloaded journal struct, poke about in the
1492  * on-disk structure to update it to the most recent supported version.
1493  */
1494 int jbd2_journal_update_format (journal_t *journal)
1495 {
1496         journal_superblock_t *sb;
1497         int err;
1498
1499         err = journal_get_superblock(journal);
1500         if (err)
1501                 return err;
1502
1503         sb = journal->j_superblock;
1504
1505         switch (be32_to_cpu(sb->s_header.h_blocktype)) {
1506         case JBD2_SUPERBLOCK_V2:
1507                 return 0;
1508         case JBD2_SUPERBLOCK_V1:
1509                 return journal_convert_superblock_v1(journal, sb);
1510         default:
1511                 break;
1512         }
1513         return -EINVAL;
1514 }
1515
1516 static int journal_convert_superblock_v1(journal_t *journal,
1517                                          journal_superblock_t *sb)
1518 {
1519         int offset, blocksize;
1520         struct buffer_head *bh;
1521
1522         printk(KERN_WARNING
1523                 "JBD: Converting superblock from version 1 to 2.\n");
1524
1525         /* Pre-initialise new fields to zero */
1526         offset = ((char *) &(sb->s_feature_compat)) - ((char *) sb);
1527         blocksize = be32_to_cpu(sb->s_blocksize);
1528         memset(&sb->s_feature_compat, 0, blocksize-offset);
1529
1530         sb->s_nr_users = cpu_to_be32(1);
1531         sb->s_header.h_blocktype = cpu_to_be32(JBD2_SUPERBLOCK_V2);
1532         journal->j_format_version = 2;
1533
1534         bh = journal->j_sb_buffer;
1535         BUFFER_TRACE(bh, "marking dirty");
1536         mark_buffer_dirty(bh);
1537         sync_dirty_buffer(bh);
1538         return 0;
1539 }
1540
1541
1542 /**
1543  * int jbd2_journal_flush () - Flush journal
1544  * @journal: Journal to act on.
1545  *
1546  * Flush all data for a given journal to disk and empty the journal.
1547  * Filesystems can use this when remounting readonly to ensure that
1548  * recovery does not need to happen on remount.
1549  */
1550
1551 int jbd2_journal_flush(journal_t *journal)
1552 {
1553         int err = 0;
1554         transaction_t *transaction = NULL;
1555         unsigned long old_tail;
1556
1557         write_lock(&journal->j_state_lock);
1558
1559         /* Force everything buffered to the log... */
1560         if (journal->j_running_transaction) {
1561                 transaction = journal->j_running_transaction;
1562                 __jbd2_log_start_commit(journal, transaction->t_tid);
1563         } else if (journal->j_committing_transaction)
1564                 transaction = journal->j_committing_transaction;
1565
1566         /* Wait for the log commit to complete... */
1567         if (transaction) {
1568                 tid_t tid = transaction->t_tid;
1569
1570                 write_unlock(&journal->j_state_lock);
1571                 jbd2_log_wait_commit(journal, tid);
1572         } else {
1573                 write_unlock(&journal->j_state_lock);
1574         }
1575
1576         /* ...and flush everything in the log out to disk. */
1577         spin_lock(&journal->j_list_lock);
1578         while (!err && journal->j_checkpoint_transactions != NULL) {
1579                 spin_unlock(&journal->j_list_lock);
1580                 mutex_lock(&journal->j_checkpoint_mutex);
1581                 err = jbd2_log_do_checkpoint(journal);
1582                 mutex_unlock(&journal->j_checkpoint_mutex);
1583                 spin_lock(&journal->j_list_lock);
1584         }
1585         spin_unlock(&journal->j_list_lock);
1586
1587         if (is_journal_aborted(journal))
1588                 return -EIO;
1589
1590         jbd2_cleanup_journal_tail(journal);
1591
1592         /* Finally, mark the journal as really needing no recovery.
1593          * This sets s_start==0 in the underlying superblock, which is
1594          * the magic code for a fully-recovered superblock.  Any future
1595          * commits of data to the journal will restore the current
1596          * s_start value. */
1597         write_lock(&journal->j_state_lock);
1598         old_tail = journal->j_tail;
1599         journal->j_tail = 0;
1600         write_unlock(&journal->j_state_lock);
1601         jbd2_journal_update_superblock(journal, 1);
1602         write_lock(&journal->j_state_lock);
1603         journal->j_tail = old_tail;
1604
1605         J_ASSERT(!journal->j_running_transaction);
1606         J_ASSERT(!journal->j_committing_transaction);
1607         J_ASSERT(!journal->j_checkpoint_transactions);
1608         J_ASSERT(journal->j_head == journal->j_tail);
1609         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1610         write_unlock(&journal->j_state_lock);
1611         return 0;
1612 }
1613
1614 /**
1615  * int jbd2_journal_wipe() - Wipe journal contents
1616  * @journal: Journal to act on.
