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54 * buffer has not been revoked, and cancel_revoke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57 * buffer has been revoked.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", __FUNCTION__);
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 == 0)
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 == 0) {
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 /* Initialise the revoke table for a given journal to a given size. */
200 int journal_init_revoke(journal_t *journal, int hash_size)
204 J_ASSERT (journal->j_revoke_table[0] == NULL);
208 while((tmp >>= 1UL) != 0UL)
211 journal->j_revoke_table[0] = kmem_cache_alloc(revoke_table_cache, GFP_KERNEL);
212 if (!journal->j_revoke_table[0])
214 journal->j_revoke = journal->j_revoke_table[0];
216 /* Check that the hash_size is a power of two */
217 J_ASSERT(is_power_of_2(hash_size));
219 journal->j_revoke->hash_size = hash_size;
221 journal->j_revoke->hash_shift = shift;
223 journal->j_revoke->hash_table =
224 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
225 if (!journal->j_revoke->hash_table) {
226 kmem_cache_free(revoke_table_cache, journal->j_revoke_table[0]);
227 journal->j_revoke = NULL;
231 for (tmp = 0; tmp < hash_size; tmp++)
232 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
234 journal->j_revoke_table[1] = kmem_cache_alloc(revoke_table_cache, GFP_KERNEL);
235 if (!journal->j_revoke_table[1]) {
236 kfree(journal->j_revoke_table[0]->hash_table);
237 kmem_cache_free(revoke_table_cache, journal->j_revoke_table[0]);
241 journal->j_revoke = journal->j_revoke_table[1];
243 /* Check that the hash_size is a power of two */
244 J_ASSERT(is_power_of_2(hash_size));
246 journal->j_revoke->hash_size = hash_size;
248 journal->j_revoke->hash_shift = shift;
250 journal->j_revoke->hash_table =
251 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
252 if (!journal->j_revoke->hash_table) {
253 kfree(journal->j_revoke_table[0]->hash_table);
254 kmem_cache_free(revoke_table_cache, journal->j_revoke_table[0]);
255 kmem_cache_free(revoke_table_cache, journal->j_revoke_table[1]);
256 journal->j_revoke = NULL;
260 for (tmp = 0; tmp < hash_size; tmp++)
261 INIT_LIST_HEAD(&journal->j_revoke->hash_table[tmp]);
263 spin_lock_init(&journal->j_revoke_lock);
268 /* Destoy a journal's revoke table. The table must already be empty! */
270 void journal_destroy_revoke(journal_t *journal)
272 struct jbd_revoke_table_s *table;
273 struct list_head *hash_list;
276 table = journal->j_revoke_table[0];
280 for (i=0; i<table->hash_size; i++) {
281 hash_list = &table->hash_table[i];
282 J_ASSERT (list_empty(hash_list));
285 kfree(table->hash_table);
286 kmem_cache_free(revoke_table_cache, table);
287 journal->j_revoke = NULL;
289 table = journal->j_revoke_table[1];
293 for (i=0; i<table->hash_size; i++) {
294 hash_list = &table->hash_table[i];
295 J_ASSERT (list_empty(hash_list));
298 kfree(table->hash_table);
299 kmem_cache_free(revoke_table_cache, table);
300 journal->j_revoke = NULL;
307 * journal_revoke: revoke a given buffer_head from the journal. This
308 * prevents the block from being replayed during recovery if we take a
309 * crash after this current transaction commits. Any subsequent
310 * metadata writes of the buffer in this transaction cancel the
313 * Note that this call may block --- it is up to the caller to make
314 * sure that there are no further calls to journal_write_metadata
315 * before the revoke is complete. In ext3, this implies calling the
316 * revoke before clearing the block bitmap when we are deleting
319 * Revoke performs a journal_forget on any buffer_head passed in as a
320 * parameter, but does _not_ forget the buffer_head if the bh was only
323 * bh_in may not be a journalled buffer - it may have come off
324 * the hash tables without an attached journal_head.
