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54 * buffer has not been revoked, and cancel_revoke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57 * buffer has been revoked.
63 #include <linux/time.h>
65 #include <linux/jbd.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 *revoke_record_cache;
74 static struct kmem_cache *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 jbd_revoke_record_s
82 struct list_head hash;
83 tid_t sequence; /* Used for recovery only */
84 unsigned long blocknr;
88 /* The revoke table is just a simple hash table of revoke records. */
89 struct jbd_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 jbd_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 block)
111 struct jbd_revoke_table_s *table = journal->j_revoke;
112 int hash_shift = table->hash_shift;
114 return ((block << (hash_shift - 6)) ^
116 (block << (hash_shift - 12))) & (table->hash_size - 1);
119 static int insert_revoke_hash(journal_t *journal, unsigned long blocknr,
122 struct list_head *hash_list;
123 struct jbd_revoke_record_s *record;
126 record = kmem_cache_alloc(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", __func__);
146 /* Find a revoke record in the journal's hash table. */
148 static struct jbd_revoke_record_s *find_revoke_record(journal_t *journal,
149 unsigned long blocknr)
151 struct list_head *hash_list;
152 struct jbd_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 jbd_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 jbd_revoke_record_s *) record->hash.next;
165 spin_unlock(&journal->j_revoke_lock);
169 int __init journal_init_revoke_caches(void)
171 revoke_record_cache = kmem_cache_create("revoke_record",
172 sizeof(struct jbd_revoke_record_s),
174 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
176 if (!revoke_record_cache)
179 revoke_table_cache = kmem_cache_create("revoke_table",
180 sizeof(struct jbd_revoke_table_s),
181 0, SLAB_TEMPORARY, NULL);
182 if (!revoke_table_cache) {
183 kmem_cache_destroy(revoke_record_cache);
184 revoke_record_cache = NULL;
190 void journal_destroy_revoke_caches(void)
192 kmem_cache_destroy(revoke_record_cache);
193 revoke_record_cache = NULL;
194 kmem_cache_destroy(revoke_table_cache);
195 revoke_table_cache = NULL;
198 static struct jbd_revoke_table_s *journal_init_revoke_table(int hash_size)
202 struct jbd_revoke_table_s *table;
204 table = kmem_cache_alloc(revoke_table_cache, GFP_KERNEL);
208 while((tmp >>= 1UL) != 0UL)
211 table->hash_size = hash_size;
212 table->hash_shift = shift;
214 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
215 if (!table->hash_table) {
216 kmem_cache_free(revoke_table_cache, table);
221 for (tmp = 0; tmp < hash_size; tmp++)
222 INIT_LIST_HEAD(&table->hash_table[tmp]);
228 static void journal_destroy_revoke_table(struct jbd_revoke_table_s *table)
231 struct list_head *hash_list;
233 for (i = 0; i < table->hash_size; i++) {
234 hash_list = &table->hash_table[i];
235 J_ASSERT(list_empty(hash_list));
238 kfree(table->hash_table);
239 kmem_cache_free(revoke_table_cache, table);
242 /* Initialise the revoke table for a given journal to a given size. */
243 int journal_init_revoke(journal_t *journal, int hash_size)
245 J_ASSERT(journal->j_revoke_table[0] == NULL);
246 J_ASSERT(is_power_of_2(hash_size));
248 journal->j_revoke_table[0] = journal_init_revoke_table(hash_size);
249 if (!journal->j_revoke_table[0])
252 journal->j_revoke_table[1] = journal_init_revoke_table(hash_size);
253 if (!journal->j_revoke_table[1])
256 journal->j_revoke = journal->j_revoke_table[1];
258 spin_lock_init(&journal->j_revoke_lock);
263 journal_destroy_revoke_table(journal->j_revoke_table[0]);
268 /* Destroy a journal's revoke table. The table must already be empty! */
269 void journal_destroy_revoke(journal_t *journal)
271 journal->j_revoke = NULL;
272 if (journal->j_revoke_table[0])
273 journal_destroy_revoke_table(journal->j_revoke_table[0]);
274 if (journal->j_revoke_table[1])
275 journal_destroy_revoke_table(journal->j_revoke_table[1]);
282 * journal_revoke: revoke a given buffer_head from the journal. This
283 * prevents the block from being replayed during recovery if we take a
284 * crash after this current transaction commits. Any subsequent
285 * metadata writes of the buffer in this transaction cancel the
288 * Note that this call may block --- it is up to the caller to make
289 * sure that there are no further calls to journal_write_metadata
290 * before the revoke is complete. In ext3, this implies calling the
291 * revoke before clearing the block bitmap when we are deleting
294 * Revoke performs a journal_forget on any buffer_head passed in as a
295 * parameter, but does _not_ forget the buffer_head if the bh was only
298 * bh_in may not be a journalled buffer - it may have come off
299 * the hash tables without an attached journal_head.
