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54 * buffer has not been revoked, and cancel_revoke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57 * buffer has been revoked.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>
72 static struct kmem_cache *jbd2_revoke_record_cache;
73 static struct kmem_cache *jbd2_revoke_table_cache;
75 /* Each revoke record represents one single revoked block. During
76 journal replay, this involves recording the transaction ID of the
77 last transaction to revoke this block. */
79 struct jbd2_revoke_record_s
81 struct list_head hash;
82 tid_t sequence; /* Used for recovery only */
83 unsigned long long blocknr;
87 /* The revoke table is just a simple hash table of revoke records. */
88 struct jbd2_revoke_table_s
90 /* It is conceivable that we might want a larger hash table
91 * for recovery. Must be a power of two. */
94 struct list_head *hash_table;
99 static void write_one_revoke_record(journal_t *, transaction_t *,
100 struct journal_head **, int *,
101 struct jbd2_revoke_record_s *);
102 static void flush_descriptor(journal_t *, struct journal_head *, int);
105 /* Utility functions to maintain the revoke table */
107 /* Borrowed from buffer.c: this is a tried and tested block hash function */
108 static inline int hash(journal_t *journal, unsigned long long block)
110 struct jbd2_revoke_table_s *table = journal->j_revoke;
111 int hash_shift = table->hash_shift;
112 int hash = (int)block ^ (int)((block >> 31) >> 1);
114 return ((hash << (hash_shift - 6)) ^
116 (hash << (hash_shift - 12))) & (table->hash_size - 1);
119 static int insert_revoke_hash(journal_t *journal, unsigned long long blocknr,
122 struct list_head *hash_list;
123 struct jbd2_revoke_record_s *record;
126 record = kmem_cache_alloc(jbd2_revoke_record_cache, GFP_NOFS);
130 record->sequence = seq;
131 record->blocknr = blocknr;
132 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
133 spin_lock(&journal->j_revoke_lock);
134 list_add(&record->hash, hash_list);
135 spin_unlock(&journal->j_revoke_lock);
139 if (!journal_oom_retry)
141 jbd_debug(1, "ENOMEM in %s, retrying\n", __FUNCTION__);
146 /* Find a revoke record in the journal's hash table. */
148 static struct jbd2_revoke_record_s *find_revoke_record(journal_t *journal,
149 unsigned long long blocknr)
151 struct list_head *hash_list;
152 struct jbd2_revoke_record_s *record;
154 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
156 spin_lock(&journal->j_revoke_lock);
157 record = (struct jbd2_revoke_record_s *) hash_list->next;
158 while (&(record->hash) != hash_list) {
159 if (record->blocknr == blocknr) {
160 spin_unlock(&journal->j_revoke_lock);
163 record = (struct jbd2_revoke_record_s *) record->hash.next;
165 spin_unlock(&journal->j_revoke_lock);
169 int __init jbd2_journal_init_revoke_caches(void)
171 jbd2_revoke_record_cache = kmem_cache_create("jbd2_revoke_record",
172 sizeof(struct jbd2_revoke_record_s),
173 0, SLAB_HWCACHE_ALIGN, NULL, NULL);
174 if (jbd2_revoke_record_cache == 0)
177 jbd2_revoke_table_cache = kmem_cache_create("jbd2_revoke_table",
178 sizeof(struct jbd2_revoke_table_s),
180 if (jbd2_revoke_table_cache == 0) {
181 kmem_cache_destroy(jbd2_revoke_record_cache);
182 jbd2_revoke_record_cache = NULL;
188 void jbd2_journal_destroy_revoke_caches(void)
190 kmem_cache_destroy(jbd2_revoke_record_cache);
191 jbd2_revoke_record_cache = NULL;
192 kmem_cache_destroy(jbd2_revoke_table_cache);
193 jbd2_revoke_table_cache = NULL;
196 /* Initialise the revoke table for a given journal to a given size. */
198 int jbd2_journal_init_revoke(journal_t *journal, int hash_size)
202 J_ASSERT (journal->j_revoke_table[0] == NULL);
206 while((tmp >>= 1UL) != 0UL)
209 journal->j_revoke_table[0] = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
210 if (!journal->j_revoke_table[0])
212 journal->j_revoke = journal->j_revoke_table[0];
214 /* Check that the hash_size is a power of two */
215 J_ASSERT ((hash_size & (hash_size-1)) == 0);
217 journal->j_revoke->hash_size = hash_size;
219 journal->j_revoke->hash_shift = shift;
221 journal->j_revoke->hash_table =
222 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
223 if (!journal->j_revoke->hash_table) {
224 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[0]);
225 journal->j_revoke = NULL;
229 for (tmp = 0; tmp < hash_size; tmp++)
230 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