326 * If bh_in is non-zero, journal_revoke() will decrement its b_count
330 int journal_revoke(handle_t *handle, unsigned long blocknr,
331 struct buffer_head *bh_in)
333 struct buffer_head *bh = NULL;
335 struct block_device *bdev;
340 BUFFER_TRACE(bh_in, "enter");
342 journal = handle->h_transaction->t_journal;
343 if (!journal_set_features(journal, 0, 0, JFS_FEATURE_INCOMPAT_REVOKE)){
344 J_ASSERT (!"Cannot set revoke feature!");
348 bdev = journal->j_fs_dev;
352 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
354 BUFFER_TRACE(bh, "found on hash");
356 #ifdef JBD_EXPENSIVE_CHECKING
358 struct buffer_head *bh2;
360 /* If there is a different buffer_head lying around in
361 * memory anywhere... */
362 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
364 /* ... and it has RevokeValid status... */
365 if (bh2 != bh && buffer_revokevalid(bh2))
366 /* ...then it better be revoked too,
367 * since it's illegal to create a revoke
368 * record against a buffer_head which is
369 * not marked revoked --- that would
370 * risk missing a subsequent revoke
372 J_ASSERT_BH(bh2, buffer_revoked(bh2));
378 /* We really ought not ever to revoke twice in a row without
379 first having the revoke cancelled: it's illegal to free a
380 block twice without allocating it in between! */
382 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
383 "inconsistent data on disk")) {
388 set_buffer_revoked(bh);
389 set_buffer_revokevalid(bh);
391 BUFFER_TRACE(bh_in, "call journal_forget");
392 journal_forget(handle, bh_in);
394 BUFFER_TRACE(bh, "call brelse");
399 jbd_debug(2, "insert revoke for block %lu, bh_in=%p\n", blocknr, bh_in);
400 err = insert_revoke_hash(journal, blocknr,
401 handle->h_transaction->t_tid);
402 BUFFER_TRACE(bh_in, "exit");
407 * Cancel an outstanding revoke. For use only internally by the
408 * journaling code (called from journal_get_write_access).
410 * We trust buffer_revoked() on the buffer if the buffer is already
411 * being journaled: if there is no revoke pending on the buffer, then we
412 * don't do anything here.
414 * This would break if it were possible for a buffer to be revoked and
415 * discarded, and then reallocated within the same transaction. In such
416 * a case we would have lost the revoked bit, but when we arrived here
417 * the second time we would still have a pending revoke to cancel. So,
418 * do not trust the Revoked bit on buffers unless RevokeValid is also
421 * The caller must have the journal locked.
423 int journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
425 struct jbd_revoke_record_s *record;
426 journal_t *journal = handle->h_transaction->t_journal;
428 int did_revoke = 0; /* akpm: debug */
429 struct buffer_head *bh = jh2bh(jh);
431 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
433 /* Is the existing Revoke bit valid? If so, we trust it, and
434 * only perform the full cancel if the revoke bit is set. If
435 * not, we can't trust the revoke bit, and we need to do the
436 * full search for a revoke record. */
437 if (test_set_buffer_revokevalid(bh)) {
438 need_cancel = test_clear_buffer_revoked(bh);
441 clear_buffer_revoked(bh);
445 record = find_revoke_record(journal, bh->b_blocknr);
447 jbd_debug(4, "cancelled existing revoke on "
448 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
449 spin_lock(&journal->j_revoke_lock);
450 list_del(&record->hash);
451 spin_unlock(&journal->j_revoke_lock);
452 kmem_cache_free(revoke_record_cache, record);
457 #ifdef JBD_EXPENSIVE_CHECKING
458 /* There better not be one left behind by now! */
459 record = find_revoke_record(journal, bh->b_blocknr);
460 J_ASSERT_JH(jh, record == NULL);
463 /* Finally, have we just cleared revoke on an unhashed
464 * buffer_head? If so, we'd better make sure we clear the
465 * revoked status on any hashed alias too, otherwise the revoke
466 * state machine will get very upset later on. */
468 struct buffer_head *bh2;
469 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
472 clear_buffer_revoked(bh2);
479 /* journal_switch_revoke table select j_revoke for next transaction
480 * we do not want to suspend any processing until all revokes are
483 void journal_switch_revoke_table(journal_t *journal)
487 if (journal->j_revoke == journal->j_revoke_table[0])
488 journal->j_revoke = journal->j_revoke_table[1];
490 journal->j_revoke = journal->j_revoke_table[0];
492 for (i = 0; i < journal->j_revoke->hash_size; i++)
493 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
497 * Write revoke records to the journal for all entries in the current
498 * revoke hash, deleting the entries as we go.