1617  * @write: flag (see below)
1618  *
1619  * Wipe out all of the contents of a journal, safely.  This will produce
1620  * a warning if the journal contains any valid recovery information.
1621  * Must be called between journal_init_*() and jbd2_journal_load().
1622  *
1623  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1624  * we merely suppress recovery.
1625  */
1626
1627 int jbd2_journal_wipe(journal_t *journal, int write)
1628 {
1629         int err = 0;
1630
1631         J_ASSERT (!(journal->j_flags & JBD2_LOADED));
1632
1633         err = load_superblock(journal);
1634         if (err)
1635                 return err;
1636
1637         if (!journal->j_tail)
1638                 goto no_recovery;
1639
1640         printk (KERN_WARNING "JBD: %s recovery information on journal\n",
1641                 write ? "Clearing" : "Ignoring");
1642
1643         err = jbd2_journal_skip_recovery(journal);
1644         if (write)
1645                 jbd2_journal_update_superblock(journal, 1);
1646
1647  no_recovery:
1648         return err;
1649 }
1650
1651 /*
1652  * Journal abort has very specific semantics, which we describe
1653  * for journal abort.
1654  *
1655  * Two internal functions, which provide abort to the jbd layer
1656  * itself are here.
1657  */
1658
1659 /*
1660  * Quick version for internal journal use (doesn't lock the journal).
1661  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1662  * and don't attempt to make any other journal updates.
1663  */
1664 void __jbd2_journal_abort_hard(journal_t *journal)
1665 {
1666         transaction_t *transaction;
1667
1668         if (journal->j_flags & JBD2_ABORT)
1669                 return;
1670
1671         printk(KERN_ERR "Aborting journal on device %s.\n",
1672                journal->j_devname);
1673
1674         write_lock(&journal->j_state_lock);
1675         journal->j_flags |= JBD2_ABORT;
1676         transaction = journal->j_running_transaction;
1677         if (transaction)
1678                 __jbd2_log_start_commit(journal, transaction->t_tid);
1679         write_unlock(&journal->j_state_lock);
1680 }
1681
1682 /* Soft abort: record the abort error status in the journal superblock,
1683  * but don't do any other IO. */
1684 static void __journal_abort_soft (journal_t *journal, int errno)
1685 {
1686         if (journal->j_flags & JBD2_ABORT)
1687                 return;
1688
1689         if (!journal->j_errno)
1690                 journal->j_errno = errno;
1691
1692         __jbd2_journal_abort_hard(journal);
1693
1694         if (errno)
1695                 jbd2_journal_update_superblock(journal, 1);
1696 }
1697
1698 /**
1699  * void jbd2_journal_abort () - Shutdown the journal immediately.
1700  * @journal: the journal to shutdown.
1701  * @errno:   an error number to record in the journal indicating
1702  *           the reason for the shutdown.
1703  *
1704  * Perform a complete, immediate shutdown of the ENTIRE
1705  * journal (not of a single transaction).  This operation cannot be
1706  * undone without closing and reopening the journal.
1707  *
1708  * The jbd2_journal_abort function is intended to support higher level error
1709  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1710  * mode.
1711  *
1712  * Journal abort has very specific semantics.  Any existing dirty,
1713  * unjournaled buffers in the main filesystem will still be written to
1714  * disk by bdflush, but the journaling mechanism will be suspended
1715  * immediately and no further transaction commits will be honoured.
1716  *
1717  * Any dirty, journaled buffers will be written back to disk without
1718  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1719  * filesystem, but we _do_ attempt to leave as much data as possible
1720  * behind for fsck to use for cleanup.
1721  *
1722  * Any attempt to get a new transaction handle on a journal which is in
1723  * ABORT state will just result in an -EROFS error return.  A
1724  * jbd2_journal_stop on an existing handle will return -EIO if we have
1725  * entered abort state during the update.
1726  *
1727  * Recursive transactions are not disturbed by journal abort until the
1728  * final jbd2_journal_stop, which will receive the -EIO error.