232 journal->j_revoke_table[1] = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
233 if (!journal->j_revoke_table[1]) {
234 kfree(journal->j_revoke_table[0]->hash_table);
235 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[0]);
239 journal->j_revoke = journal->j_revoke_table[1];
241 /* Check that the hash_size is a power of two */
242 J_ASSERT ((hash_size & (hash_size-1)) == 0);
244 journal->j_revoke->hash_size = hash_size;
246 journal->j_revoke->hash_shift = shift;
248 journal->j_revoke->hash_table =
249 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
250 if (!journal->j_revoke->hash_table) {
251 kfree(journal->j_revoke_table[0]->hash_table);
252 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[0]);
253 kmem_cache_free(jbd2_revoke_table_cache, journal->j_revoke_table[1]);
254 journal->j_revoke = NULL;
258 for (tmp = 0; tmp < hash_size; tmp++)
259 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
261 spin_lock_init(&journal->j_revoke_lock);
266 /* Destoy a journal's revoke table. The table must already be empty! */
268 void jbd2_journal_destroy_revoke(journal_t *journal)
270 struct jbd2_revoke_table_s *table;
271 struct list_head *hash_list;
274 table = journal->j_revoke_table[0];
278 for (i=0; i<table->hash_size; i++) {
279 hash_list = &table->hash_table[i];
280 J_ASSERT (list_empty(hash_list));
283 kfree(table->hash_table);
284 kmem_cache_free(jbd2_revoke_table_cache, table);
285 journal->j_revoke = NULL;
287 table = journal->j_revoke_table[1];
291 for (i=0; i<table->hash_size; i++) {
292 hash_list = &table->hash_table[i];
293 J_ASSERT (list_empty(hash_list));
296 kfree(table->hash_table);
297 kmem_cache_free(jbd2_revoke_table_cache, table);
298 journal->j_revoke = NULL;
305 * jbd2_journal_revoke: revoke a given buffer_head from the journal. This
306 * prevents the block from being replayed during recovery if we take a
307 * crash after this current transaction commits. Any subsequent
308 * metadata writes of the buffer in this transaction cancel the
311 * Note that this call may block --- it is up to the caller to make
312 * sure that there are no further calls to journal_write_metadata
313 * before the revoke is complete. In ext3, this implies calling the
314 * revoke before clearing the block bitmap when we are deleting
317 * Revoke performs a jbd2_journal_forget on any buffer_head passed in as a
318 * parameter, but does _not_ forget the buffer_head if the bh was only
321 * bh_in may not be a journalled buffer - it may have come off
322 * the hash tables without an attached journal_head.
324 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
328 int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
329 struct buffer_head *bh_in)
331 struct buffer_head *bh = NULL;
333 struct block_device *bdev;
338 BUFFER_TRACE(bh_in, "enter");
340 journal = handle->h_transaction->t_journal;
341 if (!jbd2_journal_set_features(journal, 0, 0, JBD2_FEATURE_INCOMPAT_REVOKE)){
342 J_ASSERT (!"Cannot set revoke feature!");
346 bdev = journal->j_fs_dev;
350 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
352 BUFFER_TRACE(bh, "found on hash");
354 #ifdef JBD_EXPENSIVE_CHECKING
356 struct buffer_head *bh2;
358 /* If there is a different buffer_head lying around in
359 * memory anywhere... */
360 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
362 /* ... and it has RevokeValid status... */
363 if (bh2 != bh && buffer_revokevalid(bh2))
364 /* ...then it better be revoked too,
365 * since it's illegal to create a revoke
366 * record against a buffer_head which is
367 * not marked revoked --- that would
368 * risk missing a subsequent revoke
370 J_ASSERT_BH(bh2, buffer_revoked(bh2));
376 /* We really ought not ever to revoke twice in a row without
377 first having the revoke cancelled: it's illegal to free a
378 block twice without allocating it in between! */
380 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
381 "inconsistent data on disk")) {
386 set_buffer_revoked(bh);
387 set_buffer_revokevalid(bh);
389 BUFFER_TRACE(bh_in, "call jbd2_journal_forget");
390 jbd2_journal_forget(handle, bh_in);
392 BUFFER_TRACE(bh, "call brelse");
397 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr, bh_in);
398 err = insert_revoke_hash(journal, blocknr,
399 handle->h_transaction->t_tid);
400 BUFFER_TRACE(bh_in, "exit");
405 * Cancel an outstanding revoke. For use only internally by the
406 * journaling code (called from jbd2_journal_get_write_access).
