Loading...
1/*
2 * linux/fs/jbd2/revoke.c
3 *
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 2000
5 *
6 * Copyright 2000 Red Hat corp --- All Rights Reserved
7 *
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.
11 *
12 * Journal revoke routines for the generic filesystem journaling code;
13 * part of the ext2fs journaling system.
14 *
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:
18 *
19 * + Commit: during commit we write the entire list of the current
20 * transaction's revoked blocks to the journal
21 *
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
26 * gets replayed.
27 *
28 * We can get interactions between revokes and new log data within a
29 * single transaction:
30 *
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.
34 *
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.
43 *
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
48 * bit here.
49 *
50 * We cache revoke status of a buffer in the current transaction in b_states
51 * bits. As the name says, revokevalid flag indicates that the cached revoke
52 * status of a buffer is valid and we can rely on the cached status.
53 *
54 * Revoke information on buffers is a tri-state value:
55 *
56 * RevokeValid clear: no cached revoke status, need to look it up
57 * RevokeValid set, Revoked clear:
58 * buffer has not been revoked, and cancel_revoke
59 * need do nothing.
60 * RevokeValid set, Revoked set:
61 * buffer has been revoked.
62 *
63 * Locking rules:
64 * We keep two hash tables of revoke records. One hashtable belongs to the
65 * running transaction (is pointed to by journal->j_revoke), the other one
66 * belongs to the committing transaction. Accesses to the second hash table
67 * happen only from the kjournald and no other thread touches this table. Also
68 * journal_switch_revoke_table() which switches which hashtable belongs to the
69 * running and which to the committing transaction is called only from
70 * kjournald. Therefore we need no locks when accessing the hashtable belonging
71 * to the committing transaction.
72 *
73 * All users operating on the hash table belonging to the running transaction
74 * have a handle to the transaction. Therefore they are safe from kjournald
75 * switching hash tables under them. For operations on the lists of entries in
76 * the hash table j_revoke_lock is used.
77 *
78 * Finally, also replay code uses the hash tables but at this moment no one else
79 * can touch them (filesystem isn't mounted yet) and hence no locking is
80 * needed.
81 */
82
83#ifndef __KERNEL__
84#include "jfs_user.h"
85#else
86#include <linux/time.h>
87#include <linux/fs.h>
88#include <linux/jbd2.h>
89#include <linux/errno.h>
90#include <linux/slab.h>
91#include <linux/list.h>
92#include <linux/init.h>
93#include <linux/bio.h>
94#endif
95#include <linux/log2.h>
96
97static struct kmem_cache *jbd2_revoke_record_cache;
98static struct kmem_cache *jbd2_revoke_table_cache;
99
100/* Each revoke record represents one single revoked block. During
101 journal replay, this involves recording the transaction ID of the
102 last transaction to revoke this block. */
103
104struct jbd2_revoke_record_s
105{
106 struct list_head hash;
107 tid_t sequence; /* Used for recovery only */
108 unsigned long long blocknr;
109};
110
111
112/* The revoke table is just a simple hash table of revoke records. */
113struct jbd2_revoke_table_s
114{
115 /* It is conceivable that we might want a larger hash table
116 * for recovery. Must be a power of two. */
117 int hash_size;
118 int hash_shift;
119 struct list_head *hash_table;
120};
121
122
123#ifdef __KERNEL__
124static void write_one_revoke_record(journal_t *, transaction_t *,
125 struct list_head *,
126 struct buffer_head **, int *,
127 struct jbd2_revoke_record_s *, int);
128static void flush_descriptor(journal_t *, struct buffer_head *, int, int);
129#endif
130
131/* Utility functions to maintain the revoke table */
132
133/* Borrowed from buffer.c: this is a tried and tested block hash function */
134static inline int hash(journal_t *journal, unsigned long long block)
135{
136 struct jbd2_revoke_table_s *table = journal->j_revoke;
137 int hash_shift = table->hash_shift;
138 int hash = (int)block ^ (int)((block >> 31) >> 1);
139
140 return ((hash << (hash_shift - 6)) ^
141 (hash >> 13) ^
142 (hash << (hash_shift - 12))) & (table->hash_size - 1);
143}
144
145static int insert_revoke_hash(journal_t *journal, unsigned long long blocknr,
146 tid_t seq)
147{
148 struct list_head *hash_list;
149 struct jbd2_revoke_record_s *record;
150
151repeat:
152 record = kmem_cache_alloc(jbd2_revoke_record_cache, GFP_NOFS);
153 if (!record)
154 goto oom;
155
156 record->sequence = seq;
157 record->blocknr = blocknr;
158 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
159 spin_lock(&journal->j_revoke_lock);
160 list_add(&record->hash, hash_list);
161 spin_unlock(&journal->j_revoke_lock);
162 return 0;
163
164oom:
165 if (!journal_oom_retry)
