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 * Revoke information on buffers is a tri-state value:
52 * RevokeValid clear: no cached revoke status, need to look it up
53 * RevokeValid set, Revoked clear:
54 * buffer has not been revoked, and cancel_revoke
56 * RevokeValid set, Revoked set:
57 * buffer has been revoked.
63 #include <linux/time.h>
65 #include <linux/jbd2.h>
66 #include <linux/errno.h>
67 #include <linux/slab.h>
68 #include <linux/list.h>
69 #include <linux/init.h>
71 #include <linux/log2.h>
73 static struct kmem_cache *jbd2_revoke_record_cache;
74 static struct kmem_cache *jbd2_revoke_table_cache;
76 /* Each revoke record represents one single revoked block. During
77 journal replay, this involves recording the transaction ID of the
78 last transaction to revoke this block. */
80 struct jbd2_revoke_record_s
82 struct list_head hash;
83 tid_t sequence; /* Used for recovery only */
84 unsigned long long blocknr;
88 /* The revoke table is just a simple hash table of revoke records. */
89 struct jbd2_revoke_table_s
91 /* It is conceivable that we might want a larger hash table
92 * for recovery. Must be a power of two. */
95 struct list_head *hash_table;
100 static void write_one_revoke_record(journal_t *, transaction_t *,
101 struct journal_head **, int *,
102 struct jbd2_revoke_record_s *);
103 static void flush_descriptor(journal_t *, struct journal_head *, int);
106 /* Utility functions to maintain the revoke table */
108 /* Borrowed from buffer.c: this is a tried and tested block hash function */
109 static inline int hash(journal_t *journal, unsigned long long block)
111 struct jbd2_revoke_table_s *table = journal->j_revoke;
112 int hash_shift = table->hash_shift;
113 int hash = (int)block ^ (int)((block >> 31) >> 1);
115 return ((hash << (hash_shift - 6)) ^
117 (hash << (hash_shift - 12))) & (table->hash_size - 1);
120 static int insert_revoke_hash(journal_t *journal, unsigned long long blocknr,
123 struct list_head *hash_list;
124 struct jbd2_revoke_record_s *record;
127 record = kmem_cache_alloc(jbd2_revoke_record_cache, GFP_NOFS);
131 record->sequence = seq;
132 record->blocknr = blocknr;
133 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
134 spin_lock(&journal->j_revoke_lock);
135 list_add(&record->hash, hash_list);
136 spin_unlock(&journal->j_revoke_lock);
140 if (!journal_oom_retry)
142 jbd_debug(1, "ENOMEM in %s, retrying\n", __FUNCTION__);
147 /* Find a revoke record in the journal's hash table. */
149 static struct jbd2_revoke_record_s *find_revoke_record(journal_t *journal,
150 unsigned long long blocknr)
152 struct list_head *hash_list;
153 struct jbd2_revoke_record_s *record;
155 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
157 spin_lock(&journal->j_revoke_lock);
158 record = (struct jbd2_revoke_record_s *) hash_list->next;
159 while (&(record->hash) != hash_list) {
160 if (record->blocknr == blocknr) {
161 spin_unlock(&journal->j_revoke_lock);
164 record = (struct jbd2_revoke_record_s *) record->hash.next;
166 spin_unlock(&journal->j_revoke_lock);
170 void jbd2_journal_destroy_revoke_caches(void)
172 if (jbd2_revoke_record_cache) {
173 kmem_cache_destroy(jbd2_revoke_record_cache);
174 jbd2_revoke_record_cache = NULL;
176 if (jbd2_revoke_table_cache) {
177 kmem_cache_destroy(jbd2_revoke_table_cache);
178 jbd2_revoke_table_cache = NULL;
182 int __init jbd2_journal_init_revoke_caches(void)
184 J_ASSERT(!jbd2_revoke_record_cache);
185 J_ASSERT(!jbd2_revoke_table_cache);
187 jbd2_revoke_record_cache = kmem_cache_create("jbd2_revoke_record",
188 sizeof(struct jbd2_revoke_record_s),
190 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
192 if (!jbd2_revoke_record_cache)
193 goto record_cache_failure;
195 jbd2_revoke_table_cache = kmem_cache_create("jbd2_revoke_table",
196 sizeof(struct jbd2_revoke_table_s),
197 0, SLAB_TEMPORARY, NULL);
198 if (!jbd2_revoke_table_cache)
199 goto table_cache_failure;
202 jbd2_journal_destroy_revoke_caches();
203 record_cache_failure:
207 /* Initialise the revoke table for a given journal to a given size. */
209 int jbd2_journal_init_revoke(journal_t *journal, int hash_size)
213 J_ASSERT (journal->j_revoke_table[0] == NULL);
217 while((tmp >>= 1UL) != 0UL)
220 journal->j_revoke_table[0] = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
221 if (!journal->j_revoke_table[0])
223 journal->j_revoke = journal->j_revoke_table[0];
225 /* Check that the hash_size is a power of two */
226 J_ASSERT(is_power_of_2(hash_size));
228 journal->j_revoke->hash_size = hash_size;
230 journal->j_revoke->hash_shift = shift;
232 journal->j_revoke->hash_table =
233 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
234 if (!journal->j_revoke->hash_table) {
235 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[0]);
