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58 * buffer has not been revoked, and cancel_revoke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>
95 #include <linux/log2.h>
97 static struct kmem_cache
*jbd2_revoke_record_cache
;
98 static struct kmem_cache
*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
;
137 int hash_shift
= table
->hash_shift
;
138 int hash
= (int)block
^ (int)((block
>> 31) >> 1);
140 return ((hash
<< (hash_shift
- 6)) ^
142 (hash
<< (hash_shift
- 12))) & (table
->hash_size
- 1);
145 static int insert_revoke_hash(journal_t
*journal
, unsigned long long blocknr
,
148 struct list_head
*hash_list
;
149 struct jbd2_revoke_record_s
*record
;
152 record
= kmem_cache_alloc(jbd2_revoke_record_cache
, GFP_NOFS
);
156 record
->sequence
= seq
;
157 record
->blocknr
= blocknr
;
158 hash_list
= &journal
->j_revoke
->hash_table
[hash(journal
, blocknr
)];
159 spin_lock(&journal
->j_revoke_lock
);
160 list_add(&record
->hash
, hash_list
);
161 spin_unlock(&journal
->j_revoke_lock
);
165 if (!journal_oom_retry
)
167 jbd_debug(1, "ENOMEM in %s, retrying\n", __func__
);
172 /* Find a revoke record in the journal's hash table. */
174 static struct jbd2_revoke_record_s
*find_revoke_record(journal_t
*journal
,
175 unsigned long long blocknr
)
177 struct list_head
*hash_list
;
178 struct jbd2_revoke_record_s
*record
;
180 hash_list
= &journal
->j_revoke
->hash_table
[hash(journal
, blocknr
)];
182 spin_lock(&journal
->j_revoke_lock
);
183 record
= (struct jbd2_revoke_record_s
*) hash_list
->next
;
184 while (&(record
->hash
) != hash_list
) {
185 if (record
->blocknr
== blocknr
) {
186 spin_unlock(&journal
->j_revoke_lock
);
189 record
= (struct jbd2_revoke_record_s
*) record
->hash
.next
;
191 spin_unlock(&journal
->j_revoke_lock
);
195 void jbd2_journal_destroy_revoke_caches(void)
197 if (jbd2_revoke_record_cache
) {
198 kmem_cache_destroy(jbd2_revoke_record_cache
);
199 jbd2_revoke_record_cache
= NULL
;
201 if (jbd2_revoke_table_cache
) {
202 kmem_cache_destroy(jbd2_revoke_table_cache
);
203 jbd2_revoke_table_cache
= NULL
;
207 int __init
jbd2_journal_init_revoke_caches(void)
209 J_ASSERT(!jbd2_revoke_record_cache
);
210 J_ASSERT(!jbd2_revoke_table_cache
);
212 jbd2_revoke_record_cache
= KMEM_CACHE(jbd2_revoke_record_s
,
213 SLAB_HWCACHE_ALIGN
|SLAB_TEMPORARY
);
214 if (!jbd2_revoke_record_cache
)
215 goto record_cache_failure
;
217 jbd2_revoke_table_cache
= KMEM_CACHE(jbd2_revoke_table_s
,
219 if (!jbd2_revoke_table_cache
)
220 goto table_cache_failure
;
223 jbd2_journal_destroy_revoke_caches();
224 record_cache_failure
:
228 static struct jbd2_revoke_table_s
*jbd2_journal_init_revoke_table(int hash_size
)
232 struct jbd2_revoke_table_s
*table
;
234 table
= kmem_cache_alloc(jbd2_revoke_table_cache
, GFP_KERNEL
);
238 while((tmp
>>= 1UL) != 0UL)
241 table
->hash_size
= hash_size
;
242 table
->hash_shift
= shift
;
244 kmalloc(hash_size
* sizeof(struct list_head
), GFP_KERNEL
);
245 if (!table
->hash_table
) {
246 kmem_cache_free(jbd2_revoke_table_cache
, table
);
251 for (tmp
= 0; tmp
< hash_size
; tmp
++)
252 INIT_LIST_HEAD(&table
->hash_table
[tmp
]);
258 static void jbd2_journal_destroy_revoke_table(struct jbd2_revoke_table_s
*table
)
261 struct list_head
*hash_list
;
263 for (i
= 0; i
< table
->hash_size
; i
++) {
264 hash_list
= &table
->hash_table
[i
];
265 J_ASSERT(list_empty(hash_list
));
268 kfree(table
->hash_table
);
269 kmem_cache_free(jbd2_revoke_table_cache
, table
);
272 /* Initialise the revoke table for a given journal to a given size. */
273 int jbd2_journal_init_revoke(journal_t
*journal
, int hash_size
)
275 J_ASSERT(journal
