jbd2: Change j_state_lock to be a rwlock_t
[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
46 #define CREATE_TRACE_POINTS
47 #include <trace/events/jbd2.h>
48
49 #include <asm/uaccess.h>
50 #include <asm/page.h>
51
52 EXPORT_SYMBOL(jbd2_journal_extend);
53 EXPORT_SYMBOL(jbd2_journal_stop);
54 EXPORT_SYMBOL(jbd2_journal_lock_updates);
55 EXPORT_SYMBOL(jbd2_journal_unlock_updates);
56 EXPORT_SYMBOL(jbd2_journal_get_write_access);
57 EXPORT_SYMBOL(jbd2_journal_get_create_access);
58 EXPORT_SYMBOL(jbd2_journal_get_undo_access);
59 EXPORT_SYMBOL(jbd2_journal_set_triggers);
60 EXPORT_SYMBOL(jbd2_journal_dirty_metadata);
61 EXPORT_SYMBOL(jbd2_journal_release_buffer);
62 EXPORT_SYMBOL(jbd2_journal_forget);
63 #if 0
64 EXPORT_SYMBOL(journal_sync_buffer);
65 #endif
66 EXPORT_SYMBOL(jbd2_journal_flush);
67 EXPORT_SYMBOL(jbd2_journal_revoke);
68
69 EXPORT_SYMBOL(jbd2_journal_init_dev);
70 EXPORT_SYMBOL(jbd2_journal_init_inode);
71 EXPORT_SYMBOL(jbd2_journal_update_format);
72 EXPORT_SYMBOL(jbd2_journal_check_used_features);
73 EXPORT_SYMBOL(jbd2_journal_check_available_features);
74 EXPORT_SYMBOL(jbd2_journal_set_features);
75 EXPORT_SYMBOL(jbd2_journal_load);
76 EXPORT_SYMBOL(jbd2_journal_destroy);
77 EXPORT_SYMBOL(jbd2_journal_abort);
78 EXPORT_SYMBOL(jbd2_journal_errno);
79 EXPORT_SYMBOL(jbd2_journal_ack_err);
80 EXPORT_SYMBOL(jbd2_journal_clear_err);
81 EXPORT_SYMBOL(jbd2_log_wait_commit);
82 EXPORT_SYMBOL(jbd2_log_start_commit);
83 EXPORT_SYMBOL(jbd2_journal_start_commit);
84 EXPORT_SYMBOL(jbd2_journal_force_commit_nested);
85 EXPORT_SYMBOL(jbd2_journal_wipe);
86 EXPORT_SYMBOL(jbd2_journal_blocks_per_page);
87 EXPORT_SYMBOL(jbd2_journal_invalidatepage);
88 EXPORT_SYMBOL(jbd2_journal_try_to_free_buffers);
89 EXPORT_SYMBOL(jbd2_journal_force_commit);
90 EXPORT_SYMBOL(jbd2_journal_file_inode);
91 EXPORT_SYMBOL(jbd2_journal_init_jbd_inode);
92 EXPORT_SYMBOL(jbd2_journal_release_jbd_inode);
93 EXPORT_SYMBOL(jbd2_journal_begin_ordered_truncate);
94
95 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
96 static void __journal_abort_soft (journal_t *journal, int errno);
97 static int jbd2_journal_create_slab(size_t slab_size);
98
99 /*
100  * Helper function used to manage commit timeouts
101  */
102
103 static void commit_timeout(unsigned long __data)
104 {
105         struct task_struct * p = (struct task_struct *) __data;
106
107         wake_up_process(p);
108 }
109
110 /*
111  * kjournald2: The main thread function used to manage a logging device
112  * journal.
113  *
114  * This kernel thread is responsible for two things:
115  *
116  * 1) COMMIT:  Every so often we need to commit the current state of the
117  *    filesystem to disk.  The journal thread is responsible for writing
118  *    all of the metadata buffers to disk.
119  *
120  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
121  *    of the data in that part of the log has been rewritten elsewhere on
122  *    the disk.  Flushing these old buffers to reclaim space in the log is
123  *    known as checkpointing, and this thread is responsible for that job.
124  */
125
126 static int kjournald2(void *arg)
127 {
128         journal_t *journal = arg;
129         transaction_t *transaction;
130
131         /*
132          * Set up an interval timer which can be used to trigger a commit wakeup
133          * after the commit interval expires
134          */
135         setup_timer(&journal->j_commit_timer, commit_timeout,
136                         (unsigned long)current);
137
138         /* Record that the journal thread is running */
139         journal->j_task = current;
140         wake_up(&journal->j_wait_done_commit);
141
142         /*
143          * And now, wait forever for commit wakeup events.
144          */
145         write_lock(&journal->j_state_lock);
146
147 loop:
148         if (journal->j_flags & JBD2_UNMOUNT)
149                 goto end_loop;
150
151         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
152                 journal->j_commit_sequence, journal->j_commit_request);
153
154         if (journal->j_commit_sequence != journal->j_commit_request) {
155                 jbd_debug(1, "OK, requests differ\n");
156                 write_unlock(&journal->j_state_lock);
157                 del_timer_sync(&journal->j_commit_timer);
158                 jbd2_journal_commit_transaction(journal);
159                 write_lock(&journal->j_state_lock);
160                 goto loop;
161         }
162
163         wake_up(&journal->j_wait_done_commit);
164         if (freezing(current)) {
165                 /*
166                  * The simpler the better. Flushing journal isn't a
167                  * good idea, because that depends on threads that may
168                  * be already stopped.
169                  */
170                 jbd_debug(1, "Now suspending kjournald2\n");
171                 write_unlock(&journal->j_state_lock);
172                 refrigerator();
173                 write_lock(&journal->j_state_lock);
174         } else {
175                 /*
176                  * We assume on resume that commits are already there,
177                  * so we don't sleep
178                  */
179                 DEFINE_WAIT(wait);
180                 int should_sleep = 1;
181
182                 prepare_to_wait(&journal->j_wait_commit, &wait,
183                                 TASK_INTERRUPTIBLE);
184                 if (journal->j_commit_sequence != journal->j_commit_request)
185                         should_sleep = 0;
186                 transaction = journal->j_running_transaction;
187                 if (transaction && time_after_eq(jiffies,
188                                                 transaction->t_expires))
189                         should_sleep = 0;
190                 if (journal->j_flags & JBD2_UNMOUNT)
191                         should_sleep = 0;
192                 if (should_sleep) {
193                         write_unlock(&journal->j_state_lock);
194                         schedule();
195                         write_lock(&journal->j_state_lock);
196                 }
197                 finish_wait(&journal->j_wait_commit, &wait);
198         }
199
200         jbd_debug(1, "kjournald2 wakes\n");
201
202         /*
203          * Were we woken up by a commit wakeup event?
204          */
205         transaction = journal->j_running_transaction;
206         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
207                 journal->j_commit_request = transaction->t_tid;
208                 jbd_debug(1, "woke because of timeout\n");
209         }
210         goto loop;
211
212 end_loop:
213         write_unlock(&journal->j_state_lock);
214         del_timer_sync(&journal->j_commit_timer);
215         journal->j_task = NULL;
216         wake_up(&journal->j_wait_done_commit);
217         jbd_debug(1, "Journal thread exiting.\n");
218         return 0;
219 }
220
221 static int jbd2_journal_start_thread(journal_t *journal)
222 {
223         struct task_struct *t;
224
225         t = kthread_run(kjournald2, journal, "jbd2/%s",
226                         journal->j_devname);
227         if (IS_ERR(t))
228                 return PTR_ERR(t);
229
230         wait_event(journal->j_wait_done_commit, journal->j_task != NULL);
231         return 0;
232 }
233
234 static void journal_kill_thread(journal_t *journal)
235 {
236         write_lock(&journal->j_state_lock);
237         journal->j_flags |= JBD2_UNMOUNT;
238
239         while (journal->j_task) {
240                 wake_up(&journal->j_wait_commit);
241                 write_unlock(&journal->j_state_lock);
242                 wait_event(journal->j_wait_done_commit, journal->j_task == NULL);
243                 write_lock(&journal->j_state_lock);
244         }
245         write_unlock(&journal->j_state_lock);
246 }
247
248 /*
249  * jbd2_journal_write_metadata_buffer: write a metadata buffer to the journal.
250  *
251  * Writes a metadata buffer to a given disk block.  The actual IO is not
252  * performed but a new buffer_head is constructed which labels the data
253  * to be written with the correct destination disk block.
254  *
255  * Any magic-number escaping which needs to be done will cause a
256  * copy-out here.  If the buffer happens to start with the
257  * JBD2_MAGIC_NUMBER, then we can't write it to the log directly: the
258  * magic number is only written to the log for descripter blocks.  In
259  * this case, we copy the data and replace the first word with 0, and we
260  * return a result code which indicates that this buffer needs to be
261  * marked as an escaped buffer in the corresponding log descriptor
262  * block.  The missing word can then be restored when the block is read
263  * during recovery.
264  *
265  * If the source buffer has already been modified by a new transaction
266  * since we took the last commit snapshot, we use the frozen copy of
267  * that data for IO.  If we end up using the existing buffer_head's data
268  * for the write, then we *have* to lock the buffer to prevent anyone
269  * else from using and possibly modifying it while the IO is in
270  * progress.
271  *
272  * The function returns a pointer to the buffer_heads to be used for IO.
273  *
274  * We assume that the journal has already been locked in this function.
275  *
276  * Return value:
277  *  <0: Error
278  * >=0: Finished OK
279  *
280  * On success:
281  * Bit 0 set == escape performed on the data
282  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
283  */
284
285 int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
286                                   struct journal_head  *jh_in,
287                                   struct journal_head **jh_out,
288                                   unsigned long long blocknr)
289 {
290         int need_copy_out = 0;
291         int done_copy_out = 0;
292         int do_escape = 0;
293         char *mapped_data;
294         struct buffer_head *new_bh;
295         struct journal_head *new_jh;
296         struct page *new_page;
297         unsigned int new_offset;
298         struct buffer_head *bh_in = jh2bh(jh_in);
299         struct jbd2_buffer_trigger_type *triggers;
300         journal_t *journal = transaction->t_journal;
301
302         /*
303          * The buffer really shouldn't be locked: only the current committing
304          * transaction is allowed to write it, so nobody else is allowed
305          * to do any IO.
306          *
307          * akpm: except if we're journalling data, and write() output is
308          * also part of a shared mapping, and another thread has
309          * decided to launch a writepage() against this buffer.
310          */
311         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
312
313 retry_alloc:
314         new_bh = alloc_buffer_head(GFP_NOFS);
315         if (!new_bh) {
316                 /*
317                  * Failure is not an option, but __GFP_NOFAIL is going
318                  * away; so we retry ourselves here.