166 return -ENOMEM;
167 jbd_debug(1, "ENOMEM in %s, retrying\n", __func__);
168 yield();
169 goto repeat;
170}
171
172/* Find a revoke record in the journal's hash table. */
173
174static struct jbd2_revoke_record_s *find_revoke_record(journal_t *journal,
175 unsigned long long blocknr)
176{
177 struct list_head *hash_list;
178 struct jbd2_revoke_record_s *record;
179
180 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
181
182 spin_lock(&journal->j_revoke_lock);
183 record = (struct jbd2_revoke_record_s *) hash_list->next;
184 while (&(record->hash) != hash_list) {
185 if (record->blocknr == blocknr) {
186 spin_unlock(&journal->j_revoke_lock);
187 return record;
188 }
189 record = (struct jbd2_revoke_record_s *) record->hash.next;
190 }
191 spin_unlock(&journal->j_revoke_lock);
192 return NULL;
193}
194
195void jbd2_journal_destroy_revoke_caches(void)
196{
197 if (jbd2_revoke_record_cache) {
198 kmem_cache_destroy(jbd2_revoke_record_cache);
199 jbd2_revoke_record_cache = NULL;
200 }
201 if (jbd2_revoke_table_cache) {
202 kmem_cache_destroy(jbd2_revoke_table_cache);
203 jbd2_revoke_table_cache = NULL;
204 }
205}
206
207int __init jbd2_journal_init_revoke_caches(void)
208{
209 J_ASSERT(!jbd2_revoke_record_cache);
210 J_ASSERT(!jbd2_revoke_table_cache);
211
212 jbd2_revoke_record_cache = KMEM_CACHE(jbd2_revoke_record_s,
213 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY);
214 if (!jbd2_revoke_record_cache)
215 goto record_cache_failure;
216
217 jbd2_revoke_table_cache = KMEM_CACHE(jbd2_revoke_table_s,
218 SLAB_TEMPORARY);
219 if (!jbd2_revoke_table_cache)
220 goto table_cache_failure;
221 return 0;
222table_cache_failure:
223 jbd2_journal_destroy_revoke_caches();
224record_cache_failure:
225 return -ENOMEM;
226}
227
228static struct jbd2_revoke_table_s *jbd2_journal_init_revoke_table(int hash_size)
229{
230 int shift = 0;
231 int tmp = hash_size;
232 struct jbd2_revoke_table_s *table;
233
234 table = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
235 if (!table)
236 goto out;
237
238 while((tmp >>= 1UL) != 0UL)
239 shift++;
240
241 table->hash_size = hash_size;
242 table->hash_shift = shift;
243 table->hash_table =
244 kmalloc(hash_size * sizeof(struct list_head), GFP_KERNEL);
245 if (!table->hash_table) {
246 kmem_cache_free(jbd2_revoke_table_cache, table);
247 table = NULL;
248 goto out;
249 }
250
251 for (tmp = 0; tmp < hash_size; tmp++)
252 INIT_LIST_HEAD(&table->hash_table[tmp]);
253
254out:
255 return table;
256}
257
258static void jbd2_journal_destroy_revoke_table(struct jbd2_revoke_table_s *table)
259{
260 int i;
261 struct list_head *hash_list;
262
263 for (i = 0; i < table->hash_size; i++) {
264 hash_list = &table->hash_table[i];
265 J_ASSERT(list_empty(hash_list));
266 }
267
268 kfree(table->hash_table);
269 kmem_cache_free(jbd2_revoke_table_cache, table);
270}
271
272/* Initialise the revoke table for a given journal to a given size. */
273int jbd2_journal_init_revoke(journal_t *journal, int hash_size)
274{
275 J_ASSERT(journal->j_revoke_table[0] == NULL);
276 J_ASSERT(is_power_of_2(hash_size));
277
278 journal->j_revoke_table[0] = jbd2_journal_init_revoke_table(hash_size);
279 if (!journal->j_revoke_table[0])
280 goto fail0;
281
282 journal->j_revoke_table[1] = jbd2_journal_init_revoke_table(hash_size);
283 if (!journal->j_revoke_table[1])
284 goto fail1;
285
286 journal->j_revoke = journal->j_revoke_table[1];
287
288 spin_lock_init(&journal->j_revoke_lock);
289
290 return 0;
291
292fail1:
293 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
294fail0:
295 return -ENOMEM;
296}
297
298/* Destroy a journal's revoke table. The table must already be empty! */
299void jbd2_journal_destroy_revoke(journal_t *journal)
300{
301 journal->j_revoke = NULL;
302 if (journal->j_revoke_table[0])
303 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
304 if (journal->j_revoke_table[1])
305 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[1]);
306}
307
308
309#ifdef __KERNEL__
310
311/*
312 * jbd2_journal_revoke: revoke a given buffer_head from the journal. This
313 * prevents the block from being replayed during recovery if we take a
314 * crash after this current transaction commits. Any subsequent
315 * metadata writes of the buffer in this transaction cancel the
316 * revoke.
317 *
318 * Note that this call may block --- it is up to the caller to make
319 * sure that there are no further calls to journal_write_metadata
320 * before the revoke is complete. In ext3, this implies calling the
321 * revoke before clearing the block bitmap when we are deleting
322 * metadata.
323 *
324 * Revoke performs a jbd2_journal_forget on any buffer_head passed in as a
325 * parameter, but does _not_ forget the buffer_head if the bh was only
326 * found implicitly.
327 *
328 * bh_in may not be a journalled buffer - it may have come off
329 * the hash tables without an attached journal_head.
330 *
331 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
332 * by one.
