2 * linux/fs/jbd2/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58 * buffer has not been revoked, and cancel_revoke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/jbd2.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>
94 #include <linux/log2.h>
95 #include <linux/hash.h>
98 static struct kmem_cache
*jbd2_revoke_record_cache
;
99 static struct kmem_cache
*jbd2_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 jbd2_revoke_record_s
107 struct list_head hash
;
108 tid_t sequence
; /* Used for recovery only */
109 unsigned long long blocknr
;
113 /* The revoke table is just a simple hash table of revoke records. */
114 struct jbd2_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
*,
127 struct buffer_head
**, int *,
128 struct jbd2_revoke_record_s
*, int);
129 static void flush_descriptor(journal_t
*, struct buffer_head
*, int, int);
132 /* Utility functions to maintain the revoke table */
134 static inline int hash(journal_t
*journal
, unsigned long long block
)
136 return hash_64(block
, journal
->j_revoke
->hash_shift
);
139 static int insert_revoke_hash(journal_t
*journal
, unsigned long long blocknr
,
142 struct list_head
*hash_list
;
143 struct jbd2_revoke_record_s
*record
;
146 record
= kmem_cache_alloc(jbd2_revoke_record_cache
, GFP_NOFS
);
150 record
->sequence
= seq
;
151 record
->blocknr
= blocknr
;
152 hash_list
= &journal
->j_revoke
->hash_table
[hash(journal
, blocknr
)];
153 spin_lock(&journal
->j_revoke_lock
);
154 list_add(&record
->hash
, hash_list
);
155 spin_unlock(&journal
->j_revoke_lock
);
159 if (!journal_oom_retry
)
161 jbd_debug(1, "ENOMEM in %s, retrying\n", __func__
);
166 /* Find a revoke record in the journal's hash table. */
168 static struct jbd2_revoke_record_s
*find_revoke_record(journal_t
*journal
,
169 unsigned long long blocknr
)
171 struct list_head
*hash_list
;
172 struct jbd2_revoke_record_s
*record
;
174 hash_list
= &journal
->j_revoke
->hash_table
[hash(journal
, blocknr
)];
176 spin_lock(&journal
->j_revoke_lock
);
177 record
= (struct jbd2_revoke_record_s
*) hash_list
->next
;
178 while (&(record
->hash
) != hash_list
) {
179 if (record
->blocknr
== blocknr
) {
180 spin_unlock(&journal
->j_revoke_lock
);
183 record
= (struct jbd2_revoke_record_s
*) record
->hash
.next
;
185 spin_unlock(&journal
->j_revoke_lock
);
189 void jbd2_journal_destroy_revoke_caches(void)
191 if (jbd2_revoke_record_cache
) {
192 kmem_cache_destroy(jbd2_revoke_record_cache
);
193 jbd2_revoke_record_cache
= NULL
;
195 if (jbd2_revoke_table_cache
) {
196 kmem_cache_destroy(jbd2_revoke_table_cache
);
197 jbd2_revoke_table_cache
= NULL
;
201 int __init
jbd2_journal_init_revoke_caches(void)
203 J_ASSERT(!jbd2_revoke_record_cache
);
204 J_ASSERT(!jbd2_revoke_table_cache
);
206 jbd2_revoke_record_cache
= KMEM_CACHE(jbd2_revoke_record_s
,
207 SLAB_HWCACHE_ALIGN
|SLAB_TEMPORARY
);
208 if (!jbd2_revoke_record_cache
)
209 goto record_cache_failure
;
211 jbd2_revoke_table_cache
= KMEM_CACHE(jbd2_revoke_table_s
,
213 if (!jbd2_revoke_table_cache
)
214 goto table_cache_failure
;
217 jbd2_journal_destroy_revoke_caches();
218 record_cache_failure
:
222 static struct jbd2_revoke_table_s
*jbd2_journal_init_revoke_table(int hash_size
)
226 struct jbd2_revoke_table_s
*table
;
228 table
= kmem_cache_alloc(jbd2_revoke_table_cache
, GFP_KERNEL
);
232 while((tmp
>>= 1UL) != 0UL)
235 table
->hash_size
= hash_size
;
236 table
->hash_shift
= shift
;
238 kmalloc(hash_size
* sizeof(struct list_head
), GFP_KERNEL
);
239 if (!table
->hash_table
) {
240 kmem_cache_free(jbd2_revoke_table_cache
, table
);
245 for (tmp
= 0; tmp
< hash_size
; tmp
++)
246 INIT_LIST_HEAD(&table
->hash_table
[tmp
]);
252 static void jbd2_journal_destroy_revoke_table(struct jbd2_revoke_table_s
*table
)
255 struct list_head
*hash_list
;
257 for (i
= 0; i
< table
->hash_size
; i
++) {
258 hash_list
= &table
->hash_table
[i
];
259 J_ASSERT(list_empty(hash_list
