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54 * buffer has not been revoked, and cancel_revoke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57 * buffer has been revoked.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", __func__);
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 static struct jbd2_revoke_table_s *jbd2_journal_init_revoke_table(int hash_size)
211 struct jbd2_revoke_table_s *table;
213 table = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
217 while((tmp >>= 1UL) != 0UL)
220 table->hash_size = hash_size;
221 table->hash_shift = shift;
223 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
224 if (!table->hash_table) {
225 kmem_cache_free(jbd2_revoke_table_cache, table);
230 for (tmp = 0; tmp < hash_size; tmp++)
231 INIT_LIST_HEAD(&table->hash_table[tmp]);
237 static void jbd2_journal_destroy_revoke_table(struct jbd2_revoke_table_s *table)
240 struct list_head *hash_list;
242 for (i = 0; i < table->hash_size; i++) {
243 hash_list = &table->hash_table[i];
244 J_ASSERT(list_empty(hash_list));
247 kfree(table->hash_table);
248 kmem_cache_free(jbd2_revoke_table_cache, table);
251 /* Initialise the revoke table for a given journal to a given size. */
252 int jbd2_journal_init_revoke(journal_t *journal, int hash_size)
254 J_ASSERT(journal->j_revoke_table[0] == NULL);
255 J_ASSERT(is_power_of_2(hash_size));
257 journal->j_revoke_table[0] = jbd2_journal_init_revoke_table(hash_size);
258 if (!journal->j_revoke_table[0])
261 journal->j_revoke_table[1] = jbd2_journal_init_revoke_table(hash_size);
262 if (!journal->j_revoke_table[1])
265 journal->j_revoke = journal->j_revoke_table[1];
267 spin_lock_init(&journal->j_revoke_lock);
272 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
277 /* Destroy a journal's revoke table. The table must already be empty! */
278 void jbd2_journal_destroy_revoke(journal_t *journal)
280 journal->j_revoke = NULL;
281 if (journal->j_revoke_table[0])
282 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
283 if (journal->j_revoke_table[1])
284 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[1]);
291 * jbd2_journal_revoke: revoke a given buffer_head from the journal. This
292 * prevents the block from being replayed during recovery if we take a
293 * crash after this current transaction commits. Any subsequent
294 * metadata writes of the buffer in this transaction cancel the
297 * Note that this call may block --- it is up to the caller to make
298 * sure that there are no further calls to journal_write_metadata
299 * before the revoke is complete. In ext3, this implies calling the
300 * revoke before clearing the block bitmap when we are deleting
303 * Revoke performs a jbd2_journal_forget on any buffer_head passed in as a
304 * parameter, but does _not_ forget the buffer_head if the bh was only
307 * bh_in may not be a journalled buffer - it may have come off
308 * the hash tables without an attached journal_head.
310 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
314 int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
315 struct buffer_head *bh_in)
317 struct buffer_head *bh = NULL;
319 struct block_device *bdev;
324 BUFFER_TRACE(bh_in, "enter");
326 journal = handle->h_transaction->t_journal;
327 if (!jbd2_journal_set_features(journal, 0, 0, JBD2_FEATURE_INCOMPAT_REVOKE)){
328 J_ASSERT (!"Cannot set revoke feature!");
332 bdev = journal->j_fs_dev;
336 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
338 BUFFER_TRACE(bh, "found on hash");
340 #ifdef JBD2_EXPENSIVE_CHECKING
342 struct buffer_head *bh2;
344 /* If there is a different buffer_head lying around in
345 * memory anywhere... */
346 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
348 /* ... and it has RevokeValid status... */
349 if (bh2 != bh && buffer_revokevalid(bh2))
350 /* ...then it better be revoked too,
351 * since it's illegal to create a revoke
352 * record against a buffer_head which is
353 * not marked revoked --- that would
354 * risk missing a subsequent revoke
356 J_ASSERT_BH(bh2, buffer_revoked(bh2));
362 /* We really ought not ever to revoke twice in a row without
363 first having the revoke cancelled: it's illegal to free a
364 block twice without allocating it in between! */
366 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
367 "inconsistent data on disk")) {
372 set_buffer_revoked(bh);
373 set_buffer_revokevalid(bh);
375 BUFFER_TRACE(bh_in, "call jbd2_journal_forget");
376 jbd2_journal_forget(handle, bh_in);
378 BUFFER_TRACE(bh, "call brelse");
383 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr, bh_in);
384 err = insert_revoke_hash(journal, blocknr,
385 handle->h_transaction->t_tid);
386 BUFFER_TRACE(bh_in, "exit");
391 * Cancel an outstanding revoke. For use only internally by the
392 * journaling code (called from jbd2_journal_get_write_access).
