Merge git://git.kernel.org/pub/scm/linux/kernel/git/agk/linux-2.6-dm
[sfrench/cifs-2.6.git] / fs / xfs / xfs_log_recover.c
1 /*
2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_dir2_sf.h"
35 #include "xfs_attr_sf.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
38 #include "xfs_inode_item.h"
39 #include "xfs_alloc.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_log_priv.h"
42 #include "xfs_buf_item.h"
43 #include "xfs_log_recover.h"
44 #include "xfs_extfree_item.h"
45 #include "xfs_trans_priv.h"
46 #include "xfs_quota.h"
47 #include "xfs_rw.h"
48 #include "xfs_utils.h"
49
50 STATIC int      xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
51 STATIC int      xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
52 STATIC void     xlog_recover_insert_item_backq(xlog_recover_item_t **q,
53                                                xlog_recover_item_t *item);
54 #if defined(DEBUG)
55 STATIC void     xlog_recover_check_summary(xlog_t *);
56 #else
57 #define xlog_recover_check_summary(log)
58 #endif
59
60
61 /*
62  * Sector aligned buffer routines for buffer create/read/write/access
63  */
64
65 #define XLOG_SECTOR_ROUNDUP_BBCOUNT(log, bbs)   \
66         ( ((log)->l_sectbb_mask && (bbs & (log)->l_sectbb_mask)) ? \
67         ((bbs + (log)->l_sectbb_mask + 1) & ~(log)->l_sectbb_mask) : (bbs) )
68 #define XLOG_SECTOR_ROUNDDOWN_BLKNO(log, bno)   ((bno) & ~(log)->l_sectbb_mask)
69
70 xfs_buf_t *
71 xlog_get_bp(
72         xlog_t          *log,
73         int             nbblks)
74 {
75         if (nbblks <= 0 || nbblks > log->l_logBBsize) {
76                 xlog_warn("XFS: Invalid block length (0x%x) given for buffer", nbblks);
77                 XFS_ERROR_REPORT("xlog_get_bp(1)",
78                                  XFS_ERRLEVEL_HIGH, log->l_mp);
79                 return NULL;
80         }
81
82         if (log->l_sectbb_log) {
83                 if (nbblks > 1)
84                         nbblks += XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
85                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
86         }
87         return xfs_buf_get_noaddr(BBTOB(nbblks), log->l_mp->m_logdev_targp);
88 }
89
90 void
91 xlog_put_bp(
92         xfs_buf_t       *bp)
93 {
94         xfs_buf_free(bp);
95 }
96
97 STATIC xfs_caddr_t
98 xlog_align(
99         xlog_t          *log,
100         xfs_daddr_t     blk_no,
101         int             nbblks,
102         xfs_buf_t       *bp)
103 {
104         xfs_caddr_t     ptr;
105
106         if (!log->l_sectbb_log)
107                 return XFS_BUF_PTR(bp);
108
109         ptr = XFS_BUF_PTR(bp) + BBTOB((int)blk_no & log->l_sectbb_mask);
110         ASSERT(XFS_BUF_SIZE(bp) >=
111                 BBTOB(nbblks + (blk_no & log->l_sectbb_mask)));
112         return ptr;
113 }
114
115
116 /*
117  * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
118  */
119 STATIC int
120 xlog_bread_noalign(
121         xlog_t          *log,
122         xfs_daddr_t     blk_no,
123         int             nbblks,
124         xfs_buf_t       *bp)
125 {
126         int             error;
127
128         if (nbblks <= 0 || nbblks > log->l_logBBsize) {
129                 xlog_warn("XFS: Invalid block length (0x%x) given for buffer", nbblks);
130                 XFS_ERROR_REPORT("xlog_bread(1)",
131                                  XFS_ERRLEVEL_HIGH, log->l_mp);
132                 return EFSCORRUPTED;
133         }
134
135         if (log->l_sectbb_log) {
136                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
137                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
138         }
139
140         ASSERT(nbblks > 0);
141         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
142         ASSERT(bp);
143
144         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
145         XFS_BUF_READ(bp);
146         XFS_BUF_BUSY(bp);
147         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
148         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
149
150         xfsbdstrat(log->l_mp, bp);
151         error = xfs_iowait(bp);
152         if (error)
153                 xfs_ioerror_alert("xlog_bread", log->l_mp,
154                                   bp, XFS_BUF_ADDR(bp));
155         return error;
156 }
157
158 STATIC int
159 xlog_bread(
160         xlog_t          *log,
161         xfs_daddr_t     blk_no,
162         int             nbblks,
163         xfs_buf_t       *bp,
164         xfs_caddr_t     *offset)
165 {
166         int             error;
167
168         error = xlog_bread_noalign(log, blk_no, nbblks, bp);
169         if (error)
170                 return error;
171
172         *offset = xlog_align(log, blk_no, nbblks, bp);
173         return 0;
174 }
175
176 /*
177  * Write out the buffer at the given block for the given number of blocks.
178  * The buffer is kept locked across the write and is returned locked.
179  * This can only be used for synchronous log writes.
180  */
181 STATIC int
182 xlog_bwrite(
183         xlog_t          *log,
184         xfs_daddr_t     blk_no,
185         int             nbblks,
186         xfs_buf_t       *bp)
187 {
188         int             error;
189
190         if (nbblks <= 0 || nbblks > log->l_logBBsize) {
191                 xlog_warn("XFS: Invalid block length (0x%x) given for buffer", nbblks);
192                 XFS_ERROR_REPORT("xlog_bwrite(1)",
193                                  XFS_ERRLEVEL_HIGH, log->l_mp);
194                 return EFSCORRUPTED;
195         }
196
197         if (log->l_sectbb_log) {
198                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
199                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
200         }
201
202         ASSERT(nbblks > 0);
203         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
204
205         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
206         XFS_BUF_ZEROFLAGS(bp);
207         XFS_BUF_BUSY(bp);
208         XFS_BUF_HOLD(bp);
209         XFS_BUF_PSEMA(bp, PRIBIO);
210         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
211         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
212
213         if ((error = xfs_bwrite(log->l_mp, bp)))
214                 xfs_ioerror_alert("xlog_bwrite", log->l_mp,
215                                   bp, XFS_BUF_ADDR(bp));
216         return error;
217 }
218
219 #ifdef DEBUG
220 /*
221  * dump debug superblock and log record information
222  */
223 STATIC void
224 xlog_header_check_dump(
225         xfs_mount_t             *mp,
226         xlog_rec_header_t       *head)
227 {
228         cmn_err(CE_DEBUG, "%s:  SB : uuid = %pU, fmt = %d\n",
229                 __func__, &mp->m_sb.sb_uuid, XLOG_FMT);
230         cmn_err(CE_DEBUG, "    log : uuid = %pU, fmt = %d\n",
231                 &head->h_fs_uuid, be32_to_cpu(head->h_fmt));
232 }
233 #else
234 #define xlog_header_check_dump(mp, head)
235 #endif
236
237 /*
238  * check log record header for recovery
239  */
240 STATIC int
241 xlog_header_check_recover(
242         xfs_mount_t             *mp,
243         xlog_rec_header_t       *head)
244 {
245         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
246
247         /*
248          * IRIX doesn't write the h_fmt field and leaves it zeroed
249          * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
250          * a dirty log created in IRIX.
251          */
252         if (unlikely(be32_to_cpu(head->h_fmt) != XLOG_FMT)) {
253                 xlog_warn(
254         "XFS: dirty log written in incompatible format - can't recover");
255                 xlog_header_check_dump(mp, head);
256                 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
257                                  XFS_ERRLEVEL_HIGH, mp);
258                 return XFS_ERROR(EFSCORRUPTED);
259         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
260                 xlog_warn(
261         "XFS: dirty log entry has mismatched uuid - can't recover");
262                 xlog_header_check_dump(mp, head);
263                 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
264                                  XFS_ERRLEVEL_HIGH, mp);
265                 return XFS_ERROR(EFSCORRUPTED);
266         }
267         return 0;
268 }
269
270 /*
271  * read the head block of the log and check the header
272  */
273 STATIC int
274 xlog_header_check_mount(
275         xfs_mount_t             *mp,
276         xlog_rec_header_t       *head)
277 {
278         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
279
280         if (uuid_is_nil(&head->h_fs_uuid)) {
281                 /*
282                  * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
283                  * h_fs_uuid is nil, we assume this log was last mounted
284                  * by IRIX and continue.
285                  */
286                 xlog_warn("XFS: nil uuid in log - IRIX style log");
287         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
288                 xlog_warn("XFS: log has mismatched uuid - can't recover");
289                 xlog_header_check_dump(mp, head);
290                 XFS_ERROR_REPORT("xlog_header_check_mount",
291                                  XFS_ERRLEVEL_HIGH, mp);
292                 return XFS_ERROR(EFSCORRUPTED);
293         }
294         return 0;
295 }
296
297 STATIC void
298 xlog_recover_iodone(
299         struct xfs_buf  *bp)
300 {
301         if (XFS_BUF_GETERROR(bp)) {
302                 /*
303                  * We're not going to bother about retrying
304                  * this during recovery. One strike!
305                  */
306                 xfs_ioerror_alert("xlog_recover_iodone",
307                                   bp->b_mount, bp, XFS_BUF_ADDR(bp));
308                 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
309         }
310         bp->b_mount = NULL;
311         XFS_BUF_CLR_IODONE_FUNC(bp);
312         xfs_biodone(bp);
313 }
314
315 /*
316  * This routine finds (to an approximation) the first block in the physical
317  * log which contains the given cycle.  It uses a binary search algorithm.
318  * Note that the algorithm can not be perfect because the disk will not
319  * necessarily be perfect.
320  */
321 STATIC int
322 xlog_find_cycle_start(
323         xlog_t          *log,
324         xfs_buf_t       *bp,
325         xfs_daddr_t     first_blk,
326         xfs_daddr_t     *last_blk,
327         uint            cycle)
328 {
329         xfs_caddr_t     offset;
330         xfs_daddr_t     mid_blk;
331         uint            mid_cycle;
332         int             error;
333
334         mid_blk = BLK_AVG(first_blk, *last_blk);
335         while (mid_blk != first_blk && mid_blk != *last_blk) {
336                 error = xlog_bread(log, mid_blk, 1, bp, &offset);
337                 if (error)
338                         return error;
339                 mid_cycle = xlog_get_cycle(offset);
340                 if (mid_cycle == cycle) {
341                         *last_blk = mid_blk;
342                         /* last_half_cycle == mid_cycle */
343                 } else {
344                         first_blk = mid_blk;
345                         /* first_half_cycle == mid_cycle */
346                 }
347                 mid_blk = BLK_AVG(first_blk, *last_blk);
348         }
349         ASSERT((mid_blk == first_blk && mid_blk+1 == *last_blk) ||
350                (mid_blk == *last_blk && mid_blk-1 == first_blk));
351
352         return 0;
353 }
354
355 /*
356  * Check that the range of blocks does not contain the cycle number
357  * given.  The scan needs to occur from front to back and the ptr into the
358  * region must be updated since a later routine will need to perform another
359  * test.  If the region is completely good, we end up returning the same
360  * last block number.
361  *
362  * Set blkno to -1 if we encounter no errors.  This is an invalid block number
363  * since we don't ever expect logs to get this large.
364  */
365 STATIC int
366 xlog_find_verify_cycle(
367         xlog_t          *log,
368         xfs_daddr_t     start_blk,
369         int             nbblks,
370         uint            stop_on_cycle_no,
371         xfs_daddr_t     *new_blk)
372 {
373         xfs_daddr_t     i, j;
374         uint            cycle;
375         xfs_buf_t       *bp;
376         xfs_daddr_t     bufblks;
377         xfs_caddr_t     buf = NULL;
378         int             error = 0;
379
380         bufblks = 1 << ffs(nbblks);
381
382         while (!(bp = xlog_get_bp(log, bufblks))) {
383                 /* can't get enough memory to do everything in one big buffer */
384                 bufblks >>= 1;
385                 if (bufblks <= log->l_sectbb_log)
386                         return ENOMEM;
387         }
388
389         for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
390                 int     bcount;
391
392                 bcount = min(bufblks, (start_blk + nbblks - i));
393
394                 error = xlog_bread(log, i, bcount, bp, &buf);
395                 if (error)
396                         goto out;
397
398                 for (j = 0; j < bcount; j++) {
399                         cycle = xlog_get_cycle(buf);
400                         if (cycle == stop_on_cycle_no) {
401                                 *new_blk = i+j;
402                                 goto out;
403                         }
404
405                         buf += BBSIZE;
406                 }
407         }
408
409         *new_blk = -1;
410
411 out:
412         xlog_put_bp(bp);
413         return error;
414 }
415
416 /*
417  * Potentially backup over partial log record write.
418  *
419  * In the typical case, last_blk is the number of the block directly after
420  * a good log record.  Therefore, we subtract one to get the block number
421  * of the last block in the given buffer.  extra_bblks contains the number
422  * of blocks we would have read on a previous read.  This happens when the
423  * last log record is split over the end of the physical log.
424  *
425  * extra_bblks is the number of blocks potentially verified on a previous
426  * call to this routine.
