]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - fs/jbd2/journal.c
839377e3d6244ac540e0a573bf9c4adc679225a8
[~shefty/rdma-dev.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 #include <linux/ratelimit.h>
47
48 #define CREATE_TRACE_POINTS
49 #include <trace/events/jbd2.h>
50
51 #include <asm/uaccess.h>
52 #include <asm/page.h>
53 #include <asm/system.h>
54
55 EXPORT_SYMBOL(jbd2_journal_extend);
56 EXPORT_SYMBOL(jbd2_journal_stop);
57 EXPORT_SYMBOL(jbd2_journal_lock_updates);
58 EXPORT_SYMBOL(jbd2_journal_unlock_updates);
59 EXPORT_SYMBOL(jbd2_journal_get_write_access);
60 EXPORT_SYMBOL(jbd2_journal_get_create_access);
61 EXPORT_SYMBOL(jbd2_journal_get_undo_access);
62 EXPORT_SYMBOL(jbd2_journal_set_triggers);
63 EXPORT_SYMBOL(jbd2_journal_dirty_metadata);
64 EXPORT_SYMBOL(jbd2_journal_release_buffer);
65 EXPORT_SYMBOL(jbd2_journal_forget);
66 #if 0
67 EXPORT_SYMBOL(journal_sync_buffer);
68 #endif
69 EXPORT_SYMBOL(jbd2_journal_flush);
70 EXPORT_SYMBOL(jbd2_journal_revoke);
71
72 EXPORT_SYMBOL(jbd2_journal_init_dev);
73 EXPORT_SYMBOL(jbd2_journal_init_inode);
74 EXPORT_SYMBOL(jbd2_journal_update_format);
75 EXPORT_SYMBOL(jbd2_journal_check_used_features);
76 EXPORT_SYMBOL(jbd2_journal_check_available_features);
77 EXPORT_SYMBOL(jbd2_journal_set_features);
78 EXPORT_SYMBOL(jbd2_journal_load);
79 EXPORT_SYMBOL(jbd2_journal_destroy);
80 EXPORT_SYMBOL(jbd2_journal_abort);
81 EXPORT_SYMBOL(jbd2_journal_errno);
82 EXPORT_SYMBOL(jbd2_journal_ack_err);
83 EXPORT_SYMBOL(jbd2_journal_clear_err);
84 EXPORT_SYMBOL(jbd2_log_wait_commit);
85 EXPORT_SYMBOL(jbd2_log_start_commit);
86 EXPORT_SYMBOL(jbd2_journal_start_commit);
87 EXPORT_SYMBOL(jbd2_journal_force_commit_nested);
88 EXPORT_SYMBOL(jbd2_journal_wipe);
89 EXPORT_SYMBOL(jbd2_journal_blocks_per_page);
90 EXPORT_SYMBOL(jbd2_journal_invalidatepage);
91 EXPORT_SYMBOL(jbd2_journal_try_to_free_buffers);
92 EXPORT_SYMBOL(jbd2_journal_force_commit);
93 EXPORT_SYMBOL(jbd2_journal_file_inode);
94 EXPORT_SYMBOL(jbd2_journal_init_jbd_inode);
95 EXPORT_SYMBOL(jbd2_journal_release_jbd_inode);
96 EXPORT_SYMBOL(jbd2_journal_begin_ordered_truncate);
97 EXPORT_SYMBOL(jbd2_inode_cache);
98
99 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
100 static void __journal_abort_soft (journal_t *journal, int errno);
101 static int jbd2_journal_create_slab(size_t slab_size);
102
103 /*
104  * Helper function used to manage commit timeouts
105  */
106
107 static void commit_timeout(unsigned long __data)
108 {
109         struct task_struct * p = (struct task_struct *) __data;
110
111         wake_up_process(p);
112 }
113
114 /*
115  * kjournald2: The main thread function used to manage a logging device
116  * journal.
117  *
118  * This kernel thread is responsible for two things:
119  *
120  * 1) COMMIT:  Every so often we need to commit the current state of the
121  *    filesystem to disk.  The journal thread is responsible for writing
122  *    all of the metadata buffers to disk.
123  *
124  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
125  *    of the data in that part of the log has been rewritten elsewhere on
126  *    the disk.  Flushing these old buffers to reclaim space in the log is
127  *    known as checkpointing, and this thread is responsible for that job.
128  */
129
130 static int kjournald2(void *arg)
131 {
132         journal_t *journal = arg;
133         transaction_t *transaction;
134
135         /*
136          * Set up an interval timer which can be used to trigger a commit wakeup
137          * after the commit interval expires
138          */
139         setup_timer(&journal->j_commit_timer, commit_timeout,
140                         (unsigned long)current);
141
142         set_freezable();
143
144         /* Record that the journal thread is running */
145         journal->j_task = current;
146         wake_up(&journal->j_wait_done_commit);
147
148         /*
149          * And now, wait forever for commit wakeup events.
150          */
151         write_lock(&journal->j_state_lock);
152
153 loop:
154         if (journal->j_flags & JBD2_UNMOUNT)
155                 goto end_loop;
156
157         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
158                 journal->j_commit_sequence, journal->j_commit_request);
159
160         if (journal->j_commit_sequence != journal->j_commit_request) {
161                 jbd_debug(1, "OK, requests differ\n");
162                 write_unlock(&journal->j_state_lock);
163                 del_timer_sync(&journal->j_commit_timer);
164                 jbd2_journal_commit_transaction(journal);
165                 write_lock(&journal->j_state_lock);
166                 goto loop;
167         }
168
169         wake_up(&journal->j_wait_done_commit);
170         if (freezing(current)) {
171                 /*
172                  * The simpler the better. Flushing journal isn't a
173                  * good idea, because that depends on threads that may
174                  * be already stopped.
175                  */
176                 jbd_debug(1, "Now suspending kjournald2\n");
177                 write_unlock(&journal->j_state_lock);
178                 try_to_freeze();
179                 write_lock(&journal->j_state_lock);
180         } else {
181                 /*
182                  * We assume on resume that commits are already there,
183                  * so we don't sleep
184                  */
185                 DEFINE_WAIT(wait);
186                 int should_sleep = 1;
187
188                 prepare_to_wait(&journal->j_wait_commit, &wait,
189                                 TASK_INTERRUPTIBLE);
190                 if (journal->j_commit_sequence != journal->j_commit_request)
191                         should_sleep = 0;
192                 transaction = journal->j_running_transaction;
193                 if (transaction && time_after_eq(jiffies,
194                                                 transaction->t_expires))
195                         should_sleep = 0;
196                 if (journal->j_flags & JBD2_UNMOUNT)
197                         should_sleep = 0;
198                 if (should_sleep) {
199                         write_unlock(&journal->j_state_lock);
200                         schedule();
201                         write_lock(&journal->j_state_lock);
202                 }
203                 finish_wait(&journal->j_wait_commit, &wait);
204         }
205
206         jbd_debug(1, "kjournald2 wakes\n");
207
208         /*
209          * Were we woken up by a commit wakeup event?
210          */
211         transaction = journal->j_running_transaction;
212         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
213                 journal->j_commit_request = transaction->t_tid;
214                 jbd_debug(1, "woke because of timeout\n");
215         }
216         goto loop;
217
218 end_loop:
219         write_unlock(&journal->j_state_lock);
220         del_timer_sync(&journal->j_commit_timer);
221         journal->j_task = NULL;
222         wake_up(&journal->j_wait_done_commit);
223         jbd_debug(1, "Journal thread exiting.\n");
224         return 0;
225 }
226
227 static int jbd2_journal_start_thread(journal_t *journal)
228 {
229         struct task_struct *t;
230
231         t = kthread_run(kjournald2, journal, "jbd2/%s",
232                         journal->j_devname);
233         if (IS_ERR(t))
234                 return PTR_ERR(t);
235
236         wait_event(journal->j_wait_done_commit, journal->j_task != NULL);
237         return 0;
238 }
239
240 static void journal_kill_thread(journal_t *journal)
241 {
242         write_lock(&journal->j_state_lock);
243         journal->j_flags |= JBD2_UNMOUNT;
244
245         while (journal->j_task) {
246                 wake_up(&journal->j_wait_commit);
247                 write_unlock(&journal->j_state_lock);
248                 wait_event(journal->j_wait_done_commit, journal->j_task == NULL);
249                 write_lock(&journal->j_state_lock);
250         }
251         write_unlock(&journal->j_state_lock);
252 }
253
254 /*
255  * jbd2_journal_write_metadata_buffer: write a metadata buffer to the journal.
256  *
257  * Writes a metadata buffer to a given disk block.  The actual IO is not
258  * performed but a new buffer_head is constructed which labels the data
259  * to be written with the correct destination disk block.
260  *
261  * Any magic-number escaping which needs to be done will cause a
262  * copy-out here.  If the buffer happens to start with the
263  * JBD2_MAGIC_NUMBER, then we can't write it to the log directly: the
264  * magic number is only written to the log for descripter blocks.  In
265  * this case, we copy the data and replace the first word with 0, and we
266  * return a result code which indicates that this buffer needs to be
267  * marked as an escaped buffer in the corresponding log descriptor
268  * block.  The missing word can then be restored when the block is read
269  * during recovery.
270  *
271  * If the source buffer has already been modified by a new transaction
272  * since we took the last commit snapshot, we use the frozen copy of
273  * that data for IO.  If we end up using the existing buffer_head's data
274  * for the write, then we *have* to lock the buffer to prevent anyone
275  * else from using and possibly modifying it while the IO is in
276  * progress.
277  *
278  * The function returns a pointer to the buffer_heads to be used for IO.
279  *
280  * We assume that the journal has already been locked in this function.
281  *
282  * Return value:
283  *  <0: Error
284  * >=0: Finished OK
285  *
286  * On success:
287  * Bit 0 set == escape performed on the data
288  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
289  */
290
291 int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
292                                   struct journal_head  *jh_in,
293                                   struct journal_head **jh_out,
294                                   unsigned long long blocknr)
295 {
296         int need_copy_out = 0;
297         int done_copy_out = 0;
298         int do_escape = 0;
299         char *mapped_data;
300         struct buffer_head *new_bh;
301         struct journal_head *new_jh;
302         struct page *new_page;
303         unsigned int new_offset;
304         struct buffer_head *bh_in = jh2bh(jh_in);
305         journal_t *journal = transaction->t_journal;
306
307         /*
308          * The buffer really shouldn't be locked: only the current committing
309          * transaction is allowed to write it, so nobody else is allowed
310          * to do any IO.
311          *
312          * akpm: except if we're journalling data, and write() output is
313          * also part of a shared mapping, and another thread has
314          * decided to launch a writepage() against this buffer.
