blob: 9824b5bf0ec03ea6af15803978106adeedc5a1d9 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Copyright (c) 2000-2003 Silicon Graphics, Inc. All Rights Reserved.
3 *
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms of version 2 of the GNU General Public License as
6 * published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it would be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
11 *
12 * Further, this software is distributed without any warranty that it is
13 * free of the rightful claim of any third person regarding infringement
14 * or the like. Any license provided herein, whether implied or
15 * otherwise, applies only to this software file. Patent licenses, if
16 * any, provided herein do not apply to combinations of this program with
17 * other software, or any other product whatsoever.
18 *
19 * You should have received a copy of the GNU General Public License along
20 * with this program; if not, write the Free Software Foundation, Inc., 59
21 * Temple Place - Suite 330, Boston MA 02111-1307, USA.
22 *
23 * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
24 * Mountain View, CA 94043, or:
25 *
26 * http://www.sgi.com
27 *
28 * For further information regarding this notice, see:
29 *
30 * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
31 */
32
33#include "xfs.h"
34#include "xfs_macros.h"
35#include "xfs_types.h"
36#include "xfs_inum.h"
37#include "xfs_log.h"
38#include "xfs_ag.h"
39#include "xfs_sb.h"
40#include "xfs_trans.h"
41#include "xfs_dir.h"
42#include "xfs_dir2.h"
43#include "xfs_dmapi.h"
44#include "xfs_mount.h"
45#include "xfs_error.h"
46#include "xfs_bmap_btree.h"
47#include "xfs_alloc.h"
48#include "xfs_attr_sf.h"
49#include "xfs_dir_sf.h"
50#include "xfs_dir2_sf.h"
51#include "xfs_dinode.h"
52#include "xfs_imap.h"
53#include "xfs_inode_item.h"
54#include "xfs_inode.h"
55#include "xfs_ialloc_btree.h"
56#include "xfs_ialloc.h"
57#include "xfs_log_priv.h"
58#include "xfs_buf_item.h"
59#include "xfs_alloc_btree.h"
60#include "xfs_log_recover.h"
61#include "xfs_extfree_item.h"
62#include "xfs_trans_priv.h"
63#include "xfs_bit.h"
64#include "xfs_quota.h"
65#include "xfs_rw.h"
66
67STATIC int xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
68STATIC int xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
69STATIC void xlog_recover_insert_item_backq(xlog_recover_item_t **q,
70 xlog_recover_item_t *item);
71#if defined(DEBUG)
72STATIC void xlog_recover_check_summary(xlog_t *);
73STATIC void xlog_recover_check_ail(xfs_mount_t *, xfs_log_item_t *, int);
74#else
75#define xlog_recover_check_summary(log)
76#define xlog_recover_check_ail(mp, lip, gen)
77#endif
78
79
80/*
81 * Sector aligned buffer routines for buffer create/read/write/access
82 */
83
84#define XLOG_SECTOR_ROUNDUP_BBCOUNT(log, bbs) \
85 ( ((log)->l_sectbb_mask && (bbs & (log)->l_sectbb_mask)) ? \
86 ((bbs + (log)->l_sectbb_mask + 1) & ~(log)->l_sectbb_mask) : (bbs) )
87#define XLOG_SECTOR_ROUNDDOWN_BLKNO(log, bno) ((bno) & ~(log)->l_sectbb_mask)
88
89xfs_buf_t *
90xlog_get_bp(
91 xlog_t *log,
92 int num_bblks)
93{
94 ASSERT(num_bblks > 0);
95
96 if (log->l_sectbb_log) {
97 if (num_bblks > 1)
98 num_bblks += XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
99 num_bblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, num_bblks);
100 }
101 return xfs_buf_get_noaddr(BBTOB(num_bblks), log->l_mp->m_logdev_targp);
102}
103
104void
105xlog_put_bp(
106 xfs_buf_t *bp)
107{
108 xfs_buf_free(bp);
109}
110
111
112/*
113 * nbblks should be uint, but oh well. Just want to catch that 32-bit length.
114 */
115int
116xlog_bread(
117 xlog_t *log,
118 xfs_daddr_t blk_no,
119 int nbblks,
120 xfs_buf_t *bp)
121{
122 int error;
123
124 if (log->l_sectbb_log) {
125 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
126 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
127 }
128
129 ASSERT(nbblks > 0);
130 ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
131 ASSERT(bp);
132
133 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
134 XFS_BUF_READ(bp);
135 XFS_BUF_BUSY(bp);
136 XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
137 XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
138
139 xfsbdstrat(log->l_mp, bp);
140 if ((error = xfs_iowait(bp)))
141 xfs_ioerror_alert("xlog_bread", log->l_mp,
142 bp, XFS_BUF_ADDR(bp));
143 return error;
144}
145
146/*
147 * Write out the buffer at the given block for the given number of blocks.
148 * The buffer is kept locked across the write and is returned locked.
149 * This can only be used for synchronous log writes.
150 */
151int
152xlog_bwrite(
153 xlog_t *log,
154 xfs_daddr_t blk_no,
155 int nbblks,
156 xfs_buf_t *bp)
157{
158 int error;
159
160 if (log->l_sectbb_log) {
161 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
162 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
163 }
164
165 ASSERT(nbblks > 0);
166 ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
167
168 XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
169 XFS_BUF_ZEROFLAGS(bp);
170 XFS_BUF_BUSY(bp);
171 XFS_BUF_HOLD(bp);
172 XFS_BUF_PSEMA(bp, PRIBIO);
173 XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
174 XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
175
176 if ((error = xfs_bwrite(log->l_mp, bp)))
177 xfs_ioerror_alert("xlog_bwrite", log->l_mp,
178 bp, XFS_BUF_ADDR(bp));
179 return error;
180}
181
182xfs_caddr_t
183xlog_align(
184 xlog_t *log,
185 xfs_daddr_t blk_no,
186 int nbblks,
187 xfs_buf_t *bp)
188{
189 xfs_caddr_t ptr;
190
191 if (!log->l_sectbb_log)
192 return XFS_BUF_PTR(bp);
193
194 ptr = XFS_BUF_PTR(bp) + BBTOB((int)blk_no & log->l_sectbb_mask);
195 ASSERT(XFS_BUF_SIZE(bp) >=
196 BBTOB(nbblks + (blk_no & log->l_sectbb_mask)));
197 return ptr;
198}
199
200#ifdef DEBUG
201/*
202 * dump debug superblock and log record information
203 */
204STATIC void
205xlog_header_check_dump(
206 xfs_mount_t *mp,
207 xlog_rec_header_t *head)
208{
209 int b;
210
211 printk("%s: SB : uuid = ", __FUNCTION__);
212 for (b = 0; b < 16; b++)
213 printk("%02x",((unsigned char *)&mp->m_sb.sb_uuid)[b]);
214 printk(", fmt = %d\n", XLOG_FMT);
215 printk(" log : uuid = ");
216 for (b = 0; b < 16; b++)
217 printk("%02x",((unsigned char *)&head->h_fs_uuid)[b]);
218 printk(", fmt = %d\n", INT_GET(head->h_fmt, ARCH_CONVERT));
219}
220#else
221#define xlog_header_check_dump(mp, head)
222#endif
223
224/*
225 * check log record header for recovery
226 */
227STATIC int
228xlog_header_check_recover(
229 xfs_mount_t *mp,
230 xlog_rec_header_t *head)
231{
232 ASSERT(INT_GET(head->h_magicno, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM);
233
234 /*
235 * IRIX doesn't write the h_fmt field and leaves it zeroed
236 * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
237 * a dirty log created in IRIX.
238 */
239 if (unlikely(INT_GET(head->h_fmt, ARCH_CONVERT) != XLOG_FMT)) {
240 xlog_warn(
241 "XFS: dirty log written in incompatible format - can't recover");
242 xlog_header_check_dump(mp, head);
243 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
244 XFS_ERRLEVEL_HIGH, mp);
245 return XFS_ERROR(EFSCORRUPTED);
246 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
247 xlog_warn(
248 "XFS: dirty log entry has mismatched uuid - can't recover");
249 xlog_header_check_dump(mp, head);
250 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
251 XFS_ERRLEVEL_HIGH, mp);
252 return XFS_ERROR(EFSCORRUPTED);
253 }
254 return 0;
255}
256
257/*
258 * read the head block of the log and check the header
259 */
260STATIC int
261xlog_header_check_mount(
262 xfs_mount_t *mp,
263 xlog_rec_header_t *head)
264{
265 ASSERT(INT_GET(head->h_magicno, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM);
266
267 if (uuid_is_nil(&head->h_fs_uuid)) {
268 /*
269 * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
270 * h_fs_uuid is nil, we assume this log was last mounted
271 * by IRIX and continue.
272 */
273 xlog_warn("XFS: nil uuid in log - IRIX style log");
274 } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
275 xlog_warn("XFS: log has mismatched uuid - can't recover");
276 xlog_header_check_dump(mp, head);
277 XFS_ERROR_REPORT("xlog_header_check_mount",
278 XFS_ERRLEVEL_HIGH, mp);
279 return XFS_ERROR(EFSCORRUPTED);
280 }
281 return 0;
282}
283
284STATIC void
285xlog_recover_iodone(
286 struct xfs_buf *bp)
287{
288 xfs_mount_t *mp;
289
290 ASSERT(XFS_BUF_FSPRIVATE(bp, void *));
291
292 if (XFS_BUF_GETERROR(bp)) {
293 /*
294 * We're not going to bother about retrying
295 * this during recovery. One strike!
296 */
297 mp = XFS_BUF_FSPRIVATE(bp, xfs_mount_t *);
298 xfs_ioerror_alert("xlog_recover_iodone",
299 mp, bp, XFS_BUF_ADDR(bp));
300 xfs_force_shutdown(mp, XFS_METADATA_IO_ERROR);
301 }
302 XFS_BUF_SET_FSPRIVATE(bp, NULL);
303 XFS_BUF_CLR_IODONE_FUNC(bp);
304 xfs_biodone(bp);
305}
306
307/*
308 * This routine finds (to an approximation) the first block in the physical
309 * log which contains the given cycle. It uses a binary search algorithm.
310 * Note that the algorithm can not be perfect because the disk will not
311 * necessarily be perfect.
312 */
313int
314xlog_find_cycle_start(
315 xlog_t *log,
316 xfs_buf_t *bp,
317 xfs_daddr_t first_blk,
318 xfs_daddr_t *last_blk,
319 uint cycle)
320{
321 xfs_caddr_t offset;
322 xfs_daddr_t mid_blk;
323 uint mid_cycle;
324 int error;
325
326 mid_blk = BLK_AVG(first_blk, *last_blk);
327 while (mid_blk != first_blk && mid_blk != *last_blk) {
328 if ((error = xlog_bread(log, mid_blk, 1, bp)))
329 return error;
330 offset = xlog_align(log, mid_blk, 1, bp);
331 mid_cycle = GET_CYCLE(offset, ARCH_CONVERT);
332 if (mid_cycle == cycle) {
333 *last_blk = mid_blk;
334 /* last_half_cycle == mid_cycle */
335 } else {
336 first_blk = mid_blk;
337 /* first_half_cycle == mid_cycle */
338 }
339 mid_blk = BLK_AVG(first_blk, *last_blk);
340 }
341 ASSERT((mid_blk == first_blk && mid_blk+1 == *last_blk) ||
342 (mid_blk == *last_blk && mid_blk-1 == first_blk));
343
344 return 0;
345}
346
347/*
348 * Check that the range of blocks does not contain the cycle number
349 * given. The scan needs to occur from front to back and the ptr into the
350 * region must be updated since a later routine will need to perform another
351 * test. If the region is completely good, we end up returning the same
352 * last block number.
353 *
354 * Set blkno to -1 if we encounter no errors. This is an invalid block number
355 * since we don't ever expect logs to get this large.
356 */
357STATIC int
358xlog_find_verify_cycle(
359 xlog_t *log,
360 xfs_daddr_t start_blk,
361 int nbblks,
362 uint stop_on_cycle_no,
363 xfs_daddr_t *new_blk)
364{
365 xfs_daddr_t i, j;
366 uint cycle;
367 xfs_buf_t *bp;
368 xfs_daddr_t bufblks;
369 xfs_caddr_t buf = NULL;
370 int error = 0;
371
372 bufblks = 1 << ffs(nbblks);
373
374 while (!(bp = xlog_get_bp(log, bufblks))) {
375 /* can't get enough memory to do everything in one big buffer */
376 bufblks >>= 1;
377 if (bufblks <= log->l_sectbb_log)
378 return ENOMEM;
379 }
380
381 for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
382 int bcount;
383
384 bcount = min(bufblks, (start_blk + nbblks - i));
385
386 if ((error = xlog_bread(log, i, bcount, bp)))
387 goto out;
388
389 buf = xlog_align(log, i, bcount, bp);
390 for (j = 0; j < bcount; j++) {
391 cycle = GET_CYCLE(buf, ARCH_CONVERT);
392 if (cycle == stop_on_cycle_no) {
393 *new_blk = i+j;
394 goto out;
395 }
396
397 buf += BBSIZE;
398 }
399 }
400
401 *new_blk = -1;
402
403out:
404 xlog_put_bp(bp);
405 return error;
406}
407
408/*
409 * Potentially backup over partial log record write.
410 *
411 * In the typical case, last_blk is the number of the block directly after
412 * a good log record. Therefore, we subtract one to get the block number
413 * of the last block in the given buffer. extra_bblks contains the number
414 * of blocks we would have read on a previous read. This happens when the
415 * last log record is split over the end of the physical log.
416 *
417 * extra_bblks is the number of blocks potentially verified on a previous
418 * call to this routine.
419 */
420STATIC int
421xlog_find_verify_log_record(
422 xlog_t *log,
423 xfs_daddr_t start_blk,
424 xfs_daddr_t *last_blk,
425 int extra_bblks)
426{
427 xfs_daddr_t i;
428 xfs_buf_t *bp;
429 xfs_caddr_t offset = NULL;
430 xlog_rec_header_t *head = NULL;
431 int error = 0;
432 int smallmem = 0;
433 int num_blks = *last_blk - start_blk;
434 int xhdrs;
435
436 ASSERT(start_blk != 0 || *last_blk != start_blk);
437
438 if (!(bp = xlog_get_bp(log, num_blks))) {
439 if (!(bp = xlog_get_bp(log, 1)))
440 return ENOMEM;
441 smallmem = 1;
442 } else {
443 if ((error = xlog_bread(log, start_blk, num_blks, bp)))
444 goto out;
445 offset = xlog_align(log, start_blk, num_blks, bp);
446 offset += ((num_blks - 1) << BBSHIFT);
447 }
448
449 for (i = (*last_blk) - 1; i >= 0; i--) {
450 if (i < start_blk) {
451 /* valid log record not found */
452 xlog_warn(
453 "XFS: Log inconsistent (didn't find previous header)");
454 ASSERT(0);
455 error = XFS_ERROR(EIO);
456 goto out;
457 }
458
459 if (smallmem) {
460 if ((error = xlog_bread(log, i, 1, bp)))
461 goto out;
462 offset = xlog_align(log, i, 1, bp);
463 }
464
465 head = (xlog_rec_header_t *)offset;
466
467 if (XLOG_HEADER_MAGIC_NUM ==
468 INT_GET(head->h_magicno, ARCH_CONVERT))
469 break;
470
471 if (!smallmem)
472 offset -= BBSIZE;
473 }
474
475 /*
476 * We hit the beginning of the physical log & still no header. Return
477 * to caller. If caller can handle a return of -1, then this routine
478 * will be called again for the end of the physical log.
479 */
480 if (i == -1) {
481 error = -1;
482 goto out;
483 }
484
485 /*
486 * We have the final block of the good log (the first block
487 * of the log record _before_ the head. So we check the uuid.
488 */
489 if ((error = xlog_header_check_mount(log->l_mp, head)))
490 goto out;
491
492 /*
493 * We may have found a log record header before we expected one.
494 * last_blk will be the 1st block # with a given cycle #. We may end
495 * up reading an entire log record. In this case, we don't want to
496 * reset last_blk. Only when last_blk points in the middle of a log
497 * record do we update last_blk.
498 */
499 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
500 uint h_size = INT_GET(head->h_size, ARCH_CONVERT);
501
502 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
503 if (h_size % XLOG_HEADER_CYCLE_SIZE)
504 xhdrs++;
505 } else {
506 xhdrs = 1;
507 }
508
509 if (*last_blk - i + extra_bblks
510 != BTOBB(INT_GET(head->h_len, ARCH_CONVERT)) + xhdrs)
511 *last_blk = i;
512
513out:
514 xlog_put_bp(bp);
515 return error;
516}
517
518/*
519 * Head is defined to be the point of the log where the next log write
520 * write could go. This means that incomplete LR writes at the end are
521 * eliminated when calculating the head. We aren't guaranteed that previous
522 * LR have complete transactions. We only know that a cycle number of
523 * current cycle number -1 won't be present in the log if we start writing
524 * from our current block number.
525 *
526 * last_blk contains the block number of the first block with a given
527 * cycle number.
528 *
529 * Return: zero if normal, non-zero if error.
530 */
531int
532xlog_find_head(
533 xlog_t *log,
534 xfs_daddr_t *return_head_blk)
535{
536 xfs_buf_t *bp;
537 xfs_caddr_t offset;
538 xfs_daddr_t new_blk, first_blk, start_blk, last_blk, head_blk;
539 int num_scan_bblks;
540 uint first_half_cycle, last_half_cycle;
541 uint stop_on_cycle;
542 int error, log_bbnum = log->l_logBBsize;
543
544 /* Is the end of the log device zeroed? */
545 if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
546 *return_head_blk = first_blk;
547
548 /* Is the whole lot zeroed? */
549 if (!first_blk) {
550 /* Linux XFS shouldn't generate totally zeroed logs -
551 * mkfs etc write a dummy unmount record to a fresh
552 * log so we can store the uuid in there
553 */
554 xlog_warn("XFS: totally zeroed log");
555 }
556
557 return 0;
558 } else if (error) {
559 xlog_warn("XFS: empty log check failed");
560 return error;
561 }
562
563 first_blk = 0; /* get cycle # of 1st block */
564 bp = xlog_get_bp(log, 1);
565 if (!bp)
566 return ENOMEM;
567 if ((error = xlog_bread(log, 0, 1, bp)))
568 goto bp_err;
569 offset = xlog_align(log, 0, 1, bp);
570 first_half_cycle = GET_CYCLE(offset, ARCH_CONVERT);
571
572 last_blk = head_blk = log_bbnum - 1; /* get cycle # of last block */
573 if ((error = xlog_bread(log, last_blk, 1, bp)))
574 goto bp_err;
575 offset = xlog_align(log, last_blk, 1, bp);
576 last_half_cycle = GET_CYCLE(offset, ARCH_CONVERT);
577 ASSERT(last_half_cycle != 0);
578
579 /*
580 * If the 1st half cycle number is equal to the last half cycle number,
581 * then the entire log is stamped with the same cycle number. In this
582 * case, head_blk can't be set to zero (which makes sense). The below
583 * math doesn't work out properly with head_blk equal to zero. Instead,
584 * we set it to log_bbnum which is an invalid block number, but this
585 * value makes the math correct. If head_blk doesn't changed through
586 * all the tests below, *head_blk is set to zero at the very end rather
587 * than log_bbnum. In a sense, log_bbnum and zero are the same block
588 * in a circular file.