427  */
428 STATIC int
429 xlog_find_verify_log_record(
430         xlog_t                  *log,
431         xfs_daddr_t             start_blk,
432         xfs_daddr_t             *last_blk,
433         int                     extra_bblks)
434 {
435         xfs_daddr_t             i;
436         xfs_buf_t               *bp;
437         xfs_caddr_t             offset = NULL;
438         xlog_rec_header_t       *head = NULL;
439         int                     error = 0;
440         int                     smallmem = 0;
441         int                     num_blks = *last_blk - start_blk;
442         int                     xhdrs;
443
444         ASSERT(start_blk != 0 || *last_blk != start_blk);
445
446         if (!(bp = xlog_get_bp(log, num_blks))) {
447                 if (!(bp = xlog_get_bp(log, 1)))
448                         return ENOMEM;
449                 smallmem = 1;
450         } else {
451                 error = xlog_bread(log, start_blk, num_blks, bp, &offset);
452                 if (error)
453                         goto out;
454                 offset += ((num_blks - 1) << BBSHIFT);
455         }
456
457         for (i = (*last_blk) - 1; i >= 0; i--) {
458                 if (i < start_blk) {
459                         /* valid log record not found */
460                         xlog_warn(
461                 "XFS: Log inconsistent (didn't find previous header)");
462                         ASSERT(0);
463                         error = XFS_ERROR(EIO);
464                         goto out;
465                 }
466
467                 if (smallmem) {
468                         error = xlog_bread(log, i, 1, bp, &offset);
469                         if (error)
470                                 goto out;
471                 }
472
473                 head = (xlog_rec_header_t *)offset;
474
475                 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(head->h_magicno))
476                         break;
477
478                 if (!smallmem)
479                         offset -= BBSIZE;
480         }
481
482         /*
483          * We hit the beginning of the physical log & still no header.  Return
484          * to caller.  If caller can handle a return of -1, then this routine
485          * will be called again for the end of the physical log.
486          */
487         if (i == -1) {
488                 error = -1;
489                 goto out;
490         }
491
492         /*
493          * We have the final block of the good log (the first block
494          * of the log record _before_ the head. So we check the uuid.
495          */
496         if ((error = xlog_header_check_mount(log->l_mp, head)))
497                 goto out;
498
499         /*
500          * We may have found a log record header before we expected one.
501          * last_blk will be the 1st block # with a given cycle #.  We may end
502          * up reading an entire log record.  In this case, we don't want to
503          * reset last_blk.  Only when last_blk points in the middle of a log
504          * record do we update last_blk.
505          */
506         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
507                 uint    h_size = be32_to_cpu(head->h_size);
508
509                 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
510                 if (h_size % XLOG_HEADER_CYCLE_SIZE)
511                         xhdrs++;
512         } else {
513                 xhdrs = 1;
514         }
515
516         if (*last_blk - i + extra_bblks !=
517             BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
518                 *last_blk = i;
519
520 out:
521         xlog_put_bp(bp);
522         return error;
523 }
524
525 /*
526  * Head is defined to be the point of the log where the next log write
527  * write could go.  This means that incomplete LR writes at the end are
528  * eliminated when calculating the head.  We aren't guaranteed that previous
529  * LR have complete transactions.  We only know that a cycle number of
530  * current cycle number -1 won't be present in the log if we start writing
531  * from our current block number.
532  *
533  * last_blk contains the block number of the first block with a given
534  * cycle number.
535  *
536  * Return: zero if normal, non-zero if error.
537  */
538 STATIC int
539 xlog_find_head(
540         xlog_t          *log,
541         xfs_daddr_t     *return_head_blk)
542 {
543         xfs_buf_t       *bp;
544         xfs_caddr_t     offset;
545         xfs_daddr_t     new_blk, first_blk, start_blk, last_blk, head_blk;
546         int             num_scan_bblks;
547         uint            first_half_cycle, last_half_cycle;
548         uint            stop_on_cycle;
549         int             error, log_bbnum = log->l_logBBsize;
550
551         /* Is the end of the log device zeroed? */
552         if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
553                 *return_head_blk = first_blk;
554
555                 /* Is the whole lot zeroed? */
556                 if (!first_blk) {
557                         /* Linux XFS shouldn't generate totally zeroed logs -
558                          * mkfs etc write a dummy unmount record to a fresh
559                          * log so we can store the uuid in there
560                          */
561                         xlog_warn("XFS: totally zeroed log");
562                 }
563
564                 return 0;
565         } else if (error) {
566                 xlog_warn("XFS: empty log check failed");
567                 return error;
568         }
569
570         first_blk = 0;                  /* get cycle # of 1st block */
571         bp = xlog_get_bp(log, 1);
572         if (!bp)
573                 return ENOMEM;
574
575         error = xlog_bread(log, 0, 1, bp, &offset);
576         if (error)
577                 goto bp_err;
578
579         first_half_cycle = xlog_get_cycle(offset);
580
581         last_blk = head_blk = log_bbnum - 1;    /* get cycle # of last block */
582         error = xlog_bread(log, last_blk, 1, bp, &offset);
583         if (error)
584                 goto bp_err;
585
586         last_half_cycle = xlog_get_cycle(offset);
587         ASSERT(last_half_cycle != 0);
588
589         /*
590          * If the 1st half cycle number is equal to the last half cycle number,
591          * then the entire log is stamped with the same cycle number.  In this
592          * case, head_blk can't be set to zero (which makes sense).  The below
593          * math doesn't work out properly with head_blk equal to zero.  Instead,
594          * we set it to log_bbnum which is an invalid block number, but this
595          * value makes the math correct.  If head_blk doesn't changed through
596          * all the tests below, *head_blk is set to zero at the very end rather
597          * than log_bbnum.  In a sense, log_bbnum and zero are the same block
598          * in a circular file.
599          */
600         if (first_half_cycle == last_half_cycle) {
601                 /*
602                  * In this case we believe that the entire log should have
603                  * cycle number last_half_cycle.  We need to scan backwards
604                  * from the end verifying that there are no holes still
605                  * containing last_half_cycle - 1.  If we find such a hole,
606                  * then the start of that hole will be the new head.  The
607                  * simple case looks like
608                  *        x | x ... | x - 1 | x
609                  * Another case that fits this picture would be
610                  *        x | x + 1 | x ... | x
611                  * In this case the head really is somewhere at the end of the
612                  * log, as one of the latest writes at the beginning was
613                  * incomplete.
614                  * One more case is
615                  *        x | x + 1 | x ... | x - 1 | x
616                  * This is really the combination of the above two cases, and
617                  * the head has to end up at the start of the x-1 hole at the
618                  * end of the log.
619                  *
620                  * In the 256k log case, we will read from the beginning to the
621                  * end of the log and search for cycle numbers equal to x-1.
622                  * We don't worry about the x+1 blocks that we encounter,
623                  * because we know that they cannot be the head since the log
624                  * started with x.
625                  */
626                 head_blk = log_bbnum;
627                 stop_on_cycle = last_half_cycle - 1;
628         } else {
629                 /*
630                  * In this case we want to find the first block with cycle
631                  * number matching last_half_cycle.  We expect the log to be
632                  * some variation on
633                  *        x + 1 ... | x ...
634                  * The first block with cycle number x (last_half_cycle) will
635                  * be where the new head belongs.  First we do a binary search
636                  * for the first occurrence of last_half_cycle.  The binary
637                  * search may not be totally accurate, so then we scan back
638                  * from there looking for occurrences of last_half_cycle before
639                  * us.  If that backwards scan wraps around the beginning of
640                  * the log, then we look for occurrences of last_half_cycle - 1
641                  * at the end of the log.  The cases we're looking for look
642                  * like
643                  *        x + 1 ... | x | x + 1 | x ...
644                  *                               ^ binary search stopped here
645                  * or
646                  *        x + 1 ... | x ... | x - 1 | x
647                  *        <---------> less than scan distance
648                  */
649                 stop_on_cycle = last_half_cycle;
650                 if ((error = xlog_find_cycle_start(log, bp, first_blk,
651                                                 &head_blk, last_half_cycle)))
652                         goto bp_err;
653         }
654
655         /*
656          * Now validate the answer.  Scan back some number of maximum possible
657          * blocks and make sure each one has the expected cycle number.  The
658          * maximum is determined by the total possible amount of buffering
659          * in the in-core log.  The following number can be made tighter if
660          * we actually look at the block size of the filesystem.
661          */
662         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
663         if (head_blk >= num_scan_bblks) {
664                 /*
665                  * We are guaranteed that the entire check can be performed
666                  * in one buffer.
667                  */
668                 start_blk = head_blk - num_scan_bblks;
669                 if ((error = xlog_find_verify_cycle(log,
670                                                 start_blk, num_scan_bblks,
671                                                 stop_on_cycle, &new_blk)))
672                         goto bp_err;
673                 if (new_blk != -1)
674                         head_blk = new_blk;
675         } else {                /* need to read 2 parts of log */
676                 /*
677                  * We are going to scan backwards in the log in two parts.
678                  * First we scan the physical end of the log.  In this part
679                  * of the log, we are looking for blocks with cycle number
680                  * last_half_cycle - 1.
681                  * If we find one, then we know that the log starts there, as
682                  * we've found a hole that didn't get written in going around
683                  * the end of the physical log.  The simple case for this is
684                  *        x + 1 ... | x ... | x - 1 | x
685                  *        <---------> less than scan distance
686                  * If all of the blocks at the end of the log have cycle number
687                  * last_half_cycle, then we check the blocks at the start of
688                  * the log looking for occurrences of last_half_cycle.  If we
689                  * find one, then our current estimate for the location of the
690                  * first occurrence of last_half_cycle is wrong and we move
691                  * back to the hole we've found.  This case looks like
692                  *        x + 1 ... | x | x + 1 | x ...
693                  *                               ^ binary search stopped here
694                  * Another case we need to handle that only occurs in 256k
695                  * logs is
696                  *        x + 1 ... | x ... | x+1 | x ...
697                  *                   ^ binary search stops here
698                  * In a 256k log, the scan at the end of the log will see the
699                  * x + 1 blocks.  We need to skip past those since that is
700                  * certainly not the head of the log.  By searching for
701                  * last_half_cycle-1 we accomplish that.
702                  */
703                 start_blk = log_bbnum - num_scan_bblks + head_blk;
704                 ASSERT(head_blk <= INT_MAX &&
705                         (xfs_daddr_t) num_scan_bblks - head_blk >= 0);
706                 if ((error = xlog_find_verify_cycle(log, start_blk,
707                                         num_scan_bblks - (int)head_blk,
708                                         (stop_on_cycle - 1), &new_blk)))
709                         goto bp_err;
710                 if (new_blk != -1) {
711                         head_blk = new_blk;
712                         goto bad_blk;
713                 }
714
715                 /*
716                  * Scan beginning of log now.  The last part of the physical
717                  * log is good.  This scan needs to verify that it doesn't find
718                  * the last_half_cycle.
719                  */
720                 start_blk = 0;
721                 ASSERT(head_blk <= INT_MAX);
722                 if ((error = xlog_find_verify_cycle(log,
723                                         start_blk, (int)head_blk,
724                                         stop_on_cycle, &new_blk)))
725                         goto bp_err;
726                 if (new_blk != -1)
727                         head_blk = new_blk;
728         }
729
730  bad_blk:
731         /*
732          * Now we need to make sure head_blk is not pointing to a block in
733          * the middle of a log record.
734          */
735         num_scan_bblks = XLOG_REC_SHIFT(log);
736         if (head_blk >= num_scan_bblks) {
737                 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
738
739                 /* start ptr at last block ptr before head_blk */
740                 if ((error = xlog_find_verify_log_record(log, start_blk,
741                                                         &head_blk, 0)) == -1) {
742                         error = XFS_ERROR(EIO);
743                         goto bp_err;
744                 } else if (error)
745                         goto bp_err;
746         } else {
747                 start_blk = 0;
748                 ASSERT(head_blk <= INT_MAX);
749                 if ((error = xlog_find_verify_log_record(log, start_blk,
750                                                         &head_blk, 0)) == -1) {
751                         /* We hit the beginning of the log during our search */
752                         start_blk = log_bbnum - num_scan_bblks + head_blk;
753                         new_blk = log_bbnum;
754                         ASSERT(start_blk <= INT_MAX &&
755                                 (xfs_daddr_t) log_bbnum-start_blk >= 0);
756                         ASSERT(head_blk <= INT_MAX);
757                         if ((error = xlog_find_verify_log_record(log,
758                                                         start_blk, &new_blk,
759                                                         (int)head_blk)) == -1) {
760                                 error = XFS_ERROR(EIO);
761                                 goto bp_err;
762                         } else if (error)
763                                 goto bp_err;
764                         if (new_blk != log_bbnum)
765                                 head_blk = new_blk;
766                 } else if (error)
767                         goto bp_err;
768         }
769
770         xlog_put_bp(bp);
771         if (head_blk == log_bbnum)
772                 *return_head_blk = 0;
773         else
774                 *return_head_blk = head_blk;
775         /*
776          * When returning here, we have a good block number.  Bad block
777          * means that during a previous crash, we didn't have a clean break
778          * from cycle number N to cycle number N-1.  In this case, we need
779          * to find the first block with cycle number N-1.