));
262 kfree(table
->hash_table
);
263 kmem_cache_free(jbd2_revoke_table_cache
, table
);
266 /* Initialise the revoke table for a given journal to a given size. */
267 int jbd2_journal_init_revoke(journal_t
*journal
, int hash_size
)
269 J_ASSERT(journal
->j_revoke_table
[0] == NULL
);
270 J_ASSERT(is_power_of_2(hash_size
));
272 journal
->j_revoke_table
[0] = jbd2_journal_init_revoke_table(hash_size
);
273 if (!journal
->j_revoke_table
[0])
276 journal
->j_revoke_table
[1] = jbd2_journal_init_revoke_table(hash_size
);
277 if (!journal
->j_revoke_table
[1])
280 journal
->j_revoke
= journal
->j_revoke_table
[1];
282 spin_lock_init(&journal
->j_revoke_lock
);
287 jbd2_journal_destroy_revoke_table(journal
->j_revoke_table
[0]);
292 /* Destroy a journal's revoke table. The table must already be empty! */
293 void jbd2_journal_destroy_revoke(journal_t
*journal
)
295 journal
->j_revoke
= NULL
;
296 if (journal
->j_revoke_table
[0])
297 jbd2_journal_destroy_revoke_table(journal
->j_revoke_table
[0]);
298 if (journal
->j_revoke_table
[1])
299 jbd2_journal_destroy_revoke_table(journal
->j_revoke_table
[1]);
306 * jbd2_journal_revoke: revoke a given buffer_head from the journal. This
307 * prevents the block from being replayed during recovery if we take a
308 * crash after this current transaction commits. Any subsequent
309 * metadata writes of the buffer in this transaction cancel the
312 * Note that this call may block --- it is up to the caller to make
313 * sure that there are no further calls to journal_write_metadata
314 * before the revoke is complete. In ext3, this implies calling the
315 * revoke before clearing the block bitmap when we are deleting
318 * Revoke performs a jbd2_journal_forget on any buffer_head passed in as a
319 * parameter, but does _not_ forget the buffer_head if the bh was only
322 * bh_in may not be a journalled buffer - it may have come off
323 * the hash tables without an attached journal_head.
325 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
329 int jbd2_journal_revoke(handle_t
*handle
, unsigned long long blocknr
,
330 struct buffer_head
*bh_in
)
332 struct buffer_head
*bh
= NULL
;
334 struct block_device
*bdev
;
339 BUFFER_TRACE(bh_in
, "enter");
341 journal
= handle
->h_transaction
->t_journal
;
342 if (!jbd2_journal_set_features(journal
, 0, 0, JBD2_FEATURE_INCOMPAT_REVOKE
)){
343 J_ASSERT (!"Cannot set revoke feature!");
347 bdev
= journal
->j_fs_dev
;
351 bh
= __find_get_block(bdev
, blocknr
, journal
->j_blocksize
);
353 BUFFER_TRACE(bh
, "found on hash");
355 #ifdef JBD2_EXPENSIVE_CHECKING
357 struct buffer_head
*bh2
;
359 /* If there is a different buffer_head lying around in
360 * memory anywhere... */
361 bh2
= __find_get_block(bdev
, blocknr
, journal
->j_blocksize
);
363 /* ... and it has RevokeValid status... */
364 if (bh2
!= bh
&& buffer_revokevalid(bh2
))
365 /* ...then it better be revoked too,
366 * since it's illegal to create a revoke
367 * record against a buffer_head which is
368 * not marked revoked --- that would
369 * risk missing a subsequent revoke
371 J_ASSERT_BH(bh2
, buffer_revoked(bh2
));
377 /* We really ought not ever to revoke twice in a row without
378 first having the revoke cancelled: it's illegal to free a
379 block twice without allocating it in between! */
381 if (!J_EXPECT_BH(bh
, !buffer_revoked(bh
),
382 "inconsistent data on disk")) {
387 set_buffer_revoked(bh
);
388 set_buffer_revokevalid(bh
);
390 BUFFER_TRACE(bh_in
, "call jbd2_journal_forget");
391 jbd2_journal_forget(handle
, bh_in
);
393 BUFFER_TRACE(bh
, "call brelse");
398 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr
, bh_in
);
399 err
= insert_revoke_hash(journal
, blocknr
,
400 handle
->h_transaction
->t_tid
);
401 BUFFER_TRACE(bh_in
, "exit");
406 * Cancel an outstanding revoke. For use only internally by the
407 * journaling code (called from jbd2_journal_get_write_access).
