2 * linux/fs/jbd/revoke.c
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 2000
6 * Copyright 2000 Red Hat corp --- All Rights Reserved
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.
12 * Journal revoke routines for the generic filesystem journaling code;
13 * part of the ext2fs journaling system.
15 * Revoke is the mechanism used to prevent old log records for deleted
16 * metadata from being replayed on top of newer data using the same
17 * blocks. The revoke mechanism is used in two separate places:
19 * + Commit: during commit we write the entire list of the current
20 * transaction's revoked blocks to the journal
22 * + Recovery: during recovery we record the transaction ID of all
23 * revoked blocks. If there are multiple revoke records in the log
24 * for a single block, only the last one counts, and if there is a log
25 * entry for a block beyond the last revoke, then that log entry still
28 * We can get interactions between revokes and new log data within a
31 * Block is revoked and then journaled:
32 * The desired end result is the journaling of the new block, so we
33 * cancel the revoke before the transaction commits.
35 * Block is journaled and then revoked:
36 * The revoke must take precedence over the write of the block, so we
37 * need either to cancel the journal entry or to write the revoke
38 * later in the log than the log block. In this case, we choose the
39 * latter: journaling a block cancels any revoke record for that block
40 * in the current transaction, so any revoke for that block in the
41 * transaction must have happened after the block was journaled and so
42 * the revoke must take precedence.
44 * Block is revoked and then written as data:
45 * The data write is allowed to succeed, but the revoke is _not_
46 * cancelled. We still need to prevent old log records from
47 * overwriting the new data. We don't even need to clear the revoke
50 * We cache revoke status of a buffer in the current transaction in b_states
51 * bits. As the name says, revokevalid flag indicates that the cached revoke
52 * status of a buffer is valid and we can rely on the cached status.
54 * Revoke information on buffers is a tri-state value:
56 * RevokeValid clear: no cached revoke status, need to look it up
57 * RevokeValid set, Revoked clear:
58 * buffer has not been revoked, and cancel_revoke
60 * RevokeValid set, Revoked set:
61 * buffer has been revoked.
64 * We keep two hash tables of revoke records. One hashtable belongs to the
65 * running transaction (is pointed to by journal->j_revoke), the other one
66 * belongs to the committing transaction. Accesses to the second hash table
67 * happen only from the kjournald and no other thread touches this table. Also
68 * journal_switch_revoke_table() which switches which hashtable belongs to the
69 * running and which to the committing transaction is called only from
70 * kjournald. Therefore we need no locks when accessing the hashtable belonging
71 * to the committing transaction.
73 * All users operating on the hash table belonging to the running transaction
74 * have a handle to the transaction. Therefore they are safe from kjournald
75 * switching hash tables under them. For operations on the lists of entries in
76 * the hash table j_revoke_lock is used.
78 * Finally, also replay code uses the hash tables but at this moment no one else
79 * can touch them (filesystem isn't mounted yet) and hence no locking is
86 #include <linux/time.h>
88 #include <linux/jbd.h>
89 #include <linux/errno.h>
90 #include <linux/slab.h>
91 #include <linux/list.h>
92 #include <linux/init.h>
93 #include <linux/bio.h>
95 #include <linux/log2.h>
96 #include <linux/hash.h>
98 static struct kmem_cache
*revoke_record_cache
;
99 static struct kmem_cache
*revoke_table_cache
;
101 /* Each revoke record represents one single revoked block. During
102 journal replay, this involves recording the transaction ID of the
103 last transaction to revoke this block. */
105 struct jbd_revoke_record_s
107 struct list_head hash
;
108 tid_t sequence
; /* Used for recovery only */
109 unsigned int blocknr
;
113 /* The revoke table is just a simple hash table of revoke records. */
114 struct jbd_revoke_table_s
116 /* It is conceivable that we might want a larger hash table
117 * for recovery. Must be a power of two. */
120 struct list_head
*hash_table
;
125 static void write_one_revoke_record(journal_t
*, transaction_t
*,
126 struct journal_head
**, int *,
127 struct jbd_revoke_record_s
*, int);
128 static void flush_descriptor(journal_t
*, struct journal_head
*, int, int);
131 /* Utility functions to maintain the revoke table */
133 static inline int hash(journal_t
*journal
, unsigned int block
)
135 struct jbd_revoke_table_s
*table
= journal
->j_revoke