1729  *
1730  * Finally, the jbd2_journal_abort call allows the caller to supply an errno
1731  * which will be recorded (if possible) in the journal superblock.  This
1732  * allows a client to record failure conditions in the middle of a
1733  * transaction without having to complete the transaction to record the
1734  * failure to disk.  ext3_error, for example, now uses this
1735  * functionality.
1736  *
1737  * Errors which originate from within the journaling layer will NOT
1738  * supply an errno; a null errno implies that absolutely no further
1739  * writes are done to the journal (unless there are any already in
1740  * progress).
1741  *
1742  */
1743
1744 void jbd2_journal_abort(journal_t *journal, int errno)
1745 {
1746         __journal_abort_soft(journal, errno);
1747 }
1748
1749 /**
1750  * int jbd2_journal_errno () - returns the journal's error state.
1751  * @journal: journal to examine.
1752  *
1753  * This is the errno number set with jbd2_journal_abort(), the last
1754  * time the journal was mounted - if the journal was stopped
1755  * without calling abort this will be 0.
1756  *
1757  * If the journal has been aborted on this mount time -EROFS will
1758  * be returned.
1759  */
1760 int jbd2_journal_errno(journal_t *journal)
1761 {
1762         int err;
1763
1764         read_lock(&journal->j_state_lock);
1765         if (journal->j_flags & JBD2_ABORT)
1766                 err = -EROFS;
1767         else
1768                 err = journal->j_errno;
1769         read_unlock(&journal->j_state_lock);
1770         return err;
1771 }
1772
1773 /**
1774  * int jbd2_journal_clear_err () - clears the journal's error state
1775  * @journal: journal to act on.
1776  *
1777  * An error must be cleared or acked to take a FS out of readonly
1778  * mode.
1779  */
1780 int jbd2_journal_clear_err(journal_t *journal)
1781 {
1782         int err = 0;
1783
1784         write_lock(&journal->j_state_lock);
1785         if (journal->j_flags & JBD2_ABORT)
1786                 err = -EROFS;
1787         else
1788                 journal->j_errno = 0;
1789         write_unlock(&journal->j_state_lock);
1790         return err;
1791 }
1792
1793 /**
1794  * void jbd2_journal_ack_err() - Ack journal err.
1795  * @journal: journal to act on.
1796  *
1797  * An error must be cleared or acked to take a FS out of readonly
1798  * mode.
1799  */
1800 void jbd2_journal_ack_err(journal_t *journal)
1801 {
1802         write_lock(&journal->j_state_lock);
1803         if (journal->j_errno)
1804                 journal->j_flags |= JBD2_ACK_ERR;
1805         write_unlock(&journal->j_state_lock);
1806 }
1807
1808 int jbd2_journal_blocks_per_page(struct inode *inode)
1809 {
1810         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1811 }
1812
1813 /*
1814  * helper functions to deal with 32 or 64bit block numbers.
1815  */
1816 size_t journal_tag_bytes(journal_t *journal)
1817 {
1818         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
1819                 return JBD2_TAG_SIZE64;
1820         else
1821                 return JBD2_TAG_SIZE32;
1822 }
1823
1824 /*
1825  * JBD memory management
1826  *
1827  * These functions are used to allocate block-sized chunks of memory
1828  * used for making copies of buffer_head data.  Very often it will be
1829  * page-sized chunks of data, but sometimes it will be in
1830  * sub-page-size chunks.  (For example, 16k pages on Power systems
1831  * with a 4k block file system.)  For blocks smaller than a page, we
1832  * use a SLAB allocator.  There are slab caches for each block size,
1833  * which are allocated at mount time, if necessary, and we only free
1834  * (all of) the slab caches when/if the jbd2 module is unloaded.  For
1835  * this reason we don't need to a mutex to protect access to
1836  * jbd2_slab[] allocating or releasing memory; only in
1837  * jbd2_journal_create_slab().
