| blob | d44e8b4a33919d68ca9b58f2766da865e10b6829 |
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
26 /*
27 * This structure is used during recovery to record the buf log items which
28 * have been canceled and should not be replayed.
29 */
31 xfs_daddr_t bc_blkno;
32 uint bc_len;
35 };
40 xfs_daddr_t blkno,
41 uint len)
42 {
53 }
56 }
58 static bool
61 xfs_daddr_t blkno,
62 uint len)
63 {
66 /*
67 * If we find an existing cancel record, this indicates that the buffer
68 * was cancelled multiple times. To ensure that during pass 2 we keep
69 * the record in the table until we reach its last occurrence in the
70 * log, a reference count is kept to tell how many times we expect to
71 * see this record during the second pass.
72 */
77 }
85 }
87 /*
88 * Check if there is and entry for blkno, len in the buffer cancel record table.
89 */
90 bool
93 xfs_daddr_t blkno,
94 uint len)
95 {
97 }
99 /*
100 * Check if there is and entry for blkno, len in the buffer cancel record table,
101 * and decremented the reference count on it if there is one.
102 *
103 * Remove the cancel record once the refcount hits zero, so that if the same
104 * buffer is re-used again after its last cancellation we actually replay the
105 * changes made at that point.
106 */
107 static bool
110 xfs_daddr_t blkno,
111 uint len)
112 {
119 }
124 }
126 }
128 /* log buffer item recovery */
130 /*
131 * Sort buffer items for log recovery. Most buffer items should end up on the
132 * buffer list and are recovered first, with the following exceptions:
133 *
134 * 1. XFS_BLF_CANCEL buffers must be processed last because some log items
135 * might depend on the incor ecancellation record, and replaying a cancelled
136 * buffer item can remove the incore record.
137 *
138 * 2. XFS_BLF_INODE_BUF buffers are handled after most regular items so that
139 * we replay di_next_unlinked only after flushing the inode 'free' state
140 * to the inode buffer.
141 *
142 * See xlog_recover_reorder_trans for more details.
143 */
144 STATIC enum xlog_recover_reorder
147 {
155 }
157 STATIC void
161 {
165 }
167 /*
168 * Build up the table of buf cancel records so that we don't replay cancelled
169 * data in the second pass.
170 */
171 static int
175 {
182 }
188 else
191 }
193 /*
194 * Validate the recovered buffer is of the correct type and attach the
195 * appropriate buffer operations to them for writeback. Magic numbers are in a
196 * few places:
197 * the first 16 bits of the buffer (inode buffer, dquot buffer),
198 * the first 32 bits of the buffer (most blocks),
199 * inside a struct xfs_da_blkinfo at the start of the buffer.
200 */
201 static void
206 xfs_lsn_t current_lsn)
207 {
214 /*
215 * We can only do post recovery validation on items on CRC enabled
216 * fielsystems as we need to know when the buffer was written to be able
217 * to determine if we should have replayed the item. If we replay old
218 * metadata over a newer buffer, then it will enter a temporarily
219 * inconsistent state resulting in verification failures. Hence for now
220 * just avoid the verification stage for non-crc filesystems
221 */
260 }
266 }
273 }
280 }
286 #ifdef CONFIG_XFS_QUOTA
290 }
292 #else
296 #endif
302 }
309 }
317 }
325 }
333 }
341 }
349 }
357 }
365 }
372 }
379 }
382 #ifdef CONFIG_XFS_RT
385 /* no magic numbers for verification of RT buffers */
393 }
395 /*
396 * Nothing else to do in the case of a NULL current LSN as this means
397 * the buffer is more recent than the change in the log and will be
398 * skipped.
399 */
406 }
408 /*
409 * We must update the metadata LSN of the buffer as it is written out to
410 * ensure that older transactions never replay over this one and corrupt
411 * the buffer. This can occur if log recovery is interrupted at some
412 * point after the current transaction completes, at which point a
413 * subsequent mount starts recovery from the beginning.
414 *
415 * Write verifiers update the metadata LSN from log items attached to
416 * the buffer. Therefore, initialize a bli purely to carry the LSN to
417 * the verifier.
