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 * 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54 * buffer has not been revoked, and cancel_revoke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57 * buffer has been revoked.
60 * We keep two hash tables of revoke records. One hashtable belongs to the
61 * running transaction (is pointed to by journal->j_revoke), the other one
62 * belongs to the committing transaction. Accesses to the second hash table
63 * happen only from the kjournald and no other thread touches this table. Also
64 * journal_switch_revoke_table() which switches which hashtable belongs to the
65 * running and which to the committing transaction is called only from
66 * kjournald. Therefore we need no locks when accessing the hashtable belonging
67 * to the committing transaction.
69 * All users operating on the hash table belonging to the running transaction
70 * have a handle to the transaction. Therefore they are safe from kjournald
71 * switching hash tables under them. For operations on the lists of entries in
72 * the hash table j_revoke_lock is used.
74 * Finally, also replay code uses the hash tables but at this moment no one else
75 * can touch them (filesystem isn't mounted yet) and hence no locking is
82 #include <linux/time.h>
84 #include <linux/jbd.h>
85 #include <linux/errno.h>
86 #include <linux/slab.h>
87 #include <linux/list.h>
88 #include <linux/init.h>
89 #include <linux/bio.h>
91 #include <linux/log2.h>
93 static struct kmem_cache *revoke_record_cache;
94 static struct kmem_cache *revoke_table_cache;
96 /* Each revoke record represents one single revoked block. During
97 journal replay, this involves recording the transaction ID of the
98 last transaction to revoke this block. */
100 struct jbd_revoke_record_s
102 struct list_head hash;
103 tid_t sequence; /* Used for recovery only */
104 unsigned int blocknr;
108 /* The revoke table is just a simple hash table of revoke records. */
109 struct jbd_revoke_table_s
111 /* It is conceivable that we might want a larger hash table
112 * for recovery. Must be a power of two. */
115 struct list_head *hash_table;
120 static void write_one_revoke_record(journal_t *, transaction_t *,
121 struct journal_head **, int *,
122 struct jbd_revoke_record_s *, int);
123 static void flush_descriptor(journal_t *, struct journal_head *, int, int);
126 /* Utility functions to maintain the revoke table */
128 /* Borrowed from buffer.c: this is a tried and tested block hash function */
129 static inline int hash(journal_t *journal, unsigned int block)
131 struct jbd_revoke_table_s *table = journal->j_revoke;
132 int hash_shift = table->hash_shift;
134 return ((block << (hash_shift - 6)) ^
136 (block << (hash_shift - 12))) & (table->hash_size - 1);
139 static int insert_revoke_hash(journal_t *journal, unsigned int blocknr,
142 struct list_head *hash_list;
143 struct jbd_revoke_record_s *record;
146 record = kmem_cache_alloc(revoke_record_cache, GFP_NOFS);
150 record->sequence = seq;
151 record->blocknr = blocknr;
152 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
153 spin_lock(&journal->j_revoke_lock);
154 list_add(&record->hash, hash_list);
155 spin_unlock(&journal->j_revoke_lock);
159 if (!journal_oom_retry)
161 jbd_debug(1, "ENOMEM in %s, retrying\n", __func__);
166 /* Find a revoke record in the journal's hash table. */
168 static struct jbd_revoke_record_s *find_revoke_record(journal_t *journal,
169 unsigned int blocknr)
171 struct list_head *hash_list;
172 struct jbd_revoke_record_s *record;
174 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
176 spin_lock(&journal->j_revoke_lock);
177 record = (struct jbd_revoke_record_s *) hash_list->next;
178 while (&(record->hash) != hash_list) {
179 if (record->blocknr == blocknr) {
180 spin_unlock(&journal->j_revoke_lock);
183 record = (struct jbd_revoke_record_s *) record->hash.next;
185 spin_unlock(&journal->j_revoke_lock);
189 void journal_destroy_revoke_caches(void)
191 if (revoke_record_cache) {
192 kmem_cache_destroy(revoke_record_cache);
193 revoke_record_cache = NULL;
195 if (revoke_table_cache) {
196 kmem_cache_destroy(revoke_table_cache);
197 revoke_table_cache = NULL;
201 int __init journal_init_revoke_caches(void)
203 J_ASSERT(!revoke_record_cache);
204 J_ASSERT(!revoke_table_cache);
206 revoke_record_cache = kmem_cache_create("revoke_record",
207 sizeof(struct jbd_revoke_record_s),
209 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
211 if (!revoke_record_cache)
212 goto record_cache_failure;
214 revoke_table_cache = kmem_cache_create("revoke_table",
215 sizeof(struct jbd_revoke_table_s),