1838  */
1839 #define JBD2_MAX_SLABS 8
1840 static struct kmem_cache *jbd2_slab[JBD2_MAX_SLABS];
1841 static DECLARE_MUTEX(jbd2_slab_create_sem);
1842
1843 static const char *jbd2_slab_names[JBD2_MAX_SLABS] = {
1844         "jbd2_1k", "jbd2_2k", "jbd2_4k", "jbd2_8k",
1845         "jbd2_16k", "jbd2_32k", "jbd2_64k", "jbd2_128k"
1846 };
1847
1848
1849 static void jbd2_journal_destroy_slabs(void)
1850 {
1851         int i;
1852
1853         for (i = 0; i < JBD2_MAX_SLABS; i++) {
1854                 if (jbd2_slab[i])
1855                         kmem_cache_destroy(jbd2_slab[i]);
1856                 jbd2_slab[i] = NULL;
1857         }
1858 }
1859
1860 static int jbd2_journal_create_slab(size_t size)
1861 {
1862         int i = order_base_2(size) - 10;
1863         size_t slab_size;
1864
1865         if (size == PAGE_SIZE)
1866                 return 0;
1867
1868         if (i >= JBD2_MAX_SLABS)
1869                 return -EINVAL;
1870
1871         if (unlikely(i < 0))
1872                 i = 0;
1873         down(&jbd2_slab_create_sem);
1874         if (jbd2_slab[i]) {
1875                 up(&jbd2_slab_create_sem);
1876                 return 0;       /* Already created */
1877         }
1878
1879         slab_size = 1 << (i+10);
1880         jbd2_slab[i] = kmem_cache_create(jbd2_slab_names[i], slab_size,
1881                                          slab_size, 0, NULL);
1882         up(&jbd2_slab_create_sem);
1883         if (!jbd2_slab[i]) {
1884                 printk(KERN_EMERG "JBD2: no memory for jbd2_slab cache\n");
1885                 return -ENOMEM;
1886         }
1887         return 0;
1888 }
1889
1890 static struct kmem_cache *get_slab(size_t size)
1891 {
1892         int i = order_base_2(size) - 10;
1893
1894         BUG_ON(i >= JBD2_MAX_SLABS);
1895         if (unlikely(i < 0))
1896                 i = 0;
1897         BUG_ON(jbd2_slab[i] == NULL);
1898         return jbd2_slab[i];
1899 }
1900
1901 void *jbd2_alloc(size_t size, gfp_t flags)
1902 {
1903         void *ptr;
1904
1905         BUG_ON(size & (size-1)); /* Must be a power of 2 */
1906
1907         flags |= __GFP_REPEAT;
1908         if (size == PAGE_SIZE)
1909                 ptr = (void *)__get_free_pages(flags, 0);
1910         else if (size > PAGE_SIZE) {
1911                 int order = get_order(size);
1912
1913                 if (order < 3)
1914                         ptr = (void *)__get_free_pages(flags, order);
1915                 else
1916                         ptr = vmalloc(size);
1917         } else
1918                 ptr = kmem_cache_alloc(get_slab(size), flags);
1919
1920         /* Check alignment; SLUB has gotten this wrong in the past,
1921          * and this can lead to user data corruption! */
1922         BUG_ON(((unsigned long) ptr) & (size-1));
1923
1924         return ptr;
1925 }
1926
1927 void jbd2_free(void *ptr, size_t size)
1928 {
1929         if (size == PAGE_SIZE) {
1930                 free_pages((unsigned long)ptr, 0);
1931                 return;
1932         }
1933         if (size > PAGE_SIZE) {
1934                 int order = get_order(size);
1935
1936                 if (order < 3)
1937                         free_pages((unsigned long)ptr, order);
1938                 else
1939                         vfree(ptr);
1940                 return;
1941         }
1942         kmem_cache_free(get_slab(size), ptr);
1943 };
1944
1945 /*
1946  * Journal_head storage management
1947  */
1948 static struct kmem_cache *jbd2_journal_head_cache;
1949 #ifdef CONFIG_JBD2_DEBUG
1950 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
1951 #endif
1952
1953 static int journal_init_jbd2_journal_head_cache(void)
1954 {
1955         int retval;
1956
1957         J_ASSERT(jbd2_journal_head_cache == NULL);
1958         jbd2_journal_head_cache = kmem_cache_create("jbd2_journal_head",
