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55 * buffer has not been revoked, and cancel_revoke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58 * buffer has been revoked.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 lkmem_cache_t *jbd2_revoke_record_cache;
96 static lkmem_cache_t *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;
143 record = kmem_cache_alloc(jbd2_revoke_record_cache, GFP_NOFS);
147 record->sequence = seq;
148 record->blocknr = blocknr;
149 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
150 spin_lock(&journal->j_revoke_lock);
151 list_add(&record->hash, hash_list);
152 spin_unlock(&journal->j_revoke_lock);
156 if (!journal_oom_retry)
158 jbd_debug(1, "ENOMEM in %s, retrying\n", __func__);
163 /* Find a revoke record in the journal's hash table. */
165 static struct jbd2_revoke_record_s *find_revoke_record(journal_t *journal,
166 unsigned long long blocknr)
168 struct list_head *hash_list;
169 struct jbd2_revoke_record_s *record;
171 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
173 spin_lock(&journal->j_revoke_lock);
174 record = (struct jbd2_revoke_record_s *) hash_list->next;
175 while (&(record->hash) != hash_list) {
176 if (record->blocknr == blocknr) {
177 spin_unlock(&journal->j_revoke_lock);
180 record = (struct jbd2_revoke_record_s *) record->hash.next;
182 spin_unlock(&journal->j_revoke_lock);
186 void jbd2_journal_destroy_revoke_caches(void)
188 kmem_cache_destroy(jbd2_revoke_record_cache);
189 jbd2_revoke_record_cache = NULL;
190 kmem_cache_destroy(jbd2_revoke_table_cache);
191 jbd2_revoke_table_cache = NULL;
194 int __init jbd2_journal_init_revoke_caches(void)
196 J_ASSERT(!jbd2_revoke_record_cache);
197 J_ASSERT(!jbd2_revoke_table_cache);
199 jbd2_revoke_record_cache = KMEM_CACHE(jbd2_revoke_record_s,
200 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY);
201 if (!jbd2_revoke_record_cache)
202 goto record_cache_failure;
204 jbd2_revoke_table_cache = KMEM_CACHE(jbd2_revoke_table_s,
206 if (!jbd2_revoke_table_cache)
207 goto table_cache_failure;
210 jbd2_journal_destroy_revoke_caches();
211 record_cache_failure:
215 static struct jbd2_revoke_table_s *jbd2_journal_init_revoke_table(int hash_size)
219 struct jbd2_revoke_table_s *table;
221 table = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
225 while((tmp >>= 1UL) != 0UL)
228 table->hash_size = hash_size;
229 table->hash_shift = shift;
231 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
232 if (!table->hash_table) {
233 kmem_cache_free(jbd2_revoke_table_cache, table);
238 for (tmp = 0; tmp < hash_size; tmp++)
239 INIT_LIST_HEAD(&table->hash_table[tmp]);
245 static void jbd2_journal_destroy_revoke_table(struct jbd2_revoke_table_s *table)
248 struct list_head *hash_list;
250 for (i = 0; i < table->hash_size; i++) {
251 hash_list = &table->hash_table[i];
252 J_ASSERT(list_empty(hash_list));
255 kfree(table->hash_table);
256 kmem_cache_free(jbd2_revoke_table_cache, table);
259 /* Initialise the revoke table for a given journal to a given size. */
260 int jbd2_journal_init_revoke(journal_t *journal, int hash_size)
262 J_ASSERT(journal->j_revoke_table[0] == NULL);
263 J_ASSERT(is_power_of_2(hash_size));
265 journal->j_revoke_table[0] = jbd2_journal_init_revoke_table(hash_size);
266 if (!journal->j_revoke_table[0])
269 journal->j_revoke_table[1] = jbd2_journal_init_revoke_table(hash_size);
270 if (!journal->j_revoke_table[1])
273 journal->j_revoke = journal->j_revoke_table[1];
275 spin_lock_init(&journal->j_revoke_lock);
280 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
281 journal->j_revoke_table[0] = NULL;
286 /* Destroy a journal's revoke table. The table must already be empty! */
287 void jbd2_journal_destroy_revoke(journal_t *journal)
289 journal->j_revoke = NULL;
290 if (journal->j_revoke_table[0])
291 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
292 if (journal->j_revoke_table[1])
293 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[1]);
300 * jbd2_journal_revoke: revoke a given buffer_head from the journal. This
301 * prevents the block from being replayed during recovery if we take a
302 * crash after this current transaction commits. Any subsequent
303 * metadata writes of the buffer in this transaction cancel the
306 * Note that this call may block --- it is up to the caller to make
307 * sure that there are no further calls to journal_write_metadata
308 * before the revoke is complete. In ext3, this implies calling the
309 * revoke before clearing the block bitmap when we are deleting
312 * Revoke performs a jbd2_journal_forget on any buffer_head passed in as a
313 * parameter, but does _not_ forget the buffer_head if the bh was only
316 * bh_in may not be a journalled buffer - it may have come off
317 * the hash tables without an attached journal_head.
