Merge git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / 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_imap.h"
40 #include "xfs_alloc.h"
41 #include "xfs_ialloc.h"
42 #include "xfs_log_priv.h"
43 #include "xfs_buf_item.h"
44 #include "xfs_log_recover.h"
45 #include "xfs_extfree_item.h"
46 #include "xfs_trans_priv.h"
47 #include "xfs_quota.h"
48 #include "xfs_rw.h"
49 #include "xfs_utils.h"
50
51 STATIC int      xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
52 STATIC int      xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
53 STATIC void     xlog_recover_insert_item_backq(xlog_recover_item_t **q,
54                                                xlog_recover_item_t *item);
55 #if defined(DEBUG)
56 STATIC void     xlog_recover_check_summary(xlog_t *);
57 #else
58 #define xlog_recover_check_summary(log)
59 #endif
60
61
62 /*
63  * Sector aligned buffer routines for buffer create/read/write/access
64  */
65
66 #define XLOG_SECTOR_ROUNDUP_BBCOUNT(log, bbs)   \
67         ( ((log)->l_sectbb_mask && (bbs & (log)->l_sectbb_mask)) ? \
68         ((bbs + (log)->l_sectbb_mask + 1) & ~(log)->l_sectbb_mask) : (bbs) )
69 #define XLOG_SECTOR_ROUNDDOWN_BLKNO(log, bno)   ((bno) & ~(log)->l_sectbb_mask)
70
71 xfs_buf_t *
72 xlog_get_bp(
73         xlog_t          *log,
74         int             num_bblks)
75 {
76         ASSERT(num_bblks > 0);
77
78         if (log->l_sectbb_log) {
79                 if (num_bblks > 1)
80                         num_bblks += XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
81                 num_bblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, num_bblks);
82         }
83         return xfs_buf_get_noaddr(BBTOB(num_bblks), log->l_mp->m_logdev_targp);
84 }
85
86 void
87 xlog_put_bp(
88         xfs_buf_t       *bp)
89 {
90         xfs_buf_free(bp);
91 }
92
93
94 /*
95  * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
96  */
97 int
98 xlog_bread(
99         xlog_t          *log,
100         xfs_daddr_t     blk_no,
101         int             nbblks,
102         xfs_buf_t       *bp)
103 {
104         int             error;
105
106         if (log->l_sectbb_log) {
107                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
108                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
109         }
110
111         ASSERT(nbblks > 0);
112         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
113         ASSERT(bp);
114
115         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
116         XFS_BUF_READ(bp);
117         XFS_BUF_BUSY(bp);
118         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
119         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
120
121         xfsbdstrat(log->l_mp, bp);
122         error = xfs_iowait(bp);
123         if (error)
124                 xfs_ioerror_alert("xlog_bread", log->l_mp,
125                                   bp, XFS_BUF_ADDR(bp));
126         return error;
127 }
128
129 /*
130  * Write out the buffer at the given block for the given number of blocks.
131  * The buffer is kept locked across the write and is returned locked.
132  * This can only be used for synchronous log writes.
133  */
134 STATIC int
135 xlog_bwrite(
136         xlog_t          *log,
137         xfs_daddr_t     blk_no,
138         int             nbblks,
139         xfs_buf_t       *bp)
140 {
141         int             error;
142
143         if (log->l_sectbb_log) {
144                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
145                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
146         }
147
148         ASSERT(nbblks > 0);
149         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
150
151         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
152         XFS_BUF_ZEROFLAGS(bp);
153         XFS_BUF_BUSY(bp);
154         XFS_BUF_HOLD(bp);
155         XFS_BUF_PSEMA(bp, PRIBIO);
156         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
157         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
158
159         if ((error = xfs_bwrite(log->l_mp, bp)))
160                 xfs_ioerror_alert("xlog_bwrite", log->l_mp,
161                                   bp, XFS_BUF_ADDR(bp));
162         return error;
163 }
164
165 STATIC xfs_caddr_t
166 xlog_align(
167         xlog_t          *log,
168         xfs_daddr_t     blk_no,
169         int             nbblks,
170         xfs_buf_t       *bp)
171 {
172         xfs_caddr_t     ptr;
173
174         if (!log->l_sectbb_log)
175                 return XFS_BUF_PTR(bp);
176
177         ptr = XFS_BUF_PTR(bp) + BBTOB((int)blk_no & log->l_sectbb_mask);
178         ASSERT(XFS_BUF_SIZE(bp) >=
179                 BBTOB(nbblks + (blk_no & log->l_sectbb_mask)));
180         return ptr;
181 }
182
183 #ifdef DEBUG
184 /*
185  * dump debug superblock and log record information
186  */
187 STATIC void
188 xlog_header_check_dump(
189         xfs_mount_t             *mp,
190         xlog_rec_header_t       *head)
191 {
192         int                     b;
193
194         cmn_err(CE_DEBUG, "%s:  SB : uuid = ", __func__);
195         for (b = 0; b < 16; b++)
196                 cmn_err(CE_DEBUG, "%02x", ((uchar_t *)&mp->m_sb.sb_uuid)[b]);
197         cmn_err(CE_DEBUG, ", fmt = %d\n", XLOG_FMT);
198         cmn_err(CE_DEBUG, "    log : uuid = ");
199         for (b = 0; b < 16; b++)
200                 cmn_err(CE_DEBUG, "%02x",((uchar_t *)&head->h_fs_uuid)[b]);
201         cmn_err(CE_DEBUG, ", fmt = %d\n", be32_to_cpu(head->h_fmt));
202 }
203 #else
204 #define xlog_header_check_dump(mp, head)
205 #endif
206
207 /*
208  * check log record header for recovery
209  */
210 STATIC int
211 xlog_header_check_recover(
212         xfs_mount_t             *mp,
213         xlog_rec_header_t       *head)
214 {
215         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
216
217         /*
218          * IRIX doesn't write the h_fmt field and leaves it zeroed
219          * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
220          * a dirty log created in IRIX.
221          */
222         if (unlikely(be32_to_cpu(head->h_fmt) != XLOG_FMT)) {
223                 xlog_warn(
224         "XFS: dirty log written in incompatible format - can't recover");
225                 xlog_header_check_dump(mp, head);
226                 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
227                                  XFS_ERRLEVEL_HIGH, mp);
228                 return XFS_ERROR(EFSCORRUPTED);
229         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
230                 xlog_warn(
231         "XFS: dirty log entry has mismatched uuid - can't recover");
232                 xlog_header_check_dump(mp, head);
233                 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
234                                  XFS_ERRLEVEL_HIGH, mp);
235                 return XFS_ERROR(EFSCORRUPTED);
236         }
237         return 0;
238 }
239
240 /*
241  * read the head block of the log and check the header
242  */
243 STATIC int
244 xlog_header_check_mount(
245         xfs_mount_t             *mp,
246         xlog_rec_header_t       *head)
247 {
248         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
249
250         if (uuid_is_nil(&head->h_fs_uuid)) {
251                 /*
252                  * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
253                  * h_fs_uuid is nil, we assume this log was last mounted
254                  * by IRIX and continue.
255                  */
256                 xlog_warn("XFS: nil uuid in log - IRIX style log");
257         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
258                 xlog_warn("XFS: log has mismatched uuid - can't recover");
259                 xlog_header_check_dump(mp, head);
260                 XFS_ERROR_REPORT("xlog_header_check_mount",
261                                  XFS_ERRLEVEL_HIGH, mp);
262                 return XFS_ERROR(EFSCORRUPTED);
263         }
264         return 0;
265 }
266
267 STATIC void
268 xlog_recover_iodone(
269         struct xfs_buf  *bp)
270 {
271         xfs_mount_t     *mp;
272
273         ASSERT(XFS_BUF_FSPRIVATE(bp, void *));
274
275         if (XFS_BUF_GETERROR(bp)) {
276                 /*
277                  * We're not going to bother about retrying
278                  * this during recovery. One strike!
279                  */
280                 mp = XFS_BUF_FSPRIVATE(bp, xfs_mount_t *);
281                 xfs_ioerror_alert("xlog_recover_iodone",
282                                   mp, bp, XFS_BUF_ADDR(bp));
283                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
284         }
285         XFS_BUF_SET_FSPRIVATE(bp, NULL);
286         XFS_BUF_CLR_IODONE_FUNC(bp);
287         xfs_biodone(bp);
288 }
289
290 /*
291  * This routine finds (to an approximation) the first block in the physical
292  * log which contains the given cycle.  It uses a binary search algorithm.
293  * Note that the algorithm can not be perfect because the disk will not
294  * necessarily be perfect.
295  */
296 STATIC int
297 xlog_find_cycle_start(
298         xlog_t          *log,
299         xfs_buf_t       *bp,
300         xfs_daddr_t     first_blk,
301         xfs_daddr_t     *last_blk,
302         uint            cycle)
303 {
304         xfs_caddr_t     offset;
305         xfs_daddr_t     mid_blk;
306         uint            mid_cycle;
307         int             error;
308
309         mid_blk = BLK_AVG(first_blk, *last_blk);
310         while (mid_blk != first_blk && mid_blk != *last_blk) {
311                 if ((error = xlog_bread(log, mid_blk, 1, bp)))
312                         return error;
313                 offset = xlog_align(log, mid_blk, 1, bp);
314                 mid_cycle = xlog_get_cycle(offset);
315                 if (mid_cycle == cycle) {
316                         *last_blk = mid_blk;
317                         /* last_half_cycle == mid_cycle */
318                 } else {
319                         first_blk = mid_blk;
320                         /* first_half_cycle == mid_cycle */
321                 }
322                 mid_blk = BLK_AVG(first_blk, *last_blk);
323         }
324         ASSERT((mid_blk == first_blk && mid_blk+1 == *last_blk) ||
325                (mid_blk == *last_blk && mid_blk-1 == first_blk));
326
327         return 0;
328 }
329
330 /*
331  * Check that the range of blocks does not contain the cycle number
332  * given.  The scan needs to occur from front to back and the ptr into the
333  * region must be updated since a later routine will need to perform another
334  * test.  If the region is completely good, we end up returning the same
335  * last block number.
336  *
337  * Set blkno to -1 if we encounter no errors.  This is an invalid block number
338  * since we don't ever expect logs to get this large.
339  */
340 STATIC int
341 xlog_find_verify_cycle(
342         xlog_t          *log,
343         xfs_daddr_t     start_blk,
344         int             nbblks,
345         uint            stop_on_cycle_no,
346         xfs_daddr_t     *new_blk)
347 {
348         xfs_daddr_t     i, j;
349         uint            cycle;
350         xfs_buf_t       *bp;
351         xfs_daddr_t     bufblks;
352         xfs_caddr_t     buf = NULL;
353         int             error = 0;
354
355         bufblks = 1 << ffs(nbblks);
356
357         while (!(bp = xlog_get_bp(log, bufblks))) {
358                 /* can't get enough memory to do everything in one big buffer */
359                 bufblks >>= 1;
360                 if (bufblks <= log->l_sectbb_log)
361                         return ENOMEM;
362         }
363
364         for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
365                 int     bcount;
366
367                 bcount = min(bufblks, (start_blk + nbblks - i));
368
369                 if ((error = xlog_bread(log, i, bcount, bp)))
370                         goto out;
371
372                 buf = xlog_align(log, i, bcount, bp);
373                 for (j = 0; j < bcount; j++) {
374                         cycle = xlog_get_cycle(buf);
375                         if (cycle == stop_on_cycle_no) {
376                                 *new_blk = i+j;
377                                 goto out;
378                         }
379
380                         buf += BBSIZE;
381                 }
382         }
383
384         *new_blk = -1;
385
386 out:
387         xlog_put_bp(bp);
388         return error;
389 }
390
391 /*
392  * Potentially backup over partial log record write.
393  *
394  * In the typical case, last_blk is the number of the block directly after
395  * a good log record.  Therefore, we subtract one to get the block number
396  * of the last block in the given buffer.  extra_bblks contains the number
397  * of blocks we would have read on a previous read.  This happens when the
398  * last log record is split over the end of the physical log.
399  *
400  * extra_bblks is the number of blocks potentially verified on a previous
401  * call to this routine.
402  */
403 STATIC int
404 xlog_find_verify_log_record(
405         xlog_t                  *log,
406         xfs_daddr_t             start_blk,
407         xfs_daddr_t             *last_blk,
408         int                     extra_bblks)
409 {
410         xfs_daddr_t             i;
411         xfs_buf_t               *bp;
412         xfs_caddr_t             offset = NULL;
413         xlog_rec_header_t       *head = NULL;
414         int                     error = 0;
415         int                     smallmem = 0;
416         int                     num_blks = *last_blk - start_blk;
417         int                     xhdrs;
418
419         ASSERT(start_blk != 0 || *last_blk != start_blk);
420
421         if (!(bp = xlog_get_bp(log, num_blks))) {
422                 if (!(bp = xlog_get_bp(log, 1)))
423                         return ENOMEM;
424                 smallmem = 1;
425         } else {
426                 if ((error = xlog_bread(log, start_blk, num_blks, bp)))
427                         goto out;
428                 offset = xlog_align(log, start_blk, num_blks, bp);
429                 offset += ((num_blks - 1) << BBSHIFT);
430         }
431
432         for (i = (*last_blk) - 1; i >= 0; i--) {
433                 if (i < start_blk) {
434                         /* valid log record not found */
435                         xlog_warn(
436                 "XFS: Log inconsistent (didn't find previous header)");
437                         ASSERT(0);
438                         error = XFS_ERROR(EIO);
439                         goto out;
440                 }
441
442                 if (smallmem) {
443                         if ((error = xlog_bread(log, i, 1, bp)))
444                                 goto out;
445                         offset = xlog_align(log, i, 1, bp);
446                 }
447
448                 head = (xlog_rec_header_t *)offset;
449
450                 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(head->h_magicno))
451                         break;
452
453                 if (!smallmem)
454                         offset -= BBSIZE;
455         }
456
457         /*
458          * We hit the beginning of the physical log & still no header.  Return
459          * to caller.  If caller can handle a return of -1, then this routine
460          * will be called again for the end of the physical log.
461          */
462         if (i == -1) {
463                 error = -1;
464                 goto out;
465         }
466
467         /*
468          * We have the final block of the good log (the first block
469          * of the log record _before_ the head. So we check the uuid.
470          */
471         if ((error = xlog_header_check_mount(log->l_mp, head)))
472                 goto out;
473
474         /*
475          * We may have found a log record header before we expected one.
476          * last_blk will be the 1st block # with a given cycle #.  We may end
477          * up reading an entire log record.  In this case, we don't want to
478          * reset last_blk.  Only when last_blk points in the middle of a log
479          * record do we update last_blk.
480          */
481         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
482                 uint    h_size = be32_to_cpu(head->h_size);
483
484                 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
485                 if (h_size % XLOG_HEADER_CYCLE_SIZE)
486                         xhdrs++;
487         } else {
488                 xhdrs = 1;
489         }
490
491         if (*last_blk - i + extra_bblks !=
492             BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
493                 *last_blk = i;
494
495 out:
496         xlog_put_bp(bp);
497         return error;
498 }
499
500 /*
501  * Head is defined to be the point of the log where the next log write
502  * write could go.  This means that incomplete LR writes at the end are
503  * eliminated when calculating the head.  We aren't guaranteed that previous
504  * LR have complete transactions.  We only know that a cycle number of
505  * current cycle number -1 won't be present in the log if we start writing
506  * from our current block number.
507  *
508  * last_blk contains the block number of the first block with a given
509  * cycle number.
510  *
511  * Return: zero if normal, non-zero if error.
512  */
513 STATIC int
514 xlog_find_head(
515         xlog_t          *log,
516         xfs_daddr_t     *return_head_blk)
517 {
518         xfs_buf_t       *bp;
519         xfs_caddr_t     offset;
520         xfs_daddr_t     new_blk, first_blk, start_blk, last_blk, head_blk;
521         int             num_scan_bblks;
522         uint            first_half_cycle, last_half_cycle;
523         uint            stop_on_cycle;
524         int             error, log_bbnum = log->l_logBBsize;
525
526         /* Is the end of the log device zeroed? */
527         if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
528                 *return_head_blk = first_blk;
529
530                 /* Is the whole lot zeroed? */
531                 if (!first_blk) {
532                         /* Linux XFS shouldn't generate totally zeroed logs -
533                          * mkfs etc write a dummy unmount record to a fresh
534                          * log so we can store the uuid in there
535                          */
536                         xlog_warn("XFS: totally zeroed log");
537                 }
538
539                 return 0;
540         } else if (error) {
541                 xlog_warn("XFS: empty log check failed");
542                 return error;
543         }
544
545         first_blk = 0;                  /* get cycle # of 1st block */
546         bp = xlog_get_bp(log, 1);
547         if (!bp)
548                 return ENOMEM;
549         if ((error = xlog_bread(log, 0, 1, bp)))
550                 goto bp_err;
551         offset = xlog_align(log, 0, 1, bp);
552         first_half_cycle = xlog_get_cycle(offset);
553
554         last_blk = head_blk = log_bbnum - 1;    /* get cycle # of last block */
555         if ((error = xlog_bread(log, last_blk, 1, bp)))
556                 goto bp_err;
557         offset = xlog_align(log, last_blk, 1, bp);
558         last_half_cycle = xlog_get_cycle(offset);
559         ASSERT(last_half_cycle != 0);
560
561         /*
562          * If the 1st half cycle number is equal to the last half cycle number,
563          * then the entire log is stamped with the same cycle number.  In this
564          * case, head_blk can't be set to zero (which makes sense).  The below
565          * math doesn't work out properly with head_blk equal to zero.  Instead,
566          * we set it to log_bbnum which is an invalid block number, but this
567          * value makes the math correct.  If head_blk doesn't changed through
568          * all the tests below, *head_blk is set to zero at the very end rather
569          * than log_bbnum.  In a sense, log_bbnum and zero are the same block
570          * in a circular file.
