Merge branch 'firewire-kernel-streaming' of git://git.alsa-project.org/alsa-kprivate
[firewire-audio.git] / fs / xfs / xfs_log_recover.c
blobaa0ebb776903317a8d29916396eb547c3ce296c9
1 /*
2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3 * All Rights Reserved.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_mount.h"
28 #include "xfs_error.h"
29 #include "xfs_bmap_btree.h"
30 #include "xfs_alloc_btree.h"
31 #include "xfs_ialloc_btree.h"
32 #include "xfs_dinode.h"
33 #include "xfs_inode.h"
34 #include "xfs_inode_item.h"
35 #include "xfs_alloc.h"
36 #include "xfs_ialloc.h"
37 #include "xfs_log_priv.h"
38 #include "xfs_buf_item.h"
39 #include "xfs_log_recover.h"
40 #include "xfs_extfree_item.h"
41 #include "xfs_trans_priv.h"
42 #include "xfs_quota.h"
43 #include "xfs_rw.h"
44 #include "xfs_utils.h"
45 #include "xfs_trace.h"
47 STATIC int xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
48 STATIC int xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
49 #if defined(DEBUG)
50 STATIC void xlog_recover_check_summary(xlog_t *);
51 #else
52 #define xlog_recover_check_summary(log)
53 #endif
56 * This structure is used during recovery to record the buf log items which
57 * have been canceled and should not be replayed.
59 struct xfs_buf_cancel {
60 xfs_daddr_t bc_blkno;
61 uint bc_len;
62 int bc_refcount;
63 struct list_head bc_list;
67 * Sector aligned buffer routines for buffer create/read/write/access
71 * Verify the given count of basic blocks is valid number of blocks
72 * to specify for an operation involving the given XFS log buffer.
73 * Returns nonzero if the count is valid, 0 otherwise.
76 static inline int
77 xlog_buf_bbcount_valid(
78 xlog_t *log,
79 int bbcount)
81 return bbcount > 0 && bbcount <= log->l_logBBsize;
85 * Allocate a buffer to hold log data. The buffer needs to be able
86 * to map to a range of nbblks basic blocks at any valid (basic
87 * block) offset within the log.
89 STATIC xfs_buf_t *
90 xlog_get_bp(
91 xlog_t *log,
92 int nbblks)
94 if (!xlog_buf_bbcount_valid(log, nbblks)) {
95 xlog_warn("XFS: Invalid block length (0x%x) given for buffer",
96 nbblks);
97 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
98 return NULL;
102 * We do log I/O in units of log sectors (a power-of-2
103 * multiple of the basic block size), so we round up the
104 * requested size to acommodate the basic blocks required
105 * for complete log sectors.
107 * In addition, the buffer may be used for a non-sector-
108 * aligned block offset, in which case an I/O of the
109 * requested size could extend beyond the end of the
110 * buffer. If the requested size is only 1 basic block it
111 * will never straddle a sector boundary, so this won't be
112 * an issue. Nor will this be a problem if the log I/O is
113 * done in basic blocks (sector size 1). But otherwise we
114 * extend the buffer by one extra log sector to ensure
115 * there's space to accomodate this possiblility.
117 if (nbblks > 1 && log->l_sectBBsize > 1)
118 nbblks += log->l_sectBBsize;
119 nbblks = round_up(nbblks, log->l_sectBBsize);
121 return xfs_buf_get_uncached(log->l_mp->m_logdev_targp,
122 BBTOB(nbblks), 0);
125 STATIC void
126 xlog_put_bp(
127 xfs_buf_t *bp)
129 xfs_buf_free(bp);
133 * Return the address of the start of the given block number's data
134 * in a log buffer. The buffer covers a log sector-aligned region.
136 STATIC xfs_caddr_t
137 xlog_align(
138 xlog_t *log,
139 xfs_daddr_t blk_no,
140 int nbblks,
141 xfs_buf_t *bp)
143 xfs_daddr_t offset = blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1);
145 ASSERT(BBTOB(offset + nbblks) <= XFS_BUF_SIZE(bp));
146 return XFS_BUF_PTR(bp) + BBTOB(offset);
151 * nbblks should be uint, but oh well. Just want to catch that 32-bit length.
153 STATIC int
154 xlog_bread_noalign(
155 xlog_t *log,
156 xfs_daddr_t blk_no,
157 int nbblks,
158 xfs_buf_t *bp)
160 int error;
162 if (!xlog_buf_bbcount_valid(log, nbblks)) {
163 xlog_warn("XFS: Invalid block length (0x%x) given for buffer",
164 nbblks);
165 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
166 return EFSCORRUPTED;
169 blk_no = round_down(blk_no, log->l_sectBBsize);
170 nbblks = round_up(nbblks, log->l_sectBBsize);
172 ASSERT(nbblks > 0);
173 ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
175 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
176 XFS_BUF_READ(bp);
177 XFS_BUF_BUSY(bp);
178 XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
179 XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
181 xfsbdstrat(log->l_mp, bp);
182 error = xfs_buf_iowait(bp);
183 if (error)
184 xfs_ioerror_alert("xlog_bread", log->l_mp,
185 bp, XFS_BUF_ADDR(bp));
186 return error;
189 STATIC int
190 xlog_bread(
191 xlog_t *log,
192 xfs_daddr_t blk_no,
193 int nbblks,
194 xfs_buf_t *bp,
195 xfs_caddr_t *offset)
197 int error;
199 error = xlog_bread_noalign(log, blk_no, nbblks, bp);
200 if (error)
201 return error;
203 *offset = xlog_align(log, blk_no, nbblks, bp);
204 return 0;
208 * Write out the buffer at the given block for the given number of blocks.
209 * The buffer is kept locked across the write and is returned locked.
210 * This can only be used for synchronous log writes.
212 STATIC int
213 xlog_bwrite(
214 xlog_t *log,
215 xfs_daddr_t blk_no,
216 int nbblks,
217 xfs_buf_t *bp)
219 int error;
221 if (!xlog_buf_bbcount_valid(log, nbblks)) {
222 xlog_warn("XFS: Invalid block length (0x%x) given for buffer",
223 nbblks);
224 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
225 return EFSCORRUPTED;
228 blk_no = round_down(blk_no, log->l_sectBBsize);
229 nbblks = round_up(nbblks, log->l_sectBBsize);
231 ASSERT(nbblks > 0);
232 ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
234 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
235 XFS_BUF_ZEROFLAGS(bp);
236 XFS_BUF_BUSY(bp);
237 XFS_BUF_HOLD(bp);
238 XFS_BUF_PSEMA(bp, PRIBIO);
239 XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
240 XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
242 if ((error = xfs_bwrite(log->l_mp, bp)))
243 xfs_ioerror_alert("xlog_bwrite", log->l_mp,
244 bp, XFS_BUF_ADDR(bp));
245 return error;
248 #ifdef DEBUG
250 * dump debug superblock and log record information
252 STATIC void
253 xlog_header_check_dump(
254 xfs_mount_t *mp,
255 xlog_rec_header_t *head)
257 cmn_err(CE_DEBUG, "%s: SB : uuid = %pU, fmt = %d\n",
258 __func__, &mp->m_sb.sb_uuid, XLOG_FMT);
259 cmn_err(CE_DEBUG, " log : uuid = %pU, fmt = %d\n",
260 &head->h_fs_uuid, be32_to_cpu(head->h_fmt));
262 #else
263 #define xlog_header_check_dump(mp, head)
264 #endif
267 * check log record header for recovery
269 STATIC int
270 xlog_header_check_recover(
271 xfs_mount_t *mp,
272 xlog_rec_header_t *head)
274 ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
277 * IRIX doesn't write the h_fmt field and leaves it zeroed
278 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
279 * a dirty log created in IRIX.
281 if (unlikely(be32_to_cpu(head->h_fmt) != XLOG_FMT)) {
282 xlog_warn(
283 "XFS: dirty log written in incompatible format - can't recover");
284 xlog_header_check_dump(mp, head);
285 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
286 XFS_ERRLEVEL_HIGH, mp);
287 return XFS_ERROR(EFSCORRUPTED);
288 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
289 xlog_warn(
290 "XFS: dirty log entry has mismatched uuid - can't recover");
291 xlog_header_check_dump(mp, head);
292 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
293 XFS_ERRLEVEL_HIGH, mp);
294 return XFS_ERROR(EFSCORRUPTED);
296 return 0;
300 * read the head block of the log and check the header
302 STATIC int
303 xlog_header_check_mount(
304 xfs_mount_t *mp,
305 xlog_rec_header_t *head)
307 ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
309 if (uuid_is_nil(&head->h_fs_uuid)) {
311 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
312 * h_fs_uuid is nil, we assume this log was last mounted
313 * by IRIX and continue.
315 xlog_warn("XFS: nil uuid in log - IRIX style log");
316 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
317 xlog_warn("XFS: log has mismatched uuid - can't recover");
318 xlog_header_check_dump(mp, head);
319 XFS_ERROR_REPORT("xlog_header_check_mount",
320 XFS_ERRLEVEL_HIGH, mp);
321 return XFS_ERROR(EFSCORRUPTED);
323 return 0;
326 STATIC void
327 xlog_recover_iodone(
328 struct xfs_buf *bp)
330 if (XFS_BUF_GETERROR(bp)) {
332 * We're not going to bother about retrying
333 * this during recovery. One strike!
335 xfs_ioerror_alert("xlog_recover_iodone",
336 bp->b_target->bt_mount, bp,
337 XFS_BUF_ADDR(bp));
338 xfs_force_shutdown(bp->b_target->bt_mount,
339 SHUTDOWN_META_IO_ERROR);
341 XFS_BUF_CLR_IODONE_FUNC(bp);
342 xfs_buf_ioend(bp, 0);
346 * This routine finds (to an approximation) the first block in the physical
347 * log which contains the given cycle. It uses a binary search algorithm.
348 * Note that the algorithm can not be perfect because the disk will not
349 * necessarily be perfect.
351 STATIC int
352 xlog_find_cycle_start(
353 xlog_t *log,
354 xfs_buf_t *bp,
355 xfs_daddr_t first_blk,
356 xfs_daddr_t *last_blk,
357 uint cycle)
359 xfs_caddr_t offset;
360 xfs_daddr_t mid_blk;
361 xfs_daddr_t end_blk;
362 uint mid_cycle;
363 int error;
365 end_blk = *last_blk;
366 mid_blk = BLK_AVG(first_blk, end_blk);
367 while (mid_blk != first_blk && mid_blk != end_blk) {
368 error = xlog_bread(log, mid_blk, 1, bp, &offset);
369 if (error)
370 return error;
371 mid_cycle = xlog_get_cycle(offset);
372 if (mid_cycle == cycle)
373 end_blk = mid_blk; /* last_half_cycle == mid_cycle */
374 else
375 first_blk = mid_blk; /* first_half_cycle == mid_cycle */
376 mid_blk = BLK_AVG(first_blk, end_blk);
378 ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
379 (mid_blk == end_blk && mid_blk-1 == first_blk));
381 *last_blk = end_blk;
383 return 0;
387 * Check that a range of blocks does not contain stop_on_cycle_no.
388 * Fill in *new_blk with the block offset where such a block is
389 * found, or with -1 (an invalid block number) if there is no such
390 * block in the range. The scan needs to occur from front to back
391 * and the pointer into the region must be updated since a later
392 * routine will need to perform another test.
394 STATIC int
395 xlog_find_verify_cycle(
396 xlog_t *log,
397 xfs_daddr_t start_blk,
398 int nbblks,
399 uint stop_on_cycle_no,
400 xfs_daddr_t *new_blk)
402 xfs_daddr_t i, j;
403 uint cycle;
404 xfs_buf_t *bp;
405 xfs_daddr_t bufblks;
406 xfs_caddr_t buf = NULL;
407 int error = 0;
410 * Greedily allocate a buffer big enough to handle the full
411 * range of basic blocks we'll be examining. If that fails,
412 * try a smaller size. We need to be able to read at least
413 * a log sector, or we're out of luck.
415 bufblks = 1 << ffs(nbblks);
416 while (!(bp = xlog_get_bp(log, bufblks))) {
417 bufblks >>= 1;
418 if (bufblks < log->l_sectBBsize)
419 return ENOMEM;
422 for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
423 int bcount;
425 bcount = min(bufblks, (start_blk + nbblks - i));
427 error = xlog_bread(log, i, bcount, bp, &buf);
428 if (error)
429 goto out;
431 for (j = 0; j < bcount; j++) {
432 cycle = xlog_get_cycle(buf);
433 if (cycle == stop_on_cycle_no) {
434 *new_blk = i+j;
435 goto out;
438 buf += BBSIZE;
442 *new_blk = -1;
444 out:
445 xlog_put_bp(bp);
446 return error;
450 * Potentially backup over partial log record write.
452 * In the typical case, last_blk is the number of the block directly after
453 * a good log record. Therefore, we subtract one to get the block number
454 * of the last block in the given buffer. extra_bblks contains the number
455 * of blocks we would have read on a previous read. This happens when the
456 * last log record is split over the end of the physical log.
458 * extra_bblks is the number of blocks potentially verified on a previous
459 * call to this routine.
461 STATIC int
462 xlog_find_verify_log_record(
463 xlog_t *log,
464 xfs_daddr_t start_blk,
465 xfs_daddr_t *last_blk,
466 int extra_bblks)
468 xfs_daddr_t i;
469 xfs_buf_t *bp;
470 xfs_caddr_t offset = NULL;
471 xlog_rec_header_t *head = NULL;
472 int error = 0;
473 int smallmem = 0;
474 int num_blks = *last_blk - start_blk;
475 int xhdrs;
477 ASSERT(start_blk != 0 || *last_blk != start_blk);
479 if (!(bp = xlog_get_bp(log, num_blks))) {
480 if (!(bp = xlog_get_bp(log, 1)))
481 return ENOMEM;
482 smallmem = 1;
483 } else {
484 error = xlog_bread(log, start_blk, num_blks, bp, &offset);
485 if (error)
486 goto out;
487 offset += ((num_blks - 1) << BBSHIFT);
490 for (i = (*last_blk) - 1; i >= 0; i--) {
491 if (i < start_blk) {
492 /* valid log record not found */
493 xlog_warn(
494 "XFS: Log inconsistent (didn't find previous header)");
495 ASSERT(0);
496 error = XFS_ERROR(EIO);
497 goto out;
500 if (smallmem) {
501 error = xlog_bread(log, i, 1, bp, &offset);
502 if (error)
503 goto out;
506 head = (xlog_rec_header_t *)offset;
508 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(head->h_magicno))
509 break;
511 if (!smallmem)
512 offset -= BBSIZE;
516 * We hit the beginning of the physical log & still no header. Return
517 * to caller. If caller can handle a return of -1, then this routine
518 * will be called again for the end of the physical log.