409 * We trust buffer_revoked() on the buffer if the buffer is already
410 * being journaled: if there is no revoke pending on the buffer, then we
411 * don't do anything here.
413 * This would break if it were possible for a buffer to be revoked and
414 * discarded, and then reallocated within the same transaction. In such
415 * a case we would have lost the revoked bit, but when we arrived here
416 * the second time we would still have a pending revoke to cancel. So,
417 * do not trust the Revoked bit on buffers unless RevokeValid is also
420 int jbd2_journal_cancel_revoke(handle_t
*handle
, struct journal_head
*jh
)
422 struct jbd2_revoke_record_s
*record
;
423 journal_t
*journal
= handle
->h_transaction
->t_journal
;
425 int did_revoke
= 0; /* akpm: debug */
426 struct buffer_head
*bh
= jh2bh(jh
);
428 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh
);
430 /* Is the existing Revoke bit valid? If so, we trust it, and
431 * only perform the full cancel if the revoke bit is set. If
432 * not, we can't trust the revoke bit, and we need to do the
433 * full search for a revoke record. */
434 if (test_set_buffer_revokevalid(bh
)) {
435 need_cancel
= test_clear_buffer_revoked(bh
);
438 clear_buffer_revoked(bh
);
442 record
= find_revoke_record(journal
, bh
->b_blocknr
);
444 jbd_debug(4, "cancelled existing revoke on "
445 "blocknr %llu\n", (unsigned long long)bh
->b_blocknr
);
446 spin_lock(&journal
->j_revoke_lock
);
447 list_del(&record
->hash
);
448 spin_unlock(&journal
->j_revoke_lock
);
449 kmem_cache_free(jbd2_revoke_record_cache
, record
);
454 #ifdef JBD2_EXPENSIVE_CHECKING
455 /* There better not be one left behind by now! */
456 record
= find_revoke_record(journal
, bh
->b_blocknr
);
457 J_ASSERT_JH(jh
, record
== NULL
);
460 /* Finally, have we just cleared revoke on an unhashed
461 * buffer_head? If so, we'd better make sure we clear the
462 * revoked status on any hashed alias too, otherwise the revoke
463 * state machine will get very upset later on. */
465 struct buffer_head
*bh2
;
466 bh2
= __find_get_block(bh
->b_bdev
, bh
->b_blocknr
, bh
->b_size
);
469 clear_buffer_revoked(bh2
);
477 * journal_clear_revoked_flag clears revoked flag of buffers in
478 * revoke table to reflect there is no revoked buffers in the next
479 * transaction which is going to be started.