571          */
572         if (first_half_cycle == last_half_cycle) {
573                 /*
574                  * In this case we believe that the entire log should have
575                  * cycle number last_half_cycle.  We need to scan backwards
576                  * from the end verifying that there are no holes still
577                  * containing last_half_cycle - 1.  If we find such a hole,
578                  * then the start of that hole will be the new head.  The
579                  * simple case looks like
580                  *        x | x ... | x - 1 | x
581                  * Another case that fits this picture would be
582                  *        x | x + 1 | x ... | x
583                  * In this case the head really is somewhere at the end of the
584                  * log, as one of the latest writes at the beginning was
585                  * incomplete.
586                  * One more case is
587                  *        x | x + 1 | x ... | x - 1 | x
588                  * This is really the combination of the above two cases, and
589                  * the head has to end up at the start of the x-1 hole at the
590                  * end of the log.
591                  *
592                  * In the 256k log case, we will read from the beginning to the
593                  * end of the log and search for cycle numbers equal to x-1.
594                  * We don't worry about the x+1 blocks that we encounter,
595                  * because we know that they cannot be the head since the log
596                  * started with x.
597                  */
598                 head_blk = log_bbnum;
599                 stop_on_cycle = last_half_cycle - 1;
600         } else {
601                 /*
602                  * In this case we want to find the first block with cycle
603                  * number matching last_half_cycle.  We expect the log to be
604                  * some variation on
605                  *        x + 1 ... | x ...
606                  * The first block with cycle number x (last_half_cycle) will
607                  * be where the new head belongs.  First we do a binary search
608                  * for the first occurrence of last_half_cycle.  The binary
609                  * search may not be totally accurate, so then we scan back
610                  * from there looking for occurrences of last_half_cycle before
611                  * us.  If that backwards scan wraps around the beginning of
612                  * the log, then we look for occurrences of last_half_cycle - 1
613                  * at the end of the log.  The cases we're looking for look
614                  * like
615                  *        x + 1 ... | x | x + 1 | x ...
616                  *                               ^ binary search stopped here
617                  * or
618                  *        x + 1 ... | x ... | x - 1 | x
619                  *        <---------> less than scan distance
620                  */
621                 stop_on_cycle = last_half_cycle;
622                 if ((error = xlog_find_cycle_start(log, bp, first_blk,
623                                                 &head_blk, last_half_cycle)))
624                         goto bp_err;
625         }
626
627         /*
628          * Now validate the answer.  Scan back some number of maximum possible
629          * blocks and make sure each one has the expected cycle number.  The
630          * maximum is determined by the total possible amount of buffering
631          * in the in-core log.  The following number can be made tighter if
632          * we actually look at the block size of the filesystem.
633          */
634         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
635         if (head_blk >= num_scan_bblks) {
636                 /*
637                  * We are guaranteed that the entire check can be performed
638                  * in one buffer.
639                  */
640                 start_blk = head_blk - num_scan_bblks;
641                 if ((error = xlog_find_verify_cycle(log,
642                                                 start_blk, num_scan_bblks,
643                                                 stop_on_cycle, &new_blk)))
644                         goto bp_err;
645                 if (new_blk != -1)
646                         head_blk = new_blk;
647         } else {                /* need to read 2 parts of log */
648                 /*
649                  * We are going to scan backwards in the log in two parts.
650                  * First we scan the physical end of the log.  In this part
651                  * of the log, we are looking for blocks with cycle number
652                  * last_half_cycle - 1.
653                  * If we find one, then we know that the log starts there, as
654                  * we've found a hole that didn't get written in going around
655                  * the end of the physical log.  The simple case for this is
656                  *        x + 1 ... | x ... | x - 1 | x
657                  *        <---------> less than scan distance
658                  * If all of the blocks at the end of the log have cycle number
659                  * last_half_cycle, then we check the blocks at the start of
660                  * the log looking for occurrences of last_half_cycle.  If we
661                  * find one, then our current estimate for the location of the
662                  * first occurrence of last_half_cycle is wrong and we move
663                  * back to the hole we've found.  This case looks like
664                  *        x + 1 ... | x | x + 1 | x ...
665                  *                               ^ binary search stopped here
666                  * Another case we need to handle that only occurs in 256k
667                  * logs is
668                  *        x + 1 ... | x ... | x+1 | x ...
669                  *                   ^ binary search stops here
670                  * In a 256k log, the scan at the end of the log will see the
671                  * x + 1 blocks.  We need to skip past those since that is
672                  * certainly not the head of the log.  By searching for
673                  * last_half_cycle-1 we accomplish that.
674                  */
675                 start_blk = log_bbnum - num_scan_bblks + head_blk;
676                 ASSERT(head_blk <= INT_MAX &&
677                         (xfs_daddr_t) num_scan_bblks - head_blk >= 0);
678                 if ((error = xlog_find_verify_cycle(log, start_blk,
679                                         num_scan_bblks - (int)head_blk,
680                                         (stop_on_cycle - 1), &new_blk)))
681                         goto bp_err;
682                 if (new_blk != -1) {
683                         head_blk = new_blk;
684                         goto bad_blk;
685                 }
686
687                 /*
688                  * Scan beginning of log now.  The last part of the physical
689                  * log is good.  This scan needs to verify that it doesn't find
690                  * the last_half_cycle.
691                  */
692                 start_blk = 0;
693                 ASSERT(head_blk <= INT_MAX);
694                 if ((error = xlog_find_verify_cycle(log,
695                                         start_blk, (int)head_blk,
696                                         stop_on_cycle, &new_blk)))
697                         goto bp_err;
698                 if (new_blk != -1)
699                         head_blk = new_blk;
700         }
701
702  bad_blk:
703         /*
704          * Now we need to make sure head_blk is not pointing to a block in
705          * the middle of a log record.
706          */
707         num_scan_bblks = XLOG_REC_SHIFT(log);
708         if (head_blk >= num_scan_bblks) {
709                 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
710
711                 /* start ptr at last block ptr before head_blk */
712                 if ((error = xlog_find_verify_log_record(log, start_blk,
713                                                         &head_blk, 0)) == -1) {
714                         error = XFS_ERROR(EIO);
715                         goto bp_err;
716                 } else if (error)
717                         goto bp_err;
718         } else {
719                 start_blk = 0;
720                 ASSERT(head_blk <= INT_MAX);
721                 if ((error = xlog_find_verify_log_record(log, start_blk,
722                                                         &head_blk, 0)) == -1) {
723                         /* We hit the beginning of the log during our search */
724                         start_blk = log_bbnum - num_scan_bblks + head_blk;
725                         new_blk = log_bbnum;
726                         ASSERT(start_blk <= INT_MAX &&
727                                 (xfs_daddr_t) log_bbnum-start_blk >= 0);
728                         ASSERT(head_blk <= INT_MAX);
729                         if ((error = xlog_find_verify_log_record(log,
730                                                         start_blk, &new_blk,
731                                                         (int)head_blk)) == -1) {
732                                 error = XFS_ERROR(EIO);
733                                 goto bp_err;
734                         } else if (error)
735                                 goto bp_err;
736                         if (new_blk != log_bbnum)
737                                 head_blk = new_blk;
738                 } else if (error)
739                         goto bp_err;
740         }
741
742         xlog_put_bp(bp);
743         if (head_blk == log_bbnum)
744                 *return_head_blk = 0;
745         else
746                 *return_head_blk = head_blk;
747         /*
748          * When returning here, we have a good block number.  Bad block
749          * means that during a previous crash, we didn't have a clean break
750          * from cycle number N to cycle number N-1.  In this case, we need
751          * to find the first block with cycle number N-1.
752          */
753         return 0;
754
755  bp_err:
756         xlog_put_bp(bp);
757
758         if (error)
759             xlog_warn("XFS: failed to find log head");
760         return error;
761 }
762
763 /*
764  * Find the sync block number or the tail of the log.
765  *
766  * This will be the block number of the last record to have its
767  * associated buffers synced to disk.  Every log record header has
768  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
769  * to get a sync block number.  The only concern is to figure out which
770  * log record header to believe.
771  *
772  * The following algorithm uses the log record header with the largest
773  * lsn.  The entire log record does not need to be valid.  We only care
774  * that the header is valid.
775  *
776  * We could speed up search by using current head_blk buffer, but it is not
777  * available.
778  */
779 int
780 xlog_find_tail(
781         xlog_t                  *log,
782         xfs_daddr_t             *head_blk,
783         xfs_daddr_t             *tail_blk)
784 {
785         xlog_rec_header_t       *rhead;
786         xlog_op_header_t        *op_head;
787         xfs_caddr_t             offset = NULL;
788         xfs_buf_t               *bp;
789         int                     error, i, found;
790         xfs_daddr_t             umount_data_blk;
791         xfs_daddr_t             after_umount_blk;
792         xfs_lsn_t               tail_lsn;
793         int                     hblks;
794
795         found = 0;
796
797         /*
798          * Find previous log record
799          */
800         if ((error = xlog_find_head(log, head_blk)))
801                 return error;
802
803         bp = xlog_get_bp(log, 1);
804         if (!bp)
805                 return ENOMEM;
806         if (*head_blk == 0) {                           /* special case */
807                 if ((error = xlog_bread(log, 0, 1, bp)))
808                         goto bread_err;
809                 offset = xlog_align(log, 0, 1, bp);
810                 if (xlog_get_cycle(offset) == 0) {
811                         *tail_blk = 0;
812                         /* leave all other log inited values alone */
813                         goto exit;
814                 }
815         }
816
817         /*
818          * Search backwards looking for log record header block
819          */
820         ASSERT(*head_blk < INT_MAX);
821         for (i = (int)(*head_blk) - 1; i >= 0; i--) {
822                 if ((error = xlog_bread(log, i, 1, bp)))
823                         goto bread_err;
824                 offset = xlog_align(log, i, 1, bp);
825                 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(*(__be32 *)offset)) {
826                         found = 1;
827                         break;
828                 }
829         }
830         /*
831          * If we haven't found the log record header block, start looking
832          * again from the end of the physical log.  XXXmiken: There should be
833          * a check here to make sure we didn't search more than N blocks in
834          * the previous code.
835          */
836         if (!found) {
837                 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
838                         if ((error = xlog_bread(log, i, 1, bp)))
839                                 goto bread_err;
840                         offset = xlog_align(log, i, 1, bp);
841                         if (XLOG_HEADER_MAGIC_NUM ==
842                             be32_to_cpu(*(__be32 *)offset)) {
843                                 found = 2;
844                                 break;
845                         }
846                 }
847         }
848         if (!found) {
849                 xlog_warn("XFS: xlog_find_tail: couldn't find sync record");
850                 ASSERT(0);
851                 return XFS_ERROR(EIO);
852         }
853
854         /* find blk_no of tail of log */
855         rhead = (xlog_rec_header_t *)offset;
856         *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
857
858         /*
859          * Reset log values according to the state of the log when we
860          * crashed.  In the case where head_blk == 0, we bump curr_cycle
861          * one because the next write starts a new cycle rather than
862          * continuing the cycle of the last good log record.  At this
863          * point we have guaranteed that all partial log records have been
864          * accounted for.  Therefore, we know that the last good log record
865          * written was complete and ended exactly on the end boundary
866          * of the physical log.
867          */
868         log->l_prev_block = i;
869         log->l_curr_block = (int)*head_blk;
870         log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
871         if (found == 2)
872                 log->l_curr_cycle++;
873         log->l_tail_lsn = be64_to_cpu(rhead->h_tail_lsn);
874         log->l_last_sync_lsn = be64_to_cpu(rhead->h_lsn);
875         log->l_grant_reserve_cycle = log->l_curr_cycle;
876         log->l_grant_reserve_bytes = BBTOB(log->l_curr_block);
877         log->l_grant_write_cycle = log->l_curr_cycle;
878         log->l_grant_write_bytes = BBTOB(log->l_curr_block);
879
880         /*
881          * Look for unmount record.  If we find it, then we know there
882          * was a clean unmount.  Since 'i' could be the last block in
883          * the physical log, we convert to a log block before comparing
884          * to the head_blk.
885          *
886          * Save the current tail lsn to use to pass to
887          * xlog_clear_stale_blocks() below.  We won't want to clear the
888          * unmount record if there is one, so we pass the lsn of the
889          * unmount record rather than the block after it.
890          */
891         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
892                 int     h_size = be32_to_cpu(rhead->h_size);
893                 int     h_version = be32_to_cpu(rhead->h_version);
894
895                 if ((h_version & XLOG_VERSION_2) &&
896                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
897                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
898                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
899                                 hblks++;
900                 } else {
901                         hblks = 1;
902                 }
903         } else {
904                 hblks = 1;
905         }
906         after_umount_blk = (i + hblks + (int)
907                 BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
908         tail_lsn = log->l_tail_lsn;
909         if (*head_blk == after_umount_blk &&
910             be32_to_cpu(rhead->h_num_logops) == 1) {
911                 umount_data_blk = (i + hblks) % log->l_logBBsize;
912                 if ((error = xlog_bread(log, umount_data_blk, 1, bp))) {
913                         goto bread_err;
914                 }
915                 offset = xlog_align(log, umount_data_blk, 1, bp);
916                 op_head = (xlog_op_header_t *)offset;
917                 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
918                         /*
919                          * Set tail and last sync so that newly written
920                          * log records will point recovery to after the
921                          * current unmount record.
922                          */
923                         log->l_tail_lsn =
924                                 xlog_assign_lsn(log->l_curr_cycle,
925                                                 after_umount_blk);
926                         log->l_last_sync_lsn =
927                                 xlog_assign_lsn(log->l_curr_cycle,
928                                                 after_umount_blk);
929                         *tail_blk = after_umount_blk;
930
931                         /*
932                          * Note that the unmount was clean. If the unmount
933                          * was not clean, we need to know this to rebuild the
934                          * superblock counters from the perag headers if we
935                          * have a filesystem using non-persistent counters.
936                          */
937                         log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
938                 }
939         }
940
941         /*
942          * Make sure that there are no blocks in front of the head
943          * with the same cycle number as the head.  This can happen
944          * because we allow multiple outstanding log writes concurrently,
945          * and the later writes might make it out before earlier ones.
946          *
947          * We use the lsn from before modifying it so that we'll never
948          * overwrite the unmount record after a clean unmount.
949          *
950          * Do this only if we are going to recover the filesystem
951          *
952          * NOTE: This used to say "if (!readonly)"
953          * However on Linux, we can & do recover a read-only filesystem.
954          * We only skip recovery if NORECOVERY is specified on mount,
955          * in which case we would not be here.
956          *
957          * But... if the -device- itself is readonly, just skip this.
958          * We can't recover this device anyway, so it won't matter.
959          */
960         if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
961                 error = xlog_clear_stale_blocks(log, tail_lsn);
962         }
963
964 bread_err:
965 exit:
966         xlog_put_bp(bp);
967
968         if (error)
969                 xlog_warn("XFS: failed to locate log tail");
970         return error;
971 }
972
973 /*
974  * Is the log zeroed at all?
975  *
976  * The last binary search should be changed to perform an X block read
977  * once X becomes small enough.  You can then search linearly through
978  * the X blocks.  This will cut down on the number of reads we need to do.
979  *
980  * If the log is partially zeroed, this routine will pass back the blkno
981  * of the first block with cycle number 0.  It won't have a complete LR
982  * preceding it.
983  *
984  * Return:
985  *      0  => the log is completely written to
986  *      -1 => use *blk_no as the first block of the log
987  *      >0 => error has occurred
988  */
989 STATIC int
990 xlog_find_zeroed(
991         xlog_t          *log,
992         xfs_daddr_t     *blk_no)
993 {
994         xfs_buf_t       *bp;
995         xfs_caddr_t     offset;
996         uint            first_cycle, last_cycle;
997         xfs_daddr_t     new_blk, last_blk, start_blk;
998         xfs_daddr_t     num_scan_bblks;
999         int             error, log_bbnum = log->l_logBBsize;
1000
1001         *blk_no = 0;
1002
1003         /* check totally zeroed log */
1004         bp = xlog_get_bp(log, 1);
1005         if (!bp)
1006                 return ENOMEM;
1007         if ((error = xlog_bread(log, 0, 1, bp)))
1008                 goto bp_err;
1009         offset = xlog_align(log, 0, 1, bp);
1010         first_cycle = xlog_get_cycle(offset);
1011         if (first_cycle == 0) {         /* completely zeroed log */
1012                 *blk_no = 0;
1013                 xlog_put_bp(bp);
1014                 return -1;
1015         }
1016
1017         /* check partially zeroed log */
1018         if ((error = xlog_bread(log, log_bbnum-1, 1, bp)))
1019                 goto bp_err;
1020         offset = xlog_align(log, log_bbnum-1, 1, bp);
1021         last_cycle = xlog_get_cycle(offset);
1022         if (last_cycle != 0) {          /* log completely written to */
1023                 xlog_put_bp(bp);
1024                 return 0;
1025         } else if (first_cycle != 1) {
1026                 /*
1027                  * If the cycle of the last block is zero, the cycle of
1028                  * the first block must be 1. If it's not, maybe we're
1029                  * not looking at a log... Bail out.
1030                  */
1031                 xlog_warn("XFS: Log inconsistent or not a log (last==0, first!=1)");
1032                 return XFS_ERROR(EINVAL);
1033         }
1034
1035         /* we have a partially zeroed log */
1036         last_blk = log_bbnum-1;
1037         if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1038                 goto bp_err;
1039
1040         /*
1041          * Validate the answer.  Because there is no way to guarantee that
1042          * the entire log is made up of log records which are the same size,
1043          * we scan over the defined maximum blocks.  At this point, the maximum
1044          * is not chosen to mean anything special.   XXXmiken
1045          */
1046         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1047         ASSERT(num_scan_bblks <= INT_MAX);
1048
1049         if (last_blk < num_scan_bblks)
1050                 num_scan_bblks = last_blk;
1051         start_blk = last_blk - num_scan_bblks;
1052
1053         /*
1054          * We search for any instances of cycle number 0 that occur before
1055          * our current estimate of the head.  What we're trying to detect is
1056          *        1 ... | 0 | 1 | 0...
1057          *                       ^ binary search ends here
1058          */
1059         if ((error = xlog_find_verify_cycle(log, start_blk,
1060                                          (int)num_scan_bblks, 0, &new_blk)))
1061                 goto bp_err;
1062         if (new_blk != -1)
1063                 last_blk = new_blk;
1064
1065         /*
1066          * Potentially backup over partial log record write.  We don't need
1067          * to search the end of the log because we know it is zero.