;
137 return hash_32(block
, table
->hash_shift
);
140 static int insert_revoke_hash(journal_t
*journal
, unsigned int blocknr
,
143 struct list_head
*hash_list
;
144 struct jbd_revoke_record_s
*record
;
147 record
= kmem_cache_alloc(revoke_record_cache
, GFP_NOFS
);
151 record
->sequence
= seq
;
152 record
->blocknr
= blocknr
;
153 hash_list
= &journal
->j_revoke
->hash_table
[hash(journal
, blocknr
)];
154 spin_lock(&journal
->j_revoke_lock
);
155 list_add(&record
->hash
, hash_list
);
156 spin_unlock(&journal
->j_revoke_lock
);
160 if (!journal_oom_retry
)
162 jbd_debug(1, "ENOMEM in %s, retrying\n", __func__
);
167 /* Find a revoke record in the journal's hash table. */
169 static struct jbd_revoke_record_s
*find_revoke_record(journal_t
*journal
,
170 unsigned int blocknr
)
172 struct list_head
*hash_list
;
173 struct jbd_revoke_record_s
*record
;
175 hash_list
= &journal
->j_revoke
->hash_table
[hash(journal
, blocknr
)];
177 spin_lock(&journal
->j_revoke_lock
);
178 record
= (struct jbd_revoke_record_s
*) hash_list
->next
;
179 while (&(record
->hash
) != hash_list
) {
180 if (record
->blocknr
== blocknr
) {
181 spin_unlock(&journal
->j_revoke_lock
);
184 record
= (struct jbd_revoke_record_s
*) record
->hash
.next
;
186 spin_unlock(&journal
->j_revoke_lock
);
190 void journal_destroy_revoke_caches(void)
192 if (revoke_record_cache
) {
193 kmem_cache_destroy(revoke_record_cache
);
194 revoke_record_cache
= NULL
;
196 if (revoke_table_cache
) {
197 kmem_cache_destroy(revoke_table_cache
);
198 revoke_table_cache
= NULL
;
202 int __init
journal_init_revoke_caches(void)
204 J_ASSERT(!revoke_record_cache
);
205 J_ASSERT(!revoke_table_cache
);
207 revoke_record_cache
= kmem_cache_create("revoke_record",
208 sizeof(struct jbd_revoke_record_s
),
210 SLAB_HWCACHE_ALIGN
|SLAB_TEMPORARY
,
212 if (!revoke_record_cache
)
213 goto record_cache_failure
;
215 revoke_table_cache
= kmem_cache_create("revoke_table",
216 sizeof(struct jbd_revoke_table_s
),
217 0, SLAB_TEMPORARY
, NULL
);
218 if (!revoke_table_cache
)
219 goto table_cache_failure
;
224 journal_destroy_revoke_caches();
225 record_cache_failure
:
229 static struct jbd_revoke_table_s
*journal_init_revoke_table(int hash_size
)
232 struct jbd_revoke_table_s
*table
;
234 table
= kmem_cache_alloc(revoke_table_cache
, GFP_KERNEL
);
238 table
->hash_size
= hash_size
;
239 table
->hash_shift
= ilog2(hash_size
);
241 kmalloc(hash_size
* sizeof(struct list_head
), GFP_KERNEL
);
242 if (!table
->hash_table
) {
243 kmem_cache_free(revoke_table_cache
, table
);
248 for (i
= 0; i
< hash_size
; i
++)
249 INIT_LIST_HEAD(&table
->hash_table
[i
]);
255 static void journal_destroy_revoke_table(struct jbd_revoke_table_s
*table
)
258 struct list_head
*hash_list
;
260 for (i
= 0; i
< table
->hash_size
; i
++) {
261 hash_list
= &table
->hash_table
[i
];
262 J_ASSERT(list_empty(hash_list
));
265 kfree(table
->hash_table
);
266 kmem_cache_free(revoke_table_cache
, table
);
269 /* Initialise the revoke table for a given journal to a given size. */
270 int journal_init_revoke(journal_t
*journal
, int hash_size
)
272 J_ASSERT(journal
->j_revoke_table
[0] == NULL
);
273 J_ASSERT(is_power_of_2(hash_size
));
275 journal
->j_revoke_table
[0] = journal_init_revoke_table(hash_size
);
276 if (!journal
->j_revoke_table
[0])
279 journal
->j_revoke_table
[1] = journal_init_revoke_table(hash_size
);
280 if (!journal
->j_revoke_table
[1])
283 journal
->j_revoke
= journal
->j_revoke_table
[1];
285 spin_lock_init(&journal
->j_revoke_lock
);
290 journal_destroy_revoke_table(journal
->j_revoke_table
[0]);
295 /* Destroy a journal's revoke table. The table must already be empty! */
296 void journal_destroy_revoke(journal_t
*journal
)
298 journal
->j_revoke
= NULL
;
299 if (journal
->j_revoke_table
[0])
300 journal_destroy_revoke_table(journal
->j_revoke_table
[0]);
301 if (journal
->j_revoke_table
[1])
302 journal_destroy_revoke_table(journal
->j_revoke_table
[1]);
309 * journal_revoke: revoke a given buffer_head from the journal. This
310 * prevents the block from being replayed during recovery if we take a
311 * crash after this current transaction commits. Any subsequent
312 * metadata writes of the buffer in this transaction cancel the
315 * Note that this call may block --- it is up to the caller to make
316 * sure that there are no further calls to journal_write_metadata
317 * before the revoke is complete. In ext3, this implies calling the
318 * revoke before clearing the block bitmap when we are deleting
321 * Revoke performs a journal_forget on any buffer_head passed in as a
322 * parameter, but does _not_ forget the buffer_head if the bh was only
325 * bh_in may not be a journalled buffer - it may have come off
326 * the hash tables without an attached journal_head.