589 */
590 if (first_half_cycle == last_half_cycle) {
591 /*
592 * In this case we believe that the entire log should have
593 * cycle number last_half_cycle. We need to scan backwards
594 * from the end verifying that there are no holes still
595 * containing last_half_cycle - 1. If we find such a hole,
596 * then the start of that hole will be the new head. The
597 * simple case looks like
598 * x | x ... | x - 1 | x
599 * Another case that fits this picture would be
600 * x | x + 1 | x ... | x
601 * In this case the head really is somwhere at the end of the
602 * log, as one of the latest writes at the beginning was
603 * incomplete.
604 * One more case is
605 * x | x + 1 | x ... | x - 1 | x
606 * This is really the combination of the above two cases, and
607 * the head has to end up at the start of the x-1 hole at the
608 * end of the log.
609 *
610 * In the 256k log case, we will read from the beginning to the
611 * end of the log and search for cycle numbers equal to x-1.
612 * We don't worry about the x+1 blocks that we encounter,
613 * because we know that they cannot be the head since the log
614 * started with x.
615 */
616 head_blk = log_bbnum;
617 stop_on_cycle = last_half_cycle - 1;
618 } else {
619 /*
620 * In this case we want to find the first block with cycle
621 * number matching last_half_cycle. We expect the log to be
622 * some variation on
623 * x + 1 ... | x ...
624 * The first block with cycle number x (last_half_cycle) will
625 * be where the new head belongs. First we do a binary search
626 * for the first occurrence of last_half_cycle. The binary
627 * search may not be totally accurate, so then we scan back
628 * from there looking for occurrences of last_half_cycle before
629 * us. If that backwards scan wraps around the beginning of
630 * the log, then we look for occurrences of last_half_cycle - 1
631 * at the end of the log. The cases we're looking for look
632 * like
633 * x + 1 ... | x | x + 1 | x ...
634 * ^ binary search stopped here
635 * or
636 * x + 1 ... | x ... | x - 1 | x
637 * <---------> less than scan distance
638 */
639 stop_on_cycle = last_half_cycle;
640 if ((error = xlog_find_cycle_start(log, bp, first_blk,
641 &head_blk, last_half_cycle)))
642 goto bp_err;
643 }
644
645 /*
646 * Now validate the answer. Scan back some number of maximum possible
647 * blocks and make sure each one has the expected cycle number. The
648 * maximum is determined by the total possible amount of buffering
649 * in the in-core log. The following number can be made tighter if
650 * we actually look at the block size of the filesystem.
651 */
652 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
653 if (head_blk >= num_scan_bblks) {
654 /*
655 * We are guaranteed that the entire check can be performed
656 * in one buffer.
657 */
658 start_blk = head_blk - num_scan_bblks;
659 if ((error = xlog_find_verify_cycle(log,
660 start_blk, num_scan_bblks,
661 stop_on_cycle, &new_blk)))
662 goto bp_err;
663 if (new_blk != -1)
664 head_blk = new_blk;
665 } else { /* need to read 2 parts of log */
666 /*
667 * We are going to scan backwards in the log in two parts.
668 * First we scan the physical end of the log. In this part
669 * of the log, we are looking for blocks with cycle number
670 * last_half_cycle - 1.
671 * If we find one, then we know that the log starts there, as
672 * we've found a hole that didn't get written in going around
673 * the end of the physical log. The simple case for this is
674 * x + 1 ... | x ... | x - 1 | x
675 * <---------> less than scan distance
676 * If all of the blocks at the end of the log have cycle number
677 * last_half_cycle, then we check the blocks at the start of
678 * the log looking for occurrences of last_half_cycle. If we
679 * find one, then our current estimate for the location of the
680 * first occurrence of last_half_cycle is wrong and we move
681 * back to the hole we've found. This case looks like
682 * x + 1 ... | x | x + 1 | x ...
683 * ^ binary search stopped here
684 * Another case we need to handle that only occurs in 256k
685 * logs is
686 * x + 1 ... | x ... | x+1 | x ...
687 * ^ binary search stops here
688 * In a 256k log, the scan at the end of the log will see the
689 * x + 1 blocks. We need to skip past those since that is
690 * certainly not the head of the log. By searching for
691 * last_half_cycle-1 we accomplish that.
692 */
693 start_blk = log_bbnum - num_scan_bblks + head_blk;
694 ASSERT(head_blk <= INT_MAX &&
695 (xfs_daddr_t) num_scan_bblks - head_blk >= 0);
696 if ((error = xlog_find_verify_cycle(log, start_blk,
697 num_scan_bblks - (int)head_blk,
698 (stop_on_cycle - 1), &new_blk)))
699 goto bp_err;
700 if (new_blk != -1) {
701 head_blk = new_blk;
702 goto bad_blk;
703 }
704
705 /*
706 * Scan beginning of log now. The last part of the physical
707 * log is good. This scan needs to verify that it doesn't find
708 * the last_half_cycle.
709 */
710 start_blk = 0;
711 ASSERT(head_blk <= INT_MAX);
712 if ((error = xlog_find_verify_cycle(log,
713 start_blk, (int)head_blk,
714 stop_on_cycle, &new_blk)))
715 goto bp_err;
716 if (new_blk != -1)
717 head_blk = new_blk;
718 }
719
720 bad_blk:
721 /*
722 * Now we need to make sure head_blk is not pointing to a block in
723 * the middle of a log record.
724 */
725 num_scan_bblks = XLOG_REC_SHIFT(log);
726 if (head_blk >= num_scan_bblks) {
727 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
728
729 /* start ptr at last block ptr before head_blk */
730 if ((error = xlog_find_verify_log_record(log, start_blk,
731 &head_blk, 0)) == -1) {
732 error = XFS_ERROR(EIO);
733 goto bp_err;
734 } else if (error)
735 goto bp_err;
736 } else {
737 start_blk = 0;
738 ASSERT(head_blk <= INT_MAX);
739 if ((error = xlog_find_verify_log_record(log, start_blk,
740 &head_blk, 0)) == -1) {
741 /* We hit the beginning of the log during our search */
742 start_blk = log_bbnum - num_scan_bblks + head_blk;
743 new_blk = log_bbnum;
744 ASSERT(start_blk <= INT_MAX &&
745 (xfs_daddr_t) log_bbnum-start_blk >= 0);
746 ASSERT(head_blk <= INT_MAX);
747 if ((error = xlog_find_verify_log_record(log,
748 start_blk, &new_blk,
749 (int)head_blk)) == -1) {
750 error = XFS_ERROR(EIO);
751 goto bp_err;
752 } else if (error)
753 goto bp_err;
754 if (new_blk != log_bbnum)
755 head_blk = new_blk;
756 } else if (error)
757 goto bp_err;
758 }
759
760 xlog_put_bp(bp);
761 if (head_blk == log_bbnum)
762 *return_head_blk = 0;
763 else
764 *return_head_blk = head_blk;
765 /*
766 * When returning here, we have a good block number. Bad block
767 * means that during a previous crash, we didn't have a clean break
768 * from cycle number N to cycle number N-1. In this case, we need
769 * to find the first block with cycle number N-1.
770 */
771 return 0;
772
773 bp_err:
774 xlog_put_bp(bp);
775
776 if (error)
777 xlog_warn("XFS: failed to find log head");
778 return error;
779}
780
781/*
782 * Find the sync block number or the tail of the log.
783 *
784 * This will be the block number of the last record to have its
785 * associated buffers synced to disk. Every log record header has
786 * a sync lsn embedded in it. LSNs hold block numbers, so it is easy
787 * to get a sync block number. The only concern is to figure out which
788 * log record header to believe.
789 *
790 * The following algorithm uses the log record header with the largest
791 * lsn. The entire log record does not need to be valid. We only care
792 * that the header is valid.
793 *
794 * We could speed up search by using current head_blk buffer, but it is not
795 * available.
796 */
797int
798xlog_find_tail(
799 xlog_t *log,
800 xfs_daddr_t *head_blk,
801 xfs_daddr_t *tail_blk,
802 int readonly)
803{
804 xlog_rec_header_t *rhead;
805 xlog_op_header_t *op_head;
806 xfs_caddr_t offset = NULL;
807 xfs_buf_t *bp;
808 int error, i, found;
809 xfs_daddr_t umount_data_blk;
810 xfs_daddr_t after_umount_blk;
811 xfs_lsn_t tail_lsn;
812 int hblks;
813
814 found = 0;
815
816 /*
817 * Find previous log record
818 */
819 if ((error = xlog_find_head(log, head_blk)))
820 return error;
821
822 bp = xlog_get_bp(log, 1);
823 if (!bp)
824 return ENOMEM;
825 if (*head_blk == 0) { /* special case */
826 if ((error = xlog_bread(log, 0, 1, bp)))
827 goto bread_err;
828 offset = xlog_align(log, 0, 1, bp);
829 if (GET_CYCLE(offset, ARCH_CONVERT) == 0) {
830 *tail_blk = 0;
831 /* leave all other log inited values alone */
832 goto exit;
833 }
834 }
835
836 /*
837 * Search backwards looking for log record header block
838 */
839 ASSERT(*head_blk < INT_MAX);
840 for (i = (int)(*head_blk) - 1; i >= 0; i--) {
841 if ((error = xlog_bread(log, i, 1, bp)))
842 goto bread_err;
843 offset = xlog_align(log, i, 1, bp);
844 if (XLOG_HEADER_MAGIC_NUM ==
845 INT_GET(*(uint *)offset, ARCH_CONVERT)) {
846 found = 1;
847 break;
848 }
849 }
850 /*
851 * If we haven't found the log record header block, start looking
852 * again from the end of the physical log. XXXmiken: There should be
853 * a check here to make sure we didn't search more than N blocks in
854 * the previous code.
855 */
856 if (!found) {
857 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
858 if ((error = xlog_bread(log, i, 1, bp)))
859 goto bread_err;
860 offset = xlog_align(log, i, 1, bp);
861 if (XLOG_HEADER_MAGIC_NUM ==
862 INT_GET(*(uint*)offset, ARCH_CONVERT)) {
863 found = 2;
864 break;
865 }
866 }
867 }
868 if (!found) {
869 xlog_warn("XFS: xlog_find_tail: couldn't find sync record");
870 ASSERT(0);
871 return XFS_ERROR(EIO);
872 }
873
874 /* find blk_no of tail of log */
875 rhead = (xlog_rec_header_t *)offset;
876 *tail_blk = BLOCK_LSN(INT_GET(rhead->h_tail_lsn, ARCH_CONVERT));
877
878 /*
879 * Reset log values according to the state of the log when we
880 * crashed. In the case where head_blk == 0, we bump curr_cycle
881 * one because the next write starts a new cycle rather than
882 * continuing the cycle of the last good log record. At this
883 * point we have guaranteed that all partial log records have been
884 * accounted for. Therefore, we know that the last good log record
885 * written was complete and ended exactly on the end boundary
886 * of the physical log.
887 */
888 log->l_prev_block = i;
889 log->l_curr_block = (int)*head_blk;
890 log->l_curr_cycle = INT_GET(rhead->h_cycle, ARCH_CONVERT);
891 if (found == 2)
892 log->l_curr_cycle++;
893 log->l_tail_lsn = INT_GET(rhead->h_tail_lsn, ARCH_CONVERT);
894 log->l_last_sync_lsn = INT_GET(rhead->h_lsn, ARCH_CONVERT);
895 log->l_grant_reserve_cycle = log->l_curr_cycle;
896 log->l_grant_reserve_bytes = BBTOB(log->l_curr_block);
897 log->l_grant_write_cycle = log->l_curr_cycle;
898 log->l_grant_write_bytes = BBTOB(log->l_curr_block);
899
900 /*
901 * Look for unmount record. If we find it, then we know there
902 * was a clean unmount. Since 'i' could be the last block in
903 * the physical log, we convert to a log block before comparing
904 * to the head_blk.
905 *
906 * Save the current tail lsn to use to pass to
907 * xlog_clear_stale_blocks() below. We won't want to clear the
908 * unmount record if there is one, so we pass the lsn of the
909 * unmount record rather than the block after it.
910 */
911 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
912 int h_size = INT_GET(rhead->h_size, ARCH_CONVERT);
913 int h_version = INT_GET(rhead->h_version, ARCH_CONVERT);
914
915 if ((h_version & XLOG_VERSION_2) &&
916 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
917 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
918 if (h_size % XLOG_HEADER_CYCLE_SIZE)
919 hblks++;
920 } else {
921 hblks = 1;
922 }
923 } else {
924 hblks = 1;
925 }
926 after_umount_blk = (i + hblks + (int)
927 BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT))) % log->l_logBBsize;
928 tail_lsn = log->l_tail_lsn;
929 if (*head_blk == after_umount_blk &&
930 INT_GET(rhead->h_num_logops, ARCH_CONVERT) == 1) {
931 umount_data_blk = (i + hblks) % log->l_logBBsize;
932 if ((error = xlog_bread(log, umount_data_blk, 1, bp))) {
933 goto bread_err;
934 }
935 offset = xlog_align(log, umount_data_blk, 1, bp);
936 op_head = (xlog_op_header_t *)offset;
937 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
938 /*
939 * Set tail and last sync so that newly written
940 * log records will point recovery to after the
941 * current unmount record.
942 */
943 ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, log->l_curr_cycle,
944 after_umount_blk);
945 ASSIGN_ANY_LSN_HOST(log->l_last_sync_lsn, log->l_curr_cycle,
946 after_umount_blk);
947 *tail_blk = after_umount_blk;
948 }
949 }
950
951 /*
952 * Make sure that there are no blocks in front of the head
953 * with the same cycle number as the head. This can happen
954 * because we allow multiple outstanding log writes concurrently,
955 * and the later writes might make it out before earlier ones.
956 *
957 * We use the lsn from before modifying it so that we'll never
958 * overwrite the unmount record after a clean unmount.
959 *
960 * Do this only if we are going to recover the filesystem
961 *
962 * NOTE: This used to say "if (!readonly)"
963 * However on Linux, we can & do recover a read-only filesystem.
964 * We only skip recovery if NORECOVERY is specified on mount,
965 * in which case we would not be here.
966 *
967 * But... if the -device- itself is readonly, just skip this.
968 * We can't recover this device anyway, so it won't matter.
969 */
970 if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
971 error = xlog_clear_stale_blocks(log, tail_lsn);
972 }
973
974bread_err:
975exit:
976 xlog_put_bp(bp);
977
978 if (error)
979 xlog_warn("XFS: failed to locate log tail");
980 return error;
981}
982
983/*
984 * Is the log zeroed at all?
985 *
986 * The last binary search should be changed to perform an X block read
987 * once X becomes small enough. You can then search linearly through
988 * the X blocks. This will cut down on the number of reads we need to do.
989 *
990 * If the log is partially zeroed, this routine will pass back the blkno
991 * of the first block with cycle number 0. It won't have a complete LR
992 * preceding it.
993 *
994 * Return:
995 * 0 => the log is completely written to
996 * -1 => use *blk_no as the first block of the log
997 * >0 => error has occurred
998 */
999int
1000xlog_find_zeroed(
1001 xlog_t *log,
1002 xfs_daddr_t *blk_no)
1003{
1004 xfs_buf_t *bp;
1005 xfs_caddr_t offset;
1006 uint first_cycle, last_cycle;
1007 xfs_daddr_t new_blk, last_blk, start_blk;
1008 xfs_daddr_t num_scan_bblks;
1009 int error, log_bbnum = log->l_logBBsize;
1010
1011 /* check totally zeroed log */
1012 bp = xlog_get_bp(log, 1);
1013 if (!bp)
1014 return ENOMEM;
1015 if ((error = xlog_bread(log, 0, 1, bp)))
1016 goto bp_err;
1017 offset = xlog_align(log, 0, 1, bp);
1018 first_cycle = GET_CYCLE(offset, ARCH_CONVERT);
1019 if (first_cycle == 0) { /* completely zeroed log */
1020 *blk_no = 0;
1021 xlog_put_bp(bp);
1022 return -1;
1023 }
1024
1025 /* check partially zeroed log */
1026 if ((error = xlog_bread(log, log_bbnum-1, 1, bp)))
1027 goto bp_err;
1028 offset = xlog_align(log, log_bbnum-1, 1, bp);
1029 last_cycle = GET_CYCLE(offset, ARCH_CONVERT);
1030 if (last_cycle != 0) { /* log completely written to */
1031 xlog_put_bp(bp);
1032 return 0;
1033 } else if (first_cycle != 1) {
1034 /*
1035 * If the cycle of the last block is zero, the cycle of
1036 * the first block must be 1. If it's not, maybe we're
1037 * not looking at a log... Bail out.
1038 */
1039 xlog_warn("XFS: Log inconsistent or not a log (last==0, first!=1)");
1040 return XFS_ERROR(EINVAL);
1041 }
1042
1043 /* we have a partially zeroed log */
1044 last_blk = log_bbnum-1;
1045 if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1046 goto bp_err;
1047
1048 /*
1049 * Validate the answer. Because there is no way to guarantee that
1050 * the entire log is made up of log records which are the same size,
1051 * we scan over the defined maximum blocks. At this point, the maximum
1052 * is not chosen to mean anything special. XXXmiken
1053 */
1054 num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1055 ASSERT(num_scan_bblks <= INT_MAX);
1056
1057 if (last_blk < num_scan_bblks)
1058 num_scan_bblks = last_blk;
1059 start_blk = last_blk - num_scan_bblks;
1060
1061 /*
1062 * We search for any instances of cycle number 0 that occur before
1063 * our current estimate of the head. What we're trying to detect is
1064 * 1 ... | 0 | 1 | 0...