1068          */
1069         if ((error = xlog_find_verify_log_record(log, start_blk,
1070                                 &last_blk, 0)) == -1) {
1071             error = XFS_ERROR(EIO);
1072             goto bp_err;
1073         } else if (error)
1074             goto bp_err;
1075
1076         *blk_no = last_blk;
1077 bp_err:
1078         xlog_put_bp(bp);
1079         if (error)
1080                 return error;
1081         return -1;
1082 }
1083
1084 /*
1085  * These are simple subroutines used by xlog_clear_stale_blocks() below
1086  * to initialize a buffer full of empty log record headers and write
1087  * them into the log.
1088  */
1089 STATIC void
1090 xlog_add_record(
1091         xlog_t                  *log,
1092         xfs_caddr_t             buf,
1093         int                     cycle,
1094         int                     block,
1095         int                     tail_cycle,
1096         int                     tail_block)
1097 {
1098         xlog_rec_header_t       *recp = (xlog_rec_header_t *)buf;
1099
1100         memset(buf, 0, BBSIZE);
1101         recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1102         recp->h_cycle = cpu_to_be32(cycle);
1103         recp->h_version = cpu_to_be32(
1104                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1105         recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1106         recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1107         recp->h_fmt = cpu_to_be32(XLOG_FMT);
1108         memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1109 }
1110
1111 STATIC int
1112 xlog_write_log_records(
1113         xlog_t          *log,
1114         int             cycle,
1115         int             start_block,
1116         int             blocks,
1117         int             tail_cycle,
1118         int             tail_block)
1119 {
1120         xfs_caddr_t     offset;
1121         xfs_buf_t       *bp;
1122         int             balign, ealign;
1123         int             sectbb = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
1124         int             end_block = start_block + blocks;
1125         int             bufblks;
1126         int             error = 0;
1127         int             i, j = 0;
1128
1129         bufblks = 1 << ffs(blocks);
1130         while (!(bp = xlog_get_bp(log, bufblks))) {
1131                 bufblks >>= 1;
1132                 if (bufblks <= log->l_sectbb_log)
1133                         return ENOMEM;
1134         }
1135
1136         /* We may need to do a read at the start to fill in part of
1137          * the buffer in the starting sector not covered by the first
1138          * write below.
1139          */
1140         balign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, start_block);
1141         if (balign != start_block) {
1142                 if ((error = xlog_bread(log, start_block, 1, bp))) {
1143                         xlog_put_bp(bp);
1144                         return error;
1145                 }
1146                 j = start_block - balign;
1147         }
1148
1149         for (i = start_block; i < end_block; i += bufblks) {
1150                 int             bcount, endcount;
1151
1152                 bcount = min(bufblks, end_block - start_block);
1153                 endcount = bcount - j;
1154
1155                 /* We may need to do a read at the end to fill in part of
1156                  * the buffer in the final sector not covered by the write.
1157                  * If this is the same sector as the above read, skip it.
1158                  */
1159                 ealign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, end_block);
1160                 if (j == 0 && (start_block + endcount > ealign)) {
1161                         offset = XFS_BUF_PTR(bp);
1162                         balign = BBTOB(ealign - start_block);
1163                         error = XFS_BUF_SET_PTR(bp, offset + balign,
1164                                                 BBTOB(sectbb));
1165                         if (!error)
1166                                 error = xlog_bread(log, ealign, sectbb, bp);
1167                         if (!error)
1168                                 error = XFS_BUF_SET_PTR(bp, offset, bufblks);
1169                         if (error)
1170                                 break;
1171                 }
1172
1173                 offset = xlog_align(log, start_block, endcount, bp);
1174                 for (; j < endcount; j++) {
1175                         xlog_add_record(log, offset, cycle, i+j,
1176                                         tail_cycle, tail_block);
1177                         offset += BBSIZE;
1178                 }
1179                 error = xlog_bwrite(log, start_block, endcount, bp);
1180                 if (error)
1181                         break;
1182                 start_block += endcount;
1183                 j = 0;
1184         }
1185         xlog_put_bp(bp);
1186         return error;
1187 }
1188
1189 /*
1190  * This routine is called to blow away any incomplete log writes out
1191  * in front of the log head.  We do this so that we won't become confused
1192  * if we come up, write only a little bit more, and then crash again.
1193  * If we leave the partial log records out there, this situation could
1194  * cause us to think those partial writes are valid blocks since they
1195  * have the current cycle number.  We get rid of them by overwriting them
1196  * with empty log records with the old cycle number rather than the
1197  * current one.
1198  *
1199  * The tail lsn is passed in rather than taken from
1200  * the log so that we will not write over the unmount record after a
1201  * clean unmount in a 512 block log.  Doing so would leave the log without
1202  * any valid log records in it until a new one was written.  If we crashed
1203  * during that time we would not be able to recover.
1204  */
1205 STATIC int
1206 xlog_clear_stale_blocks(
1207         xlog_t          *log,
1208         xfs_lsn_t       tail_lsn)
1209 {
1210         int             tail_cycle, head_cycle;
1211         int             tail_block, head_block;
1212         int             tail_distance, max_distance;
1213         int             distance;
1214         int             error;
1215
1216         tail_cycle = CYCLE_LSN(tail_lsn);
1217         tail_block = BLOCK_LSN(tail_lsn);
1218         head_cycle = log->l_curr_cycle;
1219         head_block = log->l_curr_block;
1220
1221         /*
1222          * Figure out the distance between the new head of the log
1223          * and the tail.  We want to write over any blocks beyond the
1224          * head that we may have written just before the crash, but
1225          * we don't want to overwrite the tail of the log.
1226          */
1227         if (head_cycle == tail_cycle) {
1228                 /*
1229                  * The tail is behind the head in the physical log,
1230                  * so the distance from the head to the tail is the
1231                  * distance from the head to the end of the log plus
1232                  * the distance from the beginning of the log to the
1233                  * tail.
1234                  */
1235                 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1236                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1237                                          XFS_ERRLEVEL_LOW, log->l_mp);
1238                         return XFS_ERROR(EFSCORRUPTED);
1239                 }
1240                 tail_distance = tail_block + (log->l_logBBsize - head_block);
1241         } else {
1242                 /*
1243                  * The head is behind the tail in the physical log,
1244                  * so the distance from the head to the tail is just
1245                  * the tail block minus the head block.
1246                  */
1247                 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1248                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1249                                          XFS_ERRLEVEL_LOW, log->l_mp);
1250                         return XFS_ERROR(EFSCORRUPTED);
1251                 }
1252                 tail_distance = tail_block - head_block;
1253         }
1254
1255         /*
1256          * If the head is right up against the tail, we can't clear
1257          * anything.
1258          */
1259         if (tail_distance <= 0) {
1260                 ASSERT(tail_distance == 0);
1261                 return 0;
1262         }
1263
1264         max_distance = XLOG_TOTAL_REC_SHIFT(log);
1265         /*
1266          * Take the smaller of the maximum amount of outstanding I/O
1267          * we could have and the distance to the tail to clear out.
1268          * We take the smaller so that we don't overwrite the tail and
1269          * we don't waste all day writing from the head to the tail
1270          * for no reason.
1271          */
1272         max_distance = MIN(max_distance, tail_distance);
1273
1274         if ((head_block + max_distance) <= log->l_logBBsize) {
1275                 /*
1276                  * We can stomp all the blocks we need to without
1277                  * wrapping around the end of the log.  Just do it
1278                  * in a single write.  Use the cycle number of the
1279                  * current cycle minus one so that the log will look like:
1280                  *     n ... | n - 1 ...
1281                  */
1282                 error = xlog_write_log_records(log, (head_cycle - 1),
1283                                 head_block, max_distance, tail_cycle,
1284                                 tail_block);
1285                 if (error)
1286                         return error;
1287         } else {
1288                 /*
1289                  * We need to wrap around the end of the physical log in
1290                  * order to clear all the blocks.  Do it in two separate
1291                  * I/Os.  The first write should be from the head to the
1292                  * end of the physical log, and it should use the current
1293                  * cycle number minus one just like above.
1294                  */
1295                 distance = log->l_logBBsize - head_block;
1296                 error = xlog_write_log_records(log, (head_cycle - 1),
1297                                 head_block, distance, tail_cycle,
1298                                 tail_block);
1299
1300                 if (error)
1301                         return error;
1302
1303                 /*
1304                  * Now write the blocks at the start of the physical log.
1305                  * This writes the remainder of the blocks we want to clear.
1306                  * It uses the current cycle number since we're now on the
1307                  * same cycle as the head so that we get:
1308                  *    n ... n ... | n - 1 ...
1309                  *    ^^^^^ blocks we're writing
1310                  */
1311                 distance = max_distance - (log->l_logBBsize - head_block);
1312                 error = xlog_write_log_records(log, head_cycle, 0, distance,
1313                                 tail_cycle, tail_block);
1314                 if (error)
1315                         return error;
1316         }
1317
1318         return 0;
1319 }
1320
1321 /******************************************************************************
1322  *
1323  *              Log recover routines
1324  *
1325  ******************************************************************************
1326  */
1327
1328 STATIC xlog_recover_t *
1329 xlog_recover_find_tid(
1330         xlog_recover_t          *q,
1331         xlog_tid_t              tid)
1332 {
1333         xlog_recover_t          *p = q;
1334
1335         while (p != NULL) {
1336                 if (p->r_log_tid == tid)
1337                     break;
1338                 p = p->r_next;
1339         }
1340         return p;
1341 }
1342
1343 STATIC void
1344 xlog_recover_put_hashq(
1345         xlog_recover_t          **q,
1346         xlog_recover_t          *trans)
1347 {
1348         trans->r_next = *q;
1349         *q = trans;
1350 }
1351
1352 STATIC void
1353 xlog_recover_add_item(
1354         xlog_recover_item_t     **itemq)
1355 {
1356         xlog_recover_item_t     *item;
1357
1358         item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1359         xlog_recover_insert_item_backq(itemq, item);
1360 }
1361
1362 STATIC int
1363 xlog_recover_add_to_cont_trans(
1364         xlog_recover_t          *trans,
1365         xfs_caddr_t             dp,
1366         int                     len)
1367 {
1368         xlog_recover_item_t     *item;
1369         xfs_caddr_t             ptr, old_ptr;
1370         int                     old_len;
1371
1372         item = trans->r_itemq;
1373         if (item == NULL) {
1374                 /* finish copying rest of trans header */
1375                 xlog_recover_add_item(&trans->r_itemq);
1376                 ptr = (xfs_caddr_t) &trans->r_theader +
1377                                 sizeof(xfs_trans_header_t) - len;
1378                 memcpy(ptr, dp, len); /* d, s, l */
1379                 return 0;
1380         }
1381         item = item->ri_prev;
1382
1383         old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1384         old_len = item->ri_buf[item->ri_cnt-1].i_len;
1385
1386         ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1387         memcpy(&ptr[old_len], dp, len); /* d, s, l */
1388         item->ri_buf[item->ri_cnt-1].i_len += len;
1389         item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1390         return 0;
1391 }
1392
1393 /*
1394  * The next region to add is the start of a new region.  It could be
1395  * a whole region or it could be the first part of a new region.  Because
1396  * of this, the assumption here is that the type and size fields of all
1397  * format structures fit into the first 32 bits of the structure.
1398  *
1399  * This works because all regions must be 32 bit aligned.  Therefore, we
1400  * either have both fields or we have neither field.  In the case we have
1401  * neither field, the data part of the region is zero length.  We only have
1402  * a log_op_header and can throw away the header since a new one will appear
1403  * later.  If we have at least 4 bytes, then we can determine how many regions
1404  * will appear in the current log item.
1405  */
1406 STATIC int
1407 xlog_recover_add_to_trans(
1408         xlog_recover_t          *trans,
1409         xfs_caddr_t             dp,
1410         int                     len)
1411 {
1412         xfs_inode_log_format_t  *in_f;                  /* any will do */
1413         xlog_recover_item_t     *item;
1414         xfs_caddr_t             ptr;
1415
1416         if (!len)
1417                 return 0;
1418         item = trans->r_itemq;
1419         if (item == NULL) {
1420                 /* we need to catch log corruptions here */
1421                 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1422                         xlog_warn("XFS: xlog_recover_add_to_trans: "
1423                                   "bad header magic number");
1424                         ASSERT(0);
1425                         return XFS_ERROR(EIO);
1426                 }
1427                 if (len == sizeof(xfs_trans_header_t))
1428                         xlog_recover_add_item(&trans->r_itemq);
1429                 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1430                 return 0;
1431         }
1432
1433         ptr = kmem_alloc(len, KM_SLEEP);
1434         memcpy(ptr, dp, len);
1435         in_f = (xfs_inode_log_format_t *)ptr;
1436
1437         if (item->ri_prev->ri_total != 0 &&
1438              item->ri_prev->ri_total == item->ri_prev->ri_cnt) {
1439                 xlog_recover_add_item(&trans->r_itemq);
1440         }
1441         item = trans->r_itemq;
1442         item = item->ri_prev;
1443
1444         if (item->ri_total == 0) {              /* first region to be added */
1445                 item->ri_total  = in_f->ilf_size;
1446                 ASSERT(item->ri_total <= XLOG_MAX_REGIONS_IN_ITEM);
1447                 item->ri_buf = kmem_zalloc((item->ri_total *
1448                                             sizeof(xfs_log_iovec_t)), KM_SLEEP);
1449         }
1450         ASSERT(item->ri_total > item->ri_cnt);
1451         /* Description region is ri_buf[0] */
1452         item->ri_buf[item->ri_cnt].i_addr = ptr;
1453         item->ri_buf[item->ri_cnt].i_len  = len;
1454         item->ri_cnt++;
1455         return 0;
1456 }
1457
1458 STATIC void
1459 xlog_recover_new_tid(
1460         xlog_recover_t          **q,
1461         xlog_tid_t              tid,
1462         xfs_lsn_t               lsn)
1463 {
1464         xlog_recover_t          *trans;
1465
1466         trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1467         trans->r_log_tid   = tid;
1468         trans->r_lsn       = lsn;
1469         xlog_recover_put_hashq(q, trans);
1470 }
1471
1472 STATIC int
1473 xlog_recover_unlink_tid(
1474         xlog_recover_t          **q,
1475         xlog_recover_t          *trans)
1476 {
1477         xlog_recover_t          *tp;
1478         int                     found = 0;
1479
1480         ASSERT(trans != NULL);
1481         if (trans == *q) {
1482                 *q = (*q)->r_next;
1483         } else {
1484                 tp = *q;
1485                 while (tp) {
1486                         if (tp->r_next == trans) {
1487                                 found = 1;
1488                                 break;
1489                         }
1490                         tp = tp->r_next;
1491                 }
1492                 if (!found) {
1493                         xlog_warn(
1494                              "XFS: xlog_recover_unlink_tid: trans not found");
1495                         ASSERT(0);
1496                         return XFS_ERROR(EIO);
1497                 }
1498                 tp->r_next = tp->r_next->r_next;
1499         }
1500         return 0;
1501 }
1502
1503 STATIC void
1504 xlog_recover_insert_item_backq(
1505         xlog_recover_item_t     **q,
1506         xlog_recover_item_t     *item)
1507 {
1508         if (*q == NULL) {
1509                 item->ri_prev = item->ri_next = item;
1510                 *q = item;
1511         } else {
1512                 item->ri_next           = *q;
1513                 item->ri_prev           = (*q)->ri_prev;
1514                 (*q)->ri_prev           = item;
1515                 item->ri_prev->ri_next  = item;
1516         }
1517 }
1518
1519 STATIC void
1520 xlog_recover_insert_item_frontq(
1521         xlog_recover_item_t     **q,
1522         xlog_recover_item_t     *item)
1523 {
1524         xlog_recover_insert_item_backq(q, item);
1525         *q = item;
1526 }
1527
1528 STATIC int
1529 xlog_recover_reorder_trans(
1530         xlog_recover_t          *trans)
1531 {
1532         xlog_recover_item_t     *first_item, *itemq, *itemq_next;
1533         xfs_buf_log_format_t    *buf_f;
1534         ushort                  flags = 0;
1535
1536         first_item = itemq = trans->r_itemq;
1537         trans->r_itemq = NULL;
1538         do {
1539                 itemq_next = itemq->ri_next;
1540                 buf_f = (xfs_buf_log_format_t *)itemq->ri_buf[0].i_addr;
1541
1542                 switch (ITEM_TYPE(itemq)) {
1543                 case XFS_LI_BUF:
1544                         flags = buf_f->blf_flags;
1545                         if (!(flags & XFS_BLI_CANCEL)) {
1546                                 xlog_recover_insert_item_frontq(&trans->r_itemq,
1547                                                                 itemq);
1548                                 break;
1549                         }
1550                 case XFS_LI_INODE:
1551                 case XFS_LI_DQUOT:
1552                 case XFS_LI_QUOTAOFF:
1553                 case XFS_LI_EFD:
1554                 case XFS_LI_EFI:
1555                         xlog_recover_insert_item_backq(&trans->r_itemq, itemq);
1556                         break;
1557                 default:
1558                         xlog_warn(
1559         "XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1560                         ASSERT(0);
1561                         return XFS_ERROR(EIO);
1562                 }
1563                 itemq = itemq_next;
1564         } while (first_item != itemq);
1565         return 0;
1566 }
1567
1568 /*
1569  * Build up the table of buf cancel records so that we don't replay
1570  * cancelled data in the second pass.  For buffer records that are
1571  * not cancel records, there is nothing to do here so we just return.
1572  *
1573  * If we get a cancel record which is already in the table, this indicates
1574  * that the buffer was cancelled multiple times.  In order to ensure
1575  * that during pass 2 we keep the record in the table until we reach its
1576  * last occurrence in the log, we keep a reference count in the cancel
1577  * record in the table to tell us how many times we expect to see this
1578  * record during the second pass.
