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58 * buffer has not been revoked, and cancel_revoke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>
97 static lkmem_cache_t
*jbd2_revoke_record_cache
;
98 static lkmem_cache_t
*jbd2_revoke_table_cache
;
100 /* Each revoke record represents one single revoked block. During
101 journal replay, this involves recording the transaction ID of the
102 last transaction to revoke this block. */
104 struct jbd2_revoke_record_s
106 struct list_head hash
;
107 tid_t sequence
; /* Used for recovery only */
108 unsigned long long blocknr
;
112 /* The revoke table is just a simple hash table of revoke records. */
113 struct jbd2_revoke_table_s
115 /* It is conceivable that we might want a larger hash table
116 * for recovery. Must be a power of two. */
119 struct list_head
*hash_table
;
124 static void write_one_revoke_record(journal_t
*, transaction_t
*,
126 struct buffer_head
**, int *,
127 struct jbd2_revoke_record_s
*, int);
128 static void flush_descriptor(journal_t
*, struct buffer_head
*, int, int);
131 /* Utility functions to maintain the revoke table */
133 /* Borrowed from buffer.c: this is a tried and tested block hash function */
134 static inline int hash(journal_t
*journal
, unsigned long long block
)
136 struct jbd2_revoke_table_s
*table
= journal
->j_revoke
;
138 return (hash_64(block
, table
->hash_shift
));
141 static int insert_revoke_hash(journal_t
*journal
, unsigned long long blocknr
,
144 struct list_head
*hash_list
;
145 struct jbd2_revoke_record_s
*record
;
148 record
= kmem_cache_alloc(jbd2_revoke_record_cache
, GFP_NOFS
);
152 record
->sequence
= seq
;
153 record
->blocknr
= blocknr
;
154 hash_list
= &journal
->j_revoke
->hash_table
[hash(journal
, blocknr
)];
155 spin_lock(&journal
->j_revoke_lock
);
156 list_add(&record
->hash
, hash_list
);
157 spin_unlock(&journal
->j_revoke_lock
);
161 if (!journal_oom_retry
)
163 jbd_debug(1, "ENOMEM in %s, retrying\n", __func__
);
168 /* Find a revoke record in the journal's hash table. */
170 static struct jbd2_revoke_record_s
*find_revoke_record(journal_t
*journal
,
171 unsigned long long blocknr
)
173 struct list_head
*hash_list
;
174 struct jbd2_revoke_record_s
*record
;
176 hash_list
= &journal
->j_revoke
->hash_table
[hash(journal
, blocknr
)];
178 spin_lock(&journal
->j_revoke_lock
);
179 record
= (struct jbd2_revoke_record_s
*) hash_list
->next
;
180 while (&(record
->hash
) != hash_list
) {
181 if (record
->blocknr
== blocknr
) {
182 spin_unlock(&journal
->j_revoke_lock
);
185 record
= (struct jbd2_revoke_record_s
*) record
->hash
.next
;
187 spin_unlock(&journal
->j_revoke_lock
);
191 void journal_destroy_revoke_caches(void)
193 if (jbd2_revoke_record_cache
) {
194 kmem_cache_destroy(jbd2_revoke_record_cache
);
195 jbd2_revoke_record_cache
= NULL
;
197 if (jbd2_revoke_table_cache
) {
198 kmem_cache_destroy(jbd2_revoke_table_cache
);
199 jbd2_revoke_table_cache
= NULL
;
203 int __init
journal_init_revoke_caches(void)
205 J_ASSERT(!jbd2_revoke_record_cache
);
206 J_ASSERT(!jbd2_revoke_table_cache