333 */
334
335int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
336 struct buffer_head *bh_in)
337{
338 struct buffer_head *bh = NULL;
339 journal_t *journal;
340 struct block_device *bdev;
341 int err;
342
343 might_sleep();
344 if (bh_in)
345 BUFFER_TRACE(bh_in, "enter");
346
347 journal = handle->h_transaction->t_journal;
348 if (!jbd2_journal_set_features(journal, 0, 0, JBD2_FEATURE_INCOMPAT_REVOKE)){
349 J_ASSERT (!"Cannot set revoke feature!");
350 return -EINVAL;
351 }
352
353 bdev = journal->j_fs_dev;
354 bh = bh_in;
355
356 if (!bh) {
357 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
358 if (bh)
359 BUFFER_TRACE(bh, "found on hash");
360 }
361#ifdef JBD2_EXPENSIVE_CHECKING
362 else {
363 struct buffer_head *bh2;
364
365 /* If there is a different buffer_head lying around in
366 * memory anywhere... */
367 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
368 if (bh2) {
369 /* ... and it has RevokeValid status... */
370 if (bh2 != bh && buffer_revokevalid(bh2))
371 /* ...then it better be revoked too,
372 * since it's illegal to create a revoke
373 * record against a buffer_head which is
374 * not marked revoked --- that would
375 * risk missing a subsequent revoke
376 * cancel. */
377 J_ASSERT_BH(bh2, buffer_revoked(bh2));
378 put_bh(bh2);
379 }
380 }
381#endif
382
383 /* We really ought not ever to revoke twice in a row without
384 first having the revoke cancelled: it's illegal to free a
385 block twice without allocating it in between! */
386 if (bh) {
387 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
388 "inconsistent data on disk")) {
389 if (!bh_in)
390 brelse(bh);
391 return -EIO;
392 }
393 set_buffer_revoked(bh);
394 set_buffer_revokevalid(bh);
395 if (bh_in) {
396 BUFFER_TRACE(bh_in, "call jbd2_journal_forget");
397 jbd2_journal_forget(handle, bh_in);
398 } else {
399 BUFFER_TRACE(bh, "call brelse");
400 __brelse(bh);
401 }
402 }
403
404 jbd_debug(2, "insert revoke for block %llu, bh_in=%p\n",blocknr, bh_in);
405 err = insert_revoke_hash(journal, blocknr,
406 handle->h_transaction->t_tid);
407 BUFFER_TRACE(bh_in, "exit");
408 return err;
409}
410
411/*
412 * Cancel an outstanding revoke. For use only internally by the
413 * journaling code (called from jbd2_journal_get_write_access).
414 *
415 * We trust buffer_revoked() on the buffer if the buffer is already
416 * being journaled: if there is no revoke pending on the buffer, then we
417 * don't do anything here.
418 *
419 * This would break if it were possible for a buffer to be revoked and
420 * discarded, and then reallocated within the same transaction. In such
421 * a case we would have lost the revoked bit, but when we arrived here
422 * the second time we would still have a pending revoke to cancel. So,
423 * do not trust the Revoked bit on buffers unless RevokeValid is also
424 * set.
425 */
426int jbd2_journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
427{
428 struct jbd2_revoke_record_s *record;
429 journal_t *journal = handle->h_transaction->t_journal;
430 int need_cancel;
431 int did_revoke = 0; /* akpm: debug */
432 struct buffer_head *bh = jh2bh(jh);
433
434 jbd_debug(4, "journal_head %p, cancelling revoke\n", jh);
435
436 /* Is the existing Revoke bit valid? If so, we trust it, and
437 * only perform the full cancel if the revoke bit is set. If
438 * not, we can't trust the revoke bit, and we need to do the
439 * full search for a revoke record. */
440 if (test_set_buffer_revokevalid(bh)) {
441 need_cancel = test_clear_buffer_revoked(bh);
442 } else {
443 need_cancel = 1;
444 clear_buffer_revoked(bh);
445 }
446
447 if (need_cancel) {
448 record = find_revoke_record(journal, bh->b_blocknr);
449 if (record) {
450 jbd_debug(4, "cancelled existing revoke on "
451 "blocknr %llu\n", (unsigned long long)bh->b_blocknr);
452 spin_lock(&journal->j_revoke_lock);
453 list_del(&record->hash);
454 spin_unlock(&journal->j_revoke_lock);
455 kmem_cache_free(jbd2_revoke_record_cache, record);
456 did_revoke = 1;
457 }
458 }
459
460#ifdef JBD2_EXPENSIVE_CHECKING
461 /* There better not be one left behind by now! */
462 record = find_revoke_record(journal, bh->b_blocknr);
463 J_ASSERT_JH(jh, record == NULL);
464#endif
465
466 /* Finally, have we just cleared revoke on an unhashed
467 * buffer_head? If so, we'd better make sure we clear the
468 * revoked status on any hashed alias too, otherwise the revoke
469 * state machine will get very upset later on. */
470 if (need_cancel) {
471 struct buffer_head *bh2;
472 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
473 if (bh2) {
474 if (bh2 != bh)
475 clear_buffer_revoked(bh2);
476 __brelse(bh2);
477 }
478 }
479 return did_revoke;
480}
481
482/*
483 * journal_clear_revoked_flag clears revoked flag of buffers in
484 * revoke table to reflect there is no revoked buffers in the next
485 * transaction which is going to be started.
