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58 * buffer has not been revoked, and cancel_revoke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>
97 static struct kmem_cache *revoke_record_cache;
98 static struct kmem_cache *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 jbd_revoke_record_s
106 struct list_head hash;
107 tid_t sequence; /* Used for recovery only */
108 unsigned int blocknr;
112 /* The revoke table is just a simple hash table of revoke records. */
113 struct jbd_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 jbd_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 int block)
135 struct jbd_revoke_table_s *table = journal->j_revoke;
136 int hash_shift = table->hash_shift;
138 return ((block << (hash_shift - 6)) ^
140 (block << (hash_shift - 12))) & (table->hash_size - 1);
143 static int insert_revoke_hash(journal_t *journal, unsigned int blocknr,
146 struct list_head *hash_list;
147 struct jbd_revoke_record_s *record;
150 record = kmem_cache_alloc(revoke_record_cache, GFP_NOFS);
154 record->sequence = seq;
155 record->blocknr = blocknr;
156 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
157 spin_lock(&journal->j_revoke_lock);
158 list_add(&record->hash, hash_list);
159 spin_unlock(&journal->j_revoke_lock);
163 if (!journal_oom_retry)
165 jbd_debug(1, "ENOMEM in %s, retrying\n", __func__);
170 /* Find a revoke record in the journal's hash table. */
172 static struct jbd_revoke_record_s *find_revoke_record(journal_t *journal,
173 unsigned int blocknr)
175 struct list_head *hash_list;
176 struct jbd_revoke_record_s *record;
178 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
180 spin_lock(&journal->j_revoke_lock);
181 record = (struct jbd_revoke_record_s *) hash_list->next;
182 while (&(record->hash) != hash_list) {
183 if (record->blocknr == blocknr) {
184 spin_unlock(&journal->j_revoke_lock);
187 record = (struct jbd_revoke_record_s *) record->hash.next;
189 spin_unlock(&journal->j_revoke_lock);
193 void journal_destroy_revoke_caches(void)
195 if (revoke_record_cache) {
196 kmem_cache_destroy(revoke_record_cache);
197 revoke_record_cache = NULL;
199 if (revoke_table_cache) {
200 kmem_cache_destroy(revoke_table_cache);
201 revoke_table_cache = NULL;
205 int __init journal_init_revoke_caches(void)
207 J_ASSERT(!revoke_record_cache);
208 J_ASSERT(!revoke_table_cache);
210 revoke_record_cache = kmem_cache_create("revoke_record",
211 sizeof(struct jbd_revoke_record_s),
213 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
215 if (!revoke_record_cache)
216 goto record_cache_failure;
218 revoke_table_cache = kmem_cache_create("revoke_table",
219 sizeof(struct jbd_revoke_table_s),
220 0, SLAB_TEMPORARY, NULL);
221 if (!revoke_table_cache)
222 goto table_cache_failure;
227 journal_destroy_revoke_caches();
228 record_cache_failure:
232 static struct jbd_revoke_table_s *journal_init_revoke_table(int hash_size)
236 struct jbd_revoke_table_s *table;
238 table = kmem_cache_alloc(revoke_table_cache, GFP_KERNEL);
242 while((tmp >>= 1UL) != 0UL)
245 table->hash_size = hash_size;
246 table->hash_shift = shift;
248 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
249 if (!table->hash_table) {
250 kmem_cache_free(revoke_table_cache, table);
255 for (tmp = 0; tmp < hash_size; tmp++)
256 INIT_LIST_HEAD(&table->hash_table[tmp]);
262 static void journal_destroy_revoke_table(struct jbd_revoke_table_s *table)
265 struct list_head *hash_list;
267 for (i = 0; i < table->hash_size; i++) {
268 hash_list = &table->hash_table[i];
269 J_ASSERT(list_empty(hash_list));
272 kfree(table->hash_table);
273 kmem_cache_free(revoke_table_cache, table);
276 /* Initialise the revoke table for a given journal to a given size. */
277 int journal_init_revoke(journal_t *journal, int hash_size)
279 J_ASSERT(journal->j_revoke_table[0] == NULL);
280 J_ASSERT(is_power_of_2(hash_size));
282 journal->j_revoke_table[0] = journal_init_revoke_table(hash_size);
283 if (!journal->j_revoke_table[0])
286 journal->j_revoke_table[1] = journal_init_revoke_table(hash_size);
287 if (!journal->j_revoke_table[1])
290 journal->j_revoke = journal->j_revoke_table[1];
292 spin_lock_init(&journal->j_revoke_lock);
297 journal_destroy_revoke_table(journal->j_revoke_table[0]);
302 /* Destroy a journal's revoke table. The table must already be empty! */
303 void journal_destroy_revoke(journal_t *journal)
305 journal->j_revoke = NULL;
306 if (journal->j_revoke_table[0])
307 journal_destroy_revoke_table(journal->j_revoke_table[0]);
308 if (journal->j_revoke_table[1])
309 journal_destroy_revoke_table(journal->j_revoke_table[1]);
316 * journal_revoke: revoke a given buffer_head from the journal. This
317 * prevents the block from being replayed during recovery if we take a
318 * crash after this current transaction commits. Any subsequent
319 * metadata writes of the buffer in this transaction cancel the
322 * Note that this call may block --- it is up to the caller to make
323 * sure that there are no further calls to journal_write_metadata
324 * before the revoke is complete. In ext3, this implies calling the
325 * revoke before clearing the block bitmap when we are deleting
328 * Revoke performs a journal_forget on any buffer_head passed in as a
329 * parameter, but does _not_ forget the buffer_head if the bh was only
332 * bh_in may not be a journalled buffer - it may have come off
333 * the hash tables without an attached journal_head.