);
208 jbd2_revoke_record_cache
= KMEM_CACHE(jbd2_revoke_record_s
,
209 SLAB_HWCACHE_ALIGN
|SLAB_TEMPORARY
);
210 if (!jbd2_revoke_record_cache
)
211 goto record_cache_failure
;
213 jbd2_revoke_table_cache
= KMEM_CACHE(jbd2_revoke_table_s
,
215 if (!jbd2_revoke_table_cache
)
216 goto table_cache_failure
;
219 journal_destroy_revoke_caches();
220 record_cache_failure
:
224 static struct jbd2_revoke_table_s
*journal_init_revoke_table(int hash_size
)
228 struct jbd2_revoke_table_s
*table
;
230 table
= kmem_cache_alloc(jbd2_revoke_table_cache
, GFP_KERNEL
);
234 while((tmp
>>= 1UL) != 0UL)
237 table
->hash_size
= hash_size
;
238 table
->hash_shift
= shift
;
240 kmalloc(hash_size
* sizeof(struct list_head
), GFP_KERNEL
);
241 if (!table
->hash_table
) {
242 kmem_cache_free(jbd2_revoke_table_cache
, table
);
247 for (tmp
= 0; tmp
< hash_size
; tmp
++)
248 INIT_LIST_HEAD(&table
->hash_table
[tmp
]);
254 static void journal_destroy_revoke_table(struct jbd2_revoke_table_s
*table
)
257 struct list_head
*hash_list
;
259 for (i
= 0; i
< table
->hash_size
; i
++) {
260 hash_list
= &table
->hash_table
[i
];
261 J_ASSERT(list_empty(hash_list
));
264 kfree(table
->hash_table
);
265 kmem_cache_free(jbd2_revoke_table_cache
, table
);
268 /* Initialise the revoke table for a given journal to a given size. */
269 int journal_init_revoke(journal_t
*journal
, int hash_size
)
271 J_ASSERT(journal
->j_revoke_table
[0] == NULL
);
272 J_ASSERT(is_power_of_2(hash_size
));
274 journal
->j_revoke_table
[0] = journal_init_revoke_table(hash_size
);
275 if (!journal
->j_revoke_table
[0])
278 journal
->j_revoke_table
[1] = journal_init_revoke_table(hash_size
);
279 if (!journal
->j_revoke_table
[1])
282 journal
->j_revoke
= journal
->j_revoke_table
[1];
284 spin_lock_init(&journal
->j_revoke_lock
);
289 journal_destroy_revoke_table(journal
->j_revoke_table
[0]);
294 /* Destroy a journal's revoke table. The table must already be empty! */
295 void journal_destroy_revoke(journal_t
*journal
)
297 journal
->j_revoke
= NULL
;
298 if (journal
->j_revoke_table
[0])
299 journal_destroy_revoke_table(journal
->j_revoke_table
[0]);
300 if (journal
->j_revoke_table
[1])
301 journal_destroy_revoke_table(journal
->j_revoke_table
[1]);
308 * journal_revoke: revoke a given buffer_head from the journal. This
309 * prevents the block from being replayed during recovery if we take a
310 * crash after this current transaction commits. Any subsequent
311 * metadata writes of the buffer in this transaction cancel the
314 * Note that this call may block --- it is up to the caller to make
315 * sure that there are no further calls to journal_write_metadata
316 * before the revoke is complete. In ext3, this implies calling the
317 * revoke before clearing the block bitmap when we are deleting
320 * Revoke performs a journal_forget on any buffer_head passed in as a
321 * parameter, but does _not_ forget the buffer_head if the bh was only
324 * bh_in may not be a journalled buffer - it may have come off
325 * the hash tables without an attached journal_head.