328 * If bh_in is non-zero, journal_revoke() will decrement its b_count
332 int journal_revoke(handle_t
*handle
, unsigned int blocknr
,
333 struct buffer_head
*bh_in
)
335 struct buffer_head
*bh
= NULL
;
337 struct block_device
*bdev
;
342 BUFFER_TRACE(bh_in
, "enter");
344 journal
= handle
->h_transaction
->t_journal
;
345 if (!journal_set_features(journal
, 0, 0, JFS_FEATURE_INCOMPAT_REVOKE
)){
346 J_ASSERT (!"Cannot set revoke feature!");
350 bdev
= journal
->j_fs_dev
;
354 bh
= __find_get_block(bdev
, blocknr
, journal
->j_blocksize
);
356 BUFFER_TRACE(bh
, "found on hash");
358 #ifdef JBD_EXPENSIVE_CHECKING
360 struct buffer_head
*bh2
;
362 /* If there is a different buffer_head lying around in
363 * memory anywhere... */
364 bh2
= __find_get_block(bdev
, blocknr
, journal
->j_blocksize
);
366 /* ... and it has RevokeValid status... */
367 if (bh2
!= bh
&& buffer_revokevalid(bh2
))
368 /* ...then it better be revoked too,
369 * since it's illegal to create a revoke
370 * record against a buffer_head which is
371 * not marked revoked --- that would
372 * risk missing a subsequent revoke
374 J_ASSERT_BH(bh2
, buffer_revoked(bh2
));
380 /* We really ought not ever to revoke twice in a row without
381 first having the revoke cancelled: it's illegal to free a
382 block twice without allocating it in between! */
384 if (!J_EXPECT_BH(bh
, !buffer_revoked(bh
),
385 "inconsistent data on disk")) {
390 set_buffer_revoked(bh
);
391 set_buffer_revokevalid(bh
);
393 BUFFER_TRACE(bh_in
, "call journal_forget");
394 journal_forget(handle
, bh_in
);
396 BUFFER_TRACE(bh
, "call brelse");
401 jbd_debug(2, "insert revoke for block %u, bh_in=%p\n", blocknr
, bh_in
);
402 err
= insert_revoke_hash(journal
, blocknr
,
403 handle
->h_transaction
->t_tid
);
404 BUFFER_TRACE(bh_in
, "exit");
409 * Cancel an outstanding revoke. For use only internally by the
410 * journaling code (called from journal_get_write_access).
412 * We trust buffer_revoked() on the buffer if the buffer is already
413 * being journaled: if there is no revoke pending on the buffer, then we
414 * don't do anything here.