->j_revoke_table
[0] == NULL
);
276 J_ASSERT(is_power_of_2(hash_size
));
278 journal
->j_revoke_table
[0] = jbd2_journal_init_revoke_table(hash_size
);
279 if (!journal
->j_revoke_table
[0])
282 journal
->j_revoke_table
[1] = jbd2_journal_init_revoke_table(hash_size
);
283 if (!journal
->j_revoke_table
[1])
286 journal
->j_revoke
= journal
->j_revoke_table
[1];
288 spin_lock_init(&journal
->j_revoke_lock
);
293 jbd2_journal_destroy_revoke_table(journal
->j_revoke_table
[0]);
298 /* Destroy a journal's revoke table. The table must already be empty! */
299 void jbd2_journal_destroy_revoke(journal_t
*journal
)
301 journal
->j_revoke
= NULL
;
302 if (journal
->j_revoke_table
[0])
303 jbd2_journal_destroy_revoke_table(journal
->j_revoke_table
[0]);
304 if (journal
->j_revoke_table
[1])
305 jbd2_journal_destroy_revoke_table(journal
->j_revoke_table
[1]);
312 * jbd2_journal_revoke: revoke a given buffer_head from the journal. This
313 * prevents the block from being replayed during recovery if we take a
314 * crash after this current transaction commits. Any subsequent
315 * metadata writes of the buffer in this transaction cancel the
318 * Note that this call may block --- it is up to the caller to make
319 * sure that there are no further calls to journal_write_metadata
320 * before the revoke is complete. In ext3, this implies calling the
321 * revoke before clearing the block bitmap when we are deleting
324 * Revoke performs a jbd2_journal_forget on any buffer_head passed in as a
325 * parameter, but does _not_ forget the buffer_head if the bh was only
328 * bh_in may not be a journalled buffer - it may have come off
329 * the hash tables without an attached journal_head.
331 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
335 int jbd2_journal_revoke(handle_t
*handle
, unsigned long long blocknr
,
336 struct buffer_head
*bh_in
)
338 struct buffer_head
*bh
= NULL
;
340 struct block_device
*bdev
;
345 BUFFER_TRACE(bh_in
, "enter");
347 journal
= handle
->h_transaction
->t_journal
;
348 if (!jbd2_journal_set_features(journal
, 0, 0, JBD2_FEATURE_INCOMPAT_REVOKE
)){
349 J_ASSERT (!"Cannot set revoke feature!");
353 bdev
= journal
->j_fs_dev
;
357 bh
= __find_get_block(bdev
, blocknr
, journal
->j_blocksize
);
359 BUFFER_TRACE(bh
, "found on hash");
361 #ifdef JBD2_EXPENSIVE_CHECKING
363 struct buffer_head
*bh2
;
365 /* If there is a different buffer_head lying around in
366 * memory anywhere... */
367 bh2
= __find_get_block(bdev
, blocknr
, journal
->j_blocksize
);
369 /* ... and it has RevokeValid status... */
370 if (bh2
!= bh
&& buffer_revokevalid(bh2
))
371 /* ...then it better be revoked too,
372 * since it's illegal to create a revoke
373 * record against a buffer_head which is
374 * not marked revoked --- that would
375 * risk missing a subsequent revoke
377 J_ASSERT_BH(bh2
, buffer_revoked(bh2
));
383 /* We really ought not ever to revoke twice in a row without
384 first having the revoke cancelled: it's illegal to free a
385 block twice without allocating it in between! */
387 if (!J_EXPECT_BH(bh
, !buffer_revoked(bh
),
388 "inconsistent data on disk")) {
393 set_buffer_revoked(bh
);
394 set_buffer_revokevalid(bh
);
396 BUFFER_TRACE(bh_in
, "call jbd2_journal_forget");
397 jbd2_journal_forget(handle
, bh_in
);
399 BUFFER_TRACE(bh
, "call brelse");
404 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr
, bh_in
);
405 err
= insert_revoke_hash(journal
, blocknr
,
406 handle
->h_transaction
->t_tid
);
407 BUFFER_TRACE(bh_in
, "exit");
412 * Cancel an outstanding revoke. For use only internally by the
413 * journaling code (called from jbd2_journal_get_write_access).