1065 * ^ binary search ends here
1066 */
1067 if ((error = xlog_find_verify_cycle(log, start_blk,
1068 (int)num_scan_bblks, 0, &new_blk)))
1069 goto bp_err;
1070 if (new_blk != -1)
1071 last_blk = new_blk;
1072
1073 /*
1074 * Potentially backup over partial log record write. We don't need
1075 * to search the end of the log because we know it is zero.
1076 */
1077 if ((error = xlog_find_verify_log_record(log, start_blk,
1078 &last_blk, 0)) == -1) {
1079 error = XFS_ERROR(EIO);
1080 goto bp_err;
1081 } else if (error)
1082 goto bp_err;
1083
1084 *blk_no = last_blk;
1085bp_err:
1086 xlog_put_bp(bp);
1087 if (error)
1088 return error;
1089 return -1;
1090}
1091
1092/*
1093 * These are simple subroutines used by xlog_clear_stale_blocks() below
1094 * to initialize a buffer full of empty log record headers and write
1095 * them into the log.
1096 */
1097STATIC void
1098xlog_add_record(
1099 xlog_t *log,
1100 xfs_caddr_t buf,
1101 int cycle,
1102 int block,
1103 int tail_cycle,
1104 int tail_block)
1105{
1106 xlog_rec_header_t *recp = (xlog_rec_header_t *)buf;
1107
1108 memset(buf, 0, BBSIZE);
1109 INT_SET(recp->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1110 INT_SET(recp->h_cycle, ARCH_CONVERT, cycle);
1111 INT_SET(recp->h_version, ARCH_CONVERT,
1112 XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1113 ASSIGN_ANY_LSN_DISK(recp->h_lsn, cycle, block);
1114 ASSIGN_ANY_LSN_DISK(recp->h_tail_lsn, tail_cycle, tail_block);
1115 INT_SET(recp->h_fmt, ARCH_CONVERT, XLOG_FMT);
1116 memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1117}
1118
1119STATIC int
1120xlog_write_log_records(
1121 xlog_t *log,
1122 int cycle,
1123 int start_block,
1124 int blocks,
1125 int tail_cycle,
1126 int tail_block)
1127{
1128 xfs_caddr_t offset;
1129 xfs_buf_t *bp;
1130 int balign, ealign;
1131 int sectbb = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
1132 int end_block = start_block + blocks;
1133 int bufblks;
1134 int error = 0;
1135 int i, j = 0;
1136
1137 bufblks = 1 << ffs(blocks);
1138 while (!(bp = xlog_get_bp(log, bufblks))) {
1139 bufblks >>= 1;
1140 if (bufblks <= log->l_sectbb_log)
1141 return ENOMEM;
1142 }
1143
1144 /* We may need to do a read at the start to fill in part of
1145 * the buffer in the starting sector not covered by the first
1146 * write below.
1147 */
1148 balign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, start_block);
1149 if (balign != start_block) {
1150 if ((error = xlog_bread(log, start_block, 1, bp))) {
1151 xlog_put_bp(bp);
1152 return error;
1153 }
1154 j = start_block - balign;
1155 }
1156
1157 for (i = start_block; i < end_block; i += bufblks) {
1158 int bcount, endcount;
1159
1160 bcount = min(bufblks, end_block - start_block);
1161 endcount = bcount - j;
1162
1163 /* We may need to do a read at the end to fill in part of
1164 * the buffer in the final sector not covered by the write.
1165 * If this is the same sector as the above read, skip it.
1166 */
1167 ealign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, end_block);
1168 if (j == 0 && (start_block + endcount > ealign)) {
1169 offset = XFS_BUF_PTR(bp);
1170 balign = BBTOB(ealign - start_block);
1171 XFS_BUF_SET_PTR(bp, offset + balign, BBTOB(sectbb));
1172 if ((error = xlog_bread(log, ealign, sectbb, bp)))
1173 break;
1174 XFS_BUF_SET_PTR(bp, offset, bufblks);
1175 }
1176
1177 offset = xlog_align(log, start_block, endcount, bp);
1178 for (; j < endcount; j++) {
1179 xlog_add_record(log, offset, cycle, i+j,
1180 tail_cycle, tail_block);
1181 offset += BBSIZE;
1182 }
1183 error = xlog_bwrite(log, start_block, endcount, bp);
1184 if (error)
1185 break;
1186 start_block += endcount;
1187 j = 0;
1188 }
1189 xlog_put_bp(bp);
1190 return error;
1191}
1192
1193/*
1194 * This routine is called to blow away any incomplete log writes out
1195 * in front of the log head. We do this so that we won't become confused
1196 * if we come up, write only a little bit more, and then crash again.
1197 * If we leave the partial log records out there, this situation could
1198 * cause us to think those partial writes are valid blocks since they
1199 * have the current cycle number. We get rid of them by overwriting them
1200 * with empty log records with the old cycle number rather than the
1201 * current one.
1202 *
1203 * The tail lsn is passed in rather than taken from
1204 * the log so that we will not write over the unmount record after a
1205 * clean unmount in a 512 block log. Doing so would leave the log without
1206 * any valid log records in it until a new one was written. If we crashed
1207 * during that time we would not be able to recover.
1208 */
1209STATIC int
1210xlog_clear_stale_blocks(
1211 xlog_t *log,
1212 xfs_lsn_t tail_lsn)
1213{
1214 int tail_cycle, head_cycle;
1215 int tail_block, head_block;
1216 int tail_distance, max_distance;
1217 int distance;
1218 int error;
1219
1220 tail_cycle = CYCLE_LSN(tail_lsn);
1221 tail_block = BLOCK_LSN(tail_lsn);
1222 head_cycle = log->l_curr_cycle;
1223 head_block = log->l_curr_block;
1224
1225 /*
1226 * Figure out the distance between the new head of the log
1227 * and the tail. We want to write over any blocks beyond the
1228 * head that we may have written just before the crash, but
1229 * we don't want to overwrite the tail of the log.
1230 */
1231 if (head_cycle == tail_cycle) {
1232 /*
1233 * The tail is behind the head in the physical log,
1234 * so the distance from the head to the tail is the
1235 * distance from the head to the end of the log plus
1236 * the distance from the beginning of the log to the
1237 * tail.
1238 */
1239 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1240 XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1241 XFS_ERRLEVEL_LOW, log->l_mp);
1242 return XFS_ERROR(EFSCORRUPTED);
1243 }
1244 tail_distance = tail_block + (log->l_logBBsize - head_block);
1245 } else {
1246 /*
1247 * The head is behind the tail in the physical log,
1248 * so the distance from the head to the tail is just
1249 * the tail block minus the head block.
1250 */
1251 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1252 XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1253 XFS_ERRLEVEL_LOW, log->l_mp);
1254 return XFS_ERROR(EFSCORRUPTED);
1255 }
1256 tail_distance = tail_block - head_block;
1257 }
1258
1259 /*
1260 * If the head is right up against the tail, we can't clear
1261 * anything.
1262 */
1263 if (tail_distance <= 0) {
1264 ASSERT(tail_distance == 0);
1265 return 0;
1266 }
1267
1268 max_distance = XLOG_TOTAL_REC_SHIFT(log);
1269 /*
1270 * Take the smaller of the maximum amount of outstanding I/O
1271 * we could have and the distance to the tail to clear out.
1272 * We take the smaller so that we don't overwrite the tail and
1273 * we don't waste all day writing from the head to the tail
1274 * for no reason.
1275 */
1276 max_distance = MIN(max_distance, tail_distance);
1277
1278 if ((head_block + max_distance) <= log->l_logBBsize) {
1279 /*
1280 * We can stomp all the blocks we need to without
1281 * wrapping around the end of the log. Just do it
1282 * in a single write. Use the cycle number of the
1283 * current cycle minus one so that the log will look like:
1284 * n ... | n - 1 ...
1285 */
1286 error = xlog_write_log_records(log, (head_cycle - 1),
1287 head_block, max_distance, tail_cycle,
1288 tail_block);
1289 if (error)
1290 return error;
1291 } else {
1292 /*
1293 * We need to wrap around the end of the physical log in
1294 * order to clear all the blocks. Do it in two separate
1295 * I/Os. The first write should be from the head to the
1296 * end of the physical log, and it should use the current
1297 * cycle number minus one just like above.
1298 */
1299 distance = log->l_logBBsize - head_block;
1300 error = xlog_write_log_records(log, (head_cycle - 1),
1301 head_block, distance, tail_cycle,
1302 tail_block);
1303
1304 if (error)
1305 return error;
1306
1307 /*
1308 * Now write the blocks at the start of the physical log.
1309 * This writes the remainder of the blocks we want to clear.
1310 * It uses the current cycle number since we're now on the
1311 * same cycle as the head so that we get:
1312 * n ... n ... | n - 1 ...
1313 * ^^^^^ blocks we're writing
1314 */
1315 distance = max_distance - (log->l_logBBsize - head_block);
1316 error = xlog_write_log_records(log, head_cycle, 0, distance,
1317 tail_cycle, tail_block);
1318 if (error)
1319 return error;
1320 }
1321
1322 return 0;
1323}
1324
1325/******************************************************************************
1326 *
1327 * Log recover routines
1328 *
1329 ******************************************************************************
1330 */
1331
1332STATIC xlog_recover_t *
1333xlog_recover_find_tid(
1334 xlog_recover_t *q,
1335 xlog_tid_t tid)
1336{
1337 xlog_recover_t *p = q;
1338
1339 while (p != NULL) {
1340 if (p->r_log_tid == tid)
1341 break;
1342 p = p->r_next;
1343 }
1344 return p;
1345}
1346
1347STATIC void
1348xlog_recover_put_hashq(
1349 xlog_recover_t **q,
1350 xlog_recover_t *trans)
1351{
1352 trans->r_next = *q;
1353 *q = trans;
1354}
1355
1356STATIC void
1357xlog_recover_add_item(
1358 xlog_recover_item_t **itemq)
1359{
1360 xlog_recover_item_t *item;
1361
1362 item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1363 xlog_recover_insert_item_backq(itemq, item);
1364}
1365
1366STATIC int
1367xlog_recover_add_to_cont_trans(
1368 xlog_recover_t *trans,
1369 xfs_caddr_t dp,
1370 int len)
1371{
1372 xlog_recover_item_t *item;
1373 xfs_caddr_t ptr, old_ptr;
1374 int old_len;
1375
1376 item = trans->r_itemq;
1377 if (item == 0) {
1378 /* finish copying rest of trans header */
1379 xlog_recover_add_item(&trans->r_itemq);
1380 ptr = (xfs_caddr_t) &trans->r_theader +
1381 sizeof(xfs_trans_header_t) - len;
1382 memcpy(ptr, dp, len); /* d, s, l */
1383 return 0;
1384 }
1385 item = item->ri_prev;
1386
1387 old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1388 old_len = item->ri_buf[item->ri_cnt-1].i_len;
1389
1390 ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0);
1391 memcpy(&ptr[old_len], dp, len); /* d, s, l */
1392 item->ri_buf[item->ri_cnt-1].i_len += len;
1393 item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1394 return 0;
1395}
1396
1397/*
1398 * The next region to add is the start of a new region. It could be
1399 * a whole region or it could be the first part of a new region. Because
1400 * of this, the assumption here is that the type and size fields of all
1401 * format structures fit into the first 32 bits of the structure.
1402 *
1403 * This works because all regions must be 32 bit aligned. Therefore, we
1404 * either have both fields or we have neither field. In the case we have
1405 * neither field, the data part of the region is zero length. We only have
1406 * a log_op_header and can throw away the header since a new one will appear
1407 * later. If we have at least 4 bytes, then we can determine how many regions
1408 * will appear in the current log item.
1409 */
1410STATIC int
1411xlog_recover_add_to_trans(
1412 xlog_recover_t *trans,
1413 xfs_caddr_t dp,
1414 int len)
1415{
1416 xfs_inode_log_format_t *in_f; /* any will do */
1417 xlog_recover_item_t *item;
1418 xfs_caddr_t ptr;
1419
1420 if (!len)
1421 return 0;
1422 item = trans->r_itemq;
1423 if (item == 0) {
1424 ASSERT(*(uint *)dp == XFS_TRANS_HEADER_MAGIC);
1425 if (len == sizeof(xfs_trans_header_t))
1426 xlog_recover_add_item(&trans->r_itemq);
1427 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1428 return 0;
1429 }
1430
1431 ptr = kmem_alloc(len, KM_SLEEP);
1432 memcpy(ptr, dp, len);
1433 in_f = (xfs_inode_log_format_t *)ptr;
1434
1435 if (item->ri_prev->ri_total != 0 &&
1436 item->ri_prev->ri_total == item->ri_prev->ri_cnt) {
1437 xlog_recover_add_item(&trans->r_itemq);
1438 }
1439 item = trans->r_itemq;
1440 item = item->ri_prev;
1441
1442 if (item->ri_total == 0) { /* first region to be added */
1443 item->ri_total = in_f->ilf_size;
1444 ASSERT(item->ri_total <= XLOG_MAX_REGIONS_IN_ITEM);
1445 item->ri_buf = kmem_zalloc((item->ri_total *
1446 sizeof(xfs_log_iovec_t)), KM_SLEEP);
1447 }
1448 ASSERT(item->ri_total > item->ri_cnt);
1449 /* Description region is ri_buf[0] */
1450 item->ri_buf[item->ri_cnt].i_addr = ptr;
1451 item->ri_buf[item->ri_cnt].i_len = len;
1452 item->ri_cnt++;
1453 return 0;
1454}
1455
1456STATIC void
1457xlog_recover_new_tid(
1458 xlog_recover_t **q,
1459 xlog_tid_t tid,
1460 xfs_lsn_t lsn)
1461{
1462 xlog_recover_t *trans;
1463
1464 trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1465 trans->r_log_tid = tid;
1466 trans->r_lsn = lsn;
1467 xlog_recover_put_hashq(q, trans);
1468}
1469
1470STATIC int
1471xlog_recover_unlink_tid(
1472 xlog_recover_t **q,
1473 xlog_recover_t *trans)
1474{
1475 xlog_recover_t *tp;
1476 int found = 0;
1477
1478 ASSERT(trans != 0);
1479 if (trans == *q) {
1480 *q = (*q)->r_next;
1481 } else {
1482 tp = *q;
1483 while (tp != 0) {
1484 if (tp->r_next == trans) {
1485 found = 1;
1486 break;
1487 }
1488 tp = tp->r_next;
1489 }
1490 if (!found) {
1491 xlog_warn(
1492 "XFS: xlog_recover_unlink_tid: trans not found");
1493 ASSERT(0);
1494 return XFS_ERROR(EIO);
1495 }
1496 tp->r_next = tp->r_next->r_next;
1497 }
1498 return 0;
1499}
1500
1501STATIC void
1502xlog_recover_insert_item_backq(
1503 xlog_recover_item_t **q,
1504 xlog_recover_item_t *item)
1505{
1506 if (*q == 0) {
1507 item->ri_prev = item->ri_next = item;
1508 *q = item;
1509 } else {
1510 item->ri_next = *q;
1511 item->ri_prev = (*q)->ri_prev;
1512 (*q)->ri_prev = item;
1513 item->ri_prev->ri_next = item;
1514 }
1515}
1516
1517STATIC void
1518xlog_recover_insert_item_frontq(
1519 xlog_recover_item_t **q,
1520 xlog_recover_item_t *item)
1521{
1522 xlog_recover_insert_item_backq(q, item);
1523 *q = item;
1524}
1525
1526STATIC int
1527xlog_recover_reorder_trans(
1528 xlog_t *log,
1529 xlog_recover_t *trans)
1530{
1531 xlog_recover_item_t *first_item, *itemq, *itemq_next;
1532 xfs_buf_log_format_t *buf_f;
1533 xfs_buf_log_format_v1_t *obuf_f;
1534 ushort flags = 0;
1535
1536 first_item = itemq = trans->r_itemq;
1537 trans->r_itemq = NULL;
1538 do {
1539 itemq_next = itemq->ri_next;
1540 buf_f = (xfs_buf_log_format_t *)itemq->ri_buf[0].i_addr;
1541 switch (ITEM_TYPE(itemq)) {
1542 case XFS_LI_BUF:
1543 flags = buf_f->blf_flags;
1544 break;
1545 case XFS_LI_6_1_BUF:
1546 case XFS_LI_5_3_BUF:
1547 obuf_f = (xfs_buf_log_format_v1_t*)buf_f;
1548 flags = obuf_f->blf_flags;
1549 break;
1550 }
1551
1552 switch (ITEM_TYPE(itemq)) {
1553 case XFS_LI_BUF:
1554 case XFS_LI_6_1_BUF:
1555 case XFS_LI_5_3_BUF:
1556 if (!(flags & XFS_BLI_CANCEL)) {
1557 xlog_recover_insert_item_frontq(&trans->r_itemq,
1558 itemq);
1559 break;
1560 }
1561 case XFS_LI_INODE:
1562 case XFS_LI_6_1_INODE:
1563 case XFS_LI_5_3_INODE:
1564 case XFS_LI_DQUOT:
1565 case XFS_LI_QUOTAOFF:
1566 case XFS_LI_EFD:
1567 case XFS_LI_EFI:
1568 xlog_recover_insert_item_backq(&trans->r_itemq, itemq);
1569 break;
1570 default:
1571 xlog_warn(
1572 "XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1573 ASSERT(0);
1574 return XFS_ERROR(EIO);
1575 }
1576 itemq = itemq_next;
1577 } while (first_item != itemq);
1578 return 0;
1579}
1580
1581/*
1582 * Build up the table of buf cancel records so that we don't replay
1583 * cancelled data in the second pass. For buffer records that are
1584 * not cancel records, there is nothing to do here so we just return.
1585 *
1586 * If we get a cancel record which is already in the table, this indicates
1587 * that the buffer was cancelled multiple times. In order to ensure
1588 * that during pass 2 we keep the record in the table until we reach its
1589 * last occurrence in the log, we keep a reference count in the cancel
1590 * record in the table to tell us how many times we expect to see this
1591 * record during the second pass.