1579  */
1580 STATIC void
1581 xlog_recover_do_buffer_pass1(
1582         xlog_t                  *log,
1583         xfs_buf_log_format_t    *buf_f)
1584 {
1585         xfs_buf_cancel_t        *bcp;
1586         xfs_buf_cancel_t        *nextp;
1587         xfs_buf_cancel_t        *prevp;
1588         xfs_buf_cancel_t        **bucket;
1589         xfs_daddr_t             blkno = 0;
1590         uint                    len = 0;
1591         ushort                  flags = 0;
1592
1593         switch (buf_f->blf_type) {
1594         case XFS_LI_BUF:
1595                 blkno = buf_f->blf_blkno;
1596                 len = buf_f->blf_len;
1597                 flags = buf_f->blf_flags;
1598                 break;
1599         }
1600
1601         /*
1602          * If this isn't a cancel buffer item, then just return.
1603          */
1604         if (!(flags & XFS_BLI_CANCEL))
1605                 return;
1606
1607         /*
1608          * Insert an xfs_buf_cancel record into the hash table of
1609          * them.  If there is already an identical record, bump
1610          * its reference count.
1611          */
1612         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1613                                           XLOG_BC_TABLE_SIZE];
1614         /*
1615          * If the hash bucket is empty then just insert a new record into
1616          * the bucket.
1617          */
1618         if (*bucket == NULL) {
1619                 bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1620                                                      KM_SLEEP);
1621                 bcp->bc_blkno = blkno;
1622                 bcp->bc_len = len;
1623                 bcp->bc_refcount = 1;
1624                 bcp->bc_next = NULL;
1625                 *bucket = bcp;
1626                 return;
1627         }
1628
1629         /*
1630          * The hash bucket is not empty, so search for duplicates of our
1631          * record.  If we find one them just bump its refcount.  If not
1632          * then add us at the end of the list.
1633          */
1634         prevp = NULL;
1635         nextp = *bucket;
1636         while (nextp != NULL) {
1637                 if (nextp->bc_blkno == blkno && nextp->bc_len == len) {
1638                         nextp->bc_refcount++;
1639                         return;
1640                 }
1641                 prevp = nextp;
1642                 nextp = nextp->bc_next;
1643         }
1644         ASSERT(prevp != NULL);
1645         bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1646                                              KM_SLEEP);
1647         bcp->bc_blkno = blkno;
1648         bcp->bc_len = len;
1649         bcp->bc_refcount = 1;
1650         bcp->bc_next = NULL;
1651         prevp->bc_next = bcp;
1652 }
1653
1654 /*
1655  * Check to see whether the buffer being recovered has a corresponding
1656  * entry in the buffer cancel record table.  If it does then return 1
1657  * so that it will be cancelled, otherwise return 0.  If the buffer is
1658  * actually a buffer cancel item (XFS_BLI_CANCEL is set), then decrement
1659  * the refcount on the entry in the table and remove it from the table
1660  * if this is the last reference.
1661  *
1662  * We remove the cancel record from the table when we encounter its
1663  * last occurrence in the log so that if the same buffer is re-used
1664  * again after its last cancellation we actually replay the changes
1665  * made at that point.
1666  */
1667 STATIC int
1668 xlog_check_buffer_cancelled(
1669         xlog_t                  *log,
1670         xfs_daddr_t             blkno,
1671         uint                    len,
1672         ushort                  flags)
1673 {
1674         xfs_buf_cancel_t        *bcp;
1675         xfs_buf_cancel_t        *prevp;
1676         xfs_buf_cancel_t        **bucket;
1677
1678         if (log->l_buf_cancel_table == NULL) {
1679                 /*
1680                  * There is nothing in the table built in pass one,
1681                  * so this buffer must not be cancelled.
1682                  */
1683                 ASSERT(!(flags & XFS_BLI_CANCEL));
1684                 return 0;
1685         }
1686
1687         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1688                                           XLOG_BC_TABLE_SIZE];
1689         bcp = *bucket;
1690         if (bcp == NULL) {
1691                 /*
1692                  * There is no corresponding entry in the table built
1693                  * in pass one, so this buffer has not been cancelled.
1694                  */
1695                 ASSERT(!(flags & XFS_BLI_CANCEL));
1696                 return 0;
1697         }
1698
1699         /*
1700          * Search for an entry in the buffer cancel table that
1701          * matches our buffer.
1702          */
1703         prevp = NULL;
1704         while (bcp != NULL) {
1705                 if (bcp->bc_blkno == blkno && bcp->bc_len == len) {
1706                         /*
1707                          * We've go a match, so return 1 so that the
1708                          * recovery of this buffer is cancelled.
1709                          * If this buffer is actually a buffer cancel
1710                          * log item, then decrement the refcount on the
1711                          * one in the table and remove it if this is the
1712                          * last reference.
1713                          */
1714                         if (flags & XFS_BLI_CANCEL) {
1715                                 bcp->bc_refcount--;
1716                                 if (bcp->bc_refcount == 0) {
1717                                         if (prevp == NULL) {
1718                                                 *bucket = bcp->bc_next;
1719                                         } else {
1720                                                 prevp->bc_next = bcp->bc_next;
1721                                         }
1722                                         kmem_free(bcp);
1723                                 }
1724                         }
1725                         return 1;
1726                 }
1727                 prevp = bcp;
1728                 bcp = bcp->bc_next;
1729         }
1730         /*
1731          * We didn't find a corresponding entry in the table, so
1732          * return 0 so that the buffer is NOT cancelled.
1733          */
1734         ASSERT(!(flags & XFS_BLI_CANCEL));
1735         return 0;
1736 }
1737
1738 STATIC int
1739 xlog_recover_do_buffer_pass2(
1740         xlog_t                  *log,
1741         xfs_buf_log_format_t    *buf_f)
1742 {
1743         xfs_daddr_t             blkno = 0;
1744         ushort                  flags = 0;
1745         uint                    len = 0;
1746
1747         switch (buf_f->blf_type) {
1748         case XFS_LI_BUF:
1749                 blkno = buf_f->blf_blkno;
1750                 flags = buf_f->blf_flags;
1751                 len = buf_f->blf_len;
1752                 break;
1753         }
1754
1755         return xlog_check_buffer_cancelled(log, blkno, len, flags);
1756 }
1757
1758 /*
1759  * Perform recovery for a buffer full of inodes.  In these buffers,
1760  * the only data which should be recovered is that which corresponds
1761  * to the di_next_unlinked pointers in the on disk inode structures.
1762  * The rest of the data for the inodes is always logged through the
1763  * inodes themselves rather than the inode buffer and is recovered
1764  * in xlog_recover_do_inode_trans().
1765  *
1766  * The only time when buffers full of inodes are fully recovered is
1767  * when the buffer is full of newly allocated inodes.  In this case
1768  * the buffer will not be marked as an inode buffer and so will be
1769  * sent to xlog_recover_do_reg_buffer() below during recovery.
1770  */
1771 STATIC int
1772 xlog_recover_do_inode_buffer(
1773         xfs_mount_t             *mp,
1774         xlog_recover_item_t     *item,
1775         xfs_buf_t               *bp,
1776         xfs_buf_log_format_t    *buf_f)
1777 {
1778         int                     i;
1779         int                     item_index;
1780         int                     bit;
1781         int                     nbits;
1782         int                     reg_buf_offset;
1783         int                     reg_buf_bytes;
1784         int                     next_unlinked_offset;
1785         int                     inodes_per_buf;
1786         xfs_agino_t             *logged_nextp;
1787         xfs_agino_t             *buffer_nextp;
1788         unsigned int            *data_map = NULL;
1789         unsigned int            map_size = 0;
1790
1791         switch (buf_f->blf_type) {
1792         case XFS_LI_BUF:
1793                 data_map = buf_f->blf_data_map;
1794                 map_size = buf_f->blf_map_size;
1795                 break;
1796         }
1797         /*
1798          * Set the variables corresponding to the current region to
1799          * 0 so that we'll initialize them on the first pass through
1800          * the loop.
1801          */
1802         reg_buf_offset = 0;
1803         reg_buf_bytes = 0;
1804         bit = 0;
1805         nbits = 0;
1806         item_index = 0;
1807         inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1808         for (i = 0; i < inodes_per_buf; i++) {
1809                 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1810                         offsetof(xfs_dinode_t, di_next_unlinked);
1811
1812                 while (next_unlinked_offset >=
1813                        (reg_buf_offset + reg_buf_bytes)) {
1814                         /*
1815                          * The next di_next_unlinked field is beyond
1816                          * the current logged region.  Find the next
1817                          * logged region that contains or is beyond
1818                          * the current di_next_unlinked field.
1819                          */
1820                         bit += nbits;
1821                         bit = xfs_next_bit(data_map, map_size, bit);
1822
1823                         /*
1824                          * If there are no more logged regions in the
1825                          * buffer, then we're done.
1826                          */
1827                         if (bit == -1) {
1828                                 return 0;
1829                         }
1830
1831                         nbits = xfs_contig_bits(data_map, map_size,
1832                                                          bit);
1833                         ASSERT(nbits > 0);
1834                         reg_buf_offset = bit << XFS_BLI_SHIFT;
1835                         reg_buf_bytes = nbits << XFS_BLI_SHIFT;
1836                         item_index++;
1837                 }
1838
1839                 /*
1840                  * If the current logged region starts after the current
1841                  * di_next_unlinked field, then move on to the next
1842                  * di_next_unlinked field.
1843                  */
1844                 if (next_unlinked_offset < reg_buf_offset) {
1845                         continue;
1846                 }
1847
1848                 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1849                 ASSERT((item->ri_buf[item_index].i_len % XFS_BLI_CHUNK) == 0);
1850                 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1851
1852                 /*
1853                  * The current logged region contains a copy of the
1854                  * current di_next_unlinked field.  Extract its value
1855                  * and copy it to the buffer copy.
1856                  */
1857                 logged_nextp = (xfs_agino_t *)
1858                                ((char *)(item->ri_buf[item_index].i_addr) +
1859                                 (next_unlinked_offset - reg_buf_offset));
1860                 if (unlikely(*logged_nextp == 0)) {
1861                         xfs_fs_cmn_err(CE_ALERT, mp,
1862                                 "bad inode buffer log record (ptr = 0x%p, bp = 0x%p).  XFS trying to replay bad (0) inode di_next_unlinked field",
1863                                 item, bp);
1864                         XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1865                                          XFS_ERRLEVEL_LOW, mp);
1866                         return XFS_ERROR(EFSCORRUPTED);
1867                 }
1868
1869                 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1870                                               next_unlinked_offset);
1871                 *buffer_nextp = *logged_nextp;
1872         }
1873
1874         return 0;
1875 }
1876
1877 /*
1878  * Perform a 'normal' buffer recovery.  Each logged region of the
1879  * buffer should be copied over the corresponding region in the
1880  * given buffer.  The bitmap in the buf log format structure indicates
1881  * where to place the logged data.
1882  */
1883 /*ARGSUSED*/
1884 STATIC void
1885 xlog_recover_do_reg_buffer(
1886         xlog_recover_item_t     *item,
1887         xfs_buf_t               *bp,
1888         xfs_buf_log_format_t    *buf_f)
1889 {
1890         int                     i;
1891         int                     bit;
1892         int                     nbits;
1893         unsigned int            *data_map = NULL;
1894         unsigned int            map_size = 0;
1895         int                     error;
1896
1897         switch (buf_f->blf_type) {
1898         case XFS_LI_BUF:
1899                 data_map = buf_f->blf_data_map;
1900                 map_size = buf_f->blf_map_size;
1901                 break;
1902         }
1903         bit = 0;
1904         i = 1;  /* 0 is the buf format structure */
1905         while (1) {
1906                 bit = xfs_next_bit(data_map, map_size, bit);
1907                 if (bit == -1)
1908                         break;
1909                 nbits = xfs_contig_bits(data_map, map_size, bit);
1910                 ASSERT(nbits > 0);
1911                 ASSERT(item->ri_buf[i].i_addr != NULL);
1912                 ASSERT(item->ri_buf[i].i_len % XFS_BLI_CHUNK == 0);
1913                 ASSERT(XFS_BUF_COUNT(bp) >=
1914                        ((uint)bit << XFS_BLI_SHIFT)+(nbits<<XFS_BLI_SHIFT));
1915
1916                 /*
1917                  * Do a sanity check if this is a dquot buffer. Just checking
1918                  * the first dquot in the buffer should do. XXXThis is
1919                  * probably a good thing to do for other buf types also.
1920                  */
1921                 error = 0;
1922                 if (buf_f->blf_flags &
1923                    (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
1924                         error = xfs_qm_dqcheck((xfs_disk_dquot_t *)
1925                                                item->ri_buf[i].i_addr,
1926                                                -1, 0, XFS_QMOPT_DOWARN,
1927                                                "dquot_buf_recover");
1928                 }
1929                 if (!error)
1930                         memcpy(xfs_buf_offset(bp,
1931                                 (uint)bit << XFS_BLI_SHIFT),    /* dest */
1932                                 item->ri_buf[i].i_addr,         /* source */
1933                                 nbits<<XFS_BLI_SHIFT);          /* length */
1934                 i++;
1935                 bit += nbits;
1936         }
1937
1938         /* Shouldn't be any more regions */
1939         ASSERT(i == item->ri_total);
1940 }
1941
1942 /*
1943  * Do some primitive error checking on ondisk dquot data structures.
1944  */
1945 int
1946 xfs_qm_dqcheck(
1947         xfs_disk_dquot_t *ddq,
1948         xfs_dqid_t       id,
1949         uint             type,    /* used only when IO_dorepair is true */
1950         uint             flags,
1951         char             *str)
1952 {
1953         xfs_dqblk_t      *d = (xfs_dqblk_t *)ddq;
1954         int             errs = 0;
1955
1956         /*
1957          * We can encounter an uninitialized dquot buffer for 2 reasons:
1958          * 1. If we crash while deleting the quotainode(s), and those blks got
1959          *    used for user data. This is because we take the path of regular
1960          *    file deletion; however, the size field of quotainodes is never
1961          *    updated, so all the tricks that we play in itruncate_finish
1962          *    don't quite matter.
1963          *
1964          * 2. We don't play the quota buffers when there's a quotaoff logitem.
1965          *    But the allocation will be replayed so we'll end up with an
1966          *    uninitialized quota block.
1967          *
1968          * This is all fine; things are still consistent, and we haven't lost
1969          * any quota information. Just don't complain about bad dquot blks.
1970          */
1971         if (be16_to_cpu(ddq->d_magic) != XFS_DQUOT_MAGIC) {
1972                 if (flags & XFS_QMOPT_DOWARN)
1973                         cmn_err(CE_ALERT,
1974                         "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
1975                         str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
1976                 errs++;
1977         }
1978         if (ddq->d_version != XFS_DQUOT_VERSION) {
1979                 if (flags & XFS_QMOPT_DOWARN)
1980                         cmn_err(CE_ALERT,
1981                         "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
1982                         str, id, ddq->d_version, XFS_DQUOT_VERSION);
1983                 errs++;
1984         }
1985
1986         if (ddq->d_flags != XFS_DQ_USER &&
1987             ddq->d_flags != XFS_DQ_PROJ &&
1988             ddq->d_flags != XFS_DQ_GROUP) {
1989                 if (flags & XFS_QMOPT_DOWARN)
1990                         cmn_err(CE_ALERT,
1991                         "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
1992                         str, id, ddq->d_flags);
1993                 errs++;
1994         }
1995
1996         if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
1997                 if (flags & XFS_QMOPT_DOWARN)
1998                         cmn_err(CE_ALERT,
1999                         "%s : ondisk-dquot 0x%p, ID mismatch: "
2000                         "0x%x expected, found id 0x%x",
2001                         str, ddq, id, be32_to_cpu(ddq->d_id));
2002                 errs++;
2003         }
2004
2005         if (!errs && ddq->d_id) {
2006                 if (ddq->d_blk_softlimit &&
2007                     be64_to_cpu(ddq->d_bcount) >=
2008                                 be64_to_cpu(ddq->d_blk_softlimit)) {
2009                         if (!ddq->d_btimer) {
2010                                 if (flags & XFS_QMOPT_DOWARN)
2011                                         cmn_err(CE_ALERT,
2012                                         "%s : Dquot ID 0x%x (0x%p) "
2013                                         "BLK TIMER NOT STARTED",
2014                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2015                                 errs++;
2016                         }
2017                 }
2018                 if (ddq->d_ino_softlimit &&
2019                     be64_to_cpu(ddq->d_icount) >=
2020                                 be64_to_cpu(ddq->d_ino_softlimit)) {
2021                         if (!ddq->d_itimer) {
2022                                 if (flags & XFS_QMOPT_DOWARN)
2023                                         cmn_err(CE_ALERT,
2024                                         "%s : Dquot ID 0x%x (0x%p) "
2025                                         "INODE TIMER NOT STARTED",
2026                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2027                                 errs++;
2028                         }
2029                 }
2030                 if (ddq->d_rtb_softlimit &&
2031                     be64_to_cpu(ddq->d_rtbcount) >=
2032                                 be64_to_cpu(ddq->d_rtb_softlimit)) {
2033                         if (!ddq->d_rtbtimer) {
2034                                 if (flags & XFS_QMOPT_DOWARN)
2035                                         cmn_err(CE_ALERT,
2036                                         "%s : Dquot ID 0x%x (0x%p) "
2037                                         "RTBLK TIMER NOT STARTED",
2038                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2039                                 errs++;
2040                         }
2041                 }
2042         }
2043
2044         if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2045                 return errs;
2046
2047         if (flags & XFS_QMOPT_DOWARN)
2048                 cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2049
2050         /*
2051          * Typically, a repair is only requested by quotacheck.
2052          */
2053         ASSERT(id != -1);
2054         ASSERT(flags & XFS_QMOPT_DQREPAIR);
2055         memset(d, 0, sizeof(xfs_dqblk_t));
2056
2057         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2058         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2059         d->dd_diskdq.d_flags = type;
2060         d->dd_diskdq.d_id = cpu_to_be32(id);
2061
2062         return errs;
2063 }
2064
2065 /*
2066  * Perform a dquot buffer recovery.
2067  * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2068  * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2069  * Else, treat it as a regular buffer and do recovery.