319                  */
320                 congestion_wait(BLK_RW_ASYNC, HZ/50);
321                 goto retry_alloc;
322         }
323
324         /* keep subsequent assertions sane */
325         new_bh->b_state = 0;
326         init_buffer(new_bh, NULL, NULL);
327         atomic_set(&new_bh->b_count, 1);
328         new_jh = jbd2_journal_add_journal_head(new_bh); /* This sleeps */
329
330         /*
331          * If a new transaction has already done a buffer copy-out, then
332          * we use that version of the data for the commit.
333          */
334         jbd_lock_bh_state(bh_in);
335 repeat:
336         if (jh_in->b_frozen_data) {
337                 done_copy_out = 1;
338                 new_page = virt_to_page(jh_in->b_frozen_data);
339                 new_offset = offset_in_page(jh_in->b_frozen_data);
340                 triggers = jh_in->b_frozen_triggers;
341         } else {
342                 new_page = jh2bh(jh_in)->b_page;
343                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
344                 triggers = jh_in->b_triggers;
345         }
346
347         mapped_data = kmap_atomic(new_page, KM_USER0);
348         /*
349          * Fire any commit trigger.  Do this before checking for escaping,
350          * as the trigger may modify the magic offset.  If a copy-out
351          * happens afterwards, it will have the correct data in the buffer.
352          */
353         jbd2_buffer_commit_trigger(jh_in, mapped_data + new_offset,
354                                    triggers);
355
356         /*
357          * Check for escaping
358          */
359         if (*((__be32 *)(mapped_data + new_offset)) ==
360                                 cpu_to_be32(JBD2_MAGIC_NUMBER)) {
361                 need_copy_out = 1;
362                 do_escape = 1;
363         }
364         kunmap_atomic(mapped_data, KM_USER0);
365
366         /*
367          * Do we need to do a data copy?
368          */
369         if (need_copy_out && !done_copy_out) {
370                 char *tmp;
371
372                 jbd_unlock_bh_state(bh_in);
373                 tmp = jbd2_alloc(bh_in->b_size, GFP_NOFS);
374                 if (!tmp) {
375                         jbd2_journal_put_journal_head(new_jh);
376                         return -ENOMEM;
377                 }
378                 jbd_lock_bh_state(bh_in);
379                 if (jh_in->b_frozen_data) {
380                         jbd2_free(tmp, bh_in->b_size);
381                         goto repeat;
382                 }
383
384                 jh_in->b_frozen_data = tmp;
385                 mapped_data = kmap_atomic(new_page, KM_USER0);
386                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
387                 kunmap_atomic(mapped_data, KM_USER0);
388
389                 new_page = virt_to_page(tmp);
390                 new_offset = offset_in_page(tmp);
391                 done_copy_out = 1;
392
393                 /*
394                  * This isn't strictly necessary, as we're using frozen
395                  * data for the escaping, but it keeps consistency with
396                  * b_frozen_data usage.
397                  */
398                 jh_in->b_frozen_triggers = jh_in->b_triggers;
399         }
400
401         /*
402          * Did we need to do an escaping?  Now we've done all the
403          * copying, we can finally do so.
404          */
405         if (do_escape) {
406                 mapped_data = kmap_atomic(new_page, KM_USER0);
407                 *((unsigned int *)(mapped_data + new_offset)) = 0;
408                 kunmap_atomic(mapped_data, KM_USER0);
409         }
410
411         set_bh_page(new_bh, new_page, new_offset);
412         new_jh->b_transaction = NULL;
413         new_bh->b_size = jh2bh(jh_in)->b_size;
414         new_bh->b_bdev = transaction->t_journal->j_dev;
415         new_bh->b_blocknr = blocknr;
416         set_buffer_mapped(new_bh);
417         set_buffer_dirty(new_bh);
418
419         *jh_out = new_jh;
420
421         /*
422          * The to-be-written buffer needs to get moved to the io queue,
423          * and the original buffer whose contents we are shadowing or
424          * copying is moved to the transaction's shadow queue.
425          */
426         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
427         spin_lock(&journal->j_list_lock);
428         __jbd2_journal_file_buffer(jh_in, transaction, BJ_Shadow);
429         spin_unlock(&journal->j_list_lock);
430         jbd_unlock_bh_state(bh_in);
431
432         JBUFFER_TRACE(new_jh, "file as BJ_IO");
433         jbd2_journal_file_buffer(new_jh, transaction, BJ_IO);
434
435         return do_escape | (done_copy_out << 1);
436 }
437
438 /*
439  * Allocation code for the journal file.  Manage the space left in the
440  * journal, so that we can begin checkpointing when appropriate.
441  */
442
443 /*
444  * __jbd2_log_space_left: Return the number of free blocks left in the journal.
445  *
446  * Called with the journal already locked.
447  *
448  * Called under j_state_lock
449  */
450
451 int __jbd2_log_space_left(journal_t *journal)
452 {
453         int left = journal->j_free;
454
455         /* assert_spin_locked(&journal->j_state_lock); */
456
457         /*
458          * Be pessimistic here about the number of those free blocks which
459          * might be required for log descriptor control blocks.
460          */
461
462 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
463
464         left -= MIN_LOG_RESERVED_BLOCKS;
465
466         if (left <= 0)
467                 return 0;
468         left -= (left >> 3);
469         return left;
470 }
471
472 /*
473  * Called under j_state_lock.  Returns true if a transaction commit was started.
474  */
475 int __jbd2_log_start_commit(journal_t *journal, tid_t target)
476 {
477         /*
478          * Are we already doing a recent enough commit?
479          */
480         if (!tid_geq(journal->j_commit_request, target)) {
481                 /*
482                  * We want a new commit: OK, mark the request and wakup the
483                  * commit thread.  We do _not_ do the commit ourselves.
484                  */
485
486                 journal->j_commit_request = target;
487                 jbd_debug(1, "JBD: requesting commit %d/%d\n",
488                           journal->j_commit_request,
489                           journal->j_commit_sequence);
490                 wake_up(&journal->j_wait_commit);
491                 return 1;
492         }
493         return 0;
494 }
495
496 int jbd2_log_start_commit(journal_t *journal, tid_t tid)
497 {
498         int ret;
499
500         write_lock(&journal->j_state_lock);
501         ret = __jbd2_log_start_commit(journal, tid);
502         write_unlock(&journal->j_state_lock);
503         return ret;
504 }
505
506 /*
507  * Force and wait upon a commit if the calling process is not within
508  * transaction.  This is used for forcing out undo-protected data which contains
509  * bitmaps, when the fs is running out of space.
510  *
511  * We can only force the running transaction if we don't have an active handle;
512  * otherwise, we will deadlock.
513  *
514  * Returns true if a transaction was started.
515  */
516 int jbd2_journal_force_commit_nested(journal_t *journal)
517 {
518         transaction_t *transaction = NULL;
519         tid_t tid;
520
521         read_lock(&journal->j_state_lock);
522         if (journal->j_running_transaction && !current->journal_info) {
523                 transaction = journal->j_running_transaction;
524                 __jbd2_log_start_commit(journal, transaction->t_tid);
525         } else if (journal->j_committing_transaction)
526                 transaction = journal->j_committing_transaction;
527
528         if (!transaction) {
529                 read_unlock(&journal->j_state_lock);
530                 return 0;       /* Nothing to retry */
531         }
532
533         tid = transaction->t_tid;
534         read_unlock(&journal->j_state_lock);
535         jbd2_log_wait_commit(journal, tid);
536         return 1;
537 }
538
539 /*
540  * Start a commit of the current running transaction (if any).  Returns true
541  * if a transaction is going to be committed (or is currently already
542  * committing), and fills its tid in at *ptid
543  */
544 int jbd2_journal_start_commit(journal_t *journal, tid_t *ptid)
545 {
546         int ret = 0;
547
548         write_lock(&journal->j_state_lock);
549         if (journal->j_running_transaction) {
550                 tid_t tid = journal->j_running_transaction->t_tid;
551
552                 __jbd2_log_start_commit(journal, tid);
553                 /* There's a running transaction and we've just made sure
554                  * it's commit has been scheduled. */
555                 if (ptid)
556                         *ptid = tid;
557                 ret = 1;
558         } else if (journal->j_committing_transaction) {
559                 /*
560                  * If ext3_write_super() recently started a commit, then we
561                  * have to wait for completion of that transaction
562                  */
563                 if (ptid)
564                         *ptid = journal->j_committing_transaction->t_tid;
565                 ret = 1;
566         }
567         write_unlock(&journal->j_state_lock);
568         return ret;
569 }
570
571 /*
572  * Wait for a specified commit to complete.
573  * The caller may not hold the journal lock.
574  */
575 int jbd2_log_wait_commit(journal_t *journal, tid_t tid)
576 {
577         int err = 0;
578
579         read_lock(&journal->j_state_lock);
580 #ifdef CONFIG_JBD2_DEBUG
581         if (!tid_geq(journal->j_commit_request, tid)) {
582                 printk(KERN_EMERG
583                        "%s: error: j_commit_request=%d, tid=%d\n",
584                        __func__, journal->j_commit_request, tid);
585         }
586 #endif
587         while (tid_gt(tid, journal->j_commit_sequence)) {
588                 jbd_debug(1, "JBD: want %d, j_commit_sequence=%d\n",
589                                   tid, journal->j_commit_sequence);
590                 wake_up(&journal->j_wait_commit);
591                 read_unlock(&journal->j_state_lock);
592                 wait_event(journal->j_wait_done_commit,
593                                 !tid_gt(tid, journal->j_commit_sequence));
594                 read_lock(&journal->j_state_lock);
595         }
596         read_unlock(&journal->j_state_lock);
597
598         if (unlikely(is_journal_aborted(journal))) {
599                 printk(KERN_EMERG "journal commit I/O error\n");
600                 err = -EIO;
601         }
602         return err;
603 }
604
605 /*
606  * Log buffer allocation routines:
607  */
608
609 int jbd2_journal_next_log_block(journal_t *journal, unsigned long long *retp)
610 {
611         unsigned long blocknr;
612
613         write_lock(&journal->j_state_lock);
614         J_ASSERT(journal->j_free > 1);
615
616         blocknr = journal->j_head;
617         journal->j_head++;
618         journal->j_free--;
619         if (journal->j_head == journal->j_last)
620                 journal->j_head = journal->j_first;
621         write_unlock(&journal->j_state_lock);
622         return jbd2_journal_bmap(journal, blocknr, retp);
623 }
624
625 /*
626  * Conversion of logical to physical block numbers for the journal
627  *
628  * On external journals the journal blocks are identity-mapped, so
629  * this is a no-op.  If needed, we can use j_blk_offset - everything is
630  * ready.
