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58 * buffer has not been revoked, and cancel_revoke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 *,
125 struct journal_head **, int *,
126 struct jbd2_revoke_record_s *, int);
127 static void flush_descriptor(journal_t *, struct journal_head *, int, int);
130 /* Utility functions to maintain the revoke table */
132 /* Borrowed from buffer.c: this is a tried and tested block hash function */
133 static inline int hash(journal_t *journal, unsigned long long block)
135 struct jbd2_revoke_table_s *table = journal->j_revoke;
136 int hash_shift = table->hash_shift;
137 int hash = (int)block ^ (int)((block >> 31) >> 1);
139 return ((hash << (hash_shift - 6)) ^
141 (hash << (hash_shift - 12))) & (table->hash_size - 1);
144 static int insert_revoke_hash(journal_t *journal, unsigned long long blocknr,
147 struct list_head *hash_list;
148 struct jbd2_revoke_record_s *record;
151 record = kmem_cache_alloc(jbd2_revoke_record_cache, GFP_NOFS);
155 record->sequence = seq;
156 record->blocknr = blocknr;
157 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
158 spin_lock(&journal->j_revoke_lock);
159 list_add(&record->hash, hash_list);
160 spin_unlock(&journal->j_revoke_lock);
164 if (!journal_oom_retry)
166 jbd_debug(1, "ENOMEM in %s, retrying\n", __func__);
171 /* Find a revoke record in the journal's hash table. */
173 static struct jbd2_revoke_record_s *find_revoke_record(journal_t *journal,
174 unsigned long long blocknr)
176 struct list_head *hash_list;
177 struct jbd2_revoke_record_s *record;
179 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
181 spin_lock(&journal->j_revoke_lock);
182 record = (struct jbd2_revoke_record_s *) hash_list->next;
183 while (&(record->hash) != hash_list) {
184 if (record->blocknr == blocknr) {
185 spin_unlock(&journal->j_revoke_lock);
188 record = (struct jbd2_revoke_record_s *) record->hash.next;
190 spin_unlock(&journal->j_revoke_lock);
194 void jbd2_journal_destroy_revoke_caches(void)
196 if (jbd2_revoke_record_cache) {
197 kmem_cache_destroy(jbd2_revoke_record_cache);
198 jbd2_revoke_record_cache = NULL;
200 if (jbd2_revoke_table_cache) {
201 kmem_cache_destroy(jbd2_revoke_table_cache);
202 jbd2_revoke_table_cache = NULL;
206 int __init jbd2_journal_init_revoke_caches(void)
208 J_ASSERT(!jbd2_revoke_record_cache);
209 J_ASSERT(!jbd2_revoke_table_cache);
211 jbd2_revoke_record_cache = kmem_cache_create("jbd2_revoke_record",
212 sizeof(struct jbd2_revoke_record_s),
214 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
216 if (!jbd2_revoke_record_cache)
217 goto record_cache_failure;
219 jbd2_revoke_table_cache = kmem_cache_create("jbd2_revoke_table",
220 sizeof(struct jbd2_revoke_table_s),
221 0, SLAB_TEMPORARY, NULL);
222 if (!jbd2_revoke_table_cache)
223 goto table_cache_failure;
226 jbd2_journal_destroy_revoke_caches();
227 record_cache_failure:
231 static struct jbd2_revoke_table_s *jbd2_journal_init_revoke_table(int hash_size)
235 struct jbd2_revoke_table_s *table;
237 table = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
241 while((tmp >>= 1UL) != 0UL)
244 table->hash_size = hash_size;
245 table->hash_shift = shift;
247 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
248 if (!table->hash_table) {
249 kmem_cache_free(jbd2_revoke_table_cache, table);
254 for (tmp = 0; tmp < hash_size; tmp++)
255 INIT_LIST_HEAD(&table->hash_table[tmp]);
261 static void jbd2_journal_destroy_revoke_table(struct jbd2_revoke_table_s *table)
264 struct list_head *hash_list;
266 for (i = 0; i < table->hash_size; i++) {
267 hash_list = &table->hash_table[i];
268 J_ASSERT(list_empty(hash_list));
271 kfree(table->hash_table);
272 kmem_cache_free(jbd2_revoke_table_cache, table);
275 /* Initialise the revoke table for a given journal to a given size. */
