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54 * buffer has not been revoked, and cancel_revoke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57 * buffer has been revoked.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/smp_lock.h>
70 #include <linux/init.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 int __init jbd2_journal_init_revoke_caches(void)
172 jbd2_revoke_record_cache = kmem_cache_create("jbd2_revoke_record",
173 sizeof(struct jbd2_revoke_record_s),
174 0, SLAB_HWCACHE_ALIGN, NULL, NULL);
175 if (jbd2_revoke_record_cache == 0)
178 jbd2_revoke_table_cache = kmem_cache_create("jbd2_revoke_table",
179 sizeof(struct jbd2_revoke_table_s),
181 if (jbd2_revoke_table_cache == 0) {
182 kmem_cache_destroy(jbd2_revoke_record_cache);
183 jbd2_revoke_record_cache = NULL;
189 void jbd2_journal_destroy_revoke_caches(void)
191 kmem_cache_destroy(jbd2_revoke_record_cache);
192 jbd2_revoke_record_cache = NULL;
193 kmem_cache_destroy(jbd2_revoke_table_cache);
194 jbd2_revoke_table_cache = NULL;
197 /* Initialise the revoke table for a given journal to a given size. */
199 int jbd2_journal_init_revoke(journal_t *journal, int hash_size)
203 J_ASSERT (journal->j_revoke_table[0] == NULL);
207 while((tmp >>= 1UL) != 0UL)
210 journal->j_revoke_table[0] = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
211 if (!journal->j_revoke_table[0])
213 journal->j_revoke = journal->j_revoke_table[0];
215 /* Check that the hash_size is a power of two */
216 J_ASSERT ((hash_size & (hash_size-1)) == 0);
218 journal->j_revoke->hash_size = hash_size;
220 journal->j_revoke->hash_shift = shift;
222 journal->j_revoke->hash_table =
223 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
224 if (!journal->j_revoke->hash_table) {
225 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[0]);
226 journal->j_revoke = NULL;
230 for (tmp = 0; tmp < hash_size; tmp++)
231 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
233 journal->j_revoke_table[1] = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
234 if (!journal->j_revoke_table[1]) {
235 kfree(journal->j_revoke_table[0]->hash_table);
236 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[0]);
240 journal->j_revoke = journal->j_revoke_table[1];
242 /* Check that the hash_size is a power of two */
243 J_ASSERT ((hash_size & (hash_size-1)) == 0);
245 journal->j_revoke->hash_size = hash_size;
247 journal->j_revoke->hash_shift = shift;
249 journal->j_revoke->hash_table =
250 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
251 if (!journal->j_revoke->hash_table) {
252 kfree(journal->j_revoke_table[0]->hash_table);
253 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[0]);
254 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[1]);
255 journal->j_revoke = NULL;
259 for (tmp = 0; tmp < hash_size; tmp++)
260 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
262 spin_lock_init(&journal->j_revoke_lock);
267 /* Destoy a journal's revoke table. The table must already be empty! */
269 void jbd2_journal_destroy_revoke(journal_t *journal)
271 struct jbd2_revoke_table_s *table;
272 struct list_head *hash_list;
275 table = journal->j_revoke_table[0];
279 for (i=0; i<table->hash_size; i++) {
280 hash_list = &table->hash_table[i];
281 J_ASSERT (list_empty(hash_list));
284 kfree(table->hash_table);
285 kmem_cache_free(jbd2_revoke_table_cache, table);
286 journal->j_revoke = NULL;
288 table = journal->j_revoke_table[1];
292 for (i=0; i<table->hash_size; i++) {
293 hash_list = &table->hash_table[i];
294 J_ASSERT (list_empty(hash_list));
297 kfree(table->hash_table);
298 kmem_cache_free(jbd2_revoke_table_cache, table);
299 journal->j_revoke = NULL;
306 * jbd2_journal_revoke: revoke a given buffer_head from the journal. This
307 * prevents the block from being replayed during recovery if we take a
308 * crash after this current transaction commits. Any subsequent
309 * metadata writes of the buffer in this transaction cancel the
312 * Note that this call may block --- it is up to the caller to make
313 * sure that there are no further calls to journal_write_metadata
314 * before the revoke is complete. In ext3, this implies calling the
315 * revoke before clearing the block bitmap when we are deleting
318 * Revoke performs a jbd2_journal_forget on any buffer_head passed in as a
319 * parameter, but does _not_ forget the buffer_head if the bh was only
322 * bh_in may not be a journalled buffer - it may have come off
323 * the hash tables without an attached journal_head.