631  */
632 int jbd2_journal_bmap(journal_t *journal, unsigned long blocknr,
633                  unsigned long long *retp)
634 {
635         int err = 0;
636         unsigned long long ret;
637
638         if (journal->j_inode) {
639                 ret = bmap(journal->j_inode, blocknr);
640                 if (ret)
641                         *retp = ret;
642                 else {
643                         printk(KERN_ALERT "%s: journal block not found "
644                                         "at offset %lu on %s\n",
645                                __func__, blocknr, journal->j_devname);
646                         err = -EIO;
647                         __journal_abort_soft(journal, err);
648                 }
649         } else {
650                 *retp = blocknr; /* +journal->j_blk_offset */
651         }
652         return err;
653 }
654
655 /*
656  * We play buffer_head aliasing tricks to write data/metadata blocks to
657  * the journal without copying their contents, but for journal
658  * descriptor blocks we do need to generate bona fide buffers.
659  *
660  * After the caller of jbd2_journal_get_descriptor_buffer() has finished modifying
661  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
662  * But we don't bother doing that, so there will be coherency problems with
663  * mmaps of blockdevs which hold live JBD-controlled filesystems.
664  */
665 struct journal_head *jbd2_journal_get_descriptor_buffer(journal_t *journal)
666 {
667         struct buffer_head *bh;
668         unsigned long long blocknr;
669         int err;
670
671         err = jbd2_journal_next_log_block(journal, &blocknr);
672
673         if (err)
674                 return NULL;
675
676         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
677         if (!bh)
678                 return NULL;
679         lock_buffer(bh);
680         memset(bh->b_data, 0, journal->j_blocksize);
681         set_buffer_uptodate(bh);
682         unlock_buffer(bh);
683         BUFFER_TRACE(bh, "return this buffer");
684         return jbd2_journal_add_journal_head(bh);
685 }
686
687 struct jbd2_stats_proc_session {
688         journal_t *journal;
689         struct transaction_stats_s *stats;
690         int start;
691         int max;
692 };
693
694 static void *jbd2_seq_info_start(struct seq_file *seq, loff_t *pos)
695 {
696         return *pos ? NULL : SEQ_START_TOKEN;
697 }
698
699 static void *jbd2_seq_info_next(struct seq_file *seq, void *v, loff_t *pos)
700 {
701         return NULL;
702 }
703
704 static int jbd2_seq_info_show(struct seq_file *seq, void *v)
705 {
706         struct jbd2_stats_proc_session *s = seq->private;
707
708         if (v != SEQ_START_TOKEN)
709                 return 0;
710         seq_printf(seq, "%lu transaction, each up to %u blocks\n",
711                         s->stats->ts_tid,
712                         s->journal->j_max_transaction_buffers);
713         if (s->stats->ts_tid == 0)
714                 return 0;
715         seq_printf(seq, "average: \n  %ums waiting for transaction\n",
716             jiffies_to_msecs(s->stats->run.rs_wait / s->stats->ts_tid));
717         seq_printf(seq, "  %ums running transaction\n",
718             jiffies_to_msecs(s->stats->run.rs_running / s->stats->ts_tid));
719         seq_printf(seq, "  %ums transaction was being locked\n",
720             jiffies_to_msecs(s->stats->run.rs_locked / s->stats->ts_tid));
721         seq_printf(seq, "  %ums flushing data (in ordered mode)\n",
722             jiffies_to_msecs(s->stats->run.rs_flushing / s->stats->ts_tid));
723         seq_printf(seq, "  %ums logging transaction\n",
724             jiffies_to_msecs(s->stats->run.rs_logging / s->stats->ts_tid));
725         seq_printf(seq, "  %lluus average transaction commit time\n",
726                    div_u64(s->journal->j_average_commit_time, 1000));
727         seq_printf(seq, "  %lu handles per transaction\n",
728             s->stats->run.rs_handle_count / s->stats->ts_tid);
729         seq_printf(seq, "  %lu blocks per transaction\n",
730             s->stats->run.rs_blocks / s->stats->ts_tid);
731         seq_printf(seq, "  %lu logged blocks per transaction\n",
732             s->stats->run.rs_blocks_logged / s->stats->ts_tid);
733         return 0;
734 }
735
736 static void jbd2_seq_info_stop(struct seq_file *seq, void *v)
737 {
738 }
739
740 static const struct seq_operations jbd2_seq_info_ops = {
741         .start  = jbd2_seq_info_start,
742         .next   = jbd2_seq_info_next,
743         .stop   = jbd2_seq_info_stop,
744         .show   = jbd2_seq_info_show,
745 };
746
747 static int jbd2_seq_info_open(struct inode *inode, struct file *file)
748 {
749         journal_t *journal = PDE(inode)->data;
750         struct jbd2_stats_proc_session *s;
751         int rc, size;
752
753         s = kmalloc(sizeof(*s), GFP_KERNEL);
754         if (s == NULL)
755                 return -ENOMEM;
756         size = sizeof(struct transaction_stats_s);
757         s->stats = kmalloc(size, GFP_KERNEL);
758         if (s->stats == NULL) {
759                 kfree(s);
760                 return -ENOMEM;
761         }
762         spin_lock(&journal->j_history_lock);
763         memcpy(s->stats, &journal->j_stats, size);
764         s->journal = journal;
765         spin_unlock(&journal->j_history_lock);
766
767         rc = seq_open(file, &jbd2_seq_info_ops);
768         if (rc == 0) {
769                 struct seq_file *m = file->private_data;
770                 m->private = s;
771         } else {
772                 kfree(s->stats);
773                 kfree(s);
774         }
775         return rc;
776
777 }
778
779 static int jbd2_seq_info_release(struct inode *inode, struct file *file)
780 {
781         struct seq_file *seq = file->private_data;
782         struct jbd2_stats_proc_session *s = seq->private;
783         kfree(s->stats);
784         kfree(s);
785         return seq_release(inode, file);
786 }
787
788 static const struct file_operations jbd2_seq_info_fops = {
789         .owner          = THIS_MODULE,
790         .open           = jbd2_seq_info_open,
791         .read           = seq_read,
792         .llseek         = seq_lseek,
793         .release        = jbd2_seq_info_release,
794 };
795
796 static struct proc_dir_entry *proc_jbd2_stats;
797
798 static void jbd2_stats_proc_init(journal_t *journal)
799 {
800         journal->j_proc_entry = proc_mkdir(journal->j_devname, proc_jbd2_stats);
801         if (journal->j_proc_entry) {
802                 proc_create_data("info", S_IRUGO, journal->j_proc_entry,
803                                  &jbd2_seq_info_fops, journal);
804         }
805 }
806
807 static void jbd2_stats_proc_exit(journal_t *journal)
808 {
809         remove_proc_entry("info", journal->j_proc_entry);
810         remove_proc_entry(journal->j_devname, proc_jbd2_stats);
811 }
812
813 /*
814  * Management for journal control blocks: functions to create and
815  * destroy journal_t structures, and to initialise and read existing
816  * journal blocks from disk.  */
817
818 /* First: create and setup a journal_t object in memory.  We initialise
819  * very few fields yet: that has to wait until we have created the
820  * journal structures from from scratch, or loaded them from disk. */
821
822 static journal_t * journal_init_common (void)
823 {
824         journal_t *journal;
825         int err;
826
827         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
828         if (!journal)
829                 goto fail;
830
831         init_waitqueue_head(&journal->j_wait_transaction_locked);
832         init_waitqueue_head(&journal->j_wait_logspace);
833         init_waitqueue_head(&journal->j_wait_done_commit);
834         init_waitqueue_head(&journal->j_wait_checkpoint);
835         init_waitqueue_head(&journal->j_wait_commit);
836         init_waitqueue_head(&journal->j_wait_updates);
837         mutex_init(&journal->j_barrier);
838         mutex_init(&journal->j_checkpoint_mutex);
839         spin_lock_init(&journal->j_revoke_lock);
840         spin_lock_init(&journal->j_list_lock);
841         rwlock_init(&journal->j_state_lock);
842
843         journal->j_commit_interval = (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE);
844         journal->j_min_batch_time = 0;
845         journal->j_max_batch_time = 15000; /* 15ms */
846
847         /* The journal is marked for error until we succeed with recovery! */
848         journal->j_flags = JBD2_ABORT;
849
850         /* Set up a default-sized revoke table for the new mount. */
851         err = jbd2_journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
852         if (err) {
853                 kfree(journal);
854                 goto fail;
855         }
856
857         spin_lock_init(&journal->j_history_lock);
858
859         return journal;
860 fail:
861         return NULL;
862 }
863
864 /* jbd2_journal_init_dev and jbd2_journal_init_inode:
865  *
866  * Create a journal structure assigned some fixed set of disk blocks to
867  * the journal.  We don't actually touch those disk blocks yet, but we
868  * need to set up all of the mapping information to tell the journaling
869  * system where the journal blocks are.
870  *
871  */
872
873 /**
874  *  journal_t * jbd2_journal_init_dev() - creates and initialises a journal structure
875  *  @bdev: Block device on which to create the journal
876  *  @fs_dev: Device which hold journalled filesystem for this journal.
877  *  @start: Block nr Start of journal.
878  *  @len:  Length of the journal in blocks.
879  *  @blocksize: blocksize of journalling device
880  *
881  *  Returns: a newly created journal_t *
882  *
883  *  jbd2_journal_init_dev creates a journal which maps a fixed contiguous
884  *  range of blocks on an arbitrary block device.
885  *
886  */
887 journal_t * jbd2_journal_init_dev(struct block_device *bdev,
888                         struct block_device *fs_dev,
889                         unsigned long long start, int len, int blocksize)
890 {
891         journal_t *journal = journal_init_common();
892         struct buffer_head *bh;
893         char *p;
894         int n;
895
896         if (!journal)
897                 return NULL;
898
899         /* journal descriptor can store up to n blocks -bzzz */
900         journal->j_blocksize = blocksize;
901         jbd2_stats_proc_init(journal);
902         n = journal->j_blocksize / sizeof(journal_block_tag_t);
903         journal->j_wbufsize = n;
904         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
905         if (!journal->j_wbuf) {
906                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
907                         __func__);
908                 goto out_err;
909         }
910         journal->j_dev = bdev;
911         journal->j_fs_dev = fs_dev;
912         journal->j_blk_offset = start;
913         journal->j_maxlen = len;
914         bdevname(journal->j_dev, journal->j_devname);
915         p = journal->j_devname;
916         while ((p = strchr(p, '/')))
917                 *p = '!';
918
919         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
920         if (!bh) {
921                 printk(KERN_ERR
922                        "%s: Cannot get buffer for journal superblock\n",
923                        __func__);
924                 goto out_err;
925         }
926         journal->j_sb_buffer = bh;
927         journal->j_superblock = (journal_superblock_t *)bh->b_data;
928
929         return journal;
930 out_err:
931         kfree(journal->j_wbuf);
932         jbd2_stats_proc_exit(journal);
933         kfree(journal);
934         return NULL;
935 }
936
937 /**
938  *  journal_t * jbd2_journal_init_inode () - creates a journal which maps to a inode.