1959                                 sizeof(struct journal_head),
1960                                 0,              /* offset */
1961                                 SLAB_TEMPORARY, /* flags */
1962                                 NULL);          /* ctor */
1963         retval = 0;
1964         if (!jbd2_journal_head_cache) {
1965                 retval = -ENOMEM;
1966                 printk(KERN_EMERG "JBD: no memory for journal_head cache\n");
1967         }
1968         return retval;
1969 }
1970
1971 static void jbd2_journal_destroy_jbd2_journal_head_cache(void)
1972 {
1973         if (jbd2_journal_head_cache) {
1974                 kmem_cache_destroy(jbd2_journal_head_cache);
1975                 jbd2_journal_head_cache = NULL;
1976         }
1977 }
1978
1979 /*
1980  * journal_head splicing and dicing
1981  */
1982 static struct journal_head *journal_alloc_journal_head(void)
1983 {
1984         struct journal_head *ret;
1985         static unsigned long last_warning;
1986
1987 #ifdef CONFIG_JBD2_DEBUG
1988         atomic_inc(&nr_journal_heads);
1989 #endif
1990         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
1991         if (!ret) {
1992                 jbd_debug(1, "out of memory for journal_head\n");
1993                 if (time_after(jiffies, last_warning + 5*HZ)) {
1994                         printk(KERN_NOTICE "ENOMEM in %s, retrying.\n",
1995                                __func__);
1996                         last_warning = jiffies;
1997                 }
1998                 while (!ret) {
1999                         yield();
2000                         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
2001                 }
2002         }
2003         return ret;
2004 }
2005
2006 static void journal_free_journal_head(struct journal_head *jh)
2007 {
2008 #ifdef CONFIG_JBD2_DEBUG
2009         atomic_dec(&nr_journal_heads);
2010         memset(jh, JBD2_POISON_FREE, sizeof(*jh));
2011 #endif
2012         kmem_cache_free(jbd2_journal_head_cache, jh);
2013 }
2014
2015 /*
2016  * A journal_head is attached to a buffer_head whenever JBD has an
2017  * interest in the buffer.
2018  *
2019  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
2020  * is set.  This bit is tested in core kernel code where we need to take
2021  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
2022  * there.
2023  *
2024  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
2025  *
2026  * When a buffer has its BH_JBD bit set it is immune from being released by
2027  * core kernel code, mainly via ->b_count.
2028  *
2029  * A journal_head may be detached from its buffer_head when the journal_head's
2030  * b_transaction, b_cp_transaction and b_next_transaction pointers are NULL.
2031  * Various places in JBD call jbd2_journal_remove_journal_head() to indicate that the
2032  * journal_head can be dropped if needed.
2033  *
2034  * Various places in the kernel want to attach a journal_head to a buffer_head
2035  * _before_ attaching the journal_head to a transaction.  To protect the
2036  * journal_head in this situation, jbd2_journal_add_journal_head elevates the
2037  * journal_head's b_jcount refcount by one.  The caller must call
2038  * jbd2_journal_put_journal_head() to undo this.
2039  *
2040  * So the typical usage would be:
2041  *
2042  *      (Attach a journal_head if needed.  Increments b_jcount)
2043  *      struct journal_head *jh = jbd2_journal_add_journal_head(bh);
2044  *      ...
2045  *      jh->b_transaction = xxx;
2046  *      jbd2_journal_put_journal_head(jh);
2047  *
2048  * Now, the journal_head's b_jcount is zero, but it is safe from being released
2049  * because it has a non-zero b_transaction.
2050  */
2051
2052 /*
2053  * Give a buffer_head a journal_head.
2054  *
2055  * Doesn't need the journal lock.
2056  * May sleep.