216 0, SLAB_TEMPORARY, NULL);
217 if (!revoke_table_cache)
218 goto table_cache_failure;
223 journal_destroy_revoke_caches();
224 record_cache_failure:
228 static struct jbd_revoke_table_s *journal_init_revoke_table(int hash_size)
232 struct jbd_revoke_table_s *table;
234 table = kmem_cache_alloc(revoke_table_cache, GFP_KERNEL);
238 while((tmp >>= 1UL) != 0UL)
241 table->hash_size = hash_size;
242 table->hash_shift = shift;
244 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
245 if (!table->hash_table) {
246 kmem_cache_free(revoke_table_cache, table);
251 for (tmp = 0; tmp < hash_size; tmp++)
252 INIT_LIST_HEAD(&table->hash_table[tmp]);
258 static void journal_destroy_revoke_table(struct jbd_revoke_table_s *table)
261 struct list_head *hash_list;
263 for (i = 0; i < table->hash_size; i++) {
264 hash_list = &table->hash_table[i];
265 J_ASSERT(list_empty(hash_list));
268 kfree(table->hash_table);
269 kmem_cache_free(revoke_table_cache, table);
272 /* Initialise the revoke table for a given journal to a given size. */
273 int journal_init_revoke(journal_t *journal, int hash_size)
275 J_ASSERT(journal->j_revoke_table[0] == NULL);
276 J_ASSERT(is_power_of_2(hash_size));
278 journal->j_revoke_table[0] = journal_init_revoke_table(hash_size);
279 if (!journal->j_revoke_table[0])
282 journal->j_revoke_table[1] = journal_init_revoke_table(hash_size);
283 if (!journal->j_revoke_table[1])
286 journal->j_revoke = journal->j_revoke_table[1];
288 spin_lock_init(&journal->j_revoke_lock);
293 journal_destroy_revoke_table(journal->j_revoke_table[0]);
298 /* Destroy a journal's revoke table. The table must already be empty! */
299 void journal_destroy_revoke(journal_t *journal)
301 journal->j_revoke = NULL;
302 if (journal->j_revoke_table[0])
303 journal_destroy_revoke_table(journal->j_revoke_table[0]);
304 if (journal->j_revoke_table[1])
305 journal_destroy_revoke_table(journal->j_revoke_table[1]);
312 * journal_revoke: revoke a given buffer_head from the journal. This
313 * prevents the block from being replayed during recovery if we take a
314 * crash after this current transaction commits. Any subsequent
315 * metadata writes of the buffer in this transaction cancel the
318 * Note that this call may block --- it is up to the caller to make
319 * sure that there are no further calls to journal_write_metadata
320 * before the revoke is complete. In ext3, this implies calling the
321 * revoke before clearing the block bitmap when we are deleting
324 * Revoke performs a journal_forget on any buffer_head passed in as a
325 * parameter, but does _not_ forget the buffer_head if the bh was only
328 * bh_in may not be a journalled buffer - it may have come off
329 * the hash tables without an attached journal_head.
331 * If bh_in is non-zero, journal_revoke() will decrement its b_count
335 int journal_revoke(handle_t *handle, unsigned int blocknr,
336 struct buffer_head *bh_in)
338 struct buffer_head *bh = NULL;
340 struct block_device *bdev;
345 BUFFER_TRACE(bh_in, "enter");
347 journal = handle->h_transaction->t_journal;
348 if (!journal_set_features(journal, 0, 0, JFS_FEATURE_INCOMPAT_REVOKE)){
349 J_ASSERT (!"Cannot set revoke feature!");
353 bdev = journal->j_fs_dev;
357 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
359 BUFFER_TRACE(bh, "found on hash");
361 #ifdef JBD_EXPENSIVE_CHECKING
363 struct buffer_head *bh2;
365 /* If there is a different buffer_head lying around in
366 * memory anywhere... */
367 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
369 /* ... and it has RevokeValid status... */
370 if (bh2 != bh && buffer_revokevalid(bh2))
371 /* ...then it better be revoked too,
372 * since it's illegal to create a revoke
373 * record against a buffer_head which is
374 * not marked revoked --- that would
375 * risk missing a subsequent revoke
377 J_ASSERT_BH(bh2, buffer_revoked(bh2));
383 /* We really ought not ever to revoke twice in a row without
384 first having the revoke cancelled: it's illegal to free a
385 block twice without allocating it in between! */
387 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
388 "inconsistent data on disk")) {
393 set_buffer_revoked(bh);
394 set_buffer_revokevalid(bh);
396 BUFFER_TRACE(bh_in, "call journal_forget");
397 journal_forget(handle, bh_in);
399 BUFFER_TRACE(bh, "call brelse");
404 jbd_debug(2, "insert revoke for block %u, bh_in=%p\n", blocknr, bh_in);
405 err = insert_revoke_hash(journal, blocknr,
406 handle->h_transaction->t_tid);
407 BUFFER_TRACE(bh_in, "exit");
412 * Cancel an outstanding revoke. For use only internally by the
413 * journaling code (called from journal_get_write_access).