780          */
781         return 0;
782
783  bp_err:
784         xlog_put_bp(bp);
785
786         if (error)
787             xlog_warn("XFS: failed to find log head");
788         return error;
789 }
790
791 /*
792  * Find the sync block number or the tail of the log.
793  *
794  * This will be the block number of the last record to have its
795  * associated buffers synced to disk.  Every log record header has
796  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
797  * to get a sync block number.  The only concern is to figure out which
798  * log record header to believe.
799  *
800  * The following algorithm uses the log record header with the largest
801  * lsn.  The entire log record does not need to be valid.  We only care
802  * that the header is valid.
803  *
804  * We could speed up search by using current head_blk buffer, but it is not
805  * available.
806  */
807 int
808 xlog_find_tail(
809         xlog_t                  *log,
810         xfs_daddr_t             *head_blk,
811         xfs_daddr_t             *tail_blk)
812 {
813         xlog_rec_header_t       *rhead;
814         xlog_op_header_t        *op_head;
815         xfs_caddr_t             offset = NULL;
816         xfs_buf_t               *bp;
817         int                     error, i, found;
818         xfs_daddr_t             umount_data_blk;
819         xfs_daddr_t             after_umount_blk;
820         xfs_lsn_t               tail_lsn;
821         int                     hblks;
822
823         found = 0;
824
825         /*
826          * Find previous log record
827          */
828         if ((error = xlog_find_head(log, head_blk)))
829                 return error;
830
831         bp = xlog_get_bp(log, 1);
832         if (!bp)
833                 return ENOMEM;
834         if (*head_blk == 0) {                           /* special case */
835                 error = xlog_bread(log, 0, 1, bp, &offset);
836                 if (error)
837                         goto bread_err;
838
839                 if (xlog_get_cycle(offset) == 0) {
840                         *tail_blk = 0;
841                         /* leave all other log inited values alone */
842                         goto exit;
843                 }
844         }
845
846         /*
847          * Search backwards looking for log record header block
848          */
849         ASSERT(*head_blk < INT_MAX);
850         for (i = (int)(*head_blk) - 1; i >= 0; i--) {
851                 error = xlog_bread(log, i, 1, bp, &offset);
852                 if (error)
853                         goto bread_err;
854
855                 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(*(__be32 *)offset)) {
856                         found = 1;
857                         break;
858                 }
859         }
860         /*
861          * If we haven't found the log record header block, start looking
862          * again from the end of the physical log.  XXXmiken: There should be
863          * a check here to make sure we didn't search more than N blocks in
864          * the previous code.
865          */
866         if (!found) {
867                 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
868                         error = xlog_bread(log, i, 1, bp, &offset);
869                         if (error)
870                                 goto bread_err;
871
872                         if (XLOG_HEADER_MAGIC_NUM ==
873                             be32_to_cpu(*(__be32 *)offset)) {
874                                 found = 2;
875                                 break;
876                         }
877                 }
878         }
879         if (!found) {
880                 xlog_warn("XFS: xlog_find_tail: couldn't find sync record");
881                 ASSERT(0);
882                 return XFS_ERROR(EIO);
883         }
884
885         /* find blk_no of tail of log */
886         rhead = (xlog_rec_header_t *)offset;
887         *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
888
889         /*
890          * Reset log values according to the state of the log when we
891          * crashed.  In the case where head_blk == 0, we bump curr_cycle
892          * one because the next write starts a new cycle rather than
893          * continuing the cycle of the last good log record.  At this
894          * point we have guaranteed that all partial log records have been
895          * accounted for.  Therefore, we know that the last good log record
896          * written was complete and ended exactly on the end boundary
897          * of the physical log.
898          */
899         log->l_prev_block = i;
900         log->l_curr_block = (int)*head_blk;
901         log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
902         if (found == 2)
903                 log->l_curr_cycle++;
904         log->l_tail_lsn = be64_to_cpu(rhead->h_tail_lsn);
905         log->l_last_sync_lsn = be64_to_cpu(rhead->h_lsn);
906         log->l_grant_reserve_cycle = log->l_curr_cycle;
907         log->l_grant_reserve_bytes = BBTOB(log->l_curr_block);
908         log->l_grant_write_cycle = log->l_curr_cycle;
909         log->l_grant_write_bytes = BBTOB(log->l_curr_block);
910
911         /*
912          * Look for unmount record.  If we find it, then we know there
913          * was a clean unmount.  Since 'i' could be the last block in
914          * the physical log, we convert to a log block before comparing
915          * to the head_blk.
916          *
917          * Save the current tail lsn to use to pass to
918          * xlog_clear_stale_blocks() below.  We won't want to clear the
919          * unmount record if there is one, so we pass the lsn of the
920          * unmount record rather than the block after it.
921          */
922         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
923                 int     h_size = be32_to_cpu(rhead->h_size);
924                 int     h_version = be32_to_cpu(rhead->h_version);
925
926                 if ((h_version & XLOG_VERSION_2) &&
927                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
928                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
929                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
930                                 hblks++;
931                 } else {
932                         hblks = 1;
933                 }
934         } else {
935                 hblks = 1;
936         }
937         after_umount_blk = (i + hblks + (int)
938                 BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
939         tail_lsn = log->l_tail_lsn;
940         if (*head_blk == after_umount_blk &&
941             be32_to_cpu(rhead->h_num_logops) == 1) {
942                 umount_data_blk = (i + hblks) % log->l_logBBsize;
943                 error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
944                 if (error)
945                         goto bread_err;
946
947                 op_head = (xlog_op_header_t *)offset;
948                 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
949                         /*
950                          * Set tail and last sync so that newly written
951                          * log records will point recovery to after the
952                          * current unmount record.
953                          */
954                         log->l_tail_lsn =
955                                 xlog_assign_lsn(log->l_curr_cycle,
956                                                 after_umount_blk);
957                         log->l_last_sync_lsn =
958                                 xlog_assign_lsn(log->l_curr_cycle,
959                                                 after_umount_blk);
960                         *tail_blk = after_umount_blk;
961
962                         /*
963                          * Note that the unmount was clean. If the unmount
964                          * was not clean, we need to know this to rebuild the
965                          * superblock counters from the perag headers if we
966                          * have a filesystem using non-persistent counters.
967                          */
968                         log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
969                 }
970         }
971
972         /*
973          * Make sure that there are no blocks in front of the head
974          * with the same cycle number as the head.  This can happen
975          * because we allow multiple outstanding log writes concurrently,
976          * and the later writes might make it out before earlier ones.
977          *
978          * We use the lsn from before modifying it so that we'll never
979          * overwrite the unmount record after a clean unmount.
980          *
981          * Do this only if we are going to recover the filesystem
982          *
983          * NOTE: This used to say "if (!readonly)"
984          * However on Linux, we can & do recover a read-only filesystem.
985          * We only skip recovery if NORECOVERY is specified on mount,
986          * in which case we would not be here.
987          *
988          * But... if the -device- itself is readonly, just skip this.
989          * We can't recover this device anyway, so it won't matter.
990          */
991         if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
992                 error = xlog_clear_stale_blocks(log, tail_lsn);
993         }
994
995 bread_err:
996 exit:
997         xlog_put_bp(bp);
998
999         if (error)
1000                 xlog_warn("XFS: failed to locate log tail");
1001         return error;
1002 }
1003
1004 /*
1005  * Is the log zeroed at all?
1006  *
1007  * The last binary search should be changed to perform an X block read
1008  * once X becomes small enough.  You can then search linearly through
1009  * the X blocks.  This will cut down on the number of reads we need to do.
1010  *
1011  * If the log is partially zeroed, this routine will pass back the blkno
1012  * of the first block with cycle number 0.  It won't have a complete LR
1013  * preceding it.
1014  *
1015  * Return:
1016  *      0  => the log is completely written to
1017  *      -1 => use *blk_no as the first block of the log
1018  *      >0 => error has occurred
1019  */
1020 STATIC int
1021 xlog_find_zeroed(
1022         xlog_t          *log,
1023         xfs_daddr_t     *blk_no)
1024 {
1025         xfs_buf_t       *bp;
1026         xfs_caddr_t     offset;
1027         uint            first_cycle, last_cycle;
1028         xfs_daddr_t     new_blk, last_blk, start_blk;
1029         xfs_daddr_t     num_scan_bblks;
1030         int             error, log_bbnum = log->l_logBBsize;
1031
1032         *blk_no = 0;
1033
1034         /* check totally zeroed log */
1035         bp = xlog_get_bp(log, 1);
1036         if (!bp)
1037                 return ENOMEM;
1038         error = xlog_bread(log, 0, 1, bp, &offset);
1039         if (error)
1040                 goto bp_err;
1041
1042         first_cycle = xlog_get_cycle(offset);
1043         if (first_cycle == 0) {         /* completely zeroed log */
1044                 *blk_no = 0;
1045                 xlog_put_bp(bp);
1046                 return -1;
1047         }
1048
1049         /* check partially zeroed log */
1050         error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1051         if (error)
1052                 goto bp_err;
1053
1054         last_cycle = xlog_get_cycle(offset);
1055         if (last_cycle != 0) {          /* log completely written to */
1056                 xlog_put_bp(bp);
1057                 return 0;
1058         } else if (first_cycle != 1) {
1059                 /*
1060                  * If the cycle of the last block is zero, the cycle of
1061                  * the first block must be 1. If it's not, maybe we're
1062                  * not looking at a log... Bail out.
1063                  */
1064                 xlog_warn("XFS: Log inconsistent or not a log (last==0, first!=1)");
1065                 return XFS_ERROR(EINVAL);
1066         }
1067
1068         /* we have a partially zeroed log */
1069         last_blk = log_bbnum-1;
1070         if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1071                 goto bp_err;
1072
1073         /*
1074          * Validate the answer.  Because there is no way to guarantee that
1075          * the entire log is made up of log records which are the same size,
1076          * we scan over the defined maximum blocks.  At this point, the maximum
1077          * is not chosen to mean anything special.   XXXmiken
1078          */
1079         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1080         ASSERT(num_scan_bblks <= INT_MAX);
1081
1082         if (last_blk < num_scan_bblks)
1083                 num_scan_bblks = last_blk;
1084         start_blk = last_blk - num_scan_bblks;
1085
1086         /*
1087          * We search for any instances of cycle number 0 that occur before
1088          * our current estimate of the head.  What we're trying to detect is
1089          *        1 ... | 0 | 1 | 0...
1090          *                       ^ binary search ends here
1091          */
1092         if ((error = xlog_find_verify_cycle(log, start_blk,
1093                                          (int)num_scan_bblks, 0, &new_blk)))
1094                 goto bp_err;
1095         if (new_blk != -1)
1096                 last_blk = new_blk;
1097
1098         /*
1099          * Potentially backup over partial log record write.  We don't need
1100          * to search the end of the log because we know it is zero.
1101          */
1102         if ((error = xlog_find_verify_log_record(log, start_blk,
1103                                 &last_blk, 0)) == -1) {
1104             error = XFS_ERROR(EIO);
1105             goto bp_err;
1106         } else if (error)
1107             goto bp_err;
1108
1109         *blk_no = last_blk;
1110 bp_err:
1111         xlog_put_bp(bp);
1112         if (error)
1113                 return error;
1114         return -1;
1115 }
1116
1117 /*
1118  * These are simple subroutines used by xlog_clear_stale_blocks() below
1119  * to initialize a buffer full of empty log record headers and write
1120  * them into the log.
1121  */
1122 STATIC void
1123 xlog_add_record(
1124         xlog_t                  *log,
1125         xfs_caddr_t             buf,
1126         int                     cycle,
1127         int                     block,
1128         int                     tail_cycle,
1129         int                     tail_block)
1130 {
1131         xlog_rec_header_t       *recp = (xlog_rec_header_t *)buf;
1132
1133         memset(buf, 0, BBSIZE);
1134         recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1135         recp->h_cycle = cpu_to_be32(cycle);
1136         recp->h_version = cpu_to_be32(
1137                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1138         recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1139         recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1140         recp->h_fmt = cpu_to_be32(XLOG_FMT);
1141         memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1142 }
1143
1144 STATIC int
1145 xlog_write_log_records(
1146         xlog_t          *log,
1147         int             cycle,
1148         int             start_block,
1149         int             blocks,
1150         int             tail_cycle,
1151         int             tail_block)
1152 {
1153         xfs_caddr_t     offset;
1154         xfs_buf_t       *bp;
1155         int             balign, ealign;
1156         int             sectbb = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
1157         int             end_block = start_block + blocks;
1158         int             bufblks;
1159         int             error = 0;
1160         int             i, j = 0;
1161
1162         bufblks = 1 << ffs(blocks);
1163         while (!(bp = xlog_get_bp(log, bufblks))) {
1164                 bufblks >>= 1;
1165                 if (bufblks <= log->l_sectbb_log)
1166                         return ENOMEM;
1167         }
1168
1169         /* We may need to do a read at the start to fill in part of
1170          * the buffer in the starting sector not covered by the first
1171          * write below.