939  *  @inode: An inode to create the journal in
940  *
941  * jbd2_journal_init_inode creates a journal which maps an on-disk inode as
942  * the journal.  The inode must exist already, must support bmap() and
943  * must have all data blocks preallocated.
944  */
945 journal_t * jbd2_journal_init_inode (struct inode *inode)
946 {
947         struct buffer_head *bh;
948         journal_t *journal = journal_init_common();
949         char *p;
950         int err;
951         int n;
952         unsigned long long blocknr;
953
954         if (!journal)
955                 return NULL;
956
957         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
958         journal->j_inode = inode;
959         bdevname(journal->j_dev, journal->j_devname);
960         p = journal->j_devname;
961         while ((p = strchr(p, '/')))
962                 *p = '!';
963         p = journal->j_devname + strlen(journal->j_devname);
964         sprintf(p, "-%lu", journal->j_inode->i_ino);
965         jbd_debug(1,
966                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
967                   journal, inode->i_sb->s_id, inode->i_ino,
968                   (long long) inode->i_size,
969                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
970
971         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
972         journal->j_blocksize = inode->i_sb->s_blocksize;
973         jbd2_stats_proc_init(journal);
974
975         /* journal descriptor can store up to n blocks -bzzz */
976         n = journal->j_blocksize / sizeof(journal_block_tag_t);
977         journal->j_wbufsize = n;
978         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
979         if (!journal->j_wbuf) {
980                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
981                         __func__);
982                 goto out_err;
983         }
984
985         err = jbd2_journal_bmap(journal, 0, &blocknr);
986         /* If that failed, give up */
987         if (err) {
988                 printk(KERN_ERR "%s: Cannnot locate journal superblock\n",
989                        __func__);
990                 goto out_err;
991         }
992
993         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
994         if (!bh) {
995                 printk(KERN_ERR
996                        "%s: Cannot get buffer for journal superblock\n",
997                        __func__);
998                 goto out_err;
999         }
1000         journal->j_sb_buffer = bh;
1001         journal->j_superblock = (journal_superblock_t *)bh->b_data;
1002
1003         return journal;
1004 out_err:
1005         kfree(journal->j_wbuf);
1006         jbd2_stats_proc_exit(journal);
1007         kfree(journal);
1008         return NULL;
1009 }
1010
1011 /*
1012  * If the journal init or create aborts, we need to mark the journal
1013  * superblock as being NULL to prevent the journal destroy from writing
1014  * back a bogus superblock.
1015  */
1016 static void journal_fail_superblock (journal_t *journal)
1017 {
1018         struct buffer_head *bh = journal->j_sb_buffer;
1019         brelse(bh);
1020         journal->j_sb_buffer = NULL;
1021 }
1022
1023 /*
1024  * Given a journal_t structure, initialise the various fields for
1025  * startup of a new journaling session.  We use this both when creating
1026  * a journal, and after recovering an old journal to reset it for
1027  * subsequent use.
1028  */
1029
1030 static int journal_reset(journal_t *journal)
1031 {
1032         journal_superblock_t *sb = journal->j_superblock;
1033         unsigned long long first, last;
1034
1035         first = be32_to_cpu(sb->s_first);
1036         last = be32_to_cpu(sb->s_maxlen);
1037         if (first + JBD2_MIN_JOURNAL_BLOCKS > last + 1) {
1038                 printk(KERN_ERR "JBD: Journal too short (blocks %llu-%llu).\n",
1039                        first, last);
1040                 journal_fail_superblock(journal);
1041                 return -EINVAL;
1042         }
1043
1044         journal->j_first = first;
1045         journal->j_last = last;
1046
1047         journal->j_head = first;
1048         journal->j_tail = first;
1049         journal->j_free = last - first;
1050
1051         journal->j_tail_sequence = journal->j_transaction_sequence;
1052         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
1053         journal->j_commit_request = journal->j_commit_sequence;
1054
1055         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
1056
1057         /* Add the dynamic fields and write it to disk. */
1058         jbd2_journal_update_superblock(journal, 1);
1059         return jbd2_journal_start_thread(journal);
1060 }
1061
1062 /**
1063  * void jbd2_journal_update_superblock() - Update journal sb on disk.
1064  * @journal: The journal to update.
1065  * @wait: Set to '0' if you don't want to wait for IO completion.
1066  *
1067  * Update a journal's dynamic superblock fields and write it to disk,
1068  * optionally waiting for the IO to complete.
1069  */
1070 void jbd2_journal_update_superblock(journal_t *journal, int wait)
1071 {
1072         journal_superblock_t *sb = journal->j_superblock;
1073         struct buffer_head *bh = journal->j_sb_buffer;
1074
1075         /*
1076          * As a special case, if the on-disk copy is already marked as needing
1077          * no recovery (s_start == 0) and there are no outstanding transactions
1078          * in the filesystem, then we can safely defer the superblock update
1079          * until the next commit by setting JBD2_FLUSHED.  This avoids
1080          * attempting a write to a potential-readonly device.
1081          */
1082         if (sb->s_start == 0 && journal->j_tail_sequence ==
1083                                 journal->j_transaction_sequence) {
1084                 jbd_debug(1,"JBD: Skipping superblock update on recovered sb "
1085                         "(start %ld, seq %d, errno %d)\n",
1086                         journal->j_tail, journal->j_tail_sequence,
1087                         journal->j_errno);
1088                 goto out;
1089         }
1090
1091         if (buffer_write_io_error(bh)) {
1092                 /*
1093                  * Oh, dear.  A previous attempt to write the journal
1094                  * superblock failed.  This could happen because the
1095                  * USB device was yanked out.  Or it could happen to
1096                  * be a transient write error and maybe the block will
1097                  * be remapped.  Nothing we can do but to retry the
1098                  * write and hope for the best.
1099                  */
1100                 printk(KERN_ERR "JBD2: previous I/O error detected "
1101                        "for journal superblock update for %s.\n",
1102                        journal->j_devname);
1103                 clear_buffer_write_io_error(bh);
1104                 set_buffer_uptodate(bh);
1105         }
1106
1107         read_lock(&journal->j_state_lock);
1108         jbd_debug(1,"JBD: updating superblock (start %ld, seq %d, errno %d)\n",
1109                   journal->j_tail, journal->j_tail_sequence, journal->j_errno);
1110
1111         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
1112         sb->s_start    = cpu_to_be32(journal->j_tail);
1113         sb->s_errno    = cpu_to_be32(journal->j_errno);
1114         read_unlock(&journal->j_state_lock);
1115
1116         BUFFER_TRACE(bh, "marking dirty");
1117         mark_buffer_dirty(bh);
1118         if (wait) {
1119                 sync_dirty_buffer(bh);
1120                 if (buffer_write_io_error(bh)) {
1121                         printk(KERN_ERR "JBD2: I/O error detected "
1122                                "when updating journal superblock for %s.\n",
1123                                journal->j_devname);
1124                         clear_buffer_write_io_error(bh);
1125                         set_buffer_uptodate(bh);
1126                 }
1127         } else
1128                 ll_rw_block(SWRITE, 1, &bh);
1129
1130 out:
1131         /* If we have just flushed the log (by marking s_start==0), then
1132          * any future commit will have to be careful to update the
1133          * superblock again to re-record the true start of the log. */
1134
1135         write_lock(&journal->j_state_lock);
1136         if (sb->s_start)
1137                 journal->j_flags &= ~JBD2_FLUSHED;
1138         else
1139                 journal->j_flags |= JBD2_FLUSHED;
1140         write_unlock(&journal->j_state_lock);
1141 }
1142
1143 /*
1144  * Read the superblock for a given journal, performing initial
1145  * validation of the format.
1146  */
1147
1148 static int journal_get_superblock(journal_t *journal)
1149 {
1150         struct buffer_head *bh;
1151         journal_superblock_t *sb;
1152         int err = -EIO;
1153
1154         bh = journal->j_sb_buffer;
1155
1156         J_ASSERT(bh != NULL);
1157         if (!buffer_uptodate(bh)) {
1158                 ll_rw_block(READ, 1, &bh);
1159                 wait_on_buffer(bh);
1160                 if (!buffer_uptodate(bh)) {
1161                         printk (KERN_ERR
1162                                 "JBD: IO error reading journal superblock\n");
1163                         goto out;
1164                 }
1165         }
1166
1167         sb = journal->j_superblock;
1168
1169         err = -EINVAL;
1170
1171         if (sb->s_header.h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER) ||
1172             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1173                 printk(KERN_WARNING "JBD: no valid journal superblock found\n");
1174                 goto out;
1175         }
1176
1177         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1178         case JBD2_SUPERBLOCK_V1:
1179                 journal->j_format_version = 1;
1180                 break;
1181         case JBD2_SUPERBLOCK_V2:
1182                 journal->j_format_version = 2;
1183                 break;
1184         default:
1185                 printk(KERN_WARNING "JBD: unrecognised superblock format ID\n");
1186                 goto out;
1187         }
1188
1189         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1190                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1191         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1192                 printk (KERN_WARNING "JBD: journal file too short\n");
1193                 goto out;
1194         }
1195
1196         return 0;
1197
1198 out:
1199         journal_fail_superblock(journal);
1200         return err;
1201 }
1202
1203 /*
1204  * Load the on-disk journal superblock and read the key fields into the
1205  * journal_t.
1206  */
1207
1208 static int load_superblock(journal_t *journal)
1209 {
1210         int err;
1211         journal_superblock_t *sb;
1212
1213         err = journal_get_superblock(journal);
1214         if (err)
1215                 return err;
1216
1217         sb = journal->j_superblock;
1218
1219         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1220         journal->j_tail = be32_to_cpu(sb->s_start);
1221         journal->j_first = be32_to_cpu(sb->s_first);
1222         journal->j_last = be32_to_cpu(sb->s_maxlen);
1223         journal->j_errno = be32_to_cpu(sb->s_errno);
1224
1225         return 0;
1226 }
1227
1228
1229 /**
1230  * int jbd2_journal_load() - Read journal from disk.
1231  * @journal: Journal to act on.
1232  *
1233  * Given a journal_t structure which tells us which disk blocks contain
1234  * a journal, read the journal from disk to initialise the in-memory
1235  * structures.