416 * This would break if it were possible for a buffer to be revoked and
417 * discarded, and then reallocated within the same transaction. In such
418 * a case we would have lost the revoked bit, but when we arrived here
419 * the second time we would still have a pending revoke to cancel. So,
420 * do not trust the Revoked bit on buffers unless RevokeValid is also
423 int journal_cancel_revoke(handle_t
*handle
, struct journal_head
*jh
)
425 struct jbd_revoke_record_s
*record
;
426 journal_t
*journal
= handle
->h_transaction
->t_journal
;
428 int did_revoke
= 0; /* akpm: debug */
429 struct buffer_head
*bh
= jh2bh(jh
);
431 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh
);
433 /* Is the existing Revoke bit valid? If so, we trust it, and
434 * only perform the full cancel if the revoke bit is set. If
435 * not, we can't trust the revoke bit, and we need to do the
436 * full search for a revoke record. */
437 if (test_set_buffer_revokevalid(bh
)) {
438 need_cancel
= test_clear_buffer_revoked(bh
);
441 clear_buffer_revoked(bh
);
445 record
= find_revoke_record(journal
, bh
->b_blocknr
);
447 jbd_debug(4, "cancelled existing revoke on "
448 "blocknr %llu\n", (unsigned long long)bh
->b_blocknr
);
449 spin_lock(&journal
->j_revoke_lock
);
450 list_del(&record
->hash
);
451 spin_unlock(&journal
->j_revoke_lock
);
452 kmem_cache_free(revoke_record_cache
, record
);
457 #ifdef JBD_EXPENSIVE_CHECKING
458 /* There better not be one left behind by now! */
459 record
= find_revoke_record(journal
, bh
->b_blocknr
);
460 J_ASSERT_JH(jh
, record
== NULL
);
463 /* Finally, have we just cleared revoke on an unhashed
464 * buffer_head? If so, we'd better make sure we clear the
465 * revoked status on any hashed alias too, otherwise the revoke
466 * state machine will get very upset later on. */
468 struct buffer_head
*bh2
;
469 bh2
= __find_get_block(bh
->b_bdev
, bh
->b_blocknr
, bh
->b_size
);
472 clear_buffer_revoked(bh2
);
480 * journal_clear_revoked_flags clears revoked flag of buffers in
481 * revoke table to reflect there is no revoked buffer in the next
482 * transaction which is going to be started.
484 void journal_clear_buffer_revoked_flags(journal_t
*journal
)
486 struct jbd_revoke_table_s
*revoke
= journal
->j_revoke
;
489 for (i
= 0; i
< revoke
->hash_size
; i
++) {
490 struct list_head
*hash_list
;
491 struct list_head
*list_entry
;
492 hash_list
= &revoke
->hash_table
[i
];
494 list_for_each(list_entry
, hash_list
) {
495 struct jbd_revoke_record_s
*record
;
496 struct buffer_head
*bh
;
497 record
= (struct jbd_revoke_record_s
*)list_entry
;
498 bh
= __find_get_block(journal
->j_fs_dev
,
500 journal
->j_blocksize
);
502 clear_buffer_revoked(bh
);
509 /* journal_switch_revoke table select j_revoke for next transaction
510 * we do not want to suspend any processing until all revokes are
513 void journal_switch_revoke_table(journal_t
*journal
)
517 if (journal
->j_revoke
== journal
->j_revoke_table
[0])
518 journal
->j_revoke
= journal
->j_revoke_table
[1];
520 journal
->j_revoke
= journal
->j_revoke_table
[0];
522 for (i
= 0; i
< journal
->j_revoke
->hash_size
; i
++)
523 INIT_LIST_HEAD(&journal
->j_revoke
->hash_table
[i
]);
527 * Write revoke records to the journal for all entries in the current
528 * revoke hash, deleting the entries as we go.
530 void journal_write_revoke_records(journal_t
*journal
,
531 transaction_t
*transaction
, int write_op
)
533 struct journal_head
*descriptor
;
534 struct jbd_revoke_record_s
*record
;
535 struct jbd_revoke_table_s
*revoke
;
536 struct list_head
*hash_list
;
537 int i
, offset
, count
;
543 /* select revoke table for committing transaction */
544 revoke
= journal
->j_revoke
== journal
->j_revoke_table
[0] ?
545 journal
->j_revoke_table
[1] : journal
->j_revoke_table
[0];
547 for (i
= 0; i
< revoke
->hash_size
; i
++) {
548 hash_list
= &revoke
->hash_table
[i
];
550 while (!list_empty(hash_list
)) {
551 record
= (struct jbd_revoke_record_s
*)
553 write_one_revoke_record(journal
, transaction
,
554 &descriptor
, &offset
,
557 list_del(&record
->hash
);
558 kmem_cache_free(revoke_record_cache
, record
);
562 flush_descriptor(journal
, descriptor
, offset
, write_op
);
563 jbd_debug(1, "Wrote %d revoke records\n", count
);
567 * Write out one revoke record. We need to create a new descriptor
568 * block if the old one is full or if we have not already created one.