236 journal->j_revoke = NULL;
240 for (tmp = 0; tmp < hash_size; tmp++)
241 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
243 journal->j_revoke_table[1] = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
244 if (!journal->j_revoke_table[1]) {
245 kfree(journal->j_revoke_table[0]->hash_table);
246 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[0]);
250 journal->j_revoke = journal->j_revoke_table[1];
252 /* Check that the hash_size is a power of two */
253 J_ASSERT(is_power_of_2(hash_size));
255 journal->j_revoke->hash_size = hash_size;
257 journal->j_revoke->hash_shift = shift;
259 journal->j_revoke->hash_table =
260 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
261 if (!journal->j_revoke->hash_table) {
262 kfree(journal->j_revoke_table[0]->hash_table);
263 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[0]);
264 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[1]);
265 journal->j_revoke = NULL;
269 for (tmp = 0; tmp < hash_size; tmp++)
270 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
272 spin_lock_init(&journal->j_revoke_lock);
277 /* Destoy a journal's revoke table. The table must already be empty! */
279 void jbd2_journal_destroy_revoke(journal_t *journal)
281 struct jbd2_revoke_table_s *table;
282 struct list_head *hash_list;
285 table = journal->j_revoke_table[0];
289 for (i=0; i<table->hash_size; i++) {
290 hash_list = &table->hash_table[i];
291 J_ASSERT (list_empty(hash_list));
294 kfree(table->hash_table);
295 kmem_cache_free(jbd2_revoke_table_cache, table);
296 journal->j_revoke = NULL;
298 table = journal->j_revoke_table[1];
302 for (i=0; i<table->hash_size; i++) {
303 hash_list = &table->hash_table[i];
304 J_ASSERT (list_empty(hash_list));
307 kfree(table->hash_table);
308 kmem_cache_free(jbd2_revoke_table_cache, table);
309 journal->j_revoke = NULL;
316 * jbd2_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 jbd2_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, jbd2_journal_revoke() will decrement its b_count
339 int jbd2_journal_revoke(handle_t *handle, unsigned long long 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 (!jbd2_journal_set_features(journal, 0, 0, JBD2_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 JBD2_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 jbd2_journal_forget");
401 jbd2_journal_forget(handle, bh_in);
403 BUFFER_TRACE(bh, "call brelse");
408 jbd_debug(2, "insert revoke for block %llu, 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 jbd2_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 * The caller must have the journal locked.
432 int jbd2_journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
434 struct jbd2_revoke_record_s *record;
435 journal_t *journal = handle->h_transaction->t_journal;
437 int did_revoke = 0; /* akpm: debug */
438 struct buffer_head *bh = jh2bh(jh);
440 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
442 /* Is the existing Revoke bit valid? If so, we trust it, and
443 * only perform the full cancel if the revoke bit is set. If
444 * not, we can't trust the revoke bit, and we need to do the
445 * full search for a revoke record. */
446 if (test_set_buffer_revokevalid(bh)) {
447 need_cancel = test_clear_buffer_revoked(bh);
450 clear_buffer_revoked(bh);
454 record = find_revoke_record(journal, bh->b_blocknr);
456 jbd_debug(4, "cancelled existing revoke on "
457 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
458 spin_lock(&journal->j_revoke_lock);
459 list_del(&record->hash);
460 spin_unlock(&journal->j_revoke_lock);
461 kmem_cache_free(jbd2_revoke_record_cache, record);
466 #ifdef JBD2_EXPENSIVE_CHECKING
467 /* There better not be one left behind by now! */
468 record = find_revoke_record(journal, bh->b_blocknr);
469 J_ASSERT_JH(jh, record == NULL);
472 /* Finally, have we just cleared revoke on an unhashed
473 * buffer_head? If so, we'd better make sure we clear the
474 * revoked status on any hashed alias too, otherwise the revoke
475 * state machine will get very upset later on. */
477 struct buffer_head *bh2;
478 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
481 clear_buffer_revoked(bh2);
488 /* journal_switch_revoke table select j_revoke for next transaction
489 * we do not want to suspend any processing until all revokes are
492 void jbd2_journal_switch_revoke_table(journal_t *journal)
496 if (journal->j_revoke == journal->j_revoke_table[0])
497 journal->j_revoke = journal->j_revoke_table[1];
499 journal->j_revoke = journal->j_revoke_table[0];
501 for (i = 0; i < journal->j_revoke->hash_size; i++)
502 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
506 * Write revoke records to the journal for all entries in the current
507 * revoke hash, deleting the entries as we go.
509 * Called with the journal lock held.