276 int jbd2_journal_init_revoke(journal_t *journal, int hash_size)
278 J_ASSERT(journal->j_revoke_table[0] == NULL);
279 J_ASSERT(is_power_of_2(hash_size));
281 journal->j_revoke_table[0] = jbd2_journal_init_revoke_table(hash_size);
282 if (!journal->j_revoke_table[0])
285 journal->j_revoke_table[1] = jbd2_journal_init_revoke_table(hash_size);
286 if (!journal->j_revoke_table[1])
289 journal->j_revoke = journal->j_revoke_table[1];
291 spin_lock_init(&journal->j_revoke_lock);
296 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
301 /* Destroy a journal's revoke table. The table must already be empty! */
302 void jbd2_journal_destroy_revoke(journal_t *journal)
304 journal->j_revoke = NULL;
305 if (journal->j_revoke_table[0])
306 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
307 if (journal->j_revoke_table[1])
308 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[1]);
315 * jbd2_journal_revoke: revoke a given buffer_head from the journal. This
316 * prevents the block from being replayed during recovery if we take a
317 * crash after this current transaction commits. Any subsequent
318 * metadata writes of the buffer in this transaction cancel the
321 * Note that this call may block --- it is up to the caller to make
322 * sure that there are no further calls to journal_write_metadata
323 * before the revoke is complete. In ext3, this implies calling the
324 * revoke before clearing the block bitmap when we are deleting
327 * Revoke performs a jbd2_journal_forget on any buffer_head passed in as a
328 * parameter, but does _not_ forget the buffer_head if the bh was only
331 * bh_in may not be a journalled buffer - it may have come off
332 * the hash tables without an attached journal_head.
334 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
338 int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
339 struct buffer_head *bh_in)
341 struct buffer_head *bh = NULL;
343 struct block_device *bdev;
348 BUFFER_TRACE(bh_in, "enter");
350 journal = handle->h_transaction->t_journal;
351 if (!jbd2_journal_set_features(journal, 0, 0, JBD2_FEATURE_INCOMPAT_REVOKE)){
352 J_ASSERT (!"Cannot set revoke feature!");
356 bdev = journal->j_fs_dev;
360 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
362 BUFFER_TRACE(bh, "found on hash");
364 #ifdef JBD2_EXPENSIVE_CHECKING
366 struct buffer_head *bh2;
368 /* If there is a different buffer_head lying around in
369 * memory anywhere... */
370 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
372 /* ... and it has RevokeValid status... */
373 if (bh2 != bh && buffer_revokevalid(bh2))
374 /* ...then it better be revoked too,
375 * since it's illegal to create a revoke
376 * record against a buffer_head which is
377 * not marked revoked --- that would
378 * risk missing a subsequent revoke
380 J_ASSERT_BH(bh2, buffer_revoked(bh2));
386 /* We really ought not ever to revoke twice in a row without
387 first having the revoke cancelled: it's illegal to free a
388 block twice without allocating it in between! */
390 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
391 "inconsistent data on disk")) {
396 set_buffer_revoked(bh);
397 set_buffer_revokevalid(bh);
399 BUFFER_TRACE(bh_in, "call jbd2_journal_forget");
400 jbd2_journal_forget(handle, bh_in);
402 BUFFER_TRACE(bh, "call brelse");
407 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr, bh_in);
408 err = insert_revoke_hash(journal, blocknr,
409 handle->h_transaction->t_tid);
410 BUFFER_TRACE(bh_in, "exit");
415 * Cancel an outstanding revoke. For use only internally by the
416 * journaling code (called from jbd2_journal_get_write_access).
418 * We trust buffer_revoked() on the buffer if the buffer is already
419 * being journaled: if there is no revoke pending on the buffer, then we
420 * don't do anything here.