1592 */
1593STATIC void
1594xlog_recover_do_buffer_pass1(
1595 xlog_t *log,
1596 xfs_buf_log_format_t *buf_f)
1597{
1598 xfs_buf_cancel_t *bcp;
1599 xfs_buf_cancel_t *nextp;
1600 xfs_buf_cancel_t *prevp;
1601 xfs_buf_cancel_t **bucket;
1602 xfs_buf_log_format_v1_t *obuf_f;
1603 xfs_daddr_t blkno = 0;
1604 uint len = 0;
1605 ushort flags = 0;
1606
1607 switch (buf_f->blf_type) {
1608 case XFS_LI_BUF:
1609 blkno = buf_f->blf_blkno;
1610 len = buf_f->blf_len;
1611 flags = buf_f->blf_flags;
1612 break;
1613 case XFS_LI_6_1_BUF:
1614 case XFS_LI_5_3_BUF:
1615 obuf_f = (xfs_buf_log_format_v1_t*)buf_f;
1616 blkno = (xfs_daddr_t) obuf_f->blf_blkno;
1617 len = obuf_f->blf_len;
1618 flags = obuf_f->blf_flags;
1619 break;
1620 }
1621
1622 /*
1623 * If this isn't a cancel buffer item, then just return.
1624 */
1625 if (!(flags & XFS_BLI_CANCEL))
1626 return;
1627
1628 /*
1629 * Insert an xfs_buf_cancel record into the hash table of
1630 * them. If there is already an identical record, bump
1631 * its reference count.
1632 */
1633 bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1634 XLOG_BC_TABLE_SIZE];
1635 /*
1636 * If the hash bucket is empty then just insert a new record into
1637 * the bucket.
1638 */
1639 if (*bucket == NULL) {
1640 bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1641 KM_SLEEP);
1642 bcp->bc_blkno = blkno;
1643 bcp->bc_len = len;
1644 bcp->bc_refcount = 1;
1645 bcp->bc_next = NULL;
1646 *bucket = bcp;
1647 return;
1648 }
1649
1650 /*
1651 * The hash bucket is not empty, so search for duplicates of our
1652 * record. If we find one them just bump its refcount. If not
1653 * then add us at the end of the list.
1654 */
1655 prevp = NULL;
1656 nextp = *bucket;
1657 while (nextp != NULL) {
1658 if (nextp->bc_blkno == blkno && nextp->bc_len == len) {
1659 nextp->bc_refcount++;
1660 return;
1661 }
1662 prevp = nextp;
1663 nextp = nextp->bc_next;
1664 }
1665 ASSERT(prevp != NULL);
1666 bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1667 KM_SLEEP);
1668 bcp->bc_blkno = blkno;
1669 bcp->bc_len = len;
1670 bcp->bc_refcount = 1;
1671 bcp->bc_next = NULL;
1672 prevp->bc_next = bcp;
1673}
1674
1675/*
1676 * Check to see whether the buffer being recovered has a corresponding
1677 * entry in the buffer cancel record table. If it does then return 1
1678 * so that it will be cancelled, otherwise return 0. If the buffer is
1679 * actually a buffer cancel item (XFS_BLI_CANCEL is set), then decrement
1680 * the refcount on the entry in the table and remove it from the table
1681 * if this is the last reference.
1682 *
1683 * We remove the cancel record from the table when we encounter its
1684 * last occurrence in the log so that if the same buffer is re-used
1685 * again after its last cancellation we actually replay the changes
1686 * made at that point.
1687 */
1688STATIC int
1689xlog_check_buffer_cancelled(
1690 xlog_t *log,
1691 xfs_daddr_t blkno,
1692 uint len,
1693 ushort flags)
1694{
1695 xfs_buf_cancel_t *bcp;
1696 xfs_buf_cancel_t *prevp;
1697 xfs_buf_cancel_t **bucket;
1698
1699 if (log->l_buf_cancel_table == NULL) {
1700 /*
1701 * There is nothing in the table built in pass one,
1702 * so this buffer must not be cancelled.
1703 */
1704 ASSERT(!(flags & XFS_BLI_CANCEL));
1705 return 0;
1706 }
1707
1708 bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1709 XLOG_BC_TABLE_SIZE];
1710 bcp = *bucket;
1711 if (bcp == NULL) {
1712 /*
1713 * There is no corresponding entry in the table built
1714 * in pass one, so this buffer has not been cancelled.
1715 */
1716 ASSERT(!(flags & XFS_BLI_CANCEL));
1717 return 0;
1718 }
1719
1720 /*
1721 * Search for an entry in the buffer cancel table that
1722 * matches our buffer.
1723 */
1724 prevp = NULL;
1725 while (bcp != NULL) {
1726 if (bcp->bc_blkno == blkno && bcp->bc_len == len) {
1727 /*
1728 * We've go a match, so return 1 so that the
1729 * recovery of this buffer is cancelled.
1730 * If this buffer is actually a buffer cancel
1731 * log item, then decrement the refcount on the
1732 * one in the table and remove it if this is the
1733 * last reference.
1734 */
1735 if (flags & XFS_BLI_CANCEL) {
1736 bcp->bc_refcount--;
1737 if (bcp->bc_refcount == 0) {
1738 if (prevp == NULL) {
1739 *bucket = bcp->bc_next;
1740 } else {
1741 prevp->bc_next = bcp->bc_next;
1742 }
1743 kmem_free(bcp,
1744 sizeof(xfs_buf_cancel_t));
1745 }
1746 }
1747 return 1;
1748 }
1749 prevp = bcp;
1750 bcp = bcp->bc_next;
1751 }
1752 /*
1753 * We didn't find a corresponding entry in the table, so
1754 * return 0 so that the buffer is NOT cancelled.
1755 */
1756 ASSERT(!(flags & XFS_BLI_CANCEL));
1757 return 0;
1758}
1759
1760STATIC int
1761xlog_recover_do_buffer_pass2(
1762 xlog_t *log,
1763 xfs_buf_log_format_t *buf_f)
1764{
1765 xfs_buf_log_format_v1_t *obuf_f;
1766 xfs_daddr_t blkno = 0;
1767 ushort flags = 0;
1768 uint len = 0;
1769
1770 switch (buf_f->blf_type) {
1771 case XFS_LI_BUF:
1772 blkno = buf_f->blf_blkno;
1773 flags = buf_f->blf_flags;
1774 len = buf_f->blf_len;
1775 break;
1776 case XFS_LI_6_1_BUF:
1777 case XFS_LI_5_3_BUF:
1778 obuf_f = (xfs_buf_log_format_v1_t*)buf_f;
1779 blkno = (xfs_daddr_t) obuf_f->blf_blkno;
1780 flags = obuf_f->blf_flags;
1781 len = (xfs_daddr_t) obuf_f->blf_len;
1782 break;
1783 }
1784
1785 return xlog_check_buffer_cancelled(log, blkno, len, flags);
1786}
1787
1788/*
1789 * Perform recovery for a buffer full of inodes. In these buffers,
1790 * the only data which should be recovered is that which corresponds
1791 * to the di_next_unlinked pointers in the on disk inode structures.
1792 * The rest of the data for the inodes is always logged through the
1793 * inodes themselves rather than the inode buffer and is recovered
1794 * in xlog_recover_do_inode_trans().
1795 *
1796 * The only time when buffers full of inodes are fully recovered is
1797 * when the buffer is full of newly allocated inodes. In this case
1798 * the buffer will not be marked as an inode buffer and so will be
1799 * sent to xlog_recover_do_reg_buffer() below during recovery.
1800 */
1801STATIC int
1802xlog_recover_do_inode_buffer(
1803 xfs_mount_t *mp,
1804 xlog_recover_item_t *item,
1805 xfs_buf_t *bp,
1806 xfs_buf_log_format_t *buf_f)
1807{
1808 int i;
1809 int item_index;
1810 int bit;
1811 int nbits;
1812 int reg_buf_offset;
1813 int reg_buf_bytes;
1814 int next_unlinked_offset;
1815 int inodes_per_buf;
1816 xfs_agino_t *logged_nextp;
1817 xfs_agino_t *buffer_nextp;
1818 xfs_buf_log_format_v1_t *obuf_f;
1819 unsigned int *data_map = NULL;
1820 unsigned int map_size = 0;
1821
1822 switch (buf_f->blf_type) {
1823 case XFS_LI_BUF:
1824 data_map = buf_f->blf_data_map;
1825 map_size = buf_f->blf_map_size;
1826 break;
1827 case XFS_LI_6_1_BUF:
1828 case XFS_LI_5_3_BUF:
1829 obuf_f = (xfs_buf_log_format_v1_t*)buf_f;
1830 data_map = obuf_f->blf_data_map;
1831 map_size = obuf_f->blf_map_size;
1832 break;
1833 }
1834 /*
1835 * Set the variables corresponding to the current region to
1836 * 0 so that we'll initialize them on the first pass through
1837 * the loop.
1838 */
1839 reg_buf_offset = 0;
1840 reg_buf_bytes = 0;
1841 bit = 0;
1842 nbits = 0;
1843 item_index = 0;
1844 inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1845 for (i = 0; i < inodes_per_buf; i++) {
1846 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1847 offsetof(xfs_dinode_t, di_next_unlinked);
1848
1849 while (next_unlinked_offset >=
1850 (reg_buf_offset + reg_buf_bytes)) {
1851 /*
1852 * The next di_next_unlinked field is beyond
1853 * the current logged region. Find the next
1854 * logged region that contains or is beyond
1855 * the current di_next_unlinked field.
1856 */
1857 bit += nbits;
1858 bit = xfs_next_bit(data_map, map_size, bit);
1859
1860 /*
1861 * If there are no more logged regions in the
1862 * buffer, then we're done.
1863 */
1864 if (bit == -1) {
1865 return 0;
1866 }
1867
1868 nbits = xfs_contig_bits(data_map, map_size,
1869 bit);
1870 ASSERT(nbits > 0);
1871 reg_buf_offset = bit << XFS_BLI_SHIFT;
1872 reg_buf_bytes = nbits << XFS_BLI_SHIFT;
1873 item_index++;
1874 }
1875
1876 /*
1877 * If the current logged region starts after the current
1878 * di_next_unlinked field, then move on to the next
1879 * di_next_unlinked field.
1880 */
1881 if (next_unlinked_offset < reg_buf_offset) {
1882 continue;
1883 }
1884
1885 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1886 ASSERT((item->ri_buf[item_index].i_len % XFS_BLI_CHUNK) == 0);
1887 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1888
1889 /*
1890 * The current logged region contains a copy of the
1891 * current di_next_unlinked field. Extract its value
1892 * and copy it to the buffer copy.
1893 */
1894 logged_nextp = (xfs_agino_t *)
1895 ((char *)(item->ri_buf[item_index].i_addr) +
1896 (next_unlinked_offset - reg_buf_offset));
1897 if (unlikely(*logged_nextp == 0)) {
1898 xfs_fs_cmn_err(CE_ALERT, mp,
1899 "bad inode buffer log record (ptr = 0x%p, bp = 0x%p). XFS trying to replay bad (0) inode di_next_unlinked field",
1900 item, bp);
1901 XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1902 XFS_ERRLEVEL_LOW, mp);
1903 return XFS_ERROR(EFSCORRUPTED);
1904 }
1905
1906 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1907 next_unlinked_offset);
1908 INT_SET(*buffer_nextp, ARCH_CONVERT, *logged_nextp);
1909 }
1910
1911 return 0;
1912}
1913
1914/*
1915 * Perform a 'normal' buffer recovery. Each logged region of the
1916 * buffer should be copied over the corresponding region in the
1917 * given buffer. The bitmap in the buf log format structure indicates
1918 * where to place the logged data.
1919 */
1920/*ARGSUSED*/
1921STATIC void
1922xlog_recover_do_reg_buffer(
1923 xfs_mount_t *mp,
1924 xlog_recover_item_t *item,
1925 xfs_buf_t *bp,
1926 xfs_buf_log_format_t *buf_f)
1927{
1928 int i;
1929 int bit;
1930 int nbits;
1931 xfs_buf_log_format_v1_t *obuf_f;
1932 unsigned int *data_map = NULL;
1933 unsigned int map_size = 0;
1934 int error;
1935
1936 switch (buf_f->blf_type) {
1937 case XFS_LI_BUF:
1938 data_map = buf_f->blf_data_map;
1939 map_size = buf_f->blf_map_size;
1940 break;
1941 case XFS_LI_6_1_BUF:
1942 case XFS_LI_5_3_BUF:
1943 obuf_f = (xfs_buf_log_format_v1_t*)buf_f;
1944 data_map = obuf_f->blf_data_map;
1945 map_size = obuf_f->blf_map_size;
1946 break;
1947 }
1948 bit = 0;
1949 i = 1; /* 0 is the buf format structure */
1950 while (1) {
1951 bit = xfs_next_bit(data_map, map_size, bit);
1952 if (bit == -1)
1953 break;
1954 nbits = xfs_contig_bits(data_map, map_size, bit);
1955 ASSERT(nbits > 0);
1956 ASSERT(item->ri_buf[i].i_addr != 0);
1957 ASSERT(item->ri_buf[i].i_len % XFS_BLI_CHUNK == 0);
1958 ASSERT(XFS_BUF_COUNT(bp) >=
1959 ((uint)bit << XFS_BLI_SHIFT)+(nbits<<XFS_BLI_SHIFT));
1960
1961 /*
1962 * Do a sanity check if this is a dquot buffer. Just checking
1963 * the first dquot in the buffer should do. XXXThis is
1964 * probably a good thing to do for other buf types also.
1965 */
1966 error = 0;
1967 if (buf_f->blf_flags & (XFS_BLI_UDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
1968 error = xfs_qm_dqcheck((xfs_disk_dquot_t *)
1969 item->ri_buf[i].i_addr,
1970 -1, 0, XFS_QMOPT_DOWARN,
1971 "dquot_buf_recover");
1972 }
1973 if (!error)
1974 memcpy(xfs_buf_offset(bp,
1975 (uint)bit << XFS_BLI_SHIFT), /* dest */
1976 item->ri_buf[i].i_addr, /* source */
1977 nbits<<XFS_BLI_SHIFT); /* length */
1978 i++;
1979 bit += nbits;
1980 }
1981
1982 /* Shouldn't be any more regions */
1983 ASSERT(i == item->ri_total);
1984}
1985
1986/*
1987 * Do some primitive error checking on ondisk dquot data structures.
1988 */
1989int
1990xfs_qm_dqcheck(
1991 xfs_disk_dquot_t *ddq,
1992 xfs_dqid_t id,
1993 uint type, /* used only when IO_dorepair is true */
1994 uint flags,
1995 char *str)
1996{
1997 xfs_dqblk_t *d = (xfs_dqblk_t *)ddq;
1998 int errs = 0;
1999
2000 /*
2001 * We can encounter an uninitialized dquot buffer for 2 reasons:
2002 * 1. If we crash while deleting the quotainode(s), and those blks got
2003 * used for user data. This is because we take the path of regular
2004 * file deletion; however, the size field of quotainodes is never
2005 * updated, so all the tricks that we play in itruncate_finish
2006 * don't quite matter.
2007 *
2008 * 2. We don't play the quota buffers when there's a quotaoff logitem.
2009 * But the allocation will be replayed so we'll end up with an
2010 * uninitialized quota block.
2011 *
2012 * This is all fine; things are still consistent, and we haven't lost
2013 * any quota information. Just don't complain about bad dquot blks.
2014 */
2015 if (INT_GET(ddq->d_magic, ARCH_CONVERT) != XFS_DQUOT_MAGIC) {
2016 if (flags & XFS_QMOPT_DOWARN)
2017 cmn_err(CE_ALERT,
2018 "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
2019 str, id,
2020 INT_GET(ddq->d_magic, ARCH_CONVERT), XFS_DQUOT_MAGIC);
2021 errs++;
2022 }
2023 if (INT_GET(ddq->d_version, ARCH_CONVERT) != XFS_DQUOT_VERSION) {
2024 if (flags & XFS_QMOPT_DOWARN)
2025 cmn_err(CE_ALERT,
2026 "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
2027 str, id,
2028 INT_GET(ddq->d_magic, ARCH_CONVERT), XFS_DQUOT_VERSION);
2029 errs++;
2030 }
2031
2032 if (INT_GET(ddq->d_flags, ARCH_CONVERT) != XFS_DQ_USER &&
2033 INT_GET(ddq->d_flags, ARCH_CONVERT) != XFS_DQ_GROUP) {
2034 if (flags & XFS_QMOPT_DOWARN)
2035 cmn_err(CE_ALERT,
2036 "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
2037 str, id, INT_GET(ddq->d_flags, ARCH_CONVERT));
2038 errs++;
2039 }
2040
2041 if (id != -1 && id != INT_GET(ddq->d_id, ARCH_CONVERT)) {
2042 if (flags & XFS_QMOPT_DOWARN)
2043 cmn_err(CE_ALERT,
2044 "%s : ondisk-dquot 0x%p, ID mismatch: "
2045 "0x%x expected, found id 0x%x",
2046 str, ddq, id, INT_GET(ddq->d_id, ARCH_CONVERT));
2047 errs++;
2048 }
2049
2050 if (!errs && ddq->d_id) {
2051 if (INT_GET(ddq->d_blk_softlimit, ARCH_CONVERT) &&
2052 INT_GET(ddq->d_bcount, ARCH_CONVERT) >=
2053 INT_GET(ddq->d_blk_softlimit, ARCH_CONVERT)) {
2054 if (!ddq->d_btimer) {
2055 if (flags & XFS_QMOPT_DOWARN)
2056 cmn_err(CE_ALERT,
2057 "%s : Dquot ID 0x%x (0x%p) "
2058 "BLK TIMER NOT STARTED",
2059 str, (int)
2060 INT_GET(ddq->d_id, ARCH_CONVERT), ddq);
2061 errs++;
2062 }
2063 }
2064 if (INT_GET(ddq->d_ino_softlimit, ARCH_CONVERT) &&
2065 INT_GET(ddq->d_icount, ARCH_CONVERT) >=
2066 INT_GET(ddq->d_ino_softlimit, ARCH_CONVERT)) {
2067 if (!ddq->d_itimer) {
2068 if (flags & XFS_QMOPT_DOWARN)
2069 cmn_err(CE_ALERT,
2070 "%s : Dquot ID 0x%x (0x%p) "
2071 "INODE TIMER NOT STARTED",
2072 str, (int)
2073 INT_GET(ddq->d_id, ARCH_CONVERT), ddq);
2074 errs++;
2075 }
2076 }
2077 if (INT_GET(ddq->d_rtb_softlimit, ARCH_CONVERT) &&
2078 INT_GET(ddq->d_rtbcount, ARCH_CONVERT) >=
2079 INT_GET(ddq->d_rtb_softlimit, ARCH_CONVERT)) {
2080 if (!ddq->d_rtbtimer) {
2081 if (flags & XFS_QMOPT_DOWARN)
2082 cmn_err(CE_ALERT,
2083 "%s : Dquot ID 0x%x (0x%p) "
2084 "RTBLK TIMER NOT STARTED",
2085 str, (int)
2086 INT_GET(ddq->d_id, ARCH_CONVERT), ddq);
2087 errs++;
2088 }
2089 }
2090 }
2091
2092 if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2093 return errs;
2094
2095 if (flags & XFS_QMOPT_DOWARN)
2096 cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2097
2098 /*
2099 * Typically, a repair is only requested by quotacheck.