520 if (i == -1) {
521 error = -1;
522 goto out;
526 * We have the final block of the good log (the first block
527 * of the log record _before_ the head. So we check the uuid.
529 if ((error = xlog_header_check_mount(log->l_mp, head)))
530 goto out;
533 * We may have found a log record header before we expected one.
534 * last_blk will be the 1st block # with a given cycle #. We may end
535 * up reading an entire log record. In this case, we don't want to
536 * reset last_blk. Only when last_blk points in the middle of a log
537 * record do we update last_blk.
539 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
540 uint h_size = be32_to_cpu(head->h_size);
542 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
543 if (h_size % XLOG_HEADER_CYCLE_SIZE)
544 xhdrs++;
545 } else {
546 xhdrs = 1;
549 if (*last_blk - i + extra_bblks !=
550 BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
551 *last_blk = i;
553 out:
554 xlog_put_bp(bp);
555 return error;
559 * Head is defined to be the point of the log where the next log write
560 * write could go. This means that incomplete LR writes at the end are
561 * eliminated when calculating the head. We aren't guaranteed that previous
562 * LR have complete transactions. We only know that a cycle number of
563 * current cycle number -1 won't be present in the log if we start writing
564 * from our current block number.
566 * last_blk contains the block number of the first block with a given
567 * cycle number.
569 * Return: zero if normal, non-zero if error.
571 STATIC int
572 xlog_find_head(
573 xlog_t *log,
574 xfs_daddr_t *return_head_blk)
576 xfs_buf_t *bp;
577 xfs_caddr_t offset;
578 xfs_daddr_t new_blk, first_blk, start_blk, last_blk, head_blk;
579 int num_scan_bblks;
580 uint first_half_cycle, last_half_cycle;
581 uint stop_on_cycle;
582 int error, log_bbnum = log->l_logBBsize;
584 /* Is the end of the log device zeroed? */
585 if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
586 *return_head_blk = first_blk;
588 /* Is the whole lot zeroed? */
589 if (!first_blk) {
590 /* Linux XFS shouldn't generate totally zeroed logs -
591 * mkfs etc write a dummy unmount record to a fresh
592 * log so we can store the uuid in there
594 xlog_warn("XFS: totally zeroed log");
597 return 0;
598 } else if (error) {
599 xlog_warn("XFS: empty log check failed");
600 return error;
603 first_blk = 0; /* get cycle # of 1st block */
604 bp = xlog_get_bp(log, 1);
605 if (!bp)
606 return ENOMEM;
608 error = xlog_bread(log, 0, 1, bp, &offset);
609 if (error)
610 goto bp_err;
612 first_half_cycle = xlog_get_cycle(offset);
614 last_blk = head_blk = log_bbnum - 1; /* get cycle # of last block */
615 error = xlog_bread(log, last_blk, 1, bp, &offset);
616 if (error)
617 goto bp_err;
619 last_half_cycle = xlog_get_cycle(offset);
620 ASSERT(last_half_cycle != 0);
623 * If the 1st half cycle number is equal to the last half cycle number,
624 * then the entire log is stamped with the same cycle number. In this
625 * case, head_blk can't be set to zero (which makes sense). The below
626 * math doesn't work out properly with head_blk equal to zero. Instead,
627 * we set it to log_bbnum which is an invalid block number, but this
628 * value makes the math correct. If head_blk doesn't changed through
629 * all the tests below, *head_blk is set to zero at the very end rather
630 * than log_bbnum. In a sense, log_bbnum and zero are the same block
631 * in a circular file.
633 if (first_half_cycle == last_half_cycle) {
635 * In this case we believe that the entire log should have
636 * cycle number last_half_cycle. We need to scan backwards
637 * from the end verifying that there are no holes still
638 * containing last_half_cycle - 1. If we find such a hole,
639 * then the start of that hole will be the new head. The
640 * simple case looks like
641 * x | x ... | x - 1 | x
642 * Another case that fits this picture would be
643 * x | x + 1 | x ... | x
644 * In this case the head really is somewhere at the end of the
645 * log, as one of the latest writes at the beginning was
646 * incomplete.
647 * One more case is
648 * x | x + 1 | x ... | x - 1 | x
649 * This is really the combination of the above two cases, and
650 * the head has to end up at the start of the x-1 hole at the
651 * end of the log.
653 * In the 256k log case, we will read from the beginning to the
654 * end of the log and search for cycle numbers equal to x-1.
655 * We don't worry about the x+1 blocks that we encounter,
656 * because we know that they cannot be the head since the log
657 * started with x.
659 head_blk = log_bbnum;
660 stop_on_cycle = last_half_cycle - 1;
661 } else {
663 * In this case we want to find the first block with cycle
664 * number matching last_half_cycle. We expect the log to be
665 * some variation on
666 * x + 1 ... | x ... | x
667 * The first block with cycle number x (last_half_cycle) will
668 * be where the new head belongs. First we do a binary search
669 * for the first occurrence of last_half_cycle. The binary
670 * search may not be totally accurate, so then we scan back
671 * from there looking for occurrences of last_half_cycle before
672 * us. If that backwards scan wraps around the beginning of
673 * the log, then we look for occurrences of last_half_cycle - 1
674 * at the end of the log. The cases we're looking for look
675 * like
676 * v binary search stopped here
677 * x + 1 ... | x | x + 1 | x ... | x
678 * ^ but we want to locate this spot
679 * or
680 * <---------> less than scan distance
681 * x + 1 ... | x ... | x - 1 | x
682 * ^ we want to locate this spot
684 stop_on_cycle = last_half_cycle;
685 if ((error = xlog_find_cycle_start(log, bp, first_blk,
686 &head_blk, last_half_cycle)))
687 goto bp_err;
691 * Now validate the answer. Scan back some number of maximum possible
692 * blocks and make sure each one has the expected cycle number. The
693 * maximum is determined by the total possible amount of buffering
694 * in the in-core log. The following number can be made tighter if
695 * we actually look at the block size of the filesystem.
697 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
698 if (head_blk >= num_scan_bblks) {
700 * We are guaranteed that the entire check can be performed
701 * in one buffer.
703 start_blk = head_blk - num_scan_bblks;
704 if ((error = xlog_find_verify_cycle(log,
705 start_blk, num_scan_bblks,
706 stop_on_cycle, &new_blk)))
707 goto bp_err;
708 if (new_blk != -1)
709 head_blk = new_blk;
710 } else { /* need to read 2 parts of log */
712 * We are going to scan backwards in the log in two parts.
713 * First we scan the physical end of the log. In this part
714 * of the log, we are looking for blocks with cycle number
715 * last_half_cycle - 1.
716 * If we find one, then we know that the log starts there, as
717 * we've found a hole that didn't get written in going around
718 * the end of the physical log. The simple case for this is
719 * x + 1 ... | x ... | x - 1 | x
720 * <---------> less than scan distance
721 * If all of the blocks at the end of the log have cycle number
722 * last_half_cycle, then we check the blocks at the start of
723 * the log looking for occurrences of last_half_cycle. If we
724 * find one, then our current estimate for the location of the
725 * first occurrence of last_half_cycle is wrong and we move
726 * back to the hole we've found. This case looks like
727 * x + 1 ... | x | x + 1 | x ...
728 * ^ binary search stopped here
729 * Another case we need to handle that only occurs in 256k
730 * logs is
731 * x + 1 ... | x ... | x+1 | x ...
732 * ^ binary search stops here
733 * In a 256k log, the scan at the end of the log will see the
734 * x + 1 blocks. We need to skip past those since that is
735 * certainly not the head of the log. By searching for
736 * last_half_cycle-1 we accomplish that.
738 ASSERT(head_blk <= INT_MAX &&
739 (xfs_daddr_t) num_scan_bblks >= head_blk);
740 start_blk = log_bbnum - (num_scan_bblks - head_blk);
741 if ((error = xlog_find_verify_cycle(log, start_blk,
742 num_scan_bblks - (int)head_blk,
743 (stop_on_cycle - 1), &new_blk)))
744 goto bp_err;
745 if (new_blk != -1) {
746 head_blk = new_blk;
747 goto validate_head;
751 * Scan beginning of log now. The last part of the physical
752 * log is good. This scan needs to verify that it doesn't find
753 * the last_half_cycle.
755 start_blk = 0;
756 ASSERT(head_blk <= INT_MAX);
757 if ((error = xlog_find_verify_cycle(log,
758 start_blk, (int)head_blk,
759 stop_on_cycle, &new_blk)))
760 goto bp_err;
761 if (new_blk != -1)
762 head_blk = new_blk;
765 validate_head:
767 * Now we need to make sure head_blk is not pointing to a block in
768 * the middle of a log record.
770 num_scan_bblks = XLOG_REC_SHIFT(log);
771 if (head_blk >= num_scan_bblks) {
772 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
774 /* start ptr at last block ptr before head_blk */
775 if ((error = xlog_find_verify_log_record(log, start_blk,
776 &head_blk, 0)) == -1) {
777 error = XFS_ERROR(EIO);
778 goto bp_err;
779 } else if (error)
780 goto bp_err;
781 } else {
782 start_blk = 0;
783 ASSERT(head_blk <= INT_MAX);
784 if ((error = xlog_find_verify_log_record(log, start_blk,
785 &head_blk, 0)) == -1) {
786 /* We hit the beginning of the log during our search */
787 start_blk = log_bbnum - (num_scan_bblks - head_blk);
788 new_blk = log_bbnum;
789 ASSERT(start_blk <= INT_MAX &&
790 (xfs_daddr_t) log_bbnum-start_blk >= 0);
791 ASSERT(head_blk <= INT_MAX);
792 if ((error = xlog_find_verify_log_record(log,
793 start_blk, &new_blk,
794 (int)head_blk)) == -1) {
795 error = XFS_ERROR(EIO);
796 goto bp_err;
797 } else if (error)
798 goto bp_err;
799 if (new_blk != log_bbnum)
800 head_blk = new_blk;
801 } else if (error)
802 goto bp_err;
805 xlog_put_bp(bp);
806 if (head_blk == log_bbnum)
807 *return_head_blk = 0;
808 else
809 *return_head_blk = head_blk;
811 * When returning here, we have a good block number. Bad block
812 * means that during a previous crash, we didn't have a clean break
813 * from cycle number N to cycle number N-1. In this case, we need
814 * to find the first block with cycle number N-1.
816 return 0;
818 bp_err:
819 xlog_put_bp(bp);
821 if (error)
822 xlog_warn("XFS: failed to find log head");
823 return error;
827 * Find the sync block number or the tail of the log.
829 * This will be the block number of the last record to have its
830 * associated buffers synced to disk. Every log record header has
831 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
832 * to get a sync block number. The only concern is to figure out which
833 * log record header to believe.
835 * The following algorithm uses the log record header with the largest
836 * lsn. The entire log record does not need to be valid. We only care
837 * that the header is valid.
839 * We could speed up search by using current head_blk buffer, but it is not
840 * available.
842 STATIC int
843 xlog_find_tail(
844 xlog_t *log,
845 xfs_daddr_t *head_blk,
846 xfs_daddr_t *tail_blk)
848 xlog_rec_header_t *rhead;
849 xlog_op_header_t *op_head;
850 xfs_caddr_t offset = NULL;
851 xfs_buf_t *bp;
852 int error, i, found;
853 xfs_daddr_t umount_data_blk;
854 xfs_daddr_t after_umount_blk;
855 xfs_lsn_t tail_lsn;
856 int hblks;
858 found = 0;
861 * Find previous log record
863 if ((error = xlog_find_head(log, head_blk)))
864 return error;
866 bp = xlog_get_bp(log, 1);
867 if (!bp)
868 return ENOMEM;
869 if (*head_blk == 0) { /* special case */
870 error = xlog_bread(log, 0, 1, bp, &offset);
871 if (error)
872 goto done;
874 if (xlog_get_cycle(offset) == 0) {
875 *tail_blk = 0;
876 /* leave all other log inited values alone */
877 goto done;
882 * Search backwards looking for log record header block
884 ASSERT(*head_blk < INT_MAX);
885 for (i = (int)(*head_blk) - 1; i >= 0; i--) {
886 error = xlog_bread(log, i, 1, bp, &offset);
887 if (error)
888 goto done;
890 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(*(__be32 *)offset)) {
891 found = 1;
892 break;
896 * If we haven't found the log record header block, start looking
897 * again from the end of the physical log. XXXmiken: There should be
898 * a check here to make sure we didn't search more than N blocks in
899 * the previous code.
901 if (!found) {
902 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
903 error = xlog_bread(log, i, 1, bp, &offset);
904 if (error)
905 goto done;
907 if (XLOG_HEADER_MAGIC_NUM ==
908 be32_to_cpu(*(__be32 *)offset)) {
909 found = 2;
910 break;
914 if (!found) {
915 xlog_warn("XFS: xlog_find_tail: couldn't find sync record");
916 ASSERT(0);
917 return XFS_ERROR(EIO);
920 /* find blk_no of tail of log */
921 rhead = (xlog_rec_header_t *)offset;
922 *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
925 * Reset log values according to the state of the log when we
926 * crashed. In the case where head_blk == 0, we bump curr_cycle
927 * one because the next write starts a new cycle rather than
928 * continuing the cycle of the last good log record. At this
929 * point we have guaranteed that all partial log records have been
930 * accounted for. Therefore, we know that the last good log record
931 * written was complete and ended exactly on the end boundary
932 * of the physical log.
934 log->l_prev_block = i;
935 log->l_curr_block = (int)*head_blk;
936 log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
937 if (found == 2)
938 log->l_curr_cycle++;
939 atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
940 atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
941 xlog_assign_grant_head(&log->l_grant_reserve_head, log->l_curr_cycle,
942 BBTOB(log->l_curr_block));
943 xlog_assign_grant_head(&log->l_grant_write_head, log->l_curr_cycle,
944 BBTOB(log->l_curr_block));
947 * Look for unmount record. If we find it, then we know there
948 * was a clean unmount. Since 'i' could be the last block in
949 * the physical log, we convert to a log block before comparing
950 * to the head_blk.
952 * Save the current tail lsn to use to pass to
953 * xlog_clear_stale_blocks() below. We won't want to clear the
954 * unmount record if there is one, so we pass the lsn of the
955 * unmount record rather than the block after it.