2070  */
2071 STATIC void
2072 xlog_recover_do_dquot_buffer(
2073         xfs_mount_t             *mp,
2074         xlog_t                  *log,
2075         xlog_recover_item_t     *item,
2076         xfs_buf_t               *bp,
2077         xfs_buf_log_format_t    *buf_f)
2078 {
2079         uint                    type;
2080
2081         /*
2082          * Filesystems are required to send in quota flags at mount time.
2083          */
2084         if (mp->m_qflags == 0) {
2085                 return;
2086         }
2087
2088         type = 0;
2089         if (buf_f->blf_flags & XFS_BLI_UDQUOT_BUF)
2090                 type |= XFS_DQ_USER;
2091         if (buf_f->blf_flags & XFS_BLI_PDQUOT_BUF)
2092                 type |= XFS_DQ_PROJ;
2093         if (buf_f->blf_flags & XFS_BLI_GDQUOT_BUF)
2094                 type |= XFS_DQ_GROUP;
2095         /*
2096          * This type of quotas was turned off, so ignore this buffer
2097          */
2098         if (log->l_quotaoffs_flag & type)
2099                 return;
2100
2101         xlog_recover_do_reg_buffer(item, bp, buf_f);
2102 }
2103
2104 /*
2105  * This routine replays a modification made to a buffer at runtime.
2106  * There are actually two types of buffer, regular and inode, which
2107  * are handled differently.  Inode buffers are handled differently
2108  * in that we only recover a specific set of data from them, namely
2109  * the inode di_next_unlinked fields.  This is because all other inode
2110  * data is actually logged via inode records and any data we replay
2111  * here which overlaps that may be stale.
2112  *
2113  * When meta-data buffers are freed at run time we log a buffer item
2114  * with the XFS_BLI_CANCEL bit set to indicate that previous copies
2115  * of the buffer in the log should not be replayed at recovery time.
2116  * This is so that if the blocks covered by the buffer are reused for
2117  * file data before we crash we don't end up replaying old, freed
2118  * meta-data into a user's file.
2119  *
2120  * To handle the cancellation of buffer log items, we make two passes
2121  * over the log during recovery.  During the first we build a table of
2122  * those buffers which have been cancelled, and during the second we
2123  * only replay those buffers which do not have corresponding cancel
2124  * records in the table.  See xlog_recover_do_buffer_pass[1,2] above
2125  * for more details on the implementation of the table of cancel records.
2126  */
2127 STATIC int
2128 xlog_recover_do_buffer_trans(
2129         xlog_t                  *log,
2130         xlog_recover_item_t     *item,
2131         int                     pass)
2132 {
2133         xfs_buf_log_format_t    *buf_f;
2134         xfs_mount_t             *mp;
2135         xfs_buf_t               *bp;
2136         int                     error;
2137         int                     cancel;
2138         xfs_daddr_t             blkno;
2139         int                     len;
2140         ushort                  flags;
2141
2142         buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
2143
2144         if (pass == XLOG_RECOVER_PASS1) {
2145                 /*
2146                  * In this pass we're only looking for buf items
2147                  * with the XFS_BLI_CANCEL bit set.
2148                  */
2149                 xlog_recover_do_buffer_pass1(log, buf_f);
2150                 return 0;
2151         } else {
2152                 /*
2153                  * In this pass we want to recover all the buffers
2154                  * which have not been cancelled and are not
2155                  * cancellation buffers themselves.  The routine
2156                  * we call here will tell us whether or not to
2157                  * continue with the replay of this buffer.
2158                  */
2159                 cancel = xlog_recover_do_buffer_pass2(log, buf_f);
2160                 if (cancel) {
2161                         return 0;
2162                 }
2163         }
2164         switch (buf_f->blf_type) {
2165         case XFS_LI_BUF:
2166                 blkno = buf_f->blf_blkno;
2167                 len = buf_f->blf_len;
2168                 flags = buf_f->blf_flags;
2169                 break;
2170         default:
2171                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
2172                         "xfs_log_recover: unknown buffer type 0x%x, logdev %s",
2173                         buf_f->blf_type, log->l_mp->m_logname ?
2174                         log->l_mp->m_logname : "internal");
2175                 XFS_ERROR_REPORT("xlog_recover_do_buffer_trans",
2176                                  XFS_ERRLEVEL_LOW, log->l_mp);
2177                 return XFS_ERROR(EFSCORRUPTED);
2178         }
2179
2180         mp = log->l_mp;
2181         if (flags & XFS_BLI_INODE_BUF) {
2182                 bp = xfs_buf_read_flags(mp->m_ddev_targp, blkno, len,
2183                                                                 XFS_BUF_LOCK);
2184         } else {
2185                 bp = xfs_buf_read(mp->m_ddev_targp, blkno, len, 0);
2186         }
2187         if (XFS_BUF_ISERROR(bp)) {
2188                 xfs_ioerror_alert("xlog_recover_do..(read#1)", log->l_mp,
2189                                   bp, blkno);
2190                 error = XFS_BUF_GETERROR(bp);
2191                 xfs_buf_relse(bp);
2192                 return error;
2193         }
2194
2195         error = 0;
2196         if (flags & XFS_BLI_INODE_BUF) {
2197                 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2198         } else if (flags &
2199                   (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
2200                 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2201         } else {
2202                 xlog_recover_do_reg_buffer(item, bp, buf_f);
2203         }
2204         if (error)
2205                 return XFS_ERROR(error);
2206
2207         /*
2208          * Perform delayed write on the buffer.  Asynchronous writes will be
2209          * slower when taking into account all the buffers to be flushed.
2210          *
2211          * Also make sure that only inode buffers with good sizes stay in
2212          * the buffer cache.  The kernel moves inodes in buffers of 1 block
2213          * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger.  The inode
2214          * buffers in the log can be a different size if the log was generated
2215          * by an older kernel using unclustered inode buffers or a newer kernel
2216          * running with a different inode cluster size.  Regardless, if the
2217          * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2218          * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2219          * the buffer out of the buffer cache so that the buffer won't
2220          * overlap with future reads of those inodes.
2221          */
2222         if (XFS_DINODE_MAGIC ==
2223             be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2224             (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2225                         (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2226                 XFS_BUF_STALE(bp);
2227                 error = xfs_bwrite(mp, bp);
2228         } else {
2229                 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2230                        XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2231                 XFS_BUF_SET_FSPRIVATE(bp, mp);
2232                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2233                 xfs_bdwrite(mp, bp);
2234         }
2235
2236         return (error);
2237 }
2238
2239 STATIC int
2240 xlog_recover_do_inode_trans(
2241         xlog_t                  *log,
2242         xlog_recover_item_t     *item,
2243         int                     pass)
2244 {
2245         xfs_inode_log_format_t  *in_f;
2246         xfs_mount_t             *mp;
2247         xfs_buf_t               *bp;
2248         xfs_imap_t              imap;
2249         xfs_dinode_t            *dip;
2250         xfs_ino_t               ino;
2251         int                     len;
2252         xfs_caddr_t             src;
2253         xfs_caddr_t             dest;
2254         int                     error;
2255         int                     attr_index;
2256         uint                    fields;
2257         xfs_icdinode_t          *dicp;
2258         int                     need_free = 0;
2259
2260         if (pass == XLOG_RECOVER_PASS1) {
2261                 return 0;
2262         }
2263
2264         if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2265                 in_f = (xfs_inode_log_format_t *)item->ri_buf[0].i_addr;
2266         } else {
2267                 in_f = (xfs_inode_log_format_t *)kmem_alloc(
2268                         sizeof(xfs_inode_log_format_t), KM_SLEEP);
2269                 need_free = 1;
2270                 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2271                 if (error)
2272                         goto error;
2273         }
2274         ino = in_f->ilf_ino;
2275         mp = log->l_mp;
2276         if (ITEM_TYPE(item) == XFS_LI_INODE) {
2277                 imap.im_blkno = (xfs_daddr_t)in_f->ilf_blkno;
2278                 imap.im_len = in_f->ilf_len;
2279                 imap.im_boffset = in_f->ilf_boffset;
2280         } else {
2281                 /*
2282                  * It's an old inode format record.  We don't know where
2283                  * its cluster is located on disk, and we can't allow
2284                  * xfs_imap() to figure it out because the inode btrees
2285                  * are not ready to be used.  Therefore do not pass the
2286                  * XFS_IMAP_LOOKUP flag to xfs_imap().  This will give
2287                  * us only the single block in which the inode lives
2288                  * rather than its cluster, so we must make sure to
2289                  * invalidate the buffer when we write it out below.
2290                  */
2291                 imap.im_blkno = 0;
2292                 error = xfs_imap(log->l_mp, NULL, ino, &imap, 0);
2293                 if (error)
2294                         goto error;
2295         }
2296
2297         /*
2298          * Inode buffers can be freed, look out for it,
2299          * and do not replay the inode.
2300          */
2301         if (xlog_check_buffer_cancelled(log, imap.im_blkno, imap.im_len, 0)) {
2302                 error = 0;
2303                 goto error;
2304         }
2305
2306         bp = xfs_buf_read_flags(mp->m_ddev_targp, imap.im_blkno, imap.im_len,
2307                                                                 XFS_BUF_LOCK);
2308         if (XFS_BUF_ISERROR(bp)) {
2309                 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2310                                   bp, imap.im_blkno);
2311                 error = XFS_BUF_GETERROR(bp);
2312                 xfs_buf_relse(bp);
2313                 goto error;
2314         }
2315         error = 0;
2316         ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2317         dip = (xfs_dinode_t *)xfs_buf_offset(bp, imap.im_boffset);
2318
2319         /*
2320          * Make sure the place we're flushing out to really looks
2321          * like an inode!
2322          */
2323         if (unlikely(be16_to_cpu(dip->di_core.di_magic) != XFS_DINODE_MAGIC)) {
2324                 xfs_buf_relse(bp);
2325                 xfs_fs_cmn_err(CE_ALERT, mp,
2326                         "xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2327                         dip, bp, ino);
2328                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(1)",
2329                                  XFS_ERRLEVEL_LOW, mp);
2330                 error = EFSCORRUPTED;
2331                 goto error;
2332         }
2333         dicp = (xfs_icdinode_t *)(item->ri_buf[1].i_addr);
2334         if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2335                 xfs_buf_relse(bp);
2336                 xfs_fs_cmn_err(CE_ALERT, mp,
2337                         "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2338                         item, ino);
2339                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(2)",
2340                                  XFS_ERRLEVEL_LOW, mp);
2341                 error = EFSCORRUPTED;
2342                 goto error;
2343         }
2344
2345         /* Skip replay when the on disk inode is newer than the log one */
2346         if (dicp->di_flushiter < be16_to_cpu(dip->di_core.di_flushiter)) {
2347                 /*
2348                  * Deal with the wrap case, DI_MAX_FLUSH is less
2349                  * than smaller numbers
2350                  */
2351                 if (be16_to_cpu(dip->di_core.di_flushiter) == DI_MAX_FLUSH &&
2352                     dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2353                         /* do nothing */
2354                 } else {
2355                         xfs_buf_relse(bp);
2356                         error = 0;
2357                         goto error;
2358                 }
2359         }
2360         /* Take the opportunity to reset the flush iteration count */
2361         dicp->di_flushiter = 0;
2362
2363         if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2364                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2365                     (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2366                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(3)",
2367                                          XFS_ERRLEVEL_LOW, mp, dicp);
2368                         xfs_buf_relse(bp);
2369                         xfs_fs_cmn_err(CE_ALERT, mp,
2370                                 "xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2371                                 item, dip, bp, ino);
2372                         error = EFSCORRUPTED;
2373                         goto error;
2374                 }
2375         } else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2376                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2377                     (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2378                     (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2379                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(4)",
2380                                              XFS_ERRLEVEL_LOW, mp, dicp);
2381                         xfs_buf_relse(bp);
2382                         xfs_fs_cmn_err(CE_ALERT, mp,
2383                                 "xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2384                                 item, dip, bp, ino);
2385                         error = EFSCORRUPTED;
2386                         goto error;
2387                 }
2388         }
2389         if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2390                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(5)",
2391                                      XFS_ERRLEVEL_LOW, mp, dicp);
2392                 xfs_buf_relse(bp);
2393                 xfs_fs_cmn_err(CE_ALERT, mp,
2394                         "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",
2395                         item, dip, bp, ino,
2396                         dicp->di_nextents + dicp->di_anextents,
2397                         dicp->di_nblocks);
2398                 error = EFSCORRUPTED;
2399                 goto error;
2400         }
2401         if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2402                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(6)",
2403                                      XFS_ERRLEVEL_LOW, mp, dicp);
2404                 xfs_buf_relse(bp);
2405                 xfs_fs_cmn_err(CE_ALERT, mp,
2406                         "xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2407                         item, dip, bp, ino, dicp->di_forkoff);
2408                 error = EFSCORRUPTED;
2409                 goto error;
2410         }
2411         if (unlikely(item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t))) {
2412                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(7)",
2413                                      XFS_ERRLEVEL_LOW, mp, dicp);
2414                 xfs_buf_relse(bp);
2415                 xfs_fs_cmn_err(CE_ALERT, mp,
2416                         "xfs_inode_recover: Bad inode log record length %d, rec ptr 0x%p",
2417                         item->ri_buf[1].i_len, item);
2418                 error = EFSCORRUPTED;
2419                 goto error;
2420         }
2421
2422         /* The core is in in-core format */
2423         xfs_dinode_to_disk(&dip->di_core,
2424                 (xfs_icdinode_t *)item->ri_buf[1].i_addr);
2425
2426         /* the rest is in on-disk format */
2427         if (item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t)) {
2428                 memcpy((xfs_caddr_t) dip + sizeof(xfs_dinode_core_t),
2429                         item->ri_buf[1].i_addr + sizeof(xfs_dinode_core_t),
2430                         item->ri_buf[1].i_len  - sizeof(xfs_dinode_core_t));
2431         }
2432
2433         fields = in_f->ilf_fields;
2434         switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2435         case XFS_ILOG_DEV:
2436                 dip->di_u.di_dev = cpu_to_be32(in_f->ilf_u.ilfu_rdev);
2437                 break;
2438         case XFS_ILOG_UUID:
2439                 dip->di_u.di_muuid = in_f->ilf_u.ilfu_uuid;
2440                 break;
2441         }
2442
2443         if (in_f->ilf_size == 2)
2444                 goto write_inode_buffer;
2445         len = item->ri_buf[2].i_len;
2446         src = item->ri_buf[2].i_addr;
2447         ASSERT(in_f->ilf_size <= 4);
2448         ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2449         ASSERT(!(fields & XFS_ILOG_DFORK) ||
2450                (len == in_f->ilf_dsize));
2451
2452         switch (fields & XFS_ILOG_DFORK) {
2453         case XFS_ILOG_DDATA:
2454         case XFS_ILOG_DEXT:
2455                 memcpy(&dip->di_u, src, len);
2456                 break;
2457
2458         case XFS_ILOG_DBROOT:
2459                 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
2460                                  &dip->di_u.di_bmbt,
2461                                  XFS_DFORK_DSIZE(dip, mp));
2462                 break;
2463
2464         default:
2465                 /*
2466                  * There are no data fork flags set.
2467                  */
2468                 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2469                 break;
2470         }
2471
2472         /*
2473          * If we logged any attribute data, recover it.  There may or
2474          * may not have been any other non-core data logged in this
2475          * transaction.
2476          */
2477         if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2478                 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2479                         attr_index = 3;
2480                 } else {
2481                         attr_index = 2;
2482                 }
2483                 len = item->ri_buf[attr_index].i_len;
2484                 src = item->ri_buf[attr_index].i_addr;
2485                 ASSERT(len == in_f->ilf_asize);
2486
2487                 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2488                 case XFS_ILOG_ADATA:
2489                 case XFS_ILOG_AEXT:
2490                         dest = XFS_DFORK_APTR(dip);
2491                         ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2492                         memcpy(dest, src, len);
2493                         break;
2494
2495                 case XFS_ILOG_ABROOT:
2496                         dest = XFS_DFORK_APTR(dip);
2497                         xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
2498                                          len, (xfs_bmdr_block_t*)dest,
2499                                          XFS_DFORK_ASIZE(dip, mp));
2500                         break;
2501
2502                 default:
2503                         xlog_warn("XFS: xlog_recover_do_inode_trans: Invalid flag");
2504                         ASSERT(0);
2505                         xfs_buf_relse(bp);
2506                         error = EIO;
2507                         goto error;
2508                 }
2509         }
2510
2511 write_inode_buffer:
2512         if (ITEM_TYPE(item) == XFS_LI_INODE) {
2513                 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2514                        XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2515                 XFS_BUF_SET_FSPRIVATE(bp, mp);
2516                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2517                 xfs_bdwrite(mp, bp);
2518         } else {
2519                 XFS_BUF_STALE(bp);
2520                 error = xfs_bwrite(mp, bp);
2521         }
2522
2523 error:
2524         if (need_free)
2525                 kmem_free(in_f);
2526         return XFS_ERROR(error);
2527 }
2528
2529 /*
2530  * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2531  * structure, so that we know not to do any dquot item or dquot buffer recovery,
2532  * of that type.
2533  */
2534 STATIC int
2535 xlog_recover_do_quotaoff_trans(
2536         xlog_t                  *log,
2537         xlog_recover_item_t     *item,
2538         int                     pass)
2539 {
2540         xfs_qoff_logformat_t    *qoff_f;
2541
2542         if (pass == XLOG_RECOVER_PASS2) {
2543                 return (0);
2544         }
2545
2546         qoff_f = (xfs_qoff_logformat_t *)item->ri_buf[0].i_addr;
2547         ASSERT(qoff_f);
2548
2549         /*
2550          * The logitem format's flag tells us if this was user quotaoff,
2551          * group/project quotaoff or both.