301 * If bh_in is non-zero, journal_revoke() will decrement its b_count
305 int journal_revoke(handle_t *handle, unsigned long blocknr,
306 struct buffer_head *bh_in)
308 struct buffer_head *bh = NULL;
310 struct block_device *bdev;
315 BUFFER_TRACE(bh_in, "enter");
317 journal = handle->h_transaction->t_journal;
318 if (!journal_set_features(journal, 0, 0, JFS_FEATURE_INCOMPAT_REVOKE)){
319 J_ASSERT (!"Cannot set revoke feature!");
323 bdev = journal->j_fs_dev;
327 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
329 BUFFER_TRACE(bh, "found on hash");
331 #ifdef JBD_EXPENSIVE_CHECKING
333 struct buffer_head *bh2;
335 /* If there is a different buffer_head lying around in
336 * memory anywhere... */
337 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
339 /* ... and it has RevokeValid status... */
340 if (bh2 != bh && buffer_revokevalid(bh2))
341 /* ...then it better be revoked too,
342 * since it's illegal to create a revoke
343 * record against a buffer_head which is
344 * not marked revoked --- that would
345 * risk missing a subsequent revoke
347 J_ASSERT_BH(bh2, buffer_revoked(bh2));
353 /* We really ought not ever to revoke twice in a row without
354 first having the revoke cancelled: it's illegal to free a
355 block twice without allocating it in between! */
357 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
358 "inconsistent data on disk")) {
363 set_buffer_revoked(bh);
364 set_buffer_revokevalid(bh);
366 BUFFER_TRACE(bh_in, "call journal_forget");
367 journal_forget(handle, bh_in);
369 BUFFER_TRACE(bh, "call brelse");
374 jbd_debug(2, "insert revoke for block %lu, bh_in=%p\n", blocknr, bh_in);
375 err = insert_revoke_hash(journal, blocknr,
376 handle->h_transaction->t_tid);
377 BUFFER_TRACE(bh_in, "exit");
382 * Cancel an outstanding revoke. For use only internally by the
383 * journaling code (called from journal_get_write_access).
385 * We trust buffer_revoked() on the buffer if the buffer is already
386 * being journaled: if there is no revoke pending on the buffer, then we
387 * don't do anything here.
389 * This would break if it were possible for a buffer to be revoked and
390 * discarded, and then reallocated within the same transaction. In such
391 * a case we would have lost the revoked bit, but when we arrived here
392 * the second time we would still have a pending revoke to cancel. So,
393 * do not trust the Revoked bit on buffers unless RevokeValid is also
396 * The caller must have the journal locked.
398 int journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
400 struct jbd_revoke_record_s *record;
401 journal_t *journal = handle->h_transaction->t_journal;
403 int did_revoke = 0; /* akpm: debug */
404 struct buffer_head *bh = jh2bh(jh);
406 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
408 /* Is the existing Revoke bit valid? If so, we trust it, and
409 * only perform the full cancel if the revoke bit is set. If
410 * not, we can't trust the revoke bit, and we need to do the
411 * full search for a revoke record. */
412 if (test_set_buffer_revokevalid(bh)) {
413 need_cancel = test_clear_buffer_revoked(bh);
416 clear_buffer_revoked(bh);
420 record = find_revoke_record(journal, bh->b_blocknr);
422 jbd_debug(4, "cancelled existing revoke on "
423 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
424 spin_lock(&journal->j_revoke_lock);
425 list_del(&record->hash);
426 spin_unlock(&journal->j_revoke_lock);
427 kmem_cache_free(revoke_record_cache, record);
432 #ifdef JBD_EXPENSIVE_CHECKING
433 /* There better not be one left behind by now! */
434 record = find_revoke_record(journal, bh->b_blocknr);
435 J_ASSERT_JH(jh, record == NULL);
438 /* Finally, have we just cleared revoke on an unhashed
439 * buffer_head? If so, we'd better make sure we clear the
440 * revoked status on any hashed alias too, otherwise the revoke
441 * state machine will get very upset later on. */
443 struct buffer_head *bh2;
444 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
447 clear_buffer_revoked(bh2);
454 /* journal_switch_revoke table select j_revoke for next transaction
455 * we do not want to suspend any processing until all revokes are
458 void journal_switch_revoke_table(journal_t *journal)
462 if (journal->j_revoke == journal->j_revoke_table[0])
463 journal->j_revoke = journal->j_revoke_table[1];
465 journal->j_revoke = journal->j_revoke_table[0];
467 for (i = 0; i < journal->j_revoke->hash_size; i++)
468 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
472 * Write revoke records to the journal for all entries in the current
473 * revoke hash, deleting the entries as we go.
475 * Called with the journal lock held.
