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54 * buffer has not been revoked, and cancel_revoke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57 * buffer has been revoked.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/smp_lock.h>
70 #include <linux/init.h>
73 static kmem_cache_t *revoke_record_cache;
74 static kmem_cache_t *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 int insert_revoke_hash(journal_t *journal, unsigned long blocknr, tid_t seq)
121 struct list_head *hash_list;
122 struct jbd_revoke_record_s *record;
125 record = kmem_cache_alloc(revoke_record_cache, GFP_NOFS);
129 record->sequence = seq;
130 record->blocknr = blocknr;
131 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
132 spin_lock(&journal->j_revoke_lock);
133 list_add(&record->hash, hash_list);
134 spin_unlock(&journal->j_revoke_lock);
138 if (!journal_oom_retry)
140 jbd_debug(1, "ENOMEM in %s, retrying\n", __FUNCTION__);
145 /* Find a revoke record in the journal's hash table. */
147 static struct jbd_revoke_record_s *find_revoke_record(journal_t *journal,
148 unsigned long blocknr)
150 struct list_head *hash_list;
151 struct jbd_revoke_record_s *record;
153 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
155 spin_lock(&journal->j_revoke_lock);
156 record = (struct jbd_revoke_record_s *) hash_list->next;
157 while (&(record->hash) != hash_list) {
158 if (record->blocknr == blocknr) {
159 spin_unlock(&journal->j_revoke_lock);
162 record = (struct jbd_revoke_record_s *) record->hash.next;
164 spin_unlock(&journal->j_revoke_lock);
168 int __init journal_init_revoke_caches(void)
170 revoke_record_cache = kmem_cache_create("revoke_record",
171 sizeof(struct jbd_revoke_record_s),
172 0, SLAB_HWCACHE_ALIGN, NULL, NULL);
173 if (revoke_record_cache == 0)
176 revoke_table_cache = kmem_cache_create("revoke_table",
177 sizeof(struct jbd_revoke_table_s),
179 if (revoke_table_cache == 0) {
180 kmem_cache_destroy(revoke_record_cache);
181 revoke_record_cache = NULL;
187 void journal_destroy_revoke_caches(void)
189 kmem_cache_destroy(revoke_record_cache);
190 revoke_record_cache = NULL;
191 kmem_cache_destroy(revoke_table_cache);
192 revoke_table_cache = NULL;
195 /* Initialise the revoke table for a given journal to a given size. */
197 int journal_init_revoke(journal_t *journal, int hash_size)
201 J_ASSERT (journal->j_revoke_table[0] == NULL);
205 while((tmp >>= 1UL) != 0UL)
208 journal->j_revoke_table[0] = kmem_cache_alloc(revoke_table_cache, GFP_KERNEL);
209 if (!journal->j_revoke_table[0])
211 journal->j_revoke = journal->j_revoke_table[0];
213 /* Check that the hash_size is a power of two */
214 J_ASSERT ((hash_size & (hash_size-1)) == 0);
216 journal->j_revoke->hash_size = hash_size;
218 journal->j_revoke->hash_shift = shift;
220 journal->j_revoke->hash_table =
221 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
222 if (!journal->j_revoke->hash_table) {
223 kmem_cache_free(revoke_table_cache, journal->j_revoke_table[0]);
224 journal->j_revoke = NULL;
228 for (tmp = 0; tmp < hash_size; tmp++)
229 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
231 journal->j_revoke_table[1] = kmem_cache_alloc(revoke_table_cache, GFP_KERNEL);
232 if (!journal->j_revoke_table[1]) {
233 kfree(journal->j_revoke_table[0]->hash_table);
234 kmem_cache_free(revoke_table_cache, journal->j_revoke_table[0]);
238 journal->j_revoke = journal->j_revoke_table[1];
240 /* Check that the hash_size is a power of two */
241 J_ASSERT ((hash_size & (hash_size-1)) == 0);
243 journal->j_revoke->hash_size = hash_size;
245 journal->j_revoke->hash_shift = shift;
247 journal->j_revoke->hash_table =
248 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
249 if (!journal->j_revoke->hash_table) {
250 kfree(journal->j_revoke_table[0]->hash_table);