486 */
487void jbd2_clear_buffer_revoked_flags(journal_t *journal)
488{
489 struct jbd2_revoke_table_s *revoke = journal->j_revoke;
490 int i = 0;
491
492 for (i = 0; i < revoke->hash_size; i++) {
493 struct list_head *hash_list;
494 struct list_head *list_entry;
495 hash_list = &revoke->hash_table[i];
496
497 list_for_each(list_entry, hash_list) {
498 struct jbd2_revoke_record_s *record;
499 struct buffer_head *bh;
500 record = (struct jbd2_revoke_record_s *)list_entry;
501 bh = __find_get_block(journal->j_fs_dev,
502 record->blocknr,
503 journal->j_blocksize);
504 if (bh) {
505 clear_buffer_revoked(bh);
506 __brelse(bh);
507 }
508 }
509 }
510}
511
512/* journal_switch_revoke table select j_revoke for next transaction
513 * we do not want to suspend any processing until all revokes are
514 * written -bzzz
515 */
516void jbd2_journal_switch_revoke_table(journal_t *journal)
517{
518 int i;
519
520 if (journal->j_revoke == journal->j_revoke_table[0])
521 journal->j_revoke = journal->j_revoke_table[1];
522 else
523 journal->j_revoke = journal->j_revoke_table[0];
524
525 for (i = 0; i < journal->j_revoke->hash_size; i++)
526 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
527}
528
529/*
530 * Write revoke records to the journal for all entries in the current
531 * revoke hash, deleting the entries as we go.
532 */
533void jbd2_journal_write_revoke_records(journal_t *journal,
534 transaction_t *transaction,
535 struct list_head *log_bufs,
536 int write_op)
537{
538 struct buffer_head *descriptor;
539 struct jbd2_revoke_record_s *record;
540 struct jbd2_revoke_table_s *revoke;
541 struct list_head *hash_list;
542 int i, offset, count;
543
544 descriptor = NULL;
545 offset = 0;
546 count = 0;
547
548 /* select revoke table for committing transaction */
549 revoke = journal->j_revoke == journal->j_revoke_table[0] ?
550 journal->j_revoke_table[1] : journal->j_revoke_table[0];
551
552 for (i = 0; i < revoke->hash_size; i++) {
553 hash_list = &revoke->hash_table[i];
554
555 while (!list_empty(hash_list)) {
556 record = (struct jbd2_revoke_record_s *)
557 hash_list->next;
558 write_one_revoke_record(journal, transaction, log_bufs,
559 &descriptor, &offset,
560 record, write_op);
561 count++;
562 list_del(&record->hash);
563 kmem_cache_free(jbd2_revoke_record_cache, record);
564 }
565 }
566 if (descriptor)
567 flush_descriptor(journal, descriptor, offset, write_op);
568 jbd_debug(1, "Wrote %d revoke records\n", count);
569}
570
571/*
572 * Write out one revoke record. We need to create a new descriptor
573 * block if the old one is full or if we have not already created one.
574 */
575
576static void write_one_revoke_record(journal_t *journal,
577 transaction_t *transaction,
578 struct list_head *log_bufs,
579 struct buffer_head **descriptorp,
580 int *offsetp,
581 struct jbd2_revoke_record_s *record,
582 int write_op)
583{
584 int csum_size = 0;
585 struct buffer_head *descriptor;
586 int offset;
587 journal_header_t *header;
588
589 /* If we are already aborting, this all becomes a noop. We
590 still need to go round the loop in
591 jbd2_journal_write_revoke_records in order to free all of the
592 revoke records: only the IO to the journal is omitted. */
593 if (is_journal_aborted(journal))
594 return;
595
596 descriptor = *descriptorp;
597 offset = *offsetp;
598
599 /* Do we need to leave space at the end for a checksum? */
600 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_CSUM_V2))
601 csum_size = sizeof(struct jbd2_journal_revoke_tail);
602
603 /* Make sure we have a descriptor with space left for the record */
604 if (descriptor) {
605 if (offset >= journal->j_blocksize - csum_size) {
606 flush_descriptor(journal, descriptor, offset, write_op);
607 descriptor = NULL;
608 }
609 }
610
611 if (!descriptor) {
612 descriptor = jbd2_journal_get_descriptor_buffer(journal);
613 if (!descriptor)
614 return;
615 header = (journal_header_t *)descriptor->b_data;
616 header->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
617 header->h_blocktype = cpu_to_be32(JBD2_REVOKE_BLOCK);
618 header->h_sequence = cpu_to_be32(transaction->t_tid);
619
620 /* Record it so that we can wait for IO completion later */
621 BUFFER_TRACE(descriptor, "file in log_bufs");
622 jbd2_file_log_bh(log_bufs, descriptor);
623
624 offset = sizeof(jbd2_journal_revoke_header_t);
625 *descriptorp = descriptor;
626 }
627
628 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) {
629 * ((__be64 *)(&descriptor->b_data[offset])) =
630 cpu_to_be64(record->blocknr);
631 offset += 8;
632
633 } else {
634 * ((__be32 *)(&descriptor->b_data[offset])) =
635 cpu_to_be32(record->blocknr);
636 offset += 4;
637 }
638
639 *offsetp = offset;
640}
641
642static void jbd2_revoke_csum_set(journal_t *j, struct buffer_head *bh)
643{
644 struct jbd2_journal_revoke_tail *tail;
645 __u32 csum;
646
647 if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
648 return;
649
650 tail = (struct jbd2_journal_revoke_tail *)(bh->b_data + j->j_blocksize -
651 sizeof(struct jbd2_journal_revoke_tail));
652 tail->r_checksum = 0;
653 csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
654 tail->r_checksum = cpu_to_be32(csum);
655}
656
657/*
658 * Flush a revoke descriptor out to the journal. If we are aborting,
659 * this is a noop; otherwise we are generating a buffer which needs to
660 * be waited for during commit, so it has to go onto the appropriate
661 * journal buffer list.