2100 */
2101 ASSERT(id != -1);
2102 ASSERT(flags & XFS_QMOPT_DQREPAIR);
2103 memset(d, 0, sizeof(xfs_dqblk_t));
2104 INT_SET(d->dd_diskdq.d_magic, ARCH_CONVERT, XFS_DQUOT_MAGIC);
2105 INT_SET(d->dd_diskdq.d_version, ARCH_CONVERT, XFS_DQUOT_VERSION);
2106 INT_SET(d->dd_diskdq.d_id, ARCH_CONVERT, id);
2107 INT_SET(d->dd_diskdq.d_flags, ARCH_CONVERT, type);
2108
2109 return errs;
2110}
2111
2112/*
2113 * Perform a dquot buffer recovery.
2114 * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2115 * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2116 * Else, treat it as a regular buffer and do recovery.
2117 */
2118STATIC void
2119xlog_recover_do_dquot_buffer(
2120 xfs_mount_t *mp,
2121 xlog_t *log,
2122 xlog_recover_item_t *item,
2123 xfs_buf_t *bp,
2124 xfs_buf_log_format_t *buf_f)
2125{
2126 uint type;
2127
2128 /*
2129 * Filesystems are required to send in quota flags at mount time.
2130 */
2131 if (mp->m_qflags == 0) {
2132 return;
2133 }
2134
2135 type = 0;
2136 if (buf_f->blf_flags & XFS_BLI_UDQUOT_BUF)
2137 type |= XFS_DQ_USER;
2138 if (buf_f->blf_flags & XFS_BLI_GDQUOT_BUF)
2139 type |= XFS_DQ_GROUP;
2140 /*
2141 * This type of quotas was turned off, so ignore this buffer
2142 */
2143 if (log->l_quotaoffs_flag & type)
2144 return;
2145
2146 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2147}
2148
2149/*
2150 * This routine replays a modification made to a buffer at runtime.
2151 * There are actually two types of buffer, regular and inode, which
2152 * are handled differently. Inode buffers are handled differently
2153 * in that we only recover a specific set of data from them, namely
2154 * the inode di_next_unlinked fields. This is because all other inode
2155 * data is actually logged via inode records and any data we replay
2156 * here which overlaps that may be stale.
2157 *
2158 * When meta-data buffers are freed at run time we log a buffer item
2159 * with the XFS_BLI_CANCEL bit set to indicate that previous copies
2160 * of the buffer in the log should not be replayed at recovery time.
2161 * This is so that if the blocks covered by the buffer are reused for
2162 * file data before we crash we don't end up replaying old, freed
2163 * meta-data into a user's file.
2164 *
2165 * To handle the cancellation of buffer log items, we make two passes
2166 * over the log during recovery. During the first we build a table of
2167 * those buffers which have been cancelled, and during the second we
2168 * only replay those buffers which do not have corresponding cancel
2169 * records in the table. See xlog_recover_do_buffer_pass[1,2] above
2170 * for more details on the implementation of the table of cancel records.
2171 */
2172STATIC int
2173xlog_recover_do_buffer_trans(
2174 xlog_t *log,
2175 xlog_recover_item_t *item,
2176 int pass)
2177{
2178 xfs_buf_log_format_t *buf_f;
2179 xfs_buf_log_format_v1_t *obuf_f;
2180 xfs_mount_t *mp;
2181 xfs_buf_t *bp;
2182 int error;
2183 int cancel;
2184 xfs_daddr_t blkno;
2185 int len;
2186 ushort flags;
2187
2188 buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
2189
2190 if (pass == XLOG_RECOVER_PASS1) {
2191 /*
2192 * In this pass we're only looking for buf items
2193 * with the XFS_BLI_CANCEL bit set.
2194 */
2195 xlog_recover_do_buffer_pass1(log, buf_f);
2196 return 0;
2197 } else {
2198 /*
2199 * In this pass we want to recover all the buffers
2200 * which have not been cancelled and are not
2201 * cancellation buffers themselves. The routine
2202 * we call here will tell us whether or not to
2203 * continue with the replay of this buffer.
2204 */
2205 cancel = xlog_recover_do_buffer_pass2(log, buf_f);
2206 if (cancel) {
2207 return 0;
2208 }
2209 }
2210 switch (buf_f->blf_type) {
2211 case XFS_LI_BUF:
2212 blkno = buf_f->blf_blkno;
2213 len = buf_f->blf_len;
2214 flags = buf_f->blf_flags;
2215 break;
2216 case XFS_LI_6_1_BUF:
2217 case XFS_LI_5_3_BUF:
2218 obuf_f = (xfs_buf_log_format_v1_t*)buf_f;
2219 blkno = obuf_f->blf_blkno;
2220 len = obuf_f->blf_len;
2221 flags = obuf_f->blf_flags;
2222 break;
2223 default:
2224 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
2225 "xfs_log_recover: unknown buffer type 0x%x, dev %s",
2226 buf_f->blf_type, XFS_BUFTARG_NAME(log->l_targ));
2227 XFS_ERROR_REPORT("xlog_recover_do_buffer_trans",
2228 XFS_ERRLEVEL_LOW, log->l_mp);
2229 return XFS_ERROR(EFSCORRUPTED);
2230 }
2231
2232 mp = log->l_mp;
2233 if (flags & XFS_BLI_INODE_BUF) {
2234 bp = xfs_buf_read_flags(mp->m_ddev_targp, blkno, len,
2235 XFS_BUF_LOCK);
2236 } else {
2237 bp = xfs_buf_read(mp->m_ddev_targp, blkno, len, 0);
2238 }
2239 if (XFS_BUF_ISERROR(bp)) {
2240 xfs_ioerror_alert("xlog_recover_do..(read#1)", log->l_mp,
2241 bp, blkno);
2242 error = XFS_BUF_GETERROR(bp);
2243 xfs_buf_relse(bp);
2244 return error;
2245 }
2246
2247 error = 0;
2248 if (flags & XFS_BLI_INODE_BUF) {
2249 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2250 } else if (flags & (XFS_BLI_UDQUOT_BUF | XFS_BLI_GDQUOT_BUF)) {
2251 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2252 } else {
2253 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2254 }
2255 if (error)
2256 return XFS_ERROR(error);
2257
2258 /*
2259 * Perform delayed write on the buffer. Asynchronous writes will be
2260 * slower when taking into account all the buffers to be flushed.
2261 *
2262 * Also make sure that only inode buffers with good sizes stay in
2263 * the buffer cache. The kernel moves inodes in buffers of 1 block
2264 * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger. The inode
2265 * buffers in the log can be a different size if the log was generated
2266 * by an older kernel using unclustered inode buffers or a newer kernel
2267 * running with a different inode cluster size. Regardless, if the
2268 * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2269 * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2270 * the buffer out of the buffer cache so that the buffer won't
2271 * overlap with future reads of those inodes.
2272 */
2273 if (XFS_DINODE_MAGIC ==
2274 INT_GET(*((__uint16_t *)(xfs_buf_offset(bp, 0))), ARCH_CONVERT) &&
2275 (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2276 (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2277 XFS_BUF_STALE(bp);
2278 error = xfs_bwrite(mp, bp);
2279 } else {
2280 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2281 XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2282 XFS_BUF_SET_FSPRIVATE(bp, mp);
2283 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2284 xfs_bdwrite(mp, bp);
2285 }
2286
2287 return (error);
2288}
2289
2290STATIC int
2291xlog_recover_do_inode_trans(
2292 xlog_t *log,
2293 xlog_recover_item_t *item,
2294 int pass)
2295{
2296 xfs_inode_log_format_t *in_f;
2297 xfs_mount_t *mp;
2298 xfs_buf_t *bp;
2299 xfs_imap_t imap;
2300 xfs_dinode_t *dip;
2301 xfs_ino_t ino;
2302 int len;
2303 xfs_caddr_t src;
2304 xfs_caddr_t dest;
2305 int error;
2306 int attr_index;
2307 uint fields;
2308 xfs_dinode_core_t *dicp;
2309
2310 if (pass == XLOG_RECOVER_PASS1) {
2311 return 0;
2312 }
2313
2314 in_f = (xfs_inode_log_format_t *)item->ri_buf[0].i_addr;
2315 ino = in_f->ilf_ino;
2316 mp = log->l_mp;
2317 if (ITEM_TYPE(item) == XFS_LI_INODE) {
2318 imap.im_blkno = (xfs_daddr_t)in_f->ilf_blkno;
2319 imap.im_len = in_f->ilf_len;
2320 imap.im_boffset = in_f->ilf_boffset;
2321 } else {
2322 /*
2323 * It's an old inode format record. We don't know where
2324 * its cluster is located on disk, and we can't allow
2325 * xfs_imap() to figure it out because the inode btrees
2326 * are not ready to be used. Therefore do not pass the
2327 * XFS_IMAP_LOOKUP flag to xfs_imap(). This will give
2328 * us only the single block in which the inode lives
2329 * rather than its cluster, so we must make sure to
2330 * invalidate the buffer when we write it out below.
2331 */
2332 imap.im_blkno = 0;
2333 xfs_imap(log->l_mp, NULL, ino, &imap, 0);
2334 }
2335
2336 /*
2337 * Inode buffers can be freed, look out for it,
2338 * and do not replay the inode.
2339 */
2340 if (xlog_check_buffer_cancelled(log, imap.im_blkno, imap.im_len, 0))
2341 return 0;
2342
2343 bp = xfs_buf_read_flags(mp->m_ddev_targp, imap.im_blkno, imap.im_len,
2344 XFS_BUF_LOCK);
2345 if (XFS_BUF_ISERROR(bp)) {
2346 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2347 bp, imap.im_blkno);
2348 error = XFS_BUF_GETERROR(bp);
2349 xfs_buf_relse(bp);
2350 return error;
2351 }
2352 error = 0;
2353 ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2354 dip = (xfs_dinode_t *)xfs_buf_offset(bp, imap.im_boffset);
2355
2356 /*
2357 * Make sure the place we're flushing out to really looks
2358 * like an inode!
2359 */
2360 if (unlikely(INT_GET(dip->di_core.di_magic, ARCH_CONVERT) != XFS_DINODE_MAGIC)) {
2361 xfs_buf_relse(bp);
2362 xfs_fs_cmn_err(CE_ALERT, mp,
2363 "xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2364 dip, bp, ino);
2365 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(1)",
2366 XFS_ERRLEVEL_LOW, mp);
2367 return XFS_ERROR(EFSCORRUPTED);
2368 }
2369 dicp = (xfs_dinode_core_t*)(item->ri_buf[1].i_addr);
2370 if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2371 xfs_buf_relse(bp);
2372 xfs_fs_cmn_err(CE_ALERT, mp,
2373 "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2374 item, ino);
2375 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(2)",
2376 XFS_ERRLEVEL_LOW, mp);
2377 return XFS_ERROR(EFSCORRUPTED);
2378 }
2379
2380 /* Skip replay when the on disk inode is newer than the log one */
2381 if (dicp->di_flushiter <
2382 INT_GET(dip->di_core.di_flushiter, ARCH_CONVERT)) {
2383 /*
2384 * Deal with the wrap case, DI_MAX_FLUSH is less
2385 * than smaller numbers
2386 */
2387 if ((INT_GET(dip->di_core.di_flushiter, ARCH_CONVERT)
2388 == DI_MAX_FLUSH) &&
2389 (dicp->di_flushiter < (DI_MAX_FLUSH>>1))) {
2390 /* do nothing */
2391 } else {
2392 xfs_buf_relse(bp);
2393 return 0;
2394 }
2395 }
2396 /* Take the opportunity to reset the flush iteration count */
2397 dicp->di_flushiter = 0;
2398
2399 if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2400 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2401 (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2402 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(3)",
2403 XFS_ERRLEVEL_LOW, mp, dicp);
2404 xfs_buf_relse(bp);
2405 xfs_fs_cmn_err(CE_ALERT, mp,
2406 "xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2407 item, dip, bp, ino);
2408 return XFS_ERROR(EFSCORRUPTED);
2409 }
2410 } else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2411 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2412 (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2413 (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2414 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(4)",
2415 XFS_ERRLEVEL_LOW, mp, dicp);
2416 xfs_buf_relse(bp);
2417 xfs_fs_cmn_err(CE_ALERT, mp,
2418 "xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2419 item, dip, bp, ino);
2420 return XFS_ERROR(EFSCORRUPTED);
2421 }
2422 }
2423 if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2424 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(5)",
2425 XFS_ERRLEVEL_LOW, mp, dicp);
2426 xfs_buf_relse(bp);
2427 xfs_fs_cmn_err(CE_ALERT, mp,
2428 "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",
2429 item, dip, bp, ino,
2430 dicp->di_nextents + dicp->di_anextents,
2431 dicp->di_nblocks);
2432 return XFS_ERROR(EFSCORRUPTED);
2433 }
2434 if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2435 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(6)",
2436 XFS_ERRLEVEL_LOW, mp, dicp);
2437 xfs_buf_relse(bp);
2438 xfs_fs_cmn_err(CE_ALERT, mp,
2439 "xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2440 item, dip, bp, ino, dicp->di_forkoff);
2441 return XFS_ERROR(EFSCORRUPTED);
2442 }
2443 if (unlikely(item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t))) {
2444 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(7)",
2445 XFS_ERRLEVEL_LOW, mp, dicp);
2446 xfs_buf_relse(bp);
2447 xfs_fs_cmn_err(CE_ALERT, mp,
2448 "xfs_inode_recover: Bad inode log record length %d, rec ptr 0x%p",
2449 item->ri_buf[1].i_len, item);
2450 return XFS_ERROR(EFSCORRUPTED);
2451 }
2452
2453 /* The core is in in-core format */
2454 xfs_xlate_dinode_core((xfs_caddr_t)&dip->di_core,
2455 (xfs_dinode_core_t*)item->ri_buf[1].i_addr, -1);
2456
2457 /* the rest is in on-disk format */
2458 if (item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t)) {
2459 memcpy((xfs_caddr_t) dip + sizeof(xfs_dinode_core_t),
2460 item->ri_buf[1].i_addr + sizeof(xfs_dinode_core_t),
2461 item->ri_buf[1].i_len - sizeof(xfs_dinode_core_t));
2462 }
2463
2464 fields = in_f->ilf_fields;
2465 switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2466 case XFS_ILOG_DEV:
2467 INT_SET(dip->di_u.di_dev, ARCH_CONVERT, in_f->ilf_u.ilfu_rdev);
2468
2469 break;
2470 case XFS_ILOG_UUID:
2471 dip->di_u.di_muuid = in_f->ilf_u.ilfu_uuid;
2472 break;
2473 }
2474
2475 if (in_f->ilf_size == 2)
2476 goto write_inode_buffer;
2477 len = item->ri_buf[2].i_len;
2478 src = item->ri_buf[2].i_addr;
2479 ASSERT(in_f->ilf_size <= 4);
2480 ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2481 ASSERT(!(fields & XFS_ILOG_DFORK) ||
2482 (len == in_f->ilf_dsize));
2483
2484 switch (fields & XFS_ILOG_DFORK) {
2485 case XFS_ILOG_DDATA:
2486 case XFS_ILOG_DEXT:
2487 memcpy(&dip->di_u, src, len);
2488 break;
2489
2490 case XFS_ILOG_DBROOT:
2491 xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2492 &(dip->di_u.di_bmbt),
2493 XFS_DFORK_DSIZE(dip, mp));
2494 break;
2495
2496 default:
2497 /*
2498 * There are no data fork flags set.
2499 */
2500 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2501 break;
2502 }
2503
2504 /*
2505 * If we logged any attribute data, recover it. There may or
2506 * may not have been any other non-core data logged in this
2507 * transaction.
2508 */
2509 if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2510 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2511 attr_index = 3;
2512 } else {
2513 attr_index = 2;
2514 }
2515 len = item->ri_buf[attr_index].i_len;
2516 src = item->ri_buf[attr_index].i_addr;
2517 ASSERT(len == in_f->ilf_asize);
2518
2519 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2520 case XFS_ILOG_ADATA:
2521 case XFS_ILOG_AEXT:
2522 dest = XFS_DFORK_APTR(dip);
2523 ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2524 memcpy(dest, src, len);
2525 break;
2526
2527 case XFS_ILOG_ABROOT:
2528 dest = XFS_DFORK_APTR(dip);
2529 xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2530 (xfs_bmdr_block_t*)dest,
2531 XFS_DFORK_ASIZE(dip, mp));
2532 break;
2533
2534 default:
2535 xlog_warn("XFS: xlog_recover_do_inode_trans: Invalid flag");
2536 ASSERT(0);
2537 xfs_buf_relse(bp);
2538 return XFS_ERROR(EIO);
2539 }
2540 }
2541
2542write_inode_buffer:
2543 if (ITEM_TYPE(item) == XFS_LI_INODE) {
2544 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2545 XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2546 XFS_BUF_SET_FSPRIVATE(bp, mp);
2547 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2548 xfs_bdwrite(mp, bp);
2549 } else {
2550 XFS_BUF_STALE(bp);
2551 error = xfs_bwrite(mp, bp);
2552 }
2553
2554 return (error);
2555}
2556
2557/*
2558 * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2559 * structure, so that we know not to do any dquot item or dquot buffer recovery,
2560 * of that type.
2561 */
2562STATIC int
2563xlog_recover_do_quotaoff_trans(
2564 xlog_t *log,
2565 xlog_recover_item_t *item,
2566 int pass)
2567{
2568 xfs_qoff_logformat_t *qoff_f;
2569
2570 if (pass == XLOG_RECOVER_PASS2) {
2571 return (0);
2572 }
2573
2574 qoff_f = (xfs_qoff_logformat_t *)item->ri_buf[0].i_addr;
2575 ASSERT(qoff_f);
2576
2577 /*
2578 * The logitem format's flag tells us if this was user quotaoff,
2579 * group quotaoff or both.