251 kmem_cache_free(revoke_table_cache, journal->j_revoke_table[0]);
252 kmem_cache_free(revoke_table_cache, journal->j_revoke_table[1]);
253 journal->j_revoke = NULL;
257 for (tmp = 0; tmp < hash_size; tmp++)
258 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
260 spin_lock_init(&journal->j_revoke_lock);
265 /* Destoy a journal's revoke table. The table must already be empty! */
267 void journal_destroy_revoke(journal_t *journal)
269 struct jbd_revoke_table_s *table;
270 struct list_head *hash_list;
273 table = journal->j_revoke_table[0];
277 for (i=0; i<table->hash_size; i++) {
278 hash_list = &table->hash_table[i];
279 J_ASSERT (list_empty(hash_list));
282 kfree(table->hash_table);
283 kmem_cache_free(revoke_table_cache, table);
284 journal->j_revoke = NULL;
286 table = journal->j_revoke_table[1];
290 for (i=0; i<table->hash_size; i++) {
291 hash_list = &table->hash_table[i];
292 J_ASSERT (list_empty(hash_list));
295 kfree(table->hash_table);
296 kmem_cache_free(revoke_table_cache, table);
297 journal->j_revoke = NULL;
304 * journal_revoke: revoke a given buffer_head from the journal. This
305 * prevents the block from being replayed during recovery if we take a
306 * crash after this current transaction commits. Any subsequent
307 * metadata writes of the buffer in this transaction cancel the
310 * Note that this call may block --- it is up to the caller to make
311 * sure that there are no further calls to journal_write_metadata
312 * before the revoke is complete. In ext3, this implies calling the
313 * revoke before clearing the block bitmap when we are deleting
316 * Revoke performs a journal_forget on any buffer_head passed in as a
317 * parameter, but does _not_ forget the buffer_head if the bh was only
320 * bh_in may not be a journalled buffer - it may have come off
321 * the hash tables without an attached journal_head.
323 * If bh_in is non-zero, journal_revoke() will decrement its b_count
327 int journal_revoke(handle_t *handle, unsigned long blocknr,
328 struct buffer_head *bh_in)
330 struct buffer_head *bh = NULL;
332 struct block_device *bdev;
337 BUFFER_TRACE(bh_in, "enter");
339 journal = handle->h_transaction->t_journal;
340 if (!journal_set_features(journal, 0, 0, JFS_FEATURE_INCOMPAT_REVOKE)){
341 J_ASSERT (!"Cannot set revoke feature!");
345 bdev = journal->j_fs_dev;
349 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
351 BUFFER_TRACE(bh, "found on hash");
353 #ifdef JBD_EXPENSIVE_CHECKING
355 struct buffer_head *bh2;
357 /* If there is a different buffer_head lying around in
358 * memory anywhere... */
359 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
361 /* ... and it has RevokeValid status... */
362 if (bh2 != bh && buffer_revokevalid(bh2))
363 /* ...then it better be revoked too,
364 * since it's illegal to create a revoke
365 * record against a buffer_head which is
366 * not marked revoked --- that would
367 * risk missing a subsequent revoke
369 J_ASSERT_BH(bh2, buffer_revoked(bh2));
375 /* We really ought not ever to revoke twice in a row without
376 first having the revoke cancelled: it's illegal to free a
377 block twice without allocating it in between! */
379 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
380 "inconsistent data on disk")) {
385 set_buffer_revoked(bh);
386 set_buffer_revokevalid(bh);
388 BUFFER_TRACE(bh_in, "call journal_forget");
389 journal_forget(handle, bh_in);
391 BUFFER_TRACE(bh, "call brelse");
396 jbd_debug(2, "insert revoke for block %lu, bh_in=%p\n", blocknr, bh_in);
397 err = insert_revoke_hash(journal, blocknr,
398 handle->h_transaction->t_tid);
399 BUFFER_TRACE(bh_in, "exit");
404 * Cancel an outstanding revoke. For use only internally by the
405 * journaling code (called from journal_get_write_access).