957 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
958 int h_size = be32_to_cpu(rhead->h_size);
959 int h_version = be32_to_cpu(rhead->h_version);
961 if ((h_version & XLOG_VERSION_2) &&
962 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
963 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
964 if (h_size % XLOG_HEADER_CYCLE_SIZE)
965 hblks++;
966 } else {
967 hblks = 1;
969 } else {
970 hblks = 1;
972 after_umount_blk = (i + hblks + (int)
973 BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
974 tail_lsn = atomic64_read(&log->l_tail_lsn);
975 if (*head_blk == after_umount_blk &&
976 be32_to_cpu(rhead->h_num_logops) == 1) {
977 umount_data_blk = (i + hblks) % log->l_logBBsize;
978 error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
979 if (error)
980 goto done;
982 op_head = (xlog_op_header_t *)offset;
983 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
985 * Set tail and last sync so that newly written
986 * log records will point recovery to after the
987 * current unmount record.
989 xlog_assign_atomic_lsn(&log->l_tail_lsn,
990 log->l_curr_cycle, after_umount_blk);
991 xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
992 log->l_curr_cycle, after_umount_blk);
993 *tail_blk = after_umount_blk;
996 * Note that the unmount was clean. If the unmount
997 * was not clean, we need to know this to rebuild the
998 * superblock counters from the perag headers if we
999 * have a filesystem using non-persistent counters.
1001 log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1006 * Make sure that there are no blocks in front of the head
1007 * with the same cycle number as the head. This can happen
1008 * because we allow multiple outstanding log writes concurrently,
1009 * and the later writes might make it out before earlier ones.
1011 * We use the lsn from before modifying it so that we'll never
1012 * overwrite the unmount record after a clean unmount.
1014 * Do this only if we are going to recover the filesystem
1016 * NOTE: This used to say "if (!readonly)"
1017 * However on Linux, we can & do recover a read-only filesystem.
1018 * We only skip recovery if NORECOVERY is specified on mount,
1019 * in which case we would not be here.
1021 * But... if the -device- itself is readonly, just skip this.
1022 * We can't recover this device anyway, so it won't matter.
1024 if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp))
1025 error = xlog_clear_stale_blocks(log, tail_lsn);
1027 done:
1028 xlog_put_bp(bp);
1030 if (error)
1031 xlog_warn("XFS: failed to locate log tail");
1032 return error;
1036 * Is the log zeroed at all?
1038 * The last binary search should be changed to perform an X block read
1039 * once X becomes small enough. You can then search linearly through
1040 * the X blocks. This will cut down on the number of reads we need to do.
1042 * If the log is partially zeroed, this routine will pass back the blkno
1043 * of the first block with cycle number 0. It won't have a complete LR
1044 * preceding it.
1046 * Return:
1047 * 0 => the log is completely written to
1048 * -1 => use *blk_no as the first block of the log
1049 * >0 => error has occurred
1051 STATIC int
1052 xlog_find_zeroed(
1053 xlog_t *log,
1054 xfs_daddr_t *blk_no)
1056 xfs_buf_t *bp;
1057 xfs_caddr_t offset;
1058 uint first_cycle, last_cycle;
1059 xfs_daddr_t new_blk, last_blk, start_blk;
1060 xfs_daddr_t num_scan_bblks;
1061 int error, log_bbnum = log->l_logBBsize;
1063 *blk_no = 0;
1065 /* check totally zeroed log */
1066 bp = xlog_get_bp(log, 1);
1067 if (!bp)
1068 return ENOMEM;
1069 error = xlog_bread(log, 0, 1, bp, &offset);
1070 if (error)
1071 goto bp_err;
1073 first_cycle = xlog_get_cycle(offset);
1074 if (first_cycle == 0) { /* completely zeroed log */
1075 *blk_no = 0;
1076 xlog_put_bp(bp);
1077 return -1;
1080 /* check partially zeroed log */
1081 error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1082 if (error)
1083 goto bp_err;
1085 last_cycle = xlog_get_cycle(offset);
1086 if (last_cycle != 0) { /* log completely written to */
1087 xlog_put_bp(bp);
1088 return 0;
1089 } else if (first_cycle != 1) {
1091 * If the cycle of the last block is zero, the cycle of
1092 * the first block must be 1. If it's not, maybe we're
1093 * not looking at a log... Bail out.
1095 xlog_warn("XFS: Log inconsistent or not a log (last==0, first!=1)");
1096 return XFS_ERROR(EINVAL);
1099 /* we have a partially zeroed log */
1100 last_blk = log_bbnum-1;
1101 if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1102 goto bp_err;
1105 * Validate the answer. Because there is no way to guarantee that
1106 * the entire log is made up of log records which are the same size,
1107 * we scan over the defined maximum blocks. At this point, the maximum
1108 * is not chosen to mean anything special. XXXmiken
1110 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1111 ASSERT(num_scan_bblks <= INT_MAX);
1113 if (last_blk < num_scan_bblks)
1114 num_scan_bblks = last_blk;
1115 start_blk = last_blk - num_scan_bblks;
1118 * We search for any instances of cycle number 0 that occur before
1119 * our current estimate of the head. What we're trying to detect is
1120 * 1 ... | 0 | 1 | 0...
1121 * ^ binary search ends here
1123 if ((error = xlog_find_verify_cycle(log, start_blk,
1124 (int)num_scan_bblks, 0, &new_blk)))
1125 goto bp_err;
1126 if (new_blk != -1)
1127 last_blk = new_blk;
1130 * Potentially backup over partial log record write. We don't need
1131 * to search the end of the log because we know it is zero.
1133 if ((error = xlog_find_verify_log_record(log, start_blk,
1134 &last_blk, 0)) == -1) {
1135 error = XFS_ERROR(EIO);
1136 goto bp_err;
1137 } else if (error)
1138 goto bp_err;
1140 *blk_no = last_blk;
1141 bp_err:
1142 xlog_put_bp(bp);
1143 if (error)
1144 return error;
1145 return -1;
1149 * These are simple subroutines used by xlog_clear_stale_blocks() below
1150 * to initialize a buffer full of empty log record headers and write
1151 * them into the log.
1153 STATIC void
1154 xlog_add_record(
1155 xlog_t *log,
1156 xfs_caddr_t buf,
1157 int cycle,
1158 int block,
1159 int tail_cycle,
1160 int tail_block)
1162 xlog_rec_header_t *recp = (xlog_rec_header_t *)buf;
1164 memset(buf, 0, BBSIZE);
1165 recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1166 recp->h_cycle = cpu_to_be32(cycle);
1167 recp->h_version = cpu_to_be32(
1168 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1169 recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1170 recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1171 recp->h_fmt = cpu_to_be32(XLOG_FMT);
1172 memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1175 STATIC int
1176 xlog_write_log_records(
1177 xlog_t *log,
1178 int cycle,
1179 int start_block,
1180 int blocks,
1181 int tail_cycle,
1182 int tail_block)
1184 xfs_caddr_t offset;
1185 xfs_buf_t *bp;
1186 int balign, ealign;
1187 int sectbb = log->l_sectBBsize;
1188 int end_block = start_block + blocks;
1189 int bufblks;
1190 int error = 0;
1191 int i, j = 0;
1194 * Greedily allocate a buffer big enough to handle the full
1195 * range of basic blocks to be written. If that fails, try
1196 * a smaller size. We need to be able to write at least a
1197 * log sector, or we're out of luck.
1199 bufblks = 1 << ffs(blocks);
1200 while (!(bp = xlog_get_bp(log, bufblks))) {
1201 bufblks >>= 1;
1202 if (bufblks < sectbb)
1203 return ENOMEM;
1206 /* We may need to do a read at the start to fill in part of
1207 * the buffer in the starting sector not covered by the first
1208 * write below.
1210 balign = round_down(start_block, sectbb);
1211 if (balign != start_block) {
1212 error = xlog_bread_noalign(log, start_block, 1, bp);
1213 if (error)
1214 goto out_put_bp;
1216 j = start_block - balign;
1219 for (i = start_block; i < end_block; i += bufblks) {
1220 int bcount, endcount;
1222 bcount = min(bufblks, end_block - start_block);
1223 endcount = bcount - j;
1225 /* We may need to do a read at the end to fill in part of
1226 * the buffer in the final sector not covered by the write.
1227 * If this is the same sector as the above read, skip it.
1229 ealign = round_down(end_block, sectbb);
1230 if (j == 0 && (start_block + endcount > ealign)) {
1231 offset = XFS_BUF_PTR(bp);
1232 balign = BBTOB(ealign - start_block);
1233 error = XFS_BUF_SET_PTR(bp, offset + balign,
1234 BBTOB(sectbb));
1235 if (error)
1236 break;
1238 error = xlog_bread_noalign(log, ealign, sectbb, bp);
1239 if (error)
1240 break;
1242 error = XFS_BUF_SET_PTR(bp, offset, bufblks);
1243 if (error)
1244 break;
1247 offset = xlog_align(log, start_block, endcount, bp);
1248 for (; j < endcount; j++) {
1249 xlog_add_record(log, offset, cycle, i+j,
1250 tail_cycle, tail_block);
1251 offset += BBSIZE;
1253 error = xlog_bwrite(log, start_block, endcount, bp);
1254 if (error)
1255 break;
1256 start_block += endcount;
1257 j = 0;
1260 out_put_bp:
1261 xlog_put_bp(bp);
1262 return error;
1266 * This routine is called to blow away any incomplete log writes out
1267 * in front of the log head. We do this so that we won't become confused
1268 * if we come up, write only a little bit more, and then crash again.
1269 * If we leave the partial log records out there, this situation could
1270 * cause us to think those partial writes are valid blocks since they
1271 * have the current cycle number. We get rid of them by overwriting them
1272 * with empty log records with the old cycle number rather than the
1273 * current one.
1275 * The tail lsn is passed in rather than taken from
1276 * the log so that we will not write over the unmount record after a
1277 * clean unmount in a 512 block log. Doing so would leave the log without
1278 * any valid log records in it until a new one was written. If we crashed
1279 * during that time we would not be able to recover.
1281 STATIC int
1282 xlog_clear_stale_blocks(
1283 xlog_t *log,
1284 xfs_lsn_t tail_lsn)
1286 int tail_cycle, head_cycle;
1287 int tail_block, head_block;
1288 int tail_distance, max_distance;
1289 int distance;
1290 int error;
1292 tail_cycle = CYCLE_LSN(tail_lsn);
1293 tail_block = BLOCK_LSN(tail_lsn);
1294 head_cycle = log->l_curr_cycle;
1295 head_block = log->l_curr_block;
1298 * Figure out the distance between the new head of the log
1299 * and the tail. We want to write over any blocks beyond the
1300 * head that we may have written just before the crash, but
1301 * we don't want to overwrite the tail of the log.
1303 if (head_cycle == tail_cycle) {
1305 * The tail is behind the head in the physical log,
1306 * so the distance from the head to the tail is the
1307 * distance from the head to the end of the log plus
1308 * the distance from the beginning of the log to the
1309 * tail.
1311 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1312 XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1313 XFS_ERRLEVEL_LOW, log->l_mp);
1314 return XFS_ERROR(EFSCORRUPTED);
1316 tail_distance = tail_block + (log->l_logBBsize - head_block);
1317 } else {
1319 * The head is behind the tail in the physical log,
1320 * so the distance from the head to the tail is just
1321 * the tail block minus the head block.
1323 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1324 XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1325 XFS_ERRLEVEL_LOW, log->l_mp);
1326 return XFS_ERROR(EFSCORRUPTED);
1328 tail_distance = tail_block - head_block;
1332 * If the head is right up against the tail, we can't clear
1333 * anything.
1335 if (tail_distance <= 0) {
1336 ASSERT(tail_distance == 0);
1337 return 0;
1340 max_distance = XLOG_TOTAL_REC_SHIFT(log);
1342 * Take the smaller of the maximum amount of outstanding I/O
1343 * we could have and the distance to the tail to clear out.
1344 * We take the smaller so that we don't overwrite the tail and
1345 * we don't waste all day writing from the head to the tail
1346 * for no reason.
1348 max_distance = MIN(max_distance, tail_distance);
1350 if ((head_block + max_distance) <= log->l_logBBsize) {
1352 * We can stomp all the blocks we need to without
1353 * wrapping around the end of the log. Just do it
1354 * in a single write. Use the cycle number of the
1355 * current cycle minus one so that the log will look like:
1356 * n ... | n - 1 ...
1358 error = xlog_write_log_records(log, (head_cycle - 1),
1359 head_block, max_distance, tail_cycle,
1360 tail_block);
1361 if (error)
1362 return error;
1363 } else {
1365 * We need to wrap around the end of the physical log in
1366 * order to clear all the blocks. Do it in two separate
1367 * I/Os. The first write should be from the head to the
1368 * end of the physical log, and it should use the current
1369 * cycle number minus one just like above.
1371 distance = log->l_logBBsize - head_block;
1372 error = xlog_write_log_records(log, (head_cycle - 1),
1373 head_block, distance, tail_cycle,
1374 tail_block);
1376 if (error)
1377 return error;
1380 * Now write the blocks at the start of the physical log.
1381 * This writes the remainder of the blocks we want to clear.
1382 * It uses the current cycle number since we're now on the
1383 * same cycle as the head so that we get:
1384 * n ... n ... | n - 1 ...
1385 * ^^^^^ blocks we're writing
1387 distance = max_distance - (log->l_logBBsize - head_block);
1388 error = xlog_write_log_records(log, head_cycle, 0, distance,
1389 tail_cycle, tail_block);
1390 if (error)
1391 return error;
1394 return 0;
1397 /******************************************************************************
1399 * Log recover routines
1401 ******************************************************************************
1404 STATIC xlog_recover_t *
1405 xlog_recover_find_tid(
1406 struct hlist_head *head,
1407 xlog_tid_t tid)
1409 xlog_recover_t *trans;
1410 struct hlist_node *n;
1412 hlist_for_each_entry(trans, n, head, r_list) {
1413 if (trans->r_log_tid == tid)
1414 return trans;
1416 return NULL;
1419 STATIC void
1420 xlog_recover_new_tid(
1421 struct hlist_head *head,
1422 xlog_tid_t tid,
1423 xfs_lsn_t lsn)
1425 xlog_recover_t *trans;
1427 trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1428 trans->r_log_tid = tid;
1429 trans->r_lsn = lsn;
1430 INIT_LIST_HEAD(&trans->r_itemq);
1432 INIT_HLIST_NODE(&trans->r_list);
1433 hlist_add_head(&trans->r_list, head);
1436 STATIC void
1437 xlog_recover_add_item(
1438 struct list_head *head)
1440 xlog_recover_item_t *item;
1442 item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1443 INIT_LIST_HEAD(&item->ri_list);
1444 list_add_tail(&item->ri_list, head);
1447 STATIC int
1448 xlog_recover_add_to_cont_trans(
1449 struct log *log,
1450 xlog_recover_t *trans,
1451 xfs_caddr_t dp,
1452 int len)
1454 xlog_recover_item_t *item;
1455 xfs_caddr_t ptr, old_ptr;
1456 int old_len;
1458 if (list_empty(&trans->r_itemq)) {
1459 /* finish copying rest of trans header */
1460 xlog_recover_add_item(&trans->r_itemq);
1461 ptr = (xfs_caddr_t) &trans->r_theader +
1462 sizeof(xfs_trans_header_t) - len;
1463 memcpy(ptr, dp, len); /* d, s, l */
1464 return 0;
1466 /* take the tail entry */
1467 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1469 old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1470 old_len = item->ri_buf[item->ri_cnt-1].i_len;
1472 ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1473 memcpy(&ptr[old_len], dp, len); /* d, s, l */
1474 item->ri_buf[item->ri_cnt-1].i_len += len;
1475 item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1476 trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
1477 return 0;
1481 * The next region to add is the start of a new region. It could be
1482 * a whole region or it could be the first part of a new region. Because
1483 * of this, the assumption here is that the type and size fields of all
1484 * format structures fit into the first 32 bits of the structure.