2552          */
2553         if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2554                 log->l_quotaoffs_flag |= XFS_DQ_USER;
2555         if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2556                 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2557         if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2558                 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2559
2560         return (0);
2561 }
2562
2563 /*
2564  * Recover a dquot record
2565  */
2566 STATIC int
2567 xlog_recover_do_dquot_trans(
2568         xlog_t                  *log,
2569         xlog_recover_item_t     *item,
2570         int                     pass)
2571 {
2572         xfs_mount_t             *mp;
2573         xfs_buf_t               *bp;
2574         struct xfs_disk_dquot   *ddq, *recddq;
2575         int                     error;
2576         xfs_dq_logformat_t      *dq_f;
2577         uint                    type;
2578
2579         if (pass == XLOG_RECOVER_PASS1) {
2580                 return 0;
2581         }
2582         mp = log->l_mp;
2583
2584         /*
2585          * Filesystems are required to send in quota flags at mount time.
2586          */
2587         if (mp->m_qflags == 0)
2588                 return (0);
2589
2590         recddq = (xfs_disk_dquot_t *)item->ri_buf[1].i_addr;
2591         ASSERT(recddq);
2592         /*
2593          * This type of quotas was turned off, so ignore this record.
2594          */
2595         type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2596         ASSERT(type);
2597         if (log->l_quotaoffs_flag & type)
2598                 return (0);
2599
2600         /*
2601          * At this point we know that quota was _not_ turned off.
2602          * Since the mount flags are not indicating to us otherwise, this
2603          * must mean that quota is on, and the dquot needs to be replayed.
2604          * Remember that we may not have fully recovered the superblock yet,
2605          * so we can't do the usual trick of looking at the SB quota bits.
2606          *
2607          * The other possibility, of course, is that the quota subsystem was
2608          * removed since the last mount - ENOSYS.
2609          */
2610         dq_f = (xfs_dq_logformat_t *)item->ri_buf[0].i_addr;
2611         ASSERT(dq_f);
2612         if ((error = xfs_qm_dqcheck(recddq,
2613                            dq_f->qlf_id,
2614                            0, XFS_QMOPT_DOWARN,
2615                            "xlog_recover_do_dquot_trans (log copy)"))) {
2616                 return XFS_ERROR(EIO);
2617         }
2618         ASSERT(dq_f->qlf_len == 1);
2619
2620         error = xfs_read_buf(mp, mp->m_ddev_targp,
2621                              dq_f->qlf_blkno,
2622                              XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2623                              0, &bp);
2624         if (error) {
2625                 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2626                                   bp, dq_f->qlf_blkno);
2627                 return error;
2628         }
2629         ASSERT(bp);
2630         ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2631
2632         /*
2633          * At least the magic num portion should be on disk because this
2634          * was among a chunk of dquots created earlier, and we did some
2635          * minimal initialization then.
2636          */
2637         if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2638                            "xlog_recover_do_dquot_trans")) {
2639                 xfs_buf_relse(bp);
2640                 return XFS_ERROR(EIO);
2641         }
2642
2643         memcpy(ddq, recddq, item->ri_buf[1].i_len);
2644
2645         ASSERT(dq_f->qlf_size == 2);
2646         ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2647                XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2648         XFS_BUF_SET_FSPRIVATE(bp, mp);
2649         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2650         xfs_bdwrite(mp, bp);
2651
2652         return (0);
2653 }
2654
2655 /*
2656  * This routine is called to create an in-core extent free intent
2657  * item from the efi format structure which was logged on disk.
2658  * It allocates an in-core efi, copies the extents from the format
2659  * structure into it, and adds the efi to the AIL with the given
2660  * LSN.
2661  */
2662 STATIC int
2663 xlog_recover_do_efi_trans(
2664         xlog_t                  *log,
2665         xlog_recover_item_t     *item,
2666         xfs_lsn_t               lsn,
2667         int                     pass)
2668 {
2669         int                     error;
2670         xfs_mount_t             *mp;
2671         xfs_efi_log_item_t      *efip;
2672         xfs_efi_log_format_t    *efi_formatp;
2673
2674         if (pass == XLOG_RECOVER_PASS1) {
2675                 return 0;
2676         }
2677
2678         efi_formatp = (xfs_efi_log_format_t *)item->ri_buf[0].i_addr;
2679
2680         mp = log->l_mp;
2681         efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2682         if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2683                                          &(efip->efi_format)))) {
2684                 xfs_efi_item_free(efip);
2685                 return error;
2686         }
2687         efip->efi_next_extent = efi_formatp->efi_nextents;
2688         efip->efi_flags |= XFS_EFI_COMMITTED;
2689
2690         spin_lock(&log->l_ailp->xa_lock);
2691         /*
2692          * xfs_trans_ail_update() drops the AIL lock.
2693          */
2694         xfs_trans_ail_update(log->l_ailp, (xfs_log_item_t *)efip, lsn);
2695         return 0;
2696 }
2697
2698
2699 /*
2700  * This routine is called when an efd format structure is found in
2701  * a committed transaction in the log.  It's purpose is to cancel
2702  * the corresponding efi if it was still in the log.  To do this
2703  * it searches the AIL for the efi with an id equal to that in the
2704  * efd format structure.  If we find it, we remove the efi from the
2705  * AIL and free it.
2706  */
2707 STATIC void
2708 xlog_recover_do_efd_trans(
2709         xlog_t                  *log,
2710         xlog_recover_item_t     *item,
2711         int                     pass)
2712 {
2713         xfs_efd_log_format_t    *efd_formatp;
2714         xfs_efi_log_item_t      *efip = NULL;
2715         xfs_log_item_t          *lip;
2716         __uint64_t              efi_id;
2717         struct xfs_ail_cursor   cur;
2718         struct xfs_ail          *ailp = log->l_ailp;
2719
2720         if (pass == XLOG_RECOVER_PASS1) {
2721                 return;
2722         }
2723
2724         efd_formatp = (xfs_efd_log_format_t *)item->ri_buf[0].i_addr;
2725         ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2726                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2727                (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2728                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2729         efi_id = efd_formatp->efd_efi_id;
2730
2731         /*
2732          * Search for the efi with the id in the efd format structure
2733          * in the AIL.
2734          */
2735         spin_lock(&ailp->xa_lock);
2736         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2737         while (lip != NULL) {
2738                 if (lip->li_type == XFS_LI_EFI) {
2739                         efip = (xfs_efi_log_item_t *)lip;
2740                         if (efip->efi_format.efi_id == efi_id) {
2741                                 /*
2742                                  * xfs_trans_ail_delete() drops the
2743                                  * AIL lock.
2744                                  */
2745                                 xfs_trans_ail_delete(ailp, lip);
2746                                 xfs_efi_item_free(efip);
2747                                 spin_lock(&ailp->xa_lock);
2748                                 break;
2749                         }
2750                 }
2751                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2752         }
2753         xfs_trans_ail_cursor_done(ailp, &cur);
2754         spin_unlock(&ailp->xa_lock);
2755 }
2756
2757 /*
2758  * Perform the transaction
2759  *
2760  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2761  * EFIs and EFDs get queued up by adding entries into the AIL for them.
2762  */
2763 STATIC int
2764 xlog_recover_do_trans(
2765         xlog_t                  *log,
2766         xlog_recover_t          *trans,
2767         int                     pass)
2768 {
2769         int                     error = 0;
2770         xlog_recover_item_t     *item, *first_item;
2771
2772         if ((error = xlog_recover_reorder_trans(trans)))
2773                 return error;
2774         first_item = item = trans->r_itemq;
2775         do {
2776                 /*
2777                  * we don't need to worry about the block number being
2778                  * truncated in > 1 TB buffers because in user-land,
2779                  * we're now n32 or 64-bit so xfs_daddr_t is 64-bits so
2780                  * the blknos will get through the user-mode buffer
2781                  * cache properly.  The only bad case is o32 kernels
2782                  * where xfs_daddr_t is 32-bits but mount will warn us
2783                  * off a > 1 TB filesystem before we get here.
2784                  */
2785                 if ((ITEM_TYPE(item) == XFS_LI_BUF)) {
2786                         if  ((error = xlog_recover_do_buffer_trans(log, item,
2787                                                                  pass)))
2788                                 break;
2789                 } else if ((ITEM_TYPE(item) == XFS_LI_INODE)) {
2790                         if ((error = xlog_recover_do_inode_trans(log, item,
2791                                                                 pass)))
2792                                 break;
2793                 } else if (ITEM_TYPE(item) == XFS_LI_EFI) {
2794                         if ((error = xlog_recover_do_efi_trans(log, item, trans->r_lsn,
2795                                                   pass)))
2796                                 break;
2797                 } else if (ITEM_TYPE(item) == XFS_LI_EFD) {
2798                         xlog_recover_do_efd_trans(log, item, pass);
2799                 } else if (ITEM_TYPE(item) == XFS_LI_DQUOT) {
2800                         if ((error = xlog_recover_do_dquot_trans(log, item,
2801                                                                    pass)))
2802                                         break;
2803                 } else if ((ITEM_TYPE(item) == XFS_LI_QUOTAOFF)) {
2804                         if ((error = xlog_recover_do_quotaoff_trans(log, item,
2805                                                                    pass)))
2806                                         break;
2807                 } else {
2808                         xlog_warn("XFS: xlog_recover_do_trans");
2809                         ASSERT(0);
2810                         error = XFS_ERROR(EIO);
2811                         break;
2812                 }
2813                 item = item->ri_next;
2814         } while (first_item != item);
2815
2816         return error;
2817 }
2818
2819 /*
2820  * Free up any resources allocated by the transaction
2821  *
2822  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2823  */
2824 STATIC void
2825 xlog_recover_free_trans(
2826         xlog_recover_t          *trans)
2827 {
2828         xlog_recover_item_t     *first_item, *item, *free_item;
2829         int                     i;
2830
2831         item = first_item = trans->r_itemq;
2832         do {
2833                 free_item = item;
2834                 item = item->ri_next;
2835                  /* Free the regions in the item. */
2836                 for (i = 0; i < free_item->ri_cnt; i++) {
2837                         kmem_free(free_item->ri_buf[i].i_addr);
2838                 }
2839                 /* Free the item itself */
2840                 kmem_free(free_item->ri_buf);
2841                 kmem_free(free_item);
2842         } while (first_item != item);
2843         /* Free the transaction recover structure */
2844         kmem_free(trans);
2845 }
2846
2847 STATIC int
2848 xlog_recover_commit_trans(
2849         xlog_t                  *log,
2850         xlog_recover_t          **q,
2851         xlog_recover_t          *trans,
2852         int                     pass)
2853 {
2854         int                     error;
2855
2856         if ((error = xlog_recover_unlink_tid(q, trans)))
2857                 return error;
2858         if ((error = xlog_recover_do_trans(log, trans, pass)))
2859                 return error;
2860         xlog_recover_free_trans(trans);                 /* no error */
2861         return 0;
2862 }
2863
2864 STATIC int
2865 xlog_recover_unmount_trans(
2866         xlog_recover_t          *trans)
2867 {
2868         /* Do nothing now */
2869         xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2870         return 0;
2871 }
2872
2873 /*
2874  * There are two valid states of the r_state field.  0 indicates that the
2875  * transaction structure is in a normal state.  We have either seen the
2876  * start of the transaction or the last operation we added was not a partial
2877  * operation.  If the last operation we added to the transaction was a
2878  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2879  *
2880  * NOTE: skip LRs with 0 data length.
2881  */
2882 STATIC int
2883 xlog_recover_process_data(
2884         xlog_t                  *log,
2885         xlog_recover_t          *rhash[],
2886         xlog_rec_header_t       *rhead,
2887         xfs_caddr_t             dp,
2888         int                     pass)
2889 {
2890         xfs_caddr_t             lp;
2891         int                     num_logops;
2892         xlog_op_header_t        *ohead;
2893         xlog_recover_t          *trans;
2894         xlog_tid_t              tid;
2895         int                     error;
2896         unsigned long           hash;
2897         uint                    flags;
2898
2899         lp = dp + be32_to_cpu(rhead->h_len);
2900         num_logops = be32_to_cpu(rhead->h_num_logops);
2901
2902         /* check the log format matches our own - else we can't recover */
2903         if (xlog_header_check_recover(log->l_mp, rhead))
2904                 return (XFS_ERROR(EIO));
2905
2906         while ((dp < lp) && num_logops) {
2907                 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2908                 ohead = (xlog_op_header_t *)dp;
2909                 dp += sizeof(xlog_op_header_t);
2910                 if (ohead->oh_clientid != XFS_TRANSACTION &&
2911                     ohead->oh_clientid != XFS_LOG) {
2912                         xlog_warn(
2913                 "XFS: xlog_recover_process_data: bad clientid");
2914                         ASSERT(0);
2915                         return (XFS_ERROR(EIO));
2916                 }
2917                 tid = be32_to_cpu(ohead->oh_tid);
2918                 hash = XLOG_RHASH(tid);
2919                 trans = xlog_recover_find_tid(rhash[hash], tid);
2920                 if (trans == NULL) {               /* not found; add new tid */
2921                         if (ohead->oh_flags & XLOG_START_TRANS)
2922                                 xlog_recover_new_tid(&rhash[hash], tid,
2923                                         be64_to_cpu(rhead->h_lsn));
2924                 } else {
2925                         if (dp + be32_to_cpu(ohead->oh_len) > lp) {
2926                                 xlog_warn(
2927                         "XFS: xlog_recover_process_data: bad length");
2928                                 WARN_ON(1);
2929                                 return (XFS_ERROR(EIO));
2930                         }
2931                         flags = ohead->oh_flags & ~XLOG_END_TRANS;
2932                         if (flags & XLOG_WAS_CONT_TRANS)
2933                                 flags &= ~XLOG_CONTINUE_TRANS;
2934                         switch (flags) {
2935                         case XLOG_COMMIT_TRANS:
2936                                 error = xlog_recover_commit_trans(log,
2937                                                 &rhash[hash], trans, pass);
2938                                 break;
2939                         case XLOG_UNMOUNT_TRANS:
2940                                 error = xlog_recover_unmount_trans(trans);
2941                                 break;
2942                         case XLOG_WAS_CONT_TRANS:
2943                                 error = xlog_recover_add_to_cont_trans(trans,
2944                                                 dp, be32_to_cpu(ohead->oh_len));
2945                                 break;
2946                         case XLOG_START_TRANS:
2947                                 xlog_warn(
2948                         "XFS: xlog_recover_process_data: bad transaction");
2949                                 ASSERT(0);
2950                                 error = XFS_ERROR(EIO);
2951                                 break;
2952                         case 0:
2953                         case XLOG_CONTINUE_TRANS:
2954                                 error = xlog_recover_add_to_trans(trans,
2955                                                 dp, be32_to_cpu(ohead->oh_len));
2956                                 break;
2957                         default:
2958                                 xlog_warn(
2959                         "XFS: xlog_recover_process_data: bad flag");
2960                                 ASSERT(0);
2961                                 error = XFS_ERROR(EIO);
2962                                 break;
2963                         }
2964                         if (error)
2965                                 return error;
2966                 }
2967                 dp += be32_to_cpu(ohead->oh_len);
2968                 num_logops--;
2969         }
2970         return 0;
2971 }
2972
2973 /*
2974  * Process an extent free intent item that was recovered from
2975  * the log.  We need to free the extents that it describes.
2976  */
2977 STATIC int
2978 xlog_recover_process_efi(
2979         xfs_mount_t             *mp,
2980         xfs_efi_log_item_t      *efip)
2981 {
2982         xfs_efd_log_item_t      *efdp;
2983         xfs_trans_t             *tp;
2984         int                     i;
2985         int                     error = 0;
2986         xfs_extent_t            *extp;
2987         xfs_fsblock_t           startblock_fsb;
2988
2989         ASSERT(!(efip->efi_flags & XFS_EFI_RECOVERED));
2990
2991         /*
2992          * First check the validity of the extents described by the
2993          * EFI.  If any are bad, then assume that all are bad and
2994          * just toss the EFI.
2995          */
2996         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2997                 extp = &(efip->efi_format.efi_extents[i]);
2998                 startblock_fsb = XFS_BB_TO_FSB(mp,
2999                                    XFS_FSB_TO_DADDR(mp, extp->ext_start));
3000                 if ((startblock_fsb == 0) ||
3001                     (extp->ext_len == 0) ||
3002                     (startblock_fsb >= mp->m_sb.sb_dblocks) ||
3003                     (extp->ext_len >= mp->m_sb.sb_agblocks)) {
3004                         /*
3005                          * This will pull the EFI from the AIL and
3006                          * free the memory associated with it.
3007                          */
3008                         xfs_efi_release(efip, efip->efi_format.efi_nextents);
3009                         return XFS_ERROR(EIO);
3010                 }
3011         }
3012
3013         tp = xfs_trans_alloc(mp, 0);
3014         error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
3015         if (error)
3016                 goto abort_error;
3017         efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3018
3019         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3020                 extp = &(efip->efi_format.efi_extents[i]);
3021                 error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3022                 if (error)
3023                         goto abort_error;
3024                 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3025                                          extp->ext_len);
3026         }
3027
3028         efip->efi_flags |= XFS_EFI_RECOVERED;
3029         error = xfs_trans_commit(tp, 0);
3030         return error;
3031
3032 abort_error:
3033         xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3034         return error;
3035 }
3036
3037 /*
3038  * When this is called, all of the EFIs which did not have
3039  * corresponding EFDs should be in the AIL.  What we do now
3040  * is free the extents associated with each one.
3041  *
3042  * Since we process the EFIs in normal transactions, they
3043  * will be removed at some point after the commit.  This prevents
3044  * us from just walking down the list processing each one.
3045  * We'll use a flag in the EFI to skip those that we've already
3046  * processed and use the AIL iteration mechanism's generation
3047  * count to try to speed this up at least a bit.
3048  *
3049  * When we start, we know that the EFIs are the only things in
3050  * the AIL.  As we process them, however, other items are added
3051  * to the AIL.  Since everything added to the AIL must come after
3052  * everything already in the AIL, we stop processing as soon as
3053  * we see something other than an EFI in the AIL.
3054  */
3055 STATIC int
3056 xlog_recover_process_efis(
3057         xlog_t                  *log)
3058 {
3059         xfs_log_item_t          *lip;
3060         xfs_efi_log_item_t      *efip;
3061         int                     error = 0;
3062         struct xfs_ail_cursor   cur;
3063         struct xfs_ail          *ailp;
3064
3065         ailp = log->l_ailp;
3066         spin_lock(&ailp->xa_lock);
3067         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3068         while (lip != NULL) {
3069                 /*
3070                  * We're done when we see something other than an EFI.