407 * We trust buffer_revoked() on the buffer if the buffer is already
408 * being journaled: if there is no revoke pending on the buffer, then we
409 * don't do anything here.
411 * This would break if it were possible for a buffer to be revoked and
412 * discarded, and then reallocated within the same transaction. In such
413 * a case we would have lost the revoked bit, but when we arrived here
414 * the second time we would still have a pending revoke to cancel. So,
415 * do not trust the Revoked bit on buffers unless RevokeValid is also
418 * The caller must have the journal locked.
420 int journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
422 struct jbd_revoke_record_s *record;
423 journal_t *journal = handle->h_transaction->t_journal;
425 int did_revoke = 0; /* akpm: debug */
426 struct buffer_head *bh = jh2bh(jh);
428 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
430 /* Is the existing Revoke bit valid? If so, we trust it, and
431 * only perform the full cancel if the revoke bit is set. If
432 * not, we can't trust the revoke bit, and we need to do the
433 * full search for a revoke record. */
434 if (test_set_buffer_revokevalid(bh)) {
435 need_cancel = test_clear_buffer_revoked(bh);
438 clear_buffer_revoked(bh);
442 record = find_revoke_record(journal, bh->b_blocknr);
444 jbd_debug(4, "cancelled existing revoke on "
445 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
446 spin_lock(&journal->j_revoke_lock);
447 list_del(&record->hash);
448 spin_unlock(&journal->j_revoke_lock);
449 kmem_cache_free(revoke_record_cache, record);
454 #ifdef JBD_EXPENSIVE_CHECKING
455 /* There better not be one left behind by now! */
456 record = find_revoke_record(journal, bh->b_blocknr);
457 J_ASSERT_JH(jh, record == NULL);
460 /* Finally, have we just cleared revoke on an unhashed
461 * buffer_head? If so, we'd better make sure we clear the
462 * revoked status on any hashed alias too, otherwise the revoke
463 * state machine will get very upset later on. */
465 struct buffer_head *bh2;
466 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
469 clear_buffer_revoked(bh2);
476 /* journal_switch_revoke table select j_revoke for next transaction
477 * we do not want to suspend any processing until all revokes are
480 void journal_switch_revoke_table(journal_t *journal)
484 if (journal->j_revoke == journal->j_revoke_table[0])
485 journal->j_revoke = journal->j_revoke_table[1];
487 journal->j_revoke = journal->j_revoke_table[0];
489 for (i = 0; i < journal->j_revoke->hash_size; i++)
490 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
494 * Write revoke records to the journal for all entries in the current
495 * revoke hash, deleting the entries as we go.
497 * Called with the journal lock held.
500 void journal_write_revoke_records(journal_t *journal,
501 transaction_t *transaction)
503 struct journal_head *descriptor;
504 struct jbd_revoke_record_s *record;
505 struct jbd_revoke_table_s *revoke;
506 struct list_head *hash_list;
507 int i, offset, count;
513 /* select revoke table for committing transaction */
514 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
515 journal->j_revoke_table[1] : journal->j_revoke_table[0];
517 for (i = 0; i < revoke->hash_size; i++) {
518 hash_list = &revoke->hash_table[i];
520 while (!list_empty(hash_list)) {
521 record = (struct jbd_revoke_record_s *)
523 write_one_revoke_record(journal, transaction,
524 &descriptor, &offset,
527 list_del(&record->hash);
528 kmem_cache_free(revoke_record_cache, record);
532 flush_descriptor(journal, descriptor, offset);
533 jbd_debug(1, "Wrote %d revoke records\n", count);
537 * Write out one revoke record. We need to create a new descriptor
538 * block if the old one is full or if we have not already created one.