408 * We trust buffer_revoked() on the buffer if the buffer is already
409 * being journaled: if there is no revoke pending on the buffer, then we
410 * don't do anything here.
412 * This would break if it were possible for a buffer to be revoked and
413 * discarded, and then reallocated within the same transaction. In such
414 * a case we would have lost the revoked bit, but when we arrived here
415 * the second time we would still have a pending revoke to cancel. So,
416 * do not trust the Revoked bit on buffers unless RevokeValid is also
419 * The caller must have the journal locked.
421 int jbd2_journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
423 struct jbd2_revoke_record_s *record;
424 journal_t *journal = handle->h_transaction->t_journal;
426 int did_revoke = 0; /* akpm: debug */
427 struct buffer_head *bh = jh2bh(jh);
429 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
431 /* Is the existing Revoke bit valid? If so, we trust it, and
432 * only perform the full cancel if the revoke bit is set. If
433 * not, we can't trust the revoke bit, and we need to do the
434 * full search for a revoke record. */
435 if (test_set_buffer_revokevalid(bh)) {
436 need_cancel = test_clear_buffer_revoked(bh);
439 clear_buffer_revoked(bh);
443 record = find_revoke_record(journal, bh->b_blocknr);
445 jbd_debug(4, "cancelled existing revoke on "
446 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
447 spin_lock(&journal->j_revoke_lock);
448 list_del(&record->hash);
449 spin_unlock(&journal->j_revoke_lock);
450 kmem_cache_free(jbd2_revoke_record_cache, record);
455 #ifdef JBD_EXPENSIVE_CHECKING
456 /* There better not be one left behind by now! */
457 record = find_revoke_record(journal, bh->b_blocknr);
458 J_ASSERT_JH(jh, record == NULL);
461 /* Finally, have we just cleared revoke on an unhashed
462 * buffer_head? If so, we'd better make sure we clear the
463 * revoked status on any hashed alias too, otherwise the revoke
464 * state machine will get very upset later on. */
466 struct buffer_head *bh2;
467 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
470 clear_buffer_revoked(bh2);
477 /* journal_switch_revoke table select j_revoke for next transaction
478 * we do not want to suspend any processing until all revokes are
481 void jbd2_journal_switch_revoke_table(journal_t *journal)
485 if (journal->j_revoke == journal->j_revoke_table[0])
486 journal->j_revoke = journal->j_revoke_table[1];
488 journal->j_revoke = journal->j_revoke_table[0];
490 for (i = 0; i < journal->j_revoke->hash_size; i++)
491 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
495 * Write revoke records to the journal for all entries in the current
496 * revoke hash, deleting the entries as we go.
498 * Called with the journal lock held.
501 void jbd2_journal_write_revoke_records(journal_t *journal,
502 transaction_t *transaction)
504 struct journal_head *descriptor;
505 struct jbd2_revoke_record_s *record;
506 struct jbd2_revoke_table_s *revoke;
507 struct list_head *hash_list;
508 int i, offset, count;
514 /* select revoke table for committing transaction */
515 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
516 journal->j_revoke_table[1] : journal->j_revoke_table[0];
518 for (i = 0; i < revoke->hash_size; i++) {
519 hash_list = &revoke->hash_table[i];
521 while (!list_empty(hash_list)) {
522 record = (struct jbd2_revoke_record_s *)
524 write_one_revoke_record(journal, transaction,
525 &descriptor, &offset,
528 list_del(&record->hash);
529 kmem_cache_free(jbd2_revoke_record_cache, record);
533 flush_descriptor(journal, descriptor, offset);
534 jbd_debug(1, "Wrote %d revoke records\n", count);
538 * Write out one revoke record. We need to create a new descriptor
539 * block if the old one is full or if we have not already created one.