3071                  * There should be no EFIs left in the AIL now.
3072                  */
3073                 if (lip->li_type != XFS_LI_EFI) {
3074 #ifdef DEBUG
3075                         for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
3076                                 ASSERT(lip->li_type != XFS_LI_EFI);
3077 #endif
3078                         break;
3079                 }
3080
3081                 /*
3082                  * Skip EFIs that we've already processed.
3083                  */
3084                 efip = (xfs_efi_log_item_t *)lip;
3085                 if (efip->efi_flags & XFS_EFI_RECOVERED) {
3086                         lip = xfs_trans_ail_cursor_next(ailp, &cur);
3087                         continue;
3088                 }
3089
3090                 spin_unlock(&ailp->xa_lock);
3091                 error = xlog_recover_process_efi(log->l_mp, efip);
3092                 spin_lock(&ailp->xa_lock);
3093                 if (error)
3094                         goto out;
3095                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3096         }
3097 out:
3098         xfs_trans_ail_cursor_done(ailp, &cur);
3099         spin_unlock(&ailp->xa_lock);
3100         return error;
3101 }
3102
3103 /*
3104  * This routine performs a transaction to null out a bad inode pointer
3105  * in an agi unlinked inode hash bucket.
3106  */
3107 STATIC void
3108 xlog_recover_clear_agi_bucket(
3109         xfs_mount_t     *mp,
3110         xfs_agnumber_t  agno,
3111         int             bucket)
3112 {
3113         xfs_trans_t     *tp;
3114         xfs_agi_t       *agi;
3115         xfs_buf_t       *agibp;
3116         int             offset;
3117         int             error;
3118
3119         tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3120         error = xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp), 0, 0, 0);
3121         if (!error)
3122                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
3123                                    XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3124                                    XFS_FSS_TO_BB(mp, 1), 0, &agibp);
3125         if (error)
3126                 goto out_abort;
3127
3128         error = EINVAL;
3129         agi = XFS_BUF_TO_AGI(agibp);
3130         if (be32_to_cpu(agi->agi_magicnum) != XFS_AGI_MAGIC)
3131                 goto out_abort;
3132
3133         agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3134         offset = offsetof(xfs_agi_t, agi_unlinked) +
3135                  (sizeof(xfs_agino_t) * bucket);
3136         xfs_trans_log_buf(tp, agibp, offset,
3137                           (offset + sizeof(xfs_agino_t) - 1));
3138
3139         error = xfs_trans_commit(tp, 0);
3140         if (error)
3141                 goto out_error;
3142         return;
3143
3144 out_abort:
3145         xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3146 out_error:
3147         xfs_fs_cmn_err(CE_WARN, mp, "xlog_recover_clear_agi_bucket: "
3148                         "failed to clear agi %d. Continuing.", agno);
3149         return;
3150 }
3151
3152 /*
3153  * xlog_iunlink_recover
3154  *
3155  * This is called during recovery to process any inodes which
3156  * we unlinked but not freed when the system crashed.  These
3157  * inodes will be on the lists in the AGI blocks.  What we do
3158  * here is scan all the AGIs and fully truncate and free any
3159  * inodes found on the lists.  Each inode is removed from the
3160  * lists when it has been fully truncated and is freed.  The
3161  * freeing of the inode and its removal from the list must be
3162  * atomic.
3163  */
3164 void
3165 xlog_recover_process_iunlinks(
3166         xlog_t          *log)
3167 {
3168         xfs_mount_t     *mp;
3169         xfs_agnumber_t  agno;
3170         xfs_agi_t       *agi;
3171         xfs_buf_t       *agibp;
3172         xfs_buf_t       *ibp;
3173         xfs_dinode_t    *dip;
3174         xfs_inode_t     *ip;
3175         xfs_agino_t     agino;
3176         xfs_ino_t       ino;
3177         int             bucket;
3178         int             error;
3179         uint            mp_dmevmask;
3180
3181         mp = log->l_mp;
3182
3183         /*
3184          * Prevent any DMAPI event from being sent while in this function.
3185          */
3186         mp_dmevmask = mp->m_dmevmask;
3187         mp->m_dmevmask = 0;
3188
3189         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3190                 /*
3191                  * Find the agi for this ag.
3192                  */
3193                 agibp = xfs_buf_read(mp->m_ddev_targp,
3194                                 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3195                                 XFS_FSS_TO_BB(mp, 1), 0);
3196                 if (XFS_BUF_ISERROR(agibp)) {
3197                         xfs_ioerror_alert("xlog_recover_process_iunlinks(#1)",
3198                                 log->l_mp, agibp,
3199                                 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)));
3200                 }
3201                 agi = XFS_BUF_TO_AGI(agibp);
3202                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agi->agi_magicnum));
3203
3204                 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3205
3206                         agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3207                         while (agino != NULLAGINO) {
3208
3209                                 /*
3210                                  * Release the agi buffer so that it can
3211                                  * be acquired in the normal course of the
3212                                  * transaction to truncate and free the inode.
3213                                  */
3214                                 xfs_buf_relse(agibp);
3215
3216                                 ino = XFS_AGINO_TO_INO(mp, agno, agino);
3217                                 error = xfs_iget(mp, NULL, ino, 0, 0, &ip, 0);
3218                                 ASSERT(error || (ip != NULL));
3219
3220                                 if (!error) {
3221                                         /*
3222                                          * Get the on disk inode to find the
3223                                          * next inode in the bucket.
3224                                          */
3225                                         error = xfs_itobp(mp, NULL, ip, &dip,
3226                                                         &ibp, 0, 0,
3227                                                         XFS_BUF_LOCK);
3228                                         ASSERT(error || (dip != NULL));
3229                                 }
3230
3231                                 if (!error) {
3232                                         ASSERT(ip->i_d.di_nlink == 0);
3233
3234                                         /* setup for the next pass */
3235                                         agino = be32_to_cpu(
3236                                                         dip->di_next_unlinked);
3237                                         xfs_buf_relse(ibp);
3238                                         /*
3239                                          * Prevent any DMAPI event from
3240                                          * being sent when the
3241                                          * reference on the inode is
3242                                          * dropped.
3243                                          */
3244                                         ip->i_d.di_dmevmask = 0;
3245
3246                                         /*
3247                                          * If this is a new inode, handle
3248                                          * it specially.  Otherwise,
3249                                          * just drop our reference to the
3250                                          * inode.  If there are no
3251                                          * other references, this will
3252                                          * send the inode to
3253                                          * xfs_inactive() which will
3254                                          * truncate the file and free
3255                                          * the inode.
3256                                          */
3257                                         if (ip->i_d.di_mode == 0)
3258                                                 xfs_iput_new(ip, 0);
3259                                         else
3260                                                 IRELE(ip);
3261                                 } else {
3262                                         /*
3263                                          * We can't read in the inode
3264                                          * this bucket points to, or
3265                                          * this inode is messed up.  Just
3266                                          * ditch this bucket of inodes.  We
3267                                          * will lose some inodes and space,
3268                                          * but at least we won't hang.  Call
3269                                          * xlog_recover_clear_agi_bucket()
3270                                          * to perform a transaction to clear
3271                                          * the inode pointer in the bucket.
3272                                          */
3273                                         xlog_recover_clear_agi_bucket(mp, agno,
3274                                                         bucket);
3275
3276                                         agino = NULLAGINO;
3277                                 }
3278
3279                                 /*
3280                                  * Reacquire the agibuffer and continue around
3281                                  * the loop.
3282                                  */
3283                                 agibp = xfs_buf_read(mp->m_ddev_targp,
3284                                                 XFS_AG_DADDR(mp, agno,
3285                                                         XFS_AGI_DADDR(mp)),
3286                                                 XFS_FSS_TO_BB(mp, 1), 0);
3287                                 if (XFS_BUF_ISERROR(agibp)) {
3288                                         xfs_ioerror_alert(
3289                                 "xlog_recover_process_iunlinks(#2)",
3290                                                 log->l_mp, agibp,
3291                                                 XFS_AG_DADDR(mp, agno,
3292                                                         XFS_AGI_DADDR(mp)));
3293                                 }
3294                                 agi = XFS_BUF_TO_AGI(agibp);
3295                                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(
3296                                         agi->agi_magicnum));
3297                         }
3298                 }
3299
3300                 /*
3301                  * Release the buffer for the current agi so we can
3302                  * go on to the next one.
3303                  */
3304                 xfs_buf_relse(agibp);
3305         }
3306
3307         mp->m_dmevmask = mp_dmevmask;
3308 }
3309
3310
3311 #ifdef DEBUG
3312 STATIC void
3313 xlog_pack_data_checksum(
3314         xlog_t          *log,
3315         xlog_in_core_t  *iclog,
3316         int             size)
3317 {
3318         int             i;
3319         __be32          *up;
3320         uint            chksum = 0;
3321
3322         up = (__be32 *)iclog->ic_datap;
3323         /* divide length by 4 to get # words */
3324         for (i = 0; i < (size >> 2); i++) {
3325                 chksum ^= be32_to_cpu(*up);
3326                 up++;
3327         }
3328         iclog->ic_header.h_chksum = cpu_to_be32(chksum);
3329 }
3330 #else
3331 #define xlog_pack_data_checksum(log, iclog, size)
3332 #endif
3333
3334 /*
3335  * Stamp cycle number in every block
3336  */
3337 void
3338 xlog_pack_data(
3339         xlog_t                  *log,
3340         xlog_in_core_t          *iclog,
3341         int                     roundoff)
3342 {
3343         int                     i, j, k;
3344         int                     size = iclog->ic_offset + roundoff;
3345         __be32                  cycle_lsn;
3346         xfs_caddr_t             dp;
3347         xlog_in_core_2_t        *xhdr;
3348
3349         xlog_pack_data_checksum(log, iclog, size);
3350
3351         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3352
3353         dp = iclog->ic_datap;
3354         for (i = 0; i < BTOBB(size) &&
3355                 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3356                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
3357                 *(__be32 *)dp = cycle_lsn;
3358                 dp += BBSIZE;
3359         }
3360
3361         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3362                 xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3363                 for ( ; i < BTOBB(size); i++) {
3364                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3365                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3366                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
3367                         *(__be32 *)dp = cycle_lsn;
3368                         dp += BBSIZE;
3369                 }
3370
3371                 for (i = 1; i < log->l_iclog_heads; i++) {
3372                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3373                 }
3374         }
3375 }
3376
3377 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
3378 STATIC void
3379 xlog_unpack_data_checksum(
3380         xlog_rec_header_t       *rhead,
3381         xfs_caddr_t             dp,
3382         xlog_t                  *log)
3383 {
3384         __be32                  *up = (__be32 *)dp;
3385         uint                    chksum = 0;
3386         int                     i;
3387
3388         /* divide length by 4 to get # words */
3389         for (i=0; i < be32_to_cpu(rhead->h_len) >> 2; i++) {
3390                 chksum ^= be32_to_cpu(*up);
3391                 up++;
3392         }
3393         if (chksum != be32_to_cpu(rhead->h_chksum)) {
3394             if (rhead->h_chksum ||
3395                 ((log->l_flags & XLOG_CHKSUM_MISMATCH) == 0)) {
3396                     cmn_err(CE_DEBUG,
3397                         "XFS: LogR chksum mismatch: was (0x%x) is (0x%x)\n",
3398                             be32_to_cpu(rhead->h_chksum), chksum);
3399                     cmn_err(CE_DEBUG,
3400 "XFS: Disregard message if filesystem was created with non-DEBUG kernel");
3401                     if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3402                             cmn_err(CE_DEBUG,
3403                                 "XFS: LogR this is a LogV2 filesystem\n");
3404                     }
3405                     log->l_flags |= XLOG_CHKSUM_MISMATCH;
3406             }
3407         }
3408 }
3409 #else
3410 #define xlog_unpack_data_checksum(rhead, dp, log)
3411 #endif
3412
3413 STATIC void
3414 xlog_unpack_data(
3415         xlog_rec_header_t       *rhead,
3416         xfs_caddr_t             dp,
3417         xlog_t                  *log)
3418 {
3419         int                     i, j, k;
3420         xlog_in_core_2_t        *xhdr;
3421
3422         for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3423                   i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3424                 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3425                 dp += BBSIZE;
3426         }
3427
3428         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3429                 xhdr = (xlog_in_core_2_t *)rhead;
3430                 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3431                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3432                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3433                         *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3434                         dp += BBSIZE;
3435                 }
3436         }
3437
3438         xlog_unpack_data_checksum(rhead, dp, log);
3439 }
3440
3441 STATIC int
3442 xlog_valid_rec_header(
3443         xlog_t                  *log,
3444         xlog_rec_header_t       *rhead,
3445         xfs_daddr_t             blkno)
3446 {
3447         int                     hlen;
3448
3449         if (unlikely(be32_to_cpu(rhead->h_magicno) != XLOG_HEADER_MAGIC_NUM)) {
3450                 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3451                                 XFS_ERRLEVEL_LOW, log->l_mp);
3452                 return XFS_ERROR(EFSCORRUPTED);
3453         }
3454         if (unlikely(
3455             (!rhead->h_version ||
3456             (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
3457                 xlog_warn("XFS: %s: unrecognised log version (%d).",
3458                         __func__, be32_to_cpu(rhead->h_version));
3459                 return XFS_ERROR(EIO);
3460         }
3461
3462         /* LR body must have data or it wouldn't have been written */
3463         hlen = be32_to_cpu(rhead->h_len);
3464         if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3465                 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3466                                 XFS_ERRLEVEL_LOW, log->l_mp);
3467                 return XFS_ERROR(EFSCORRUPTED);
3468         }
3469         if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3470                 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3471                                 XFS_ERRLEVEL_LOW, log->l_mp);
3472                 return XFS_ERROR(EFSCORRUPTED);
3473         }
3474         return 0;
3475 }
3476
3477 /*
3478  * Read the log from tail to head and process the log records found.
3479  * Handle the two cases where the tail and head are in the same cycle
3480  * and where the active portion of the log wraps around the end of
3481  * the physical log separately.  The pass parameter is passed through
3482  * to the routines called to process the data and is not looked at
3483  * here.
3484  */
3485 STATIC int
3486 xlog_do_recovery_pass(
3487         xlog_t                  *log,
3488         xfs_daddr_t             head_blk,
3489         xfs_daddr_t             tail_blk,
3490         int                     pass)
3491 {
3492         xlog_rec_header_t       *rhead;
3493         xfs_daddr_t             blk_no;
3494         xfs_caddr_t             bufaddr, offset;
3495         xfs_buf_t               *hbp, *dbp;
3496         int                     error = 0, h_size;
3497         int                     bblks, split_bblks;
3498         int                     hblks, split_hblks, wrapped_hblks;
3499         xlog_recover_t          *rhash[XLOG_RHASH_SIZE];
3500
3501         ASSERT(head_blk != tail_blk);
3502
3503         /*
3504          * Read the header of the tail block and get the iclog buffer size from
3505          * h_size.  Use this to tell how many sectors make up the log header.
3506          */
3507         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3508                 /*
3509                  * When using variable length iclogs, read first sector of
3510                  * iclog header and extract the header size from it.  Get a
3511                  * new hbp that is the correct size.
3512                  */
3513                 hbp = xlog_get_bp(log, 1);
3514                 if (!hbp)
3515                         return ENOMEM;
3516                 if ((error = xlog_bread(log, tail_blk, 1, hbp)))
3517                         goto bread_err1;
3518                 offset = xlog_align(log, tail_blk, 1, hbp);
3519                 rhead = (xlog_rec_header_t *)offset;
3520                 error = xlog_valid_rec_header(log, rhead, tail_blk);
3521                 if (error)
3522                         goto bread_err1;
3523                 h_size = be32_to_cpu(rhead->h_size);
3524                 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3525                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3526                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3527                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
3528                                 hblks++;
3529                         xlog_put_bp(hbp);
3530                         hbp = xlog_get_bp(log, hblks);
3531                 } else {
3532                         hblks = 1;
3533                 }
3534         } else {
3535                 ASSERT(log->l_sectbb_log == 0);
3536                 hblks = 1;
3537                 hbp = xlog_get_bp(log, 1);
3538                 h_size = XLOG_BIG_RECORD_BSIZE;
3539         }
3540
3541         if (!hbp)
3542                 return ENOMEM;
3543         dbp = xlog_get_bp(log, BTOBB(h_size));
3544         if (!dbp) {
3545                 xlog_put_bp(hbp);
3546                 return ENOMEM;
3547         }
3548
3549         memset(rhash, 0, sizeof(rhash));
3550         if (tail_blk <= head_blk) {
3551                 for (blk_no = tail_blk; blk_no < head_blk; ) {
3552                         if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3553                                 goto bread_err2;
3554                         offset = xlog_align(log, blk_no, hblks, hbp);
3555                         rhead = (xlog_rec_header_t *)offset;
3556                         error = xlog_valid_rec_header(log, rhead, blk_no);
3557                         if (error)
3558                                 goto bread_err2;
3559
3560                         /* blocks in data section */
3561                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3562                         error = xlog_bread(log, blk_no + hblks, bblks, dbp);
3563                         if (error)
3564                                 goto bread_err2;
3565                         offset = xlog_align(log, blk_no + hblks, bblks, dbp);
3566                         xlog_unpack_data(rhead, offset, log);
3567                         if ((error = xlog_recover_process_data(log,
3568                                                 rhash, rhead, offset, pass)))
3569                                 goto bread_err2;
3570                         blk_no += bblks + hblks;
3571                 }
3572         } else {
3573                 /*
3574                  * Perform recovery around the end of the physical log.
3575                  * When the head is not on the same cycle number as the tail,
3576                  * we can't do a sequential recovery as above.