1486 * This works because all regions must be 32 bit aligned. Therefore, we
1487 * either have both fields or we have neither field. In the case we have
1488 * neither field, the data part of the region is zero length. We only have
1489 * a log_op_header and can throw away the header since a new one will appear
1490 * later. If we have at least 4 bytes, then we can determine how many regions
1491 * will appear in the current log item.
1493 STATIC int
1494 xlog_recover_add_to_trans(
1495 struct log *log,
1496 xlog_recover_t *trans,
1497 xfs_caddr_t dp,
1498 int len)
1500 xfs_inode_log_format_t *in_f; /* any will do */
1501 xlog_recover_item_t *item;
1502 xfs_caddr_t ptr;
1504 if (!len)
1505 return 0;
1506 if (list_empty(&trans->r_itemq)) {
1507 /* we need to catch log corruptions here */
1508 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1509 xlog_warn("XFS: xlog_recover_add_to_trans: "
1510 "bad header magic number");
1511 ASSERT(0);
1512 return XFS_ERROR(EIO);
1514 if (len == sizeof(xfs_trans_header_t))
1515 xlog_recover_add_item(&trans->r_itemq);
1516 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1517 return 0;
1520 ptr = kmem_alloc(len, KM_SLEEP);
1521 memcpy(ptr, dp, len);
1522 in_f = (xfs_inode_log_format_t *)ptr;
1524 /* take the tail entry */
1525 item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1526 if (item->ri_total != 0 &&
1527 item->ri_total == item->ri_cnt) {
1528 /* tail item is in use, get a new one */
1529 xlog_recover_add_item(&trans->r_itemq);
1530 item = list_entry(trans->r_itemq.prev,
1531 xlog_recover_item_t, ri_list);
1534 if (item->ri_total == 0) { /* first region to be added */
1535 if (in_f->ilf_size == 0 ||
1536 in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
1537 xlog_warn(
1538 "XFS: bad number of regions (%d) in inode log format",
1539 in_f->ilf_size);
1540 ASSERT(0);
1541 return XFS_ERROR(EIO);
1544 item->ri_total = in_f->ilf_size;
1545 item->ri_buf =
1546 kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
1547 KM_SLEEP);
1549 ASSERT(item->ri_total > item->ri_cnt);
1550 /* Description region is ri_buf[0] */
1551 item->ri_buf[item->ri_cnt].i_addr = ptr;
1552 item->ri_buf[item->ri_cnt].i_len = len;
1553 item->ri_cnt++;
1554 trace_xfs_log_recover_item_add(log, trans, item, 0);
1555 return 0;
1559 * Sort the log items in the transaction. Cancelled buffers need
1560 * to be put first so they are processed before any items that might
1561 * modify the buffers. If they are cancelled, then the modifications
1562 * don't need to be replayed.
1564 STATIC int
1565 xlog_recover_reorder_trans(
1566 struct log *log,
1567 xlog_recover_t *trans,
1568 int pass)
1570 xlog_recover_item_t *item, *n;
1571 LIST_HEAD(sort_list);
1573 list_splice_init(&trans->r_itemq, &sort_list);
1574 list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1575 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1577 switch (ITEM_TYPE(item)) {
1578 case XFS_LI_BUF:
1579 if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1580 trace_xfs_log_recover_item_reorder_head(log,
1581 trans, item, pass);
1582 list_move(&item->ri_list, &trans->r_itemq);
1583 break;
1585 case XFS_LI_INODE:
1586 case XFS_LI_DQUOT:
1587 case XFS_LI_QUOTAOFF:
1588 case XFS_LI_EFD:
1589 case XFS_LI_EFI:
1590 trace_xfs_log_recover_item_reorder_tail(log,
1591 trans, item, pass);
1592 list_move_tail(&item->ri_list, &trans->r_itemq);
1593 break;
1594 default:
1595 xlog_warn(
1596 "XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1597 ASSERT(0);
1598 return XFS_ERROR(EIO);
1601 ASSERT(list_empty(&sort_list));
1602 return 0;
1606 * Build up the table of buf cancel records so that we don't replay
1607 * cancelled data in the second pass. For buffer records that are
1608 * not cancel records, there is nothing to do here so we just return.
1610 * If we get a cancel record which is already in the table, this indicates
1611 * that the buffer was cancelled multiple times. In order to ensure
1612 * that during pass 2 we keep the record in the table until we reach its
1613 * last occurrence in the log, we keep a reference count in the cancel
1614 * record in the table to tell us how many times we expect to see this
1615 * record during the second pass.
1617 STATIC int
1618 xlog_recover_buffer_pass1(
1619 struct log *log,
1620 xlog_recover_item_t *item)
1622 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
1623 struct list_head *bucket;
1624 struct xfs_buf_cancel *bcp;
1627 * If this isn't a cancel buffer item, then just return.
1629 if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
1630 trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1631 return 0;
1635 * Insert an xfs_buf_cancel record into the hash table of them.
1636 * If there is already an identical record, bump its reference count.
1638 bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
1639 list_for_each_entry(bcp, bucket, bc_list) {
1640 if (bcp->bc_blkno == buf_f->blf_blkno &&
1641 bcp->bc_len == buf_f->blf_len) {
1642 bcp->bc_refcount++;
1643 trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
1644 return 0;
1648 bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), KM_SLEEP);
1649 bcp->bc_blkno = buf_f->blf_blkno;
1650 bcp->bc_len = buf_f->blf_len;
1651 bcp->bc_refcount = 1;
1652 list_add_tail(&bcp->bc_list, bucket);
1654 trace_xfs_log_recover_buf_cancel_add(log, buf_f);
1655 return 0;
1659 * Check to see whether the buffer being recovered has a corresponding
1660 * entry in the buffer cancel record table. If it does then return 1
1661 * so that it will be cancelled, otherwise return 0. If the buffer is
1662 * actually a buffer cancel item (XFS_BLF_CANCEL is set), then decrement
1663 * the refcount on the entry in the table and remove it from the table
1664 * if this is the last reference.
1666 * We remove the cancel record from the table when we encounter its
1667 * last occurrence in the log so that if the same buffer is re-used
1668 * again after its last cancellation we actually replay the changes
1669 * made at that point.
1671 STATIC int
1672 xlog_check_buffer_cancelled(
1673 struct log *log,
1674 xfs_daddr_t blkno,
1675 uint len,
1676 ushort flags)
1678 struct list_head *bucket;
1679 struct xfs_buf_cancel *bcp;
1681 if (log->l_buf_cancel_table == NULL) {
1683 * There is nothing in the table built in pass one,
1684 * so this buffer must not be cancelled.
1686 ASSERT(!(flags & XFS_BLF_CANCEL));
1687 return 0;
1691 * Search for an entry in the cancel table that matches our buffer.
1693 bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
1694 list_for_each_entry(bcp, bucket, bc_list) {
1695 if (bcp->bc_blkno == blkno && bcp->bc_len == len)
1696 goto found;
1700 * We didn't find a corresponding entry in the table, so return 0 so
1701 * that the buffer is NOT cancelled.
1703 ASSERT(!(flags & XFS_BLF_CANCEL));
1704 return 0;
1706 found:
1708 * We've go a match, so return 1 so that the recovery of this buffer
1709 * is cancelled. If this buffer is actually a buffer cancel log
1710 * item, then decrement the refcount on the one in the table and
1711 * remove it if this is the last reference.
1713 if (flags & XFS_BLF_CANCEL) {
1714 if (--bcp->bc_refcount == 0) {
1715 list_del(&bcp->bc_list);
1716 kmem_free(bcp);
1719 return 1;
1723 * Perform recovery for a buffer full of inodes. In these buffers, the only
1724 * data which should be recovered is that which corresponds to the
1725 * di_next_unlinked pointers in the on disk inode structures. The rest of the
1726 * data for the inodes is always logged through the inodes themselves rather
1727 * than the inode buffer and is recovered in xlog_recover_inode_pass2().
1729 * The only time when buffers full of inodes are fully recovered is when the
1730 * buffer is full of newly allocated inodes. In this case the buffer will
1731 * not be marked as an inode buffer and so will be sent to
1732 * xlog_recover_do_reg_buffer() below during recovery.
1734 STATIC int
1735 xlog_recover_do_inode_buffer(
1736 struct xfs_mount *mp,
1737 xlog_recover_item_t *item,
1738 struct xfs_buf *bp,
1739 xfs_buf_log_format_t *buf_f)
1741 int i;
1742 int item_index = 0;
1743 int bit = 0;
1744 int nbits = 0;
1745 int reg_buf_offset = 0;
1746 int reg_buf_bytes = 0;
1747 int next_unlinked_offset;
1748 int inodes_per_buf;
1749 xfs_agino_t *logged_nextp;
1750 xfs_agino_t *buffer_nextp;
1752 trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
1754 inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1755 for (i = 0; i < inodes_per_buf; i++) {
1756 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1757 offsetof(xfs_dinode_t, di_next_unlinked);
1759 while (next_unlinked_offset >=
1760 (reg_buf_offset + reg_buf_bytes)) {
1762 * The next di_next_unlinked field is beyond
1763 * the current logged region. Find the next
1764 * logged region that contains or is beyond
1765 * the current di_next_unlinked field.
1767 bit += nbits;
1768 bit = xfs_next_bit(buf_f->blf_data_map,
1769 buf_f->blf_map_size, bit);
1772 * If there are no more logged regions in the
1773 * buffer, then we're done.
1775 if (bit == -1)
1776 return 0;
1778 nbits = xfs_contig_bits(buf_f->blf_data_map,
1779 buf_f->blf_map_size, bit);
1780 ASSERT(nbits > 0);
1781 reg_buf_offset = bit << XFS_BLF_SHIFT;
1782 reg_buf_bytes = nbits << XFS_BLF_SHIFT;
1783 item_index++;
1787 * If the current logged region starts after the current
1788 * di_next_unlinked field, then move on to the next
1789 * di_next_unlinked field.
1791 if (next_unlinked_offset < reg_buf_offset)
1792 continue;
1794 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1795 ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
1796 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1799 * The current logged region contains a copy of the
1800 * current di_next_unlinked field. Extract its value
1801 * and copy it to the buffer copy.
1803 logged_nextp = item->ri_buf[item_index].i_addr +
1804 next_unlinked_offset - reg_buf_offset;
1805 if (unlikely(*logged_nextp == 0)) {
1806 xfs_fs_cmn_err(CE_ALERT, mp,
1807 "bad inode buffer log record (ptr = 0x%p, bp = 0x%p). XFS trying to replay bad (0) inode di_next_unlinked field",
1808 item, bp);
1809 XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1810 XFS_ERRLEVEL_LOW, mp);
1811 return XFS_ERROR(EFSCORRUPTED);
1814 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1815 next_unlinked_offset);
1816 *buffer_nextp = *logged_nextp;
1819 return 0;
1823 * Perform a 'normal' buffer recovery. Each logged region of the
1824 * buffer should be copied over the corresponding region in the
1825 * given buffer. The bitmap in the buf log format structure indicates
1826 * where to place the logged data.
1828 STATIC void
1829 xlog_recover_do_reg_buffer(
1830 struct xfs_mount *mp,
1831 xlog_recover_item_t *item,
1832 struct xfs_buf *bp,
1833 xfs_buf_log_format_t *buf_f)
1835 int i;
1836 int bit;
1837 int nbits;
1838 int error;
1840 trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
1842 bit = 0;
1843 i = 1; /* 0 is the buf format structure */
1844 while (1) {
1845 bit = xfs_next_bit(buf_f->blf_data_map,
1846 buf_f->blf_map_size, bit);
1847 if (bit == -1)
1848 break;
1849 nbits = xfs_contig_bits(buf_f->blf_data_map,
1850 buf_f->blf_map_size, bit);
1851 ASSERT(nbits > 0);
1852 ASSERT(item->ri_buf[i].i_addr != NULL);
1853 ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
1854 ASSERT(XFS_BUF_COUNT(bp) >=
1855 ((uint)bit << XFS_BLF_SHIFT)+(nbits<<XFS_BLF_SHIFT));
1858 * Do a sanity check if this is a dquot buffer. Just checking
1859 * the first dquot in the buffer should do. XXXThis is
1860 * probably a good thing to do for other buf types also.
1862 error = 0;
1863 if (buf_f->blf_flags &
1864 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
1865 if (item->ri_buf[i].i_addr == NULL) {
1866 cmn_err(CE_ALERT,
1867 "XFS: NULL dquot in %s.", __func__);
1868 goto next;
1870 if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
1871 cmn_err(CE_ALERT,
1872 "XFS: dquot too small (%d) in %s.",
1873 item->ri_buf[i].i_len, __func__);
1874 goto next;
1876 error = xfs_qm_dqcheck(item->ri_buf[i].i_addr,
1877 -1, 0, XFS_QMOPT_DOWARN,
1878 "dquot_buf_recover");
1879 if (error)
1880 goto next;
1883 memcpy(xfs_buf_offset(bp,
1884 (uint)bit << XFS_BLF_SHIFT), /* dest */
1885 item->ri_buf[i].i_addr, /* source */
1886 nbits<<XFS_BLF_SHIFT); /* length */
1887 next:
1888 i++;
1889 bit += nbits;
1892 /* Shouldn't be any more regions */
1893 ASSERT(i == item->ri_total);
1897 * Do some primitive error checking on ondisk dquot data structures.
1900 xfs_qm_dqcheck(
1901 xfs_disk_dquot_t *ddq,
1902 xfs_dqid_t id,
1903 uint type, /* used only when IO_dorepair is true */
1904 uint flags,
1905 char *str)
1907 xfs_dqblk_t *d = (xfs_dqblk_t *)ddq;
1908 int errs = 0;
1911 * We can encounter an uninitialized dquot buffer for 2 reasons:
1912 * 1. If we crash while deleting the quotainode(s), and those blks got
1913 * used for user data. This is because we take the path of regular
1914 * file deletion; however, the size field of quotainodes is never
1915 * updated, so all the tricks that we play in itruncate_finish
1916 * don't quite matter.