319 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
323 int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
324 struct buffer_head *bh_in)
326 struct buffer_head *bh = NULL;
328 struct block_device *bdev;
333 BUFFER_TRACE(bh_in, "enter");
335 journal = handle->h_transaction->t_journal;
336 if (!jbd2_journal_set_features(journal, 0, 0, JBD2_FEATURE_INCOMPAT_REVOKE)){
337 J_ASSERT (!"Cannot set revoke feature!");
341 bdev = journal->j_fs_dev;
345 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
347 BUFFER_TRACE(bh, "found on hash");
349 #ifdef JBD2_EXPENSIVE_CHECKING
351 struct buffer_head *bh2;
353 /* If there is a different buffer_head lying around in
354 * memory anywhere... */
355 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
357 /* ... and it has RevokeValid status... */
358 if (bh2 != bh && buffer_revokevalid(bh2))
359 /* ...then it better be revoked too,
360 * since it's illegal to create a revoke
361 * record against a buffer_head which is
362 * not marked revoked --- that would
363 * risk missing a subsequent revoke
365 J_ASSERT_BH(bh2, buffer_revoked(bh2));
371 if (WARN_ON_ONCE(handle->h_revoke_credits <= 0)) {
376 /* We really ought not ever to revoke twice in a row without
377 first having the revoke cancelled: it's illegal to free a
378 block twice without allocating it in between! */
380 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
381 "inconsistent data on disk")) {
386 set_buffer_revoked(bh);
387 set_buffer_revokevalid(bh);
389 BUFFER_TRACE(bh_in, "call jbd2_journal_forget");
390 jbd2_journal_forget(handle, bh_in);
392 BUFFER_TRACE(bh, "call brelse");
396 handle->h_revoke_credits--;
398 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr, bh_in);
399 err = insert_revoke_hash(journal, blocknr,
400 handle->h_transaction->t_tid);
401 BUFFER_TRACE(bh_in, "exit");
406 * Cancel an outstanding revoke. For use only internally by the
407 * journaling code (called from jbd2_journal_get_write_access).
409 * We trust buffer_revoked() on the buffer if the buffer is already
410 * being journaled: if there is no revoke pending on the buffer, then we
411 * don't do anything here.
413 * This would break if it were possible for a buffer to be revoked and
414 * discarded, and then reallocated within the same transaction. In such
415 * a case we would have lost the revoked bit, but when we arrived here
416 * the second time we would still have a pending revoke to cancel. So,
417 * do not trust the Revoked bit on buffers unless RevokeValid is also
420 int jbd2_journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
422 struct jbd2_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(jbd2_revoke_record_cache, record);
454 #ifdef JBD2_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);
477 * journal_clear_revoked_flag clears revoked flag of buffers in
478 * revoke table to reflect there is no revoked buffers in the next
479 * transaction which is going to be started.
481 void jbd2_clear_buffer_revoked_flags(journal_t *journal)
483 struct jbd2_revoke_table_s *revoke = journal->j_revoke;
486 for (i = 0; i < revoke->hash_size; i++) {
487 struct list_head *hash_list;
488 struct list_head *list_entry;
489 hash_list = &revoke->hash_table[i];
491 list_for_each(list_entry, hash_list) {
492 struct jbd2_revoke_record_s *record;
493 struct buffer_head *bh;
494 record = (struct jbd2_revoke_record_s *)list_entry;
495 bh = __find_get_block(journal->j_fs_dev,
497 journal->j_blocksize);
499 clear_buffer_revoked(bh);
506 /* journal_switch_revoke table select j_revoke for next transaction
507 * we do not want to suspend any processing until all revokes are
510 void jbd2_journal_switch_revoke_table(journal_t *journal)
514 if (journal->j_revoke == journal->j_revoke_table[0])
515 journal->j_revoke = journal->j_revoke_table[1];
517 journal->j_revoke = journal->j_revoke_table[0];
519 for (i = 0; i < journal->j_revoke->hash_size; i++)
520 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
524 * Write revoke records to the journal for all entries in the current
525 * revoke hash, deleting the entries as we go.
527 void jbd2_journal_write_revoke_records(transaction_t *transaction,
528 struct list_head *log_bufs)
530 journal_t *journal = transaction->t_journal;
531 struct buffer_head *descriptor;
532 struct jbd2_revoke_record_s *record;
533 struct jbd2_revoke_table_s *revoke;
534 struct list_head *hash_list;
535 int i, offset, count;
541 /* select revoke table for committing transaction */
542 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
543 journal->j_revoke_table[1] : journal->j_revoke_table[0];
545 for (i = 0; i < revoke->hash_size; i++) {
546 hash_list = &revoke->hash_table[i];
548 while (!list_empty(hash_list)) {
549 record = (struct jbd2_revoke_record_s *)
551 write_one_revoke_record(transaction, log_bufs,
552 &descriptor, &offset, record);
554 list_del(&record->hash);
555 kmem_cache_free(jbd2_revoke_record_cache, record);
559 flush_descriptor(journal, descriptor, offset);
560 jbd_debug(1, "Wrote %d revoke records\n", count);
564 * Write out one revoke record. We need to create a new descriptor
565 * block if the old one is full or if we have not already created one.