571 static void write_one_revoke_record(journal_t
*journal
,
572 transaction_t
*transaction
,
573 struct journal_head
**descriptorp
,
575 struct jbd_revoke_record_s
*record
,
578 struct journal_head
*descriptor
;
580 journal_header_t
*header
;
582 /* If we are already aborting, this all becomes a noop. We
583 still need to go round the loop in
584 journal_write_revoke_records in order to free all of the
585 revoke records: only the IO to the journal is omitted. */
586 if (is_journal_aborted(journal
))
589 descriptor
= *descriptorp
;
592 /* Make sure we have a descriptor with space left for the record */
594 if (offset
== journal
->j_blocksize
) {
595 flush_descriptor(journal
, descriptor
, offset
, write_op
);
601 descriptor
= journal_get_descriptor_buffer(journal
);
604 header
= (journal_header_t
*) &jh2bh(descriptor
)->b_data
[0];
605 header
->h_magic
= cpu_to_be32(JFS_MAGIC_NUMBER
);
606 header
->h_blocktype
= cpu_to_be32(JFS_REVOKE_BLOCK
);
607 header
->h_sequence
= cpu_to_be32(transaction
->t_tid
);
609 /* Record it so that we can wait for IO completion later */
610 JBUFFER_TRACE(descriptor
, "file as BJ_LogCtl");
611 journal_file_buffer(descriptor
, transaction
, BJ_LogCtl
);
613 offset
= sizeof(journal_revoke_header_t
);
614 *descriptorp
= descriptor
;
617 * ((__be32
*)(&jh2bh(descriptor
)->b_data
[offset
])) =
618 cpu_to_be32(record
->blocknr
);
624 * Flush a revoke descriptor out to the journal. If we are aborting,
625 * this is a noop; otherwise we are generating a buffer which needs to
626 * be waited for during commit, so it has to go onto the appropriate
627 * journal buffer list.
630 static void flush_descriptor(journal_t
*journal
,
631 struct journal_head
*descriptor
,
632 int offset
, int write_op
)
634 journal_revoke_header_t
*header
;
635 struct buffer_head
*bh
= jh2bh(descriptor
);
637 if (is_journal_aborted(journal
)) {
642 header
= (journal_revoke_header_t
*) jh2bh(descriptor
)->b_data
;
643 header
->r_count
= cpu_to_be32(offset
);
644 set_buffer_jwrite(bh
);
645 BUFFER_TRACE(bh
, "write");
646 set_buffer_dirty(bh
);
647 write_dirty_buffer(bh
, write_op
);
652 * Revoke support for recovery.
654 * Recovery needs to be able to:
656 * record all revoke records, including the tid of the latest instance
657 * of each revoke in the journal
659 * check whether a given block in a given transaction should be replayed
660 * (ie. has not been revoked by a revoke record in that or a subsequent
663 * empty the revoke table after recovery.
667 * First, setting revoke records. We create a new revoke record for
668 * every block ever revoked in the log as we scan it for recovery, and
669 * we update the existing records if we find multiple revokes for a
673 int journal_set_revoke(journal_t
*journal
,
674 unsigned int blocknr
,
677 struct jbd_revoke_record_s
*record
;
679 record
= find_revoke_record(journal
, blocknr
);
681 /* If we have multiple occurrences, only record the
682 * latest sequence number in the hashed record */
683 if (tid_gt(sequence
, record
->sequence
))
684 record
->sequence
= sequence
;
687 return insert_revoke_hash(journal
, blocknr
, sequence
);
691 * Test revoke records. For a given block referenced in the log, has
692 * that block been revoked? A revoke record with a given transaction
693 * sequence number revokes all blocks in that transaction and earlier
694 * ones, but later transactions still need replayed.
697 int journal_test_revoke(journal_t
*journal
,
698 unsigned int blocknr
,
701 struct jbd_revoke_record_s
*record
;
703 record
= find_revoke_record(journal
, blocknr
);
706 if (tid_gt(sequence
, record
->sequence
))
712 * Finally, once recovery is over, we need to clear the revoke table so
713 * that it can be reused by the running filesystem.
716 void journal_clear_revoke(journal_t
*journal
)
719 struct list_head
*hash_list
;
720 struct jbd_revoke_record_s
*record
;
721 struct jbd_revoke_table_s
*revoke
;
723 revoke
= journal
->j_revoke
;
725 for (i
= 0; i
< revoke
->hash_size
; i
++) {
726 hash_list
= &revoke
->hash_table
[i
];
727 while (!list_empty(hash_list
)) {
728 record
= (struct jbd_revoke_record_s
*) hash_list
->next
;
729 list_del(&record
->hash
);
730 kmem_cache_free(revoke_record_cache
, record
);