315          */
316         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
317
318 retry_alloc:
319         new_bh = alloc_buffer_head(GFP_NOFS);
320         if (!new_bh) {
321                 /*
322                  * Failure is not an option, but __GFP_NOFAIL is going
323                  * away; so we retry ourselves here.
324                  */
325                 congestion_wait(BLK_RW_ASYNC, HZ/50);
326                 goto retry_alloc;
327         }
328
329         /* keep subsequent assertions sane */
330         new_bh->b_state = 0;
331         init_buffer(new_bh, NULL, NULL);
332         atomic_set(&new_bh->b_count, 1);
333         new_jh = jbd2_journal_add_journal_head(new_bh); /* This sleeps */
334
335         /*
336          * If a new transaction has already done a buffer copy-out, then
337          * we use that version of the data for the commit.
338          */
339         jbd_lock_bh_state(bh_in);
340 repeat:
341         if (jh_in->b_frozen_data) {
342                 done_copy_out = 1;
343                 new_page = virt_to_page(jh_in->b_frozen_data);
344                 new_offset = offset_in_page(jh_in->b_frozen_data);
345         } else {
346                 new_page = jh2bh(jh_in)->b_page;
347                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
348         }
349
350         mapped_data = kmap_atomic(new_page);
351         /*
352          * Fire data frozen trigger if data already wasn't frozen.  Do this
353          * before checking for escaping, as the trigger may modify the magic
354          * offset.  If a copy-out happens afterwards, it will have the correct
355          * data in the buffer.
356          */
357         if (!done_copy_out)
358                 jbd2_buffer_frozen_trigger(jh_in, mapped_data + new_offset,
359                                            jh_in->b_triggers);
360
361         /*
362          * Check for escaping
363          */
364         if (*((__be32 *)(mapped_data + new_offset)) ==
365                                 cpu_to_be32(JBD2_MAGIC_NUMBER)) {
366                 need_copy_out = 1;
367                 do_escape = 1;
368         }
369         kunmap_atomic(mapped_data);
370
371         /*
372          * Do we need to do a data copy?
373          */
374         if (need_copy_out && !done_copy_out) {
375                 char *tmp;
376
377                 jbd_unlock_bh_state(bh_in);
378                 tmp = jbd2_alloc(bh_in->b_size, GFP_NOFS);
379                 if (!tmp) {
380                         jbd2_journal_put_journal_head(new_jh);
381                         return -ENOMEM;
382                 }
383                 jbd_lock_bh_state(bh_in);
384                 if (jh_in->b_frozen_data) {
385                         jbd2_free(tmp, bh_in->b_size);
386                         goto repeat;
387                 }
388
389                 jh_in->b_frozen_data = tmp;
390                 mapped_data = kmap_atomic(new_page);
391                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
392                 kunmap_atomic(mapped_data);
393
394                 new_page = virt_to_page(tmp);
395                 new_offset = offset_in_page(tmp);
396                 done_copy_out = 1;
397
398                 /*
399                  * This isn't strictly necessary, as we're using frozen
400                  * data for the escaping, but it keeps consistency with
401                  * b_frozen_data usage.
402                  */
403                 jh_in->b_frozen_triggers = jh_in->b_triggers;
404         }
405
406         /*
407          * Did we need to do an escaping?  Now we've done all the
408          * copying, we can finally do so.
409          */
410         if (do_escape) {
411                 mapped_data = kmap_atomic(new_page);
412                 *((unsigned int *)(mapped_data + new_offset)) = 0;
413                 kunmap_atomic(mapped_data);
414         }
415
416         set_bh_page(new_bh, new_page, new_offset);
417         new_jh->b_transaction = NULL;
418         new_bh->b_size = jh2bh(jh_in)->b_size;
419         new_bh->b_bdev = transaction->t_journal->j_dev;
420         new_bh->b_blocknr = blocknr;
421         set_buffer_mapped(new_bh);
422         set_buffer_dirty(new_bh);
423
424         *jh_out = new_jh;
425
426         /*
427          * The to-be-written buffer needs to get moved to the io queue,
428          * and the original buffer whose contents we are shadowing or
429          * copying is moved to the transaction's shadow queue.
430          */
431         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
432         spin_lock(&journal->j_list_lock);
433         __jbd2_journal_file_buffer(jh_in, transaction, BJ_Shadow);
434         spin_unlock(&journal->j_list_lock);
435         jbd_unlock_bh_state(bh_in);
436
437         JBUFFER_TRACE(new_jh, "file as BJ_IO");
438         jbd2_journal_file_buffer(new_jh, transaction, BJ_IO);
439
440         return do_escape | (done_copy_out << 1);
441 }
442
443 /*
444  * Allocation code for the journal file.  Manage the space left in the
445  * journal, so that we can begin checkpointing when appropriate.
446  */
447
448 /*
449  * __jbd2_log_space_left: Return the number of free blocks left in the journal.
450  *
451  * Called with the journal already locked.
452  *
453  * Called under j_state_lock
454  */
455
456 int __jbd2_log_space_left(journal_t *journal)
457 {
458         int left = journal->j_free;
459
460         /* assert_spin_locked(&journal->j_state_lock); */
461
462         /*
463          * Be pessimistic here about the number of those free blocks which
464          * might be required for log descriptor control blocks.
465          */
466
467 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
468
469         left -= MIN_LOG_RESERVED_BLOCKS;
470
471         if (left <= 0)
472                 return 0;
473         left -= (left >> 3);
474         return left;
475 }
476
477 /*
478  * Called with j_state_lock locked for writing.
479  * Returns true if a transaction commit was started.
480  */
481 int __jbd2_log_start_commit(journal_t *journal, tid_t target)
482 {
483         /*
484          * The only transaction we can possibly wait upon is the
485          * currently running transaction (if it exists).  Otherwise,
486          * the target tid must be an old one.
487          */
488         if (journal->j_running_transaction &&
489             journal->j_running_transaction->t_tid == target) {
490                 /*
491                  * We want a new commit: OK, mark the request and wakeup the
492                  * commit thread.  We do _not_ do the commit ourselves.
493                  */
494
495                 journal->j_commit_request = target;
496                 jbd_debug(1, "JBD2: requesting commit %d/%d\n",
497                           journal->j_commit_request,
498                           journal->j_commit_sequence);
499                 wake_up(&journal->j_wait_commit);
500                 return 1;
501         } else if (!tid_geq(journal->j_commit_request, target))
502                 /* This should never happen, but if it does, preserve
503                    the evidence before kjournald goes into a loop and
504                    increments j_commit_sequence beyond all recognition. */
505                 WARN_ONCE(1, "JBD2: bad log_start_commit: %u %u %u %u\n",
506                           journal->j_commit_request,
507                           journal->j_commit_sequence,
508                           target, journal->j_running_transaction ? 
509                           journal->j_running_transaction->t_tid : 0);
510         return 0;
511 }
512
513 int jbd2_log_start_commit(journal_t *journal, tid_t tid)
514 {
515         int ret;
516
517         write_lock(&journal->j_state_lock);
518         ret = __jbd2_log_start_commit(journal, tid);
519         write_unlock(&journal->j_state_lock);
520         return ret;
521 }
522
523 /*
524  * Force and wait upon a commit if the calling process is not within
525  * transaction.  This is used for forcing out undo-protected data which contains
526  * bitmaps, when the fs is running out of space.
527  *
528  * We can only force the running transaction if we don't have an active handle;
529  * otherwise, we will deadlock.
530  *
531  * Returns true if a transaction was started.
532  */
533 int jbd2_journal_force_commit_nested(journal_t *journal)
534 {
535         transaction_t *transaction = NULL;
536         tid_t tid;
537         int need_to_start = 0;
538
539         read_lock(&journal->j_state_lock);
540         if (journal->j_running_transaction && !current->journal_info) {
541                 transaction = journal->j_running_transaction;
542                 if (!tid_geq(journal->j_commit_request, transaction->t_tid))
543                         need_to_start = 1;
544         } else if (journal->j_committing_transaction)
545                 transaction = journal->j_committing_transaction;
546
547         if (!transaction) {
548                 read_unlock(&journal->j_state_lock);
549                 return 0;       /* Nothing to retry */
550         }
551
552         tid = transaction->t_tid;
553         read_unlock(&journal->j_state_lock);
554         if (need_to_start)
555                 jbd2_log_start_commit(journal, tid);
556         jbd2_log_wait_commit(journal, tid);
557         return 1;
558 }
559
560 /*
561  * Start a commit of the current running transaction (if any).  Returns true
562  * if a transaction is going to be committed (or is currently already
563  * committing), and fills its tid in at *ptid
564  */
565 int jbd2_journal_start_commit(journal_t *journal, tid_t *ptid)
566 {
567         int ret = 0;
568
569         write_lock(&journal->j_state_lock);
570         if (journal->j_running_transaction) {
571                 tid_t tid = journal->j_running_transaction->t_tid;
572
573                 __jbd2_log_start_commit(journal, tid);
574                 /* There's a running transaction and we've just made sure
575                  * it's commit has been scheduled. */
576                 if (ptid)
577                         *ptid = tid;
578                 ret = 1;
579         } else if (journal->j_committing_transaction) {
580                 /*
581                  * If ext3_write_super() recently started a commit, then we
582                  * have to wait for completion of that transaction
583                  */
584                 if (ptid)
585                         *ptid = journal->j_committing_transaction->t_tid;
586                 ret = 1;
587         }
588         write_unlock(&journal->j_state_lock);
589         return ret;
590 }
591
592 /*
593  * Return 1 if a given transaction has not yet sent barrier request
594  * connected with a transaction commit. If 0 is returned, transaction
595  * may or may not have sent the barrier. Used to avoid sending barrier
596  * twice in common cases.
597  */
598 int jbd2_trans_will_send_data_barrier(journal_t *journal, tid_t tid)
599 {
600         int ret = 0;
601         transaction_t *commit_trans;
602
603         if (!(journal->j_flags & JBD2_BARRIER))
604                 return 0;
605         read_lock(&journal->j_state_lock);
606         /* Transaction already committed? */
607         if (tid_geq(journal->j_commit_sequence, tid))
608                 goto out;
609         commit_trans = journal->j_committing_transaction;
610         if (!commit_trans || commit_trans->t_tid != tid) {
611                 ret = 1;
612                 goto out;
613         }
614         /*
615          * Transaction is being committed and we already proceeded to
616          * submitting a flush to fs partition?
617          */
618         if (journal->j_fs_dev != journal->j_dev) {
619                 if (!commit_trans->t_need_data_flush ||
620                     commit_trans->t_state >= T_COMMIT_DFLUSH)
621                         goto out;
622         } else {
623                 if (commit_trans->t_state >= T_COMMIT_JFLUSH)
624                         goto out;
625         }
626         ret = 1;
627 out:
628         read_unlock(&journal->j_state_lock);
629         return ret;
630 }
631 EXPORT_SYMBOL(jbd2_trans_will_send_data_barrier);
632
633 /*
634  * Wait for a specified commit to complete.