1172          */
1173         balign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, start_block);
1174         if (balign != start_block) {
1175                 error = xlog_bread_noalign(log, start_block, 1, bp);
1176                 if (error)
1177                         goto out_put_bp;
1178
1179                 j = start_block - balign;
1180         }
1181
1182         for (i = start_block; i < end_block; i += bufblks) {
1183                 int             bcount, endcount;
1184
1185                 bcount = min(bufblks, end_block - start_block);
1186                 endcount = bcount - j;
1187
1188                 /* We may need to do a read at the end to fill in part of
1189                  * the buffer in the final sector not covered by the write.
1190                  * If this is the same sector as the above read, skip it.
1191                  */
1192                 ealign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, end_block);
1193                 if (j == 0 && (start_block + endcount > ealign)) {
1194                         offset = XFS_BUF_PTR(bp);
1195                         balign = BBTOB(ealign - start_block);
1196                         error = XFS_BUF_SET_PTR(bp, offset + balign,
1197                                                 BBTOB(sectbb));
1198                         if (error)
1199                                 break;
1200
1201                         error = xlog_bread_noalign(log, ealign, sectbb, bp);
1202                         if (error)
1203                                 break;
1204
1205                         error = XFS_BUF_SET_PTR(bp, offset, bufblks);
1206                         if (error)
1207                                 break;
1208                 }
1209
1210                 offset = xlog_align(log, start_block, endcount, bp);
1211                 for (; j < endcount; j++) {
1212                         xlog_add_record(log, offset, cycle, i+j,
1213                                         tail_cycle, tail_block);
1214                         offset += BBSIZE;
1215                 }
1216                 error = xlog_bwrite(log, start_block, endcount, bp);
1217                 if (error)
1218                         break;
1219                 start_block += endcount;
1220                 j = 0;
1221         }
1222
1223  out_put_bp:
1224         xlog_put_bp(bp);
1225         return error;
1226 }
1227
1228 /*
1229  * This routine is called to blow away any incomplete log writes out
1230  * in front of the log head.  We do this so that we won't become confused
1231  * if we come up, write only a little bit more, and then crash again.
1232  * If we leave the partial log records out there, this situation could
1233  * cause us to think those partial writes are valid blocks since they
1234  * have the current cycle number.  We get rid of them by overwriting them
1235  * with empty log records with the old cycle number rather than the
1236  * current one.
1237  *
1238  * The tail lsn is passed in rather than taken from
1239  * the log so that we will not write over the unmount record after a
1240  * clean unmount in a 512 block log.  Doing so would leave the log without
1241  * any valid log records in it until a new one was written.  If we crashed
1242  * during that time we would not be able to recover.
1243  */
1244 STATIC int
1245 xlog_clear_stale_blocks(
1246         xlog_t          *log,
1247         xfs_lsn_t       tail_lsn)
1248 {
1249         int             tail_cycle, head_cycle;
1250         int             tail_block, head_block;
1251         int             tail_distance, max_distance;
1252         int             distance;
1253         int             error;
1254
1255         tail_cycle = CYCLE_LSN(tail_lsn);
1256         tail_block = BLOCK_LSN(tail_lsn);
1257         head_cycle = log->l_curr_cycle;
1258         head_block = log->l_curr_block;
1259
1260         /*
1261          * Figure out the distance between the new head of the log
1262          * and the tail.  We want to write over any blocks beyond the
1263          * head that we may have written just before the crash, but
1264          * we don't want to overwrite the tail of the log.
1265          */
1266         if (head_cycle == tail_cycle) {
1267                 /*
1268                  * The tail is behind the head in the physical log,
1269                  * so the distance from the head to the tail is the
1270                  * distance from the head to the end of the log plus
1271                  * the distance from the beginning of the log to the
1272                  * tail.
1273                  */
1274                 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1275                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1276                                          XFS_ERRLEVEL_LOW, log->l_mp);
1277                         return XFS_ERROR(EFSCORRUPTED);
1278                 }
1279                 tail_distance = tail_block + (log->l_logBBsize - head_block);
1280         } else {
1281                 /*
1282                  * The head is behind the tail in the physical log,
1283                  * so the distance from the head to the tail is just
1284                  * the tail block minus the head block.
1285                  */
1286                 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1287                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1288                                          XFS_ERRLEVEL_LOW, log->l_mp);
1289                         return XFS_ERROR(EFSCORRUPTED);
1290                 }
1291                 tail_distance = tail_block - head_block;
1292         }
1293
1294         /*
1295          * If the head is right up against the tail, we can't clear
1296          * anything.
1297          */
1298         if (tail_distance <= 0) {
1299                 ASSERT(tail_distance == 0);
1300                 return 0;
1301         }
1302
1303         max_distance = XLOG_TOTAL_REC_SHIFT(log);
1304         /*
1305          * Take the smaller of the maximum amount of outstanding I/O
1306          * we could have and the distance to the tail to clear out.
1307          * We take the smaller so that we don't overwrite the tail and
1308          * we don't waste all day writing from the head to the tail
1309          * for no reason.
1310          */
1311         max_distance = MIN(max_distance, tail_distance);
1312
1313         if ((head_block + max_distance) <= log->l_logBBsize) {
1314                 /*
1315                  * We can stomp all the blocks we need to without
1316                  * wrapping around the end of the log.  Just do it
1317                  * in a single write.  Use the cycle number of the
1318                  * current cycle minus one so that the log will look like:
1319                  *     n ... | n - 1 ...
1320                  */
1321                 error = xlog_write_log_records(log, (head_cycle - 1),
1322                                 head_block, max_distance, tail_cycle,
1323                                 tail_block);
1324                 if (error)
1325                         return error;
1326         } else {
1327                 /*
1328                  * We need to wrap around the end of the physical log in
1329                  * order to clear all the blocks.  Do it in two separate
1330                  * I/Os.  The first write should be from the head to the
1331                  * end of the physical log, and it should use the current
1332                  * cycle number minus one just like above.
1333                  */
1334                 distance = log->l_logBBsize - head_block;
1335                 error = xlog_write_log_records(log, (head_cycle - 1),
1336                                 head_block, distance, tail_cycle,
1337                                 tail_block);
1338
1339                 if (error)
1340                         return error;
1341
1342                 /*
1343                  * Now write the blocks at the start of the physical log.
1344                  * This writes the remainder of the blocks we want to clear.
1345                  * It uses the current cycle number since we're now on the
1346                  * same cycle as the head so that we get:
1347                  *    n ... n ... | n - 1 ...
1348                  *    ^^^^^ blocks we're writing
1349                  */
1350                 distance = max_distance - (log->l_logBBsize - head_block);
1351                 error = xlog_write_log_records(log, head_cycle, 0, distance,
1352                                 tail_cycle, tail_block);
1353                 if (error)
1354                         return error;
1355         }
1356
1357         return 0;
1358 }
1359
1360 /******************************************************************************
1361  *
1362  *              Log recover routines
1363  *
1364  ******************************************************************************
1365  */
1366
1367 STATIC xlog_recover_t *
1368 xlog_recover_find_tid(
1369         xlog_recover_t          *q,
1370         xlog_tid_t              tid)
1371 {
1372         xlog_recover_t          *p = q;
1373
1374         while (p != NULL) {
1375                 if (p->r_log_tid == tid)
1376                     break;
1377                 p = p->r_next;
1378         }
1379         return p;
1380 }
1381
1382 STATIC void
1383 xlog_recover_put_hashq(
1384         xlog_recover_t          **q,
1385         xlog_recover_t          *trans)
1386 {
1387         trans->r_next = *q;
1388         *q = trans;
1389 }
1390
1391 STATIC void
1392 xlog_recover_add_item(
1393         xlog_recover_item_t     **itemq)
1394 {
1395         xlog_recover_item_t     *item;
1396
1397         item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1398         xlog_recover_insert_item_backq(itemq, item);
1399 }
1400
1401 STATIC int
1402 xlog_recover_add_to_cont_trans(
1403         xlog_recover_t          *trans,
1404         xfs_caddr_t             dp,
1405         int                     len)
1406 {
1407         xlog_recover_item_t     *item;
1408         xfs_caddr_t             ptr, old_ptr;
1409         int                     old_len;
1410
1411         item = trans->r_itemq;
1412         if (item == NULL) {
1413                 /* finish copying rest of trans header */
1414                 xlog_recover_add_item(&trans->r_itemq);
1415                 ptr = (xfs_caddr_t) &trans->r_theader +
1416                                 sizeof(xfs_trans_header_t) - len;
1417                 memcpy(ptr, dp, len); /* d, s, l */
1418                 return 0;
1419         }
1420         item = item->ri_prev;
1421
1422         old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1423         old_len = item->ri_buf[item->ri_cnt-1].i_len;
1424
1425         ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1426         memcpy(&ptr[old_len], dp, len); /* d, s, l */
1427         item->ri_buf[item->ri_cnt-1].i_len += len;
1428         item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1429         return 0;
1430 }
1431
1432 /*
1433  * The next region to add is the start of a new region.  It could be
1434  * a whole region or it could be the first part of a new region.  Because
1435  * of this, the assumption here is that the type and size fields of all
1436  * format structures fit into the first 32 bits of the structure.
1437  *
1438  * This works because all regions must be 32 bit aligned.  Therefore, we
1439  * either have both fields or we have neither field.  In the case we have
1440  * neither field, the data part of the region is zero length.  We only have
1441  * a log_op_header and can throw away the header since a new one will appear
1442  * later.  If we have at least 4 bytes, then we can determine how many regions
1443  * will appear in the current log item.
1444  */
1445 STATIC int
1446 xlog_recover_add_to_trans(
1447         xlog_recover_t          *trans,
1448         xfs_caddr_t             dp,
1449         int                     len)
1450 {
1451         xfs_inode_log_format_t  *in_f;                  /* any will do */
1452         xlog_recover_item_t     *item;
1453         xfs_caddr_t             ptr;
1454
1455         if (!len)
1456                 return 0;
1457         item = trans->r_itemq;
1458         if (item == NULL) {
1459                 /* we need to catch log corruptions here */
1460                 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1461                         xlog_warn("XFS: xlog_recover_add_to_trans: "
1462                                   "bad header magic number");
1463                         ASSERT(0);
1464                         return XFS_ERROR(EIO);
1465                 }
1466                 if (len == sizeof(xfs_trans_header_t))
1467                         xlog_recover_add_item(&trans->r_itemq);
1468                 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1469                 return 0;
1470         }
1471
1472         ptr = kmem_alloc(len, KM_SLEEP);
1473         memcpy(ptr, dp, len);
1474         in_f = (xfs_inode_log_format_t *)ptr;
1475
1476         if (item->ri_prev->ri_total != 0 &&
1477              item->ri_prev->ri_total == item->ri_prev->ri_cnt) {
1478                 xlog_recover_add_item(&trans->r_itemq);
1479         }
1480         item = trans->r_itemq;
1481         item = item->ri_prev;
1482
1483         if (item->ri_total == 0) {              /* first region to be added */
1484                 if (in_f->ilf_size == 0 ||
1485                     in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
1486                         xlog_warn(
1487         "XFS: bad number of regions (%d) in inode log format",
1488                                   in_f->ilf_size);
1489                         ASSERT(0);
1490                         return XFS_ERROR(EIO);
1491                 }
1492
1493                 item->ri_total = in_f->ilf_size;
1494                 item->ri_buf =
1495                         kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
1496                                     KM_SLEEP);
1497         }
1498         ASSERT(item->ri_total > item->ri_cnt);
1499         /* Description region is ri_buf[0] */
1500         item->ri_buf[item->ri_cnt].i_addr = ptr;
1501         item->ri_buf[item->ri_cnt].i_len  = len;
1502         item->ri_cnt++;
1503         return 0;
1504 }
1505
1506 STATIC void
1507 xlog_recover_new_tid(
1508         xlog_recover_t          **q,
1509         xlog_tid_t              tid,
1510         xfs_lsn_t               lsn)
1511 {
1512         xlog_recover_t          *trans;
1513
1514         trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1515         trans->r_log_tid   = tid;
1516         trans->r_lsn       = lsn;
1517         xlog_recover_put_hashq(q, trans);
1518 }
1519
1520 STATIC int
1521 xlog_recover_unlink_tid(
1522         xlog_recover_t          **q,
1523         xlog_recover_t          *trans)
1524 {
1525         xlog_recover_t          *tp;
1526         int                     found = 0;
1527
1528         ASSERT(trans != NULL);
1529         if (trans == *q) {
1530                 *q = (*q)->r_next;
1531         } else {
1532                 tp = *q;
1533                 while (tp) {
1534                         if (tp->r_next == trans) {
1535                                 found = 1;
1536                                 break;
1537                         }
1538                         tp = tp->r_next;
1539                 }
1540                 if (!found) {
1541                         xlog_warn(
1542                              "XFS: xlog_recover_unlink_tid: trans not found");
1543                         ASSERT(0);
1544                         return XFS_ERROR(EIO);
1545                 }
1546                 tp->r_next = tp->r_next->r_next;
1547         }
1548         return 0;
1549 }
1550
1551 STATIC void
1552 xlog_recover_insert_item_backq(
1553         xlog_recover_item_t     **q,
1554         xlog_recover_item_t     *item)
1555 {
1556         if (*q == NULL) {
1557                 item->ri_prev = item->ri_next = item;
1558                 *q = item;
1559         } else {
1560                 item->ri_next           = *q;
1561                 item->ri_prev           = (*q)->ri_prev;
1562                 (*q)->ri_prev           = item;
1563                 item->ri_prev->ri_next  = item;
1564         }
1565 }
1566
1567 STATIC void
1568 xlog_recover_insert_item_frontq(
1569         xlog_recover_item_t     **q,
1570         xlog_recover_item_t     *item)
1571 {
1572         xlog_recover_insert_item_backq(q, item);
1573         *q = item;
1574 }
1575
1576 STATIC int
1577 xlog_recover_reorder_trans(
1578         xlog_recover_t          *trans)
1579 {
1580         xlog_recover_item_t     *first_item, *itemq, *itemq_next;
1581         xfs_buf_log_format_t    *buf_f;
1582         ushort                  flags = 0;
1583
1584         first_item = itemq = trans->r_itemq;
1585         trans->r_itemq = NULL;
1586         do {
1587                 itemq_next = itemq->ri_next;
1588                 buf_f = (xfs_buf_log_format_t *)itemq->ri_buf[0].i_addr;
1589
1590                 switch (ITEM_TYPE(itemq)) {
1591                 case XFS_LI_BUF:
1592                         flags = buf_f->blf_flags;
1593                         if (!(flags & XFS_BLI_CANCEL)) {
1594                                 xlog_recover_insert_item_frontq(&trans->r_itemq,
1595                                                                 itemq);
1596                                 break;
1597                         }
1598                 case XFS_LI_INODE:
1599                 case XFS_LI_DQUOT:
1600                 case XFS_LI_QUOTAOFF:
1601                 case XFS_LI_EFD:
1602                 case XFS_LI_EFI:
1603                         xlog_recover_insert_item_backq(&trans->r_itemq, itemq);
1604                         break;
1605                 default:
1606                         xlog_warn(
1607         "XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1608                         ASSERT(0);
1609                         return XFS_ERROR(EIO);
1610                 }
1611                 itemq = itemq_next;
1612         } while (first_item != itemq);
1613         return 0;
1614 }
1615
1616 /*
1617  * Build up the table of buf cancel records so that we don't replay
1618  * cancelled data in the second pass.  For buffer records that are
1619  * not cancel records, there is nothing to do here so we just return.