1236  */
1237 int jbd2_journal_load(journal_t *journal)
1238 {
1239         int err;
1240         journal_superblock_t *sb;
1241
1242         err = load_superblock(journal);
1243         if (err)
1244                 return err;
1245
1246         sb = journal->j_superblock;
1247         /* If this is a V2 superblock, then we have to check the
1248          * features flags on it. */
1249
1250         if (journal->j_format_version >= 2) {
1251                 if ((sb->s_feature_ro_compat &
1252                      ~cpu_to_be32(JBD2_KNOWN_ROCOMPAT_FEATURES)) ||
1253                     (sb->s_feature_incompat &
1254                      ~cpu_to_be32(JBD2_KNOWN_INCOMPAT_FEATURES))) {
1255                         printk (KERN_WARNING
1256                                 "JBD: Unrecognised features on journal\n");
1257                         return -EINVAL;
1258                 }
1259         }
1260
1261         /*
1262          * Create a slab for this blocksize
1263          */
1264         err = jbd2_journal_create_slab(be32_to_cpu(sb->s_blocksize));
1265         if (err)
1266                 return err;
1267
1268         /* Let the recovery code check whether it needs to recover any
1269          * data from the journal. */
1270         if (jbd2_journal_recover(journal))
1271                 goto recovery_error;
1272
1273         if (journal->j_failed_commit) {
1274                 printk(KERN_ERR "JBD2: journal transaction %u on %s "
1275                        "is corrupt.\n", journal->j_failed_commit,
1276                        journal->j_devname);
1277                 return -EIO;
1278         }
1279
1280         /* OK, we've finished with the dynamic journal bits:
1281          * reinitialise the dynamic contents of the superblock in memory
1282          * and reset them on disk. */
1283         if (journal_reset(journal))
1284                 goto recovery_error;
1285
1286         journal->j_flags &= ~JBD2_ABORT;
1287         journal->j_flags |= JBD2_LOADED;
1288         return 0;
1289
1290 recovery_error:
1291         printk (KERN_WARNING "JBD: recovery failed\n");
1292         return -EIO;
1293 }
1294
1295 /**
1296  * void jbd2_journal_destroy() - Release a journal_t structure.
1297  * @journal: Journal to act on.
1298  *
1299  * Release a journal_t structure once it is no longer in use by the
1300  * journaled object.
1301  * Return <0 if we couldn't clean up the journal.
1302  */
1303 int jbd2_journal_destroy(journal_t *journal)
1304 {
1305         int err = 0;
1306
1307         /* Wait for the commit thread to wake up and die. */
1308         journal_kill_thread(journal);
1309
1310         /* Force a final log commit */
1311         if (journal->j_running_transaction)
1312                 jbd2_journal_commit_transaction(journal);
1313
1314         /* Force any old transactions to disk */
1315
1316         /* Totally anal locking here... */
1317         spin_lock(&journal->j_list_lock);
1318         while (journal->j_checkpoint_transactions != NULL) {
1319                 spin_unlock(&journal->j_list_lock);
1320                 mutex_lock(&journal->j_checkpoint_mutex);
1321                 jbd2_log_do_checkpoint(journal);
1322                 mutex_unlock(&journal->j_checkpoint_mutex);
1323                 spin_lock(&journal->j_list_lock);
1324         }
1325
1326         J_ASSERT(journal->j_running_transaction == NULL);
1327         J_ASSERT(journal->j_committing_transaction == NULL);
1328         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1329         spin_unlock(&journal->j_list_lock);
1330
1331         if (journal->j_sb_buffer) {
1332                 if (!is_journal_aborted(journal)) {
1333                         /* We can now mark the journal as empty. */
1334                         journal->j_tail = 0;
1335                         journal->j_tail_sequence =
1336                                 ++journal->j_transaction_sequence;
1337                         jbd2_journal_update_superblock(journal, 1);
1338                 } else {
1339                         err = -EIO;
1340                 }
1341                 brelse(journal->j_sb_buffer);
1342         }
1343
1344         if (journal->j_proc_entry)
1345                 jbd2_stats_proc_exit(journal);
1346         if (journal->j_inode)
1347                 iput(journal->j_inode);
1348         if (journal->j_revoke)
1349                 jbd2_journal_destroy_revoke(journal);
1350         kfree(journal->j_wbuf);
1351         kfree(journal);
1352
1353         return err;
1354 }
1355
1356
1357 /**
1358  *int jbd2_journal_check_used_features () - Check if features specified are used.
1359  * @journal: Journal to check.
1360  * @compat: bitmask of compatible features
1361  * @ro: bitmask of features that force read-only mount
1362  * @incompat: bitmask of incompatible features
1363  *
1364  * Check whether the journal uses all of a given set of
1365  * features.  Return true (non-zero) if it does.
1366  **/
1367
1368 int jbd2_journal_check_used_features (journal_t *journal, unsigned long compat,
1369                                  unsigned long ro, unsigned long incompat)
1370 {
1371         journal_superblock_t *sb;
1372
1373         if (!compat && !ro && !incompat)
1374                 return 1;
1375         if (journal->j_format_version == 1)
1376                 return 0;
1377
1378         sb = journal->j_superblock;
1379
1380         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1381             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1382             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1383                 return 1;
1384
1385         return 0;
1386 }
1387
1388 /**
1389  * int jbd2_journal_check_available_features() - Check feature set in journalling layer
1390  * @journal: Journal to check.
1391  * @compat: bitmask of compatible features
1392  * @ro: bitmask of features that force read-only mount
1393  * @incompat: bitmask of incompatible features
1394  *
1395  * Check whether the journaling code supports the use of
1396  * all of a given set of features on this journal.  Return true
1397  * (non-zero) if it can. */
1398
1399 int jbd2_journal_check_available_features (journal_t *journal, unsigned long compat,
1400                                       unsigned long ro, unsigned long incompat)
1401 {
1402         if (!compat && !ro && !incompat)
1403                 return 1;
1404
1405         /* We can support any known requested features iff the
1406          * superblock is in version 2.  Otherwise we fail to support any
1407          * extended sb features. */
1408
1409         if (journal->j_format_version != 2)
1410                 return 0;
1411
1412         if ((compat   & JBD2_KNOWN_COMPAT_FEATURES) == compat &&
1413             (ro       & JBD2_KNOWN_ROCOMPAT_FEATURES) == ro &&
1414             (incompat & JBD2_KNOWN_INCOMPAT_FEATURES) == incompat)
1415                 return 1;
1416
1417         return 0;
1418 }
1419
1420 /**
1421  * int jbd2_journal_set_features () - Mark a given journal feature in the superblock
1422  * @journal: Journal to act on.
1423  * @compat: bitmask of compatible features
1424  * @ro: bitmask of features that force read-only mount
1425  * @incompat: bitmask of incompatible features
1426  *
1427  * Mark a given journal feature as present on the
1428  * superblock.  Returns true if the requested features could be set.
1429  *
1430  */
1431
1432 int jbd2_journal_set_features (journal_t *journal, unsigned long compat,
1433                           unsigned long ro, unsigned long incompat)
1434 {
1435         journal_superblock_t *sb;
1436
1437         if (jbd2_journal_check_used_features(journal, compat, ro, incompat))
1438                 return 1;
1439
1440         if (!jbd2_journal_check_available_features(journal, compat, ro, incompat))
1441                 return 0;
1442
1443         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1444                   compat, ro, incompat);
1445
1446         sb = journal->j_superblock;
1447
1448         sb->s_feature_compat    |= cpu_to_be32(compat);
1449         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1450         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1451
1452         return 1;
1453 }
1454
1455 /*
1456  * jbd2_journal_clear_features () - Clear a given journal feature in the
1457  *                                  superblock
1458  * @journal: Journal to act on.
1459  * @compat: bitmask of compatible features
1460  * @ro: bitmask of features that force read-only mount
1461  * @incompat: bitmask of incompatible features
1462  *
1463  * Clear a given journal feature as present on the
1464  * superblock.
1465  */
1466 void jbd2_journal_clear_features(journal_t *journal, unsigned long compat,
1467                                 unsigned long ro, unsigned long incompat)
1468 {
1469         journal_superblock_t *sb;
1470
1471         jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
1472                   compat, ro, incompat);
1473
1474         sb = journal->j_superblock;
1475
1476         sb->s_feature_compat    &= ~cpu_to_be32(compat);
1477         sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
1478         sb->s_feature_incompat  &= ~cpu_to_be32(incompat);
1479 }
1480 EXPORT_SYMBOL(jbd2_journal_clear_features);
1481
1482 /**
1483  * int jbd2_journal_update_format () - Update on-disk journal structure.
1484  * @journal: Journal to act on.
1485  *
1486  * Given an initialised but unloaded journal struct, poke about in the
1487  * on-disk structure to update it to the most recent supported version.
1488  */
1489 int jbd2_journal_update_format (journal_t *journal)
1490 {
1491         journal_superblock_t *sb;
1492         int err;
1493
1494         err = journal_get_superblock(journal);
1495         if (err)
1496                 return err;
1497
1498         sb = journal->j_superblock;
1499
1500         switch (be32_to_cpu(sb->s_header.h_blocktype)) {
1501         case JBD2_SUPERBLOCK_V2:
1502                 return 0;
1503         case JBD2_SUPERBLOCK_V1:
1504                 return journal_convert_superblock_v1(journal, sb);
1505         default:
1506                 break;
1507         }
1508         return -EINVAL;
1509 }
1510
1511 static int journal_convert_superblock_v1(journal_t *journal,
1512                                          journal_superblock_t *sb)
1513 {
1514         int offset, blocksize;
1515         struct buffer_head *bh;
1516
1517         printk(KERN_WARNING
1518                 "JBD: Converting superblock from version 1 to 2.\n");
1519
1520         /* Pre-initialise new fields to zero */
1521         offset = ((char *) &(sb->s_feature_compat)) - ((char *) sb);
1522         blocksize = be32_to_cpu(sb->s_blocksize);
1523         memset(&sb->s_feature_compat, 0, blocksize-offset);
1524
1525         sb->s_nr_users = cpu_to_be32(1);
1526         sb->s_header.h_blocktype = cpu_to_be32(JBD2_SUPERBLOCK_V2);
1527         journal->j_format_version = 2;
1528
1529         bh = journal->j_sb_buffer;
1530         BUFFER_TRACE(bh, "marking dirty");
1531         mark_buffer_dirty(bh);
1532         sync_dirty_buffer(bh);
1533         return 0;
1534 }
1535
1536
1537 /**
1538  * int jbd2_journal_flush () - Flush journal
1539  * @journal: Journal to act on.
1540  *
1541  * Flush all data for a given journal to disk and empty the journal.
1542  * Filesystems can use this when remounting readonly to ensure that
1543  * recovery does not need to happen on remount.