327 * If bh_in is non-zero, journal_revoke() will decrement its b_count
331 int journal_revoke(handle_t
*handle
, unsigned long long blocknr
,
332 struct buffer_head
*bh_in
)
334 struct buffer_head
*bh
= NULL
;
336 struct block_device
*bdev
;
341 BUFFER_TRACE(bh_in
, "enter");
343 journal
= handle
->h_transaction
->t_journal
;
344 if (!journal_set_features(journal
, 0, 0, JFS_FEATURE_INCOMPAT_REVOKE
)){
345 J_ASSERT (!"Cannot set revoke feature!");
349 bdev
= journal
->j_fs_dev
;
353 bh
= __find_get_block(bdev
, blocknr
, journal
->j_blocksize
);
355 BUFFER_TRACE(bh
, "found on hash");
357 #ifdef JFS_EXPENSIVE_CHECKING
359 struct buffer_head
*bh2
;
361 /* If there is a different buffer_head lying around in
362 * memory anywhere... */
363 bh2
= __find_get_block(bdev
, blocknr
, journal
->j_blocksize
);
365 /* ... and it has RevokeValid status... */
366 if (bh2
!= bh
&& buffer_revokevalid(bh2
))
367 /* ...then it better be revoked too,
368 * since it's illegal to create a revoke
369 * record against a buffer_head which is
370 * not marked revoked --- that would
371 * risk missing a subsequent revoke
373 J_ASSERT_BH(bh2
, buffer_revoked(bh2
));
379 /* We really ought not ever to revoke twice in a row without
380 first having the revoke cancelled: it's illegal to free a
381 block twice without allocating it in between! */
383 if (!J_EXPECT_BH(bh
, !buffer_revoked(bh
),
384 "inconsistent data on disk")) {
389 set_buffer_revoked(bh
);
390 set_buffer_revokevalid(bh
);
392 BUFFER_TRACE(bh_in
, "call journal_forget");
393 journal_forget(handle
, bh_in
);
395 BUFFER_TRACE(bh
, "call brelse");
400 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr
, bh_in
);
401 err
= insert_revoke_hash(journal
, blocknr
,
402 handle
->h_transaction
->t_tid
);
403 BUFFER_TRACE(bh_in
, "exit");
408 * Cancel an outstanding revoke. For use only internally by the
409 * journaling code (called from journal_get_write_access).
411 * We trust buffer_revoked() on the buffer if the buffer is already
412 * being journaled: if there is no revoke pending on the buffer, then we
413 * don't do anything here.
415 * This would break if it were possible for a buffer to be revoked and
416 * discarded, and then reallocated within the same transaction. In such
417 * a case we would have lost the revoked bit, but when we arrived here
418 * the second time we would still have a pending revoke to cancel. So,
419 * do not trust the Revoked bit on buffers unless RevokeValid is also
422 int journal_cancel_revoke(handle_t
*handle
, struct journal_head
*jh
)
424 struct jbd2_revoke_record_s
*record
;
425 journal_t
*journal
= handle
->h_transaction
->t_journal
;
427 int did_revoke
= 0; /* akpm: debug */
428 struct buffer_head
*bh
= jh2bh(jh
);
430 jbd_debug(4, "journal_head %p, canceling revoke\n", jh
);
432 /* Is the existing Revoke bit valid? If so, we trust it, and
433 * only perform the full cancel if the revoke bit is set. If
434 * not, we can't trust the revoke bit, and we need to do the
435 * full search for a revoke record. */
436 if (test_set_buffer_revokevalid(bh
)) {
437 need_cancel
= test_clear_buffer_revoked(bh
);
440 clear_buffer_revoked(bh
);
444 record
= find_revoke_record(journal
, bh
->b_blocknr
);
446 jbd_debug(4, "cancelled existing revoke on "
447 "blocknr %llu\n", (unsigned long long)bh
->b_blocknr
);
448 spin_lock(&journal
->j_revoke_lock
);
449 list_del(&record
->hash
);
450 spin_unlock(&journal
->j_revoke_lock
);
451 kmem_cache_free(jbd2_revoke_record_cache
, record
);
456 #ifdef JFS_EXPENSIVE_CHECKING
457 /* There better not be one left behind by now! */
458 record
= find_revoke_record(journal
, bh
->b_blocknr
);
459 J_ASSERT_JH(jh
, record
== NULL
);
462 /* Finally, have we just cleared revoke on an unhashed
463 * buffer_head? If so, we'd better make sure we clear the
464 * revoked status on any hashed alias too, otherwise the revoke
465 * state machine will get very upset later on. */
467 struct buffer_head
*bh2
;
468 bh2
= __find_get_block(bh
->b_bdev
, bh
->b_blocknr
, bh
->b_size
);
471 clear_buffer_revoked(bh2
);
479 * journal_clear_revoked_flag clears revoked flag of buffers in
480 * revoke table to reflect there is no revoked buffers in the next
481 * transaction which is going to be started.