1620  *
1621  * If we get a cancel record which is already in the table, this indicates
1622  * that the buffer was cancelled multiple times.  In order to ensure
1623  * that during pass 2 we keep the record in the table until we reach its
1624  * last occurrence in the log, we keep a reference count in the cancel
1625  * record in the table to tell us how many times we expect to see this
1626  * record during the second pass.
1627  */
1628 STATIC void
1629 xlog_recover_do_buffer_pass1(
1630         xlog_t                  *log,
1631         xfs_buf_log_format_t    *buf_f)
1632 {
1633         xfs_buf_cancel_t        *bcp;
1634         xfs_buf_cancel_t        *nextp;
1635         xfs_buf_cancel_t        *prevp;
1636         xfs_buf_cancel_t        **bucket;
1637         xfs_daddr_t             blkno = 0;
1638         uint                    len = 0;
1639         ushort                  flags = 0;
1640
1641         switch (buf_f->blf_type) {
1642         case XFS_LI_BUF:
1643                 blkno = buf_f->blf_blkno;
1644                 len = buf_f->blf_len;
1645                 flags = buf_f->blf_flags;
1646                 break;
1647         }
1648
1649         /*
1650          * If this isn't a cancel buffer item, then just return.
1651          */
1652         if (!(flags & XFS_BLI_CANCEL))
1653                 return;
1654
1655         /*
1656          * Insert an xfs_buf_cancel record into the hash table of
1657          * them.  If there is already an identical record, bump
1658          * its reference count.
1659          */
1660         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1661                                           XLOG_BC_TABLE_SIZE];
1662         /*
1663          * If the hash bucket is empty then just insert a new record into
1664          * the bucket.
1665          */
1666         if (*bucket == NULL) {
1667                 bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1668                                                      KM_SLEEP);
1669                 bcp->bc_blkno = blkno;
1670                 bcp->bc_len = len;
1671                 bcp->bc_refcount = 1;
1672                 bcp->bc_next = NULL;
1673                 *bucket = bcp;
1674                 return;
1675         }
1676
1677         /*
1678          * The hash bucket is not empty, so search for duplicates of our
1679          * record.  If we find one them just bump its refcount.  If not
1680          * then add us at the end of the list.
1681          */
1682         prevp = NULL;
1683         nextp = *bucket;
1684         while (nextp != NULL) {
1685                 if (nextp->bc_blkno == blkno && nextp->bc_len == len) {
1686                         nextp->bc_refcount++;
1687                         return;
1688                 }
1689                 prevp = nextp;
1690                 nextp = nextp->bc_next;
1691         }
1692         ASSERT(prevp != NULL);
1693         bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1694                                              KM_SLEEP);
1695         bcp->bc_blkno = blkno;
1696         bcp->bc_len = len;
1697         bcp->bc_refcount = 1;
1698         bcp->bc_next = NULL;
1699         prevp->bc_next = bcp;
1700 }
1701
1702 /*
1703  * Check to see whether the buffer being recovered has a corresponding
1704  * entry in the buffer cancel record table.  If it does then return 1
1705  * so that it will be cancelled, otherwise return 0.  If the buffer is
1706  * actually a buffer cancel item (XFS_BLI_CANCEL is set), then decrement
1707  * the refcount on the entry in the table and remove it from the table
1708  * if this is the last reference.
1709  *
1710  * We remove the cancel record from the table when we encounter its
1711  * last occurrence in the log so that if the same buffer is re-used
1712  * again after its last cancellation we actually replay the changes
1713  * made at that point.
1714  */
1715 STATIC int
1716 xlog_check_buffer_cancelled(
1717         xlog_t                  *log,
1718         xfs_daddr_t             blkno,
1719         uint                    len,
1720         ushort                  flags)
1721 {
1722         xfs_buf_cancel_t        *bcp;
1723         xfs_buf_cancel_t        *prevp;
1724         xfs_buf_cancel_t        **bucket;
1725
1726         if (log->l_buf_cancel_table == NULL) {
1727                 /*
1728                  * There is nothing in the table built in pass one,
1729                  * so this buffer must not be cancelled.
1730                  */
1731                 ASSERT(!(flags & XFS_BLI_CANCEL));
1732                 return 0;
1733         }
1734
1735         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1736                                           XLOG_BC_TABLE_SIZE];
1737         bcp = *bucket;
1738         if (bcp == NULL) {
1739                 /*
1740                  * There is no corresponding entry in the table built
1741                  * in pass one, so this buffer has not been cancelled.
1742                  */
1743                 ASSERT(!(flags & XFS_BLI_CANCEL));
1744                 return 0;
1745         }
1746
1747         /*
1748          * Search for an entry in the buffer cancel table that
1749          * matches our buffer.
1750          */
1751         prevp = NULL;
1752         while (bcp != NULL) {
1753                 if (bcp->bc_blkno == blkno && bcp->bc_len == len) {
1754                         /*
1755                          * We've go a match, so return 1 so that the
1756                          * recovery of this buffer is cancelled.
1757                          * If this buffer is actually a buffer cancel
1758                          * log item, then decrement the refcount on the
1759                          * one in the table and remove it if this is the
1760                          * last reference.
1761                          */
1762                         if (flags & XFS_BLI_CANCEL) {
1763                                 bcp->bc_refcount--;
1764                                 if (bcp->bc_refcount == 0) {
1765                                         if (prevp == NULL) {
1766                                                 *bucket = bcp->bc_next;
1767                                         } else {
1768                                                 prevp->bc_next = bcp->bc_next;
1769                                         }
1770                                         kmem_free(bcp);
1771                                 }
1772                         }
1773                         return 1;
1774                 }
1775                 prevp = bcp;
1776                 bcp = bcp->bc_next;
1777         }
1778         /*
1779          * We didn't find a corresponding entry in the table, so
1780          * return 0 so that the buffer is NOT cancelled.
1781          */
1782         ASSERT(!(flags & XFS_BLI_CANCEL));
1783         return 0;
1784 }
1785
1786 STATIC int
1787 xlog_recover_do_buffer_pass2(
1788         xlog_t                  *log,
1789         xfs_buf_log_format_t    *buf_f)
1790 {
1791         xfs_daddr_t             blkno = 0;
1792         ushort                  flags = 0;
1793         uint                    len = 0;
1794
1795         switch (buf_f->blf_type) {
1796         case XFS_LI_BUF:
1797                 blkno = buf_f->blf_blkno;
1798                 flags = buf_f->blf_flags;
1799                 len = buf_f->blf_len;
1800                 break;
1801         }
1802
1803         return xlog_check_buffer_cancelled(log, blkno, len, flags);
1804 }
1805
1806 /*
1807  * Perform recovery for a buffer full of inodes.  In these buffers,
1808  * the only data which should be recovered is that which corresponds
1809  * to the di_next_unlinked pointers in the on disk inode structures.
1810  * The rest of the data for the inodes is always logged through the
1811  * inodes themselves rather than the inode buffer and is recovered
1812  * in xlog_recover_do_inode_trans().
1813  *
1814  * The only time when buffers full of inodes are fully recovered is
1815  * when the buffer is full of newly allocated inodes.  In this case
1816  * the buffer will not be marked as an inode buffer and so will be
1817  * sent to xlog_recover_do_reg_buffer() below during recovery.
1818  */
1819 STATIC int
1820 xlog_recover_do_inode_buffer(
1821         xfs_mount_t             *mp,
1822         xlog_recover_item_t     *item,
1823         xfs_buf_t               *bp,
1824         xfs_buf_log_format_t    *buf_f)
1825 {
1826         int                     i;
1827         int                     item_index;
1828         int                     bit;
1829         int                     nbits;
1830         int                     reg_buf_offset;
1831         int                     reg_buf_bytes;
1832         int                     next_unlinked_offset;
1833         int                     inodes_per_buf;
1834         xfs_agino_t             *logged_nextp;
1835         xfs_agino_t             *buffer_nextp;
1836         unsigned int            *data_map = NULL;
1837         unsigned int            map_size = 0;
1838
1839         switch (buf_f->blf_type) {
1840         case XFS_LI_BUF:
1841                 data_map = buf_f->blf_data_map;
1842                 map_size = buf_f->blf_map_size;
1843                 break;
1844         }
1845         /*
1846          * Set the variables corresponding to the current region to
1847          * 0 so that we'll initialize them on the first pass through
1848          * the loop.
1849          */
1850         reg_buf_offset = 0;
1851         reg_buf_bytes = 0;
1852         bit = 0;
1853         nbits = 0;
1854         item_index = 0;
1855         inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1856         for (i = 0; i < inodes_per_buf; i++) {
1857                 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1858                         offsetof(xfs_dinode_t, di_next_unlinked);
1859
1860                 while (next_unlinked_offset >=
1861                        (reg_buf_offset + reg_buf_bytes)) {
1862                         /*
1863                          * The next di_next_unlinked field is beyond
1864                          * the current logged region.  Find the next
1865                          * logged region that contains or is beyond
1866                          * the current di_next_unlinked field.
1867                          */
1868                         bit += nbits;
1869                         bit = xfs_next_bit(data_map, map_size, bit);
1870
1871                         /*
1872                          * If there are no more logged regions in the
1873                          * buffer, then we're done.
1874                          */
1875                         if (bit == -1) {
1876                                 return 0;
1877                         }
1878
1879                         nbits = xfs_contig_bits(data_map, map_size,
1880                                                          bit);
1881                         ASSERT(nbits > 0);
1882                         reg_buf_offset = bit << XFS_BLI_SHIFT;
1883                         reg_buf_bytes = nbits << XFS_BLI_SHIFT;
1884                         item_index++;
1885                 }
1886
1887                 /*
1888                  * If the current logged region starts after the current
1889                  * di_next_unlinked field, then move on to the next
1890                  * di_next_unlinked field.
1891                  */
1892                 if (next_unlinked_offset < reg_buf_offset) {
1893                         continue;
1894                 }
1895
1896                 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1897                 ASSERT((item->ri_buf[item_index].i_len % XFS_BLI_CHUNK) == 0);
1898                 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1899
1900                 /*
1901                  * The current logged region contains a copy of the
1902                  * current di_next_unlinked field.  Extract its value
1903                  * and copy it to the buffer copy.
1904                  */
1905                 logged_nextp = (xfs_agino_t *)
1906                                ((char *)(item->ri_buf[item_index].i_addr) +
1907                                 (next_unlinked_offset - reg_buf_offset));
1908                 if (unlikely(*logged_nextp == 0)) {
1909                         xfs_fs_cmn_err(CE_ALERT, mp,
1910                                 "bad inode buffer log record (ptr = 0x%p, bp = 0x%p).  XFS trying to replay bad (0) inode di_next_unlinked field",
1911                                 item, bp);
1912                         XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1913                                          XFS_ERRLEVEL_LOW, mp);
1914                         return XFS_ERROR(EFSCORRUPTED);
1915                 }
1916
1917                 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1918                                               next_unlinked_offset);
1919                 *buffer_nextp = *logged_nextp;
1920         }
1921
1922         return 0;
1923 }
1924
1925 /*
1926  * Perform a 'normal' buffer recovery.  Each logged region of the
1927  * buffer should be copied over the corresponding region in the
1928  * given buffer.  The bitmap in the buf log format structure indicates
1929  * where to place the logged data.