422 * This would break if it were possible for a buffer to be revoked and
423 * discarded, and then reallocated within the same transaction. In such
424 * a case we would have lost the revoked bit, but when we arrived here
425 * the second time we would still have a pending revoke to cancel. So,
426 * do not trust the Revoked bit on buffers unless RevokeValid is also
429 int jbd2_journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
431 struct jbd2_revoke_record_s *record;
432 journal_t *journal = handle->h_transaction->t_journal;
434 int did_revoke = 0; /* akpm: debug */
435 struct buffer_head *bh = jh2bh(jh);
437 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
439 /* Is the existing Revoke bit valid? If so, we trust it, and
440 * only perform the full cancel if the revoke bit is set. If
441 * not, we can't trust the revoke bit, and we need to do the
442 * full search for a revoke record. */
443 if (test_set_buffer_revokevalid(bh)) {
444 need_cancel = test_clear_buffer_revoked(bh);
447 clear_buffer_revoked(bh);
451 record = find_revoke_record(journal, bh->b_blocknr);
453 jbd_debug(4, "cancelled existing revoke on "
454 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
455 spin_lock(&journal->j_revoke_lock);
456 list_del(&record->hash);
457 spin_unlock(&journal->j_revoke_lock);
458 kmem_cache_free(jbd2_revoke_record_cache, record);
463 #ifdef JBD2_EXPENSIVE_CHECKING
464 /* There better not be one left behind by now! */
465 record = find_revoke_record(journal, bh->b_blocknr);
466 J_ASSERT_JH(jh, record == NULL);
469 /* Finally, have we just cleared revoke on an unhashed
470 * buffer_head? If so, we'd better make sure we clear the
471 * revoked status on any hashed alias too, otherwise the revoke
472 * state machine will get very upset later on. */
474 struct buffer_head *bh2;
475 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
478 clear_buffer_revoked(bh2);
486 * journal_clear_revoked_flag clears revoked flag of buffers in
487 * revoke table to reflect there is no revoked buffers in the next
488 * transaction which is going to be started.
490 void jbd2_clear_buffer_revoked_flags(journal_t *journal)
492 struct jbd2_revoke_table_s *revoke = journal->j_revoke;
495 for (i = 0; i < revoke->hash_size; i++) {
496 struct list_head *hash_list;
497 struct list_head *list_entry;
498 hash_list = &revoke->hash_table[i];
500 list_for_each(list_entry, hash_list) {
501 struct jbd2_revoke_record_s *record;
502 struct buffer_head *bh;
503 record = (struct jbd2_revoke_record_s *)list_entry;
504 bh = __find_get_block(journal->j_fs_dev,
506 journal->j_blocksize);
508 clear_buffer_revoked(bh);
515 /* journal_switch_revoke table select j_revoke for next transaction
516 * we do not want to suspend any processing until all revokes are
519 void jbd2_journal_switch_revoke_table(journal_t *journal)
523 if (journal->j_revoke == journal->j_revoke_table[0])
524 journal->j_revoke = journal->j_revoke_table[1];
526 journal->j_revoke = journal->j_revoke_table[0];
528 for (i = 0; i < journal->j_revoke->hash_size; i++)
529 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
533 * Write revoke records to the journal for all entries in the current
534 * revoke hash, deleting the entries as we go.
536 void jbd2_journal_write_revoke_records(journal_t *journal,
537 transaction_t *transaction,
540 struct journal_head *descriptor;
541 struct jbd2_revoke_record_s *record;
542 struct jbd2_revoke_table_s *revoke;
543 struct list_head *hash_list;
544 int i, offset, count;
550 /* select revoke table for committing transaction */
551 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
552 journal->j_revoke_table[1] : journal->j_revoke_table[0];
554 for (i = 0; i < revoke->hash_size; i++) {
555 hash_list = &revoke->hash_table[i];
557 while (!list_empty(hash_list)) {
558 record = (struct jbd2_revoke_record_s *)
560 write_one_revoke_record(journal, transaction,
561 &descriptor, &offset,
564 list_del(&record->hash);
565 kmem_cache_free(jbd2_revoke_record_cache, record);
569 flush_descriptor(journal, descriptor, offset, write_op);
570 jbd_debug(1, "Wrote %d revoke records\n", count);
574 * Write out one revoke record. We need to create a new descriptor
575 * block if the old one is full or if we have not already created one.