2057  */
2058 struct journal_head *jbd2_journal_add_journal_head(struct buffer_head *bh)
2059 {
2060         struct journal_head *jh;
2061         struct journal_head *new_jh = NULL;
2062
2063 repeat:
2064         if (!buffer_jbd(bh)) {
2065                 new_jh = journal_alloc_journal_head();
2066                 memset(new_jh, 0, sizeof(*new_jh));
2067         }
2068
2069         jbd_lock_bh_journal_head(bh);
2070         if (buffer_jbd(bh)) {
2071                 jh = bh2jh(bh);
2072         } else {
2073                 J_ASSERT_BH(bh,
2074                         (atomic_read(&bh->b_count) > 0) ||
2075                         (bh->b_page && bh->b_page->mapping));
2076
2077                 if (!new_jh) {
2078                         jbd_unlock_bh_journal_head(bh);
2079                         goto repeat;
2080                 }
2081
2082                 jh = new_jh;
2083                 new_jh = NULL;          /* We consumed it */
2084                 set_buffer_jbd(bh);
2085                 bh->b_private = jh;
2086                 jh->b_bh = bh;
2087                 get_bh(bh);
2088                 BUFFER_TRACE(bh, "added journal_head");
2089         }
2090         jh->b_jcount++;
2091         jbd_unlock_bh_journal_head(bh);
2092         if (new_jh)
2093                 journal_free_journal_head(new_jh);
2094         return bh->b_private;
2095 }
2096
2097 /*
2098  * Grab a ref against this buffer_head's journal_head.  If it ended up not
2099  * having a journal_head, return NULL
2100  */
2101 struct journal_head *jbd2_journal_grab_journal_head(struct buffer_head *bh)
2102 {
2103         struct journal_head *jh = NULL;
2104
2105         jbd_lock_bh_journal_head(bh);
2106         if (buffer_jbd(bh)) {
2107                 jh = bh2jh(bh);
2108                 jh->b_jcount++;
2109         }
2110         jbd_unlock_bh_journal_head(bh);
2111         return jh;
2112 }
2113
2114 static void __journal_remove_journal_head(struct buffer_head *bh)
2115 {
2116         struct journal_head *jh = bh2jh(bh);
2117
2118         J_ASSERT_JH(jh, jh->b_jcount >= 0);
2119
2120         get_bh(bh);
2121         if (jh->b_jcount == 0) {
2122                 if (jh->b_transaction == NULL &&
2123                                 jh->b_next_transaction == NULL &&
2124                                 jh->b_cp_transaction == NULL) {
2125                         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
2126                         J_ASSERT_BH(bh, buffer_jbd(bh));
2127                         J_ASSERT_BH(bh, jh2bh(jh) == bh);
2128                         BUFFER_TRACE(bh, "remove journal_head");
2129                         if (jh->b_frozen_data) {
2130                                 printk(KERN_WARNING "%s: freeing "
2131                                                 "b_frozen_data\n",
2132                                                 __func__);
2133                                 jbd2_free(jh->b_frozen_data, bh->b_size);
2134                         }
2135                         if (jh->b_committed_data) {
2136                                 printk(KERN_WARNING "%s: freeing "
2137                                                 "b_committed_data\n",
2138                                                 __func__);
2139                                 jbd2_free(jh->b_committed_data, bh->b_size);
2140                         }
2141                         bh->b_private = NULL;
2142                         jh->b_bh = NULL;        /* debug, really */
2143                         clear_buffer_jbd(bh);
2144                         __brelse(bh);
2145                         journal_free_journal_head(jh);
2146                 } else {
2147                         BUFFER_TRACE(bh, "journal_head was locked");
2148                 }
2149         }
2150 }
2151
2152 /*
2153  * jbd2_journal_remove_journal_head(): if the buffer isn't attached to a transaction
2154  * and has a zero b_jcount then remove and release its journal_head.   If we did
2155  * see that the buffer is not used by any transaction we also "logically"
2156  * decrement ->b_count.
2157  *
2158  * We in fact take an additional increment on ->b_count as a convenience,
2159  * because the caller usually wants to do additional things with the bh
2160  * after calling here.
2161  * The caller of jbd2_journal_remove_journal_head() *must* run __brelse(bh) at some
2162  * time.  Once the caller has run __brelse(), the buffer is eligible for
2163  * reaping by try_to_free_buffers().
2164  */
2165 void jbd2_journal_remove_journal_head(struct buffer_head *bh)
2166 {
2167         jbd_lock_bh_journal_head(bh);
2168         __journal_remove_journal_head(bh);
2169         jbd_unlock_bh_journal_head(bh);
2170 }
2171
2172 /*
2173  * Drop a reference on the passed journal_head.  If it fell to zero then try to
2174  * release the journal_head from the buffer_head.
2175  */
2176 void jbd2_journal_put_journal_head(struct journal_head *jh)
2177 {
2178         struct buffer_head *bh = jh2bh(jh);
2179
2180         jbd_lock_bh_journal_head(bh);
2181         J_ASSERT_JH(jh, jh->b_jcount > 0);
2182         --jh->b_jcount;
2183         if (!jh->b_jcount && !jh->b_transaction) {
2184                 __journal_remove_journal_head(bh);
2185                 __brelse(bh);
2186         }
2187         jbd_unlock_bh_journal_head(bh);
2188 }
2189
2190 /*
2191  * Initialize jbd inode head
2192  */
2193 void jbd2_journal_init_jbd_inode(struct jbd2_inode *jinode, struct inode *inode)
2194 {
2195         jinode->i_transaction = NULL;
2196         jinode->i_next_transaction = NULL;
2197         jinode->i_vfs_inode = inode;
2198         jinode->i_flags = 0;
2199         INIT_LIST_HEAD(&jinode->i_list);
2200 }
2201
2202 /*
2203  * Function to be called before we start removing inode from memory (i.e.,
2204  * clear_inode() is a fine place to be called from). It removes inode from
2205  * transaction's lists.