542 static void write_one_revoke_record(journal_t *journal,
543 transaction_t *transaction,
544 struct journal_head **descriptorp,
546 struct jbd2_revoke_record_s *record)
548 struct journal_head *descriptor;
550 journal_header_t *header;
552 /* If we are already aborting, this all becomes a noop. We
553 still need to go round the loop in
554 jbd2_journal_write_revoke_records in order to free all of the
555 revoke records: only the IO to the journal is omitted. */
556 if (is_journal_aborted(journal))
559 descriptor = *descriptorp;
562 /* Make sure we have a descriptor with space left for the record */
564 if (offset == journal->j_blocksize) {
565 flush_descriptor(journal, descriptor, offset);
571 descriptor = jbd2_journal_get_descriptor_buffer(journal);
574 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
575 header->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
576 header->h_blocktype = cpu_to_be32(JBD2_REVOKE_BLOCK);
577 header->h_sequence = cpu_to_be32(transaction->t_tid);
579 /* Record it so that we can wait for IO completion later */
580 JBUFFER_TRACE(descriptor, "file as BJ_LogCtl");
581 jbd2_journal_file_buffer(descriptor, transaction, BJ_LogCtl);
583 offset = sizeof(jbd2_journal_revoke_header_t);
584 *descriptorp = descriptor;
587 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) {
588 * ((__be64 *)(&jh2bh(descriptor)->b_data[offset])) =
589 cpu_to_be64(record->blocknr);
593 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
594 cpu_to_be32(record->blocknr);
602 * Flush a revoke descriptor out to the journal. If we are aborting,
603 * this is a noop; otherwise we are generating a buffer which needs to
604 * be waited for during commit, so it has to go onto the appropriate
605 * journal buffer list.
608 static void flush_descriptor(journal_t *journal,
609 struct journal_head *descriptor,
612 jbd2_journal_revoke_header_t *header;
613 struct buffer_head *bh = jh2bh(descriptor);
615 if (is_journal_aborted(journal)) {
620 header = (jbd2_journal_revoke_header_t *) jh2bh(descriptor)->b_data;
621 header->r_count = cpu_to_be32(offset);
622 set_buffer_jwrite(bh);
623 BUFFER_TRACE(bh, "write");
624 set_buffer_dirty(bh);
625 ll_rw_block(SWRITE, 1, &bh);
630 * Revoke support for recovery.
632 * Recovery needs to be able to:
634 * record all revoke records, including the tid of the latest instance
635 * of each revoke in the journal
637 * check whether a given block in a given transaction should be replayed
638 * (ie. has not been revoked by a revoke record in that or a subsequent
641 * empty the revoke table after recovery.
645 * First, setting revoke records. We create a new revoke record for
646 * every block ever revoked in the log as we scan it for recovery, and
647 * we update the existing records if we find multiple revokes for a
651 int jbd2_journal_set_revoke(journal_t *journal,
652 unsigned long long blocknr,
655 struct jbd2_revoke_record_s *record;
657 record = find_revoke_record(journal, blocknr);
659 /* If we have multiple occurrences, only record the
660 * latest sequence number in the hashed record */
661 if (tid_gt(sequence, record->sequence))
662 record->sequence = sequence;
665 return insert_revoke_hash(journal, blocknr, sequence);
669 * Test revoke records. For a given block referenced in the log, has
670 * that block been revoked? A revoke record with a given transaction
671 * sequence number revokes all blocks in that transaction and earlier
672 * ones, but later transactions still need replayed.
675 int jbd2_journal_test_revoke(journal_t *journal,
676 unsigned long long blocknr,
679 struct jbd2_revoke_record_s *record;
681 record = find_revoke_record(journal, blocknr);
684 if (tid_gt(sequence, record->sequence))
690 * Finally, once recovery is over, we need to clear the revoke table so
691 * that it can be reused by the running filesystem.
694 void jbd2_journal_clear_revoke(journal_t *journal)
697 struct list_head *hash_list;
698 struct jbd2_revoke_record_s *record;
699 struct jbd2_revoke_table_s *revoke;
701 revoke = journal->j_revoke;
703 for (i = 0; i < revoke->hash_size; i++) {
704 hash_list = &revoke->hash_table[i];
705 while (!list_empty(hash_list)) {
706 record = (struct jbd2_revoke_record_s*) hash_list->next;
707 list_del(&record->hash);
708 kmem_cache_free(jbd2_revoke_record_cache, record);