500 * Called with the journal lock held.
503 void journal_write_revoke_records(journal_t *journal,
504 transaction_t *transaction)
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);
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)
550 struct journal_head *descriptor;
552 journal_header_t *header;
554 /* If we are already aborting, this all becomes a noop. We
555 still need to go round the loop in
556 journal_write_revoke_records in order to free all of the
557 revoke records: only the IO to the journal is omitted. */
558 if (is_journal_aborted(journal))
561 descriptor = *descriptorp;
564 /* Make sure we have a descriptor with space left for the record */
566 if (offset == journal->j_blocksize) {
567 flush_descriptor(journal, descriptor, offset);
573 descriptor = journal_get_descriptor_buffer(journal);
576 header = (journal_header_t *) &jh2bh(descriptor)->b_data[0];
577 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
578 header->h_blocktype = cpu_to_be32(JFS_REVOKE_BLOCK);
579 header->h_sequence = cpu_to_be32(transaction->t_tid);
581 /* Record it so that we can wait for IO completion later */
582 JBUFFER_TRACE(descriptor, "file as BJ_LogCtl");
583 journal_file_buffer(descriptor, transaction, BJ_LogCtl);
585 offset = sizeof(journal_revoke_header_t);
586 *descriptorp = descriptor;
589 * ((__be32 *)(&jh2bh(descriptor)->b_data[offset])) =
590 cpu_to_be32(record->blocknr);
596 * Flush a revoke descriptor out to the journal. If we are aborting,
597 * this is a noop; otherwise we are generating a buffer which needs to
598 * be waited for during commit, so it has to go onto the appropriate
599 * journal buffer list.
602 static void flush_descriptor(journal_t *journal,
603 struct journal_head *descriptor,
606 journal_revoke_header_t *header;
607 struct buffer_head *bh = jh2bh(descriptor);
609 if (is_journal_aborted(journal)) {
614 header = (journal_revoke_header_t *) jh2bh(descriptor)->b_data;
615 header->r_count = cpu_to_be32(offset);
616 set_buffer_jwrite(bh);
617 BUFFER_TRACE(bh, "write");
618 set_buffer_dirty(bh);
619 ll_rw_block(SWRITE, 1, &bh);
624 * Revoke support for recovery.
626 * Recovery needs to be able to:
628 * record all revoke records, including the tid of the latest instance
629 * of each revoke in the journal
631 * check whether a given block in a given transaction should be replayed
632 * (ie. has not been revoked by a revoke record in that or a subsequent
635 * empty the revoke table after recovery.
639 * First, setting revoke records. We create a new revoke record for
640 * every block ever revoked in the log as we scan it for recovery, and
641 * we update the existing records if we find multiple revokes for a
645 int journal_set_revoke(journal_t *journal,
646 unsigned long blocknr,
649 struct jbd_revoke_record_s *record;
651 record = find_revoke_record(journal, blocknr);
653 /* If we have multiple occurrences, only record the
654 * latest sequence number in the hashed record */
655 if (tid_gt(sequence, record->sequence))
656 record->sequence = sequence;
659 return insert_revoke_hash(journal, blocknr, sequence);
663 * Test revoke records. For a given block referenced in the log, has
664 * that block been revoked? A revoke record with a given transaction
665 * sequence number revokes all blocks in that transaction and earlier
666 * ones, but later transactions still need replayed.
669 int journal_test_revoke(journal_t *journal,
670 unsigned long blocknr,
673 struct jbd_revoke_record_s *record;
675 record = find_revoke_record(journal, blocknr);
678 if (tid_gt(sequence, record->sequence))
684 * Finally, once recovery is over, we need to clear the revoke table so
685 * that it can be reused by the running filesystem.
688 void journal_clear_revoke(journal_t *journal)
691 struct list_head *hash_list;
692 struct jbd_revoke_record_s *record;
693 struct jbd_revoke_table_s *revoke;
695 revoke = journal->j_revoke;
697 for (i = 0; i < revoke->hash_size; i++) {
698 hash_list = &revoke->hash_table[i];
699 while (!list_empty(hash_list)) {
700 record = (struct jbd_revoke_record_s*) hash_list->next;
701 list_del(&record->hash);
702 kmem_cache_free(revoke_record_cache, record);