415 * We trust buffer_revoked() on the buffer if the buffer is already
416 * being journaled: if there is no revoke pending on the buffer, then we
417 * don't do anything here.
419 * This would break if it were possible for a buffer to be revoked and
420 * discarded, and then reallocated within the same transaction. In such
421 * a case we would have lost the revoked bit, but when we arrived here
422 * the second time we would still have a pending revoke to cancel. So,
423 * do not trust the Revoked bit on buffers unless RevokeValid is also
426 int journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
428 struct jbd_revoke_record_s *record;
429 journal_t *journal = handle->h_transaction->t_journal;
431 int did_revoke = 0; /* akpm: debug */
432 struct buffer_head *bh = jh2bh(jh);
434 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
436 /* Is the existing Revoke bit valid? If so, we trust it, and
437 * only perform the full cancel if the revoke bit is set. If
438 * not, we can't trust the revoke bit, and we need to do the
439 * full search for a revoke record. */
440 if (test_set_buffer_revokevalid(bh)) {
441 need_cancel = test_clear_buffer_revoked(bh);
444 clear_buffer_revoked(bh);
448 record = find_revoke_record(journal, bh->b_blocknr);
450 jbd_debug(4, "cancelled existing revoke on "
451 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
452 spin_lock(&journal->j_revoke_lock);
453 list_del(&record->hash);
454 spin_unlock(&journal->j_revoke_lock);
455 kmem_cache_free(revoke_record_cache, record);
460 #ifdef JBD_EXPENSIVE_CHECKING
461 /* There better not be one left behind by now! */
462 record = find_revoke_record(journal, bh->b_blocknr);
463 J_ASSERT_JH(jh, record == NULL);
466 /* Finally, have we just cleared revoke on an unhashed
467 * buffer_head? If so, we'd better make sure we clear the
468 * revoked status on any hashed alias too, otherwise the revoke
469 * state machine will get very upset later on. */
471 struct buffer_head *bh2;
472 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
475 clear_buffer_revoked(bh2);
482 /* journal_switch_revoke table select j_revoke for next transaction
483 * we do not want to suspend any processing until all revokes are
486 void journal_switch_revoke_table(journal_t *journal)
490 if (journal->j_revoke == journal->j_revoke_table[0])
491 journal->j_revoke = journal->j_revoke_table[1];
493 journal->j_revoke = journal->j_revoke_table[0];
495 for (i = 0; i < journal->j_revoke->hash_size; i++)
496 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
500 * Write revoke records to the journal for all entries in the current
501 * revoke hash, deleting the entries as we go.
503 void journal_write_revoke_records(journal_t *journal,
504 transaction_t *transaction, int write_op)
506 struct journal_head *descriptor;
507 struct jbd_revoke_record_s *record;
508 struct jbd_revoke_table_s *revoke;
509 struct list_head *hash_list;
510 int i, offset, count;
516 /* select revoke table for committing transaction */
517 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
518 journal->j_revoke_table[1] : journal->j_revoke_table[0];
520 for (i = 0; i < revoke->hash_size; i++) {
521 hash_list = &revoke->hash_table[i];
523 while (!list_empty(hash_list)) {
524 record = (struct jbd_revoke_record_s *)
526 write_one_revoke_record(journal, transaction,
527 &descriptor, &offset,
530 list_del(&record->hash);
531 kmem_cache_free(revoke_record_cache, record);
535 flush_descriptor(journal, descriptor, offset, write_op);
536 jbd_debug(1, "Wrote %d revoke records\n", count);
540 * Write out one revoke record. We need to create a new descriptor
541 * block if the old one is full or if we have not already created one.