1544  */
1545
1546 int jbd2_journal_flush(journal_t *journal)
1547 {
1548         int err = 0;
1549         transaction_t *transaction = NULL;
1550         unsigned long old_tail;
1551
1552         write_lock(&journal->j_state_lock);
1553
1554         /* Force everything buffered to the log... */
1555         if (journal->j_running_transaction) {
1556                 transaction = journal->j_running_transaction;
1557                 __jbd2_log_start_commit(journal, transaction->t_tid);
1558         } else if (journal->j_committing_transaction)
1559                 transaction = journal->j_committing_transaction;
1560
1561         /* Wait for the log commit to complete... */
1562         if (transaction) {
1563                 tid_t tid = transaction->t_tid;
1564
1565                 write_unlock(&journal->j_state_lock);
1566                 jbd2_log_wait_commit(journal, tid);
1567         } else {
1568                 write_unlock(&journal->j_state_lock);
1569         }
1570
1571         /* ...and flush everything in the log out to disk. */
1572         spin_lock(&journal->j_list_lock);
1573         while (!err && journal->j_checkpoint_transactions != NULL) {
1574                 spin_unlock(&journal->j_list_lock);
1575                 mutex_lock(&journal->j_checkpoint_mutex);
1576                 err = jbd2_log_do_checkpoint(journal);
1577                 mutex_unlock(&journal->j_checkpoint_mutex);
1578                 spin_lock(&journal->j_list_lock);
1579         }
1580         spin_unlock(&journal->j_list_lock);
1581
1582         if (is_journal_aborted(journal))
1583                 return -EIO;
1584
1585         jbd2_cleanup_journal_tail(journal);
1586
1587         /* Finally, mark the journal as really needing no recovery.
1588          * This sets s_start==0 in the underlying superblock, which is
1589          * the magic code for a fully-recovered superblock.  Any future
1590          * commits of data to the journal will restore the current
1591          * s_start value. */
1592         write_lock(&journal->j_state_lock);
1593         old_tail = journal->j_tail;
1594         journal->j_tail = 0;
1595         write_unlock(&journal->j_state_lock);
1596         jbd2_journal_update_superblock(journal, 1);
1597         write_lock(&journal->j_state_lock);
1598         journal->j_tail = old_tail;
1599
1600         J_ASSERT(!journal->j_running_transaction);
1601         J_ASSERT(!journal->j_committing_transaction);
1602         J_ASSERT(!journal->j_checkpoint_transactions);
1603         J_ASSERT(journal->j_head == journal->j_tail);
1604         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1605         write_unlock(&journal->j_state_lock);
1606         return 0;
1607 }
1608
1609 /**
1610  * int jbd2_journal_wipe() - Wipe journal contents
1611  * @journal: Journal to act on.
1612  * @write: flag (see below)
1613  *
1614  * Wipe out all of the contents of a journal, safely.  This will produce
1615  * a warning if the journal contains any valid recovery information.
1616  * Must be called between journal_init_*() and jbd2_journal_load().
1617  *
1618  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1619  * we merely suppress recovery.
1620  */
1621
1622 int jbd2_journal_wipe(journal_t *journal, int write)
1623 {
1624         int err = 0;
1625
1626         J_ASSERT (!(journal->j_flags & JBD2_LOADED));
1627
1628         err = load_superblock(journal);
1629         if (err)
1630                 return err;
1631
1632         if (!journal->j_tail)
1633                 goto no_recovery;
1634
1635         printk (KERN_WARNING "JBD: %s recovery information on journal\n",
1636                 write ? "Clearing" : "Ignoring");
1637
1638         err = jbd2_journal_skip_recovery(journal);
1639         if (write)
1640                 jbd2_journal_update_superblock(journal, 1);
1641
1642  no_recovery:
1643         return err;
1644 }
1645
1646 /*
1647  * Journal abort has very specific semantics, which we describe
1648  * for journal abort.
1649  *
1650  * Two internal functions, which provide abort to the jbd layer
1651  * itself are here.
1652  */
1653
1654 /*
1655  * Quick version for internal journal use (doesn't lock the journal).
1656  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1657  * and don't attempt to make any other journal updates.
1658  */
1659 void __jbd2_journal_abort_hard(journal_t *journal)
1660 {
1661         transaction_t *transaction;
1662
1663         if (journal->j_flags & JBD2_ABORT)
1664                 return;
1665
1666         printk(KERN_ERR "Aborting journal on device %s.\n",
1667                journal->j_devname);
1668
1669         write_lock(&journal->j_state_lock);
1670         journal->j_flags |= JBD2_ABORT;
1671         transaction = journal->j_running_transaction;
1672         if (transaction)
1673                 __jbd2_log_start_commit(journal, transaction->t_tid);
1674         write_unlock(&journal->j_state_lock);
1675 }
1676
1677 /* Soft abort: record the abort error status in the journal superblock,
1678  * but don't do any other IO. */
1679 static void __journal_abort_soft (journal_t *journal, int errno)
1680 {
1681         if (journal->j_flags & JBD2_ABORT)
1682                 return;
1683
1684         if (!journal->j_errno)
1685                 journal->j_errno = errno;
1686
1687         __jbd2_journal_abort_hard(journal);
1688
1689         if (errno)
1690                 jbd2_journal_update_superblock(journal, 1);
1691 }
1692
1693 /**
1694  * void jbd2_journal_abort () - Shutdown the journal immediately.
1695  * @journal: the journal to shutdown.
1696  * @errno:   an error number to record in the journal indicating
1697  *           the reason for the shutdown.
1698  *
1699  * Perform a complete, immediate shutdown of the ENTIRE
1700  * journal (not of a single transaction).  This operation cannot be
1701  * undone without closing and reopening the journal.
1702  *
1703  * The jbd2_journal_abort function is intended to support higher level error
1704  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1705  * mode.
1706  *
1707  * Journal abort has very specific semantics.  Any existing dirty,
1708  * unjournaled buffers in the main filesystem will still be written to
1709  * disk by bdflush, but the journaling mechanism will be suspended
1710  * immediately and no further transaction commits will be honoured.
1711  *
1712  * Any dirty, journaled buffers will be written back to disk without
1713  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1714  * filesystem, but we _do_ attempt to leave as much data as possible
1715  * behind for fsck to use for cleanup.
1716  *
1717  * Any attempt to get a new transaction handle on a journal which is in
1718  * ABORT state will just result in an -EROFS error return.  A
1719  * jbd2_journal_stop on an existing handle will return -EIO if we have
1720  * entered abort state during the update.
1721  *
1722  * Recursive transactions are not disturbed by journal abort until the
1723  * final jbd2_journal_stop, which will receive the -EIO error.
1724  *
1725  * Finally, the jbd2_journal_abort call allows the caller to supply an errno
1726  * which will be recorded (if possible) in the journal superblock.  This
1727  * allows a client to record failure conditions in the middle of a
1728  * transaction without having to complete the transaction to record the
1729  * failure to disk.  ext3_error, for example, now uses this
1730  * functionality.
1731  *
1732  * Errors which originate from within the journaling layer will NOT
1733  * supply an errno; a null errno implies that absolutely no further
1734  * writes are done to the journal (unless there are any already in
1735  * progress).
1736  *
1737  */
1738
1739 void jbd2_journal_abort(journal_t *journal, int errno)
1740 {
1741         __journal_abort_soft(journal, errno);
1742 }
1743
1744 /**
1745  * int jbd2_journal_errno () - returns the journal's error state.
1746  * @journal: journal to examine.
1747  *
1748  * This is the errno number set with jbd2_journal_abort(), the last
1749  * time the journal was mounted - if the journal was stopped
1750  * without calling abort this will be 0.
1751  *
1752  * If the journal has been aborted on this mount time -EROFS will
1753  * be returned.
1754  */
1755 int jbd2_journal_errno(journal_t *journal)
1756 {
1757         int err;
1758
1759         read_lock(&journal->j_state_lock);
1760         if (journal->j_flags & JBD2_ABORT)
1761                 err = -EROFS;
1762         else
1763                 err = journal->j_errno;
1764         read_unlock(&journal->j_state_lock);
1765         return err;
1766 }
1767
1768 /**
1769  * int jbd2_journal_clear_err () - clears the journal's error state
1770  * @journal: journal to act on.
1771  *
1772  * An error must be cleared or acked to take a FS out of readonly
1773  * mode.
1774  */
1775 int jbd2_journal_clear_err(journal_t *journal)
1776 {
1777         int err = 0;
1778
1779         write_lock(&journal->j_state_lock);
1780         if (journal->j_flags & JBD2_ABORT)
1781                 err = -EROFS;
1782         else
1783                 journal->j_errno = 0;
1784         write_unlock(&journal->j_state_lock);
1785         return err;
1786 }
1787
1788 /**
1789  * void jbd2_journal_ack_err() - Ack journal err.
1790  * @journal: journal to act on.
1791  *
1792  * An error must be cleared or acked to take a FS out of readonly
1793  * mode.
1794  */
1795 void jbd2_journal_ack_err(journal_t *journal)
1796 {
1797         write_lock(&journal->j_state_lock);
1798         if (journal->j_errno)
1799                 journal->j_flags |= JBD2_ACK_ERR;
1800         write_unlock(&journal->j_state_lock);
1801 }
1802
1803 int jbd2_journal_blocks_per_page(struct inode *inode)
1804 {
1805         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1806 }
1807
1808 /*
1809  * helper functions to deal with 32 or 64bit block numbers.
1810  */
1811 size_t journal_tag_bytes(journal_t *journal)
1812 {
1813         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
1814                 return JBD2_TAG_SIZE64;
1815         else
1816                 return JBD2_TAG_SIZE32;
1817 }
1818
1819 /*
1820  * JBD memory management
1821  *
1822  * These functions are used to allocate block-sized chunks of memory
1823  * used for making copies of buffer_head data.  Very often it will be
1824  * page-sized chunks of data, but sometimes it will be in
1825  * sub-page-size chunks.  (For example, 16k pages on Power systems
1826  * with a 4k block file system.)  For blocks smaller than a page, we
1827  * use a SLAB allocator.  There are slab caches for each block size,
1828  * which are allocated at mount time, if necessary, and we only free
1829  * (all of) the slab caches when/if the jbd2 module is unloaded.  For
1830  * this reason we don't need to a mutex to protect access to
1831  * jbd2_slab[] allocating or releasing memory; only in
1832  * jbd2_journal_create_slab().