415 * We trust buffer_revoked() on the buffer if the buffer is already
416 * being journaled: if there is no revoke pending on the buffer, then we
417 * don't do anything here.
419 * This would break if it were possible for a buffer to be revoked and
420 * discarded, and then reallocated within the same transaction. In such
421 * a case we would have lost the revoked bit, but when we arrived here
422 * the second time we would still have a pending revoke to cancel. So,
423 * do not trust the Revoked bit on buffers unless RevokeValid is also
426 int jbd2_journal_cancel_revoke(handle_t
*handle
, struct journal_head
*jh
)
428 struct jbd2_revoke_record_s
*record
;
429 journal_t
*journal
= handle
->h_transaction
->t_journal
;
431 int did_revoke
= 0; /* akpm: debug */
432 struct buffer_head
*bh
= jh2bh(jh
);
434 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh
);
436 /* Is the existing Revoke bit valid? If so, we trust it, and
437 * only perform the full cancel if the revoke bit is set. If
438 * not, we can't trust the revoke bit, and we need to do the
439 * full search for a revoke record. */
440 if (test_set_buffer_revokevalid(bh
)) {
441 need_cancel
= test_clear_buffer_revoked(bh
);
444 clear_buffer_revoked(bh
);
448 record
= find_revoke_record(journal
, bh
->b_blocknr
);
450 jbd_debug(4, "cancelled existing revoke on "
451 "blocknr %llu\n", (unsigned long long)bh
->b_blocknr
);
452 spin_lock(&journal
->j_revoke_lock
);
453 list_del(&record
->hash
);
454 spin_unlock(&journal
->j_revoke_lock
);
455 kmem_cache_free(jbd2_revoke_record_cache
, record
);
460 #ifdef JBD2_EXPENSIVE_CHECKING
461 /* There better not be one left behind by now! */
462 record
= find_revoke_record(journal
, bh
->b_blocknr
);
463 J_ASSERT_JH(jh
, record
== NULL
);
466 /* Finally, have we just cleared revoke on an unhashed
467 * buffer_head? If so, we'd better make sure we clear the
468 * revoked status on any hashed alias too, otherwise the revoke
469 * state machine will get very upset later on. */
471 struct buffer_head
*bh2
;
472 bh2
= __find_get_block(bh
->b_bdev
, bh
->b_blocknr
, bh
->b_size
);
475 clear_buffer_revoked(bh2
);
483 * journal_clear_revoked_flag clears revoked flag of buffers in
484 * revoke table to reflect there is no revoked buffers in the next
485 * transaction which is going to be started.
487 void jbd2_clear_buffer_revoked_flags(journal_t
*journal
)
489 struct jbd2_revoke_table_s
*revoke
= journal
->j_revoke
;
492 for (i
= 0; i
< revoke
->hash_size
; i
++) {
493 struct list_head
*hash_list
;
494 struct list_head
*list_entry
;
495 hash_list
= &revoke
->hash_table
[i
];
497 list_for_each(list_entry
, hash_list
) {
498 struct jbd2_revoke_record_s
*record
;
499 struct buffer_head
*bh
;
500 record
= (struct jbd2_revoke_record_s
*)list_entry
;
501 bh
= __find_get_block(journal
->j_fs_dev
,
503 journal
->j_blocksize
);
505 clear_buffer_revoked(bh
);
512 /* journal_switch_revoke table select j_revoke for next transaction
513 * we do not want to suspend any processing until all revokes are
516 void jbd2_journal_switch_revoke_table(journal_t
*journal
)
520 if (journal
->j_revoke
== journal
->j_revoke_table
[0])
521 journal
->j_revoke
= journal
->j_revoke_table
[1];
523 journal
->j_revoke
= journal
->j_revoke_table
[0];
525 for (i
= 0; i
< journal
->j_revoke
->hash_size
; i
++)
526 INIT_LIST_HEAD(&journal
->j_revoke
->hash_table
[i
]);
530 * Write revoke records to the journal for all entries in the current
531 * revoke hash, deleting the entries as we go.