635  * The caller may not hold the journal lock.
636  */
637 int jbd2_log_wait_commit(journal_t *journal, tid_t tid)
638 {
639         int err = 0;
640
641         read_lock(&journal->j_state_lock);
642 #ifdef CONFIG_JBD2_DEBUG
643         if (!tid_geq(journal->j_commit_request, tid)) {
644                 printk(KERN_EMERG
645                        "%s: error: j_commit_request=%d, tid=%d\n",
646                        __func__, journal->j_commit_request, tid);
647         }
648 #endif
649         while (tid_gt(tid, journal->j_commit_sequence)) {
650                 jbd_debug(1, "JBD2: want %d, j_commit_sequence=%d\n",
651                                   tid, journal->j_commit_sequence);
652                 wake_up(&journal->j_wait_commit);
653                 read_unlock(&journal->j_state_lock);
654                 wait_event(journal->j_wait_done_commit,
655                                 !tid_gt(tid, journal->j_commit_sequence));
656                 read_lock(&journal->j_state_lock);
657         }
658         read_unlock(&journal->j_state_lock);
659
660         if (unlikely(is_journal_aborted(journal))) {
661                 printk(KERN_EMERG "journal commit I/O error\n");
662                 err = -EIO;
663         }
664         return err;
665 }
666
667 /*
668  * Log buffer allocation routines:
669  */
670
671 int jbd2_journal_next_log_block(journal_t *journal, unsigned long long *retp)
672 {
673         unsigned long blocknr;
674
675         write_lock(&journal->j_state_lock);
676         J_ASSERT(journal->j_free > 1);
677
678         blocknr = journal->j_head;
679         journal->j_head++;
680         journal->j_free--;
681         if (journal->j_head == journal->j_last)
682                 journal->j_head = journal->j_first;
683         write_unlock(&journal->j_state_lock);
684         return jbd2_journal_bmap(journal, blocknr, retp);
685 }
686
687 /*
688  * Conversion of logical to physical block numbers for the journal
689  *
690  * On external journals the journal blocks are identity-mapped, so
691  * this is a no-op.  If needed, we can use j_blk_offset - everything is
692  * ready.
693  */
694 int jbd2_journal_bmap(journal_t *journal, unsigned long blocknr,
695                  unsigned long long *retp)
696 {
697         int err = 0;
698         unsigned long long ret;
699
700         if (journal->j_inode) {
701                 ret = bmap(journal->j_inode, blocknr);
702                 if (ret)
703                         *retp = ret;
704                 else {
705                         printk(KERN_ALERT "%s: journal block not found "
706                                         "at offset %lu on %s\n",
707                                __func__, blocknr, journal->j_devname);
708                         err = -EIO;
709                         __journal_abort_soft(journal, err);
710                 }
711         } else {
712                 *retp = blocknr; /* +journal->j_blk_offset */
713         }
714         return err;
715 }
716
717 /*
718  * We play buffer_head aliasing tricks to write data/metadata blocks to
719  * the journal without copying their contents, but for journal
720  * descriptor blocks we do need to generate bona fide buffers.
721  *
722  * After the caller of jbd2_journal_get_descriptor_buffer() has finished modifying
723  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
724  * But we don't bother doing that, so there will be coherency problems with
725  * mmaps of blockdevs which hold live JBD-controlled filesystems.
726  */
727 struct journal_head *jbd2_journal_get_descriptor_buffer(journal_t *journal)
728 {
729         struct buffer_head *bh;
730         unsigned long long blocknr;
731         int err;
732
733         err = jbd2_journal_next_log_block(journal, &blocknr);
734
735         if (err)
736                 return NULL;
737
738         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
739         if (!bh)
740                 return NULL;
741         lock_buffer(bh);
742         memset(bh->b_data, 0, journal->j_blocksize);
743         set_buffer_uptodate(bh);
744         unlock_buffer(bh);
745         BUFFER_TRACE(bh, "return this buffer");
746         return jbd2_journal_add_journal_head(bh);
747 }
748
749 struct jbd2_stats_proc_session {
750         journal_t *journal;
751         struct transaction_stats_s *stats;
752         int start;
753         int max;
754 };
755
756 static void *jbd2_seq_info_start(struct seq_file *seq, loff_t *pos)
757 {
758         return *pos ? NULL : SEQ_START_TOKEN;
759 }
760
761 static void *jbd2_seq_info_next(struct seq_file *seq, void *v, loff_t *pos)
762 {
763         return NULL;
764 }
765
766 static int jbd2_seq_info_show(struct seq_file *seq, void *v)
767 {
768         struct jbd2_stats_proc_session *s = seq->private;
769
770         if (v != SEQ_START_TOKEN)
771                 return 0;
772         seq_printf(seq, "%lu transaction, each up to %u blocks\n",
773                         s->stats->ts_tid,
774                         s->journal->j_max_transaction_buffers);
775         if (s->stats->ts_tid == 0)
776                 return 0;
777         seq_printf(seq, "average: \n  %ums waiting for transaction\n",
778             jiffies_to_msecs(s->stats->run.rs_wait / s->stats->ts_tid));
779         seq_printf(seq, "  %ums running transaction\n",
780             jiffies_to_msecs(s->stats->run.rs_running / s->stats->ts_tid));
781         seq_printf(seq, "  %ums transaction was being locked\n",
782             jiffies_to_msecs(s->stats->run.rs_locked / s->stats->ts_tid));
783         seq_printf(seq, "  %ums flushing data (in ordered mode)\n",
784             jiffies_to_msecs(s->stats->run.rs_flushing / s->stats->ts_tid));
785         seq_printf(seq, "  %ums logging transaction\n",
786             jiffies_to_msecs(s->stats->run.rs_logging / s->stats->ts_tid));
787         seq_printf(seq, "  %lluus average transaction commit time\n",
788                    div_u64(s->journal->j_average_commit_time, 1000));
789         seq_printf(seq, "  %lu handles per transaction\n",
790             s->stats->run.rs_handle_count / s->stats->ts_tid);
791         seq_printf(seq, "  %lu blocks per transaction\n",
792             s->stats->run.rs_blocks / s->stats->ts_tid);
793         seq_printf(seq, "  %lu logged blocks per transaction\n",
794             s->stats->run.rs_blocks_logged / s->stats->ts_tid);
795         return 0;
796 }
797
798 static void jbd2_seq_info_stop(struct seq_file *seq, void *v)
799 {
800 }
801
802 static const struct seq_operations jbd2_seq_info_ops = {
803         .start  = jbd2_seq_info_start,
804         .next   = jbd2_seq_info_next,
805         .stop   = jbd2_seq_info_stop,
806         .show   = jbd2_seq_info_show,
807 };
808
809 static int jbd2_seq_info_open(struct inode *inode, struct file *file)
810 {
811         journal_t *journal = PDE(inode)->data;
812         struct jbd2_stats_proc_session *s;
813         int rc, size;
814
815         s = kmalloc(sizeof(*s), GFP_KERNEL);
816         if (s == NULL)
817                 return -ENOMEM;
818         size = sizeof(struct transaction_stats_s);
819         s->stats = kmalloc(size, GFP_KERNEL);
820         if (s->stats == NULL) {
821                 kfree(s);
822                 return -ENOMEM;
823         }
824         spin_lock(&journal->j_history_lock);
825         memcpy(s->stats, &journal->j_stats, size);
826         s->journal = journal;
827         spin_unlock(&journal->j_history_lock);
828
829         rc = seq_open(file, &jbd2_seq_info_ops);
830         if (rc == 0) {
831                 struct seq_file *m = file->private_data;
832                 m->private = s;
833         } else {
834                 kfree(s->stats);
835                 kfree(s);
836         }
837         return rc;
838
839 }
840
841 static int jbd2_seq_info_release(struct inode *inode, struct file *file)
842 {
843         struct seq_file *seq = file->private_data;
844         struct jbd2_stats_proc_session *s = seq->private;
845         kfree(s->stats);
846         kfree(s);
847         return seq_release(inode, file);
848 }
849
850 static const struct file_operations jbd2_seq_info_fops = {
851         .owner          = THIS_MODULE,
852         .open           = jbd2_seq_info_open,
853         .read           = seq_read,
854         .llseek         = seq_lseek,
855         .release        = jbd2_seq_info_release,
856 };
857
858 static struct proc_dir_entry *proc_jbd2_stats;
859
860 static void jbd2_stats_proc_init(journal_t *journal)
861 {
862         journal->j_proc_entry = proc_mkdir(journal->j_devname, proc_jbd2_stats);
863         if (journal->j_proc_entry) {
864                 proc_create_data("info", S_IRUGO, journal->j_proc_entry,
865                                  &jbd2_seq_info_fops, journal);
866         }
867 }
868
869 static void jbd2_stats_proc_exit(journal_t *journal)
870 {
871         remove_proc_entry("info", journal->j_proc_entry);
872         remove_proc_entry(journal->j_devname, proc_jbd2_stats);
873 }
874
875 /*
876  * Management for journal control blocks: functions to create and
877  * destroy journal_t structures, and to initialise and read existing
878  * journal blocks from disk.  */
879
880 /* First: create and setup a journal_t object in memory.  We initialise
881  * very few fields yet: that has to wait until we have created the
882  * journal structures from from scratch, or loaded them from disk. */
883
884 static journal_t * journal_init_common (void)
885 {
886         journal_t *journal;
887         int err;
888
889         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
890         if (!journal)
891                 return NULL;
892
893         init_waitqueue_head(&journal->j_wait_transaction_locked);
894         init_waitqueue_head(&journal->j_wait_logspace);
895         init_waitqueue_head(&journal->j_wait_done_commit);
896         init_waitqueue_head(&journal->j_wait_checkpoint);
897         init_waitqueue_head(&journal->j_wait_commit);
898         init_waitqueue_head(&journal->j_wait_updates);
899         mutex_init(&journal->j_barrier);
900         mutex_init(&journal->j_checkpoint_mutex);
901         spin_lock_init(&journal->j_revoke_lock);
902         spin_lock_init(&journal->j_list_lock);
903         rwlock_init(&journal->j_state_lock);
904
905         journal->j_commit_interval = (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE);
906         journal->j_min_batch_time = 0;
907         journal->j_max_batch_time = 15000; /* 15ms */
908
909         /* The journal is marked for error until we succeed with recovery! */
910         journal->j_flags = JBD2_ABORT;
911
912         /* Set up a default-sized revoke table for the new mount. */
913         err = jbd2_journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
914         if (err) {
915                 kfree(journal);
916                 return NULL;
917         }
918
919         spin_lock_init(&journal->j_history_lock);
920
921         return journal;
922 }
923
924 /* jbd2_journal_init_dev and jbd2_journal_init_inode:
925  *
926  * Create a journal structure assigned some fixed set of disk blocks to
927  * the journal.  We don't actually touch those disk blocks yet, but we
928  * need to set up all of the mapping information to tell the journaling
929  * system where the journal blocks are.