1918 * 2. We don't play the quota buffers when there's a quotaoff logitem.
1919 * But the allocation will be replayed so we'll end up with an
1920 * uninitialized quota block.
1922 * This is all fine; things are still consistent, and we haven't lost
1923 * any quota information. Just don't complain about bad dquot blks.
1925 if (be16_to_cpu(ddq->d_magic) != XFS_DQUOT_MAGIC) {
1926 if (flags & XFS_QMOPT_DOWARN)
1927 cmn_err(CE_ALERT,
1928 "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
1929 str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
1930 errs++;
1932 if (ddq->d_version != XFS_DQUOT_VERSION) {
1933 if (flags & XFS_QMOPT_DOWARN)
1934 cmn_err(CE_ALERT,
1935 "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
1936 str, id, ddq->d_version, XFS_DQUOT_VERSION);
1937 errs++;
1940 if (ddq->d_flags != XFS_DQ_USER &&
1941 ddq->d_flags != XFS_DQ_PROJ &&
1942 ddq->d_flags != XFS_DQ_GROUP) {
1943 if (flags & XFS_QMOPT_DOWARN)
1944 cmn_err(CE_ALERT,
1945 "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
1946 str, id, ddq->d_flags);
1947 errs++;
1950 if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
1951 if (flags & XFS_QMOPT_DOWARN)
1952 cmn_err(CE_ALERT,
1953 "%s : ondisk-dquot 0x%p, ID mismatch: "
1954 "0x%x expected, found id 0x%x",
1955 str, ddq, id, be32_to_cpu(ddq->d_id));
1956 errs++;
1959 if (!errs && ddq->d_id) {
1960 if (ddq->d_blk_softlimit &&
1961 be64_to_cpu(ddq->d_bcount) >=
1962 be64_to_cpu(ddq->d_blk_softlimit)) {
1963 if (!ddq->d_btimer) {
1964 if (flags & XFS_QMOPT_DOWARN)
1965 cmn_err(CE_ALERT,
1966 "%s : Dquot ID 0x%x (0x%p) "
1967 "BLK TIMER NOT STARTED",
1968 str, (int)be32_to_cpu(ddq->d_id), ddq);
1969 errs++;
1972 if (ddq->d_ino_softlimit &&
1973 be64_to_cpu(ddq->d_icount) >=
1974 be64_to_cpu(ddq->d_ino_softlimit)) {
1975 if (!ddq->d_itimer) {
1976 if (flags & XFS_QMOPT_DOWARN)
1977 cmn_err(CE_ALERT,
1978 "%s : Dquot ID 0x%x (0x%p) "
1979 "INODE TIMER NOT STARTED",
1980 str, (int)be32_to_cpu(ddq->d_id), ddq);
1981 errs++;
1984 if (ddq->d_rtb_softlimit &&
1985 be64_to_cpu(ddq->d_rtbcount) >=
1986 be64_to_cpu(ddq->d_rtb_softlimit)) {
1987 if (!ddq->d_rtbtimer) {
1988 if (flags & XFS_QMOPT_DOWARN)
1989 cmn_err(CE_ALERT,
1990 "%s : Dquot ID 0x%x (0x%p) "
1991 "RTBLK TIMER NOT STARTED",
1992 str, (int)be32_to_cpu(ddq->d_id), ddq);
1993 errs++;
1998 if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
1999 return errs;
2001 if (flags & XFS_QMOPT_DOWARN)
2002 cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2005 * Typically, a repair is only requested by quotacheck.
2007 ASSERT(id != -1);
2008 ASSERT(flags & XFS_QMOPT_DQREPAIR);
2009 memset(d, 0, sizeof(xfs_dqblk_t));
2011 d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2012 d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2013 d->dd_diskdq.d_flags = type;
2014 d->dd_diskdq.d_id = cpu_to_be32(id);
2016 return errs;
2020 * Perform a dquot buffer recovery.
2021 * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2022 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2023 * Else, treat it as a regular buffer and do recovery.
2025 STATIC void
2026 xlog_recover_do_dquot_buffer(
2027 xfs_mount_t *mp,
2028 xlog_t *log,
2029 xlog_recover_item_t *item,
2030 xfs_buf_t *bp,
2031 xfs_buf_log_format_t *buf_f)
2033 uint type;
2035 trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2038 * Filesystems are required to send in quota flags at mount time.
2040 if (mp->m_qflags == 0) {
2041 return;
2044 type = 0;
2045 if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2046 type |= XFS_DQ_USER;
2047 if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2048 type |= XFS_DQ_PROJ;
2049 if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2050 type |= XFS_DQ_GROUP;
2052 * This type of quotas was turned off, so ignore this buffer
2054 if (log->l_quotaoffs_flag & type)
2055 return;
2057 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2061 * This routine replays a modification made to a buffer at runtime.
2062 * There are actually two types of buffer, regular and inode, which
2063 * are handled differently. Inode buffers are handled differently
2064 * in that we only recover a specific set of data from them, namely
2065 * the inode di_next_unlinked fields. This is because all other inode
2066 * data is actually logged via inode records and any data we replay
2067 * here which overlaps that may be stale.
2069 * When meta-data buffers are freed at run time we log a buffer item
2070 * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2071 * of the buffer in the log should not be replayed at recovery time.
2072 * This is so that if the blocks covered by the buffer are reused for
2073 * file data before we crash we don't end up replaying old, freed
2074 * meta-data into a user's file.
2076 * To handle the cancellation of buffer log items, we make two passes
2077 * over the log during recovery. During the first we build a table of
2078 * those buffers which have been cancelled, and during the second we
2079 * only replay those buffers which do not have corresponding cancel
2080 * records in the table. See xlog_recover_do_buffer_pass[1,2] above
2081 * for more details on the implementation of the table of cancel records.
2083 STATIC int
2084 xlog_recover_buffer_pass2(
2085 xlog_t *log,
2086 xlog_recover_item_t *item)
2088 xfs_buf_log_format_t *buf_f = item->ri_buf[0].i_addr;
2089 xfs_mount_t *mp = log->l_mp;
2090 xfs_buf_t *bp;
2091 int error;
2092 uint buf_flags;
2095 * In this pass we only want to recover all the buffers which have
2096 * not been cancelled and are not cancellation buffers themselves.
2098 if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2099 buf_f->blf_len, buf_f->blf_flags)) {
2100 trace_xfs_log_recover_buf_cancel(log, buf_f);
2101 return 0;
2104 trace_xfs_log_recover_buf_recover(log, buf_f);
2106 buf_flags = XBF_LOCK;
2107 if (!(buf_f->blf_flags & XFS_BLF_INODE_BUF))
2108 buf_flags |= XBF_MAPPED;
2110 bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2111 buf_flags);
2112 if (XFS_BUF_ISERROR(bp)) {
2113 xfs_ioerror_alert("xlog_recover_do..(read#1)", mp,
2114 bp, buf_f->blf_blkno);
2115 error = XFS_BUF_GETERROR(bp);
2116 xfs_buf_relse(bp);
2117 return error;
2120 error = 0;
2121 if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2122 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2123 } else if (buf_f->blf_flags &
2124 (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2125 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2126 } else {
2127 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2129 if (error)
2130 return XFS_ERROR(error);
2133 * Perform delayed write on the buffer. Asynchronous writes will be
2134 * slower when taking into account all the buffers to be flushed.
2136 * Also make sure that only inode buffers with good sizes stay in
2137 * the buffer cache. The kernel moves inodes in buffers of 1 block
2138 * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger. The inode
2139 * buffers in the log can be a different size if the log was generated
2140 * by an older kernel using unclustered inode buffers or a newer kernel
2141 * running with a different inode cluster size. Regardless, if the
2142 * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2143 * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2144 * the buffer out of the buffer cache so that the buffer won't
2145 * overlap with future reads of those inodes.
2147 if (XFS_DINODE_MAGIC ==
2148 be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2149 (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2150 (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2151 XFS_BUF_STALE(bp);
2152 error = xfs_bwrite(mp, bp);
2153 } else {
2154 ASSERT(bp->b_target->bt_mount == mp);
2155 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2156 xfs_bdwrite(mp, bp);
2159 return (error);
2162 STATIC int
2163 xlog_recover_inode_pass2(
2164 xlog_t *log,
2165 xlog_recover_item_t *item)
2167 xfs_inode_log_format_t *in_f;
2168 xfs_mount_t *mp = log->l_mp;
2169 xfs_buf_t *bp;
2170 xfs_dinode_t *dip;
2171 int len;
2172 xfs_caddr_t src;
2173 xfs_caddr_t dest;
2174 int error;
2175 int attr_index;
2176 uint fields;
2177 xfs_icdinode_t *dicp;
2178 int need_free = 0;
2180 if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2181 in_f = item->ri_buf[0].i_addr;
2182 } else {
2183 in_f = kmem_alloc(sizeof(xfs_inode_log_format_t), KM_SLEEP);
2184 need_free = 1;
2185 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2186 if (error)
2187 goto error;
2191 * Inode buffers can be freed, look out for it,
2192 * and do not replay the inode.
2194 if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2195 in_f->ilf_len, 0)) {
2196 error = 0;
2197 trace_xfs_log_recover_inode_cancel(log, in_f);
2198 goto error;
2200 trace_xfs_log_recover_inode_recover(log, in_f);
2202 bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len,
2203 XBF_LOCK);
2204 if (XFS_BUF_ISERROR(bp)) {
2205 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2206 bp, in_f->ilf_blkno);
2207 error = XFS_BUF_GETERROR(bp);
2208 xfs_buf_relse(bp);
2209 goto error;
2211 error = 0;
2212 ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2213 dip = (xfs_dinode_t *)xfs_buf_offset(bp, in_f->ilf_boffset);
2216 * Make sure the place we're flushing out to really looks
2217 * like an inode!
2219 if (unlikely(be16_to_cpu(dip->di_magic) != XFS_DINODE_MAGIC)) {
2220 xfs_buf_relse(bp);
2221 xfs_fs_cmn_err(CE_ALERT, mp,
2222 "xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2223 dip, bp, in_f->ilf_ino);
2224 XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
2225 XFS_ERRLEVEL_LOW, mp);
2226 error = EFSCORRUPTED;
2227 goto error;
2229 dicp = item->ri_buf[1].i_addr;
2230 if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2231 xfs_buf_relse(bp);
2232 xfs_fs_cmn_err(CE_ALERT, mp,
2233 "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2234 item, in_f->ilf_ino);
2235 XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
2236 XFS_ERRLEVEL_LOW, mp);
2237 error = EFSCORRUPTED;
2238 goto error;
2241 /* Skip replay when the on disk inode is newer than the log one */
2242 if (dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
2244 * Deal with the wrap case, DI_MAX_FLUSH is less
2245 * than smaller numbers
2247 if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
2248 dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2249 /* do nothing */
2250 } else {
2251 xfs_buf_relse(bp);
2252 trace_xfs_log_recover_inode_skip(log, in_f);
2253 error = 0;
2254 goto error;
2257 /* Take the opportunity to reset the flush iteration count */
2258 dicp->di_flushiter = 0;
2260 if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2261 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2262 (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2263 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
2264 XFS_ERRLEVEL_LOW, mp, dicp);
2265 xfs_buf_relse(bp);
2266 xfs_fs_cmn_err(CE_ALERT, mp,
2267 "xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2268 item, dip, bp, in_f->ilf_ino);
2269 error = EFSCORRUPTED;
2270 goto error;
2272 } else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2273 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2274 (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2275 (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2276 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
2277 XFS_ERRLEVEL_LOW, mp, dicp);
2278 xfs_buf_relse(bp);
2279 xfs_fs_cmn_err(CE_ALERT, mp,
2280 "xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2281 item, dip, bp, in_f->ilf_ino);
2282 error = EFSCORRUPTED;
2283 goto error;
2286 if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2287 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
2288 XFS_ERRLEVEL_LOW, mp, dicp);
2289 xfs_buf_relse(bp);
2290 xfs_fs_cmn_err(CE_ALERT, mp,
2291 "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2292 item, dip, bp, in_f->ilf_ino,
2293 dicp->di_nextents + dicp->di_anextents,
2294 dicp->di_nblocks);
2295 error = EFSCORRUPTED;
2296 goto error;
2298 if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2299 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
2300 XFS_ERRLEVEL_LOW, mp, dicp);
2301 xfs_buf_relse(bp);
2302 xfs_fs_cmn_err(CE_ALERT, mp,
2303 "xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2304 item, dip, bp, in_f->ilf_ino, dicp->di_forkoff);
2305 error = EFSCORRUPTED;
2306 goto error;
2308 if (unlikely(item->ri_buf[1].i_len > sizeof(struct xfs_icdinode))) {
2309 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
2310 XFS_ERRLEVEL_LOW, mp, dicp);
2311 xfs_buf_relse(bp);
2312 xfs_fs_cmn_err(CE_ALERT, mp,
2313 "xfs_inode_recover: Bad inode log record length %d, rec ptr 0x%p",
2314 item->ri_buf[1].i_len, item);
2315 error = EFSCORRUPTED;
2316 goto error;
2319 /* The core is in in-core format */
2320 xfs_dinode_to_disk(dip, item->ri_buf[1].i_addr);
2322 /* the rest is in on-disk format */
2323 if (item->ri_buf[1].i_len > sizeof(struct xfs_icdinode)) {
2324 memcpy((xfs_caddr_t) dip + sizeof(struct xfs_icdinode),
2325 item->ri_buf[1].i_addr + sizeof(struct xfs_icdinode),
2326 item->ri_buf[1].i_len - sizeof(struct xfs_icdinode));
2329 fields = in_f->ilf_fields;
2330 switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2331 case XFS_ILOG_DEV:
2332 xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
2333 break;
2334 case XFS_ILOG_UUID:
2335 memcpy(XFS_DFORK_DPTR(dip),
2336 &in_f->ilf_u.ilfu_uuid,
2337 sizeof(uuid_t));
2338 break;
2341 if (in_f->ilf_size == 2)
2342 goto write_inode_buffer;
2343 len = item->ri_buf[2].i_len;
2344 src = item->ri_buf[2].i_addr;
2345 ASSERT(in_f->ilf_size <= 4);
2346 ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2347 ASSERT(!(fields & XFS_ILOG_DFORK) ||
2348 (len == in_f->ilf_dsize));
2350 switch (fields & XFS_ILOG_DFORK) {
2351 case XFS_ILOG_DDATA:
2352 case XFS_ILOG_DEXT:
2353 memcpy(XFS_DFORK_DPTR(dip), src, len);
2354 break;
2356 case XFS_ILOG_DBROOT:
2357 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
2358 (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
2359 XFS_DFORK_DSIZE(dip, mp));
2360 break;
2362 default:
2364 * There are no data fork flags set.