662 */
663
664static void flush_descriptor(journal_t *journal,
665 struct buffer_head *descriptor,
666 int offset, int write_op)
667{
668 jbd2_journal_revoke_header_t *header;
669
670 if (is_journal_aborted(journal)) {
671 put_bh(descriptor);
672 return;
673 }
674
675 header = (jbd2_journal_revoke_header_t *)descriptor->b_data;
676 header->r_count = cpu_to_be32(offset);
677 jbd2_revoke_csum_set(journal, descriptor);
678
679 set_buffer_jwrite(descriptor);
680 BUFFER_TRACE(descriptor, "write");
681 set_buffer_dirty(descriptor);
682 write_dirty_buffer(descriptor, write_op);
683}
684#endif
685
686/*
687 * Revoke support for recovery.
688 *
689 * Recovery needs to be able to:
690 *
691 * record all revoke records, including the tid of the latest instance
692 * of each revoke in the journal
693 *
694 * check whether a given block in a given transaction should be replayed
695 * (ie. has not been revoked by a revoke record in that or a subsequent
696 * transaction)
697 *
698 * empty the revoke table after recovery.
699 */
700
701/*
702 * First, setting revoke records. We create a new revoke record for
703 * every block ever revoked in the log as we scan it for recovery, and
704 * we update the existing records if we find multiple revokes for a
705 * single block.
706 */
707
708int jbd2_journal_set_revoke(journal_t *journal,
709 unsigned long long blocknr,
710 tid_t sequence)
711{
712 struct jbd2_revoke_record_s *record;
713
714 record = find_revoke_record(journal, blocknr);
715 if (record) {
716 /* If we have multiple occurrences, only record the
717 * latest sequence number in the hashed record */
718 if (tid_gt(sequence, record->sequence))
719 record->sequence = sequence;
720 return 0;
721 }
722 return insert_revoke_hash(journal, blocknr, sequence);
723}
724
725/*
726 * Test revoke records. For a given block referenced in the log, has
727 * that block been revoked? A revoke record with a given transaction
728 * sequence number revokes all blocks in that transaction and earlier
729 * ones, but later transactions still need replayed.
730 */
731
732int jbd2_journal_test_revoke(journal_t *journal,
733 unsigned long long blocknr,
734 tid_t sequence)
735{
736 struct jbd2_revoke_record_s *record;
737
738 record = find_revoke_record(journal, blocknr);
739 if (!record)
740 return 0;
741 if (tid_gt(sequence, record->sequence))
742 return 0;
743 return 1;
744}
745
746/*
747 * Finally, once recovery is over, we need to clear the revoke table so
748 * that it can be reused by the running filesystem.
749 */
750
751void jbd2_journal_clear_revoke(journal_t *journal)
752{
753 int i;
754 struct list_head *hash_list;
755 struct jbd2_revoke_record_s *record;
756 struct jbd2_revoke_table_s *revoke;
757
758 revoke = journal->j_revoke;
759
760 for (i = 0; i < revoke->hash_size; i++) {
761 hash_list = &revoke->hash_table[i];
762 while (!list_empty(hash_list)) {
763 record = (struct jbd2_revoke_record_s*) hash_list->next;
764 list_del(&record->hash);
765 kmem_cache_free(jbd2_revoke_record_cache, record);
766 }
767 }
768}
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * linux/fs/jbd2/revoke.c
4 *
5 * Written by Stephen C. Tweedie <sct@redhat.com>, 2000
6 *
7 * Copyright 2000 Red Hat corp --- All Rights Reserved
8 *
9 * Journal revoke routines for the generic filesystem journaling code;
10 * part of the ext2fs journaling system.
11 *
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:
15 *
16 * + Commit: during commit we write the entire list of the current
17 * transaction's revoked blocks to the journal
18 *
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
23 * gets replayed.
24 *
25 * We can get interactions between revokes and new log data within a
26 * single transaction:
27 *
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.
31 *
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.
40 *
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
45 * bit here.
46 *
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.
50 *
51 * Revoke information on buffers is a tri-state value:
52 *
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
56 * need do nothing.
57 * RevokeValid set, Revoked set:
58 * buffer has been revoked.
59 *
60 * Locking rules:
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.
69 *
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.
74 *
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
77 * needed.