930  *
931  */
932
933 /**
934  *  journal_t * jbd2_journal_init_dev() - creates and initialises a journal structure
935  *  @bdev: Block device on which to create the journal
936  *  @fs_dev: Device which hold journalled filesystem for this journal.
937  *  @start: Block nr Start of journal.
938  *  @len:  Length of the journal in blocks.
939  *  @blocksize: blocksize of journalling device
940  *
941  *  Returns: a newly created journal_t *
942  *
943  *  jbd2_journal_init_dev creates a journal which maps a fixed contiguous
944  *  range of blocks on an arbitrary block device.
945  *
946  */
947 journal_t * jbd2_journal_init_dev(struct block_device *bdev,
948                         struct block_device *fs_dev,
949                         unsigned long long start, int len, int blocksize)
950 {
951         journal_t *journal = journal_init_common();
952         struct buffer_head *bh;
953         char *p;
954         int n;
955
956         if (!journal)
957                 return NULL;
958
959         /* journal descriptor can store up to n blocks -bzzz */
960         journal->j_blocksize = blocksize;
961         journal->j_dev = bdev;
962         journal->j_fs_dev = fs_dev;
963         journal->j_blk_offset = start;
964         journal->j_maxlen = len;
965         bdevname(journal->j_dev, journal->j_devname);
966         p = journal->j_devname;
967         while ((p = strchr(p, '/')))
968                 *p = '!';
969         jbd2_stats_proc_init(journal);
970         n = journal->j_blocksize / sizeof(journal_block_tag_t);
971         journal->j_wbufsize = n;
972         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
973         if (!journal->j_wbuf) {
974                 printk(KERN_ERR "%s: Can't allocate bhs for commit thread\n",
975                         __func__);
976                 goto out_err;
977         }
978
979         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
980         if (!bh) {
981                 printk(KERN_ERR
982                        "%s: Cannot get buffer for journal superblock\n",
983                        __func__);
984                 goto out_err;
985         }
986         journal->j_sb_buffer = bh;
987         journal->j_superblock = (journal_superblock_t *)bh->b_data;
988
989         return journal;
990 out_err:
991         kfree(journal->j_wbuf);
992         jbd2_stats_proc_exit(journal);
993         kfree(journal);
994         return NULL;
995 }
996
997 /**
998  *  journal_t * jbd2_journal_init_inode () - creates a journal which maps to a inode.
999  *  @inode: An inode to create the journal in
1000  *
1001  * jbd2_journal_init_inode creates a journal which maps an on-disk inode as
1002  * the journal.  The inode must exist already, must support bmap() and
1003  * must have all data blocks preallocated.
1004  */
1005 journal_t * jbd2_journal_init_inode (struct inode *inode)
1006 {
1007         struct buffer_head *bh;
1008         journal_t *journal = journal_init_common();
1009         char *p;
1010         int err;
1011         int n;
1012         unsigned long long blocknr;
1013
1014         if (!journal)
1015                 return NULL;
1016
1017         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
1018         journal->j_inode = inode;
1019         bdevname(journal->j_dev, journal->j_devname);
1020         p = journal->j_devname;
1021         while ((p = strchr(p, '/')))
1022                 *p = '!';
1023         p = journal->j_devname + strlen(journal->j_devname);
1024         sprintf(p, "-%lu", journal->j_inode->i_ino);
1025         jbd_debug(1,
1026                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
1027                   journal, inode->i_sb->s_id, inode->i_ino,
1028                   (long long) inode->i_size,
1029                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
1030
1031         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
1032         journal->j_blocksize = inode->i_sb->s_blocksize;
1033         jbd2_stats_proc_init(journal);
1034
1035         /* journal descriptor can store up to n blocks -bzzz */
1036         n = journal->j_blocksize / sizeof(journal_block_tag_t);
1037         journal->j_wbufsize = n;
1038         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
1039         if (!journal->j_wbuf) {
1040                 printk(KERN_ERR "%s: Can't allocate bhs for commit thread\n",
1041                         __func__);
1042                 goto out_err;
1043         }
1044
1045         err = jbd2_journal_bmap(journal, 0, &blocknr);
1046         /* If that failed, give up */
1047         if (err) {
1048                 printk(KERN_ERR "%s: Cannot locate journal superblock\n",
1049                        __func__);
1050                 goto out_err;
1051         }
1052
1053         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
1054         if (!bh) {
1055                 printk(KERN_ERR
1056                        "%s: Cannot get buffer for journal superblock\n",
1057                        __func__);
1058                 goto out_err;
1059         }
1060         journal->j_sb_buffer = bh;
1061         journal->j_superblock = (journal_superblock_t *)bh->b_data;
1062
1063         return journal;
1064 out_err:
1065         kfree(journal->j_wbuf);
1066         jbd2_stats_proc_exit(journal);
1067         kfree(journal);
1068         return NULL;
1069 }
1070
1071 /*
1072  * If the journal init or create aborts, we need to mark the journal
1073  * superblock as being NULL to prevent the journal destroy from writing
1074  * back a bogus superblock.
1075  */
1076 static void journal_fail_superblock (journal_t *journal)
1077 {
1078         struct buffer_head *bh = journal->j_sb_buffer;
1079         brelse(bh);
1080         journal->j_sb_buffer = NULL;
1081 }
1082
1083 /*
1084  * Given a journal_t structure, initialise the various fields for
1085  * startup of a new journaling session.  We use this both when creating
1086  * a journal, and after recovering an old journal to reset it for
1087  * subsequent use.
1088  */
1089
1090 static int journal_reset(journal_t *journal)
1091 {
1092         journal_superblock_t *sb = journal->j_superblock;
1093         unsigned long long first, last;
1094
1095         first = be32_to_cpu(sb->s_first);
1096         last = be32_to_cpu(sb->s_maxlen);
1097         if (first + JBD2_MIN_JOURNAL_BLOCKS > last + 1) {
1098                 printk(KERN_ERR "JBD2: Journal too short (blocks %llu-%llu).\n",
1099                        first, last);
1100                 journal_fail_superblock(journal);
1101                 return -EINVAL;
1102         }
1103
1104         journal->j_first = first;
1105         journal->j_last = last;
1106
1107         journal->j_head = first;
1108         journal->j_tail = first;
1109         journal->j_free = last - first;
1110
1111         journal->j_tail_sequence = journal->j_transaction_sequence;
1112         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
1113         journal->j_commit_request = journal->j_commit_sequence;
1114
1115         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
1116
1117         /* Add the dynamic fields and write it to disk. */
1118         jbd2_journal_update_superblock(journal, 1);
1119         return jbd2_journal_start_thread(journal);
1120 }
1121
1122 /**
1123  * void jbd2_journal_update_superblock() - Update journal sb on disk.
1124  * @journal: The journal to update.
1125  * @wait: Set to '0' if you don't want to wait for IO completion.
1126  *
1127  * Update a journal's dynamic superblock fields and write it to disk,
1128  * optionally waiting for the IO to complete.
1129  */
1130 void jbd2_journal_update_superblock(journal_t *journal, int wait)
1131 {
1132         journal_superblock_t *sb = journal->j_superblock;
1133         struct buffer_head *bh = journal->j_sb_buffer;
1134
1135         /*
1136          * As a special case, if the on-disk copy is already marked as needing
1137          * no recovery (s_start == 0) and there are no outstanding transactions
1138          * in the filesystem, then we can safely defer the superblock update
1139          * until the next commit by setting JBD2_FLUSHED.  This avoids
1140          * attempting a write to a potential-readonly device.
1141          */
1142         if (sb->s_start == 0 && journal->j_tail_sequence ==
1143                                 journal->j_transaction_sequence) {
1144                 jbd_debug(1, "JBD2: Skipping superblock update on recovered sb "
1145                         "(start %ld, seq %d, errno %d)\n",
1146                         journal->j_tail, journal->j_tail_sequence,
1147                         journal->j_errno);
1148                 goto out;
1149         }
1150
1151         if (buffer_write_io_error(bh)) {
1152                 /*
1153                  * Oh, dear.  A previous attempt to write the journal
1154                  * superblock failed.  This could happen because the
1155                  * USB device was yanked out.  Or it could happen to
1156                  * be a transient write error and maybe the block will
1157                  * be remapped.  Nothing we can do but to retry the
1158                  * write and hope for the best.
1159                  */
1160                 printk(KERN_ERR "JBD2: previous I/O error detected "
1161                        "for journal superblock update for %s.\n",
1162                        journal->j_devname);
1163                 clear_buffer_write_io_error(bh);
1164                 set_buffer_uptodate(bh);
1165         }
1166
1167         read_lock(&journal->j_state_lock);
1168         jbd_debug(1, "JBD2: updating superblock (start %ld, seq %d, errno %d)\n",
1169                   journal->j_tail, journal->j_tail_sequence, journal->j_errno);
1170
1171         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
1172         sb->s_start    = cpu_to_be32(journal->j_tail);
1173         sb->s_errno    = cpu_to_be32(journal->j_errno);
1174         read_unlock(&journal->j_state_lock);
1175
1176         BUFFER_TRACE(bh, "marking dirty");
1177         mark_buffer_dirty(bh);
1178         if (wait) {
1179                 sync_dirty_buffer(bh);
1180                 if (buffer_write_io_error(bh)) {
1181                         printk(KERN_ERR "JBD2: I/O error detected "
1182                                "when updating journal superblock for %s.\n",
1183                                journal->j_devname);
1184                         clear_buffer_write_io_error(bh);
1185                         set_buffer_uptodate(bh);
1186                 }
1187         } else
1188                 write_dirty_buffer(bh, WRITE);
1189
1190 out:
1191         /* If we have just flushed the log (by marking s_start==0), then
1192          * any future commit will have to be careful to update the
1193          * superblock again to re-record the true start of the log. */
1194
1195         write_lock(&journal->j_state_lock);
1196         if (sb->s_start)
1197                 journal->j_flags &= ~JBD2_FLUSHED;
1198         else
1199                 journal->j_flags |= JBD2_FLUSHED;
1200         write_unlock(&journal->j_state_lock);
1201 }
1202
1203 /*
1204  * Read the superblock for a given journal, performing initial
1205  * validation of the format.