2580 */
2581 if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2582 log->l_quotaoffs_flag |= XFS_DQ_USER;
2583 if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2584 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2585
2586 return (0);
2587}
2588
2589/*
2590 * Recover a dquot record
2591 */
2592STATIC int
2593xlog_recover_do_dquot_trans(
2594 xlog_t *log,
2595 xlog_recover_item_t *item,
2596 int pass)
2597{
2598 xfs_mount_t *mp;
2599 xfs_buf_t *bp;
2600 struct xfs_disk_dquot *ddq, *recddq;
2601 int error;
2602 xfs_dq_logformat_t *dq_f;
2603 uint type;
2604
2605 if (pass == XLOG_RECOVER_PASS1) {
2606 return 0;
2607 }
2608 mp = log->l_mp;
2609
2610 /*
2611 * Filesystems are required to send in quota flags at mount time.
2612 */
2613 if (mp->m_qflags == 0)
2614 return (0);
2615
2616 recddq = (xfs_disk_dquot_t *)item->ri_buf[1].i_addr;
2617 ASSERT(recddq);
2618 /*
2619 * This type of quotas was turned off, so ignore this record.
2620 */
2621 type = INT_GET(recddq->d_flags, ARCH_CONVERT) &
2622 (XFS_DQ_USER | XFS_DQ_GROUP);
2623 ASSERT(type);
2624 if (log->l_quotaoffs_flag & type)
2625 return (0);
2626
2627 /*
2628 * At this point we know that quota was _not_ turned off.
2629 * Since the mount flags are not indicating to us otherwise, this
2630 * must mean that quota is on, and the dquot needs to be replayed.
2631 * Remember that we may not have fully recovered the superblock yet,
2632 * so we can't do the usual trick of looking at the SB quota bits.
2633 *
2634 * The other possibility, of course, is that the quota subsystem was
2635 * removed since the last mount - ENOSYS.
2636 */
2637 dq_f = (xfs_dq_logformat_t *)item->ri_buf[0].i_addr;
2638 ASSERT(dq_f);
2639 if ((error = xfs_qm_dqcheck(recddq,
2640 dq_f->qlf_id,
2641 0, XFS_QMOPT_DOWARN,
2642 "xlog_recover_do_dquot_trans (log copy)"))) {
2643 return XFS_ERROR(EIO);
2644 }
2645 ASSERT(dq_f->qlf_len == 1);
2646
2647 error = xfs_read_buf(mp, mp->m_ddev_targp,
2648 dq_f->qlf_blkno,
2649 XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2650 0, &bp);
2651 if (error) {
2652 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2653 bp, dq_f->qlf_blkno);
2654 return error;
2655 }
2656 ASSERT(bp);
2657 ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2658
2659 /*
2660 * At least the magic num portion should be on disk because this
2661 * was among a chunk of dquots created earlier, and we did some
2662 * minimal initialization then.
2663 */
2664 if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2665 "xlog_recover_do_dquot_trans")) {
2666 xfs_buf_relse(bp);
2667 return XFS_ERROR(EIO);
2668 }
2669
2670 memcpy(ddq, recddq, item->ri_buf[1].i_len);
2671
2672 ASSERT(dq_f->qlf_size == 2);
2673 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2674 XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2675 XFS_BUF_SET_FSPRIVATE(bp, mp);
2676 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2677 xfs_bdwrite(mp, bp);
2678
2679 return (0);
2680}
2681
2682/*
2683 * This routine is called to create an in-core extent free intent
2684 * item from the efi format structure which was logged on disk.
2685 * It allocates an in-core efi, copies the extents from the format
2686 * structure into it, and adds the efi to the AIL with the given
2687 * LSN.
2688 */
2689STATIC void
2690xlog_recover_do_efi_trans(
2691 xlog_t *log,
2692 xlog_recover_item_t *item,
2693 xfs_lsn_t lsn,
2694 int pass)
2695{
2696 xfs_mount_t *mp;
2697 xfs_efi_log_item_t *efip;
2698 xfs_efi_log_format_t *efi_formatp;
2699 SPLDECL(s);
2700
2701 if (pass == XLOG_RECOVER_PASS1) {
2702 return;
2703 }
2704
2705 efi_formatp = (xfs_efi_log_format_t *)item->ri_buf[0].i_addr;
2706 ASSERT(item->ri_buf[0].i_len ==
2707 (sizeof(xfs_efi_log_format_t) +
2708 ((efi_formatp->efi_nextents - 1) * sizeof(xfs_extent_t))));
2709
2710 mp = log->l_mp;
2711 efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2712 memcpy((char *)&(efip->efi_format), (char *)efi_formatp,
2713 sizeof(xfs_efi_log_format_t) +
2714 ((efi_formatp->efi_nextents - 1) * sizeof(xfs_extent_t)));
2715 efip->efi_next_extent = efi_formatp->efi_nextents;
2716 efip->efi_flags |= XFS_EFI_COMMITTED;
2717
2718 AIL_LOCK(mp,s);
2719 /*
2720 * xfs_trans_update_ail() drops the AIL lock.
2721 */
2722 xfs_trans_update_ail(mp, (xfs_log_item_t *)efip, lsn, s);
2723}
2724
2725
2726/*
2727 * This routine is called when an efd format structure is found in
2728 * a committed transaction in the log. It's purpose is to cancel
2729 * the corresponding efi if it was still in the log. To do this
2730 * it searches the AIL for the efi with an id equal to that in the
2731 * efd format structure. If we find it, we remove the efi from the
2732 * AIL and free it.
2733 */
2734STATIC void
2735xlog_recover_do_efd_trans(
2736 xlog_t *log,
2737 xlog_recover_item_t *item,
2738 int pass)
2739{
2740 xfs_mount_t *mp;
2741 xfs_efd_log_format_t *efd_formatp;
2742 xfs_efi_log_item_t *efip = NULL;
2743 xfs_log_item_t *lip;
2744 int gen;
2745 int nexts;
2746 __uint64_t efi_id;
2747 SPLDECL(s);
2748
2749 if (pass == XLOG_RECOVER_PASS1) {
2750 return;
2751 }
2752
2753 efd_formatp = (xfs_efd_log_format_t *)item->ri_buf[0].i_addr;
2754 ASSERT(item->ri_buf[0].i_len ==
2755 (sizeof(xfs_efd_log_format_t) +
2756 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_t))));
2757 efi_id = efd_formatp->efd_efi_id;
2758
2759 /*
2760 * Search for the efi with the id in the efd format structure
2761 * in the AIL.
2762 */
2763 mp = log->l_mp;
2764 AIL_LOCK(mp,s);
2765 lip = xfs_trans_first_ail(mp, &gen);
2766 while (lip != NULL) {
2767 if (lip->li_type == XFS_LI_EFI) {
2768 efip = (xfs_efi_log_item_t *)lip;
2769 if (efip->efi_format.efi_id == efi_id) {
2770 /*
2771 * xfs_trans_delete_ail() drops the
2772 * AIL lock.
2773 */
2774 xfs_trans_delete_ail(mp, lip, s);
2775 break;
2776 }
2777 }
2778 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
2779 }
2780 if (lip == NULL) {
2781 AIL_UNLOCK(mp, s);
2782 }
2783
2784 /*
2785 * If we found it, then free it up. If it wasn't there, it
2786 * must have been overwritten in the log. Oh well.
2787 */
2788 if (lip != NULL) {
2789 nexts = efip->efi_format.efi_nextents;
2790 if (nexts > XFS_EFI_MAX_FAST_EXTENTS) {
2791 kmem_free(lip, sizeof(xfs_efi_log_item_t) +
2792 ((nexts - 1) * sizeof(xfs_extent_t)));
2793 } else {
2794 kmem_zone_free(xfs_efi_zone, efip);
2795 }
2796 }
2797}
2798
2799/*
2800 * Perform the transaction
2801 *
2802 * If the transaction modifies a buffer or inode, do it now. Otherwise,
2803 * EFIs and EFDs get queued up by adding entries into the AIL for them.
2804 */
2805STATIC int
2806xlog_recover_do_trans(
2807 xlog_t *log,
2808 xlog_recover_t *trans,
2809 int pass)
2810{
2811 int error = 0;
2812 xlog_recover_item_t *item, *first_item;
2813
2814 if ((error = xlog_recover_reorder_trans(log, trans)))
2815 return error;
2816 first_item = item = trans->r_itemq;
2817 do {
2818 /*
2819 * we don't need to worry about the block number being
2820 * truncated in > 1 TB buffers because in user-land,
2821 * we're now n32 or 64-bit so xfs_daddr_t is 64-bits so
2822 * the blkno's will get through the user-mode buffer
2823 * cache properly. The only bad case is o32 kernels
2824 * where xfs_daddr_t is 32-bits but mount will warn us
2825 * off a > 1 TB filesystem before we get here.
2826 */
2827 if ((ITEM_TYPE(item) == XFS_LI_BUF) ||
2828 (ITEM_TYPE(item) == XFS_LI_6_1_BUF) ||
2829 (ITEM_TYPE(item) == XFS_LI_5_3_BUF)) {
2830 if ((error = xlog_recover_do_buffer_trans(log, item,
2831 pass)))
2832 break;
2833 } else if ((ITEM_TYPE(item) == XFS_LI_INODE) ||
2834 (ITEM_TYPE(item) == XFS_LI_6_1_INODE) ||
2835 (ITEM_TYPE(item) == XFS_LI_5_3_INODE)) {
2836 if ((error = xlog_recover_do_inode_trans(log, item,
2837 pass)))
2838 break;
2839 } else if (ITEM_TYPE(item) == XFS_LI_EFI) {
2840 xlog_recover_do_efi_trans(log, item, trans->r_lsn,
2841 pass);
2842 } else if (ITEM_TYPE(item) == XFS_LI_EFD) {
2843 xlog_recover_do_efd_trans(log, item, pass);
2844 } else if (ITEM_TYPE(item) == XFS_LI_DQUOT) {
2845 if ((error = xlog_recover_do_dquot_trans(log, item,
2846 pass)))
2847 break;
2848 } else if ((ITEM_TYPE(item) == XFS_LI_QUOTAOFF)) {
2849 if ((error = xlog_recover_do_quotaoff_trans(log, item,
2850 pass)))
2851 break;
2852 } else {
2853 xlog_warn("XFS: xlog_recover_do_trans");
2854 ASSERT(0);
2855 error = XFS_ERROR(EIO);
2856 break;
2857 }
2858 item = item->ri_next;
2859 } while (first_item != item);
2860
2861 return error;
2862}
2863
2864/*
2865 * Free up any resources allocated by the transaction
2866 *
2867 * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2868 */
2869STATIC void
2870xlog_recover_free_trans(
2871 xlog_recover_t *trans)
2872{
2873 xlog_recover_item_t *first_item, *item, *free_item;
2874 int i;
2875
2876 item = first_item = trans->r_itemq;
2877 do {
2878 free_item = item;
2879 item = item->ri_next;
2880 /* Free the regions in the item. */
2881 for (i = 0; i < free_item->ri_cnt; i++) {
2882 kmem_free(free_item->ri_buf[i].i_addr,
2883 free_item->ri_buf[i].i_len);
2884 }
2885 /* Free the item itself */
2886 kmem_free(free_item->ri_buf,
2887 (free_item->ri_total * sizeof(xfs_log_iovec_t)));
2888 kmem_free(free_item, sizeof(xlog_recover_item_t));
2889 } while (first_item != item);
2890 /* Free the transaction recover structure */
2891 kmem_free(trans, sizeof(xlog_recover_t));
2892}
2893
2894STATIC int
2895xlog_recover_commit_trans(
2896 xlog_t *log,
2897 xlog_recover_t **q,
2898 xlog_recover_t *trans,
2899 int pass)
2900{
2901 int error;
2902
2903 if ((error = xlog_recover_unlink_tid(q, trans)))
2904 return error;
2905 if ((error = xlog_recover_do_trans(log, trans, pass)))
2906 return error;
2907 xlog_recover_free_trans(trans); /* no error */
2908 return 0;
2909}
2910
2911STATIC int
2912xlog_recover_unmount_trans(
2913 xlog_recover_t *trans)
2914{
2915 /* Do nothing now */
2916 xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2917 return 0;
2918}
2919
2920/*
2921 * There are two valid states of the r_state field. 0 indicates that the
2922 * transaction structure is in a normal state. We have either seen the
2923 * start of the transaction or the last operation we added was not a partial
2924 * operation. If the last operation we added to the transaction was a
2925 * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2926 *
2927 * NOTE: skip LRs with 0 data length.
2928 */
2929STATIC int
2930xlog_recover_process_data(
2931 xlog_t *log,
2932 xlog_recover_t *rhash[],
2933 xlog_rec_header_t *rhead,
2934 xfs_caddr_t dp,
2935 int pass)
2936{
2937 xfs_caddr_t lp;
2938 int num_logops;
2939 xlog_op_header_t *ohead;
2940 xlog_recover_t *trans;
2941 xlog_tid_t tid;
2942 int error;
2943 unsigned long hash;
2944 uint flags;
2945
2946 lp = dp + INT_GET(rhead->h_len, ARCH_CONVERT);
2947 num_logops = INT_GET(rhead->h_num_logops, ARCH_CONVERT);
2948
2949 /* check the log format matches our own - else we can't recover */
2950 if (xlog_header_check_recover(log->l_mp, rhead))
2951 return (XFS_ERROR(EIO));
2952
2953 while ((dp < lp) && num_logops) {
2954 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2955 ohead = (xlog_op_header_t *)dp;
2956 dp += sizeof(xlog_op_header_t);
2957 if (ohead->oh_clientid != XFS_TRANSACTION &&
2958 ohead->oh_clientid != XFS_LOG) {
2959 xlog_warn(
2960 "XFS: xlog_recover_process_data: bad clientid");
2961 ASSERT(0);
2962 return (XFS_ERROR(EIO));
2963 }
2964 tid = INT_GET(ohead->oh_tid, ARCH_CONVERT);
2965 hash = XLOG_RHASH(tid);
2966 trans = xlog_recover_find_tid(rhash[hash], tid);
2967 if (trans == NULL) { /* not found; add new tid */
2968 if (ohead->oh_flags & XLOG_START_TRANS)
2969 xlog_recover_new_tid(&rhash[hash], tid,
2970 INT_GET(rhead->h_lsn, ARCH_CONVERT));
2971 } else {
2972 ASSERT(dp+INT_GET(ohead->oh_len, ARCH_CONVERT) <= lp);
2973 flags = ohead->oh_flags & ~XLOG_END_TRANS;
2974 if (flags & XLOG_WAS_CONT_TRANS)
2975 flags &= ~XLOG_CONTINUE_TRANS;
2976 switch (flags) {
2977 case XLOG_COMMIT_TRANS:
2978 error = xlog_recover_commit_trans(log,
2979 &rhash[hash], trans, pass);
2980 break;
2981 case XLOG_UNMOUNT_TRANS:
2982 error = xlog_recover_unmount_trans(trans);
2983 break;
2984 case XLOG_WAS_CONT_TRANS:
2985 error = xlog_recover_add_to_cont_trans(trans,
2986 dp, INT_GET(ohead->oh_len,
2987 ARCH_CONVERT));
2988 break;
2989 case XLOG_START_TRANS:
2990 xlog_warn(
2991 "XFS: xlog_recover_process_data: bad transaction");
2992 ASSERT(0);
2993 error = XFS_ERROR(EIO);
2994 break;
2995 case 0:
2996 case XLOG_CONTINUE_TRANS:
2997 error = xlog_recover_add_to_trans(trans,
2998 dp, INT_GET(ohead->oh_len,
2999 ARCH_CONVERT));
3000 break;
3001 default:
3002 xlog_warn(
3003 "XFS: xlog_recover_process_data: bad flag");
3004 ASSERT(0);
3005 error = XFS_ERROR(EIO);
3006 break;
3007 }
3008 if (error)
3009 return error;
3010 }
3011 dp += INT_GET(ohead->oh_len, ARCH_CONVERT);
3012 num_logops--;
3013 }
3014 return 0;
3015}
3016
3017/*
3018 * Process an extent free intent item that was recovered from
3019 * the log. We need to free the extents that it describes.
3020 */
3021STATIC void
3022xlog_recover_process_efi(
3023 xfs_mount_t *mp,
3024 xfs_efi_log_item_t *efip)
3025{
3026 xfs_efd_log_item_t *efdp;
3027 xfs_trans_t *tp;
3028 int i;
3029 xfs_extent_t *extp;
3030 xfs_fsblock_t startblock_fsb;
3031
3032 ASSERT(!(efip->efi_flags & XFS_EFI_RECOVERED));
3033
3034 /*
3035 * First check the validity of the extents described by the
3036 * EFI. If any are bad, then assume that all are bad and
3037 * just toss the EFI.
3038 */
3039 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3040 extp = &(efip->efi_format.efi_extents[i]);
3041 startblock_fsb = XFS_BB_TO_FSB(mp,
3042 XFS_FSB_TO_DADDR(mp, extp->ext_start));
3043 if ((startblock_fsb == 0) ||
3044 (extp->ext_len == 0) ||
3045 (startblock_fsb >= mp->m_sb.sb_dblocks) ||
3046 (extp->ext_len >= mp->m_sb.sb_agblocks)) {
3047 /*
3048 * This will pull the EFI from the AIL and
3049 * free the memory associated with it.
3050 */
3051 xfs_efi_release(efip, efip->efi_format.efi_nextents);
3052 return;
3053 }
3054 }
3055
3056 tp = xfs_trans_alloc(mp, 0);
3057 xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
3058 efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3059
3060 for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3061 extp = &(efip->efi_format.efi_extents[i]);
3062 xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3063 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3064 extp->ext_len);
3065 }
3066
3067 efip->efi_flags |= XFS_EFI_RECOVERED;
3068 xfs_trans_commit(tp, 0, NULL);
3069}
3070
3071/*
3072 * Verify that once we've encountered something other than an EFI
3073 * in the AIL that there are no more EFIs in the AIL.
3074 */
3075#if defined(DEBUG)
3076STATIC void
3077xlog_recover_check_ail(
3078 xfs_mount_t *mp,
3079 xfs_log_item_t *lip,
3080 int gen)
3081{
3082 int orig_gen = gen;
3083
3084 do {
3085 ASSERT(lip->li_type != XFS_LI_EFI);
3086 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3087 /*
3088 * The check will be bogus if we restart from the
3089 * beginning of the AIL, so ASSERT that we don't.
3090 * We never should since we're holding the AIL lock
3091 * the entire time.