418 */
426 }
427 }
429 /*
430 * Perform a 'normal' buffer recovery. Each logged region of the
431 * buffer should be copied over the corresponding region in the
432 * given buffer. The bitmap in the buf log format structure indicates
433 * where to place the logged data.
434 */
435 STATIC void
441 xfs_lsn_t current_lsn)
442 {
446 xfs_failaddr_t fa;
466 /*
467 * The dirty regions logged in the buffer, even though
468 * contiguous, may span multiple chunks. This is because the
469 * dirty region may span a physical page boundary in a buffer
470 * and hence be split into two separate vectors for writing into
471 * the log. Hence we need to trim nbits back to the length of
472 * the current region being copied out of the log.
473 */
477 /*
478 * Do a sanity check if this is a dquot buffer. Just checking
479 * the first dquot in the buffer should do. XXXThis is
480 * probably a good thing to do for other buf types also.
481 */
489 }
495 }
502 }
503 }
509 next:
510 i++;
512 }
514 /* Shouldn't be any more regions */
518 }
520 /*
521 * Perform a dquot buffer recovery.
522 * Simple algorithm: if we have found a QUOTAOFF log item of the same type
523 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
524 * Else, treat it as a regular buffer and do recovery.
525 *
526 * Return false if the buffer was tossed and true if we recovered the buffer to
527 * indicate to the caller if the buffer needs writing.
528 */
529 STATIC bool
536 {
537 uint type;
541 /*
542 * Filesystems are required to send in quota flags at mount time.
543 */
554 /*
555 * This type of quotas was turned off, so ignore this buffer
556 */
562 }
564 /*
565 * Perform recovery for a buffer full of inodes. In these buffers, the only
566 * data which should be recovered is that which corresponds to the
567 * di_next_unlinked pointers in the on disk inode structures. The rest of the
568 * data for the inodes is always logged through the inodes themselves rather
569 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
570 *
571 * The only time when buffers full of inodes are fully recovered is when the
572 * buffer is full of newly allocated inodes. In this case the buffer will
573 * not be marked as an inode buffer and so will be sent to
574 * xlog_recover_do_reg_buffer() below during recovery.
575 */
576 STATIC int
582 {
596 /*
597 * Post recovery validation only works properly on CRC enabled
598 * filesystems.
599 */
610 /*
611 * The next di_next_unlinked field is beyond
612 * the current logged region. Find the next
613 * logged region that contains or is beyond
614 * the current di_next_unlinked field.
615 */
620 /*
621 * If there are no more logged regions in the
622 * buffer, then we're done.
623 */
632 item_index++;
633 }
635 /*
636 * If the current logged region starts after the current
637 * di_next_unlinked field, then move on to the next
638 * di_next_unlinked field.
639 */
647 /*
648 * The current logged region contains a copy of the
649 * current di_next_unlinked field. Extract its value
650 * and copy it to the buffer copy.
651 */
660 }
665 /*
666 * If necessary, recalculate the CRC in the on-disk inode. We
667 * have to leave the inode in a consistent state for whoever
668 * reads it next....
669 */
673 }
676 }
678 /*
679 * V5 filesystems know the age of the buffer on disk being recovered. We can
680 * have newer objects on disk than we are replaying, and so for these cases we
681 * don't want to replay the current change as that will make the buffer contents
682 * temporarily invalid on disk.
683 *
684 * The magic number might not match the buffer type we are going to recover
685 * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags. Hence
686 * extract the LSN of the existing object in the buffer based on it's current
687 * magic number. If we don't recognise the magic number in the buffer, then
688 * return a LSN of -1 so that the caller knows it was an unrecognised block and
689 * so can recover the buffer.
690 *
691 * Note: we cannot rely solely on magic number matches to determine that the
692 * buffer has a valid LSN - we also need to verify that it belongs to this
693 * filesystem, so we need to extract the object's LSN and compare it to that
694 * which we read from the superblock. If the UUIDs don't match, then we've got a
695 * stale metadata block from an old filesystem instance that we need to recover
696 * over the top of.
697 */
698 static xfs_lsn_t
702 {
710 /* v4 filesystems always recover immediately */
731 }
739 }
763 /*
764 * Remote attr blocks are written synchronously, rather than
765 * being logged. That means they do not contain a valid LSN
766 * (i.e. transactionally ordered) in them, and hence any time we
767 * see a buffer to replay over the top of a remote attribute
768 * block we should simply do so.