481 void jbd2_clear_buffer_revoked_flags(journal_t
*journal
)
483 struct jbd2_revoke_table_s
*revoke
= journal
->j_revoke
;
486 for (i
= 0; i
< revoke
->hash_size
; i
++) {
487 struct list_head
*hash_list
;
488 struct list_head
*list_entry
;
489 hash_list
= &revoke
->hash_table
[i
];
491 list_for_each(list_entry
, hash_list
) {
492 struct jbd2_revoke_record_s
*record
;
493 struct buffer_head
*bh
;
494 record
= (struct jbd2_revoke_record_s
*)list_entry
;
495 bh
= __find_get_block(journal
->j_fs_dev
,
497 journal
->j_blocksize
);
499 clear_buffer_revoked(bh
);
506 /* journal_switch_revoke table select j_revoke for next transaction
507 * we do not want to suspend any processing until all revokes are
510 void jbd2_journal_switch_revoke_table(journal_t
*journal
)
514 if (journal
->j_revoke
== journal
->j_revoke_table
[0])
515 journal
->j_revoke
= journal
->j_revoke_table
[1];
517 journal
->j_revoke
= journal
->j_revoke_table
[0];
519 for (i
= 0; i
< journal
->j_revoke
->hash_size
; i
++)
520 INIT_LIST_HEAD(&journal
->j_revoke
->hash_table
[i
]);
524 * Write revoke records to the journal for all entries in the current
525 * revoke hash, deleting the entries as we go.
527 void jbd2_journal_write_revoke_records(journal_t
*journal
,
528 transaction_t
*transaction
,
529 struct list_head
*log_bufs
,
532 struct buffer_head
*descriptor
;
533 struct jbd2_revoke_record_s
*record
;
534 struct jbd2_revoke_table_s
*revoke
;
535 struct list_head
*hash_list
;
536 int i
, offset
, count
;
542 /* select revoke table for committing transaction */
543 revoke
= journal
->j_revoke
== journal
->j_revoke_table
[0] ?
544 journal
->j_revoke_table
[1] : journal
->j_revoke_table
[0];
546 for (i
= 0; i
< revoke
->hash_size
; i
++) {
547 hash_list
= &revoke
->hash_table
[i
];
549 while (!list_empty(hash_list
)) {
550 record
= (struct jbd2_revoke_record_s
*)
552 write_one_revoke_record(journal
, transaction
, log_bufs
,
553 &descriptor
, &offset
,
556 list_del(&record
->hash
);
557 kmem_cache_free(jbd2_revoke_record_cache
, record
);
561 flush_descriptor(journal
, descriptor
, offset
, write_op
);
562 jbd_debug(1, "Wrote %d revoke records\n", count
);
566 * Write out one revoke record. We need to create a new descriptor
567 * block if the old one is full or if we have not already created one.
570 static void write_one_revoke_record(journal_t
*journal
,
571 transaction_t
*transaction
,
572 struct list_head
*log_bufs
,
573 struct buffer_head
**descriptorp
,
575 struct jbd2_revoke_record_s
*record
,
579 struct buffer_head
*descriptor
;
581 journal_header_t
*header
;
583 /* If we are already aborting, this all becomes a noop. We
584 still need to go round the loop in
585 jbd2_journal_write_revoke_records in order to free all of the
586 revoke records: only the IO to the journal is omitted. */
587 if (is_journal_aborted(journal
))
590 descriptor
= *descriptorp
;
593 /* Do we need to leave space at the end for a checksum? */
594 if (jbd2_journal_has_csum_v2or3(journal
))
595 csum_size
= sizeof(struct jbd2_journal_revoke_tail
);
597 /* Make sure we have a descriptor with space left for the record */
599 if (offset
>= journal
->j_blocksize
- csum_size
) {
600 flush_descriptor(journal
, descriptor
, offset
, write_op
);
606 descriptor
= jbd2_journal_get_descriptor_buffer(journal
);
609 header
= (journal_header_t
*)descriptor
->b_data
;
610 header
->h_magic
= cpu_to_be32(JBD2_MAGIC_NUMBER
);
611 header
->h_blocktype
= cpu_to_be32(JBD2_REVOKE_BLOCK
);
612 header
->h_sequence
= cpu_to_be32(transaction
->t_tid
);
614 /* Record it so that we can wait for IO completion later */
615 BUFFER_TRACE(descriptor
, "file in log_bufs");
616 jbd2_file_log_bh(log_bufs
, descriptor
);
618 offset
= sizeof(jbd2_journal_revoke_header_t
);
619 *descriptorp
= descriptor
;
622 if (JBD2_HAS_INCOMPAT_FEATURE(journal
, JBD2_FEATURE_INCOMPAT_64BIT
)) {
623 * ((__be64
*)(&descriptor
->b_data
[offset
])) =
624 cpu_to_be64(record
->blocknr
);
628 * ((__be32
*)(&descriptor
->b_data
[offset
])) =
629 cpu_to_be32(record
->blocknr
);
636 static void jbd2_revoke_csum_set(journal_t
*j
, struct buffer_head
*bh
)
638 struct jbd2_journal_revoke_tail
*tail
;
641 if (!jbd2_journal_has_csum_v2or3(j
))
644 tail
= (struct jbd2_journal_revoke_tail
*)(bh
->b_data
+ j
->j_blocksize
-
645 sizeof(struct jbd2_journal_revoke_tail
));
646 tail
->r_checksum
= 0;
647 csum
= jbd2_chksum(j
, j
->j_csum_seed
, bh
->b_data
, j
->j_blocksize
);
648 tail
->r_checksum
= cpu_to_be32(csum
);
652 * Flush a revoke descriptor out to the journal. If we are aborting,
653 * this is a noop; otherwise we are generating a buffer which needs to
654 * be waited for during commit, so it has to go onto the appropriate
655 * journal buffer list.