568 static void write_one_revoke_record(transaction_t *transaction,
569 struct list_head *log_bufs,
570 struct buffer_head **descriptorp,
572 struct jbd2_revoke_record_s *record)
574 journal_t *journal = transaction->t_journal;
576 struct buffer_head *descriptor;
579 /* If we are already aborting, this all becomes a noop. We
580 still need to go round the loop in
581 jbd2_journal_write_revoke_records in order to free all of the
582 revoke records: only the IO to the journal is omitted. */
583 if (is_journal_aborted(journal))
586 descriptor = *descriptorp;
589 /* Do we need to leave space at the end for a checksum? */
590 if (jbd2_journal_has_csum_v2or3(journal))
591 csum_size = sizeof(struct jbd2_journal_block_tail);
593 if (jbd2_has_feature_64bit(journal))
598 /* Make sure we have a descriptor with space left for the record */
600 if (offset + sz > journal->j_blocksize - csum_size) {
601 flush_descriptor(journal, descriptor, offset);
607 descriptor = jbd2_journal_get_descriptor_buffer(transaction,
612 /* Record it so that we can wait for IO completion later */
613 BUFFER_TRACE(descriptor, "file in log_bufs");
614 jbd2_file_log_bh(log_bufs, descriptor);
616 offset = sizeof(jbd2_journal_revoke_header_t);
617 *descriptorp = descriptor;
620 if (jbd2_has_feature_64bit(journal))
621 * ((__be64 *)(&descriptor->b_data[offset])) =
622 cpu_to_be64(record->blocknr);
624 * ((__be32 *)(&descriptor->b_data[offset])) =
625 cpu_to_be32(record->blocknr);
632 * Flush a revoke descriptor out to the journal. If we are aborting,
633 * this is a noop; otherwise we are generating a buffer which needs to
634 * be waited for during commit, so it has to go onto the appropriate
635 * journal buffer list.
638 static void flush_descriptor(journal_t *journal,
639 struct buffer_head *descriptor,
642 jbd2_journal_revoke_header_t *header;
644 if (is_journal_aborted(journal))
647 header = (jbd2_journal_revoke_header_t *)descriptor->b_data;
648 header->r_count = ext2fs_cpu_to_be32(offset);
649 jbd2_descriptor_block_csum_set(journal, descriptor);
651 set_buffer_jwrite(descriptor);
652 BUFFER_TRACE(descriptor, "write");
653 set_buffer_dirty(descriptor);
654 write_dirty_buffer(descriptor, REQ_SYNC);
659 * Revoke support for recovery.
661 * Recovery needs to be able to:
663 * record all revoke records, including the tid of the latest instance
664 * of each revoke in the journal
666 * check whether a given block in a given transaction should be replayed
667 * (ie. has not been revoked by a revoke record in that or a subsequent
670 * empty the revoke table after recovery.
674 * First, setting revoke records. We create a new revoke record for
675 * every block ever revoked in the log as we scan it for recovery, and
676 * we update the existing records if we find multiple revokes for a
680 int jbd2_journal_set_revoke(journal_t *journal,
681 unsigned long long blocknr,
684 struct jbd2_revoke_record_s *record;
686 record = find_revoke_record(journal, blocknr);
688 /* If we have multiple occurrences, only record the
689 * latest sequence number in the hashed record */
690 if (tid_gt(sequence, record->sequence))
691 record->sequence = sequence;
694 return insert_revoke_hash(journal, blocknr, sequence);
698 * Test revoke records. For a given block referenced in the log, has
699 * that block been revoked? A revoke record with a given transaction
700 * sequence number revokes all blocks in that transaction and earlier
701 * ones, but later transactions still need replayed.
704 int jbd2_journal_test_revoke(journal_t *journal,
705 unsigned long long blocknr,
708 struct jbd2_revoke_record_s *record;
710 record = find_revoke_record(journal, blocknr);
713 if (tid_gt(sequence, record->sequence))
719 * Finally, once recovery is over, we need to clear the revoke table so
720 * that it can be reused by the running filesystem.
723 void jbd2_journal_clear_revoke(journal_t *journal)
726 struct list_head *hash_list;
727 struct jbd2_revoke_record_s *record;
728 struct jbd2_revoke_table_s *revoke;
730 revoke = journal->j_revoke;
732 for (i = 0; i < revoke->hash_size; i++) {
733 hash_list = &revoke->hash_table[i];
734 while (!list_empty(hash_list)) {
735 record = (struct jbd2_revoke_record_s*) hash_list->next;
736 list_del(&record->hash);
737 kmem_cache_free(jbd2_revoke_record_cache, record);