1206  */
1207
1208 static int journal_get_superblock(journal_t *journal)
1209 {
1210         struct buffer_head *bh;
1211         journal_superblock_t *sb;
1212         int err = -EIO;
1213
1214         bh = journal->j_sb_buffer;
1215
1216         J_ASSERT(bh != NULL);
1217         if (!buffer_uptodate(bh)) {
1218                 ll_rw_block(READ, 1, &bh);
1219                 wait_on_buffer(bh);
1220                 if (!buffer_uptodate(bh)) {
1221                         printk(KERN_ERR
1222                                 "JBD2: IO error reading journal superblock\n");
1223                         goto out;
1224                 }
1225         }
1226
1227         sb = journal->j_superblock;
1228
1229         err = -EINVAL;
1230
1231         if (sb->s_header.h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER) ||
1232             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1233                 printk(KERN_WARNING "JBD2: no valid journal superblock found\n");
1234                 goto out;
1235         }
1236
1237         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1238         case JBD2_SUPERBLOCK_V1:
1239                 journal->j_format_version = 1;
1240                 break;
1241         case JBD2_SUPERBLOCK_V2:
1242                 journal->j_format_version = 2;
1243                 break;
1244         default:
1245                 printk(KERN_WARNING "JBD2: unrecognised superblock format ID\n");
1246                 goto out;
1247         }
1248
1249         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1250                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1251         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1252                 printk(KERN_WARNING "JBD2: journal file too short\n");
1253                 goto out;
1254         }
1255
1256         if (be32_to_cpu(sb->s_first) == 0 ||
1257             be32_to_cpu(sb->s_first) >= journal->j_maxlen) {
1258                 printk(KERN_WARNING
1259                         "JBD2: Invalid start block of journal: %u\n",
1260                         be32_to_cpu(sb->s_first));
1261                 goto out;
1262         }
1263
1264         return 0;
1265
1266 out:
1267         journal_fail_superblock(journal);
1268         return err;
1269 }
1270
1271 /*
1272  * Load the on-disk journal superblock and read the key fields into the
1273  * journal_t.
1274  */
1275
1276 static int load_superblock(journal_t *journal)
1277 {
1278         int err;
1279         journal_superblock_t *sb;
1280
1281         err = journal_get_superblock(journal);
1282         if (err)
1283                 return err;
1284
1285         sb = journal->j_superblock;
1286
1287         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1288         journal->j_tail = be32_to_cpu(sb->s_start);
1289         journal->j_first = be32_to_cpu(sb->s_first);
1290         journal->j_last = be32_to_cpu(sb->s_maxlen);
1291         journal->j_errno = be32_to_cpu(sb->s_errno);
1292
1293         return 0;
1294 }
1295
1296
1297 /**
1298  * int jbd2_journal_load() - Read journal from disk.
1299  * @journal: Journal to act on.
1300  *
1301  * Given a journal_t structure which tells us which disk blocks contain
1302  * a journal, read the journal from disk to initialise the in-memory
1303  * structures.
1304  */
1305 int jbd2_journal_load(journal_t *journal)
1306 {
1307         int err;
1308         journal_superblock_t *sb;
1309
1310         err = load_superblock(journal);
1311         if (err)
1312                 return err;
1313
1314         sb = journal->j_superblock;
1315         /* If this is a V2 superblock, then we have to check the
1316          * features flags on it. */
1317
1318         if (journal->j_format_version >= 2) {
1319                 if ((sb->s_feature_ro_compat &
1320                      ~cpu_to_be32(JBD2_KNOWN_ROCOMPAT_FEATURES)) ||
1321                     (sb->s_feature_incompat &
1322                      ~cpu_to_be32(JBD2_KNOWN_INCOMPAT_FEATURES))) {
1323                         printk(KERN_WARNING
1324                                 "JBD2: Unrecognised features on journal\n");
1325                         return -EINVAL;
1326                 }
1327         }
1328
1329         /*
1330          * Create a slab for this blocksize
1331          */
1332         err = jbd2_journal_create_slab(be32_to_cpu(sb->s_blocksize));
1333         if (err)
1334                 return err;
1335
1336         /* Let the recovery code check whether it needs to recover any
1337          * data from the journal. */
1338         if (jbd2_journal_recover(journal))
1339                 goto recovery_error;
1340
1341         if (journal->j_failed_commit) {
1342                 printk(KERN_ERR "JBD2: journal transaction %u on %s "
1343                        "is corrupt.\n", journal->j_failed_commit,
1344                        journal->j_devname);
1345                 return -EIO;
1346         }
1347
1348         /* OK, we've finished with the dynamic journal bits:
1349          * reinitialise the dynamic contents of the superblock in memory
1350          * and reset them on disk. */
1351         if (journal_reset(journal))
1352                 goto recovery_error;
1353
1354         journal->j_flags &= ~JBD2_ABORT;
1355         journal->j_flags |= JBD2_LOADED;
1356         return 0;
1357
1358 recovery_error:
1359         printk(KERN_WARNING "JBD2: recovery failed\n");
1360         return -EIO;
1361 }
1362
1363 /**
1364  * void jbd2_journal_destroy() - Release a journal_t structure.
1365  * @journal: Journal to act on.
1366  *
1367  * Release a journal_t structure once it is no longer in use by the
1368  * journaled object.
1369  * Return <0 if we couldn't clean up the journal.
1370  */
1371 int jbd2_journal_destroy(journal_t *journal)
1372 {
1373         int err = 0;
1374
1375         /* Wait for the commit thread to wake up and die. */
1376         journal_kill_thread(journal);
1377
1378         /* Force a final log commit */
1379         if (journal->j_running_transaction)
1380                 jbd2_journal_commit_transaction(journal);
1381
1382         /* Force any old transactions to disk */
1383
1384         /* Totally anal locking here... */
1385         spin_lock(&journal->j_list_lock);
1386         while (journal->j_checkpoint_transactions != NULL) {
1387                 spin_unlock(&journal->j_list_lock);
1388                 mutex_lock(&journal->j_checkpoint_mutex);
1389                 jbd2_log_do_checkpoint(journal);
1390                 mutex_unlock(&journal->j_checkpoint_mutex);
1391                 spin_lock(&journal->j_list_lock);
1392         }
1393
1394         J_ASSERT(journal->j_running_transaction == NULL);
1395         J_ASSERT(journal->j_committing_transaction == NULL);
1396         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1397         spin_unlock(&journal->j_list_lock);
1398
1399         if (journal->j_sb_buffer) {
1400                 if (!is_journal_aborted(journal)) {
1401                         /* We can now mark the journal as empty. */
1402                         journal->j_tail = 0;
1403                         journal->j_tail_sequence =
1404                                 ++journal->j_transaction_sequence;
1405                         jbd2_journal_update_superblock(journal, 1);
1406                 } else {
1407                         err = -EIO;
1408                 }
1409                 brelse(journal->j_sb_buffer);
1410         }
1411
1412         if (journal->j_proc_entry)
1413                 jbd2_stats_proc_exit(journal);
1414         if (journal->j_inode)
1415                 iput(journal->j_inode);
1416         if (journal->j_revoke)
1417                 jbd2_journal_destroy_revoke(journal);
1418         kfree(journal->j_wbuf);
1419         kfree(journal);
1420
1421         return err;
1422 }
1423
1424
1425 /**
1426  *int jbd2_journal_check_used_features () - Check if features specified are used.
1427  * @journal: Journal to check.
1428  * @compat: bitmask of compatible features
1429  * @ro: bitmask of features that force read-only mount
1430  * @incompat: bitmask of incompatible features
1431  *
1432  * Check whether the journal uses all of a given set of
1433  * features.  Return true (non-zero) if it does.
1434  **/
1435
1436 int jbd2_journal_check_used_features (journal_t *journal, unsigned long compat,
1437                                  unsigned long ro, unsigned long incompat)
1438 {
1439         journal_superblock_t *sb;
1440
1441         if (!compat && !ro && !incompat)
1442                 return 1;
1443         /* Load journal superblock if it is not loaded yet. */
1444         if (journal->j_format_version == 0 &&
1445             journal_get_superblock(journal) != 0)
1446                 return 0;
1447         if (journal->j_format_version == 1)
1448                 return 0;
1449
1450         sb = journal->j_superblock;
1451
1452         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1453             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1454             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1455                 return 1;
1456
1457         return 0;
1458 }
1459
1460 /**
1461  * int jbd2_journal_check_available_features() - Check feature set in journalling layer
1462  * @journal: Journal to check.
1463  * @compat: bitmask of compatible features
1464  * @ro: bitmask of features that force read-only mount
1465  * @incompat: bitmask of incompatible features
1466  *
1467  * Check whether the journaling code supports the use of
1468  * all of a given set of features on this journal.  Return true
1469  * (non-zero) if it can. */
1470
1471 int jbd2_journal_check_available_features (journal_t *journal, unsigned long compat,
1472                                       unsigned long ro, unsigned long incompat)
1473 {
1474         if (!compat && !ro && !incompat)
1475                 return 1;
1476
1477         /* We can support any known requested features iff the
1478          * superblock is in version 2.  Otherwise we fail to support any
1479          * extended sb features. */
1480
1481         if (journal->j_format_version != 2)
1482                 return 0;
1483
1484         if ((compat   & JBD2_KNOWN_COMPAT_FEATURES) == compat &&
1485             (ro       & JBD2_KNOWN_ROCOMPAT_FEATURES) == ro &&
1486             (incompat & JBD2_KNOWN_INCOMPAT_FEATURES) == incompat)
1487                 return 1;
1488
1489         return 0;
1490 }
1491
1492 /**
1493  * int jbd2_journal_set_features () - Mark a given journal feature in the superblock
1494  * @journal: Journal to act on.
1495  * @compat: bitmask of compatible features
1496  * @ro: bitmask of features that force read-only mount
1497  * @incompat: bitmask of incompatible features
1498  *
1499  * Mark a given journal feature as present on the
1500  * superblock.  Returns true if the requested features could be set.
1501  *
1502  */
1503
1504 int jbd2_journal_set_features (journal_t *journal, unsigned long compat,
1505                           unsigned long ro, unsigned long incompat)
1506 {
1507         journal_superblock_t *sb;
1508
1509         if (jbd2_journal_check_used_features(journal, compat, ro, incompat))
1510                 return 1;
1511
1512         if (!jbd2_journal_check_available_features(journal, compat, ro, incompat))
1513                 return 0;
1514
1515         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1516                   compat, ro, incompat);
1517
1518         sb = journal->j_superblock;
1519
1520         sb->s_feature_compat    |= cpu_to_be32(compat);
1521         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1522         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1523
1524         return 1;
1525 }
1526
1527 /*
1528  * jbd2_journal_clear_features () - Clear a given journal feature in the
1529  *                                  superblock
1530  * @journal: Journal to act on.