78 */
79
80#ifndef __KERNEL__
81#include "jfs_user.h"
82#else
83#include <linux/time.h>
84#include <linux/fs.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>
93#endif
94
95static struct kmem_cache *jbd2_revoke_record_cache;
96static struct kmem_cache *jbd2_revoke_table_cache;
97
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. */
101
102struct jbd2_revoke_record_s
103{
104 struct list_head hash;
105 tid_t sequence; /* Used for recovery only */
106 unsigned long long blocknr;
107};
108
109
110/* The revoke table is just a simple hash table of revoke records. */
111struct jbd2_revoke_table_s
112{
113 /* It is conceivable that we might want a larger hash table
114 * for recovery. Must be a power of two. */
115 int hash_size;
116 int hash_shift;
117 struct list_head *hash_table;
118};
119
120
121#ifdef __KERNEL__
122static void write_one_revoke_record(transaction_t *,
123 struct list_head *,
124 struct buffer_head **, int *,
125 struct jbd2_revoke_record_s *);
126static void flush_descriptor(journal_t *, struct buffer_head *, int);
127#endif
128
129/* Utility functions to maintain the revoke table */
130
131static inline int hash(journal_t *journal, unsigned long long block)
132{
133 return hash_64(block, journal->j_revoke->hash_shift);
134}
135
136static int insert_revoke_hash(journal_t *journal, unsigned long long blocknr,
137 tid_t seq)
138{
139 struct list_head *hash_list;
140 struct jbd2_revoke_record_s *record;
141 gfp_t gfp_mask = GFP_NOFS;
142
143 if (journal_oom_retry)
144 gfp_mask |= __GFP_NOFAIL;
145 record = kmem_cache_alloc(jbd2_revoke_record_cache, gfp_mask);
146 if (!record)
147 return -ENOMEM;
148
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);
155 return 0;
156}
157
158/* Find a revoke record in the journal's hash table. */
159
160static struct jbd2_revoke_record_s *find_revoke_record(journal_t *journal,
161 unsigned long long blocknr)
162{
163 struct list_head *hash_list;
164 struct jbd2_revoke_record_s *record;
165
166 hash_list = &journal->j_revoke->hash_table[hash(journal, blocknr)];
167
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);
173 return record;
174 }
175 record = (struct jbd2_revoke_record_s *) record->hash.next;
176 }
177 spin_unlock(&journal->j_revoke_lock);
178 return NULL;
179}
180
181void jbd2_journal_destroy_revoke_record_cache(void)
182{
183 kmem_cache_destroy(jbd2_revoke_record_cache);
184 jbd2_revoke_record_cache = NULL;
185}
186
187void jbd2_journal_destroy_revoke_table_cache(void)
188{
189 kmem_cache_destroy(jbd2_revoke_table_cache);
190 jbd2_revoke_table_cache = NULL;
191}
192
193int __init jbd2_journal_init_revoke_record_cache(void)
194{
195 J_ASSERT(!jbd2_revoke_record_cache);
196 jbd2_revoke_record_cache = KMEM_CACHE(jbd2_revoke_record_s,
197 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY);
198
199 if (!jbd2_revoke_record_cache) {
200 pr_emerg("JBD2: failed to create revoke_record cache\n");
201 return -ENOMEM;
202 }
203 return 0;
204}
205
206int __init jbd2_journal_init_revoke_table_cache(void)
207{
208 J_ASSERT(!jbd2_revoke_table_cache);
209 jbd2_revoke_table_cache = KMEM_CACHE(jbd2_revoke_table_s,
210 SLAB_TEMPORARY);
211 if (!jbd2_revoke_table_cache) {
212 pr_emerg("JBD2: failed to create revoke_table cache\n");
213 return -ENOMEM;
214 }
215 return 0;
216}
217
218static struct jbd2_revoke_table_s *jbd2_journal_init_revoke_table(int hash_size)
219{
220 int shift = 0;
221 int tmp = hash_size;
222 struct jbd2_revoke_table_s *table;
223
224 table = kmem_cache_alloc(jbd2_revoke_table_cache, GFP_KERNEL);
225 if (!table)
226 goto out;
227
228 while((tmp >>= 1UL) != 0UL)
229 shift++;
230
231 table->hash_size = hash_size;
232 table->hash_shift = shift;
233 table->hash_table =
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);
237 table = NULL;
238 goto out;
239 }
240
241 for (tmp = 0; tmp < hash_size; tmp++)
242 INIT_LIST_HEAD(&table->hash_table[tmp]);
243
244out:
245 return table;
246}
247
248static void jbd2_journal_destroy_revoke_table(struct jbd2_revoke_table_s *table)
249{
250 int i;
251 struct list_head *hash_list;
252
253 for (i = 0; i < table->hash_size; i++) {
254 hash_list = &table->hash_table[i];
255 J_ASSERT(list_empty(hash_list));
256 }
257
258 kfree(table->hash_table);
259 kmem_cache_free(jbd2_revoke_table_cache, table);
260}
261
262/* Initialise the revoke table for a given journal to a given size. */
263int jbd2_journal_init_revoke(journal_t *journal, int hash_size)
264{
265 J_ASSERT(journal->j_revoke_table[0] == NULL);
266 J_ASSERT(is_power_of_2(hash_size));
267
268 journal->j_revoke_table[0] = jbd2_journal_init_revoke_table(hash_size);
269 if (!journal->j_revoke_table[0])
270 goto fail0;
271
272 journal->j_revoke_table[1] = jbd2_journal_init_revoke_table(hash_size);
273 if (!journal->j_revoke_table[1])
274 goto fail1;
275
276 journal->j_revoke = journal->j_revoke_table[1];
277
278 spin_lock_init(&journal->j_revoke_lock);
279
280 return 0;
281
282fail1:
283 jbd2_journal_destroy_revoke_table(journal->j_revoke_table[0]);
284 journal->j_revoke_table[0] = NULL;
285fail0:
286 return -ENOMEM;
287}
288
289/* Destroy a journal's revoke table. The table must already be empty! */
290void jbd2_journal_destroy_revoke(journal_t *journal)
291{
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]);
297}
298
299
300#ifdef __KERNEL__
301
302/*
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
307 * revoke.
