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