1 // SPDX-License-Identifier: GPL-2.0+
3 * linux/fs/jbd2/revoke.c
5 * Written by Stephen C. Tweedie <sct@redhat.com>, 2000
7 * Copyright 2000 Red Hat corp --- All Rights Reserved
9 * Journal revoke routines for the generic filesystem journaling code;
10 * part of the ext2fs journaling system.
12 * Revoke is the mechanism used to prevent old log records for deleted
13 * metadata from being replayed on top of newer data using the same
14 * blocks. The revoke mechanism is used in two separate places:
16 * + Commit: during commit we write the entire list of the current
17 * transaction's revoked blocks to the journal
19 * + Recovery: during recovery we record the transaction ID of all
20 * revoked blocks. If there are multiple revoke records in the log
21 * for a single block, only the last one counts, and if there is a log
22 * entry for a block beyond the last revoke, then that log entry still
25 * We can get interactions between revokes and new log data within a
28 * Block is revoked and then journaled:
29 * The desired end result is the journaling of the new block, so we
30 * cancel the revoke before the transaction commits.
32 * Block is journaled and then revoked:
33 * The revoke must take precedence over the write of the block, so we
34 * need either to cancel the journal entry or to write the revoke
35 * later in the log than the log block. In this case, we choose the
36 * latter: journaling a block cancels any revoke record for that block
37 * in the current transaction, so any revoke for that block in the
38 * transaction must have happened after the block was journaled and so
39 * the revoke must take precedence.
41 * Block is revoked and then written as data:
42 * The data write is allowed to succeed, but the revoke is _not_
43 * cancelled. We still need to prevent old log records from
44 * overwriting the new data. We don't even need to clear the revoke
47 * We cache revoke status of a buffer in the current transaction in b_states
48 * bits. As the name says, revokevalid flag indicates that the cached revoke
49 * status of a buffer is valid and we can rely on the cached status.
51 * Revoke information on buffers is a tri-state value:
53 * RevokeValid clear: no cached revoke status, need to look it up
54 * RevokeValid set, Revoked clear:
55 * buffer has not been revoked, and cancel_revoke
57 * RevokeValid set, Revoked set:
58 * buffer has been revoked.
61 * We keep two hash tables of revoke records. One hashtable belongs to the
62 * running transaction (is pointed to by journal->j_revoke), the other one
63 * belongs to the committing transaction. Accesses to the second hash table
64 * happen only from the kjournald and no other thread touches this table. Also
65 * journal_switch_revoke_table() which switches which hashtable belongs to the
66 * running and which to the committing transaction is called only from
67 * kjournald. Therefore we need no locks when accessing the hashtable belonging
68 * to the committing transaction.
70 * All users operating on the hash table belonging to the running transaction
71 * have a handle to the transaction. Therefore they are safe from kjournald
72 * switching hash tables under them. For operations on the lists of entries in
73 * the hash table j_revoke_lock is used.