769 */
772 /*
773 * superblock uuids are magic. We may or may not have a
774 * sb_meta_uuid on disk, but it will be set in the in-core
775 * superblock. We set the uuid pointer for verification
776 * according to the superblock feature mask to ensure we check
777 * the relevant UUID in the superblock.
778 */
782 else
787 }
793 }
805 }
811 }
813 /*
814 * We do individual object checks on dquot and inode buffers as they
815 * have their own individual LSN records. Also, we could have a stale
816 * buffer here, so we have to at least recognise these buffer types.
817 *
818 * A notd complexity here is inode unlinked list processing - it logs
819 * the inode directly in the buffer, but we don't know which inodes have
820 * been modified, and there is no global buffer LSN. Hence we need to
821 * recover all inode buffer types immediately. This problem will be
822 * fixed by logical logging of the unlinked list modifications.
823 */
831 }
833 /* unknown buffer contents, recover immediately */
835 recover_immediately:
838 }
840 /*
841 * This routine replays a modification made to a buffer at runtime.
842 * There are actually two types of buffer, regular and inode, which
843 * are handled differently. Inode buffers are handled differently
844 * in that we only recover a specific set of data from them, namely
845 * the inode di_next_unlinked fields. This is because all other inode
846 * data is actually logged via inode records and any data we replay
847 * here which overlaps that may be stale.
848 *
849 * When meta-data buffers are freed at run time we log a buffer item
850 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
851 * of the buffer in the log should not be replayed at recovery time.
852 * This is so that if the blocks covered by the buffer are reused for
853 * file data before we crash we don't end up replaying old, freed
854 * meta-data into a user's file.
855 *
856 * To handle the cancellation of buffer log items, we make two passes
857 * over the log during recovery. During the first we build a table of
858 * those buffers which have been cancelled, and during the second we
859 * only replay those buffers which do not have corresponding cancel
860 * records in the table. See xlog_recover_buf_pass[1,2] above
861 * for more details on the implementation of the table of cancel records.
862 */
863 STATIC int
868 xfs_lsn_t current_lsn)
869 {
874 uint buf_flags;
875 xfs_lsn_t lsn;
877 /*
878 * In this pass we only want to recover all the buffers which have
879 * not been cancelled and are not cancellation buffers themselves.
880 */
890 }
903 /*
904 * Recover the buffer only if we get an LSN from it and it's less than
905 * the lsn of the transaction we are replaying.
906 *
907 * Note that we have to be extremely careful of readahead here.
908 * Readahead does not attach verfiers to the buffers so if we don't
909 * actually do any replay after readahead because of the LSN we found
910 * in the buffer if more recent than that current transaction then we
911 * need to attach the verifier directly. Failure to do so can lead to
912 * future recovery actions (e.g. EFI and unlinked list recovery) can
913 * operate on the buffers and they won't get the verifier attached. This
914 * can lead to blocks on disk having the correct content but a stale
915 * CRC.
916 *
917 * It is safe to assume these clean buffers are currently up to date.
918 * If the buffer is dirtied by a later transaction being replayed, then
919 * the verifier will be reset to match whatever recover turns that
920 * buffer into.
921 */
927 }
942 }
944 /*
945 * Perform delayed write on the buffer. Asynchronous writes will be
946 * slower when taking into account all the buffers to be flushed.
947 *
948 * Also make sure that only inode buffers with good sizes stay in
949 * the buffer cache. The kernel moves inodes in buffers of 1 block
950 * or inode_cluster_size bytes, whichever is bigger. The inode
951 * buffers in the log can be a different size if the log was generated
952 * by an older kernel using unclustered inode buffers or a newer kernel
953 * running with a different inode cluster size. Regardless, if
954 * the inode buffer size isn't max(blocksize, inode_cluster_size)
955 * for *our* value of inode_cluster_size, then we need to keep
956 * the buffer out of the buffer cache so that the buffer won't
957 * overlap with future reads of those inodes.
958 */
968 }
970 out_release:
973 cancelled:
976 }
984 };