483 void jbd2_clear_buffer_revoked_flags(journal_t
*journal
)
485 struct jbd2_revoke_table_s
*revoke
= journal
->j_revoke
;
488 for (i
= 0; i
< revoke
->hash_size
; i
++) {
489 struct list_head
*hash_list
;
490 struct list_head
*list_entry
;
491 hash_list
= &revoke
->hash_table
[i
];
493 list_for_each(list_entry
, hash_list
) {
494 struct jbd2_revoke_record_s
*record
;
495 struct buffer_head
*bh
;
496 record
= (struct jbd2_revoke_record_s
*)list_entry
;
497 bh
= __find_get_block(journal
->j_fs_dev
,
499 journal
->j_blocksize
);
501 clear_buffer_revoked(bh
);
508 /* journal_switch_revoke table select j_revoke for next transaction
509 * we do not want to suspend any processing until all revokes are
512 void journal_switch_revoke_table(journal_t
*journal
)
516 if (journal
->j_revoke
== journal
->j_revoke_table
[0])
517 journal
->j_revoke
= journal
->j_revoke_table
[1];
519 journal
->j_revoke
= journal
->j_revoke_table
[0];
521 for (i
= 0; i
< journal
->j_revoke
->hash_size
; i
++)
522 INIT_LIST_HEAD(&journal
->j_revoke
->hash_table
[i
]);
526 * Write revoke records to the journal for all entries in the current
527 * revoke hash, deleting the entries as we go.
529 void journal_write_revoke_records(journal_t
*journal
,
530 transaction_t
*transaction
,
531 struct list_head
*log_bufs
,
534 struct buffer_head
*descriptor
;
535 struct jbd2_revoke_record_s
*record
;
536 struct jbd2_revoke_table_s
*revoke
;
537 struct list_head
*hash_list
;
538 int i
, offset
, count
;
544 /* select revoke table for committing transaction */
545 revoke
= journal
->j_revoke
== journal
->j_revoke_table
[0] ?
546 journal
->j_revoke_table
[1] : journal
->j_revoke_table
[0];
548 for (i
= 0; i
< revoke
->hash_size
; i
++) {
549 hash_list
= &revoke
->hash_table
[i
];
551 while (!list_empty(hash_list
)) {
552 record
= (struct jbd2_revoke_record_s
*)
554 write_one_revoke_record(journal
, transaction
, log_bufs
,
555 &descriptor
, &offset
,
558 list_del(&record
->hash
);
559 kmem_cache_free(jbd2_revoke_record_cache
, record
);
563 flush_descriptor(journal
, descriptor
, offset
, write_op
);
564 jbd_debug(1, "Wrote %d revoke records\n", count
);
568 * Write out one revoke record. We need to create a new descriptor
569 * block if the old one is full or if we have not already created one.
572 static void write_one_revoke_record(journal_t
*journal
,
573 transaction_t
*transaction
,
574 struct list_head
*log_bufs
,
575 struct buffer_head
**descriptorp
,
577 struct jbd2_revoke_record_s
*record
,
581 struct buffer_head
*descriptor
;
583 journal_header_t
*header
;
585 /* If we are already aborting, this all becomes a noop. We
586 still need to go round the loop in
587 journal_write_revoke_records in order to free all of the
588 revoke records: only the IO to the journal is omitted. */
589 if (is_journal_aborted(journal
))
592 descriptor
= *descriptorp
;
595 /* Do we need to leave space at the end for a checksum? */
596 if (journal_has_csum_v2or3(journal
))
597 csum_size
= sizeof(struct journal_revoke_tail
);
599 if (JBD2_HAS_INCOMPAT_FEATURE(journal
, JBD2_FEATURE_INCOMPAT_64BIT
))
604 /* Make sure we have a descriptor with space left for the record */
606 if (offset
+ sz
> journal
->j_blocksize
- csum_size
) {
607 flush_descriptor(journal
, descriptor
, offset
, write_op
);
613 descriptor
= journal_get_descriptor_buffer(journal
);
616 header
= (journal_header_t
*)descriptor
->b_data
;
617 header
->h_magic
= ext2fs_cpu_to_be32(JFS_MAGIC_NUMBER
);
618 header
->h_blocktype
= ext2fs_cpu_to_be32(JFS_REVOKE_BLOCK
);
619 header
->h_sequence
= ext2fs_cpu_to_be32(transaction
->t_tid
);
621 /* Record it so that we can wait for IO completion later */
622 BUFFER_TRACE(descriptor
, "file in log_bufs");