1833  */
1834 #define JBD2_MAX_SLABS 8
1835 static struct kmem_cache *jbd2_slab[JBD2_MAX_SLABS];
1836 static DECLARE_MUTEX(jbd2_slab_create_sem);
1837
1838 static const char *jbd2_slab_names[JBD2_MAX_SLABS] = {
1839         "jbd2_1k", "jbd2_2k", "jbd2_4k", "jbd2_8k",
1840         "jbd2_16k", "jbd2_32k", "jbd2_64k", "jbd2_128k"
1841 };
1842
1843
1844 static void jbd2_journal_destroy_slabs(void)
1845 {
1846         int i;
1847
1848         for (i = 0; i < JBD2_MAX_SLABS; i++) {
1849                 if (jbd2_slab[i])
1850                         kmem_cache_destroy(jbd2_slab[i]);
1851                 jbd2_slab[i] = NULL;
1852         }
1853 }
1854
1855 static int jbd2_journal_create_slab(size_t size)
1856 {
1857         int i = order_base_2(size) - 10;
1858         size_t slab_size;
1859
1860         if (size == PAGE_SIZE)
1861                 return 0;
1862
1863         if (i >= JBD2_MAX_SLABS)
1864                 return -EINVAL;
1865
1866         if (unlikely(i < 0))
1867                 i = 0;
1868         down(&jbd2_slab_create_sem);
1869         if (jbd2_slab[i]) {
1870                 up(&jbd2_slab_create_sem);
1871                 return 0;       /* Already created */
1872         }
1873
1874         slab_size = 1 << (i+10);
1875         jbd2_slab[i] = kmem_cache_create(jbd2_slab_names[i], slab_size,
1876                                          slab_size, 0, NULL);
1877         up(&jbd2_slab_create_sem);
1878         if (!jbd2_slab[i]) {
1879                 printk(KERN_EMERG "JBD2: no memory for jbd2_slab cache\n");
1880                 return -ENOMEM;
1881         }
1882         return 0;
1883 }
1884
1885 static struct kmem_cache *get_slab(size_t size)
1886 {
1887         int i = order_base_2(size) - 10;
1888
1889         BUG_ON(i >= JBD2_MAX_SLABS);
1890         if (unlikely(i < 0))
1891                 i = 0;
1892         BUG_ON(jbd2_slab[i] == NULL);
1893         return jbd2_slab[i];
1894 }
1895
1896 void *jbd2_alloc(size_t size, gfp_t flags)
1897 {
1898         void *ptr;
1899
1900         BUG_ON(size & (size-1)); /* Must be a power of 2 */
1901
1902         flags |= __GFP_REPEAT;
1903         if (size == PAGE_SIZE)
1904                 ptr = (void *)__get_free_pages(flags, 0);
1905         else if (size > PAGE_SIZE) {
1906                 int order = get_order(size);
1907
1908                 if (order < 3)
1909                         ptr = (void *)__get_free_pages(flags, order);
1910                 else
1911                         ptr = vmalloc(size);
1912         } else
1913                 ptr = kmem_cache_alloc(get_slab(size), flags);
1914
1915         /* Check alignment; SLUB has gotten this wrong in the past,
1916          * and this can lead to user data corruption! */
1917         BUG_ON(((unsigned long) ptr) & (size-1));
1918
1919         return ptr;
1920 }
1921
1922 void jbd2_free(void *ptr, size_t size)
1923 {
1924         if (size == PAGE_SIZE) {
1925                 free_pages((unsigned long)ptr, 0);
1926                 return;
1927         }
1928         if (size > PAGE_SIZE) {
1929                 int order = get_order(size);
1930
1931                 if (order < 3)
1932                         free_pages((unsigned long)ptr, order);
1933                 else
1934                         vfree(ptr);
1935                 return;
1936         }
1937         kmem_cache_free(get_slab(size), ptr);
1938 };
1939
1940 /*
1941  * Journal_head storage management
1942  */
1943 static struct kmem_cache *jbd2_journal_head_cache;
1944 #ifdef CONFIG_JBD2_DEBUG
1945 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
1946 #endif
1947
1948 static int journal_init_jbd2_journal_head_cache(void)
1949 {
1950         int retval;
1951
1952         J_ASSERT(jbd2_journal_head_cache == NULL);
1953         jbd2_journal_head_cache = kmem_cache_create("jbd2_journal_head",
1954                                 sizeof(struct journal_head),
1955                                 0,              /* offset */
1956                                 SLAB_TEMPORARY, /* flags */
1957                                 NULL);          /* ctor */
1958         retval = 0;
1959         if (!jbd2_journal_head_cache) {
1960                 retval = -ENOMEM;
1961                 printk(KERN_EMERG "JBD: no memory for journal_head cache\n");
1962         }
1963         return retval;
1964 }
1965
1966 static void jbd2_journal_destroy_jbd2_journal_head_cache(void)
1967 {
1968         if (jbd2_journal_head_cache) {
1969                 kmem_cache_destroy(jbd2_journal_head_cache);
1970                 jbd2_journal_head_cache = NULL;
1971         }
1972 }
1973
1974 /*
1975  * journal_head splicing and dicing
1976  */
1977 static struct journal_head *journal_alloc_journal_head(void)
1978 {
1979         struct journal_head *ret;
1980         static unsigned long last_warning;
1981
1982 #ifdef CONFIG_JBD2_DEBUG
1983         atomic_inc(&nr_journal_heads);
1984 #endif
1985         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
1986         if (!ret) {
1987                 jbd_debug(1, "out of memory for journal_head\n");
1988                 if (time_after(jiffies, last_warning + 5*HZ)) {
1989                         printk(KERN_NOTICE "ENOMEM in %s, retrying.\n",
1990                                __func__);
1991                         last_warning = jiffies;
1992                 }
1993                 while (!ret) {
1994                         yield();
1995                         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
1996                 }
1997         }
1998         return ret;
1999 }
2000
2001 static void journal_free_journal_head(struct journal_head *jh)
2002 {
2003 #ifdef CONFIG_JBD2_DEBUG
2004         atomic_dec(&nr_journal_heads);
2005         memset(jh, JBD2_POISON_FREE, sizeof(*jh));
2006 #endif
2007         kmem_cache_free(jbd2_journal_head_cache, jh);
2008 }
2009
2010 /*
2011  * A journal_head is attached to a buffer_head whenever JBD has an
2012  * interest in the buffer.
2013  *
2014  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
2015  * is set.  This bit is tested in core kernel code where we need to take
2016  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
2017  * there.
2018  *
2019  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
2020  *
2021  * When a buffer has its BH_JBD bit set it is immune from being released by
2022  * core kernel code, mainly via ->b_count.
2023  *
2024  * A journal_head may be detached from its buffer_head when the journal_head's
2025  * b_transaction, b_cp_transaction and b_next_transaction pointers are NULL.
2026  * Various places in JBD call jbd2_journal_remove_journal_head() to indicate that the
2027  * journal_head can be dropped if needed.
2028  *
2029  * Various places in the kernel want to attach a journal_head to a buffer_head
2030  * _before_ attaching the journal_head to a transaction.  To protect the
2031  * journal_head in this situation, jbd2_journal_add_journal_head elevates the
2032  * journal_head's b_jcount refcount by one.  The caller must call
2033  * jbd2_journal_put_journal_head() to undo this.
2034  *
2035  * So the typical usage would be:
2036  *
2037  *      (Attach a journal_head if needed.  Increments b_jcount)
2038  *      struct journal_head *jh = jbd2_journal_add_journal_head(bh);
2039  *      ...
2040  *      jh->b_transaction = xxx;
2041  *      jbd2_journal_put_journal_head(jh);
2042  *
2043  * Now, the journal_head's b_jcount is zero, but it is safe from being released
2044  * because it has a non-zero b_transaction.
2045  */
2046
2047 /*
2048  * Give a buffer_head a journal_head.
2049  *
2050  * Doesn't need the journal lock.
2051  * May sleep.
2052  */
2053 struct journal_head *jbd2_journal_add_journal_head(struct buffer_head *bh)
2054 {
2055         struct journal_head *jh;
2056         struct journal_head *new_jh = NULL;
2057
2058 repeat:
2059         if (!buffer_jbd(bh)) {
2060                 new_jh = journal_alloc_journal_head();
2061                 memset(new_jh, 0, sizeof(*new_jh));
2062         }
2063
2064         jbd_lock_bh_journal_head(bh);
2065         if (buffer_jbd(bh)) {
2066                 jh = bh2jh(bh);
2067         } else {
2068                 J_ASSERT_BH(bh,
2069                         (atomic_read(&bh->b_count) > 0) ||
2070                         (bh->b_page && bh->b_page->mapping));
2071
2072                 if (!new_jh) {
2073                         jbd_unlock_bh_journal_head(bh);
2074                         goto repeat;
2075                 }
2076
2077                 jh = new_jh;
2078                 new_jh = NULL;          /* We consumed it */
2079                 set_buffer_jbd(bh);
2080                 bh->b_private = jh;
2081                 jh->b_bh = bh;
2082                 get_bh(bh);
2083                 BUFFER_TRACE(bh, "added journal_head");
2084         }
2085         jh->b_jcount++;
2086         jbd_unlock_bh_journal_head(bh);
2087         if (new_jh)
2088                 journal_free_journal_head(new_jh);
2089         return bh->b_private;
2090 }
2091
2092 /*
2093  * Grab a ref against this buffer_head's journal_head.  If it ended up not
2094  * having a journal_head, return NULL
2095  */
2096 struct journal_head *jbd2_journal_grab_journal_head(struct buffer_head *bh)
2097 {
2098         struct journal_head *jh = NULL;
2099
2100         jbd_lock_bh_journal_head(bh);
2101         if (buffer_jbd(bh)) {
2102                 jh = bh2jh(bh);
2103                 jh->b_jcount++;
2104         }
2105         jbd_unlock_bh_journal_head(bh);
2106         return jh;
2107 }
2108
2109 static void __journal_remove_journal_head(struct buffer_head *bh)
2110 {
2111         struct journal_head *jh = bh2jh(bh);
2112
2113         J_ASSERT_JH(jh, jh->b_jcount >= 0);
2114
2115         get_bh(bh);
2116         if (jh->b_jcount == 0) {
2117                 if (jh->b_transaction == NULL &&
2118                                 jh->b_next_transaction == NULL &&
2119                                 jh->b_cp_transaction == NULL) {
2120                         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
2121                         J_ASSERT_BH(bh, buffer_jbd(bh));
2122                         J_ASSERT_BH(bh, jh2bh(jh) == bh);
2123                         BUFFER_TRACE(bh, "remove journal_head");
2124                         if (jh->b_frozen_data) {
2125                                 printk(KERN_WARNING "%s: freeing "
2126                                                 "b_frozen_data\n",
2127                                                 __func__);
2128                                 jbd2_free(jh->b_frozen_data, bh->b_size);
2129                         }
2130                         if (jh->b_committed_data) {
2131                                 printk(KERN_WARNING "%s: freeing "
2132                                                 "b_committed_data\n",
2133                                                 __func__);
2134                                 jbd2_free(jh->b_committed_data, bh->b_size);
2135                         }
2136                         bh->b_private = NULL;
2137                         jh->b_bh = NULL;        /* debug, really */
2138                         clear_buffer_jbd(bh);
2139                         __brelse(bh);
2140                         journal_free_journal_head(jh);
2141                 } else {
2142                         BUFFER_TRACE(bh, "journal_head was locked");
2143                 }
2144         }
2145 }
2146
2147 /*
2148  * jbd2_journal_remove_journal_head(): if the buffer isn't attached to a transaction
2149  * and has a zero b_jcount then remove and release its journal_head.   If we did
2150  * see that the buffer is not used by any transaction we also "logically"
2151  * decrement ->b_count.