512 void jbd2_journal_write_revoke_records(journal_t *journal,
513 transaction_t *transaction)
515 struct journal_head *descriptor;
516 struct jbd2_revoke_record_s *record;
517 struct jbd2_revoke_table_s *revoke;
518 struct list_head *hash_list;
519 int i, offset, count;
525 /* select revoke table for committing transaction */
526 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
527 journal->j_revoke_table[1] : journal->j_revoke_table[0];
529 for (i = 0; i < revoke->hash_size; i++) {
530 hash_list = &revoke->hash_table[i];
532 while (!list_empty(hash_list)) {
533 record = (struct jbd2_revoke_record_s *)
535 write_one_revoke_record(journal, transaction,
536 &descriptor, &offset,
539 list_del(&record->hash);
540 kmem_cache_free(jbd2_revoke_record_cache, record);
544 flush_descriptor(journal, descriptor, offset);
545 jbd_debug(1, "Wrote %d revoke records\n", count);
549 * Write out one revoke record. We need to create a new descriptor
550 * block if the old one is full or if we have not already created one.
553 static void write_one_revoke_record(journal_t *journal,
554 transaction_t *transaction,
555 struct journal_head **descriptorp,
557 struct jbd2_revoke_record_s *record)
559 struct journal_head *descriptor;
561 journal_header_t *header;
563 /* If we are already aborting, this all becomes a noop. We
564 still need to go round the loop in
565 jbd2_journal_write_revoke_records in order to free all of the
566 revoke records: only the IO to the journal is omitted. */
567 if (is_journal_aborted(journal))
570 descriptor = *descriptorp;
573 /* Make sure we have a descriptor with space left for the record */
575 if (offset == journal->j_blocksize) {
576 flush_descriptor(journal, descriptor, offset);
582 descriptor = jbd2_journal_get_descriptor_buffer(journal);
585 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
586 header->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
587 header->h_blocktype = cpu_to_be32(JBD2_REVOKE_BLOCK);
588 header->h_sequence = cpu_to_be32(transaction->t_tid);
590 /* Record it so that we can wait for IO completion later */
591 JBUFFER_TRACE(descriptor, "file as BJ_LogCtl");
592 jbd2_journal_file_buffer(descriptor, transaction, BJ_LogCtl);
594 offset = sizeof(jbd2_journal_revoke_header_t);
595 *descriptorp = descriptor;
598 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) {
599 * ((__be64 *)(&jh2bh(descriptor)->b_data[offset])) =
600 cpu_to_be64(record->blocknr);
604 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
605 cpu_to_be32(record->blocknr);
613 * Flush a revoke descriptor out to the journal. If we are aborting,
614 * this is a noop; otherwise we are generating a buffer which needs to
615 * be waited for during commit, so it has to go onto the appropriate
616 * journal buffer list.
619 static void flush_descriptor(journal_t *journal,
620 struct journal_head *descriptor,
623 jbd2_journal_revoke_header_t *header;
624 struct buffer_head *bh = jh2bh(descriptor);
626 if (is_journal_aborted(journal)) {
631 header = (jbd2_journal_revoke_header_t *) jh2bh(descriptor)->b_data;
632 header->r_count = cpu_to_be32(offset);
633 set_buffer_jwrite(bh);
634 BUFFER_TRACE(bh, "write");
635 set_buffer_dirty(bh);
636 ll_rw_block(SWRITE, 1, &bh);
641 * Revoke support for recovery.
643 * Recovery needs to be able to:
645 * record all revoke records, including the tid of the latest instance
646 * of each revoke in the journal
648 * check whether a given block in a given transaction should be replayed
649 * (ie. has not been revoked by a revoke record in that or a subsequent
652 * empty the revoke table after recovery.
656 * First, setting revoke records. We create a new revoke record for
657 * every block ever revoked in the log as we scan it for recovery, and
658 * we update the existing records if we find multiple revokes for a
662 int jbd2_journal_set_revoke(journal_t *journal,
663 unsigned long long blocknr,
666 struct jbd2_revoke_record_s *record;
668 record = find_revoke_record(journal, blocknr);
670 /* If we have multiple occurrences, only record the
671 * latest sequence number in the hashed record */
672 if (tid_gt(sequence, record->sequence))
673 record->sequence = sequence;
676 return insert_revoke_hash(journal, blocknr, sequence);
680 * Test revoke records. For a given block referenced in the log, has
681 * that block been revoked? A revoke record with a given transaction
682 * sequence number revokes all blocks in that transaction and earlier
683 * ones, but later transactions still need replayed.
686 int jbd2_journal_test_revoke(journal_t *journal,
687 unsigned long long blocknr,
690 struct jbd2_revoke_record_s *record;
692 record = find_revoke_record(journal, blocknr);
695 if (tid_gt(sequence, record->sequence))
701 * Finally, once recovery is over, we need to clear the revoke table so
702 * that it can be reused by the running filesystem.
705 void jbd2_journal_clear_revoke(journal_t *journal)
708 struct list_head *hash_list;
709 struct jbd2_revoke_record_s *record;
710 struct jbd2_revoke_table_s *revoke;
712 revoke = journal->j_revoke;
714 for (i = 0; i < revoke->hash_size; i++) {
715 hash_list = &revoke->hash_table[i];
716 while (!list_empty(hash_list)) {
717 record = (struct jbd2_revoke_record_s*) hash_list->next;
718 list_del(&record->hash);
719 kmem_cache_free(jbd2_revoke_record_cache, record);