325 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
329 int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
330 struct buffer_head *bh_in)
332 struct buffer_head *bh = NULL;
334 struct block_device *bdev;
339 BUFFER_TRACE(bh_in, "enter");
341 journal = handle->h_transaction->t_journal;
342 if (!jbd2_journal_set_features(journal, 0, 0, JBD2_FEATURE_INCOMPAT_REVOKE)){
343 J_ASSERT (!"Cannot set revoke feature!");
347 bdev = journal->j_fs_dev;
351 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
353 BUFFER_TRACE(bh, "found on hash");
355 #ifdef JBD_EXPENSIVE_CHECKING
357 struct buffer_head *bh2;
359 /* If there is a different buffer_head lying around in
360 * memory anywhere... */
361 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
363 /* ... and it has RevokeValid status... */
364 if (bh2 != bh && buffer_revokevalid(bh2))
365 /* ...then it better be revoked too,
366 * since it's illegal to create a revoke
367 * record against a buffer_head which is
368 * not marked revoked --- that would
369 * risk missing a subsequent revoke
371 J_ASSERT_BH(bh2, buffer_revoked(bh2));
377 /* We really ought not ever to revoke twice in a row without
378 first having the revoke cancelled: it's illegal to free a
379 block twice without allocating it in between! */
381 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
382 "inconsistent data on disk")) {
387 set_buffer_revoked(bh);
388 set_buffer_revokevalid(bh);
390 BUFFER_TRACE(bh_in, "call jbd2_journal_forget");
391 jbd2_journal_forget(handle, bh_in);
393 BUFFER_TRACE(bh, "call brelse");
398 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr, bh_in);
399 err = insert_revoke_hash(journal, blocknr,
400 handle->h_transaction->t_tid);
401 BUFFER_TRACE(bh_in, "exit");
406 * Cancel an outstanding revoke. For use only internally by the
407 * journaling code (called from jbd2_journal_get_write_access).
409 * We trust buffer_revoked() on the buffer if the buffer is already
410 * being journaled: if there is no revoke pending on the buffer, then we
411 * don't do anything here.
413 * This would break if it were possible for a buffer to be revoked and
414 * discarded, and then reallocated within the same transaction. In such
415 * a case we would have lost the revoked bit, but when we arrived here
416 * the second time we would still have a pending revoke to cancel. So,
417 * do not trust the Revoked bit on buffers unless RevokeValid is also
420 * The caller must have the journal locked.
422 int jbd2_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, cancelling 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 JBD_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);
478 /* journal_switch_revoke table select j_revoke for next transaction
479 * we do not want to suspend any processing until all revokes are
482 void jbd2_journal_switch_revoke_table(journal_t *journal)
486 if (journal->j_revoke == journal->j_revoke_table[0])
487 journal->j_revoke = journal->j_revoke_table[1];
489 journal->j_revoke = journal->j_revoke_table[0];
491 for (i = 0; i < journal->j_revoke->hash_size; i++)
492 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
496 * Write revoke records to the journal for all entries in the current
497 * revoke hash, deleting the entries as we go.
499 * Called with the journal lock held.
502 void jbd2_journal_write_revoke_records(journal_t *journal,
503 transaction_t *transaction)
505 struct journal_head *descriptor;
506 struct jbd2_revoke_record_s *record;
507 struct jbd2_revoke_table_s *revoke;
508 struct list_head *hash_list;
509 int i, offset, count;
515 /* select revoke table for committing transaction */
516 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
517 journal->j_revoke_table[1] : journal->j_revoke_table[0];
519 for (i = 0; i < revoke->hash_size; i++) {
520 hash_list = &revoke->hash_table[i];
522 while (!list_empty(hash_list)) {
523 record = (struct jbd2_revoke_record_s *)
525 write_one_revoke_record(journal, transaction,
526 &descriptor, &offset,
529 list_del(&record->hash);
530 kmem_cache_free(jbd2_revoke_record_cache, record);
534 flush_descriptor(journal, descriptor, offset);
535 jbd_debug(1, "Wrote %d revoke records\n", count);
539 * Write out one revoke record. We need to create a new descriptor
540 * block if the old one is full or if we have not already created one.