308 *
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
313 * metadata.
314 *
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
317 * found implicitly.
318 *
319 * bh_in may not be a journalled buffer - it may have come off
320 * the hash tables without an attached journal_head.
321 *
322 * If bh_in is non-zero, jbd2_journal_revoke() will decrement its b_count
323 * by one.
324 */
325
326int jbd2_journal_revoke(handle_t *handle, unsigned long long blocknr,
327 struct buffer_head *bh_in)
328{
329 struct buffer_head *bh = NULL;
330 journal_t *journal;
331 struct block_device *bdev;
332 int err;
333
334 might_sleep();
335 if (bh_in)
336 BUFFER_TRACE(bh_in, "enter");
337
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!");
341 return -EINVAL;
342 }
343
344 bdev = journal->j_fs_dev;
345 bh = bh_in;
346
347 if (!bh) {
348 bh = __find_get_block(bdev, blocknr, journal->j_blocksize);
349 if (bh)
350 BUFFER_TRACE(bh, "found on hash");
351 }
352#ifdef JBD2_EXPENSIVE_CHECKING
353 else {
354 struct buffer_head *bh2;
355
356 /* If there is a different buffer_head lying around in
357 * memory anywhere... */
358 bh2 = __find_get_block(bdev, blocknr, journal->j_blocksize);
359 if (bh2) {
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
367 * cancel. */
368 J_ASSERT_BH(bh2, buffer_revoked(bh2));
369 put_bh(bh2);
370 }
371 }
372#endif
373
374 if (WARN_ON_ONCE(handle->h_revoke_credits <= 0)) {
375 if (!bh_in)
376 brelse(bh);
377 return -EIO;
378 }
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! */
382 if (bh) {
383 if (!J_EXPECT_BH(bh, !buffer_revoked(bh),
384 "inconsistent data on disk")) {
385 if (!bh_in)
386 brelse(bh);
387 return -EIO;
388 }
389 set_buffer_revoked(bh);
390 set_buffer_revokevalid(bh);
391 if (bh_in) {
392 BUFFER_TRACE(bh_in, "call jbd2_journal_forget");
393 jbd2_journal_forget(handle, bh_in);
394 } else {
395 BUFFER_TRACE(bh, "call brelse");
396 __brelse(bh);
397 }
398 }
399 handle->h_revoke_credits--;
400
401 jbd2_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");
405 return err;
406}
407
408/*
409 * Cancel an outstanding revoke. For use only internally by the
410 * journaling code (called from jbd2_journal_get_write_access).
411 *
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.
415 *
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
421 * set.
422 */
423int jbd2_journal_cancel_revoke(handle_t *handle, struct journal_head *jh)
424{
425 struct jbd2_revoke_record_s *record;
426 journal_t *journal = handle->h_transaction->t_journal;
427 int need_cancel;
428 int did_revoke = 0; /* akpm: debug */
429 struct buffer_head *bh = jh2bh(jh);
430
431 jbd2_debug(4, "journal_head %p, cancelling revoke\n", jh);
432
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);
439 } else {
440 need_cancel = 1;
441 clear_buffer_revoked(bh);
442 }
443
444 if (need_cancel) {
445 record = find_revoke_record(journal, bh->b_blocknr);
446 if (record) {
447 jbd2_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);
453 did_revoke = 1;
454 }
455 }
456
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);
461#endif
462
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. */
467 if (need_cancel) {
468 struct buffer_head *bh2;
469 bh2 = __find_get_block(bh->b_bdev, bh->b_blocknr, bh->b_size);
470 if (bh2) {
471 if (bh2 != bh)
472 clear_buffer_revoked(bh2);
473 __brelse(bh2);
474 }
475 }
476 return did_revoke;
477}
478
479/*
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.