541 static void write_one_revoke_record(journal_t *journal,
542 transaction_t *transaction,
543 struct journal_head **descriptorp,
545 struct jbd_revoke_record_s *record)
547 struct journal_head *descriptor;
549 journal_header_t *header;
551 /* If we are already aborting, this all becomes a noop. We
552 still need to go round the loop in
553 journal_write_revoke_records in order to free all of the
554 revoke records: only the IO to the journal is omitted. */
555 if (is_journal_aborted(journal))
558 descriptor = *descriptorp;
561 /* Make sure we have a descriptor with space left for the record */
563 if (offset == journal->j_blocksize) {
564 flush_descriptor(journal, descriptor, offset);
570 descriptor = journal_get_descriptor_buffer(journal);
573 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
574 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
575 header->h_blocktype = cpu_to_be32(JFS_REVOKE_BLOCK);
576 header->h_sequence = cpu_to_be32(transaction->t_tid);
578 /* Record it so that we can wait for IO completion later */
579 JBUFFER_TRACE(descriptor, "file as BJ_LogCtl");
580 journal_file_buffer(descriptor, transaction, BJ_LogCtl);
582 offset = sizeof(journal_revoke_header_t);
583 *descriptorp = descriptor;
586 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
587 cpu_to_be32(record->blocknr);
593 * Flush a revoke descriptor out to the journal. If we are aborting,
594 * this is a noop; otherwise we are generating a buffer which needs to
595 * be waited for during commit, so it has to go onto the appropriate
596 * journal buffer list.
599 static void flush_descriptor(journal_t *journal,
600 struct journal_head *descriptor,
603 journal_revoke_header_t *header;
604 struct buffer_head *bh = jh2bh(descriptor);
606 if (is_journal_aborted(journal)) {
611 header = (journal_revoke_header_t *) jh2bh(descriptor)->b_data;
612 header->r_count = cpu_to_be32(offset);
613 set_buffer_jwrite(bh);
614 BUFFER_TRACE(bh, "write");
615 set_buffer_dirty(bh);
616 ll_rw_block(WRITE, 1, &bh);
621 * Revoke support for recovery.
623 * Recovery needs to be able to:
625 * record all revoke records, including the tid of the latest instance
626 * of each revoke in the journal
628 * check whether a given block in a given transaction should be replayed
629 * (ie. has not been revoked by a revoke record in that or a subsequent
632 * empty the revoke table after recovery.
636 * First, setting revoke records. We create a new revoke record for
637 * every block ever revoked in the log as we scan it for recovery, and
638 * we update the existing records if we find multiple revokes for a
642 int journal_set_revoke(journal_t *journal,
643 unsigned long blocknr,
646 struct jbd_revoke_record_s *record;
648 record = find_revoke_record(journal, blocknr);
650 /* If we have multiple occurrences, only record the
651 * latest sequence number in the hashed record */
652 if (tid_gt(sequence, record->sequence))
653 record->sequence = sequence;
656 return insert_revoke_hash(journal, blocknr, sequence);
660 * Test revoke records. For a given block referenced in the log, has
661 * that block been revoked? A revoke record with a given transaction
662 * sequence number revokes all blocks in that transaction and earlier
663 * ones, but later transactions still need replayed.
666 int journal_test_revoke(journal_t *journal,
667 unsigned long blocknr,
670 struct jbd_revoke_record_s *record;
672 record = find_revoke_record(journal, blocknr);
675 if (tid_gt(sequence, record->sequence))
681 * Finally, once recovery is over, we need to clear the revoke table so
682 * that it can be reused by the running filesystem.
685 void journal_clear_revoke(journal_t *journal)
688 struct list_head *hash_list;
689 struct jbd_revoke_record_s *record;
690 struct jbd_revoke_table_s *revoke;
692 revoke = journal->j_revoke;
694 for (i = 0; i < revoke->hash_size; i++) {
695 hash_list = &revoke->hash_table[i];
696 while (!list_empty(hash_list)) {
697 record = (struct jbd_revoke_record_s*) hash_list->next;
698 list_del(&record->hash);
699 kmem_cache_free(revoke_record_cache, record);