2206  */
2207 void jbd2_journal_release_jbd_inode(journal_t *journal,
2208                                     struct jbd2_inode *jinode)
2209 {
2210         if (!journal)
2211                 return;
2212 restart:
2213         spin_lock(&journal->j_list_lock);
2214         /* Is commit writing out inode - we have to wait */
2215         if (test_bit(__JI_COMMIT_RUNNING, &jinode->i_flags)) {
2216                 wait_queue_head_t *wq;
2217                 DEFINE_WAIT_BIT(wait, &jinode->i_flags, __JI_COMMIT_RUNNING);
2218                 wq = bit_waitqueue(&jinode->i_flags, __JI_COMMIT_RUNNING);
2219                 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2220                 spin_unlock(&journal->j_list_lock);
2221                 schedule();
2222                 finish_wait(wq, &wait.wait);
2223                 goto restart;
2224         }
2225
2226         if (jinode->i_transaction) {
2227                 list_del(&jinode->i_list);
2228                 jinode->i_transaction = NULL;
2229         }
2230         spin_unlock(&journal->j_list_lock);
2231 }
2232
2233 /*
2234  * debugfs tunables
2235  */
2236 #ifdef CONFIG_JBD2_DEBUG
2237 u8 jbd2_journal_enable_debug __read_mostly;
2238 EXPORT_SYMBOL(jbd2_journal_enable_debug);
2239
2240 #define JBD2_DEBUG_NAME "jbd2-debug"
2241
2242 static struct dentry *jbd2_debugfs_dir;
2243 static struct dentry *jbd2_debug;
2244
2245 static void __init jbd2_create_debugfs_entry(void)
2246 {
2247         jbd2_debugfs_dir = debugfs_create_dir("jbd2", NULL);
2248         if (jbd2_debugfs_dir)
2249                 jbd2_debug = debugfs_create_u8(JBD2_DEBUG_NAME,
2250                                                S_IRUGO | S_IWUSR,
2251                                                jbd2_debugfs_dir,
2252                                                &jbd2_journal_enable_debug);
2253 }
2254
2255 static void __exit jbd2_remove_debugfs_entry(void)
2256 {
2257         debugfs_remove(jbd2_debug);
2258         debugfs_remove(jbd2_debugfs_dir);
2259 }
2260
2261 #else
2262
2263 static void __init jbd2_create_debugfs_entry(void)
2264 {
2265 }
2266
2267 static void __exit jbd2_remove_debugfs_entry(void)
2268 {
2269 }
2270
2271 #endif
2272
2273 #ifdef CONFIG_PROC_FS
2274
2275 #define JBD2_STATS_PROC_NAME "fs/jbd2"
2276
2277 static void __init jbd2_create_jbd_stats_proc_entry(void)
2278 {
2279         proc_jbd2_stats = proc_mkdir(JBD2_STATS_PROC_NAME, NULL);
2280 }
2281
2282 static void __exit jbd2_remove_jbd_stats_proc_entry(void)
2283 {
2284         if (proc_jbd2_stats)
2285                 remove_proc_entry(JBD2_STATS_PROC_NAME, NULL);
2286 }
2287
2288 #else
2289
2290 #define jbd2_create_jbd_stats_proc_entry() do {} while (0)
2291 #define jbd2_remove_jbd_stats_proc_entry() do {} while (0)
2292
2293 #endif
2294
2295 struct kmem_cache *jbd2_handle_cache;
2296
2297 static int __init journal_init_handle_cache(void)
2298 {
2299         jbd2_handle_cache = kmem_cache_create("jbd2_journal_handle",
2300                                 sizeof(handle_t),
2301                                 0,              /* offset */
2302                                 SLAB_TEMPORARY, /* flags */
2303                                 NULL);          /* ctor */
2304         if (jbd2_handle_cache == NULL) {
2305                 printk(KERN_EMERG "JBD: failed to create handle cache\n");
2306                 return -ENOMEM;
2307         }
2308         return 0;
2309 }
2310
2311 static void jbd2_journal_destroy_handle_cache(void)
2312 {
2313         if (jbd2_handle_cache)
2314                 kmem_cache_destroy(jbd2_handle_cache);
2315 }
2316
2317 /*
2318  * Module startup and shutdown
2319  */
2320
2321 static int __init journal_init_caches(void)
2322 {
2323         int ret;
2324
2325         ret = jbd2_journal_init_revoke_caches();
2326         if (ret == 0)