1930  */
1931 /*ARGSUSED*/
1932 STATIC void
1933 xlog_recover_do_reg_buffer(
1934         xlog_recover_item_t     *item,
1935         xfs_buf_t               *bp,
1936         xfs_buf_log_format_t    *buf_f)
1937 {
1938         int                     i;
1939         int                     bit;
1940         int                     nbits;
1941         unsigned int            *data_map = NULL;
1942         unsigned int            map_size = 0;
1943         int                     error;
1944
1945         switch (buf_f->blf_type) {
1946         case XFS_LI_BUF:
1947                 data_map = buf_f->blf_data_map;
1948                 map_size = buf_f->blf_map_size;
1949                 break;
1950         }
1951         bit = 0;
1952         i = 1;  /* 0 is the buf format structure */
1953         while (1) {
1954                 bit = xfs_next_bit(data_map, map_size, bit);
1955                 if (bit == -1)
1956                         break;
1957                 nbits = xfs_contig_bits(data_map, map_size, bit);
1958                 ASSERT(nbits > 0);
1959                 ASSERT(item->ri_buf[i].i_addr != NULL);
1960                 ASSERT(item->ri_buf[i].i_len % XFS_BLI_CHUNK == 0);
1961                 ASSERT(XFS_BUF_COUNT(bp) >=
1962                        ((uint)bit << XFS_BLI_SHIFT)+(nbits<<XFS_BLI_SHIFT));
1963
1964                 /*
1965                  * Do a sanity check if this is a dquot buffer. Just checking
1966                  * the first dquot in the buffer should do. XXXThis is
1967                  * probably a good thing to do for other buf types also.
1968                  */
1969                 error = 0;
1970                 if (buf_f->blf_flags &
1971                    (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
1972                         if (item->ri_buf[i].i_addr == NULL) {
1973                                 cmn_err(CE_ALERT,
1974                                         "XFS: NULL dquot in %s.", __func__);
1975                                 goto next;
1976                         }
1977                         if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
1978                                 cmn_err(CE_ALERT,
1979                                         "XFS: dquot too small (%d) in %s.",
1980                                         item->ri_buf[i].i_len, __func__);
1981                                 goto next;
1982                         }
1983                         error = xfs_qm_dqcheck((xfs_disk_dquot_t *)
1984                                                item->ri_buf[i].i_addr,
1985                                                -1, 0, XFS_QMOPT_DOWARN,
1986                                                "dquot_buf_recover");
1987                         if (error)
1988                                 goto next;
1989                 }
1990
1991                 memcpy(xfs_buf_offset(bp,
1992                         (uint)bit << XFS_BLI_SHIFT),    /* dest */
1993                         item->ri_buf[i].i_addr,         /* source */
1994                         nbits<<XFS_BLI_SHIFT);          /* length */
1995  next:
1996                 i++;
1997                 bit += nbits;
1998         }
1999
2000         /* Shouldn't be any more regions */
2001         ASSERT(i == item->ri_total);
2002 }
2003
2004 /*
2005  * Do some primitive error checking on ondisk dquot data structures.
2006  */
2007 int
2008 xfs_qm_dqcheck(
2009         xfs_disk_dquot_t *ddq,
2010         xfs_dqid_t       id,
2011         uint             type,    /* used only when IO_dorepair is true */
2012         uint             flags,
2013         char             *str)
2014 {
2015         xfs_dqblk_t      *d = (xfs_dqblk_t *)ddq;
2016         int             errs = 0;
2017
2018         /*
2019          * We can encounter an uninitialized dquot buffer for 2 reasons:
2020          * 1. If we crash while deleting the quotainode(s), and those blks got
2021          *    used for user data. This is because we take the path of regular
2022          *    file deletion; however, the size field of quotainodes is never
2023          *    updated, so all the tricks that we play in itruncate_finish
2024          *    don't quite matter.
2025          *
2026          * 2. We don't play the quota buffers when there's a quotaoff logitem.
2027          *    But the allocation will be replayed so we'll end up with an
2028          *    uninitialized quota block.
2029          *
2030          * This is all fine; things are still consistent, and we haven't lost
2031          * any quota information. Just don't complain about bad dquot blks.
2032          */
2033         if (be16_to_cpu(ddq->d_magic) != XFS_DQUOT_MAGIC) {
2034                 if (flags & XFS_QMOPT_DOWARN)
2035                         cmn_err(CE_ALERT,
2036                         "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
2037                         str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
2038                 errs++;
2039         }
2040         if (ddq->d_version != XFS_DQUOT_VERSION) {
2041                 if (flags & XFS_QMOPT_DOWARN)
2042                         cmn_err(CE_ALERT,
2043                         "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
2044                         str, id, ddq->d_version, XFS_DQUOT_VERSION);
2045                 errs++;
2046         }
2047
2048         if (ddq->d_flags != XFS_DQ_USER &&
2049             ddq->d_flags != XFS_DQ_PROJ &&
2050             ddq->d_flags != XFS_DQ_GROUP) {
2051                 if (flags & XFS_QMOPT_DOWARN)
2052                         cmn_err(CE_ALERT,
2053                         "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
2054                         str, id, ddq->d_flags);
2055                 errs++;
2056         }
2057
2058         if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
2059                 if (flags & XFS_QMOPT_DOWARN)
2060                         cmn_err(CE_ALERT,
2061                         "%s : ondisk-dquot 0x%p, ID mismatch: "
2062                         "0x%x expected, found id 0x%x",
2063                         str, ddq, id, be32_to_cpu(ddq->d_id));
2064                 errs++;
2065         }
2066
2067         if (!errs && ddq->d_id) {
2068                 if (ddq->d_blk_softlimit &&
2069                     be64_to_cpu(ddq->d_bcount) >=
2070                                 be64_to_cpu(ddq->d_blk_softlimit)) {
2071                         if (!ddq->d_btimer) {
2072                                 if (flags & XFS_QMOPT_DOWARN)
2073                                         cmn_err(CE_ALERT,
2074                                         "%s : Dquot ID 0x%x (0x%p) "
2075                                         "BLK TIMER NOT STARTED",
2076                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2077                                 errs++;
2078                         }
2079                 }
2080                 if (ddq->d_ino_softlimit &&
2081                     be64_to_cpu(ddq->d_icount) >=
2082                                 be64_to_cpu(ddq->d_ino_softlimit)) {
2083                         if (!ddq->d_itimer) {
2084                                 if (flags & XFS_QMOPT_DOWARN)
2085                                         cmn_err(CE_ALERT,
2086                                         "%s : Dquot ID 0x%x (0x%p) "
2087                                         "INODE TIMER NOT STARTED",
2088                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2089                                 errs++;
2090                         }
2091                 }
2092                 if (ddq->d_rtb_softlimit &&
2093                     be64_to_cpu(ddq->d_rtbcount) >=
2094                                 be64_to_cpu(ddq->d_rtb_softlimit)) {
2095                         if (!ddq->d_rtbtimer) {
2096                                 if (flags & XFS_QMOPT_DOWARN)
2097                                         cmn_err(CE_ALERT,
2098                                         "%s : Dquot ID 0x%x (0x%p) "
2099                                         "RTBLK TIMER NOT STARTED",
2100                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2101                                 errs++;
2102                         }
2103                 }
2104         }
2105
2106         if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2107                 return errs;
2108
2109         if (flags & XFS_QMOPT_DOWARN)
2110                 cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2111
2112         /*
2113          * Typically, a repair is only requested by quotacheck.
2114          */
2115         ASSERT(id != -1);
2116         ASSERT(flags & XFS_QMOPT_DQREPAIR);
2117         memset(d, 0, sizeof(xfs_dqblk_t));
2118
2119         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2120         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2121         d->dd_diskdq.d_flags = type;
2122         d->dd_diskdq.d_id = cpu_to_be32(id);
2123
2124         return errs;
2125 }
2126
2127 /*
2128  * Perform a dquot buffer recovery.
2129  * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2130  * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2131  * Else, treat it as a regular buffer and do recovery.
2132  */
2133 STATIC void
2134 xlog_recover_do_dquot_buffer(
2135         xfs_mount_t             *mp,
2136         xlog_t                  *log,
2137         xlog_recover_item_t     *item,
2138         xfs_buf_t               *bp,
2139         xfs_buf_log_format_t    *buf_f)
2140 {
2141         uint                    type;
2142
2143         /*
2144          * Filesystems are required to send in quota flags at mount time.
2145          */
2146         if (mp->m_qflags == 0) {
2147                 return;
2148         }
2149
2150         type = 0;
2151         if (buf_f->blf_flags & XFS_BLI_UDQUOT_BUF)
2152                 type |= XFS_DQ_USER;
2153         if (buf_f->blf_flags & XFS_BLI_PDQUOT_BUF)
2154                 type |= XFS_DQ_PROJ;
2155         if (buf_f->blf_flags & XFS_BLI_GDQUOT_BUF)
2156                 type |= XFS_DQ_GROUP;
2157         /*
2158          * This type of quotas was turned off, so ignore this buffer
2159          */
2160         if (log->l_quotaoffs_flag & type)
2161                 return;
2162
2163         xlog_recover_do_reg_buffer(item, bp, buf_f);
2164 }
2165
2166 /*
2167  * This routine replays a modification made to a buffer at runtime.
2168  * There are actually two types of buffer, regular and inode, which
2169  * are handled differently.  Inode buffers are handled differently
2170  * in that we only recover a specific set of data from them, namely
2171  * the inode di_next_unlinked fields.  This is because all other inode
2172  * data is actually logged via inode records and any data we replay
2173  * here which overlaps that may be stale.
2174  *
2175  * When meta-data buffers are freed at run time we log a buffer item
2176  * with the XFS_BLI_CANCEL bit set to indicate that previous copies
2177  * of the buffer in the log should not be replayed at recovery time.
2178  * This is so that if the blocks covered by the buffer are reused for
2179  * file data before we crash we don't end up replaying old, freed
2180  * meta-data into a user's file.
2181  *
2182  * To handle the cancellation of buffer log items, we make two passes
2183  * over the log during recovery.  During the first we build a table of
2184  * those buffers which have been cancelled, and during the second we
2185  * only replay those buffers which do not have corresponding cancel
2186  * records in the table.  See xlog_recover_do_buffer_pass[1,2] above
2187  * for more details on the implementation of the table of cancel records.
2188  */
2189 STATIC int
2190 xlog_recover_do_buffer_trans(
2191         xlog_t                  *log,
2192         xlog_recover_item_t     *item,
2193         int                     pass)
2194 {
2195         xfs_buf_log_format_t    *buf_f;
2196         xfs_mount_t             *mp;
2197         xfs_buf_t               *bp;
2198         int                     error;
2199         int                     cancel;
2200         xfs_daddr_t             blkno;
2201         int                     len;
2202         ushort                  flags;
2203         uint                    buf_flags;
2204
2205         buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
2206
2207         if (pass == XLOG_RECOVER_PASS1) {
2208                 /*
2209                  * In this pass we're only looking for buf items
2210                  * with the XFS_BLI_CANCEL bit set.
2211                  */
2212                 xlog_recover_do_buffer_pass1(log, buf_f);
2213                 return 0;
2214         } else {
2215                 /*
2216                  * In this pass we want to recover all the buffers
2217                  * which have not been cancelled and are not
2218                  * cancellation buffers themselves.  The routine
2219                  * we call here will tell us whether or not to
2220                  * continue with the replay of this buffer.
2221                  */
2222                 cancel = xlog_recover_do_buffer_pass2(log, buf_f);
2223                 if (cancel) {
2224                         return 0;
2225                 }
2226         }
2227         switch (buf_f->blf_type) {
2228         case XFS_LI_BUF:
2229                 blkno = buf_f->blf_blkno;
2230                 len = buf_f->blf_len;
2231                 flags = buf_f->blf_flags;
2232                 break;
2233         default:
2234                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
2235                         "xfs_log_recover: unknown buffer type 0x%x, logdev %s",
2236                         buf_f->blf_type, log->l_mp->m_logname ?
2237                         log->l_mp->m_logname : "internal");
2238                 XFS_ERROR_REPORT("xlog_recover_do_buffer_trans",
2239                                  XFS_ERRLEVEL_LOW, log->l_mp);
2240                 return XFS_ERROR(EFSCORRUPTED);
2241         }
2242
2243         mp = log->l_mp;
2244         buf_flags = XFS_BUF_LOCK;
2245         if (!(flags & XFS_BLI_INODE_BUF))
2246                 buf_flags |= XFS_BUF_MAPPED;
2247
2248         bp = xfs_buf_read(mp->m_ddev_targp, blkno, len, buf_flags);
2249         if (XFS_BUF_ISERROR(bp)) {
2250                 xfs_ioerror_alert("xlog_recover_do..(read#1)", log->l_mp,
2251                                   bp, blkno);
2252                 error = XFS_BUF_GETERROR(bp);
2253                 xfs_buf_relse(bp);
2254                 return error;
2255         }
2256
2257         error = 0;
2258         if (flags & XFS_BLI_INODE_BUF) {
2259                 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2260         } else if (flags &
2261                   (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
2262                 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2263         } else {
2264                 xlog_recover_do_reg_buffer(item, bp, buf_f);
2265         }
2266         if (error)
2267                 return XFS_ERROR(error);
2268
2269         /*
2270          * Perform delayed write on the buffer.  Asynchronous writes will be
2271          * slower when taking into account all the buffers to be flushed.