623 jbd2_file_log_bh(log_bufs
, descriptor
);
625 offset
= sizeof(journal_revoke_header_t
);
626 *descriptorp
= descriptor
;
629 if (JBD2_HAS_INCOMPAT_FEATURE(journal
, JBD2_FEATURE_INCOMPAT_64BIT
)) {
630 * ((__be64
*)(&descriptor
->b_data
[offset
])) =
631 cpu_to_be64(record
->blocknr
);
633 * ((__be32
*)(&descriptor
->b_data
[offset
])) =
634 cpu_to_be32(record
->blocknr
);
640 static void jbd2_revoke_csum_set(journal_t
*j
, struct buffer_head
*bh
)
642 struct journal_revoke_tail
*tail
;
645 if (!journal_has_csum_v2or3(j
))
648 tail
= (struct journal_revoke_tail
*)(bh
->b_data
+ j
->j_blocksize
-
649 sizeof(struct journal_revoke_tail
));
650 tail
->r_checksum
= 0;
651 csum
= jbd2_chksum(j
, j
->j_csum_seed
, bh
->b_data
, j
->j_blocksize
);
652 tail
->r_checksum
= ext2fs_cpu_to_be32(csum
);
656 * Flush a revoke descriptor out to the journal. If we are aborting,
657 * this is a noop; otherwise we are generating a buffer which needs to
658 * be waited for during commit, so it has to go onto the appropriate
659 * journal buffer list.
662 static void flush_descriptor(journal_t
*journal
,
663 struct buffer_head
*descriptor
,
664 int offset
, int write_op
)
666 journal_revoke_header_t
*header
;
668 if (is_journal_aborted(journal
)) {
673 header
= (journal_revoke_header_t
*)descriptor
->b_data
;
674 header
->r_count
= ext2fs_cpu_to_be32(offset
);
675 jbd2_revoke_csum_set(journal
, descriptor
);
677 set_buffer_jwrite(descriptor
);
678 BUFFER_TRACE(descriptor
, "write");
679 set_buffer_dirty(descriptor
);
680 write_dirty_buffer(descriptor
, write_op
);
685 * Revoke support for recovery.
687 * Recovery needs to be able to:
689 * record all revoke records, including the tid of the latest instance
690 * of each revoke in the journal
692 * check whether a given block in a given transaction should be replayed
693 * (ie. has not been revoked by a revoke record in that or a subsequent
696 * empty the revoke table after recovery.
700 * First, setting revoke records. We create a new revoke record for
701 * every block ever revoked in the log as we scan it for recovery, and
702 * we update the existing records if we find multiple revokes for a
706 int journal_set_revoke(journal_t
*journal
,
707 unsigned long long blocknr
,
710 struct jbd2_revoke_record_s
*record
;
712 record
= find_revoke_record(journal
, blocknr
);
714 /* If we have multiple occurrences, only record the
715 * latest sequence number in the hashed record */
716 if (tid_gt(sequence
, record
->sequence
))
717 record
->sequence
= sequence
;
720 return insert_revoke_hash(journal
, blocknr
, sequence
);
724 * Test revoke records. For a given block referenced in the log, has
725 * that block been revoked? A revoke record with a given transaction
726 * sequence number revokes all blocks in that transaction and earlier
727 * ones, but later transactions still need replayed.
730 int journal_test_revoke(journal_t
*journal
,
731 unsigned long long blocknr
,
734 struct jbd2_revoke_record_s
*record
;
736 record
= find_revoke_record(journal
, blocknr
);
739 if (tid_gt(sequence
, record
->sequence
))
745 * Finally, once recovery is over, we need to clear the revoke table so
746 * that it can be reused by the running filesystem.
749 void journal_clear_revoke(journal_t
*journal
)
752 struct list_head
*hash_list
;
753 struct jbd2_revoke_record_s
*record
;
754 struct jbd2_revoke_table_s
*revoke
;
756 revoke
= journal
->j_revoke
;
758 for (i
= 0; i
< revoke
->hash_size
; i
++) {
759 hash_list
= &revoke
->hash_table
[i
];
760 while (!list_empty(hash_list
)) {
761 record
= (struct jbd2_revoke_record_s
*) hash_list
->next
;
762 list_del(&record
->hash
);
763 kmem_cache_free(jbd2_revoke_record_cache
, record
);