1531  * @compat: bitmask of compatible features
1532  * @ro: bitmask of features that force read-only mount
1533  * @incompat: bitmask of incompatible features
1534  *
1535  * Clear a given journal feature as present on the
1536  * superblock.
1537  */
1538 void jbd2_journal_clear_features(journal_t *journal, unsigned long compat,
1539                                 unsigned long ro, unsigned long incompat)
1540 {
1541         journal_superblock_t *sb;
1542
1543         jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
1544                   compat, ro, incompat);
1545
1546         sb = journal->j_superblock;
1547
1548         sb->s_feature_compat    &= ~cpu_to_be32(compat);
1549         sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
1550         sb->s_feature_incompat  &= ~cpu_to_be32(incompat);
1551 }
1552 EXPORT_SYMBOL(jbd2_journal_clear_features);
1553
1554 /**
1555  * int jbd2_journal_update_format () - Update on-disk journal structure.
1556  * @journal: Journal to act on.
1557  *
1558  * Given an initialised but unloaded journal struct, poke about in the
1559  * on-disk structure to update it to the most recent supported version.
1560  */
1561 int jbd2_journal_update_format (journal_t *journal)
1562 {
1563         journal_superblock_t *sb;
1564         int err;
1565
1566         err = journal_get_superblock(journal);
1567         if (err)
1568                 return err;
1569
1570         sb = journal->j_superblock;
1571
1572         switch (be32_to_cpu(sb->s_header.h_blocktype)) {
1573         case JBD2_SUPERBLOCK_V2:
1574                 return 0;
1575         case JBD2_SUPERBLOCK_V1:
1576                 return journal_convert_superblock_v1(journal, sb);
1577         default:
1578                 break;
1579         }
1580         return -EINVAL;
1581 }
1582
1583 static int journal_convert_superblock_v1(journal_t *journal,
1584                                          journal_superblock_t *sb)
1585 {
1586         int offset, blocksize;
1587         struct buffer_head *bh;
1588
1589         printk(KERN_WARNING
1590                 "JBD2: Converting superblock from version 1 to 2.\n");
1591
1592         /* Pre-initialise new fields to zero */
1593         offset = ((char *) &(sb->s_feature_compat)) - ((char *) sb);
1594         blocksize = be32_to_cpu(sb->s_blocksize);
1595         memset(&sb->s_feature_compat, 0, blocksize-offset);
1596
1597         sb->s_nr_users = cpu_to_be32(1);
1598         sb->s_header.h_blocktype = cpu_to_be32(JBD2_SUPERBLOCK_V2);
1599         journal->j_format_version = 2;
1600
1601         bh = journal->j_sb_buffer;
1602         BUFFER_TRACE(bh, "marking dirty");
1603         mark_buffer_dirty(bh);
1604         sync_dirty_buffer(bh);
1605         return 0;
1606 }
1607
1608
1609 /**
1610  * int jbd2_journal_flush () - Flush journal
1611  * @journal: Journal to act on.
1612  *
1613  * Flush all data for a given journal to disk and empty the journal.
1614  * Filesystems can use this when remounting readonly to ensure that
1615  * recovery does not need to happen on remount.
1616  */
1617
1618 int jbd2_journal_flush(journal_t *journal)
1619 {
1620         int err = 0;
1621         transaction_t *transaction = NULL;
1622         unsigned long old_tail;
1623
1624         write_lock(&journal->j_state_lock);
1625
1626         /* Force everything buffered to the log... */
1627         if (journal->j_running_transaction) {
1628                 transaction = journal->j_running_transaction;
1629                 __jbd2_log_start_commit(journal, transaction->t_tid);
1630         } else if (journal->j_committing_transaction)
1631                 transaction = journal->j_committing_transaction;
1632
1633         /* Wait for the log commit to complete... */
1634         if (transaction) {
1635                 tid_t tid = transaction->t_tid;
1636
1637                 write_unlock(&journal->j_state_lock);
1638                 jbd2_log_wait_commit(journal, tid);
1639         } else {
1640                 write_unlock(&journal->j_state_lock);
1641         }
1642
1643         /* ...and flush everything in the log out to disk. */
1644         spin_lock(&journal->j_list_lock);
1645         while (!err && journal->j_checkpoint_transactions != NULL) {
1646                 spin_unlock(&journal->j_list_lock);
1647                 mutex_lock(&journal->j_checkpoint_mutex);
1648                 err = jbd2_log_do_checkpoint(journal);
1649                 mutex_unlock(&journal->j_checkpoint_mutex);
1650                 spin_lock(&journal->j_list_lock);
1651         }
1652         spin_unlock(&journal->j_list_lock);
1653
1654         if (is_journal_aborted(journal))
1655                 return -EIO;
1656
1657         jbd2_cleanup_journal_tail(journal);
1658
1659         /* Finally, mark the journal as really needing no recovery.
1660          * This sets s_start==0 in the underlying superblock, which is
1661          * the magic code for a fully-recovered superblock.  Any future
1662          * commits of data to the journal will restore the current
1663          * s_start value. */
1664         write_lock(&journal->j_state_lock);
1665         old_tail = journal->j_tail;
1666         journal->j_tail = 0;
1667         write_unlock(&journal->j_state_lock);
1668         jbd2_journal_update_superblock(journal, 1);
1669         write_lock(&journal->j_state_lock);
1670         journal->j_tail = old_tail;
1671
1672         J_ASSERT(!journal->j_running_transaction);
1673         J_ASSERT(!journal->j_committing_transaction);
1674         J_ASSERT(!journal->j_checkpoint_transactions);
1675         J_ASSERT(journal->j_head == journal->j_tail);
1676         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1677         write_unlock(&journal->j_state_lock);
1678         return 0;
1679 }
1680
1681 /**
1682  * int jbd2_journal_wipe() - Wipe journal contents
1683  * @journal: Journal to act on.
1684  * @write: flag (see below)
1685  *
1686  * Wipe out all of the contents of a journal, safely.  This will produce
1687  * a warning if the journal contains any valid recovery information.
1688  * Must be called between journal_init_*() and jbd2_journal_load().
1689  *
1690  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1691  * we merely suppress recovery.
1692  */
1693
1694 int jbd2_journal_wipe(journal_t *journal, int write)
1695 {
1696         int err = 0;
1697
1698         J_ASSERT (!(journal->j_flags & JBD2_LOADED));
1699
1700         err = load_superblock(journal);
1701         if (err)
1702                 return err;
1703
1704         if (!journal->j_tail)
1705                 goto no_recovery;
1706
1707         printk(KERN_WARNING "JBD2: %s recovery information on journal\n",
1708                 write ? "Clearing" : "Ignoring");
1709
1710         err = jbd2_journal_skip_recovery(journal);
1711         if (write)
1712                 jbd2_journal_update_superblock(journal, 1);
1713
1714  no_recovery:
1715         return err;
1716 }
1717
1718 /*
1719  * Journal abort has very specific semantics, which we describe
1720  * for journal abort.
1721  *
1722  * Two internal functions, which provide abort to the jbd layer
1723  * itself are here.
1724  */
1725
1726 /*
1727  * Quick version for internal journal use (doesn't lock the journal).
1728  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1729  * and don't attempt to make any other journal updates.
1730  */
1731 void __jbd2_journal_abort_hard(journal_t *journal)
1732 {
1733         transaction_t *transaction;
1734
1735         if (journal->j_flags & JBD2_ABORT)
1736                 return;
1737
1738         printk(KERN_ERR "Aborting journal on device %s.\n",
1739                journal->j_devname);
1740
1741         write_lock(&journal->j_state_lock);
1742         journal->j_flags |= JBD2_ABORT;
1743         transaction = journal->j_running_transaction;
1744         if (transaction)
1745                 __jbd2_log_start_commit(journal, transaction->t_tid);
1746         write_unlock(&journal->j_state_lock);
1747 }
1748
1749 /* Soft abort: record the abort error status in the journal superblock,
1750  * but don't do any other IO. */
1751 static void __journal_abort_soft (journal_t *journal, int errno)
1752 {
1753         if (journal->j_flags & JBD2_ABORT)
1754                 return;
1755
1756         if (!journal->j_errno)
1757                 journal->j_errno = errno;
1758
1759         __jbd2_journal_abort_hard(journal);
1760
1761         if (errno)
1762                 jbd2_journal_update_superblock(journal, 1);
1763 }
1764
1765 /**
1766  * void jbd2_journal_abort () - Shutdown the journal immediately.
1767  * @journal: the journal to shutdown.
1768  * @errno:   an error number to record in the journal indicating
1769  *           the reason for the shutdown.
1770  *
1771  * Perform a complete, immediate shutdown of the ENTIRE
1772  * journal (not of a single transaction).  This operation cannot be
1773  * undone without closing and reopening the journal.
1774  *
1775  * The jbd2_journal_abort function is intended to support higher level error
1776  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1777  * mode.
1778  *
1779  * Journal abort has very specific semantics.  Any existing dirty,
1780  * unjournaled buffers in the main filesystem will still be written to
1781  * disk by bdflush, but the journaling mechanism will be suspended
1782  * immediately and no further transaction commits will be honoured.
1783  *
1784  * Any dirty, journaled buffers will be written back to disk without
1785  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1786  * filesystem, but we _do_ attempt to leave as much data as possible
1787  * behind for fsck to use for cleanup.
1788  *
1789  * Any attempt to get a new transaction handle on a journal which is in
1790  * ABORT state will just result in an -EROFS error return.  A
1791  * jbd2_journal_stop on an existing handle will return -EIO if we have
1792  * entered abort state during the update.
1793  *
1794  * Recursive transactions are not disturbed by journal abort until the
1795  * final jbd2_journal_stop, which will receive the -EIO error.
1796  *
1797  * Finally, the jbd2_journal_abort call allows the caller to supply an errno
1798  * which will be recorded (if possible) in the journal superblock.  This
1799  * allows a client to record failure conditions in the middle of a
1800  * transaction without having to complete the transaction to record the
1801  * failure to disk.  ext3_error, for example, now uses this
1802  * functionality.
1803  *
1804  * Errors which originate from within the journaling layer will NOT
1805  * supply an errno; a null errno implies that absolutely no further
1806  * writes are done to the journal (unless there are any already in
1807  * progress).
1808  *
1809  */
1810
1811 void jbd2_journal_abort(journal_t *journal, int errno)
1812 {
1813         __journal_abort_soft(journal, errno);
1814 }
1815
1816 /**
1817  * int jbd2_journal_errno () - returns the journal's error state.
1818  * @journal: journal to examine.