394 * We trust buffer_revoked() on the buffer if the buffer is already
395 * being journaled: if there is no revoke pending on the buffer, then we
396 * don't do anything here.
398 * This would break if it were possible for a buffer to be revoked and
399 * discarded, and then reallocated within the same transaction. In such
400 * a case we would have lost the revoked bit, but when we arrived here
401 * the second time we would still have a pending revoke to cancel. So,
402 * do not trust the Revoked bit on buffers unless RevokeValid is also
405 * The caller must have the journal locked.
407 int jbd2_journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
409 struct jbd2_revoke_record_s *record;
410 journal_t *journal = handle->h_transaction->t_journal;
412 int did_revoke = 0; /* akpm: debug */
413 struct buffer_head *bh = jh2bh(jh);
415 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
417 /* Is the existing Revoke bit valid? If so, we trust it, and
418 * only perform the full cancel if the revoke bit is set. If
419 * not, we can't trust the revoke bit, and we need to do the
420 * full search for a revoke record. */
421 if (test_set_buffer_revokevalid(bh)) {
422 need_cancel = test_clear_buffer_revoked(bh);
425 clear_buffer_revoked(bh);
429 record = find_revoke_record(journal, bh->b_blocknr);
431 jbd_debug(4, "cancelled existing revoke on "
432 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
433 spin_lock(&journal->j_revoke_lock);
434 list_del(&record->hash);
435 spin_unlock(&journal->j_revoke_lock);
436 kmem_cache_free(jbd2_revoke_record_cache, record);
441 #ifdef JBD2_EXPENSIVE_CHECKING
442 /* There better not be one left behind by now! */
443 record = find_revoke_record(journal, bh->b_blocknr);
444 J_ASSERT_JH(jh, record == NULL);
447 /* Finally, have we just cleared revoke on an unhashed
448 * buffer_head? If so, we'd better make sure we clear the
449 * revoked status on any hashed alias too, otherwise the revoke
450 * state machine will get very upset later on. */
452 struct buffer_head *bh2;
453 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
456 clear_buffer_revoked(bh2);
463 /* journal_switch_revoke table select j_revoke for next transaction
464 * we do not want to suspend any processing until all revokes are
467 void jbd2_journal_switch_revoke_table(journal_t *journal)
471 if (journal->j_revoke == journal->j_revoke_table[0])
472 journal->j_revoke = journal->j_revoke_table[1];
474 journal->j_revoke = journal->j_revoke_table[0];
476 for (i = 0; i < journal->j_revoke->hash_size; i++)
477 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
481 * Write revoke records to the journal for all entries in the current
482 * revoke hash, deleting the entries as we go.
484 * Called with the journal lock held.
487 void jbd2_journal_write_revoke_records(journal_t *journal,
488 transaction_t *transaction)
490 struct journal_head *descriptor;
491 struct jbd2_revoke_record_s *record;
492 struct jbd2_revoke_table_s *revoke;
493 struct list_head *hash_list;
494 int i, offset, count;
500 /* select revoke table for committing transaction */
501 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
502 journal->j_revoke_table[1] : journal->j_revoke_table[0];
504 for (i = 0; i < revoke->hash_size; i++) {
505 hash_list = &revoke->hash_table[i];
507 while (!list_empty(hash_list)) {
508 record = (struct jbd2_revoke_record_s *)
510 write_one_revoke_record(journal, transaction,
511 &descriptor, &offset,
514 list_del(&record->hash);
515 kmem_cache_free(jbd2_revoke_record_cache, record);
519 flush_descriptor(journal, descriptor, offset);
520 jbd_debug(1, "Wrote %d revoke records\n", count);
524 * Write out one revoke record. We need to create a new descriptor
525 * block if the old one is full or if we have not already created one.