75 * Finally, also replay code uses the hash tables but at this moment no one else
76 * can touch them (filesystem isn't mounted yet) and hence no locking is
83 #include <linux/time.h>
85 #include <linux/jbd2.h>
86 #include <linux/errno.h>
87 #include <linux/slab.h>
88 #include <linux/list.h>
89 #include <linux/init.h>
90 #include <linux/bio.h>
91 #include <linux/log2.h>
92 #include <linux/hash.h>
95 static struct kmem_cache *jbd2_revoke_record_cache;
96 static struct kmem_cache *jbd2_revoke_table_cache;
98 /* Each revoke record represents one single revoked block. During
99 journal replay, this involves recording the transaction ID of the
100 last transaction to revoke this block. */
102 struct jbd2_revoke_record_s
104 struct list_head hash;
105 tid_t sequence; /* Used for recovery only */
106 unsigned long long blocknr;
110 /* The revoke table is just a simple hash table of revoke records. */
111 struct jbd2_revoke_table_s
113 /* It is conceivable that we might want a larger hash table
114 * for recovery. Must be a power of two. */
117 struct list_head *hash_table;
122 static void write_one_revoke_record(transaction_t *,
124 struct buffer_head **, int *,
125 struct jbd2_revoke_record_s *);
126 static void flush_descriptor(journal_t *, struct buffer_head *, int);
129 /* Utility functions to maintain the revoke table */
131 static inline int hash(journal_t *journal, unsigned long long block)
133 return hash_64(block, journal->j_revoke->hash_shift);
136 static int insert_revoke_hash(journal_t *journal, unsigned long long blocknr,
139 struct list_head *hash_list;
140 struct jbd2_revoke_record_s *record;
141 gfp_t gfp_mask = GFP_NOFS;
143 if (journal_oom_retry)
144 gfp_mask |= __GFP_NOFAIL;
145 record = kmem_cache_alloc(jbd2_revoke_record_cache, gfp_mask);
149 record->sequence = seq;
150 record->blocknr = blocknr;
151 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
152 spin_lock(&journal->j_revoke_lock);
153 list_add(&record->hash, hash_list);
154 spin_unlock(&journal->j_revoke_lock);
158 /* Find a revoke record in the journal's hash table. */
160 static struct jbd2_revoke_record_s *find_revoke_record(journal_t *journal,
161 unsigned long long blocknr)
163 struct list_head *hash_list;
164 struct jbd2_revoke_record_s *record;
166 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
168 spin_lock(&journal->j_revoke_lock);
169 record = (struct jbd2_revoke_record_s *) hash_list->next;
170 while (&(record->hash) != hash_list) {
171 if (record->blocknr == blocknr) {
172 spin_unlock(&journal->j_revoke_lock);
175 record = (struct jbd2_revoke_record_s *) record->hash.next;
177 spin_unlock(&journal->j_revoke_lock);
181 void jbd2_journal_destroy_revoke_record_cache(void)
183 kmem_cache_destroy(jbd2_revoke_record_cache);
184 jbd2_revoke_record_cache = NULL;
187 void jbd2_journal_destroy_revoke_table_cache(void)
189 kmem_cache_destroy(jbd2_revoke_table_cache);
190 jbd2_revoke_table_cache = NULL;
193 int __init jbd2_journal_init_revoke_record_cache(void)
195 J_ASSERT(!jbd2_revoke_record_cache);
196 jbd2_revoke_record_cache = KMEM_CACHE(jbd2_revoke_record_s,
197 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY);
199 if (!jbd2_revoke_record_cache) {
200 pr_emerg("JBD2: failed to create revoke_record cache\n");
206 int __init jbd2_journal_init_revoke_table_cache(void)
208 J_ASSERT(!jbd2_revoke_table_cache);
209 jbd2_revoke_table_cache = KMEM_CACHE(jbd2_revoke_table_s,
211 if (!jbd2_revoke_table_cache) {
212 pr_emerg("JBD2: failed to create revoke_table cache\n");
218 static struct jbd2_revoke_table_s *jbd2_journal_init_revoke_table(int hash_size)
222 struct jbd2_revoke_table_s *table;
224 table = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
228 while((tmp >>= 1UL) != 0UL)
231 table->hash_size = hash_size;
232 table->hash_shift = shift;
234 kmalloc_array(hash_size, sizeof(struct list_head), GFP_KERNEL);
235 if (!table->hash_table) {
236 kmem_cache_free(jbd2_revoke_table_cache, table);
241 for (tmp = 0; tmp < hash_size; tmp++)
242 INIT_LIST_HEAD(&table->hash_table[tmp]);
248 static void jbd2_journal_destroy_revoke_table(struct jbd2_revoke_table_s *table)