2152  *
2153  * We in fact take an additional increment on ->b_count as a convenience,
2154  * because the caller usually wants to do additional things with the bh
2155  * after calling here.
2156  * The caller of jbd2_journal_remove_journal_head() *must* run __brelse(bh) at some
2157  * time.  Once the caller has run __brelse(), the buffer is eligible for
2158  * reaping by try_to_free_buffers().
2159  */
2160 void jbd2_journal_remove_journal_head(struct buffer_head *bh)
2161 {
2162         jbd_lock_bh_journal_head(bh);
2163         __journal_remove_journal_head(bh);
2164         jbd_unlock_bh_journal_head(bh);
2165 }
2166
2167 /*
2168  * Drop a reference on the passed journal_head.  If it fell to zero then try to
2169  * release the journal_head from the buffer_head.
2170  */
2171 void jbd2_journal_put_journal_head(struct journal_head *jh)
2172 {
2173         struct buffer_head *bh = jh2bh(jh);
2174
2175         jbd_lock_bh_journal_head(bh);
2176         J_ASSERT_JH(jh, jh->b_jcount > 0);
2177         --jh->b_jcount;
2178         if (!jh->b_jcount && !jh->b_transaction) {
2179                 __journal_remove_journal_head(bh);
2180                 __brelse(bh);
2181         }
2182         jbd_unlock_bh_journal_head(bh);
2183 }
2184
2185 /*
2186  * Initialize jbd inode head
2187  */
2188 void jbd2_journal_init_jbd_inode(struct jbd2_inode *jinode, struct inode *inode)
2189 {
2190         jinode->i_transaction = NULL;
2191         jinode->i_next_transaction = NULL;
2192         jinode->i_vfs_inode = inode;
2193         jinode->i_flags = 0;
2194         INIT_LIST_HEAD(&jinode->i_list);
2195 }
2196
2197 /*
2198  * Function to be called before we start removing inode from memory (i.e.,
2199  * clear_inode() is a fine place to be called from). It removes inode from
2200  * transaction's lists.
2201  */
2202 void jbd2_journal_release_jbd_inode(journal_t *journal,
2203                                     struct jbd2_inode *jinode)
2204 {
2205         if (!journal)
2206                 return;
2207 restart:
2208         spin_lock(&journal->j_list_lock);
2209         /* Is commit writing out inode - we have to wait */
2210         if (jinode->i_flags & JI_COMMIT_RUNNING) {
2211                 wait_queue_head_t *wq;
2212                 DEFINE_WAIT_BIT(wait, &jinode->i_flags, __JI_COMMIT_RUNNING);
2213                 wq = bit_waitqueue(&jinode->i_flags, __JI_COMMIT_RUNNING);
2214                 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2215                 spin_unlock(&journal->j_list_lock);
2216                 schedule();
2217                 finish_wait(wq, &wait.wait);
2218                 goto restart;
2219         }
2220
2221         if (jinode->i_transaction) {
2222                 list_del(&jinode->i_list);
2223                 jinode->i_transaction = NULL;
2224         }
2225         spin_unlock(&journal->j_list_lock);
2226 }
2227
2228 /*
2229  * debugfs tunables
2230  */
2231 #ifdef CONFIG_JBD2_DEBUG
2232 u8 jbd2_journal_enable_debug __read_mostly;
2233 EXPORT_SYMBOL(jbd2_journal_enable_debug);
2234
2235 #define JBD2_DEBUG_NAME "jbd2-debug"
2236
2237 static struct dentry *jbd2_debugfs_dir;
2238 static struct dentry *jbd2_debug;
2239
2240 static void __init jbd2_create_debugfs_entry(void)
2241 {
2242         jbd2_debugfs_dir = debugfs_create_dir("jbd2", NULL);
2243         if (jbd2_debugfs_dir)
2244                 jbd2_debug = debugfs_create_u8(JBD2_DEBUG_NAME,
2245                                                S_IRUGO | S_IWUSR,
2246                                                jbd2_debugfs_dir,
2247                                                &jbd2_journal_enable_debug);
2248 }
2249
2250 static void __exit jbd2_remove_debugfs_entry(void)
2251 {
2252         debugfs_remove(jbd2_debug);
2253         debugfs_remove(jbd2_debugfs_dir);
2254 }
2255
2256 #else
2257
2258 static void __init jbd2_create_debugfs_entry(void)
2259 {
2260 }
2261
2262 static void __exit jbd2_remove_debugfs_entry(void)
2263 {
2264 }
2265
2266 #endif
2267
2268 #ifdef CONFIG_PROC_FS
2269
2270 #define JBD2_STATS_PROC_NAME "fs/jbd2"
2271
2272 static void __init jbd2_create_jbd_stats_proc_entry(void)
2273 {
2274         proc_jbd2_stats = proc_mkdir(JBD2_STATS_PROC_NAME, NULL);
2275 }
2276
2277 static void __exit jbd2_remove_jbd_stats_proc_entry(void)
2278 {
2279         if (proc_jbd2_stats)
2280                 remove_proc_entry(JBD2_STATS_PROC_NAME, NULL);
2281 }
2282
2283 #else
2284
2285 #define jbd2_create_jbd_stats_proc_entry() do {} while (0)
2286 #define jbd2_remove_jbd_stats_proc_entry() do {} while (0)
2287
2288 #endif
2289
2290 struct kmem_cache *jbd2_handle_cache;
2291
2292 static int __init journal_init_handle_cache(void)
2293 {
2294         jbd2_handle_cache = kmem_cache_create("jbd2_journal_handle",
2295                                 sizeof(handle_t),
2296                                 0,              /* offset */
2297                                 SLAB_TEMPORARY, /* flags */
2298                                 NULL);          /* ctor */
2299         if (jbd2_handle_cache == NULL) {
2300                 printk(KERN_EMERG "JBD: failed to create handle cache\n");
2301                 return -ENOMEM;
2302         }
2303         return 0;
2304 }
2305
2306 static void jbd2_journal_destroy_handle_cache(void)
2307 {
2308         if (jbd2_handle_cache)
2309                 kmem_cache_destroy(jbd2_handle_cache);
2310 }
2311
2312 /*
2313  * Module startup and shutdown
2314  */
2315
2316 static int __init journal_init_caches(void)
2317 {
2318         int ret;
2319
2320         ret = jbd2_journal_init_revoke_caches();
2321         if (ret == 0)
2322                 ret = journal_init_jbd2_journal_head_cache();
2323         if (ret == 0)
2324                 ret = journal_init_handle_cache();
2325         return ret;
2326 }
2327
2328 static void jbd2_journal_destroy_caches(void)
2329 {
2330         jbd2_journal_destroy_revoke_caches();
2331         jbd2_journal_destroy_jbd2_journal_head_cache();
2332         jbd2_journal_destroy_handle_cache();
2333         jbd2_journal_destroy_slabs();
2334 }
2335
2336 static int __init journal_init(void)
2337 {
2338         int ret;
2339
2340         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
2341
2342         ret = journal_init_caches();
2343         if (ret == 0) {
2344                 jbd2_create_debugfs_entry();
2345                 jbd2_create_jbd_stats_proc_entry();
2346         } else {
2347                 jbd2_journal_destroy_caches();
2348         }
2349         return ret;
2350 }
2351
2352 static void __exit journal_exit(void)
2353 {
2354 #ifdef CONFIG_JBD2_DEBUG
2355         int n = atomic_read(&nr_journal_heads);
2356         if (n)
2357                 printk(KERN_EMERG "JBD: leaked %d journal_heads!\n", n);
2358 #endif
2359         jbd2_remove_debugfs_entry();
2360         jbd2_remove_jbd_stats_proc_entry();
2361         jbd2_journal_destroy_caches();
2362 }
2363
2364 /* 
2365  * jbd2_dev_to_name is a utility function used by the jbd2 and ext4 
2366  * tracing infrastructure to map a dev_t to a device name.
2367  *
2368  * The caller should use rcu_read_lock() in order to make sure the
2369  * device name stays valid until its done with it.  We use
2370  * rcu_read_lock() as well to make sure we're safe in case the caller
2371  * gets sloppy, and because rcu_read_lock() is cheap and can be safely
2372  * nested.
2373  */
2374 struct devname_cache {
2375         struct rcu_head rcu;
2376         dev_t           device;
2377         char            devname[BDEVNAME_SIZE];
2378 };
2379 #define CACHE_SIZE_BITS 6
2380 static struct devname_cache *devcache[1 << CACHE_SIZE_BITS];
2381 static DEFINE_SPINLOCK(devname_cache_lock);
2382
2383 static void free_devcache(struct rcu_head *rcu)
2384 {
2385         kfree(rcu);
2386 }
2387
2388 const char *jbd2_dev_to_name(dev_t device)
2389 {
2390         int     i = hash_32(device, CACHE_SIZE_BITS);
2391         char    *ret;
2392         struct block_device *bd;
2393         static struct devname_cache *new_dev;
2394
2395         rcu_read_lock();
2396         if (devcache[i] && devcache[i]->device == device) {
2397                 ret = devcache[i]->devname;
2398                 rcu_read_unlock();
2399                 return ret;
2400         }
2401         rcu_read_unlock();
2402
2403         new_dev = kmalloc(sizeof(struct devname_cache), GFP_KERNEL);
2404         if (!new_dev)
2405                 return "NODEV-ALLOCFAILURE"; /* Something non-NULL */
2406         spin_lock(&devname_cache_lock);
2407         if (devcache[i]) {
2408                 if (devcache[i]->device == device) {
2409                         kfree(new_dev);
2410                         ret = devcache[i]->devname;
2411                         spin_unlock(&devname_cache_lock);
2412                         return ret;
2413                 }
2414                 call_rcu(&devcache[i]->rcu, free_devcache);
2415         }
2416         devcache[i] = new_dev;
2417         devcache[i]->device = device;
2418         bd = bdget(device);
2419         if (bd) {
2420                 bdevname(bd, devcache[i]->devname);
2421                 bdput(bd);
2422         } else
2423                 __bdevname(device, devcache[i]->devname);
2424         ret = devcache[i]->devname;
2425         spin_unlock(&devname_cache_lock);
2426         return ret;
2427 }
2428 EXPORT_SYMBOL(jbd2_dev_to_name);
2429
2430 MODULE_LICENSE("GPL");
2431 module_init(journal_init);
2432 module_exit(journal_exit);
2433