335 * If bh_in is non-zero, journal_revoke() will decrement its b_count
339 int journal_revoke(handle_t *handle, unsigned int blocknr,
340 struct buffer_head *bh_in)
342 struct buffer_head *bh = NULL;
344 struct block_device *bdev;
349 BUFFER_TRACE(bh_in, "enter");
351 journal = handle->h_transaction->t_journal;
352 if (!journal_set_features(journal, 0, 0, JFS_FEATURE_INCOMPAT_REVOKE)){
353 J_ASSERT (!"Cannot set revoke feature!");
357 bdev = journal->j_fs_dev;
361 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
363 BUFFER_TRACE(bh, "found on hash");
365 #ifdef JBD_EXPENSIVE_CHECKING
367 struct buffer_head *bh2;
369 /* If there is a different buffer_head lying around in
370 * memory anywhere... */
371 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
373 /* ... and it has RevokeValid status... */
374 if (bh2 != bh && buffer_revokevalid(bh2))
375 /* ...then it better be revoked too,
376 * since it's illegal to create a revoke
377 * record against a buffer_head which is
378 * not marked revoked --- that would
379 * risk missing a subsequent revoke
381 J_ASSERT_BH(bh2, buffer_revoked(bh2));
387 /* We really ought not ever to revoke twice in a row without
388 first having the revoke cancelled: it's illegal to free a
389 block twice without allocating it in between! */
391 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
392 "inconsistent data on disk")) {
397 set_buffer_revoked(bh);
398 set_buffer_revokevalid(bh);
400 BUFFER_TRACE(bh_in, "call journal_forget");
401 journal_forget(handle, bh_in);
403 BUFFER_TRACE(bh, "call brelse");
408 jbd_debug(2, "insert revoke for block %u, bh_in=%p\n", blocknr, bh_in);
409 err = insert_revoke_hash(journal, blocknr,
410 handle->h_transaction->t_tid);
411 BUFFER_TRACE(bh_in, "exit");
416 * Cancel an outstanding revoke. For use only internally by the
417 * journaling code (called from journal_get_write_access).
419 * We trust buffer_revoked() on the buffer if the buffer is already
420 * being journaled: if there is no revoke pending on the buffer, then we
421 * don't do anything here.
423 * This would break if it were possible for a buffer to be revoked and
424 * discarded, and then reallocated within the same transaction. In such
425 * a case we would have lost the revoked bit, but when we arrived here
426 * the second time we would still have a pending revoke to cancel. So,
427 * do not trust the Revoked bit on buffers unless RevokeValid is also
430 int journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
432 struct jbd_revoke_record_s *record;
433 journal_t *journal = handle->h_transaction->t_journal;
435 int did_revoke = 0; /* akpm: debug */
436 struct buffer_head *bh = jh2bh(jh);
438 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
440 /* Is the existing Revoke bit valid? If so, we trust it, and
441 * only perform the full cancel if the revoke bit is set. If
442 * not, we can't trust the revoke bit, and we need to do the
443 * full search for a revoke record. */
444 if (test_set_buffer_revokevalid(bh)) {
445 need_cancel = test_clear_buffer_revoked(bh);
448 clear_buffer_revoked(bh);
452 record = find_revoke_record(journal, bh->b_blocknr);
454 jbd_debug(4, "cancelled existing revoke on "
455 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
456 spin_lock(&journal->j_revoke_lock);
457 list_del(&record->hash);
458 spin_unlock(&journal->j_revoke_lock);
459 kmem_cache_free(revoke_record_cache, record);
464 #ifdef JBD_EXPENSIVE_CHECKING
465 /* There better not be one left behind by now! */
466 record = find_revoke_record(journal, bh->b_blocknr);
467 J_ASSERT_JH(jh, record == NULL);
470 /* Finally, have we just cleared revoke on an unhashed
471 * buffer_head? If so, we'd better make sure we clear the
472 * revoked status on any hashed alias too, otherwise the revoke
473 * state machine will get very upset later on. */
475 struct buffer_head *bh2;
476 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
479 clear_buffer_revoked(bh2);
487 * journal_clear_revoked_flags clears revoked flag of buffers in
488 * revoke table to reflect there is no revoked buffer in the next
489 * transaction which is going to be started.