533 void jbd2_journal_write_revoke_records(journal_t
*journal
,
534 transaction_t
*transaction
,
535 struct list_head
*log_bufs
,
538 struct buffer_head
*descriptor
;
539 struct jbd2_revoke_record_s
*record
;
540 struct jbd2_revoke_table_s
*revoke
;
541 struct list_head
*hash_list
;
542 int i
, offset
, count
;
548 /* select revoke table for committing transaction */
549 revoke
= journal
->j_revoke
== journal
->j_revoke_table
[0] ?
550 journal
->j_revoke_table
[1] : journal
->j_revoke_table
[0];
552 for (i
= 0; i
< revoke
->hash_size
; i
++) {
553 hash_list
= &revoke
->hash_table
[i
];
555 while (!list_empty(hash_list
)) {
556 record
= (struct jbd2_revoke_record_s
*)
558 write_one_revoke_record(journal
, transaction
, log_bufs
,
559 &descriptor
, &offset
,
562 list_del(&record
->hash
);
563 kmem_cache_free(jbd2_revoke_record_cache
, record
);
567 flush_descriptor(journal
, descriptor
, offset
, write_op
);
568 jbd_debug(1, "Wrote %d revoke records\n", count
);
572 * Write out one revoke record. We need to create a new descriptor
573 * block if the old one is full or if we have not already created one.
576 static void write_one_revoke_record(journal_t
*journal
,
577 transaction_t
*transaction
,
578 struct list_head
*log_bufs
,
579 struct buffer_head
**descriptorp
,
581 struct jbd2_revoke_record_s
*record
,
585 struct buffer_head
*descriptor
;
587 journal_header_t
*header
;
589 /* If we are already aborting, this all becomes a noop. We
590 still need to go round the loop in
591 jbd2_journal_write_revoke_records in order to free all of the
592 revoke records: only the IO to the journal is omitted. */
593 if (is_journal_aborted(journal
))
596 descriptor
= *descriptorp
;
599 /* Do we need to leave space at the end for a checksum? */
600 if (JBD2_HAS_INCOMPAT_FEATURE(journal
, JBD2_FEATURE_INCOMPAT_CSUM_V2
))
601 csum_size
= sizeof(struct jbd2_journal_revoke_tail
);
603 /* Make sure we have a descriptor with space left for the record */
605 if (offset
>= journal
->j_blocksize
- csum_size
) {
606 flush_descriptor(journal
, descriptor
, offset
, write_op
);
612 descriptor
= jbd2_journal_get_descriptor_buffer(journal
);
615 header
= (journal_header_t
*)descriptor
->b_data
;
616 header
->h_magic
= cpu_to_be32(JBD2_MAGIC_NUMBER
);
617 header
->h_blocktype
= cpu_to_be32(JBD2_REVOKE_BLOCK
);
618 header
->h_sequence
= cpu_to_be32(transaction
->t_tid
);
620 /* Record it so that we can wait for IO completion later */
621 BUFFER_TRACE(descriptor
, "file in log_bufs");
622 jbd2_file_log_bh(log_bufs
, descriptor
);
624 offset
= sizeof(jbd2_journal_revoke_header_t
);
625 *descriptorp
= descriptor
;
628 if (JBD2_HAS_INCOMPAT_FEATURE(journal
, JBD2_FEATURE_INCOMPAT_64BIT
)) {
629 * ((__be64
*)(&descriptor
->b_data
[offset
])) =
630 cpu_to_be64(record
->blocknr
);
634 * ((__be32
*)(&descriptor
->b_data
[offset
])) =
635 cpu_to_be32(record
->blocknr
);
642 static void jbd2_revoke_csum_set(journal_t
*j
, struct buffer_head
*bh
)
644 struct jbd2_journal_revoke_tail
*tail
;
647 if (!JBD2_HAS_INCOMPAT_FEATURE(j
, JBD2_FEATURE_INCOMPAT_CSUM_V2
))
650 tail
= (struct jbd2_journal_revoke_tail
*)(bh
->b_data
+ j
->j_blocksize
-
651 sizeof(struct jbd2_journal_revoke_tail
));
652 tail
->r_checksum
= 0;
653 csum
= jbd2_chksum(j
, j
->j_csum_seed
, bh
->b_data
, j
->j_blocksize
);
654 tail
->r_checksum
= cpu_to_be32(csum
);
658 * Flush a revoke descriptor out to the journal. If we are aborting,
659 * this is a noop; otherwise we are generating a buffer which needs to
660 * be waited for during commit, so it has to go onto the appropriate
661 * journal buffer list.