2272          *
2273          * Also make sure that only inode buffers with good sizes stay in
2274          * the buffer cache.  The kernel moves inodes in buffers of 1 block
2275          * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger.  The inode
2276          * buffers in the log can be a different size if the log was generated
2277          * by an older kernel using unclustered inode buffers or a newer kernel
2278          * running with a different inode cluster size.  Regardless, if the
2279          * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2280          * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2281          * the buffer out of the buffer cache so that the buffer won't
2282          * overlap with future reads of those inodes.
2283          */
2284         if (XFS_DINODE_MAGIC ==
2285             be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2286             (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2287                         (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2288                 XFS_BUF_STALE(bp);
2289                 error = xfs_bwrite(mp, bp);
2290         } else {
2291                 ASSERT(bp->b_mount == NULL || bp->b_mount == mp);
2292                 bp->b_mount = mp;
2293                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2294                 xfs_bdwrite(mp, bp);
2295         }
2296
2297         return (error);
2298 }
2299
2300 STATIC int
2301 xlog_recover_do_inode_trans(
2302         xlog_t                  *log,
2303         xlog_recover_item_t     *item,
2304         int                     pass)
2305 {
2306         xfs_inode_log_format_t  *in_f;
2307         xfs_mount_t             *mp;
2308         xfs_buf_t               *bp;
2309         xfs_dinode_t            *dip;
2310         xfs_ino_t               ino;
2311         int                     len;
2312         xfs_caddr_t             src;
2313         xfs_caddr_t             dest;
2314         int                     error;
2315         int                     attr_index;
2316         uint                    fields;
2317         xfs_icdinode_t          *dicp;
2318         int                     need_free = 0;
2319
2320         if (pass == XLOG_RECOVER_PASS1) {
2321                 return 0;
2322         }
2323
2324         if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2325                 in_f = (xfs_inode_log_format_t *)item->ri_buf[0].i_addr;
2326         } else {
2327                 in_f = (xfs_inode_log_format_t *)kmem_alloc(
2328                         sizeof(xfs_inode_log_format_t), KM_SLEEP);
2329                 need_free = 1;
2330                 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2331                 if (error)
2332                         goto error;
2333         }
2334         ino = in_f->ilf_ino;
2335         mp = log->l_mp;
2336
2337         /*
2338          * Inode buffers can be freed, look out for it,
2339          * and do not replay the inode.
2340          */
2341         if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2342                                         in_f->ilf_len, 0)) {
2343                 error = 0;
2344                 goto error;
2345         }
2346
2347         bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len,
2348                           XFS_BUF_LOCK);
2349         if (XFS_BUF_ISERROR(bp)) {
2350                 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2351                                   bp, in_f->ilf_blkno);
2352                 error = XFS_BUF_GETERROR(bp);
2353                 xfs_buf_relse(bp);
2354                 goto error;
2355         }
2356         error = 0;
2357         ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2358         dip = (xfs_dinode_t *)xfs_buf_offset(bp, in_f->ilf_boffset);
2359
2360         /*
2361          * Make sure the place we're flushing out to really looks
2362          * like an inode!
2363          */
2364         if (unlikely(be16_to_cpu(dip->di_magic) != XFS_DINODE_MAGIC)) {
2365                 xfs_buf_relse(bp);
2366                 xfs_fs_cmn_err(CE_ALERT, mp,
2367                         "xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2368                         dip, bp, ino);
2369                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(1)",
2370                                  XFS_ERRLEVEL_LOW, mp);
2371                 error = EFSCORRUPTED;
2372                 goto error;
2373         }
2374         dicp = (xfs_icdinode_t *)(item->ri_buf[1].i_addr);
2375         if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2376                 xfs_buf_relse(bp);
2377                 xfs_fs_cmn_err(CE_ALERT, mp,
2378                         "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2379                         item, ino);
2380                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(2)",
2381                                  XFS_ERRLEVEL_LOW, mp);
2382                 error = EFSCORRUPTED;
2383                 goto error;
2384         }
2385
2386         /* Skip replay when the on disk inode is newer than the log one */
2387         if (dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
2388                 /*
2389                  * Deal with the wrap case, DI_MAX_FLUSH is less
2390                  * than smaller numbers
2391                  */
2392                 if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
2393                     dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2394                         /* do nothing */
2395                 } else {
2396                         xfs_buf_relse(bp);
2397                         error = 0;
2398                         goto error;
2399                 }
2400         }
2401         /* Take the opportunity to reset the flush iteration count */
2402         dicp->di_flushiter = 0;
2403
2404         if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2405                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2406                     (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2407                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(3)",
2408                                          XFS_ERRLEVEL_LOW, mp, dicp);
2409                         xfs_buf_relse(bp);
2410                         xfs_fs_cmn_err(CE_ALERT, mp,
2411                                 "xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2412                                 item, dip, bp, ino);
2413                         error = EFSCORRUPTED;
2414                         goto error;
2415                 }
2416         } else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2417                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2418                     (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2419                     (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2420                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(4)",
2421                                              XFS_ERRLEVEL_LOW, mp, dicp);
2422                         xfs_buf_relse(bp);
2423                         xfs_fs_cmn_err(CE_ALERT, mp,
2424                                 "xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2425                                 item, dip, bp, ino);
2426                         error = EFSCORRUPTED;
2427                         goto error;
2428                 }
2429         }
2430         if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2431                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(5)",
2432                                      XFS_ERRLEVEL_LOW, mp, dicp);
2433                 xfs_buf_relse(bp);
2434                 xfs_fs_cmn_err(CE_ALERT, mp,
2435                         "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",
2436                         item, dip, bp, ino,
2437                         dicp->di_nextents + dicp->di_anextents,
2438                         dicp->di_nblocks);
2439                 error = EFSCORRUPTED;
2440                 goto error;
2441         }
2442         if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2443                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(6)",
2444                                      XFS_ERRLEVEL_LOW, mp, dicp);
2445                 xfs_buf_relse(bp);
2446                 xfs_fs_cmn_err(CE_ALERT, mp,
2447                         "xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2448                         item, dip, bp, ino, dicp->di_forkoff);
2449                 error = EFSCORRUPTED;
2450                 goto error;
2451         }
2452         if (unlikely(item->ri_buf[1].i_len > sizeof(struct xfs_icdinode))) {
2453                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(7)",
2454                                      XFS_ERRLEVEL_LOW, mp, dicp);
2455                 xfs_buf_relse(bp);
2456                 xfs_fs_cmn_err(CE_ALERT, mp,
2457                         "xfs_inode_recover: Bad inode log record length %d, rec ptr 0x%p",
2458                         item->ri_buf[1].i_len, item);
2459                 error = EFSCORRUPTED;
2460                 goto error;
2461         }
2462
2463         /* The core is in in-core format */
2464         xfs_dinode_to_disk(dip, (xfs_icdinode_t *)item->ri_buf[1].i_addr);
2465
2466         /* the rest is in on-disk format */
2467         if (item->ri_buf[1].i_len > sizeof(struct xfs_icdinode)) {
2468                 memcpy((xfs_caddr_t) dip + sizeof(struct xfs_icdinode),
2469                         item->ri_buf[1].i_addr + sizeof(struct xfs_icdinode),
2470                         item->ri_buf[1].i_len  - sizeof(struct xfs_icdinode));
2471         }
2472
2473         fields = in_f->ilf_fields;
2474         switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2475         case XFS_ILOG_DEV:
2476                 xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
2477                 break;
2478         case XFS_ILOG_UUID:
2479                 memcpy(XFS_DFORK_DPTR(dip),
2480                        &in_f->ilf_u.ilfu_uuid,
2481                        sizeof(uuid_t));
2482                 break;
2483         }
2484
2485         if (in_f->ilf_size == 2)
2486                 goto write_inode_buffer;
2487         len = item->ri_buf[2].i_len;
2488         src = item->ri_buf[2].i_addr;
2489         ASSERT(in_f->ilf_size <= 4);
2490         ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2491         ASSERT(!(fields & XFS_ILOG_DFORK) ||
2492                (len == in_f->ilf_dsize));
2493
2494         switch (fields & XFS_ILOG_DFORK) {
2495         case XFS_ILOG_DDATA:
2496         case XFS_ILOG_DEXT:
2497                 memcpy(XFS_DFORK_DPTR(dip), src, len);
2498                 break;
2499
2500         case XFS_ILOG_DBROOT:
2501                 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
2502                                  (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
2503                                  XFS_DFORK_DSIZE(dip, mp));
2504                 break;
2505
2506         default:
2507                 /*
2508                  * There are no data fork flags set.
2509                  */
2510                 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2511                 break;
2512         }
2513
2514         /*
2515          * If we logged any attribute data, recover it.  There may or
2516          * may not have been any other non-core data logged in this
2517          * transaction.
2518          */
2519         if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2520                 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2521                         attr_index = 3;
2522                 } else {
2523                         attr_index = 2;
2524                 }
2525                 len = item->ri_buf[attr_index].i_len;
2526                 src = item->ri_buf[attr_index].i_addr;
2527                 ASSERT(len == in_f->ilf_asize);
2528
2529                 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2530                 case XFS_ILOG_ADATA:
2531                 case XFS_ILOG_AEXT:
2532                         dest = XFS_DFORK_APTR(dip);
2533                         ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2534                         memcpy(dest, src, len);
2535                         break;
2536
2537                 case XFS_ILOG_ABROOT:
2538                         dest = XFS_DFORK_APTR(dip);
2539                         xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
2540                                          len, (xfs_bmdr_block_t*)dest,
2541                                          XFS_DFORK_ASIZE(dip, mp));
2542                         break;
2543
2544                 default:
2545                         xlog_warn("XFS: xlog_recover_do_inode_trans: Invalid flag");
2546                         ASSERT(0);
2547                         xfs_buf_relse(bp);
2548                         error = EIO;
2549                         goto error;
2550                 }
2551         }
2552
2553 write_inode_buffer:
2554         ASSERT(bp->b_mount == NULL || bp->b_mount == mp);
2555         bp->b_mount = mp;
2556         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2557         xfs_bdwrite(mp, bp);
2558 error:
2559         if (need_free)
2560                 kmem_free(in_f);
2561         return XFS_ERROR(error);
2562 }
2563
2564 /*
2565  * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2566  * structure, so that we know not to do any dquot item or dquot buffer recovery,
2567  * of that type.
2568  */
2569 STATIC int
2570 xlog_recover_do_quotaoff_trans(
2571         xlog_t                  *log,
2572         xlog_recover_item_t     *item,
2573         int                     pass)
2574 {
2575         xfs_qoff_logformat_t    *qoff_f;
2576
2577         if (pass == XLOG_RECOVER_PASS2) {
2578                 return (0);
2579         }
2580
2581         qoff_f = (xfs_qoff_logformat_t *)item->ri_buf[0].i_addr;
2582         ASSERT(qoff_f);
2583
2584         /*
2585          * The logitem format's flag tells us if this was user quotaoff,
2586          * group/project quotaoff or both.
2587          */
2588         if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2589                 log->l_quotaoffs_flag |= XFS_DQ_USER;
2590         if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2591                 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2592         if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2593                 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2594
2595         return (0);
2596 }
2597
2598 /*
2599  * Recover a dquot record
2600  */
2601 STATIC int
2602 xlog_recover_do_dquot_trans(
2603         xlog_t                  *log,
2604         xlog_recover_item_t     *item,
2605         int                     pass)
2606 {
2607         xfs_mount_t             *mp;
2608         xfs_buf_t               *bp;
2609         struct xfs_disk_dquot   *ddq, *recddq;
2610         int                     error;
2611         xfs_dq_logformat_t      *dq_f;
2612         uint                    type;
2613
2614         if (pass == XLOG_RECOVER_PASS1) {
2615                 return 0;
2616         }
2617         mp = log->l_mp;
2618
2619         /*
2620          * Filesystems are required to send in quota flags at mount time.
2621          */
2622         if (mp->m_qflags == 0)
2623                 return (0);
2624
2625         recddq = (xfs_disk_dquot_t *)item->ri_buf[1].i_addr;
2626
2627         if (item->ri_buf[1].i_addr == NULL) {
2628                 cmn_err(CE_ALERT,
2629                         "XFS: NULL dquot in %s.", __func__);
2630                 return XFS_ERROR(EIO);
2631         }
2632         if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
2633                 cmn_err(CE_ALERT,
2634                         "XFS: dquot too small (%d) in %s.",
2635                         item->ri_buf[1].i_len, __func__);
2636                 return XFS_ERROR(EIO);
2637         }
2638
2639         /*
2640          * This type of quotas was turned off, so ignore this record.
2641          */
2642         type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2643         ASSERT(type);
2644         if (log->l_quotaoffs_flag & type)
2645                 return (0);
2646
2647         /*
2648          * At this point we know that quota was _not_ turned off.