491 void journal_clear_buffer_revoked_flags(journal_t *journal)
493 struct jbd_revoke_table_s *revoke = journal->j_revoke;
496 for (i = 0; i < revoke->hash_size; i++) {
497 struct list_head *hash_list;
498 struct list_head *list_entry;
499 hash_list = &revoke->hash_table[i];
501 list_for_each(list_entry, hash_list) {
502 struct jbd_revoke_record_s *record;
503 struct buffer_head *bh;
504 record = (struct jbd_revoke_record_s *)list_entry;
505 bh = __find_get_block(journal->j_fs_dev,
507 journal->j_blocksize);
509 clear_buffer_revoked(bh);
516 /* journal_switch_revoke table select j_revoke for next transaction
517 * we do not want to suspend any processing until all revokes are
520 void journal_switch_revoke_table(journal_t *journal)
524 if (journal->j_revoke == journal->j_revoke_table[0])
525 journal->j_revoke = journal->j_revoke_table[1];
527 journal->j_revoke = journal->j_revoke_table[0];
529 for (i = 0; i < journal->j_revoke->hash_size; i++)
530 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
534 * Write revoke records to the journal for all entries in the current
535 * revoke hash, deleting the entries as we go.
537 void journal_write_revoke_records(journal_t *journal,
538 transaction_t *transaction, int write_op)
540 struct journal_head *descriptor;
541 struct jbd_revoke_record_s *record;
542 struct jbd_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 jbd_revoke_record_s *)
560 write_one_revoke_record(journal, transaction,
561 &descriptor, &offset,
564 list_del(&record->hash);
565 kmem_cache_free(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 jbd_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 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 = journal_get_descriptor_buffer(journal);
611 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
612 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
613 header->h_blocktype = cpu_to_be32(JFS_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 journal_file_buffer(descriptor, transaction, BJ_LogCtl);
620 offset = sizeof(journal_revoke_header_t);
621 *descriptorp = descriptor;
624 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
625 cpu_to_be32(record->blocknr);
631 * Flush a revoke descriptor out to the journal. If we are aborting,
632 * this is a noop; otherwise we are generating a buffer which needs to
633 * be waited for during commit, so it has to go onto the appropriate
634 * journal buffer list.
637 static void flush_descriptor(journal_t *journal,
638 struct journal_head *descriptor,
639 int offset, int write_op)
641 journal_revoke_header_t *header;
642 struct buffer_head *bh = jh2bh(descriptor);
644 if (is_journal_aborted(journal)) {
649 header = (journal_revoke_header_t *) jh2bh(descriptor)->b_data;
650 header->r_count = cpu_to_be32(offset);
651 set_buffer_jwrite(bh);
652 BUFFER_TRACE(bh, "write");
653 set_buffer_dirty(bh);
654 write_dirty_buffer(bh, write_op);
659 * Revoke support for recovery.
661 * Recovery needs to be able to:
663 * record all revoke records, including the tid of the latest instance
664 * of each revoke in the journal
666 * check whether a given block in a given transaction should be replayed
667 * (ie. has not been revoked by a revoke record in that or a subsequent
670 * empty the revoke table after recovery.
674 * First, setting revoke records. We create a new revoke record for
675 * every block ever revoked in the log as we scan it for recovery, and
676 * we update the existing records if we find multiple revokes for a
680 int journal_set_revoke(journal_t *journal,
681 unsigned int blocknr,
684 struct jbd_revoke_record_s *record;
686 record = find_revoke_record(journal, blocknr);
688 /* If we have multiple occurrences, only record the
689 * latest sequence number in the hashed record */
690 if (tid_gt(sequence, record->sequence))
691 record->sequence = sequence;
694 return insert_revoke_hash(journal, blocknr, sequence);
698 * Test revoke records. For a given block referenced in the log, has
699 * that block been revoked? A revoke record with a given transaction
700 * sequence number revokes all blocks in that transaction and earlier
701 * ones, but later transactions still need replayed.
704 int journal_test_revoke(journal_t *journal,
705 unsigned int blocknr,
708 struct jbd_revoke_record_s *record;
710 record = find_revoke_record(journal, blocknr);
713 if (tid_gt(sequence, record->sequence))
719 * Finally, once recovery is over, we need to clear the revoke table so
720 * that it can be reused by the running filesystem.
723 void journal_clear_revoke(journal_t *journal)
726 struct list_head *hash_list;
727 struct jbd_revoke_record_s *record;
728 struct jbd_revoke_table_s *revoke;
730 revoke = journal->j_revoke;
732 for (i = 0; i < revoke->hash_size; i++) {
733 hash_list = &revoke->hash_table[i];
734 while (!list_empty(hash_list)) {
735 record = (struct jbd_revoke_record_s*) hash_list->next;
736 list_del(&record->hash);
737 kmem_cache_free(revoke_record_cache, record);