483 */
484void jbd2_clear_buffer_revoked_flags(journal_t *journal)
485{
486 struct jbd2_revoke_table_s *revoke = journal->j_revoke;
487 int i = 0;
488
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];
493
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,
499 record->blocknr,
500 journal->j_blocksize);
501 if (bh) {
502 clear_buffer_revoked(bh);
503 __brelse(bh);
504 }
505 }
506 }
507}
508
509/* journal_switch_revoke table select j_revoke for next transaction
510 * we do not want to suspend any processing until all revokes are
511 * written -bzzz
512 */
513void jbd2_journal_switch_revoke_table(journal_t *journal)
514{
515 int i;
516
517 if (journal->j_revoke == journal->j_revoke_table[0])
518 journal->j_revoke = journal->j_revoke_table[1];
519 else
520 journal->j_revoke = journal->j_revoke_table[0];
521
522 for (i = 0; i < journal->j_revoke->hash_size; i++)
523 INIT_LIST_HEAD(&journal->j_revoke->hash_table[i]);
524}
525
526/*
527 * Write revoke records to the journal for all entries in the current
528 * revoke hash, deleting the entries as we go.
529 */
530void jbd2_journal_write_revoke_records(transaction_t *transaction,
531 struct list_head *log_bufs)
532{
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;
539
540 descriptor = NULL;
541 offset = 0;
542 count = 0;
543
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];
547
548 for (i = 0; i < revoke->hash_size; i++) {
549 hash_list = &revoke->hash_table[i];
550
551 while (!list_empty(hash_list)) {
552 record = (struct jbd2_revoke_record_s *)
553 hash_list->next;
554 write_one_revoke_record(transaction, log_bufs,
555 &descriptor, &offset, record);
556 count++;
557 list_del(&record->hash);
558 kmem_cache_free(jbd2_revoke_record_cache, record);
559 }
560 }
561 if (descriptor)
562 flush_descriptor(journal, descriptor, offset);
563 jbd2_debug(1, "Wrote %d revoke records\n", count);
564}
565
566/*
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.
569 */
570
571static void write_one_revoke_record(transaction_t *transaction,
572 struct list_head *log_bufs,
573 struct buffer_head **descriptorp,
574 int *offsetp,
575 struct jbd2_revoke_record_s *record)
576{
577 journal_t *journal = transaction->t_journal;
578 int csum_size = 0;
579 struct buffer_head *descriptor;
580 int sz, offset;
581
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))
587 return;
588
589 descriptor = *descriptorp;
590 offset = *offsetp;
591
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);
595
596 if (jbd2_has_feature_64bit(journal))
597 sz = 8;
598 else
599 sz = 4;
600
601 /* Make sure we have a descriptor with space left for the record */
602 if (descriptor) {
603 if (offset + sz > journal->j_blocksize - csum_size) {
604 flush_descriptor(journal, descriptor, offset);
605 descriptor = NULL;
606 }
607 }
608
609 if (!descriptor) {
610 descriptor = jbd2_journal_get_descriptor_buffer(transaction,
611 JBD2_REVOKE_BLOCK);
612 if (!descriptor)
613 return;
614
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);
618
619 offset = sizeof(jbd2_journal_revoke_header_t);
620 *descriptorp = descriptor;
621 }
622
623 if (jbd2_has_feature_64bit(journal))
624 * ((__be64 *)(&descriptor->b_data[offset])) =
625 cpu_to_be64(record->blocknr);
626 else
627 * ((__be32 *)(&descriptor->b_data[offset])) =
628 cpu_to_be32(record->blocknr);
629 offset += sz;
630
631 *offsetp = offset;
632}
633
634/*
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.
639 */
640
641static void flush_descriptor(journal_t *journal,
642 struct buffer_head *descriptor,
643 int offset)
644{
645 jbd2_journal_revoke_header_t *header;
646
647 if (is_journal_aborted(journal))
648 return;
649
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);
653
654 set_buffer_jwrite(descriptor);
655 BUFFER_TRACE(descriptor, "write");
656 set_buffer_dirty(descriptor);
657 write_dirty_buffer(descriptor, JBD2_JOURNAL_REQ_FLAGS);
658}
659#endif
660
661/*
662 * Revoke support for recovery.
663 *
664 * Recovery needs to be able to:
665 *
666 * record all revoke records, including the tid of the latest instance
667 * of each revoke in the journal
668 *
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
671 * transaction)
672 *
673 * empty the revoke table after recovery.
674 */
675
676/*
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
680 * single block.
681 */
682
683int jbd2_journal_set_revoke(journal_t *journal,
684 unsigned long long blocknr,
685 tid_t sequence)
686{
687 struct jbd2_revoke_record_s *record;
688
689 record = find_revoke_record(journal, blocknr);
690 if (record) {
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;
695 return 0;
696 }
697 return insert_revoke_hash(journal, blocknr, sequence);
698}
699
700/*
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.
705 */
706
707int jbd2_journal_test_revoke(journal_t *journal,
708 unsigned long long blocknr,
709 tid_t sequence)
710{
711 struct jbd2_revoke_record_s *record;
712
713 record = find_revoke_record(journal, blocknr);
714 if (!record)
715 return 0;
716 if (tid_gt(sequence, record->sequence))
717 return 0;
718 return 1;
719}
720
721/*
722 * Finally, once recovery is over, we need to clear the revoke table so
723 * that it can be reused by the running filesystem.
724 */
725
726void jbd2_journal_clear_revoke(journal_t *journal)
727{
728 int i;
729 struct list_head *hash_list;
730 struct jbd2_revoke_record_s *record;
731 struct jbd2_revoke_table_s *revoke;
732
733 revoke = journal->j_revoke;
734
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);
741 }
742 }
743}