2366 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2367 break;
2371 * If we logged any attribute data, recover it. There may or
2372 * may not have been any other non-core data logged in this
2373 * transaction.
2375 if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2376 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2377 attr_index = 3;
2378 } else {
2379 attr_index = 2;
2381 len = item->ri_buf[attr_index].i_len;
2382 src = item->ri_buf[attr_index].i_addr;
2383 ASSERT(len == in_f->ilf_asize);
2385 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2386 case XFS_ILOG_ADATA:
2387 case XFS_ILOG_AEXT:
2388 dest = XFS_DFORK_APTR(dip);
2389 ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2390 memcpy(dest, src, len);
2391 break;
2393 case XFS_ILOG_ABROOT:
2394 dest = XFS_DFORK_APTR(dip);
2395 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
2396 len, (xfs_bmdr_block_t*)dest,
2397 XFS_DFORK_ASIZE(dip, mp));
2398 break;
2400 default:
2401 xlog_warn("XFS: xlog_recover_inode_pass2: Invalid flag");
2402 ASSERT(0);
2403 xfs_buf_relse(bp);
2404 error = EIO;
2405 goto error;
2409 write_inode_buffer:
2410 ASSERT(bp->b_target->bt_mount == mp);
2411 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2412 xfs_bdwrite(mp, bp);
2413 error:
2414 if (need_free)
2415 kmem_free(in_f);
2416 return XFS_ERROR(error);
2420 * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2421 * structure, so that we know not to do any dquot item or dquot buffer recovery,
2422 * of that type.
2424 STATIC int
2425 xlog_recover_quotaoff_pass1(
2426 xlog_t *log,
2427 xlog_recover_item_t *item)
2429 xfs_qoff_logformat_t *qoff_f = item->ri_buf[0].i_addr;
2430 ASSERT(qoff_f);
2433 * The logitem format's flag tells us if this was user quotaoff,
2434 * group/project quotaoff or both.
2436 if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2437 log->l_quotaoffs_flag |= XFS_DQ_USER;
2438 if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2439 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2440 if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2441 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2443 return (0);
2447 * Recover a dquot record
2449 STATIC int
2450 xlog_recover_dquot_pass2(
2451 xlog_t *log,
2452 xlog_recover_item_t *item)
2454 xfs_mount_t *mp = log->l_mp;
2455 xfs_buf_t *bp;
2456 struct xfs_disk_dquot *ddq, *recddq;
2457 int error;
2458 xfs_dq_logformat_t *dq_f;
2459 uint type;
2463 * Filesystems are required to send in quota flags at mount time.
2465 if (mp->m_qflags == 0)
2466 return (0);
2468 recddq = item->ri_buf[1].i_addr;
2469 if (recddq == NULL) {
2470 cmn_err(CE_ALERT,
2471 "XFS: NULL dquot in %s.", __func__);
2472 return XFS_ERROR(EIO);
2474 if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
2475 cmn_err(CE_ALERT,
2476 "XFS: dquot too small (%d) in %s.",
2477 item->ri_buf[1].i_len, __func__);
2478 return XFS_ERROR(EIO);
2482 * This type of quotas was turned off, so ignore this record.
2484 type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2485 ASSERT(type);
2486 if (log->l_quotaoffs_flag & type)
2487 return (0);
2490 * At this point we know that quota was _not_ turned off.
2491 * Since the mount flags are not indicating to us otherwise, this
2492 * must mean that quota is on, and the dquot needs to be replayed.
2493 * Remember that we may not have fully recovered the superblock yet,
2494 * so we can't do the usual trick of looking at the SB quota bits.
2496 * The other possibility, of course, is that the quota subsystem was
2497 * removed since the last mount - ENOSYS.
2499 dq_f = item->ri_buf[0].i_addr;
2500 ASSERT(dq_f);
2501 if ((error = xfs_qm_dqcheck(recddq,
2502 dq_f->qlf_id,
2503 0, XFS_QMOPT_DOWARN,
2504 "xlog_recover_dquot_pass2 (log copy)"))) {
2505 return XFS_ERROR(EIO);
2507 ASSERT(dq_f->qlf_len == 1);
2509 error = xfs_read_buf(mp, mp->m_ddev_targp,
2510 dq_f->qlf_blkno,
2511 XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2512 0, &bp);
2513 if (error) {
2514 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2515 bp, dq_f->qlf_blkno);
2516 return error;
2518 ASSERT(bp);
2519 ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2522 * At least the magic num portion should be on disk because this
2523 * was among a chunk of dquots created earlier, and we did some
2524 * minimal initialization then.
2526 if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2527 "xlog_recover_dquot_pass2")) {
2528 xfs_buf_relse(bp);
2529 return XFS_ERROR(EIO);
2532 memcpy(ddq, recddq, item->ri_buf[1].i_len);
2534 ASSERT(dq_f->qlf_size == 2);
2535 ASSERT(bp->b_target->bt_mount == mp);
2536 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2537 xfs_bdwrite(mp, bp);
2539 return (0);
2543 * This routine is called to create an in-core extent free intent
2544 * item from the efi format structure which was logged on disk.
2545 * It allocates an in-core efi, copies the extents from the format
2546 * structure into it, and adds the efi to the AIL with the given
2547 * LSN.
2549 STATIC int
2550 xlog_recover_efi_pass2(
2551 xlog_t *log,
2552 xlog_recover_item_t *item,
2553 xfs_lsn_t lsn)
2555 int error;
2556 xfs_mount_t *mp = log->l_mp;
2557 xfs_efi_log_item_t *efip;
2558 xfs_efi_log_format_t *efi_formatp;
2560 efi_formatp = item->ri_buf[0].i_addr;
2562 efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2563 if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2564 &(efip->efi_format)))) {
2565 xfs_efi_item_free(efip);
2566 return error;
2568 atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
2570 spin_lock(&log->l_ailp->xa_lock);
2572 * xfs_trans_ail_update() drops the AIL lock.
2574 xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
2575 return 0;
2580 * This routine is called when an efd format structure is found in
2581 * a committed transaction in the log. It's purpose is to cancel
2582 * the corresponding efi if it was still in the log. To do this
2583 * it searches the AIL for the efi with an id equal to that in the
2584 * efd format structure. If we find it, we remove the efi from the
2585 * AIL and free it.
2587 STATIC int
2588 xlog_recover_efd_pass2(
2589 xlog_t *log,
2590 xlog_recover_item_t *item)
2592 xfs_efd_log_format_t *efd_formatp;
2593 xfs_efi_log_item_t *efip = NULL;
2594 xfs_log_item_t *lip;
2595 __uint64_t efi_id;
2596 struct xfs_ail_cursor cur;
2597 struct xfs_ail *ailp = log->l_ailp;
2599 efd_formatp = item->ri_buf[0].i_addr;
2600 ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2601 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2602 (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2603 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2604 efi_id = efd_formatp->efd_efi_id;
2607 * Search for the efi with the id in the efd format structure
2608 * in the AIL.
2610 spin_lock(&ailp->xa_lock);
2611 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2612 while (lip != NULL) {
2613 if (lip->li_type == XFS_LI_EFI) {
2614 efip = (xfs_efi_log_item_t *)lip;
2615 if (efip->efi_format.efi_id == efi_id) {
2617 * xfs_trans_ail_delete() drops the
2618 * AIL lock.
2620 xfs_trans_ail_delete(ailp, lip);
2621 xfs_efi_item_free(efip);
2622 spin_lock(&ailp->xa_lock);
2623 break;
2626 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2628 xfs_trans_ail_cursor_done(ailp, &cur);
2629 spin_unlock(&ailp->xa_lock);
2631 return 0;
2635 * Free up any resources allocated by the transaction
2637 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2639 STATIC void
2640 xlog_recover_free_trans(
2641 struct xlog_recover *trans)
2643 xlog_recover_item_t *item, *n;
2644 int i;
2646 list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2647 /* Free the regions in the item. */
2648 list_del(&item->ri_list);
2649 for (i = 0; i < item->ri_cnt; i++)
2650 kmem_free(item->ri_buf[i].i_addr);
2651 /* Free the item itself */
2652 kmem_free(item->ri_buf);
2653 kmem_free(item);
2655 /* Free the transaction recover structure */
2656 kmem_free(trans);
2659 STATIC int
2660 xlog_recover_commit_pass1(
2661 struct log *log,
2662 struct xlog_recover *trans,
2663 xlog_recover_item_t *item)
2665 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
2667 switch (ITEM_TYPE(item)) {
2668 case XFS_LI_BUF:
2669 return xlog_recover_buffer_pass1(log, item);
2670 case XFS_LI_QUOTAOFF:
2671 return xlog_recover_quotaoff_pass1(log, item);
2672 case XFS_LI_INODE:
2673 case XFS_LI_EFI:
2674 case XFS_LI_EFD:
2675 case XFS_LI_DQUOT:
2676 /* nothing to do in pass 1 */
2677 return 0;
2678 default:
2679 xlog_warn(
2680 "XFS: invalid item type (%d) xlog_recover_commit_pass1",
2681 ITEM_TYPE(item));
2682 ASSERT(0);
2683 return XFS_ERROR(EIO);
2687 STATIC int
2688 xlog_recover_commit_pass2(
2689 struct log *log,
2690 struct xlog_recover *trans,
2691 xlog_recover_item_t *item)
2693 trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
2695 switch (ITEM_TYPE(item)) {
2696 case XFS_LI_BUF:
2697 return xlog_recover_buffer_pass2(log, item);
2698 case XFS_LI_INODE:
2699 return xlog_recover_inode_pass2(log, item);
2700 case XFS_LI_EFI:
2701 return xlog_recover_efi_pass2(log, item, trans->r_lsn);
2702 case XFS_LI_EFD:
2703 return xlog_recover_efd_pass2(log, item);
2704 case XFS_LI_DQUOT:
2705 return xlog_recover_dquot_pass2(log, item);
2706 case XFS_LI_QUOTAOFF:
2707 /* nothing to do in pass2 */
2708 return 0;
2709 default:
2710 xlog_warn(
2711 "XFS: invalid item type (%d) xlog_recover_commit_pass2",
2712 ITEM_TYPE(item));
2713 ASSERT(0);
2714 return XFS_ERROR(EIO);
2719 * Perform the transaction.
2721 * If the transaction modifies a buffer or inode, do it now. Otherwise,
2722 * EFIs and EFDs get queued up by adding entries into the AIL for them.
2724 STATIC int
2725 xlog_recover_commit_trans(
2726 struct log *log,
2727 struct xlog_recover *trans,
2728 int pass)
2730 int error = 0;
2731 xlog_recover_item_t *item;
2733 hlist_del(&trans->r_list);
2735 error = xlog_recover_reorder_trans(log, trans, pass);
2736 if (error)
2737 return error;
2739 list_for_each_entry(item, &trans->r_itemq, ri_list) {
2740 if (pass == XLOG_RECOVER_PASS1)
2741 error = xlog_recover_commit_pass1(log, trans, item);
2742 else
2743 error = xlog_recover_commit_pass2(log, trans, item);
2744 if (error)
2745 return error;
2748 xlog_recover_free_trans(trans);
2749 return 0;
2752 STATIC int
2753 xlog_recover_unmount_trans(
2754 xlog_recover_t *trans)
2756 /* Do nothing now */
2757 xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2758 return 0;
2762 * There are two valid states of the r_state field. 0 indicates that the
2763 * transaction structure is in a normal state. We have either seen the
2764 * start of the transaction or the last operation we added was not a partial
2765 * operation. If the last operation we added to the transaction was a
2766 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2768 * NOTE: skip LRs with 0 data length.
2770 STATIC int
2771 xlog_recover_process_data(
2772 xlog_t *log,
2773 struct hlist_head rhash[],
2774 xlog_rec_header_t *rhead,
2775 xfs_caddr_t dp,
2776 int pass)
2778 xfs_caddr_t lp;
2779 int num_logops;
2780 xlog_op_header_t *ohead;
2781 xlog_recover_t *trans;
2782 xlog_tid_t tid;
2783 int error;
2784 unsigned long hash;
2785 uint flags;
2787 lp = dp + be32_to_cpu(rhead->h_len);
2788 num_logops = be32_to_cpu(rhead->h_num_logops);
2790 /* check the log format matches our own - else we can't recover */
2791 if (xlog_header_check_recover(log->l_mp, rhead))
2792 return (XFS_ERROR(EIO));
2794 while ((dp < lp) && num_logops) {
2795 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2796 ohead = (xlog_op_header_t *)dp;
2797 dp += sizeof(xlog_op_header_t);
2798 if (ohead->oh_clientid != XFS_TRANSACTION &&
2799 ohead->oh_clientid != XFS_LOG) {
2800 xlog_warn(
2801 "XFS: xlog_recover_process_data: bad clientid");
2802 ASSERT(0);
2803 return (XFS_ERROR(EIO));
2805 tid = be32_to_cpu(ohead->oh_tid);
2806 hash = XLOG_RHASH(tid);
2807 trans = xlog_recover_find_tid(&rhash[hash], tid);
2808 if (trans == NULL) { /* not found; add new tid */
2809 if (ohead->oh_flags & XLOG_START_TRANS)
2810 xlog_recover_new_tid(&rhash[hash], tid,
2811 be64_to_cpu(rhead->h_lsn));
2812 } else {
2813 if (dp + be32_to_cpu(ohead->oh_len) > lp) {
2814 xlog_warn(
2815 "XFS: xlog_recover_process_data: bad length");
2816 WARN_ON(1);
2817 return (XFS_ERROR(EIO));
2819 flags = ohead->oh_flags & ~XLOG_END_TRANS;
2820 if (flags & XLOG_WAS_CONT_TRANS)
2821 flags &= ~XLOG_CONTINUE_TRANS;
2822 switch (flags) {
2823 case XLOG_COMMIT_TRANS:
2824 error = xlog_recover_commit_trans(log,
2825 trans, pass);
2826 break;
2827 case XLOG_UNMOUNT_TRANS:
2828 error = xlog_recover_unmount_trans(trans);
2829 break;
2830 case XLOG_WAS_CONT_TRANS:
2831 error = xlog_recover_add_to_cont_trans(log,
2832 trans, dp,
2833 be32_to_cpu(ohead->oh_len));
2834 break;
2835 case XLOG_START_TRANS:
2836 xlog_warn(
2837 "XFS: xlog_recover_process_data: bad transaction");
2838 ASSERT(0);
2839 error = XFS_ERROR(EIO);
2840 break;
2841 case 0:
2842 case XLOG_CONTINUE_TRANS:
2843 error = xlog_recover_add_to_trans(log, trans,
2844 dp, be32_to_cpu(ohead->oh_len));
2845 break;
2846 default:
2847 xlog_warn(
2848 "XFS: xlog_recover_process_data: bad flag");
2849 ASSERT(0);
2850 error = XFS_ERROR(EIO);
2851 break;
2853 if (error)
2854 return error;
2856 dp += be32_to_cpu(ohead->oh_len);
2857 num_logops--;
2859 return 0;
2863 * Process an extent free intent item that was recovered from
2864 * the log. We need to free the extents that it describes.