478 void journal_write_revoke_records(journal_t *journal,
479 transaction_t *transaction)
481 struct journal_head *descriptor;
482 struct jbd_revoke_record_s *record;
483 struct jbd_revoke_table_s *revoke;
484 struct list_head *hash_list;
485 int i, offset, count;
491 /* select revoke table for committing transaction */
492 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
493 journal->j_revoke_table[1] : journal->j_revoke_table[0];
495 for (i = 0; i < revoke->hash_size; i++) {
496 hash_list = &revoke->hash_table[i];
498 while (!list_empty(hash_list)) {
499 record = (struct jbd_revoke_record_s *)
501 write_one_revoke_record(journal, transaction,
502 &descriptor, &offset,
505 list_del(&record->hash);
506 kmem_cache_free(revoke_record_cache, record);
510 flush_descriptor(journal, descriptor, offset);
511 jbd_debug(1, "Wrote %d revoke records\n", count);
515 * Write out one revoke record. We need to create a new descriptor
516 * block if the old one is full or if we have not already created one.
519 static void write_one_revoke_record(journal_t *journal,
520 transaction_t *transaction,
521 struct journal_head **descriptorp,
523 struct jbd_revoke_record_s *record)
525 struct journal_head *descriptor;
527 journal_header_t *header;
529 /* If we are already aborting, this all becomes a noop. We
530 still need to go round the loop in
531 journal_write_revoke_records in order to free all of the
532 revoke records: only the IO to the journal is omitted. */
533 if (is_journal_aborted(journal))
536 descriptor = *descriptorp;
539 /* Make sure we have a descriptor with space left for the record */
541 if (offset == journal->j_blocksize) {
542 flush_descriptor(journal, descriptor, offset);
548 descriptor = journal_get_descriptor_buffer(journal);
551 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
552 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
553 header->h_blocktype = cpu_to_be32(JFS_REVOKE_BLOCK);
554 header->h_sequence = cpu_to_be32(transaction->t_tid);
556 /* Record it so that we can wait for IO completion later */
557 JBUFFER_TRACE(descriptor, "file as BJ_LogCtl");
558 journal_file_buffer(descriptor, transaction, BJ_LogCtl);
560 offset = sizeof(journal_revoke_header_t);
561 *descriptorp = descriptor;
564 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
565 cpu_to_be32(record->blocknr);
571 * Flush a revoke descriptor out to the journal. If we are aborting,
572 * this is a noop; otherwise we are generating a buffer which needs to
573 * be waited for during commit, so it has to go onto the appropriate
574 * journal buffer list.
577 static void flush_descriptor(journal_t *journal,
578 struct journal_head *descriptor,
581 journal_revoke_header_t *header;
582 struct buffer_head *bh = jh2bh(descriptor);
584 if (is_journal_aborted(journal)) {
589 header = (journal_revoke_header_t *) jh2bh(descriptor)->b_data;
590 header->r_count = cpu_to_be32(offset);
591 set_buffer_jwrite(bh);
592 BUFFER_TRACE(bh, "write");
593 set_buffer_dirty(bh);
594 ll_rw_block(SWRITE, 1, &bh);
599 * Revoke support for recovery.
601 * Recovery needs to be able to:
603 * record all revoke records, including the tid of the latest instance
604 * of each revoke in the journal
606 * check whether a given block in a given transaction should be replayed
607 * (ie. has not been revoked by a revoke record in that or a subsequent
610 * empty the revoke table after recovery.
614 * First, setting revoke records. We create a new revoke record for
615 * every block ever revoked in the log as we scan it for recovery, and
616 * we update the existing records if we find multiple revokes for a
620 int journal_set_revoke(journal_t *journal,
621 unsigned long blocknr,
624 struct jbd_revoke_record_s *record;
626 record = find_revoke_record(journal, blocknr);
628 /* If we have multiple occurrences, only record the
629 * latest sequence number in the hashed record */
630 if (tid_gt(sequence, record->sequence))
631 record->sequence = sequence;
634 return insert_revoke_hash(journal, blocknr, sequence);
638 * Test revoke records. For a given block referenced in the log, has
639 * that block been revoked? A revoke record with a given transaction
640 * sequence number revokes all blocks in that transaction and earlier
641 * ones, but later transactions still need replayed.
644 int journal_test_revoke(journal_t *journal,
645 unsigned long blocknr,
648 struct jbd_revoke_record_s *record;
650 record = find_revoke_record(journal, blocknr);
653 if (tid_gt(sequence, record->sequence))
659 * Finally, once recovery is over, we need to clear the revoke table so
660 * that it can be reused by the running filesystem.
663 void journal_clear_revoke(journal_t *journal)
666 struct list_head *hash_list;
667 struct jbd_revoke_record_s *record;
668 struct jbd_revoke_table_s *revoke;
670 revoke = journal->j_revoke;
672 for (i = 0; i < revoke->hash_size; i++) {
673 hash_list = &revoke->hash_table[i];
674 while (!list_empty(hash_list)) {
675 record = (struct jbd_revoke_record_s*) hash_list->next;
676 list_del(&record->hash);
677 kmem_cache_free(revoke_record_cache, record);