543 static void write_one_revoke_record(journal_t *journal,
544 transaction_t *transaction,
545 struct journal_head **descriptorp,
547 struct jbd2_revoke_record_s *record)
549 struct journal_head *descriptor;
551 journal_header_t *header;
553 /* If we are already aborting, this all becomes a noop. We
554 still need to go round the loop in
555 jbd2_journal_write_revoke_records in order to free all of the
556 revoke records: only the IO to the journal is omitted. */
557 if (is_journal_aborted(journal))
560 descriptor = *descriptorp;
563 /* Make sure we have a descriptor with space left for the record */
565 if (offset == journal->j_blocksize) {
566 flush_descriptor(journal, descriptor, offset);
572 descriptor = jbd2_journal_get_descriptor_buffer(journal);
575 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
576 header->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
577 header->h_blocktype = cpu_to_be32(JBD2_REVOKE_BLOCK);
578 header->h_sequence = cpu_to_be32(transaction->t_tid);
580 /* Record it so that we can wait for IO completion later */
581 JBUFFER_TRACE(descriptor, "file as BJ_LogCtl");
582 jbd2_journal_file_buffer(descriptor, transaction, BJ_LogCtl);
584 offset = sizeof(jbd2_journal_revoke_header_t);
585 *descriptorp = descriptor;
588 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) {
589 * ((__be64 *)(&jh2bh(descriptor)->b_data[offset])) =
590 cpu_to_be64(record->blocknr);
594 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
595 cpu_to_be32(record->blocknr);
603 * Flush a revoke descriptor out to the journal. If we are aborting,
604 * this is a noop; otherwise we are generating a buffer which needs to
605 * be waited for during commit, so it has to go onto the appropriate
606 * journal buffer list.
609 static void flush_descriptor(journal_t *journal,
610 struct journal_head *descriptor,
613 jbd2_journal_revoke_header_t *header;
614 struct buffer_head *bh = jh2bh(descriptor);
616 if (is_journal_aborted(journal)) {
621 header = (jbd2_journal_revoke_header_t *) jh2bh(descriptor)->b_data;
622 header->r_count = cpu_to_be32(offset);
623 set_buffer_jwrite(bh);
624 BUFFER_TRACE(bh, "write");
625 set_buffer_dirty(bh);
626 ll_rw_block(SWRITE, 1, &bh);
631 * Revoke support for recovery.
633 * Recovery needs to be able to:
635 * record all revoke records, including the tid of the latest instance
636 * of each revoke in the journal
638 * check whether a given block in a given transaction should be replayed
639 * (ie. has not been revoked by a revoke record in that or a subsequent
642 * empty the revoke table after recovery.
646 * First, setting revoke records. We create a new revoke record for
647 * every block ever revoked in the log as we scan it for recovery, and
648 * we update the existing records if we find multiple revokes for a
652 int jbd2_journal_set_revoke(journal_t *journal,
653 unsigned long long blocknr,
656 struct jbd2_revoke_record_s *record;
658 record = find_revoke_record(journal, blocknr);
660 /* If we have multiple occurrences, only record the
661 * latest sequence number in the hashed record */
662 if (tid_gt(sequence, record->sequence))
663 record->sequence = sequence;
666 return insert_revoke_hash(journal, blocknr, sequence);
670 * Test revoke records. For a given block referenced in the log, has
671 * that block been revoked? A revoke record with a given transaction
672 * sequence number revokes all blocks in that transaction and earlier
673 * ones, but later transactions still need replayed.
676 int jbd2_journal_test_revoke(journal_t *journal,
677 unsigned long long blocknr,
680 struct jbd2_revoke_record_s *record;
682 record = find_revoke_record(journal, blocknr);
685 if (tid_gt(sequence, record->sequence))
691 * Finally, once recovery is over, we need to clear the revoke table so
692 * that it can be reused by the running filesystem.
695 void jbd2_journal_clear_revoke(journal_t *journal)
698 struct list_head *hash_list;
699 struct jbd2_revoke_record_s *record;
700 struct jbd2_revoke_table_s *revoke;
702 revoke = journal->j_revoke;
704 for (i = 0; i < revoke->hash_size; i++) {
705 hash_list = &revoke->hash_table[i];
706 while (!list_empty(hash_list)) {
707 record = (struct jbd2_revoke_record_s*) hash_list->next;
708 list_del(&record->hash);
709 kmem_cache_free(jbd2_revoke_record_cache, record);