2866 STATIC int
2867 xlog_recover_process_efi(
2868 xfs_mount_t *mp,
2869 xfs_efi_log_item_t *efip)
2871 xfs_efd_log_item_t *efdp;
2872 xfs_trans_t *tp;
2873 int i;
2874 int error = 0;
2875 xfs_extent_t *extp;
2876 xfs_fsblock_t startblock_fsb;
2878 ASSERT(!test_bit(XFS_EFI_RECOVERED, &efip->efi_flags));
2881 * First check the validity of the extents described by the
2882 * EFI. If any are bad, then assume that all are bad and
2883 * just toss the EFI.
2885 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2886 extp = &(efip->efi_format.efi_extents[i]);
2887 startblock_fsb = XFS_BB_TO_FSB(mp,
2888 XFS_FSB_TO_DADDR(mp, extp->ext_start));
2889 if ((startblock_fsb == 0) ||
2890 (extp->ext_len == 0) ||
2891 (startblock_fsb >= mp->m_sb.sb_dblocks) ||
2892 (extp->ext_len >= mp->m_sb.sb_agblocks)) {
2894 * This will pull the EFI from the AIL and
2895 * free the memory associated with it.
2897 xfs_efi_release(efip, efip->efi_format.efi_nextents);
2898 return XFS_ERROR(EIO);
2902 tp = xfs_trans_alloc(mp, 0);
2903 error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
2904 if (error)
2905 goto abort_error;
2906 efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
2908 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2909 extp = &(efip->efi_format.efi_extents[i]);
2910 error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
2911 if (error)
2912 goto abort_error;
2913 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
2914 extp->ext_len);
2917 set_bit(XFS_EFI_RECOVERED, &efip->efi_flags);
2918 error = xfs_trans_commit(tp, 0);
2919 return error;
2921 abort_error:
2922 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
2923 return error;
2927 * When this is called, all of the EFIs which did not have
2928 * corresponding EFDs should be in the AIL. What we do now
2929 * is free the extents associated with each one.
2931 * Since we process the EFIs in normal transactions, they
2932 * will be removed at some point after the commit. This prevents
2933 * us from just walking down the list processing each one.
2934 * We'll use a flag in the EFI to skip those that we've already
2935 * processed and use the AIL iteration mechanism's generation
2936 * count to try to speed this up at least a bit.
2938 * When we start, we know that the EFIs are the only things in
2939 * the AIL. As we process them, however, other items are added
2940 * to the AIL. Since everything added to the AIL must come after
2941 * everything already in the AIL, we stop processing as soon as
2942 * we see something other than an EFI in the AIL.
2944 STATIC int
2945 xlog_recover_process_efis(
2946 xlog_t *log)
2948 xfs_log_item_t *lip;
2949 xfs_efi_log_item_t *efip;
2950 int error = 0;
2951 struct xfs_ail_cursor cur;
2952 struct xfs_ail *ailp;
2954 ailp = log->l_ailp;
2955 spin_lock(&ailp->xa_lock);
2956 lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2957 while (lip != NULL) {
2959 * We're done when we see something other than an EFI.
2960 * There should be no EFIs left in the AIL now.
2962 if (lip->li_type != XFS_LI_EFI) {
2963 #ifdef DEBUG
2964 for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
2965 ASSERT(lip->li_type != XFS_LI_EFI);
2966 #endif
2967 break;
2971 * Skip EFIs that we've already processed.
2973 efip = (xfs_efi_log_item_t *)lip;
2974 if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags)) {
2975 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2976 continue;
2979 spin_unlock(&ailp->xa_lock);
2980 error = xlog_recover_process_efi(log->l_mp, efip);
2981 spin_lock(&ailp->xa_lock);
2982 if (error)
2983 goto out;
2984 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2986 out:
2987 xfs_trans_ail_cursor_done(ailp, &cur);
2988 spin_unlock(&ailp->xa_lock);
2989 return error;
2993 * This routine performs a transaction to null out a bad inode pointer
2994 * in an agi unlinked inode hash bucket.
2996 STATIC void
2997 xlog_recover_clear_agi_bucket(
2998 xfs_mount_t *mp,
2999 xfs_agnumber_t agno,
3000 int bucket)
3002 xfs_trans_t *tp;
3003 xfs_agi_t *agi;
3004 xfs_buf_t *agibp;
3005 int offset;
3006 int error;
3008 tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3009 error = xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp),
3010 0, 0, 0);
3011 if (error)
3012 goto out_abort;
3014 error = xfs_read_agi(mp, tp, agno, &agibp);
3015 if (error)
3016 goto out_abort;
3018 agi = XFS_BUF_TO_AGI(agibp);
3019 agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3020 offset = offsetof(xfs_agi_t, agi_unlinked) +
3021 (sizeof(xfs_agino_t) * bucket);
3022 xfs_trans_log_buf(tp, agibp, offset,
3023 (offset + sizeof(xfs_agino_t) - 1));
3025 error = xfs_trans_commit(tp, 0);
3026 if (error)
3027 goto out_error;
3028 return;
3030 out_abort:
3031 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3032 out_error:
3033 xfs_fs_cmn_err(CE_WARN, mp, "xlog_recover_clear_agi_bucket: "
3034 "failed to clear agi %d. Continuing.", agno);
3035 return;
3038 STATIC xfs_agino_t
3039 xlog_recover_process_one_iunlink(
3040 struct xfs_mount *mp,
3041 xfs_agnumber_t agno,
3042 xfs_agino_t agino,
3043 int bucket)
3045 struct xfs_buf *ibp;
3046 struct xfs_dinode *dip;
3047 struct xfs_inode *ip;
3048 xfs_ino_t ino;
3049 int error;
3051 ino = XFS_AGINO_TO_INO(mp, agno, agino);
3052 error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
3053 if (error)
3054 goto fail;
3057 * Get the on disk inode to find the next inode in the bucket.
3059 error = xfs_itobp(mp, NULL, ip, &dip, &ibp, XBF_LOCK);
3060 if (error)
3061 goto fail_iput;
3063 ASSERT(ip->i_d.di_nlink == 0);
3064 ASSERT(ip->i_d.di_mode != 0);
3066 /* setup for the next pass */
3067 agino = be32_to_cpu(dip->di_next_unlinked);
3068 xfs_buf_relse(ibp);
3071 * Prevent any DMAPI event from being sent when the reference on
3072 * the inode is dropped.
3074 ip->i_d.di_dmevmask = 0;
3076 IRELE(ip);
3077 return agino;
3079 fail_iput:
3080 IRELE(ip);
3081 fail:
3083 * We can't read in the inode this bucket points to, or this inode
3084 * is messed up. Just ditch this bucket of inodes. We will lose
3085 * some inodes and space, but at least we won't hang.
3087 * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3088 * clear the inode pointer in the bucket.
3090 xlog_recover_clear_agi_bucket(mp, agno, bucket);
3091 return NULLAGINO;
3095 * xlog_iunlink_recover
3097 * This is called during recovery to process any inodes which
3098 * we unlinked but not freed when the system crashed. These
3099 * inodes will be on the lists in the AGI blocks. What we do
3100 * here is scan all the AGIs and fully truncate and free any
3101 * inodes found on the lists. Each inode is removed from the
3102 * lists when it has been fully truncated and is freed. The
3103 * freeing of the inode and its removal from the list must be
3104 * atomic.
3106 STATIC void
3107 xlog_recover_process_iunlinks(
3108 xlog_t *log)
3110 xfs_mount_t *mp;
3111 xfs_agnumber_t agno;
3112 xfs_agi_t *agi;
3113 xfs_buf_t *agibp;
3114 xfs_agino_t agino;
3115 int bucket;
3116 int error;
3117 uint mp_dmevmask;
3119 mp = log->l_mp;
3122 * Prevent any DMAPI event from being sent while in this function.
3124 mp_dmevmask = mp->m_dmevmask;
3125 mp->m_dmevmask = 0;
3127 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3129 * Find the agi for this ag.
3131 error = xfs_read_agi(mp, NULL, agno, &agibp);
3132 if (error) {
3134 * AGI is b0rked. Don't process it.
3136 * We should probably mark the filesystem as corrupt
3137 * after we've recovered all the ag's we can....
3139 continue;
3141 agi = XFS_BUF_TO_AGI(agibp);
3143 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3144 agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3145 while (agino != NULLAGINO) {
3147 * Release the agi buffer so that it can
3148 * be acquired in the normal course of the
3149 * transaction to truncate and free the inode.
3151 xfs_buf_relse(agibp);
3153 agino = xlog_recover_process_one_iunlink(mp,
3154 agno, agino, bucket);
3157 * Reacquire the agibuffer and continue around
3158 * the loop. This should never fail as we know
3159 * the buffer was good earlier on.
3161 error = xfs_read_agi(mp, NULL, agno, &agibp);
3162 ASSERT(error == 0);
3163 agi = XFS_BUF_TO_AGI(agibp);
3168 * Release the buffer for the current agi so we can
3169 * go on to the next one.
3171 xfs_buf_relse(agibp);
3174 mp->m_dmevmask = mp_dmevmask;
3178 #ifdef DEBUG
3179 STATIC void
3180 xlog_pack_data_checksum(
3181 xlog_t *log,
3182 xlog_in_core_t *iclog,
3183 int size)
3185 int i;
3186 __be32 *up;
3187 uint chksum = 0;
3189 up = (__be32 *)iclog->ic_datap;
3190 /* divide length by 4 to get # words */
3191 for (i = 0; i < (size >> 2); i++) {
3192 chksum ^= be32_to_cpu(*up);
3193 up++;
3195 iclog->ic_header.h_chksum = cpu_to_be32(chksum);
3197 #else
3198 #define xlog_pack_data_checksum(log, iclog, size)
3199 #endif
3202 * Stamp cycle number in every block
3204 void
3205 xlog_pack_data(
3206 xlog_t *log,
3207 xlog_in_core_t *iclog,
3208 int roundoff)
3210 int i, j, k;
3211 int size = iclog->ic_offset + roundoff;
3212 __be32 cycle_lsn;
3213 xfs_caddr_t dp;
3215 xlog_pack_data_checksum(log, iclog, size);
3217 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3219 dp = iclog->ic_datap;
3220 for (i = 0; i < BTOBB(size) &&
3221 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3222 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
3223 *(__be32 *)dp = cycle_lsn;
3224 dp += BBSIZE;
3227 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3228 xlog_in_core_2_t *xhdr = iclog->ic_data;
3230 for ( ; i < BTOBB(size); i++) {
3231 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3232 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3233 xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
3234 *(__be32 *)dp = cycle_lsn;
3235 dp += BBSIZE;
3238 for (i = 1; i < log->l_iclog_heads; i++) {
3239 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3244 STATIC void
3245 xlog_unpack_data(
3246 xlog_rec_header_t *rhead,
3247 xfs_caddr_t dp,
3248 xlog_t *log)
3250 int i, j, k;
3252 for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3253 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3254 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3255 dp += BBSIZE;
3258 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3259 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
3260 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3261 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3262 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3263 *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3264 dp += BBSIZE;
3269 STATIC int
3270 xlog_valid_rec_header(
3271 xlog_t *log,
3272 xlog_rec_header_t *rhead,
3273 xfs_daddr_t blkno)
3275 int hlen;
3277 if (unlikely(be32_to_cpu(rhead->h_magicno) != XLOG_HEADER_MAGIC_NUM)) {
3278 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3279 XFS_ERRLEVEL_LOW, log->l_mp);
3280 return XFS_ERROR(EFSCORRUPTED);
3282 if (unlikely(
3283 (!rhead->h_version ||
3284 (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
3285 xlog_warn("XFS: %s: unrecognised log version (%d).",
3286 __func__, be32_to_cpu(rhead->h_version));
3287 return XFS_ERROR(EIO);
3290 /* LR body must have data or it wouldn't have been written */
3291 hlen = be32_to_cpu(rhead->h_len);
3292 if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3293 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3294 XFS_ERRLEVEL_LOW, log->l_mp);
3295 return XFS_ERROR(EFSCORRUPTED);
3297 if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3298 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3299 XFS_ERRLEVEL_LOW, log->l_mp);
3300 return XFS_ERROR(EFSCORRUPTED);
3302 return 0;
3306 * Read the log from tail to head and process the log records found.
3307 * Handle the two cases where the tail and head are in the same cycle
3308 * and where the active portion of the log wraps around the end of
3309 * the physical log separately. The pass parameter is passed through
3310 * to the routines called to process the data and is not looked at
3311 * here.
3313 STATIC int
3314 xlog_do_recovery_pass(
3315 xlog_t *log,
3316 xfs_daddr_t head_blk,
3317 xfs_daddr_t tail_blk,
3318 int pass)
3320 xlog_rec_header_t *rhead;
3321 xfs_daddr_t blk_no;
3322 xfs_caddr_t offset;
3323 xfs_buf_t *hbp, *dbp;
3324 int error = 0, h_size;
3325 int bblks, split_bblks;
3326 int hblks, split_hblks, wrapped_hblks;
3327 struct hlist_head rhash[XLOG_RHASH_SIZE];
3329 ASSERT(head_blk != tail_blk);
3332 * Read the header of the tail block and get the iclog buffer size from
3333 * h_size. Use this to tell how many sectors make up the log header.
3335 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3337 * When using variable length iclogs, read first sector of
3338 * iclog header and extract the header size from it. Get a
3339 * new hbp that is the correct size.