2649          * Since the mount flags are not indicating to us otherwise, this
2650          * must mean that quota is on, and the dquot needs to be replayed.
2651          * Remember that we may not have fully recovered the superblock yet,
2652          * so we can't do the usual trick of looking at the SB quota bits.
2653          *
2654          * The other possibility, of course, is that the quota subsystem was
2655          * removed since the last mount - ENOSYS.
2656          */
2657         dq_f = (xfs_dq_logformat_t *)item->ri_buf[0].i_addr;
2658         ASSERT(dq_f);
2659         if ((error = xfs_qm_dqcheck(recddq,
2660                            dq_f->qlf_id,
2661                            0, XFS_QMOPT_DOWARN,
2662                            "xlog_recover_do_dquot_trans (log copy)"))) {
2663                 return XFS_ERROR(EIO);
2664         }
2665         ASSERT(dq_f->qlf_len == 1);
2666
2667         error = xfs_read_buf(mp, mp->m_ddev_targp,
2668                              dq_f->qlf_blkno,
2669                              XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2670                              0, &bp);
2671         if (error) {
2672                 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2673                                   bp, dq_f->qlf_blkno);
2674                 return error;
2675         }
2676         ASSERT(bp);
2677         ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2678
2679         /*
2680          * At least the magic num portion should be on disk because this
2681          * was among a chunk of dquots created earlier, and we did some
2682          * minimal initialization then.
2683          */
2684         if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2685                            "xlog_recover_do_dquot_trans")) {
2686                 xfs_buf_relse(bp);
2687                 return XFS_ERROR(EIO);
2688         }
2689
2690         memcpy(ddq, recddq, item->ri_buf[1].i_len);
2691
2692         ASSERT(dq_f->qlf_size == 2);
2693         ASSERT(bp->b_mount == NULL || bp->b_mount == mp);
2694         bp->b_mount = mp;
2695         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2696         xfs_bdwrite(mp, bp);
2697
2698         return (0);
2699 }
2700
2701 /*
2702  * This routine is called to create an in-core extent free intent
2703  * item from the efi format structure which was logged on disk.
2704  * It allocates an in-core efi, copies the extents from the format
2705  * structure into it, and adds the efi to the AIL with the given
2706  * LSN.
2707  */
2708 STATIC int
2709 xlog_recover_do_efi_trans(
2710         xlog_t                  *log,
2711         xlog_recover_item_t     *item,
2712         xfs_lsn_t               lsn,
2713         int                     pass)
2714 {
2715         int                     error;
2716         xfs_mount_t             *mp;
2717         xfs_efi_log_item_t      *efip;
2718         xfs_efi_log_format_t    *efi_formatp;
2719
2720         if (pass == XLOG_RECOVER_PASS1) {
2721                 return 0;
2722         }
2723
2724         efi_formatp = (xfs_efi_log_format_t *)item->ri_buf[0].i_addr;
2725
2726         mp = log->l_mp;
2727         efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2728         if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2729                                          &(efip->efi_format)))) {
2730                 xfs_efi_item_free(efip);
2731                 return error;
2732         }
2733         efip->efi_next_extent = efi_formatp->efi_nextents;
2734         efip->efi_flags |= XFS_EFI_COMMITTED;
2735
2736         spin_lock(&log->l_ailp->xa_lock);
2737         /*
2738          * xfs_trans_ail_update() drops the AIL lock.
2739          */
2740         xfs_trans_ail_update(log->l_ailp, (xfs_log_item_t *)efip, lsn);
2741         return 0;
2742 }
2743
2744
2745 /*
2746  * This routine is called when an efd format structure is found in
2747  * a committed transaction in the log.  It's purpose is to cancel
2748  * the corresponding efi if it was still in the log.  To do this
2749  * it searches the AIL for the efi with an id equal to that in the
2750  * efd format structure.  If we find it, we remove the efi from the
2751  * AIL and free it.
2752  */
2753 STATIC void
2754 xlog_recover_do_efd_trans(
2755         xlog_t                  *log,
2756         xlog_recover_item_t     *item,
2757         int                     pass)
2758 {
2759         xfs_efd_log_format_t    *efd_formatp;
2760         xfs_efi_log_item_t      *efip = NULL;
2761         xfs_log_item_t          *lip;
2762         __uint64_t              efi_id;
2763         struct xfs_ail_cursor   cur;
2764         struct xfs_ail          *ailp = log->l_ailp;
2765
2766         if (pass == XLOG_RECOVER_PASS1) {
2767                 return;
2768         }
2769
2770         efd_formatp = (xfs_efd_log_format_t *)item->ri_buf[0].i_addr;
2771         ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2772                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2773                (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2774                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2775         efi_id = efd_formatp->efd_efi_id;
2776
2777         /*
2778          * Search for the efi with the id in the efd format structure
2779          * in the AIL.
2780          */
2781         spin_lock(&ailp->xa_lock);
2782         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2783         while (lip != NULL) {
2784                 if (lip->li_type == XFS_LI_EFI) {
2785                         efip = (xfs_efi_log_item_t *)lip;
2786                         if (efip->efi_format.efi_id == efi_id) {
2787                                 /*
2788                                  * xfs_trans_ail_delete() drops the
2789                                  * AIL lock.
2790                                  */
2791                                 xfs_trans_ail_delete(ailp, lip);
2792                                 xfs_efi_item_free(efip);
2793                                 spin_lock(&ailp->xa_lock);
2794                                 break;
2795                         }
2796                 }
2797                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2798         }
2799         xfs_trans_ail_cursor_done(ailp, &cur);
2800         spin_unlock(&ailp->xa_lock);
2801 }
2802
2803 /*
2804  * Perform the transaction
2805  *
2806  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2807  * EFIs and EFDs get queued up by adding entries into the AIL for them.
2808  */
2809 STATIC int
2810 xlog_recover_do_trans(
2811         xlog_t                  *log,
2812         xlog_recover_t          *trans,
2813         int                     pass)
2814 {
2815         int                     error = 0;
2816         xlog_recover_item_t     *item, *first_item;
2817
2818         error = xlog_recover_reorder_trans(trans);
2819         if (error)
2820                 return error;
2821
2822         first_item = item = trans->r_itemq;
2823         do {
2824                 switch (ITEM_TYPE(item)) {
2825                 case XFS_LI_BUF:
2826                         error = xlog_recover_do_buffer_trans(log, item, pass);
2827                         break;
2828                 case XFS_LI_INODE:
2829                         error = xlog_recover_do_inode_trans(log, item, pass);
2830                         break;
2831                 case XFS_LI_EFI:
2832                         error = xlog_recover_do_efi_trans(log, item,
2833                                                           trans->r_lsn, pass);
2834                         break;
2835                 case XFS_LI_EFD:
2836                         xlog_recover_do_efd_trans(log, item, pass);
2837                         error = 0;
2838                         break;
2839                 case XFS_LI_DQUOT:
2840                         error = xlog_recover_do_dquot_trans(log, item, pass);
2841                         break;
2842                 case XFS_LI_QUOTAOFF:
2843                         error = xlog_recover_do_quotaoff_trans(log, item,
2844                                                                pass);
2845                         break;
2846                 default:
2847                         xlog_warn(
2848         "XFS: invalid item type (%d) xlog_recover_do_trans", ITEM_TYPE(item));
2849                         ASSERT(0);
2850                         error = XFS_ERROR(EIO);
2851                         break;
2852                 }
2853
2854                 if (error)
2855                         return error;
2856                 item = item->ri_next;
2857         } while (first_item != item);
2858
2859         return 0;
2860 }
2861
2862 /*
2863  * Free up any resources allocated by the transaction
2864  *
2865  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2866  */
2867 STATIC void
2868 xlog_recover_free_trans(
2869         xlog_recover_t          *trans)
2870 {
2871         xlog_recover_item_t     *first_item, *item, *free_item;
2872         int                     i;
2873
2874         item = first_item = trans->r_itemq;
2875         do {
2876                 free_item = item;
2877                 item = item->ri_next;
2878                  /* Free the regions in the item. */
2879                 for (i = 0; i < free_item->ri_cnt; i++) {
2880                         kmem_free(free_item->ri_buf[i].i_addr);
2881                 }
2882                 /* Free the item itself */
2883                 kmem_free(free_item->ri_buf);
2884                 kmem_free(free_item);
2885         } while (first_item != item);
2886         /* Free the transaction recover structure */
2887         kmem_free(trans);
2888 }
2889
2890 STATIC int
2891 xlog_recover_commit_trans(
2892         xlog_t                  *log,
2893         xlog_recover_t          **q,
2894         xlog_recover_t          *trans,
2895         int                     pass)
2896 {
2897         int                     error;
2898
2899         if ((error = xlog_recover_unlink_tid(q, trans)))
2900                 return error;
2901         if ((error = xlog_recover_do_trans(log, trans, pass)))
2902                 return error;
2903         xlog_recover_free_trans(trans);                 /* no error */
2904         return 0;
2905 }
2906
2907 STATIC int
2908 xlog_recover_unmount_trans(
2909         xlog_recover_t          *trans)
2910 {
2911         /* Do nothing now */
2912         xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2913         return 0;
2914 }
2915
2916 /*
2917  * There are two valid states of the r_state field.  0 indicates that the
2918  * transaction structure is in a normal state.  We have either seen the
2919  * start of the transaction or the last operation we added was not a partial
2920  * operation.  If the last operation we added to the transaction was a
2921  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2922  *
2923  * NOTE: skip LRs with 0 data length.
2924  */
2925 STATIC int
2926 xlog_recover_process_data(
2927         xlog_t                  *log,
2928         xlog_recover_t          *rhash[],
2929         xlog_rec_header_t       *rhead,
2930         xfs_caddr_t             dp,
2931         int                     pass)
2932 {
2933         xfs_caddr_t             lp;
2934         int                     num_logops;
2935         xlog_op_header_t        *ohead;
2936         xlog_recover_t          *trans;
2937         xlog_tid_t              tid;
2938         int                     error;
2939         unsigned long           hash;
2940         uint                    flags;
2941
2942         lp = dp + be32_to_cpu(rhead->h_len);
2943         num_logops = be32_to_cpu(rhead->h_num_logops);
2944
2945         /* check the log format matches our own - else we can't recover */
2946         if (xlog_header_check_recover(log->l_mp, rhead))
2947                 return (XFS_ERROR(EIO));
2948
2949         while ((dp < lp) && num_logops) {
2950                 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2951                 ohead = (xlog_op_header_t *)dp;
2952                 dp += sizeof(xlog_op_header_t);
2953                 if (ohead->oh_clientid != XFS_TRANSACTION &&
2954                     ohead->oh_clientid != XFS_LOG) {
2955                         xlog_warn(
2956                 "XFS: xlog_recover_process_data: bad clientid");
2957                         ASSERT(0);
2958                         return (XFS_ERROR(EIO));
2959                 }
2960                 tid = be32_to_cpu(ohead->oh_tid);
2961                 hash = XLOG_RHASH(tid);
2962                 trans = xlog_recover_find_tid(rhash[hash], tid);
2963                 if (trans == NULL) {               /* not found; add new tid */
2964                         if (ohead->oh_flags & XLOG_START_TRANS)
2965                                 xlog_recover_new_tid(&rhash[hash], tid,
2966                                         be64_to_cpu(rhead->h_lsn));
2967                 } else {
2968                         if (dp + be32_to_cpu(ohead->oh_len) > lp) {
2969                                 xlog_warn(
2970                         "XFS: xlog_recover_process_data: bad length");
2971                                 WARN_ON(1);
2972                                 return (XFS_ERROR(EIO));
2973                         }
2974                         flags = ohead->oh_flags & ~XLOG_END_TRANS;
2975                         if (flags & XLOG_WAS_CONT_TRANS)
2976                                 flags &= ~XLOG_CONTINUE_TRANS;
2977                         switch (flags) {
2978                         case XLOG_COMMIT_TRANS:
2979                                 error = xlog_recover_commit_trans(log,
2980                                                 &rhash[hash], trans, pass);
2981                                 break;
2982                         case XLOG_UNMOUNT_TRANS:
2983                                 error = xlog_recover_unmount_trans(trans);
2984                                 break;
2985                         case XLOG_WAS_CONT_TRANS:
2986                                 error = xlog_recover_add_to_cont_trans(trans,
2987                                                 dp, be32_to_cpu(ohead->oh_len));
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, be32_to_cpu(ohead->oh_len));
2999                                 break;
3000                         default:
3001                                 xlog_warn(
3002                         "XFS: xlog_recover_process_data: bad flag");
3003                                 ASSERT(0);
3004                                 error = XFS_ERROR(EIO);
3005                                 break;
3006                         }
3007                         if (error)
3008                                 return error;
3009                 }
3010                 dp += be32_to_cpu(ohead->oh_len);
3011                 num_logops--;
3012         }
3013         return 0;
3014 }
3015
3016 /*
3017  * Process an extent free intent item that was recovered from
3018  * the log.  We need to free the extents that it describes.
3019  */
3020 STATIC int
3021 xlog_recover_process_efi(
3022         xfs_mount_t             *mp,
3023         xfs_efi_log_item_t      *efip)
3024 {
3025         xfs_efd_log_item_t      *efdp;
3026         xfs_trans_t    &nbs