3092 */
3093 ASSERT(gen == orig_gen);
3094 } while (lip != NULL);
3095}
3096#endif /* DEBUG */
3097
3098/*
3099 * When this is called, all of the EFIs which did not have
3100 * corresponding EFDs should be in the AIL. What we do now
3101 * is free the extents associated with each one.
3102 *
3103 * Since we process the EFIs in normal transactions, they
3104 * will be removed at some point after the commit. This prevents
3105 * us from just walking down the list processing each one.
3106 * We'll use a flag in the EFI to skip those that we've already
3107 * processed and use the AIL iteration mechanism's generation
3108 * count to try to speed this up at least a bit.
3109 *
3110 * When we start, we know that the EFIs are the only things in
3111 * the AIL. As we process them, however, other items are added
3112 * to the AIL. Since everything added to the AIL must come after
3113 * everything already in the AIL, we stop processing as soon as
3114 * we see something other than an EFI in the AIL.
3115 */
3116STATIC void
3117xlog_recover_process_efis(
3118 xlog_t *log)
3119{
3120 xfs_log_item_t *lip;
3121 xfs_efi_log_item_t *efip;
3122 int gen;
3123 xfs_mount_t *mp;
3124 SPLDECL(s);
3125
3126 mp = log->l_mp;
3127 AIL_LOCK(mp,s);
3128
3129 lip = xfs_trans_first_ail(mp, &gen);
3130 while (lip != NULL) {
3131 /*
3132 * We're done when we see something other than an EFI.
3133 */
3134 if (lip->li_type != XFS_LI_EFI) {
3135 xlog_recover_check_ail(mp, lip, gen);
3136 break;
3137 }
3138
3139 /*
3140 * Skip EFIs that we've already processed.
3141 */
3142 efip = (xfs_efi_log_item_t *)lip;
3143 if (efip->efi_flags & XFS_EFI_RECOVERED) {
3144 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3145 continue;
3146 }
3147
3148 AIL_UNLOCK(mp, s);
3149 xlog_recover_process_efi(mp, efip);
3150 AIL_LOCK(mp,s);
3151 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3152 }
3153 AIL_UNLOCK(mp, s);
3154}
3155
3156/*
3157 * This routine performs a transaction to null out a bad inode pointer
3158 * in an agi unlinked inode hash bucket.
3159 */
3160STATIC void
3161xlog_recover_clear_agi_bucket(
3162 xfs_mount_t *mp,
3163 xfs_agnumber_t agno,
3164 int bucket)
3165{
3166 xfs_trans_t *tp;
3167 xfs_agi_t *agi;
3168 xfs_buf_t *agibp;
3169 int offset;
3170 int error;
3171
3172 tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3173 xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp), 0, 0, 0);
3174
3175 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
3176 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3177 XFS_FSS_TO_BB(mp, 1), 0, &agibp);
3178 if (error) {
3179 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3180 return;
3181 }
3182
3183 agi = XFS_BUF_TO_AGI(agibp);
3184 if (INT_GET(agi->agi_magicnum, ARCH_CONVERT) != XFS_AGI_MAGIC) {
3185 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3186 return;
3187 }
3188 ASSERT(INT_GET(agi->agi_magicnum, ARCH_CONVERT) == XFS_AGI_MAGIC);
3189
3190 INT_SET(agi->agi_unlinked[bucket], ARCH_CONVERT, NULLAGINO);
3191 offset = offsetof(xfs_agi_t, agi_unlinked) +
3192 (sizeof(xfs_agino_t) * bucket);
3193 xfs_trans_log_buf(tp, agibp, offset,
3194 (offset + sizeof(xfs_agino_t) - 1));
3195
3196 (void) xfs_trans_commit(tp, 0, NULL);
3197}
3198
3199/*
3200 * xlog_iunlink_recover
3201 *
3202 * This is called during recovery to process any inodes which
3203 * we unlinked but not freed when the system crashed. These
3204 * inodes will be on the lists in the AGI blocks. What we do
3205 * here is scan all the AGIs and fully truncate and free any
3206 * inodes found on the lists. Each inode is removed from the
3207 * lists when it has been fully truncated and is freed. The
3208 * freeing of the inode and its removal from the list must be
3209 * atomic.
3210 */
3211void
3212xlog_recover_process_iunlinks(
3213 xlog_t *log)
3214{
3215 xfs_mount_t *mp;
3216 xfs_agnumber_t agno;
3217 xfs_agi_t *agi;
3218 xfs_buf_t *agibp;
3219 xfs_buf_t *ibp;
3220 xfs_dinode_t *dip;
3221 xfs_inode_t *ip;
3222 xfs_agino_t agino;
3223 xfs_ino_t ino;
3224 int bucket;
3225 int error;
3226 uint mp_dmevmask;
3227
3228 mp = log->l_mp;
3229
3230 /*
3231 * Prevent any DMAPI event from being sent while in this function.
3232 */
3233 mp_dmevmask = mp->m_dmevmask;
3234 mp->m_dmevmask = 0;
3235
3236 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3237 /*
3238 * Find the agi for this ag.
3239 */
3240 agibp = xfs_buf_read(mp->m_ddev_targp,
3241 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3242 XFS_FSS_TO_BB(mp, 1), 0);
3243 if (XFS_BUF_ISERROR(agibp)) {
3244 xfs_ioerror_alert("xlog_recover_process_iunlinks(#1)",
3245 log->l_mp, agibp,
3246 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)));
3247 }
3248 agi = XFS_BUF_TO_AGI(agibp);
3249 ASSERT(XFS_AGI_MAGIC ==
3250 INT_GET(agi->agi_magicnum, ARCH_CONVERT));
3251
3252 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3253
3254 agino = INT_GET(agi->agi_unlinked[bucket], ARCH_CONVERT);
3255 while (agino != NULLAGINO) {
3256
3257 /*
3258 * Release the agi buffer so that it can
3259 * be acquired in the normal course of the
3260 * transaction to truncate and free the inode.
3261 */
3262 xfs_buf_relse(agibp);
3263
3264 ino = XFS_AGINO_TO_INO(mp, agno, agino);
3265 error = xfs_iget(mp, NULL, ino, 0, 0, &ip, 0);
3266 ASSERT(error || (ip != NULL));
3267
3268 if (!error) {
3269 /*
3270 * Get the on disk inode to find the
3271 * next inode in the bucket.
3272 */
3273 error = xfs_itobp(mp, NULL, ip, &dip,
3274 &ibp, 0);
3275 ASSERT(error || (dip != NULL));
3276 }
3277
3278 if (!error) {
3279 ASSERT(ip->i_d.di_nlink == 0);
3280
3281 /* setup for the next pass */
3282 agino = INT_GET(dip->di_next_unlinked,
3283 ARCH_CONVERT);
3284 xfs_buf_relse(ibp);
3285 /*
3286 * Prevent any DMAPI event from
3287 * being sent when the
3288 * reference on the inode is
3289 * dropped.
3290 */
3291 ip->i_d.di_dmevmask = 0;
3292
3293 /*
3294 * If this is a new inode, handle
3295 * it specially. Otherwise,
3296 * just drop our reference to the
3297 * inode. If there are no
3298 * other references, this will
3299 * send the inode to
3300 * xfs_inactive() which will
3301 * truncate the file and free
3302 * the inode.
3303 */
3304 if (ip->i_d.di_mode == 0)
3305 xfs_iput_new(ip, 0);
3306 else
3307 VN_RELE(XFS_ITOV(ip));
3308 } else {
3309 /*
3310 * We can't read in the inode
3311 * this bucket points to, or
3312 * this inode is messed up. Just
3313 * ditch this bucket of inodes. We
3314 * will lose some inodes and space,
3315 * but at least we won't hang. Call
3316 * xlog_recover_clear_agi_bucket()
3317 * to perform a transaction to clear
3318 * the inode pointer in the bucket.
3319 */
3320 xlog_recover_clear_agi_bucket(mp, agno,
3321 bucket);
3322
3323 agino = NULLAGINO;
3324 }
3325
3326 /*
3327 * Reacquire the agibuffer and continue around
3328 * the loop.
3329 */
3330 agibp = xfs_buf_read(mp->m_ddev_targp,
3331 XFS_AG_DADDR(mp, agno,
3332 XFS_AGI_DADDR(mp)),
3333 XFS_FSS_TO_BB(mp, 1), 0);
3334 if (XFS_BUF_ISERROR(agibp)) {
3335 xfs_ioerror_alert(
3336 "xlog_recover_process_iunlinks(#2)",
3337 log->l_mp, agibp,
3338 XFS_AG_DADDR(mp, agno,
3339 XFS_AGI_DADDR(mp)));
3340 }
3341 agi = XFS_BUF_TO_AGI(agibp);
3342 ASSERT(XFS_AGI_MAGIC == INT_GET(
3343 agi->agi_magicnum, ARCH_CONVERT));
3344 }
3345 }
3346
3347 /*
3348 * Release the buffer for the current agi so we can
3349 * go on to the next one.
3350 */
3351 xfs_buf_relse(agibp);
3352 }
3353
3354 mp->m_dmevmask = mp_dmevmask;
3355}
3356
3357
3358#ifdef DEBUG
3359STATIC void
3360xlog_pack_data_checksum(
3361 xlog_t *log,
3362 xlog_in_core_t *iclog,
3363 int size)
3364{
3365 int i;
3366 uint *up;
3367 uint chksum = 0;
3368
3369 up = (uint *)iclog->ic_datap;
3370 /* divide length by 4 to get # words */
3371 for (i = 0; i < (size >> 2); i++) {
3372 chksum ^= INT_GET(*up, ARCH_CONVERT);
3373 up++;
3374 }
3375 INT_SET(iclog->ic_header.h_chksum, ARCH_CONVERT, chksum);
3376}
3377#else
3378#define xlog_pack_data_checksum(log, iclog, size)
3379#endif
3380
3381/*
3382 * Stamp cycle number in every block
3383 */
3384void
3385xlog_pack_data(
3386 xlog_t *log,
3387 xlog_in_core_t *iclog,
3388 int roundoff)
3389{
3390 int i, j, k;
3391 int size = iclog->ic_offset + roundoff;
3392 uint cycle_lsn;
3393 xfs_caddr_t dp;
3394 xlog_in_core_2_t *xhdr;
3395
3396 xlog_pack_data_checksum(log, iclog, size);
3397
3398 cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3399
3400 dp = iclog->ic_datap;
3401 for (i = 0; i < BTOBB(size) &&
3402 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3403 iclog->ic_header.h_cycle_data[i] = *(uint *)dp;
3404 *(uint *)dp = cycle_lsn;
3405 dp += BBSIZE;
3406 }
3407
3408 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3409 xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3410 for ( ; i < BTOBB(size); i++) {
3411 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3412 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3413 xhdr[j].hic_xheader.xh_cycle_data[k] = *(uint *)dp;
3414 *(uint *)dp = cycle_lsn;
3415 dp += BBSIZE;
3416 }
3417
3418 for (i = 1; i < log->l_iclog_heads; i++) {
3419 xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3420 }
3421 }
3422}
3423
3424#if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
3425STATIC void
3426xlog_unpack_data_checksum(
3427 xlog_rec_header_t *rhead,
3428 xfs_caddr_t dp,
3429 xlog_t *log)
3430{
3431 uint *up = (uint *)dp;
3432 uint chksum = 0;
3433 int i;
3434
3435 /* divide length by 4 to get # words */
3436 for (i=0; i < INT_GET(rhead->h_len, ARCH_CONVERT) >> 2; i++) {
3437 chksum ^= INT_GET(*up, ARCH_CONVERT);
3438 up++;
3439 }
3440 if (chksum != INT_GET(rhead->h_chksum, ARCH_CONVERT)) {
3441 if (rhead->h_chksum ||
3442 ((log->l_flags & XLOG_CHKSUM_MISMATCH) == 0)) {
3443 cmn_err(CE_DEBUG,
3444 "XFS: LogR chksum mismatch: was (0x%x) is (0x%x)",
3445 INT_GET(rhead->h_chksum, ARCH_CONVERT), chksum);
3446 cmn_err(CE_DEBUG,
3447"XFS: Disregard message if filesystem was created with non-DEBUG kernel");
3448 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3449 cmn_err(CE_DEBUG,
3450 "XFS: LogR this is a LogV2 filesystem");
3451 }
3452 log->l_flags |= XLOG_CHKSUM_MISMATCH;
3453 }
3454 }
3455}
3456#else
3457#define xlog_unpack_data_checksum(rhead, dp, log)
3458#endif
3459
3460STATIC void
3461xlog_unpack_data(
3462 xlog_rec_header_t *rhead,
3463 xfs_caddr_t dp,
3464 xlog_t *log)
3465{
3466 int i, j, k;
3467 xlog_in_core_2_t *xhdr;
3468
3469 for (i = 0; i < BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT)) &&
3470 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3471 *(uint *)dp = *(uint *)&rhead->h_cycle_data[i];
3472 dp += BBSIZE;
3473 }
3474
3475 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3476 xhdr = (xlog_in_core_2_t *)rhead;
3477 for ( ; i < BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT)); i++) {
3478 j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3479 k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3480 *(uint *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3481 dp += BBSIZE;
3482 }
3483 }
3484
3485 xlog_unpack_data_checksum(rhead, dp, log);
3486}
3487
3488STATIC int
3489xlog_valid_rec_header(
3490 xlog_t *log,
3491 xlog_rec_header_t *rhead,
3492 xfs_daddr_t blkno)
3493{
3494 int hlen;
3495
3496 if (unlikely(
3497 (INT_GET(rhead->h_magicno, ARCH_CONVERT) !=
3498 XLOG_HEADER_MAGIC_NUM))) {
3499 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3500 XFS_ERRLEVEL_LOW, log->l_mp);
3501 return XFS_ERROR(EFSCORRUPTED);
3502 }
3503 if (unlikely(
3504 (!rhead->h_version ||
3505 (INT_GET(rhead->h_version, ARCH_CONVERT) &
3506 (~XLOG_VERSION_OKBITS)) != 0))) {
3507 xlog_warn("XFS: %s: unrecognised log version (%d).",
3508 __FUNCTION__, INT_GET(rhead->h_version, ARCH_CONVERT));
3509 return XFS_ERROR(EIO);
3510 }
3511
3512 /* LR body must have data or it wouldn't have been written */
3513 hlen = INT_GET(rhead->h_len, ARCH_CONVERT);
3514 if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3515 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3516 XFS_ERRLEVEL_LOW, log->l_mp);
3517 return XFS_ERROR(EFSCORRUPTED);
3518 }
3519 if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3520 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3521 XFS_ERRLEVEL_LOW, log->l_mp);
3522 return XFS_ERROR(EFSCORRUPTED);
3523 }
3524 return 0;
3525}
3526
3527/*
3528 * Read the log from tail to head and process the log records found.
3529 * Handle the two cases where the tail and head are in the same cycle
3530 * and where the active portion of the log wraps around the end of
3531 * the physical log separately. The pass parameter is passed through
3532 * to the routines called to process the data and is not looked at
3533 * here.
3534 */
3535STATIC int
3536xlog_do_recovery_pass(
3537 xlog_t *log,
3538 xfs_daddr_t head_blk,
3539 xfs_daddr_t tail_blk,
3540 int pass)
3541{
3542 xlog_rec_header_t *rhead;
3543 xfs_daddr_t blk_no;
3544 xfs_caddr_t bufaddr, offset;
3545 xfs_buf_t *hbp, *dbp;
3546 int error = 0, h_size;
3547 int bblks, split_bblks;
3548 int hblks, split_hblks, wrapped_hblks;
3549 xlog_recover_t *rhash[XLOG_RHASH_SIZE];
3550
3551 ASSERT(head_blk != tail_blk);
3552
3553 /*
3554 * Read the header of the tail block and get the iclog buffer size from
3555 * h_size. Use this to tell how many sectors make up the log header.
3556 */
3557 if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3558 /*
3559 * When using variable length iclogs, read first sector of
3560 * iclog header and extract the header size from it. Get a
3561 * new hbp that is the correct size.
3562 */
3563 hbp = xlog_get_bp(log, 1);
3564 if (!hbp)
3565 return ENOMEM;
3566 if ((error = xlog_bread(log, tail_blk, 1, hbp)))
3567 goto bread_err1;
3568 offset = xlog_align(log, tail_blk, 1, hbp);
3569 rhead = (xlog_rec_header_t *)offset;
3570 error = xlog_valid_rec_header(log, rhead, tail_blk);
3571 if (error)
3572 goto bread_err1;
3573 h_size = INT_GET(rhead->h_size, ARCH_CONVERT);
3574 if ((INT_GET(rhead->h_version, ARCH_CONVERT)
3575 & XLOG_VERSION_2) &&
3576 (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3577 hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3578 if (h_size % XLOG_HEADER_CYCLE_SIZE)
3579 hblks++;
3580 xlog_put_bp(hbp);
3581 hbp = xlog_get_bp(log, hblks);
3582 } else {
3583 hblks = 1;
3584 }
3585 } else {
3586 ASSERT(log->l_sectbb_log == 0);
3587 hblks = 1;
3588 hbp = xlog_get_bp(log, 1);
3589 h_size = XLOG_BIG_RECORD_BSIZE;
3590 }
3591
3592 if (!hbp)
3593 return ENOMEM;
3594 dbp = xlog_get_bp(log, BTOBB(h_size));
3595 if (!dbp) {
3596 xlog_put_bp(hbp);
3597 return ENOMEM;
3598 }
3599
3600 memset(rhash, 0, sizeof(rhash));
3601 if (tail_blk <= head_blk) {
3602 for (blk_no = tail_blk; blk_no < head_blk; ) {
3603 if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3604 goto bread_err2;
3605 offset = xlog_align(log, blk_no, hblks, hbp);
3606 rhead = (xlog_rec_header_t *)offset;
3607 error = xlog_valid_rec_header(log, rhead, blk_no);
3608 if (error)
3609 goto bread_err2;
3610
3611 /* blocks in data section */
3612 bblks = (int)BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT));
3613 error = xlog_bread(log, blk_no + hblks, bblks, dbp);
3614 if (error)
3615 goto bread_err2;
3616 offset = xlog_align(log, blk_no + hblks, bblks, dbp);
3617 xlog_unpack_data(rhead, offset, log);
3618 if ((error = xlog_recover_process_data(log,
3619 rhash, rhead, offset, pass)))
3620 goto bread_err2;
3621 blk_no += bblks + hblks;
3622 }
3623 } else {
3624 /*
3625 * Perform recovery around the end of the physical log.