544 static void write_one_revoke_record(journal_t *journal,
545 transaction_t *transaction,
546 struct journal_head **descriptorp,
548 struct jbd_revoke_record_s *record,
551 struct journal_head *descriptor;
553 journal_header_t *header;
555 /* If we are already aborting, this all becomes a noop. We
556 still need to go round the loop in
557 journal_write_revoke_records in order to free all of the
558 revoke records: only the IO to the journal is omitted. */
559 if (is_journal_aborted(journal))
562 descriptor = *descriptorp;
565 /* Make sure we have a descriptor with space left for the record */
567 if (offset == journal->j_blocksize) {
568 flush_descriptor(journal, descriptor, offset, write_op);
574 descriptor = journal_get_descriptor_buffer(journal);
577 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
578 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
579 header->h_blocktype = cpu_to_be32(JFS_REVOKE_BLOCK);
580 header->h_sequence = cpu_to_be32(transaction->t_tid);
582 /* Record it so that we can wait for IO completion later */
583 JBUFFER_TRACE(descriptor, "file as BJ_LogCtl");
584 journal_file_buffer(descriptor, transaction, BJ_LogCtl);
586 offset = sizeof(journal_revoke_header_t);
587 *descriptorp = descriptor;
590 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
591 cpu_to_be32(record->blocknr);
597 * Flush a revoke descriptor out to the journal. If we are aborting,
598 * this is a noop; otherwise we are generating a buffer which needs to
599 * be waited for during commit, so it has to go onto the appropriate
600 * journal buffer list.
603 static void flush_descriptor(journal_t *journal,
604 struct journal_head *descriptor,
605 int offset, int write_op)
607 journal_revoke_header_t *header;
608 struct buffer_head *bh = jh2bh(descriptor);
610 if (is_journal_aborted(journal)) {
615 header = (journal_revoke_header_t *) jh2bh(descriptor)->b_data;
616 header->r_count = cpu_to_be32(offset);
617 set_buffer_jwrite(bh);
618 BUFFER_TRACE(bh, "write");
619 set_buffer_dirty(bh);
620 write_dirty_buffer(bh, write_op);
625 * Revoke support for recovery.
627 * Recovery needs to be able to:
629 * record all revoke records, including the tid of the latest instance
630 * of each revoke in the journal
632 * check whether a given block in a given transaction should be replayed
633 * (ie. has not been revoked by a revoke record in that or a subsequent
636 * empty the revoke table after recovery.
640 * First, setting revoke records. We create a new revoke record for
641 * every block ever revoked in the log as we scan it for recovery, and
642 * we update the existing records if we find multiple revokes for a
646 int journal_set_revoke(journal_t *journal,
647 unsigned int blocknr,
650 struct jbd_revoke_record_s *record;
652 record = find_revoke_record(journal, blocknr);
654 /* If we have multiple occurrences, only record the
655 * latest sequence number in the hashed record */
656 if (tid_gt(sequence, record->sequence))
657 record->sequence = sequence;
660 return insert_revoke_hash(journal, blocknr, sequence);
664 * Test revoke records. For a given block referenced in the log, has
665 * that block been revoked? A revoke record with a given transaction
666 * sequence number revokes all blocks in that transaction and earlier
667 * ones, but later transactions still need replayed.
670 int journal_test_revoke(journal_t *journal,
671 unsigned int blocknr,
674 struct jbd_revoke_record_s *record;
676 record = find_revoke_record(journal, blocknr);
679 if (tid_gt(sequence, record->sequence))
685 * Finally, once recovery is over, we need to clear the revoke table so
686 * that it can be reused by the running filesystem.
689 void journal_clear_revoke(journal_t *journal)
692 struct list_head *hash_list;
693 struct jbd_revoke_record_s *record;
694 struct jbd_revoke_table_s *revoke;
696 revoke = journal->j_revoke;
698 for (i = 0; i < revoke->hash_size; i++) {
699 hash_list = &revoke->hash_table[i];
700 while (!list_empty(hash_list)) {
701 record = (struct jbd_revoke_record_s*) hash_list->next;
702 list_del(&record->hash);
703 kmem_cache_free(revoke_record_cache, record);