1819  *
1820  * This is the errno number set with jbd2_journal_abort(), the last
1821  * time the journal was mounted - if the journal was stopped
1822  * without calling abort this will be 0.
1823  *
1824  * If the journal has been aborted on this mount time -EROFS will
1825  * be returned.
1826  */
1827 int jbd2_journal_errno(journal_t *journal)
1828 {
1829         int err;
1830
1831         read_lock(&journal->j_state_lock);
1832         if (journal->j_flags & JBD2_ABORT)
1833                 err = -EROFS;
1834         else
1835                 err = journal->j_errno;
1836         read_unlock(&journal->j_state_lock);
1837         return err;
1838 }
1839
1840 /**
1841  * int jbd2_journal_clear_err () - clears the journal's error state
1842  * @journal: journal to act on.
1843  *
1844  * An error must be cleared or acked to take a FS out of readonly
1845  * mode.
1846  */
1847 int jbd2_journal_clear_err(journal_t *journal)
1848 {
1849         int err = 0;
1850
1851         write_lock(&journal->j_state_lock);
1852         if (journal->j_flags & JBD2_ABORT)
1853                 err = -EROFS;
1854         else
1855                 journal->j_errno = 0;
1856         write_unlock(&journal->j_state_lock);
1857         return err;
1858 }
1859
1860 /**
1861  * void jbd2_journal_ack_err() - Ack journal err.
1862  * @journal: journal to act on.
1863  *
1864  * An error must be cleared or acked to take a FS out of readonly
1865  * mode.
1866  */
1867 void jbd2_journal_ack_err(journal_t *journal)
1868 {
1869         write_lock(&journal->j_state_lock);
1870         if (journal->j_errno)
1871                 journal->j_flags |= JBD2_ACK_ERR;
1872         write_unlock(&journal->j_state_lock);
1873 }
1874
1875 int jbd2_journal_blocks_per_page(struct inode *inode)
1876 {
1877         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1878 }
1879
1880 /*
1881  * helper functions to deal with 32 or 64bit block numbers.
1882  */
1883 size_t journal_tag_bytes(journal_t *journal)
1884 {
1885         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
1886                 return JBD2_TAG_SIZE64;
1887         else
1888                 return JBD2_TAG_SIZE32;
1889 }
1890
1891 /*
1892  * JBD memory management
1893  *
1894  * These functions are used to allocate block-sized chunks of memory
1895  * used for making copies of buffer_head data.  Very often it will be
1896  * page-sized chunks of data, but sometimes it will be in
1897  * sub-page-size chunks.  (For example, 16k pages on Power systems
1898  * with a 4k block file system.)  For blocks smaller than a page, we
1899  * use a SLAB allocator.  There are slab caches for each block size,
1900  * which are allocated at mount time, if necessary, and we only free
1901  * (all of) the slab caches when/if the jbd2 module is unloaded.  For
1902  * this reason we don't need to a mutex to protect access to
1903  * jbd2_slab[] allocating or releasing memory; only in
1904  * jbd2_journal_create_slab().
1905  */
1906 #define JBD2_MAX_SLABS 8
1907 static struct kmem_cache *jbd2_slab[JBD2_MAX_SLABS];
1908
1909 static const char *jbd2_slab_names[JBD2_MAX_SLABS] = {
1910         "jbd2_1k", "jbd2_2k", "jbd2_4k", "jbd2_8k",
1911         "jbd2_16k", "jbd2_32k", "jbd2_64k", "jbd2_128k"
1912 };
1913
1914
1915 static void jbd2_journal_destroy_slabs(void)
1916 {
1917         int i;
1918
1919         for (i = 0; i < JBD2_MAX_SLABS; i++) {
1920                 if (jbd2_slab[i])
1921                         kmem_cache_destroy(jbd2_slab[i]);
1922                 jbd2_slab[i] = NULL;
1923         }
1924 }
1925
1926 static int jbd2_journal_create_slab(size_t size)
1927 {
1928         static DEFINE_MUTEX(jbd2_slab_create_mutex);
1929         int i = order_base_2(size) - 10;
1930         size_t slab_size;
1931
1932         if (size == PAGE_SIZE)
1933                 return 0;
1934
1935         if (i >= JBD2_MAX_SLABS)
1936                 return -EINVAL;
1937
1938         if (unlikely(i < 0))
1939                 i = 0;
1940         mutex_lock(&jbd2_slab_create_mutex);
1941         if (jbd2_slab[i]) {
1942                 mutex_unlock(&jbd2_slab_create_mutex);
1943                 return 0;       /* Already created */
1944         }
1945
1946         slab_size = 1 << (i+10);
1947         jbd2_slab[i] = kmem_cache_create(jbd2_slab_names[i], slab_size,
1948                                          slab_size, 0, NULL);
1949         mutex_unlock(&jbd2_slab_create_mutex);
1950         if (!jbd2_slab[i]) {
1951                 printk(KERN_EMERG "JBD2: no memory for jbd2_slab cache\n");
1952                 return -ENOMEM;
1953         }
1954         return 0;
1955 }
1956
1957 static struct kmem_cache *get_slab(size_t size)
1958 {
1959         int i = order_base_2(size) - 10;
1960
1961         BUG_ON(i >= JBD2_MAX_SLABS);
1962         if (unlikely(i < 0))
1963                 i = 0;
1964         BUG_ON(jbd2_slab[i] == NULL);
1965         return jbd2_slab[i];
1966 }
1967
1968 void *jbd2_alloc(size_t size, gfp_t flags)
1969 {
1970         void *ptr;
1971
1972         BUG_ON(size & (size-1)); /* Must be a power of 2 */
1973
1974         flags |= __GFP_REPEAT;
1975         if (size == PAGE_SIZE)
1976                 ptr = (void *)__get_free_pages(flags, 0);
1977         else if (size > PAGE_SIZE) {
1978                 int order = get_order(size);
1979
1980                 if (order < 3)
1981                         ptr = (void *)__get_free_pages(flags, order);
1982                 else
1983                         ptr = vmalloc(size);
1984         } else
1985                 ptr = kmem_cache_alloc(get_slab(size), flags);
1986
1987         /* Check alignment; SLUB has gotten this wrong in the past,
1988          * and this can lead to user data corruption! */
1989         BUG_ON(((unsigned long) ptr) & (size-1));
1990
1991         return ptr;
1992 }
1993
1994 void jbd2_free(void *ptr, size_t size)
1995 {
1996         if (size == PAGE_SIZE) {
1997                 free_pages((unsigned long)ptr, 0);
1998                 return;
1999         }
2000         if (size > PAGE_SIZE) {
2001                 int order = get_order(size);
2002
2003                 if (order < 3)
2004                         free_pages((unsigned long)ptr, order);
2005                 else
2006                         vfree(ptr);
2007                 return;
2008         }
2009         kmem_cache_free(get_slab(size), ptr);
2010 };
2011
2012 /*
2013  * Journal_head storage management
2014  */
2015 static struct kmem_cache *jbd2_journal_head_cache;
2016 #ifdef CONFIG_JBD2_DEBUG
2017 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
2018 #endif
2019
2020 static int journal_init_jbd2_journal_head_cache(void)
2021 {
2022         int retval;
2023
2024         J_ASSERT(jbd2_journal_head_cache == NULL);
2025         jbd2_journal_head_cache = kmem_cache_create("jbd2_journal_head",
2026                                 sizeof(struct journal_head),
2027                                 0,              /* offset */
2028                                 SLAB_TEMPORARY, /* flags */
2029                                 NULL);          /* ctor */
2030         retval = 0;
2031         if (!jbd2_journal_head_cache) {
2032                 retval = -ENOMEM;
2033                 printk(KERN_EMERG "JBD2: no memory for journal_head cache\n");
2034         }
2035         return retval;
2036 }
2037
2038 static void jbd2_journal_destroy_jbd2_journal_head_cache(void)
2039 {
2040         if (jbd2_journal_head_cache) {
2041                 kmem_cache_destroy(jbd2_journal_head_cache);
2042                 jbd2_journal_head_cache = NULL;
2043         }
2044 }
2045
2046 /*
2047  * journal_head splicing and dicing
2048  */
2049 static struct journal_head *journal_alloc_journal_head(void)
2050 {
2051         struct journal_head *ret;
2052
2053 #ifdef CONFIG_JBD2_DEBUG
2054         atomic_inc(&nr_journal_heads);
2055 #endif
2056         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
2057         if (!ret) {
2058                 jbd_debug(1, "out of memory for journal_head\n");
2059                 pr_notice_ratelimited("ENOMEM in %s, retrying.\n", __func__);
2060                 while (!ret) {
2061                         yield();
2062                         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
2063                 }
2064         }
2065         return ret;
2066 }
2067
2068 static void journal_free_journal_head(struct journal_head *jh)
2069 {
2070 #ifdef CONFIG_JBD2_DEBUG
2071         atomic_dec(&nr_journal_heads);
2072         memset(jh, JBD2_POISON_FREE, sizeof(*jh));
2073 #endif
2074         kmem_cache_free(jbd2_journal_head_cache, jh);
2075 }
2076
2077 /*
2078  * A journal_head is attached to a buffer_head whenever JBD has an
2079  * interest in the buffer.
2080  *
2081  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
2082  * is set.  This bit is tested in core kernel code where we need to take
2083  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
2084  * there.
2085  *
2086  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
2087  *
2088  * When a buffer has its BH_JBD bit set it is immune from being released by
2089  * core kernel code, mainly via ->b_count.
2090  *
2091  * A journal_head is detached from its buffer_head when the journal_head's
2092  * b_jcount reaches zero. Running transaction (b_transaction) and checkpoint
2093  * transaction (b_cp_transaction) hold their references to b_jcount.
2094  *
2095  * Various places in the kernel want to attach a journal_head to a buffer_head
2096  * _before_ attaching the journal_head to a transaction.  To protect the
2097  * journal_head in this situation, jbd2_journal_add_journal_head elevates the
2098  * journal_head's b_jcount refcount by one.  The caller must call
2099  * jbd2_journal_put_journal_head() to undo this.
2100  *
2101  * So the typical usage would be:
2102  *
2103  *      (Attach a journal_head if needed.  Increments b_jcount)
2104  *      struct journal_head *jh = jbd2_journal_add_journal_head(bh);
2105  *      ...
2106  *      (Get another reference for transaction)
2107  *      jbd2_journal_grab_journal_head(bh);
2108  *      jh->b_transaction = xxx;
2109  *      (Put original reference)
2110  *      jbd2_journal_put_journal_head(jh);
2111  */
2112
2113 /*
2114  * Give a buffer_head a journal_head.
2115  *
2116  * May sleep.