251 struct list_head *hash_list;
253 for (i = 0; i < table->hash_size; i++) {
254 hash_list = &table->hash_table[i];
255 J_ASSERT(list_empty(hash_list));
258 kfree(table->hash_table);
259 kmem_cache_free(jbd2_revoke_table_cache, table);
262 /* Initialise the revoke table for a given journal to a given size. */
263 int jbd2_journal_init_revoke(journal_t *journal, int hash_size)
265 J_ASSERT(journal->j_revoke_table[0] == NULL);
266 J_ASSERT(is_power_of_2(hash_size));
268 journal->j_revoke_table[0] = jbd2_journal_init_revoke_table(hash_size);
269 if (!journal->j_revoke_table[0])
272 journal->j_revoke_table[1] = jbd2_journal_init_revoke_table(hash_size);
273 if (!journal->j_revoke_table[1])
276 journal->j_revoke = journal->j_revoke_table[1];
278 spin_lock_init(&journal->j_revoke_lock);
283 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
284 journal->j_revoke_table[0] = NULL;
289 /* Destroy a journal's revoke table. The table must already be empty! */
290 void jbd2_journal_destroy_revoke(journal_t *journal)
292 journal->j_revoke = NULL;
293 if (journal->j_revoke_table[0])
294 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
295 if (journal->j_revoke_table[1])
296 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[1]);
303 * jbd2_journal_revoke: revoke a given buffer_head from the journal. This
304 * prevents the block from being replayed during recovery if we take a
305 * crash after this current transaction commits. Any subsequent
306 * metadata writes of the buffer in this transaction cancel the
309 * Note that this call may block --- it is up to the caller to make
310 * sure that there are no further calls to journal_write_metadata
311 * before the revoke is complete. In ext3, this implies calling the
312 * revoke before clearing the block bitmap when we are deleting
315 * Revoke performs a jbd2_journal_forget on any buffer_head passed in as a
316 * parameter, but does _not_ forget the buffer_head if the bh was only
319 * bh_in may not be a journalled buffer - it may have come off
320 * the hash tables without an attached journal_head.
322 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
326 int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
327 struct buffer_head *bh_in)
329 struct buffer_head *bh = NULL;
331 struct block_device *bdev;
336 BUFFER_TRACE(bh_in, "enter");
338 journal = handle->h_transaction->t_journal;
339 if (!jbd2_journal_set_features(journal, 0, 0, JBD2_FEATURE_INCOMPAT_REVOKE)){
340 J_ASSERT (!"Cannot set revoke feature!");
344 bdev = journal->j_fs_dev;
348 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
350 BUFFER_TRACE(bh, "found on hash");
352 #ifdef JBD2_EXPENSIVE_CHECKING
354 struct buffer_head *bh2;
356 /* If there is a different buffer_head lying around in
357 * memory anywhere... */
358 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
360 /* ... and it has RevokeValid status... */
361 if (bh2 != bh && buffer_revokevalid(bh2))
362 /* ...then it better be revoked too,
363 * since it's illegal to create a revoke
364 * record against a buffer_head which is
365 * not marked revoked --- that would
366 * risk missing a subsequent revoke
368 J_ASSERT_BH(bh2, buffer_revoked(bh2));
374 if (WARN_ON_ONCE(handle->h_revoke_credits <= 0)) {
379 /* We really ought not ever to revoke twice in a row without
380 first having the revoke cancelled: it's illegal to free a
381 block twice without allocating it in between! */
383 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
384 "inconsistent data on disk")) {
389 set_buffer_revoked(bh);
390 set_buffer_revokevalid(bh);
392 BUFFER_TRACE(bh_in, "call jbd2_journal_forget");
393 jbd2_journal_forget(handle, bh_in);
395 BUFFER_TRACE(bh, "call brelse");
399 handle->h_revoke_credits--;
401 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr, bh_in);
402 err = insert_revoke_hash(journal, blocknr,
403 handle->h_transaction->t_tid);
404 BUFFER_TRACE(bh_in, "exit");
409 * Cancel an outstanding revoke. For use only internally by the
410 * journaling code (called from jbd2_journal_get_write_access).