2327                 ret = journal_init_jbd2_journal_head_cache();
2328         if (ret == 0)
2329                 ret = journal_init_handle_cache();
2330         return ret;
2331 }
2332
2333 static void jbd2_journal_destroy_caches(void)
2334 {
2335         jbd2_journal_destroy_revoke_caches();
2336         jbd2_journal_destroy_jbd2_journal_head_cache();
2337         jbd2_journal_destroy_handle_cache();
2338         jbd2_journal_destroy_slabs();
2339 }
2340
2341 static int __init journal_init(void)
2342 {
2343         int ret;
2344
2345         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
2346
2347         ret = journal_init_caches();
2348         if (ret == 0) {
2349                 jbd2_create_debugfs_entry();
2350                 jbd2_create_jbd_stats_proc_entry();
2351         } else {
2352                 jbd2_journal_destroy_caches();
2353         }
2354         return ret;
2355 }
2356
2357 static void __exit journal_exit(void)
2358 {
2359 #ifdef CONFIG_JBD2_DEBUG
2360         int n = atomic_read(&nr_journal_heads);
2361         if (n)
2362                 printk(KERN_EMERG "JBD: leaked %d journal_heads!\n", n);
2363 #endif
2364         jbd2_remove_debugfs_entry();
2365         jbd2_remove_jbd_stats_proc_entry();
2366         jbd2_journal_destroy_caches();
2367 }
2368
2369 /* 
2370  * jbd2_dev_to_name is a utility function used by the jbd2 and ext4 
2371  * tracing infrastructure to map a dev_t to a device name.
2372  *
2373  * The caller should use rcu_read_lock() in order to make sure the
2374  * device name stays valid until its done with it.  We use
2375  * rcu_read_lock() as well to make sure we're safe in case the caller
2376  * gets sloppy, and because rcu_read_lock() is cheap and can be safely
2377  * nested.
2378  */
2379 struct devname_cache {
2380         struct rcu_head rcu;
2381         dev_t           device;
2382         char            devname[BDEVNAME_SIZE];
2383 };
2384 #define CACHE_SIZE_BITS 6
2385 static struct devname_cache *devcache[1 << CACHE_SIZE_BITS];
2386 static DEFINE_SPINLOCK(devname_cache_lock);
2387
2388 static void free_devcache(struct rcu_head *rcu)
2389 {
2390         kfree(rcu);
2391 }
2392
2393 const char *jbd2_dev_to_name(dev_t device)
2394 {
2395         int     i = hash_32(device, CACHE_SIZE_BITS);
2396         char    *ret;
2397         struct block_device *bd;
2398         static struct devname_cache *new_dev;
2399
2400         rcu_read_lock();
2401         if (devcache[i] && devcache[i]->device == device) {
2402                 ret = devcache[i]->devname;
2403                 rcu_read_unlock();
2404                 return ret;
2405         }
2406         rcu_read_unlock();
2407
2408         new_dev = kmalloc(sizeof(struct devname_cache), GFP_KERNEL);
2409         if (!new_dev)
2410                 return "NODEV-ALLOCFAILURE"; /* Something non-NULL */
2411         spin_lock(&devname_cache_lock);
2412         if (devcache[i]) {
2413                 if (devcache[i]->device == device) {
2414                         kfree(new_dev);
2415                         ret = devcache[i]->devname;
2416                         spin_unlock(&devname_cache_lock);
2417                         return ret;
2418                 }
2419                 call_rcu(&devcache[i]->rcu, free_devcache);
2420         }
2421         devcache[i] = new_dev;
2422         devcache[i]->device = device;
2423         bd = bdget(device);
2424         if (bd) {
2425                 bdevname(bd, devcache[i]->devname);
2426                 bdput(bd);
2427         } else
2428                 __bdevname(device, devcache[i]->devname);
2429         ret = devcache[i]->devname;
2430         spin_unlock(&devname_cache_lock);
2431         return ret;
2432 }
2433 EXPORT_SYMBOL(jbd2_dev_to_name);
2434
2435 MODULE_LICENSE("GPL");
2436 module_init(journal_init);
2437 module_exit(journal_exit);
2438