3341 hbp = xlog_get_bp(log, 1);
3342 if (!hbp)
3343 return ENOMEM;
3345 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
3346 if (error)
3347 goto bread_err1;
3349 rhead = (xlog_rec_header_t *)offset;
3350 error = xlog_valid_rec_header(log, rhead, tail_blk);
3351 if (error)
3352 goto bread_err1;
3353 h_size = be32_to_cpu(rhead->h_size);
3354 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3355 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3356 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3357 if (h_size % XLOG_HEADER_CYCLE_SIZE)
3358 hblks++;
3359 xlog_put_bp(hbp);
3360 hbp = xlog_get_bp(log, hblks);
3361 } else {
3362 hblks = 1;
3364 } else {
3365 ASSERT(log->l_sectBBsize == 1);
3366 hblks = 1;
3367 hbp = xlog_get_bp(log, 1);
3368 h_size = XLOG_BIG_RECORD_BSIZE;
3371 if (!hbp)
3372 return ENOMEM;
3373 dbp = xlog_get_bp(log, BTOBB(h_size));
3374 if (!dbp) {
3375 xlog_put_bp(hbp);
3376 return ENOMEM;
3379 memset(rhash, 0, sizeof(rhash));
3380 if (tail_blk <= head_blk) {
3381 for (blk_no = tail_blk; blk_no < head_blk; ) {
3382 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3383 if (error)
3384 goto bread_err2;
3386 rhead = (xlog_rec_header_t *)offset;
3387 error = xlog_valid_rec_header(log, rhead, blk_no);
3388 if (error)
3389 goto bread_err2;
3391 /* blocks in data section */
3392 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3393 error = xlog_bread(log, blk_no + hblks, bblks, dbp,
3394 &offset);
3395 if (error)
3396 goto bread_err2;
3398 xlog_unpack_data(rhead, offset, log);
3399 if ((error = xlog_recover_process_data(log,
3400 rhash, rhead, offset, pass)))
3401 goto bread_err2;
3402 blk_no += bblks + hblks;
3404 } else {
3406 * Perform recovery around the end of the physical log.
3407 * When the head is not on the same cycle number as the tail,
3408 * we can't do a sequential recovery as above.
3410 blk_no = tail_blk;
3411 while (blk_no < log->l_logBBsize) {
3413 * Check for header wrapping around physical end-of-log
3415 offset = XFS_BUF_PTR(hbp);
3416 split_hblks = 0;
3417 wrapped_hblks = 0;
3418 if (blk_no + hblks <= log->l_logBBsize) {
3419 /* Read header in one read */
3420 error = xlog_bread(log, blk_no, hblks, hbp,
3421 &offset);
3422 if (error)
3423 goto bread_err2;
3424 } else {
3425 /* This LR is split across physical log end */
3426 if (blk_no != log->l_logBBsize) {
3427 /* some data before physical log end */
3428 ASSERT(blk_no <= INT_MAX);
3429 split_hblks = log->l_logBBsize - (int)blk_no;
3430 ASSERT(split_hblks > 0);
3431 error = xlog_bread(log, blk_no,
3432 split_hblks, hbp,
3433 &offset);
3434 if (error)
3435 goto bread_err2;
3439 * Note: this black magic still works with
3440 * large sector sizes (non-512) only because:
3441 * - we increased the buffer size originally
3442 * by 1 sector giving us enough extra space
3443 * for the second read;
3444 * - the log start is guaranteed to be sector
3445 * aligned;
3446 * - we read the log end (LR header start)
3447 * _first_, then the log start (LR header end)
3448 * - order is important.
3450 wrapped_hblks = hblks - split_hblks;
3451 error = XFS_BUF_SET_PTR(hbp,
3452 offset + BBTOB(split_hblks),
3453 BBTOB(hblks - split_hblks));
3454 if (error)
3455 goto bread_err2;
3457 error = xlog_bread_noalign(log, 0,
3458 wrapped_hblks, hbp);
3459 if (error)
3460 goto bread_err2;
3462 error = XFS_BUF_SET_PTR(hbp, offset,
3463 BBTOB(hblks));
3464 if (error)
3465 goto bread_err2;
3467 rhead = (xlog_rec_header_t *)offset;
3468 error = xlog_valid_rec_header(log, rhead,
3469 split_hblks ? blk_no : 0);
3470 if (error)
3471 goto bread_err2;
3473 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3474 blk_no += hblks;
3476 /* Read in data for log record */
3477 if (blk_no + bblks <= log->l_logBBsize) {
3478 error = xlog_bread(log, blk_no, bblks, dbp,
3479 &offset);
3480 if (error)
3481 goto bread_err2;
3482 } else {
3483 /* This log record is split across the
3484 * physical end of log */
3485 offset = XFS_BUF_PTR(dbp);
3486 split_bblks = 0;
3487 if (blk_no != log->l_logBBsize) {
3488 /* some data is before the physical
3489 * end of log */
3490 ASSERT(!wrapped_hblks);
3491 ASSERT(blk_no <= INT_MAX);
3492 split_bblks =
3493 log->l_logBBsize - (int)blk_no;
3494 ASSERT(split_bblks > 0);
3495 error = xlog_bread(log, blk_no,
3496 split_bblks, dbp,
3497 &offset);
3498 if (error)
3499 goto bread_err2;
3503 * Note: this black magic still works with
3504 * large sector sizes (non-512) only because:
3505 * - we increased the buffer size originally
3506 * by 1 sector giving us enough extra space
3507 * for the second read;
3508 * - the log start is guaranteed to be sector
3509 * aligned;
3510 * - we read the log end (LR header start)
3511 * _first_, then the log start (LR header end)
3512 * - order is important.
3514 error = XFS_BUF_SET_PTR(dbp,
3515 offset + BBTOB(split_bblks),
3516 BBTOB(bblks - split_bblks));
3517 if (error)
3518 goto bread_err2;
3520 error = xlog_bread_noalign(log, wrapped_hblks,
3521 bblks - split_bblks,
3522 dbp);
3523 if (error)
3524 goto bread_err2;
3526 error = XFS_BUF_SET_PTR(dbp, offset, h_size);
3527 if (error)
3528 goto bread_err2;
3530 xlog_unpack_data(rhead, offset, log);
3531 if ((error = xlog_recover_process_data(log, rhash,
3532 rhead, offset, pass)))
3533 goto bread_err2;
3534 blk_no += bblks;
3537 ASSERT(blk_no >= log->l_logBBsize);
3538 blk_no -= log->l_logBBsize;
3540 /* read first part of physical log */
3541 while (blk_no < head_blk) {
3542 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3543 if (error)
3544 goto bread_err2;
3546 rhead = (xlog_rec_header_t *)offset;
3547 error = xlog_valid_rec_header(log, rhead, blk_no);
3548 if (error)
3549 goto bread_err2;
3551 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3552 error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3553 &offset);
3554 if (error)
3555 goto bread_err2;
3557 xlog_unpack_data(rhead, offset, log);
3558 if ((error = xlog_recover_process_data(log, rhash,
3559 rhead, offset, pass)))
3560 goto bread_err2;
3561 blk_no += bblks + hblks;
3565 bread_err2:
3566 xlog_put_bp(dbp);
3567 bread_err1:
3568 xlog_put_bp(hbp);
3569 return error;
3573 * Do the recovery of the log. We actually do this in two phases.
3574 * The two passes are necessary in order to implement the function
3575 * of cancelling a record written into the log. The first pass
3576 * determines those things which have been cancelled, and the
3577 * second pass replays log items normally except for those which
3578 * have been cancelled. The handling of the replay and cancellations
3579 * takes place in the log item type specific routines.
3581 * The table of items which have cancel records in the log is allocated
3582 * and freed at this level, since only here do we know when all of
3583 * the log recovery has been completed.
3585 STATIC int
3586 xlog_do_log_recovery(
3587 xlog_t *log,
3588 xfs_daddr_t head_blk,
3589 xfs_daddr_t tail_blk)
3591 int error, i;
3593 ASSERT(head_blk != tail_blk);
3596 * First do a pass to find all of the cancelled buf log items.
3597 * Store them in the buf_cancel_table for use in the second pass.
3599 log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
3600 sizeof(struct list_head),
3601 KM_SLEEP);
3602 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3603 INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
3605 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3606 XLOG_RECOVER_PASS1);
3607 if (error != 0) {
3608 kmem_free(log->l_buf_cancel_table);
3609 log->l_buf_cancel_table = NULL;
3610 return error;
3613 * Then do a second pass to actually recover the items in the log.
3614 * When it is complete free the table of buf cancel items.
3616 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3617 XLOG_RECOVER_PASS2);
3618 #ifdef DEBUG
3619 if (!error) {
3620 int i;
3622 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3623 ASSERT(list_empty(&log->l_buf_cancel_table[i]));
3625 #endif /* DEBUG */
3627 kmem_free(log->l_buf_cancel_table);
3628 log->l_buf_cancel_table = NULL;
3630 return error;
3634 * Do the actual recovery
3636 STATIC int
3637 xlog_do_recover(
3638 xlog_t *log,
3639 xfs_daddr_t head_blk,
3640 xfs_daddr_t tail_blk)
3642 int error;
3643 xfs_buf_t *bp;
3644 xfs_sb_t *sbp;
3647 * First replay the images in the log.
3649 error = xlog_do_log_recovery(log, head_blk, tail_blk);
3650 if (error) {
3651 return error;
3654 XFS_bflush(log->l_mp->m_ddev_targp);
3657 * If IO errors happened during recovery, bail out.
3659 if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3660 return (EIO);
3664 * We now update the tail_lsn since much of the recovery has completed
3665 * and there may be space available to use. If there were no extent
3666 * or iunlinks, we can free up the entire log and set the tail_lsn to
3667 * be the last_sync_lsn. This was set in xlog_find_tail to be the
3668 * lsn of the last known good LR on disk. If there are extent frees
3669 * or iunlinks they will have some entries in the AIL; so we look at
3670 * the AIL to determine how to set the tail_lsn.
3672 xlog_assign_tail_lsn(log->l_mp);
3675 * Now that we've finished replaying all buffer and inode
3676 * updates, re-read in the superblock.
3678 bp = xfs_getsb(log->l_mp, 0);
3679 XFS_BUF_UNDONE(bp);
3680 ASSERT(!(XFS_BUF_ISWRITE(bp)));
3681 ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
3682 XFS_BUF_READ(bp);
3683 XFS_BUF_UNASYNC(bp);
3684 xfsbdstrat(log->l_mp, bp);
3685 error = xfs_buf_iowait(bp);
3686 if (error) {
3687 xfs_ioerror_alert("xlog_do_recover",
3688 log->l_mp, bp, XFS_BUF_ADDR(bp));
3689 ASSERT(0);
3690 xfs_buf_relse(bp);
3691 return error;
3694 /* Convert superblock from on-disk format */
3695 sbp = &log->l_mp->m_sb;
3696 xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
3697 ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3698 ASSERT(xfs_sb_good_version(sbp));
3699 xfs_buf_relse(bp);
3701 /* We've re-read the superblock so re-initialize per-cpu counters */
3702 xfs_icsb_reinit_counters(log->l_mp);
3704 xlog_recover_check_summary(log);
3706 /* Normal transactions can now occur */
3707 log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3708 return 0;
3712 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3714 * Return error or zero.
3717 xlog_recover(
3718 xlog_t *log)
3720 xfs_daddr_t head_blk, tail_blk;
3721 int error;
3723 /* find the tail of the log */
3724 if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3725 return error;
3727 if (tail_blk != head_blk) {
3728 /* There used to be a comment here:
3730 * disallow recovery on read-only mounts. note -- mount
3731 * checks for ENOSPC and turns it into an intelligent
3732 * error message.
3733 * ...but this is no longer true. Now, unless you specify
3734 * NORECOVERY (in which case this function would never be
3735 * called), we just go ahead and recover. We do this all
3736 * under the vfs layer, so we can get away with it unless
3737 * the device itself is read-only, in which case we fail.
3739 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3740 return error;
3743 cmn_err(CE_NOTE,
3744 "Starting XFS recovery on filesystem: %s (logdev: %s)",
3745 log->l_mp->m_fsname, log->l_mp->m_logname ?
3746 log->l_mp->m_logname : "internal");
3748 error = xlog_do_recover(log, head_blk, tail_blk);
3749 log->l_flags |= XLOG_RECOVERY_NEEDED;
3751 return error;
3755 * In the first part of recovery we replay inodes and buffers and build
3756 * up the list of extent free items which need to be processed. Here
3757 * we process the extent free items and clean up the on disk unlinked
3758 * inode lists. This is separated from the first part of recovery so
3759 * that the root and real-time bitmap inodes can be read in from disk in
3760 * between the two stages. This is necessary so that we can free space
3761 * in the real-time portion of the file system.
3764 xlog_recover_finish(
3765 xlog_t *log)
3768 * Now we're ready to do the transactions needed for the
3769 * rest of recovery. Start with completing all the extent
3770 * free intent records and then process the unlinked inode
3771 * lists. At this point, we essentially run in normal mode
3772 * except that we're still performing recovery actions
3773 * rather than accepting new requests.
3775 if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3776 int error;
3777 error = xlog_recover_process_efis(log);
3778 if (error) {
3779 cmn_err(CE_ALERT,
3780 "Failed to recover EFIs on filesystem: %s",
3781 log->l_mp->m_fsname);
3782 return error;
3785 * Sync the log to get all the EFIs out of the AIL.
3786 * This isn't absolutely necessary, but it helps in
3787 * case the unlink transactions would have problems
3788 * pushing the EFIs out of the way.
3790 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3792 xlog_recover_process_iunlinks(log);
3794 xlog_recover_check_summary(log);
3796 cmn_err(CE_NOTE,
3797 "Ending XFS recovery on filesystem: %s (logdev: %s)",
3798 log->l_mp->m_fsname, log->l_mp->m_logname ?
3799 log->l_mp->m_logname : "internal");
3800 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3801 } else {
3802 cmn_err(CE_DEBUG,
3803 "Ending clean XFS mount for filesystem: %s\n",
3804 log->l_mp->m_fsname);
3806 return 0;
3810 #if defined(DEBUG)
3812 * Read all of the agf and agi counters and check that they
3813 * are consistent with the superblock counters.
3815 void
3816 xlog_recover_check_summary(
3817 xlog_t *log)
3819 xfs_mount_t *mp;
3820 xfs_agf_t *agfp;
3821 xfs_buf_t *agfbp;
3822 xfs_buf_t *agibp;
3823 xfs_agnumber_t agno;
3824 __uint64_t freeblks;
3825 __uint64_t itotal;
3826 __uint64_t ifree;
3827 int error;
3829 mp = log->l_mp;
3831 freeblks = 0LL;
3832 itotal = 0LL;
3833 ifree = 0LL;
3834 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3835 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
3836 if (error) {
3837 xfs_fs_cmn_err(CE_ALERT, mp,
3838 "xlog_recover_check_summary(agf)"
3839 "agf read failed agno %d error %d",
3840 agno, error);
3841 } else {
3842 agfp = XFS_BUF_TO_AGF(agfbp);
3843 freeblks += be32_to_cpu(agfp->agf_freeblks) +
3844 be32_to_cpu(agfp->agf_flcount);
3845 xfs_buf_relse(agfbp);
3848 error = xfs_read_agi(mp, NULL, agno, &agibp);
3849 if (!error) {
3850 struct xfs_agi *agi = XFS_BUF_TO_AGI(agibp);
3852 itotal += be32_to_cpu(agi->agi_count);
3853 ifree += be32_to_cpu(agi->agi_freecount);
3854 xfs_buf_relse(agibp);
3858 #endif /* DEBUG */