658 static void flush_descriptor(journal_t
*journal
,
659 struct buffer_head
*descriptor
,
660 int offset
, int write_op
)
662 jbd2_journal_revoke_header_t
*header
;
664 if (is_journal_aborted(journal
)) {
669 header
= (jbd2_journal_revoke_header_t
*)descriptor
->b_data
;
670 header
->r_count
= cpu_to_be32(offset
);
671 jbd2_revoke_csum_set(journal
, descriptor
);
673 set_buffer_jwrite(descriptor
);
674 BUFFER_TRACE(descriptor
, "write");
675 set_buffer_dirty(descriptor
);
676 write_dirty_buffer(descriptor
, write_op
);
681 * Revoke support for recovery.
683 * Recovery needs to be able to:
685 * record all revoke records, including the tid of the latest instance
686 * of each revoke in the journal
688 * check whether a given block in a given transaction should be replayed
689 * (ie. has not been revoked by a revoke record in that or a subsequent
692 * empty the revoke table after recovery.
696 * First, setting revoke records. We create a new revoke record for
697 * every block ever revoked in the log as we scan it for recovery, and
698 * we update the existing records if we find multiple revokes for a
702 int jbd2_journal_set_revoke(journal_t
*journal
,
703 unsigned long long blocknr
,
706 struct jbd2_revoke_record_s
*record
;
708 record
= find_revoke_record(journal
, blocknr
);
710 /* If we have multiple occurrences, only record the
711 * latest sequence number in the hashed record */
712 if (tid_gt(sequence
, record
->sequence
))
713 record
->sequence
= sequence
;
716 return insert_revoke_hash(journal
, blocknr
, sequence
);
720 * Test revoke records. For a given block referenced in the log, has
721 * that block been revoked? A revoke record with a given transaction
722 * sequence number revokes all blocks in that transaction and earlier
723 * ones, but later transactions still need replayed.
726 int jbd2_journal_test_revoke(journal_t
*journal
,
727 unsigned long long blocknr
,
730 struct jbd2_revoke_record_s
*record
;
732 record
= find_revoke_record(journal
, blocknr
);
735 if (tid_gt(sequence
, record
->sequence
))
741 * Finally, once recovery is over, we need to clear the revoke table so
742 * that it can be reused by the running filesystem.
745 void jbd2_journal_clear_revoke(journal_t
*journal
)
748 struct list_head
*hash_list
;
749 struct jbd2_revoke_record_s
*record
;
750 struct jbd2_revoke_table_s
*revoke
;
752 revoke
= journal
->j_revoke
;
754 for (i
= 0; i
< revoke
->hash_size
; i
++) {
755 hash_list
= &revoke
->hash_table
[i
];
756 while (!list_empty(hash_list
)) {
757 record
= (struct jbd2_revoke_record_s
*) hash_list
->next
;
758 list_del(&record
->hash
);
759 kmem_cache_free(jbd2_revoke_record_cache
, record
);