3626 * When the head is not on the same cycle number as the tail,
3627 * we can't do a sequential recovery as above.
3628 */
3629 blk_no = tail_blk;
3630 while (blk_no < log->l_logBBsize) {
3631 /*
3632 * Check for header wrapping around physical end-of-log
3633 */
3634 offset = NULL;
3635 split_hblks = 0;
3636 wrapped_hblks = 0;
3637 if (blk_no + hblks <= log->l_logBBsize) {
3638 /* Read header in one read */
3639 error = xlog_bread(log, blk_no, hblks, hbp);
3640 if (error)
3641 goto bread_err2;
3642 offset = xlog_align(log, blk_no, hblks, hbp);
3643 } else {
3644 /* This LR is split across physical log end */
3645 if (blk_no != log->l_logBBsize) {
3646 /* some data before physical log end */
3647 ASSERT(blk_no <= INT_MAX);
3648 split_hblks = log->l_logBBsize - (int)blk_no;
3649 ASSERT(split_hblks > 0);
3650 if ((error = xlog_bread(log, blk_no,
3651 split_hblks, hbp)))
3652 goto bread_err2;
3653 offset = xlog_align(log, blk_no,
3654 split_hblks, hbp);
3655 }
3656 /*
3657 * Note: this black magic still works with
3658 * large sector sizes (non-512) only because:
3659 * - we increased the buffer size originally
3660 * by 1 sector giving us enough extra space
3661 * for the second read;
3662 * - the log start is guaranteed to be sector
3663 * aligned;
3664 * - we read the log end (LR header start)
3665 * _first_, then the log start (LR header end)
3666 * - order is important.
3667 */
3668 bufaddr = XFS_BUF_PTR(hbp);
3669 XFS_BUF_SET_PTR(hbp,
3670 bufaddr + BBTOB(split_hblks),
3671 BBTOB(hblks - split_hblks));
3672 wrapped_hblks = hblks - split_hblks;
3673 error = xlog_bread(log, 0, wrapped_hblks, hbp);
3674 if (error)
3675 goto bread_err2;
3676 XFS_BUF_SET_PTR(hbp, bufaddr, BBTOB(hblks));
3677 if (!offset)
3678 offset = xlog_align(log, 0,
3679 wrapped_hblks, hbp);
3680 }
3681 rhead = (xlog_rec_header_t *)offset;
3682 error = xlog_valid_rec_header(log, rhead,
3683 split_hblks ? blk_no : 0);
3684 if (error)
3685 goto bread_err2;
3686
3687 bblks = (int)BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT));
3688 blk_no += hblks;
3689
3690 /* Read in data for log record */
3691 if (blk_no + bblks <= log->l_logBBsize) {
3692 error = xlog_bread(log, blk_no, bblks, dbp);
3693 if (error)
3694 goto bread_err2;
3695 offset = xlog_align(log, blk_no, bblks, dbp);
3696 } else {
3697 /* This log record is split across the
3698 * physical end of log */
3699 offset = NULL;
3700 split_bblks = 0;
3701 if (blk_no != log->l_logBBsize) {
3702 /* some data is before the physical
3703 * end of log */
3704 ASSERT(!wrapped_hblks);
3705 ASSERT(blk_no <= INT_MAX);
3706 split_bblks =
3707 log->l_logBBsize - (int)blk_no;
3708 ASSERT(split_bblks > 0);
3709 if ((error = xlog_bread(log, blk_no,
3710 split_bblks, dbp)))
3711 goto bread_err2;
3712 offset = xlog_align(log, blk_no,
3713 split_bblks, dbp);
3714 }
3715 /*
3716 * Note: this black magic still works with
3717 * large sector sizes (non-512) only because:
3718 * - we increased the buffer size originally
3719 * by 1 sector giving us enough extra space
3720 * for the second read;
3721 * - the log start is guaranteed to be sector
3722 * aligned;
3723 * - we read the log end (LR header start)
3724 * _first_, then the log start (LR header end)
3725 * - order is important.
3726 */
3727 bufaddr = XFS_BUF_PTR(dbp);
3728 XFS_BUF_SET_PTR(dbp,
3729 bufaddr + BBTOB(split_bblks),
3730 BBTOB(bblks - split_bblks));
3731 if ((error = xlog_bread(log, wrapped_hblks,
3732 bblks - split_bblks, dbp)))
3733 goto bread_err2;
3734 XFS_BUF_SET_PTR(dbp, bufaddr, h_size);
3735 if (!offset)
3736 offset = xlog_align(log, wrapped_hblks,
3737 bblks - split_bblks, dbp);
3738 }
3739 xlog_unpack_data(rhead, offset, log);
3740 if ((error = xlog_recover_process_data(log, rhash,
3741 rhead, offset, pass)))
3742 goto bread_err2;
3743 blk_no += bblks;
3744 }
3745
3746 ASSERT(blk_no >= log->l_logBBsize);
3747 blk_no -= log->l_logBBsize;
3748
3749 /* read first part of physical log */
3750 while (blk_no < head_blk) {
3751 if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3752 goto bread_err2;
3753 offset = xlog_align(log, blk_no, hblks, hbp);
3754 rhead = (xlog_rec_header_t *)offset;
3755 error = xlog_valid_rec_header(log, rhead, blk_no);
3756 if (error)
3757 goto bread_err2;
3758 bblks = (int)BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT));
3759 if ((error = xlog_bread(log, blk_no+hblks, bblks, dbp)))
3760 goto bread_err2;
3761 offset = xlog_align(log, blk_no+hblks, bblks, dbp);
3762 xlog_unpack_data(rhead, offset, log);
3763 if ((error = xlog_recover_process_data(log, rhash,
3764 rhead, offset, pass)))
3765 goto bread_err2;
3766 blk_no += bblks + hblks;
3767 }
3768 }
3769
3770 bread_err2:
3771 xlog_put_bp(dbp);
3772 bread_err1:
3773 xlog_put_bp(hbp);
3774 return error;
3775}
3776
3777/*
3778 * Do the recovery of the log. We actually do this in two phases.
3779 * The two passes are necessary in order to implement the function
3780 * of cancelling a record written into the log. The first pass
3781 * determines those things which have been cancelled, and the
3782 * second pass replays log items normally except for those which
3783 * have been cancelled. The handling of the replay and cancellations
3784 * takes place in the log item type specific routines.
3785 *
3786 * The table of items which have cancel records in the log is allocated
3787 * and freed at this level, since only here do we know when all of
3788 * the log recovery has been completed.
3789 */
3790STATIC int
3791xlog_do_log_recovery(
3792 xlog_t *log,
3793 xfs_daddr_t head_blk,
3794 xfs_daddr_t tail_blk)
3795{
3796 int error;
3797
3798 ASSERT(head_blk != tail_blk);
3799
3800 /*
3801 * First do a pass to find all of the cancelled buf log items.
3802 * Store them in the buf_cancel_table for use in the second pass.
3803 */
3804 log->l_buf_cancel_table =
3805 (xfs_buf_cancel_t **)kmem_zalloc(XLOG_BC_TABLE_SIZE *
3806 sizeof(xfs_buf_cancel_t*),
3807 KM_SLEEP);
3808 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3809 XLOG_RECOVER_PASS1);
3810 if (error != 0) {
3811 kmem_free(log->l_buf_cancel_table,
3812 XLOG_BC_TABLE_SIZE * sizeof(xfs_buf_cancel_t*));
3813 log->l_buf_cancel_table = NULL;
3814 return error;
3815 }
3816 /*
3817 * Then do a second pass to actually recover the items in the log.
3818 * When it is complete free the table of buf cancel items.
3819 */
3820 error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3821 XLOG_RECOVER_PASS2);
3822#ifdef DEBUG
3823 {
3824 int i;
3825
3826 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3827 ASSERT(log->l_buf_cancel_table[i] == NULL);
3828 }
3829#endif /* DEBUG */
3830
3831 kmem_free(log->l_buf_cancel_table,
3832 XLOG_BC_TABLE_SIZE * sizeof(xfs_buf_cancel_t*));
3833 log->l_buf_cancel_table = NULL;
3834
3835 return error;
3836}
3837
3838/*
3839 * Do the actual recovery
3840 */
3841STATIC int
3842xlog_do_recover(
3843 xlog_t *log,
3844 xfs_daddr_t head_blk,
3845 xfs_daddr_t tail_blk)
3846{
3847 int error;
3848 xfs_buf_t *bp;
3849 xfs_sb_t *sbp;
3850
3851 /*
3852 * First replay the images in the log.
3853 */
3854 error = xlog_do_log_recovery(log, head_blk, tail_blk);
3855 if (error) {
3856 return error;
3857 }
3858
3859 XFS_bflush(log->l_mp->m_ddev_targp);
3860
3861 /*
3862 * If IO errors happened during recovery, bail out.
3863 */
3864 if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3865 return (EIO);
3866 }
3867
3868 /*
3869 * We now update the tail_lsn since much of the recovery has completed
3870 * and there may be space available to use. If there were no extent
3871 * or iunlinks, we can free up the entire log and set the tail_lsn to
3872 * be the last_sync_lsn. This was set in xlog_find_tail to be the
3873 * lsn of the last known good LR on disk. If there are extent frees
3874 * or iunlinks they will have some entries in the AIL; so we look at
3875 * the AIL to determine how to set the tail_lsn.
3876 */
3877 xlog_assign_tail_lsn(log->l_mp);
3878
3879 /*
3880 * Now that we've finished replaying all buffer and inode
3881 * updates, re-read in the superblock.
3882 */
3883 bp = xfs_getsb(log->l_mp, 0);
3884 XFS_BUF_UNDONE(bp);
3885 XFS_BUF_READ(bp);
3886 xfsbdstrat(log->l_mp, bp);
3887 if ((error = xfs_iowait(bp))) {
3888 xfs_ioerror_alert("xlog_do_recover",
3889 log->l_mp, bp, XFS_BUF_ADDR(bp));
3890 ASSERT(0);
3891 xfs_buf_relse(bp);
3892 return error;
3893 }
3894
3895 /* Convert superblock from on-disk format */
3896 sbp = &log->l_mp->m_sb;
3897 xfs_xlatesb(XFS_BUF_TO_SBP(bp), sbp, 1, XFS_SB_ALL_BITS);
3898 ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3899 ASSERT(XFS_SB_GOOD_VERSION(sbp));
3900 xfs_buf_relse(bp);
3901
3902 xlog_recover_check_summary(log);
3903
3904 /* Normal transactions can now occur */
3905 log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3906 return 0;
3907}
3908
3909/*
3910 * Perform recovery and re-initialize some log variables in xlog_find_tail.
3911 *
3912 * Return error or zero.
3913 */
3914int
3915xlog_recover(
3916 xlog_t *log,
3917 int readonly)
3918{
3919 xfs_daddr_t head_blk, tail_blk;
3920 int error;
3921
3922 /* find the tail of the log */
3923 if ((error = xlog_find_tail(log, &head_blk, &tail_blk, readonly)))
3924 return error;
3925
3926 if (tail_blk != head_blk) {
3927 /* There used to be a comment here:
3928 *
3929 * disallow recovery on read-only mounts. note -- mount
3930 * checks for ENOSPC and turns it into an intelligent
3931 * error message.
3932 * ...but this is no longer true. Now, unless you specify
3933 * NORECOVERY (in which case this function would never be
3934 * called), we just go ahead and recover. We do this all
3935 * under the vfs layer, so we can get away with it unless
3936 * the device itself is read-only, in which case we fail.
3937 */
3938 if ((error = xfs_dev_is_read_only(log->l_mp,
3939 "recovery required"))) {
3940 return error;
3941 }
3942
3943 cmn_err(CE_NOTE,
3944 "Starting XFS recovery on filesystem: %s (dev: %s)",
3945 log->l_mp->m_fsname, XFS_BUFTARG_NAME(log->l_targ));
3946
3947 error = xlog_do_recover(log, head_blk, tail_blk);
3948 log->l_flags |= XLOG_RECOVERY_NEEDED;
3949 }
3950 return error;
3951}
3952
3953/*
3954 * In the first part of recovery we replay inodes and buffers and build
3955 * up the list of extent free items which need to be processed. Here
3956 * we process the extent free items and clean up the on disk unlinked
3957 * inode lists. This is separated from the first part of recovery so
3958 * that the root and real-time bitmap inodes can be read in from disk in
3959 * between the two stages. This is necessary so that we can free space
3960 * in the real-time portion of the file system.
3961 */
3962int
3963xlog_recover_finish(
3964 xlog_t *log,
3965 int mfsi_flags)
3966{
3967 /*
3968 * Now we're ready to do the transactions needed for the
3969 * rest of recovery. Start with completing all the extent
3970 * free intent records and then process the unlinked inode
3971 * lists. At this point, we essentially run in normal mode
3972 * except that we're still performing recovery actions
3973 * rather than accepting new requests.
3974 */
3975 if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3976 xlog_recover_process_efis(log);
3977 /*
3978 * Sync the log to get all the EFIs out of the AIL.
3979 * This isn't absolutely necessary, but it helps in
3980 * case the unlink transactions would have problems
3981 * pushing the EFIs out of the way.
3982 */
3983 xfs_log_force(log->l_mp, (xfs_lsn_t)0,
3984 (XFS_LOG_FORCE | XFS_LOG_SYNC));
3985
3986 if ( (mfsi_flags & XFS_MFSI_NOUNLINK) == 0 ) {
3987 xlog_recover_process_iunlinks(log);
3988 }
3989
3990 xlog_recover_check_summary(log);
3991
3992 cmn_err(CE_NOTE,
3993 "Ending XFS recovery on filesystem: %s (dev: %s)",
3994 log->l_mp->m_fsname, XFS_BUFTARG_NAME(log->l_targ));
3995 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3996 } else {
3997 cmn_err(CE_DEBUG,
3998 "!Ending clean XFS mount for filesystem: %s",
3999 log->l_mp->m_fsname);
4000 }
4001 return 0;
4002}
4003
4004
4005#if defined(DEBUG)
4006/*
4007 * Read all of the agf and agi counters and check that they
4008 * are consistent with the superblock counters.
4009 */
4010void
4011xlog_recover_check_summary(
4012 xlog_t *log)
4013{
4014 xfs_mount_t *mp;
4015 xfs_agf_t *agfp;
4016 xfs_agi_t *agip;
4017 xfs_buf_t *agfbp;
4018 xfs_buf_t *agibp;
4019 xfs_daddr_t agfdaddr;
4020 xfs_daddr_t agidaddr;
4021 xfs_buf_t *sbbp;
4022#ifdef XFS_LOUD_RECOVERY
4023 xfs_sb_t *sbp;
4024#endif
4025 xfs_agnumber_t agno;
4026 __uint64_t freeblks;
4027 __uint64_t itotal;
4028 __uint64_t ifree;
4029
4030 mp = log->l_mp;
4031
4032 freeblks = 0LL;
4033 itotal = 0LL;
4034 ifree = 0LL;
4035 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4036 agfdaddr = XFS_AG_DADDR(mp, agno, XFS_AGF_DADDR(mp));
4037 agfbp = xfs_buf_read(mp->m_ddev_targp, agfdaddr,
4038 XFS_FSS_TO_BB(mp, 1), 0);
4039 if (XFS_BUF_ISERROR(agfbp)) {
4040 xfs_ioerror_alert("xlog_recover_check_summary(agf)",
4041 mp, agfbp, agfdaddr);
4042 }
4043 agfp = XFS_BUF_TO_AGF(agfbp);
4044 ASSERT(XFS_AGF_MAGIC ==
4045 INT_GET(agfp->agf_magicnum, ARCH_CONVERT));
4046 ASSERT(XFS_AGF_GOOD_VERSION(
4047 INT_GET(agfp->agf_versionnum, ARCH_CONVERT)));
4048 ASSERT(INT_GET(agfp->agf_seqno, ARCH_CONVERT) == agno);
4049
4050 freeblks += INT_GET(agfp->agf_freeblks, ARCH_CONVERT) +
4051 INT_GET(agfp->agf_flcount, ARCH_CONVERT);
4052 xfs_buf_relse(agfbp);
4053
4054 agidaddr = XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp));
4055 agibp = xfs_buf_read(mp->m_ddev_targp, agidaddr,
4056 XFS_FSS_TO_BB(mp, 1), 0);
4057 if (XFS_BUF_ISERROR(agibp)) {
4058 xfs_ioerror_alert("xlog_recover_check_summary(agi)",
4059 mp, agibp, agidaddr);
4060 }
4061 agip = XFS_BUF_TO_AGI(agibp);
4062 ASSERT(XFS_AGI_MAGIC ==
4063 INT_GET(agip->agi_magicnum, ARCH_CONVERT));
4064 ASSERT(XFS_AGI_GOOD_VERSION(
4065 INT_GET(agip->agi_versionnum, ARCH_CONVERT)));
4066 ASSERT(INT_GET(agip->agi_seqno, ARCH_CONVERT) == agno);
4067
4068 itotal += INT_GET(agip->agi_count, ARCH_CONVERT);
4069 ifree += INT_GET(agip->agi_freecount, ARCH_CONVERT);
4070 xfs_buf_relse(agibp);
4071 }
4072
4073 sbbp = xfs_getsb(mp, 0);
4074#ifdef XFS_LOUD_RECOVERY
4075 sbp = &mp->m_sb;
4076 xfs_xlatesb(XFS_BUF_TO_SBP(sbbp), sbp, 1, XFS_SB_ALL_BITS);
4077 cmn_err(CE_NOTE,
4078 "xlog_recover_check_summary: sb_icount %Lu itotal %Lu",
4079 sbp->sb_icount, itotal);
4080 cmn_err(CE_NOTE,
4081 "xlog_recover_check_summary: sb_ifree %Lu itotal %Lu",
4082 sbp->sb_ifree, ifree);
4083 cmn_err(CE_NOTE,
4084 "xlog_recover_check_summary: sb_fdblocks %Lu freeblks %Lu",
4085 sbp->sb_fdblocks, freeblks);
4086#if 0
4087 /*
4088 * This is turned off until I account for the allocation
4089 * btree blocks which live in free space.
4090 */
4091 ASSERT(sbp->sb_icount == itotal);
4092 ASSERT(sbp->sb_ifree == ifree);
4093 ASSERT(sbp->sb_fdblocks == freeblks);
4094#endif
4095#endif
4096 xfs_buf_relse(sbbp);
4097}
4098#endif /* DEBUG */