664 static void flush_descriptor(journal_t
*journal
,
665 struct buffer_head
*descriptor
,
666 int offset
, int write_op
)
668 jbd2_journal_revoke_header_t
*header
;
670 if (is_journal_aborted(journal
)) {
675 header
= (jbd2_journal_revoke_header_t
*)descriptor
->b_data
;
676 header
->r_count
= cpu_to_be32(offset
);
677 jbd2_revoke_csum_set(journal
, descriptor
);
679 set_buffer_jwrite(descriptor
);
680 BUFFER_TRACE(descriptor
, "write");
681 set_buffer_dirty(descriptor
);
682 write_dirty_buffer(descriptor
, write_op
);
687 * Revoke support for recovery.
689 * Recovery needs to be able to:
691 * record all revoke records, including the tid of the latest instance
692 * of each revoke in the journal
694 * check whether a given block in a given transaction should be replayed
695 * (ie. has not been revoked by a revoke record in that or a subsequent
698 * empty the revoke table after recovery.
702 * First, setting revoke records. We create a new revoke record for
703 * every block ever revoked in the log as we scan it for recovery, and
704 * we update the existing records if we find multiple revokes for a
708 int jbd2_journal_set_revoke(journal_t
*journal
,
709 unsigned long long blocknr
,
712 struct jbd2_revoke_record_s
*record
;
714 record
= find_revoke_record(journal
, blocknr
);
716 /* If we have multiple occurrences, only record the
717 * latest sequence number in the hashed record */
718 if (tid_gt(sequence
, record
->sequence
))
719 record
->sequence
= sequence
;
722 return insert_revoke_hash(journal
, blocknr
, sequence
);
726 * Test revoke records. For a given block referenced in the log, has
727 * that block been revoked? A revoke record with a given transaction
728 * sequence number revokes all blocks in that transaction and earlier
729 * ones, but later transactions still need replayed.
732 int jbd2_journal_test_revoke(journal_t
*journal
,
733 unsigned long long blocknr
,
736 struct jbd2_revoke_record_s
*record
;
738 record
= find_revoke_record(journal
, blocknr
);
741 if (tid_gt(sequence
, record
->sequence
))
747 * Finally, once recovery is over, we need to clear the revoke table so
748 * that it can be reused by the running filesystem.
751 void jbd2_journal_clear_revoke(journal_t
*journal
)
754 struct list_head
*hash_list
;
755 struct jbd2_revoke_record_s
*record
;
756 struct jbd2_revoke_table_s
*revoke
;
758 revoke
= journal
->j_revoke
;
760 for (i
= 0; i
< revoke
->hash_size
; i
++) {
761 hash_list
= &revoke
->hash_table
[i
];
762 while (!list_empty(hash_list
)) {
763 record
= (struct jbd2_revoke_record_s
*) hash_list
->next
;
764 list_del(&record
->hash
);
765 kmem_cache_free(jbd2_revoke_record_cache
, record
);