528 static void write_one_revoke_record(journal_t *journal,
529 transaction_t *transaction,
530 struct journal_head **descriptorp,
532 struct jbd2_revoke_record_s *record)
534 struct journal_head *descriptor;
536 journal_header_t *header;
538 /* If we are already aborting, this all becomes a noop. We
539 still need to go round the loop in
540 jbd2_journal_write_revoke_records in order to free all of the
541 revoke records: only the IO to the journal is omitted. */
542 if (is_journal_aborted(journal))
545 descriptor = *descriptorp;
548 /* Make sure we have a descriptor with space left for the record */
550 if (offset == journal->j_blocksize) {
551 flush_descriptor(journal, descriptor, offset);
557 descriptor = jbd2_journal_get_descriptor_buffer(journal);
560 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
561 header->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
562 header->h_blocktype = cpu_to_be32(JBD2_REVOKE_BLOCK);
563 header->h_sequence = cpu_to_be32(transaction->t_tid);
565 /* Record it so that we can wait for IO completion later */
566 JBUFFER_TRACE(descriptor, "file as BJ_LogCtl");
567 jbd2_journal_file_buffer(descriptor, transaction, BJ_LogCtl);
569 offset = sizeof(jbd2_journal_revoke_header_t);
570 *descriptorp = descriptor;
573 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) {
574 * ((__be64 *)(&jh2bh(descriptor)->b_data[offset])) =
575 cpu_to_be64(record->blocknr);
579 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
580 cpu_to_be32(record->blocknr);
588 * Flush a revoke descriptor out to the journal. If we are aborting,
589 * this is a noop; otherwise we are generating a buffer which needs to
590 * be waited for during commit, so it has to go onto the appropriate
591 * journal buffer list.
594 static void flush_descriptor(journal_t *journal,
595 struct journal_head *descriptor,
598 jbd2_journal_revoke_header_t *header;
599 struct buffer_head *bh = jh2bh(descriptor);
601 if (is_journal_aborted(journal)) {
606 header = (jbd2_journal_revoke_header_t *) jh2bh(descriptor)->b_data;
607 header->r_count = cpu_to_be32(offset);
608 set_buffer_jwrite(bh);
609 BUFFER_TRACE(bh, "write");
610 set_buffer_dirty(bh);
611 ll_rw_block(SWRITE, 1, &bh);
616 * Revoke support for recovery.
618 * Recovery needs to be able to:
620 * record all revoke records, including the tid of the latest instance
621 * of each revoke in the journal
623 * check whether a given block in a given transaction should be replayed
624 * (ie. has not been revoked by a revoke record in that or a subsequent
627 * empty the revoke table after recovery.
631 * First, setting revoke records. We create a new revoke record for
632 * every block ever revoked in the log as we scan it for recovery, and
633 * we update the existing records if we find multiple revokes for a
637 int jbd2_journal_set_revoke(journal_t *journal,
638 unsigned long long blocknr,
641 struct jbd2_revoke_record_s *record;
643 record = find_revoke_record(journal, blocknr);
645 /* If we have multiple occurrences, only record the
646 * latest sequence number in the hashed record */
647 if (tid_gt(sequence, record->sequence))
648 record->sequence = sequence;
651 return insert_revoke_hash(journal, blocknr, sequence);
655 * Test revoke records. For a given block referenced in the log, has
656 * that block been revoked? A revoke record with a given transaction
657 * sequence number revokes all blocks in that transaction and earlier
658 * ones, but later transactions still need replayed.
661 int jbd2_journal_test_revoke(journal_t *journal,
662 unsigned long long blocknr,
665 struct jbd2_revoke_record_s *record;
667 record = find_revoke_record(journal, blocknr);
670 if (tid_gt(sequence, record->sequence))
676 * Finally, once recovery is over, we need to clear the revoke table so
677 * that it can be reused by the running filesystem.
680 void jbd2_journal_clear_revoke(journal_t *journal)
683 struct list_head *hash_list;
684 struct jbd2_revoke_record_s *record;
685 struct jbd2_revoke_table_s *revoke;
687 revoke = journal->j_revoke;
689 for (i = 0; i < revoke->hash_size; i++) {
690 hash_list = &revoke->hash_table[i];
691 while (!list_empty(hash_list)) {
692 record = (struct jbd2_revoke_record_s*) hash_list->next;
693 list_del(&record->hash);
694 kmem_cache_free(jbd2_revoke_record_cache, record);