412 * We trust buffer_revoked() on the buffer if the buffer is already
413 * being journaled: if there is no revoke pending on the buffer, then we
414 * don't do anything here.
416 * This would break if it were possible for a buffer to be revoked and
417 * discarded, and then reallocated within the same transaction. In such
418 * a case we would have lost the revoked bit, but when we arrived here
419 * the second time we would still have a pending revoke to cancel. So,
420 * do not trust the Revoked bit on buffers unless RevokeValid is also
423 int jbd2_journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
425 struct jbd2_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(jbd2_revoke_record_cache, record);
457 #ifdef JBD2_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);
480 * journal_clear_revoked_flag clears revoked flag of buffers in
481 * revoke table to reflect there is no revoked buffers in the next
482 * transaction which is going to be started.
484 void jbd2_clear_buffer_revoked_flags(journal_t *journal)
486 struct jbd2_revoke_table_s *revoke = journal->j_revoke;
489 for (i = 0; i < revoke->hash_size; i++) {
490 struct list_head *hash_list;
491 struct list_head *list_entry;
492 hash_list = &revoke->hash_table[i];
494 list_for_each(list_entry, hash_list) {
495 struct jbd2_revoke_record_s *record;
496 struct buffer_head *bh;
497 record = (struct jbd2_revoke_record_s *)list_entry;
498 bh = __find_get_block(journal->j_fs_dev,
500 journal->j_blocksize);
502 clear_buffer_revoked(bh);
509 /* journal_switch_revoke table select j_revoke for next transaction
510 * we do not want to suspend any processing until all revokes are
513 void jbd2_journal_switch_revoke_table(journal_t *journal)
517 if (journal->j_revoke == journal->j_revoke_table[0])
518 journal->j_revoke = journal->j_revoke_table[1];
520 journal->j_revoke = journal->j_revoke_table[0];
522 for (i = 0; i < journal->j_revoke->hash_size; i++)
523 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
527 * Write revoke records to the journal for all entries in the current
528 * revoke hash, deleting the entries as we go.
530 void jbd2_journal_write_revoke_records(transaction_t *transaction,
531 struct list_head *log_bufs)
533 journal_t *journal = transaction->t_journal;
534 struct buffer_head *descriptor;
535 struct jbd2_revoke_record_s *record;
536 struct jbd2_revoke_table_s *revoke;
537 struct list_head *hash_list;
538 int i, offset, count;
544 /* select revoke table for committing transaction */
545 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
546 journal->j_revoke_table[1] : journal->j_revoke_table[0];
548 for (i = 0; i < revoke->hash_size; i++) {
549 hash_list = &revoke->hash_table[i];
551 while (!list_empty(hash_list)) {
552 record = (struct jbd2_revoke_record_s *)
554 write_one_revoke_record(transaction, log_bufs,
555 &descriptor, &offset, record);
557 list_del(&record->hash);
558 kmem_cache_free(jbd2_revoke_record_cache, record);
562 flush_descriptor(journal, descriptor, offset);
563 jbd_debug(1, "Wrote %d revoke records\n", count);
567 * Write out one revoke record. We need to create a new descriptor
568 * block if the old one is full or if we have not already created one.