578 static void write_one_revoke_record(journal_t *journal,
579 transaction_t *transaction,
580 struct journal_head **descriptorp,
582 struct jbd2_revoke_record_s *record,
585 struct journal_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 /* Make sure we have a descriptor with space left for the record */
601 if (offset == journal->j_blocksize) {
602 flush_descriptor(journal, descriptor, offset, write_op);
608 descriptor = jbd2_journal_get_descriptor_buffer(journal);
611 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
612 header->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
613 header->h_blocktype = cpu_to_be32(JBD2_REVOKE_BLOCK);
614 header->h_sequence = cpu_to_be32(transaction->t_tid);
616 /* Record it so that we can wait for IO completion later */
617 JBUFFER_TRACE(descriptor, "file as BJ_LogCtl");
618 jbd2_journal_file_buffer(descriptor, transaction, BJ_LogCtl);
620 offset = sizeof(jbd2_journal_revoke_header_t);
621 *descriptorp = descriptor;
624 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) {
625 * ((__be64 *)(&jh2bh(descriptor)->b_data[offset])) =
626 cpu_to_be64(record->blocknr);
630 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
631 cpu_to_be32(record->blocknr);
639 * Flush a revoke descriptor out to the journal. If we are aborting,
640 * this is a noop; otherwise we are generating a buffer which needs to
641 * be waited for during commit, so it has to go onto the appropriate
642 * journal buffer list.
645 static void flush_descriptor(journal_t *journal,
646 struct journal_head *descriptor,
647 int offset, int write_op)
649 jbd2_journal_revoke_header_t *header;
650 struct buffer_head *bh = jh2bh(descriptor);
652 if (is_journal_aborted(journal)) {
657 header = (jbd2_journal_revoke_header_t *) jh2bh(descriptor)->b_data;
658 header->r_count = cpu_to_be32(offset);
659 set_buffer_jwrite(bh);
660 BUFFER_TRACE(bh, "write");
661 set_buffer_dirty(bh);
662 write_dirty_buffer(bh, write_op);
667 * Revoke support for recovery.
669 * Recovery needs to be able to:
671 * record all revoke records, including the tid of the latest instance
672 * of each revoke in the journal
674 * check whether a given block in a given transaction should be replayed
675 * (ie. has not been revoked by a revoke record in that or a subsequent
678 * empty the revoke table after recovery.
682 * First, setting revoke records. We create a new revoke record for
683 * every block ever revoked in the log as we scan it for recovery, and
684 * we update the existing records if we find multiple revokes for a
688 int jbd2_journal_set_revoke(journal_t *journal,
689 unsigned long long blocknr,
692 struct jbd2_revoke_record_s *record;
694 record = find_revoke_record(journal, blocknr);
696 /* If we have multiple occurrences, only record the
697 * latest sequence number in the hashed record */
698 if (tid_gt(sequence, record->sequence))
699 record->sequence = sequence;
702 return insert_revoke_hash(journal, blocknr, sequence);
706 * Test revoke records. For a given block referenced in the log, has
707 * that block been revoked? A revoke record with a given transaction
708 * sequence number revokes all blocks in that transaction and earlier
709 * ones, but later transactions still need replayed.
712 int jbd2_journal_test_revoke(journal_t *journal,
713 unsigned long long blocknr,
716 struct jbd2_revoke_record_s *record;
718 record = find_revoke_record(journal, blocknr);
721 if (tid_gt(sequence, record->sequence))
727 * Finally, once recovery is over, we need to clear the revoke table so
728 * that it can be reused by the running filesystem.
731 void jbd2_journal_clear_revoke(journal_t *journal)
734 struct list_head *hash_list;
735 struct jbd2_revoke_record_s *record;
736 struct jbd2_revoke_table_s *revoke;
738 revoke = journal->j_revoke;
740 for (i = 0; i < revoke->hash_size; i++) {
741 hash_list = &revoke->hash_table[i];
742 while (!list_empty(hash_list)) {
743 record = (struct jbd2_revoke_record_s*) hash_list->next;
744 list_del(&record->hash);
745 kmem_cache_free(jbd2_revoke_record_cache, record);