2117  */
2118 struct journal_head *jbd2_journal_add_journal_head(struct buffer_head *bh)
2119 {
2120         struct journal_head *jh;
2121         struct journal_head *new_jh = NULL;
2122
2123 repeat:
2124         if (!buffer_jbd(bh)) {
2125                 new_jh = journal_alloc_journal_head();
2126                 memset(new_jh, 0, sizeof(*new_jh));
2127         }
2128
2129         jbd_lock_bh_journal_head(bh);
2130         if (buffer_jbd(bh)) {
2131                 jh = bh2jh(bh);
2132         } else {
2133                 J_ASSERT_BH(bh,
2134                         (atomic_read(&bh->b_count) > 0) ||
2135                         (bh->b_page && bh->b_page->mapping));
2136
2137                 if (!new_jh) {
2138                         jbd_unlock_bh_journal_head(bh);
2139                         goto repeat;
2140                 }
2141
2142                 jh = new_jh;
2143                 new_jh = NULL;          /* We consumed it */
2144                 set_buffer_jbd(bh);
2145                 bh->b_private = jh;
2146                 jh->b_bh = bh;
2147                 get_bh(bh);
2148                 BUFFER_TRACE(bh, "added journal_head");
2149         }
2150         jh->b_jcount++;
2151         jbd_unlock_bh_journal_head(bh);
2152         if (new_jh)
2153                 journal_free_journal_head(new_jh);
2154         return bh->b_private;
2155 }
2156
2157 /*
2158  * Grab a ref against this buffer_head's journal_head.  If it ended up not
2159  * having a journal_head, return NULL
2160  */
2161 struct journal_head *jbd2_journal_grab_journal_head(struct buffer_head *bh)
2162 {
2163         struct journal_head *jh = NULL;
2164
2165         jbd_lock_bh_journal_head(bh);
2166         if (buffer_jbd(bh)) {
2167                 jh = bh2jh(bh);
2168                 jh->b_jcount++;
2169         }
2170         jbd_unlock_bh_journal_head(bh);
2171         return jh;
2172 }
2173
2174 static void __journal_remove_journal_head(struct buffer_head *bh)
2175 {
2176         struct journal_head *jh = bh2jh(bh);
2177
2178         J_ASSERT_JH(jh, jh->b_jcount >= 0);
2179         J_ASSERT_JH(jh, jh->b_transaction == NULL);
2180         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2181         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
2182         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
2183         J_ASSERT_BH(bh, buffer_jbd(bh));
2184         J_ASSERT_BH(bh, jh2bh(jh) == bh);
2185         BUFFER_TRACE(bh, "remove journal_head");
2186         if (jh->b_frozen_data) {
2187                 printk(KERN_WARNING "%s: freeing b_frozen_data\n", __func__);
2188                 jbd2_free(jh->b_frozen_data, bh->b_size);
2189         }
2190         if (jh->b_committed_data) {
2191                 printk(KERN_WARNING "%s: freeing b_committed_data\n", __func__);
2192                 jbd2_free(jh->b_committed_data, bh->b_size);
2193         }
2194         bh->b_private = NULL;
2195         jh->b_bh = NULL;        /* debug, really */
2196         clear_buffer_jbd(bh);
2197         journal_free_journal_head(jh);
2198 }
2199
2200 /*
2201  * Drop a reference on the passed journal_head.  If it fell to zero then
2202  * release the journal_head from the buffer_head.
2203  */
2204 void jbd2_journal_put_journal_head(struct journal_head *jh)
2205 {
2206         struct buffer_head *bh = jh2bh(jh);
2207
2208         jbd_lock_bh_journal_head(bh);
2209         J_ASSERT_JH(jh, jh->b_jcount > 0);
2210         --jh->b_jcount;
2211         if (!jh->b_jcount) {
2212                 __journal_remove_journal_head(bh);
2213                 jbd_unlock_bh_journal_head(bh);
2214                 __brelse(bh);
2215         } else
2216                 jbd_unlock_bh_journal_head(bh);
2217 }
2218
2219 /*
2220  * Initialize jbd inode head
2221  */
2222 void jbd2_journal_init_jbd_inode(struct jbd2_inode *jinode, struct inode *inode)
2223 {
2224         jinode->i_transaction = NULL;
2225         jinode->i_next_transaction = NULL;
2226         jinode->i_vfs_inode = inode;
2227         jinode->i_flags = 0;
2228         INIT_LIST_HEAD(&jinode->i_list);
2229 }
2230
2231 /*
2232  * Function to be called before we start removing inode from memory (i.e.,
2233  * clear_inode() is a fine place to be called from). It removes inode from
2234  * transaction's lists.
2235  */
2236 void jbd2_journal_release_jbd_inode(journal_t *journal,
2237                                     struct jbd2_inode *jinode)
2238 {
2239         if (!journal)
2240                 return;
2241 restart:
2242         spin_lock(&journal->j_list_lock);
2243         /* Is commit writing out inode - we have to wait */
2244         if (test_bit(__JI_COMMIT_RUNNING, &jinode->i_flags)) {
2245                 wait_queue_head_t *wq;
2246                 DEFINE_WAIT_BIT(wait, &jinode->i_flags, __JI_COMMIT_RUNNING);
2247                 wq = bit_waitqueue(&jinode->i_flags, __JI_COMMIT_RUNNING);
2248                 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2249                 spin_unlock(&journal->j_list_lock);
2250                 schedule();
2251                 finish_wait(wq, &wait.wait);
2252                 goto restart;
2253         }
2254
2255         if (jinode->i_transaction) {
2256                 list_del(&jinode->i_list);
2257                 jinode->i_transaction = NULL;
2258         }
2259         spin_unlock(&journal->j_list_lock);
2260 }
2261
2262 /*
2263  * debugfs tunables
2264  */
2265 #ifdef CONFIG_JBD2_DEBUG
2266 u8 jbd2_journal_enable_debug __read_mostly;
2267 EXPORT_SYMBOL(jbd2_journal_enable_debug);
2268
2269 #define JBD2_DEBUG_NAME "jbd2-debug"
2270
2271 static struct dentry *jbd2_debugfs_dir;
2272 static struct dentry *jbd2_debug;
2273
2274 static void __init jbd2_create_debugfs_entry(void)
2275 {
2276         jbd2_debugfs_dir = debugfs_create_dir("jbd2", NULL);
2277         if (jbd2_debugfs_dir)
2278                 jbd2_debug = debugfs_create_u8(JBD2_DEBUG_NAME,
2279                                                S_IRUGO | S_IWUSR,
2280                                                jbd2_debugfs_dir,
2281                                                &jbd2_journal_enable_debug);
2282 }
2283
2284 static void __exit jbd2_remove_debugfs_entry(void)
2285 {
2286         debugfs_remove(jbd2_debug);
2287         debugfs_remove(jbd2_debugfs_dir);
2288 }
2289
2290 #else
2291
2292 static void __init jbd2_create_debugfs_entry(void)
2293 {
2294 }
2295
2296 static void __exit jbd2_remove_debugfs_entry(void)
2297 {
2298 }
2299
2300 #endif
2301
2302 #ifdef CONFIG_PROC_FS
2303
2304 #define JBD2_STATS_PROC_NAME "fs/jbd2"
2305
2306 static void __init jbd2_create_jbd_stats_proc_entry(void)
2307 {
2308         proc_jbd2_stats = proc_mkdir(JBD2_STATS_PROC_NAME, NULL);
2309 }
2310
2311 static void __exit jbd2_remove_jbd_stats_proc_entry(void)
2312 {
2313         if (proc_jbd2_stats)
2314                 remove_proc_entry(JBD2_STATS_PROC_NAME, NULL);
2315 }
2316
2317 #else
2318
2319 #define jbd2_create_jbd_stats_proc_entry() do {} while (0)
2320 #define jbd2_remove_jbd_stats_proc_entry() do {} while (0)
2321
2322 #endif
2323
2324 struct kmem_cache *jbd2_handle_cache, *jbd2_inode_cache;
2325
2326 static int __init journal_init_handle_cache(void)
2327 {
2328         jbd2_handle_cache = KMEM_CACHE(jbd2_journal_handle, SLAB_TEMPORARY);
2329         if (jbd2_handle_cache == NULL) {
2330                 printk(KERN_EMERG "JBD2: failed to create handle cache\n");
2331                 return -ENOMEM;
2332         }
2333         jbd2_inode_cache = KMEM_CACHE(jbd2_inode, 0);
2334         if (jbd2_inode_cache == NULL) {
2335                 printk(KERN_EMERG "JBD2: failed to create inode cache\n");
2336                 kmem_cache_destroy(jbd2_handle_cache);
2337                 return -ENOMEM;
2338         }
2339         return 0;
2340 }
2341
2342 static void jbd2_journal_destroy_handle_cache(void)
2343 {
2344         if (jbd2_handle_cache)
2345                 kmem_cache_destroy(jbd2_handle_cache);
2346         if (jbd2_inode_cache)
2347                 kmem_cache_destroy(jbd2_inode_cache);
2348
2349 }
2350
2351 /*
2352  * Module startup and shutdown
2353  */
2354
2355 static int __init journal_init_caches(void)
2356 {
2357         int ret;
2358
2359         ret = jbd2_journal_init_revoke_caches();
2360         if (ret == 0)
2361                 ret = journal_init_jbd2_journal_head_cache();
2362         if (ret == 0)
2363                 ret = journal_init_handle_cache();
2364         return ret;
2365 }
2366
2367 static void jbd2_journal_destroy_caches(void)
2368 {
2369         jbd2_journal_destroy_revoke_caches();
2370         jbd2_journal_destroy_jbd2_journal_head_cache();
2371         jbd2_journal_destroy_handle_cache();
2372         jbd2_journal_destroy_slabs();
2373 }
2374
2375 static int __init journal_init(void)
2376 {
2377         int ret;
2378
2379         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
2380
2381         ret = journal_init_caches();
2382         if (ret == 0) {
2383                 jbd2_create_debugfs_entry();
2384                 jbd2_create_jbd_stats_proc_entry();
2385         } else {
2386                 jbd2_journal_destroy_caches();
2387         }
2388         return ret;
2389 }
2390
2391 static void __exit journal_exit(void)
2392 {
2393 #ifdef CONFIG_JBD2_DEBUG
2394         int n = atomic_read(&nr_journal_heads);
2395         if (n)
2396                 printk(KERN_EMERG "JBD2: leaked %d journal_heads!\n", n);
2397 #endif
2398         jbd2_remove_debugfs_entry();
2399         jbd2_remove_jbd_stats_proc_entry();
2400         jbd2_journal_destroy_caches();
2401 }
2402
2403 MODULE_LICENSE("GPL");
2404 module_init(journal_init);
2405 module_exit(journal_exit);
2406