571 static void write_one_revoke_record(transaction_t *transaction,
572 struct list_head *log_bufs,
573 struct buffer_head **descriptorp,
575 struct jbd2_revoke_record_s *record)
577 journal_t *journal = transaction->t_journal;
579 struct buffer_head *descriptor;
582 /* If we are already aborting, this all becomes a noop. We
583 still need to go round the loop in
584 jbd2_journal_write_revoke_records in order to free all of the
585 revoke records: only the IO to the journal is omitted. */
586 if (is_journal_aborted(journal))
589 descriptor = *descriptorp;
592 /* Do we need to leave space at the end for a checksum? */
593 if (jbd2_journal_has_csum_v2or3(journal))
594 csum_size = sizeof(struct jbd2_journal_block_tail);
596 if (jbd2_has_feature_64bit(journal))
601 /* Make sure we have a descriptor with space left for the record */
603 if (offset + sz > journal->j_blocksize - csum_size) {
604 flush_descriptor(journal, descriptor, offset);
610 descriptor = jbd2_journal_get_descriptor_buffer(transaction,
615 /* Record it so that we can wait for IO completion later */
616 BUFFER_TRACE(descriptor, "file in log_bufs");
617 jbd2_file_log_bh(log_bufs, descriptor);
619 offset = sizeof(jbd2_journal_revoke_header_t);
620 *descriptorp = descriptor;
623 if (jbd2_has_feature_64bit(journal))
624 * ((__be64 *)(&descriptor->b_data[offset])) =
625 cpu_to_be64(record->blocknr);
627 * ((__be32 *)(&descriptor->b_data[offset])) =
628 cpu_to_be32(record->blocknr);
635 * Flush a revoke descriptor out to the journal. If we are aborting,
636 * this is a noop; otherwise we are generating a buffer which needs to
637 * be waited for during commit, so it has to go onto the appropriate
638 * journal buffer list.
641 static void flush_descriptor(journal_t *journal,
642 struct buffer_head *descriptor,
645 jbd2_journal_revoke_header_t *header;
647 if (is_journal_aborted(journal))
650 header = (jbd2_journal_revoke_header_t *)descriptor->b_data;
651 header->r_count = cpu_to_be32(offset);
652 jbd2_descriptor_block_csum_set(journal, descriptor);
654 set_buffer_jwrite(descriptor);
655 BUFFER_TRACE(descriptor, "write");
656 set_buffer_dirty(descriptor);
657 write_dirty_buffer(descriptor, REQ_SYNC);
662 * Revoke support for recovery.
664 * Recovery needs to be able to:
666 * record all revoke records, including the tid of the latest instance
667 * of each revoke in the journal
669 * check whether a given block in a given transaction should be replayed
670 * (ie. has not been revoked by a revoke record in that or a subsequent
673 * empty the revoke table after recovery.
677 * First, setting revoke records. We create a new revoke record for
678 * every block ever revoked in the log as we scan it for recovery, and
679 * we update the existing records if we find multiple revokes for a
683 int jbd2_journal_set_revoke(journal_t *journal,
684 unsigned long long blocknr,
687 struct jbd2_revoke_record_s *record;
689 record = find_revoke_record(journal, blocknr);
691 /* If we have multiple occurrences, only record the
692 * latest sequence number in the hashed record */
693 if (tid_gt(sequence, record->sequence))
694 record->sequence = sequence;
697 return insert_revoke_hash(journal, blocknr, sequence);
701 * Test revoke records. For a given block referenced in the log, has
702 * that block been revoked? A revoke record with a given transaction
703 * sequence number revokes all blocks in that transaction and earlier
704 * ones, but later transactions still need replayed.
707 int jbd2_journal_test_revoke(journal_t *journal,
708 unsigned long long blocknr,
711 struct jbd2_revoke_record_s *record;
713 record = find_revoke_record(journal, blocknr);
716 if (tid_gt(sequence, record->sequence))
722 * Finally, once recovery is over, we need to clear the revoke table so
723 * that it can be reused by the running filesystem.
726 void jbd2_journal_clear_revoke(journal_t *journal)
729 struct list_head *hash_list;
730 struct jbd2_revoke_record_s *record;
731 struct jbd2_revoke_table_s *revoke;
733 revoke = journal->j_revoke;
735 for (i = 0; i < revoke->hash_size; i++) {
736 hash_list = &revoke->hash_table[i];
737 while (!list_empty(hash_list)) {
738 record = (struct jbd2_revoke_record_s*) hash_list->next;
739 list_del(&record->hash);
740 kmem_cache_free(jbd2_revoke_record_cache, record);