3577                  */
3578                 blk_no = tail_blk;
3579                 while (blk_no < log->l_logBBsize) {
3580                         /*
3581                          * Check for header wrapping around physical end-of-log
3582                          */
3583                         offset = NULL;
3584                         split_hblks = 0;
3585                         wrapped_hblks = 0;
3586                         if (blk_no + hblks <= log->l_logBBsize) {
3587                                 /* Read header in one read */
3588                                 error = xlog_bread(log, blk_no, hblks, hbp);
3589                                 if (error)
3590                                         goto bread_err2;
3591                                 offset = xlog_align(log, blk_no, hblks, hbp);
3592                         } else {
3593                                 /* This LR is split across physical log end */
3594                                 if (blk_no != log->l_logBBsize) {
3595                                         /* some data before physical log end */
3596                                         ASSERT(blk_no <= INT_MAX);
3597                                         split_hblks = log->l_logBBsize - (int)blk_no;
3598                                         ASSERT(split_hblks > 0);
3599                                         if ((error = xlog_bread(log, blk_no,
3600                                                         split_hblks, hbp)))
3601                                                 goto bread_err2;
3602                                         offset = xlog_align(log, blk_no,
3603                                                         split_hblks, hbp);
3604                                 }
3605                                 /*
3606                                  * Note: this black magic still works with
3607                                  * large sector sizes (non-512) only because:
3608                                  * - we increased the buffer size originally
3609                                  *   by 1 sector giving us enough extra space
3610                                  *   for the second read;
3611                                  * - the log start is guaranteed to be sector
3612                                  *   aligned;
3613                                  * - we read the log end (LR header start)
3614                                  *   _first_, then the log start (LR header end)
3615                                  *   - order is important.
3616                                  */
3617                                 wrapped_hblks = hblks - split_hblks;
3618                                 bufaddr = XFS_BUF_PTR(hbp);
3619                                 error = XFS_BUF_SET_PTR(hbp,
3620                                                 bufaddr + BBTOB(split_hblks),
3621                                                 BBTOB(hblks - split_hblks));
3622                                 if (!error)
3623                                         error = xlog_bread(log, 0,
3624                                                         wrapped_hblks, hbp);
3625                                 if (!error)
3626                                         error = XFS_BUF_SET_PTR(hbp, bufaddr,
3627                                                         BBTOB(hblks));
3628                                 if (error)
3629                                         goto bread_err2;
3630                                 if (!offset)
3631                                         offset = xlog_align(log, 0,
3632                                                         wrapped_hblks, hbp);
3633                         }
3634                         rhead = (xlog_rec_header_t *)offset;
3635                         error = xlog_valid_rec_header(log, rhead,
3636                                                 split_hblks ? blk_no : 0);
3637                         if (error)
3638                                 goto bread_err2;
3639
3640                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3641                         blk_no += hblks;
3642
3643                         /* Read in data for log record */
3644                         if (blk_no + bblks <= log->l_logBBsize) {
3645                                 error = xlog_bread(log, blk_no, bblks, dbp);
3646                                 if (error)
3647                                         goto bread_err2;
3648                                 offset = xlog_align(log, blk_no, bblks, dbp);
3649                         } else {
3650                                 /* This log record is split across the
3651                                  * physical end of log */
3652                                 offset = NULL;
3653                                 split_bblks = 0;
3654                                 if (blk_no != log->l_logBBsize) {
3655                                         /* some data is before the physical
3656                                          * end of log */
3657                                         ASSERT(!wrapped_hblks);
3658                                         ASSERT(blk_no <= INT_MAX);
3659                                         split_bblks =
3660                                                 log->l_logBBsize - (int)blk_no;
3661                                         ASSERT(split_bblks > 0);
3662                                         if ((error = xlog_bread(log, blk_no,
3663                                                         split_bblks, dbp)))
3664                                                 goto bread_err2;
3665                                         offset = xlog_align(log, blk_no,
3666                                                         split_bblks, dbp);
3667                                 }
3668                                 /*
3669                                  * Note: this black magic still works with
3670                                  * large sector sizes (non-512) only because:
3671                                  * - we increased the buffer size originally
3672                                  *   by 1 sector giving us enough extra space
3673                                  *   for the second read;
3674                                  * - the log start is guaranteed to be sector
3675                                  *   aligned;
3676                                  * - we read the log end (LR header start)
3677                                  *   _first_, then the log start (LR header end)
3678                                  *   - order is important.
3679                                  */
3680                                 bufaddr = XFS_BUF_PTR(dbp);
3681                                 error = XFS_BUF_SET_PTR(dbp,
3682                                                 bufaddr + BBTOB(split_bblks),
3683                                                 BBTOB(bblks - split_bblks));
3684                                 if (!error)
3685                                         error = xlog_bread(log, wrapped_hblks,
3686                                                         bblks - split_bblks,
3687                                                         dbp);
3688                                 if (!error)
3689                                         error = XFS_BUF_SET_PTR(dbp, bufaddr,
3690                                                         h_size);
3691                                 if (error)
3692                                         goto bread_err2;
3693                                 if (!offset)
3694                                         offset = xlog_align(log, wrapped_hblks,
3695                                                 bblks - split_bblks, dbp);
3696                         }
3697                         xlog_unpack_data(rhead, offset, log);
3698                         if ((error = xlog_recover_process_data(log, rhash,
3699                                                         rhead, offset, pass)))
3700                                 goto bread_err2;
3701                         blk_no += bblks;
3702                 }
3703
3704                 ASSERT(blk_no >= log->l_logBBsize);
3705                 blk_no -= log->l_logBBsize;
3706
3707                 /* read first part of physical log */
3708                 while (blk_no < head_blk) {
3709                         if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3710                                 goto bread_err2;
3711                         offset = xlog_align(log, blk_no, hblks, hbp);
3712                         rhead = (xlog_rec_header_t *)offset;
3713                         error = xlog_valid_rec_header(log, rhead, blk_no);
3714                         if (error)
3715                                 goto bread_err2;
3716                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3717                         if ((error = xlog_bread(log, blk_no+hblks, bblks, dbp)))
3718                                 goto bread_err2;
3719                         offset = xlog_align(log, blk_no+hblks, bblks, dbp);
3720                         xlog_unpack_data(rhead, offset, log);
3721                         if ((error = xlog_recover_process_data(log, rhash,
3722                                                         rhead, offset, pass)))
3723                                 goto bread_err2;
3724                         blk_no += bblks + hblks;
3725                 }
3726         }
3727
3728  bread_err2:
3729         xlog_put_bp(dbp);
3730  bread_err1:
3731         xlog_put_bp(hbp);
3732         return error;
3733 }
3734
3735 /*
3736  * Do the recovery of the log.  We actually do this in two phases.
3737  * The two passes are necessary in order to implement the function
3738  * of cancelling a record written into the log.  The first pass
3739  * determines those things which have been cancelled, and the
3740  * second pass replays log items normally except for those which
3741  * have been cancelled.  The handling of the replay and cancellations
3742  * takes place in the log item type specific routines.
3743  *
3744  * The table of items which have cancel records in the log is allocated
3745  * and freed at this level, since only here do we know when all of
3746  * the log recovery has been completed.
3747  */
3748 STATIC int
3749 xlog_do_log_recovery(
3750         xlog_t          *log,
3751         xfs_daddr_t     head_blk,
3752         xfs_daddr_t     tail_blk)
3753 {
3754         int             error;
3755
3756         ASSERT(head_blk != tail_blk);
3757
3758         /*
3759          * First do a pass to find all of the cancelled buf log items.
3760          * Store them in the buf_cancel_table for use in the second pass.
3761          */
3762         log->l_buf_cancel_table =
3763                 (xfs_buf_cancel_t **)kmem_zalloc(XLOG_BC_TABLE_SIZE *
3764                                                  sizeof(xfs_buf_cancel_t*),
3765                                                  KM_SLEEP);
3766         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3767                                       XLOG_RECOVER_PASS1);
3768         if (error != 0) {
3769                 kmem_free(log->l_buf_cancel_table);
3770                 log->l_buf_cancel_table = NULL;
3771                 return error;
3772         }
3773         /*
3774          * Then do a second pass to actually recover the items in the log.
3775          * When it is complete free the table of buf cancel items.
3776          */
3777         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3778                                       XLOG_RECOVER_PASS2);
3779 #ifdef DEBUG
3780         if (!error) {
3781                 int     i;
3782
3783                 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3784                         ASSERT(log->l_buf_cancel_table[i] == NULL);
3785         }
3786 #endif  /* DEBUG */
3787
3788         kmem_free(log->l_buf_cancel_table);
3789         log->l_buf_cancel_table = NULL;
3790
3791         return error;
3792 }
3793
3794 /*
3795  * Do the actual recovery
3796  */
3797 STATIC int
3798 xlog_do_recover(
3799         xlog_t          *log,
3800         xfs_daddr_t     head_blk,
3801         xfs_daddr_t     tail_blk)
3802 {
3803         int             error;
3804         xfs_buf_t       *bp;
3805         xfs_sb_t        *sbp;
3806
3807         /*
3808          * First replay the images in the log.
3809          */
3810         error = xlog_do_log_recovery(log, head_blk, tail_blk);
3811         if (error) {
3812                 return error;
3813         }
3814
3815         XFS_bflush(log->l_mp->m_ddev_targp);
3816
3817         /*
3818          * If IO errors happened during recovery, bail out.
3819          */
3820         if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3821                 return (EIO);
3822         }
3823
3824         /*
3825          * We now update the tail_lsn since much of the recovery has completed
3826          * and there may be space available to use.  If there were no extent
3827          * or iunlinks, we can free up the entire log and set the tail_lsn to
3828          * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3829          * lsn of the last known good LR on disk.  If there are extent frees
3830          * or iunlinks they will have some entries in the AIL; so we look at
3831          * the AIL to determine how to set the tail_lsn.
3832          */
3833         xlog_assign_tail_lsn(log->l_mp);
3834
3835         /*
3836          * Now that we've finished replaying all buffer and inode
3837          * updates, re-read in the superblock.
3838          */
3839         bp = xfs_getsb(log->l_mp, 0);
3840         XFS_BUF_UNDONE(bp);
3841         ASSERT(!(XFS_BUF_ISWRITE(bp)));
3842         ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
3843         XFS_BUF_READ(bp);
3844         XFS_BUF_UNASYNC(bp);
3845         xfsbdstrat(log->l_mp, bp);
3846         error = xfs_iowait(bp);
3847         if (error) {
3848                 xfs_ioerror_alert("xlog_do_recover",
3849                                   log->l_mp, bp, XFS_BUF_ADDR(bp));
3850                 ASSERT(0);
3851                 xfs_buf_relse(bp);
3852                 return error;
3853         }
3854
3855         /* Convert superblock from on-disk format */
3856         sbp = &log->l_mp->m_sb;
3857         xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
3858         ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3859         ASSERT(xfs_sb_good_version(sbp));
3860         xfs_buf_relse(bp);
3861
3862         /* We've re-read the superblock so re-initialize per-cpu counters */
3863         xfs_icsb_reinit_counters(log->l_mp);
3864
3865         xlog_recover_check_summary(log);
3866
3867         /* Normal transactions can now occur */
3868         log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3869         return 0;
3870 }
3871
3872 /*
3873  * Perform recovery and re-initialize some log variables in xlog_find_tail.
3874  *
3875  * Return error or zero.
3876  */
3877 int
3878 xlog_recover(
3879         xlog_t          *log)
3880 {
3881         xfs_daddr_t     head_blk, tail_blk;
3882         int             error;
3883
3884         /* find the tail of the log */
3885         if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3886                 return error;
3887
3888         if (tail_blk != head_blk) {
3889                 /* There used to be a comment here:
3890                  *
3891                  * disallow recovery on read-only mounts.  note -- mount
3892                  * checks for ENOSPC and turns it into an intelligent
3893                  * error message.
3894                  * ...but this is no longer true.  Now, unless you specify
3895                  * NORECOVERY (in which case this function would never be
3896                  * called), we just go ahead and recover.  We do this all
3897                  * under the vfs layer, so we can get away with it unless
3898                  * the device itself is read-only, in which case we fail.
3899                  */
3900                 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3901                         return error;
3902                 }
3903
3904                 cmn_err(CE_NOTE,
3905                         "Starting XFS recovery on filesystem: %s (logdev: %s)",
3906                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3907                         log->l_mp->m_logname : "internal");
3908
3909                 error = xlog_do_recover(log, head_blk, tail_blk);
3910                 log->l_flags |= XLOG_RECOVERY_NEEDED;
3911         }
3912         return error;
3913 }
3914
3915 /*
3916  * In the first part of recovery we replay inodes and buffers and build
3917  * up the list of extent free items which need to be processed.  Here
3918  * we process the extent free items and clean up the on disk unlinked
3919  * inode lists.  This is separated from the first part of recovery so
3920  * that the root and real-time bitmap inodes can be read in from disk in
3921  * between the two stages.  This is necessary so that we can free space
3922  * in the real-time portion of the file system.
3923  */
3924 int
3925 xlog_recover_finish(
3926         xlog_t          *log)
3927 {
3928         /*
3929          * Now we're ready to do the transactions needed for the
3930          * rest of recovery.  Start with completing all the extent
3931          * free intent records and then process the unlinked inode
3932          * lists.  At this point, we essentially run in normal mode
3933          * except that we're still performing recovery actions
3934          * rather than accepting new requests.
3935          */
3936         if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3937                 int     error;
3938                 error = xlog_recover_process_efis(log);
3939                 if (error) {
3940                         cmn_err(CE_ALERT,
3941                                 "Failed to recover EFIs on filesystem: %s",
3942                                 log->l_mp->m_fsname);
3943                         return error;
3944                 }
3945                 /*
3946                  * Sync the log to get all the EFIs out of the AIL.
3947                  * This isn't absolutely necessary, but it helps in
3948                  * case the unlink transactions would have problems
3949                  * pushing the EFIs out of the way.
3950                  */
3951                 xfs_log_force(log->l_mp, (xfs_lsn_t)0,
3952                               (XFS_LOG_FORCE | XFS_LOG_SYNC));
3953
3954                 xlog_recover_process_iunlinks(log);
3955
3956                 xlog_recover_check_summary(log);
3957
3958                 cmn_err(CE_NOTE,
3959                         "Ending XFS recovery on filesystem: %s (logdev: %s)",
3960                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3961                         log->l_mp->m_logname : "internal");
3962                 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3963         } else {
3964                 cmn_err(CE_DEBUG,
3965                         "!Ending clean XFS mount for filesystem: %s\n",
3966                         log->l_mp->m_fsname);
3967         }
3968         return 0;
3969 }
3970
3971
3972 #if defined(DEBUG)
3973 /*
3974  * Read all of the agf and agi counters and check that they
3975  * are consistent with the superblock counters.
3976  */
3977 void
3978 xlog_recover_check_summary(
3979         xlog_t          *log)
3980 {
3981         xfs_mount_t     *mp;
3982         xfs_agf_t       *agfp;
3983         xfs_agi_t       *agip;
3984         xfs_buf_t       *agfbp;
3985         xfs_buf_t       *agibp;
3986         xfs_daddr_t     agfdaddr;
3987         xfs_daddr_t     agidaddr;
3988         xfs_buf_t       *sbbp;
3989 #ifdef XFS_LOUD_RECOVERY
3990         xfs_sb_t        *sbp;
3991 #endif
3992         xfs_agnumber_t  agno;
3993         __uint64_t      freeblks;
3994         __uint64_t      itotal;
3995         __uint64_t      ifree;
3996
3997         mp = log->l_mp;
3998
3999         freeblks = 0LL;
4000         itotal = 0LL;
4001         ifree = 0LL;
4002         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4003                 agfdaddr = XFS_AG_DADDR(mp, agno, XFS_AGF_DADDR(mp));
4004                 agfbp = xfs_buf_read(mp->m_ddev_targp, agfdaddr,
4005                                 XFS_FSS_TO_BB(mp, 1), 0);
4006                 if (XFS_BUF_ISERROR(agfbp)) {
4007                         xfs_ioerror_alert("xlog_recover_check_summary(agf)",
4008                                                 mp, agfbp, agfdaddr);
4009                 }
4010                 agfp = XFS_BUF_TO_AGF(agfbp);
4011                 ASSERT(XFS_AGF_MAGIC == be32_to_cpu(agfp->agf_magicnum));
4012                 ASSERT(XFS_AGF_GOOD_VERSION(be32_to_cpu(agfp->agf_versionnum)));
4013                 ASSERT(be32_to_cpu(agfp->agf_seqno) == agno);
4014
4015                 freeblks += be32_to_cpu(agfp->agf_freeblks) +
4016                             be32_to_cpu(agfp->agf_flcount);
4017                 xfs_buf_relse(agfbp);
4018
4019                 agidaddr = XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp));
4020                 agibp = xfs_buf_read(mp->m_ddev_targp, agidaddr,
4021                                 XFS_FSS_TO_BB(mp, 1), 0);
4022                 if (XFS_BUF_ISERROR(agibp)) {
4023                         xfs_ioerror_alert("xlog_recover_check_summary(agi)",
4024                                           mp, agibp, agidaddr);
4025                 }
4026                 agip = XFS_BUF_TO_AGI(agibp);
4027                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agip->agi_magicnum));
4028                 ASSERT(XFS_AGI_GOOD_VERSION(be32_to_cpu(agip->agi_versionnum)));
4029                 ASSERT(be32_to_cpu(agip->agi_seqno) == agno);
4030
4031                 itotal += be32_to_cpu(agip->agi_count);
4032                 ifree += be32_to_cpu(agip->agi_freecount);
4033                 xfs_buf_relse(agibp);
4034         }
4035
4036         sbbp = xfs_getsb(mp, 0);
4037 #ifdef XFS_LOUD_RECOVERY
4038         sbp = &mp->m_sb;
4039         xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(sbbp));
4040         cmn_err(CE_NOTE,
4041                 "xlog_recover_check_summary: sb_icount %Lu itotal %Lu",
4042                 sbp->sb_icount, itotal);
4043         cmn_err(CE_NOTE,
4044                 "xlog_recover_check_summary: sb_ifree %Lu itotal %Lu",
4045                 sbp->sb_ifree, ifree);
4046         cmn_err(CE_NOTE,
4047                 "xlog_recover_check_summary: sb_fdblocks %Lu freeblks %Lu",
4048                 sbp->sb_fdblocks, freeblks);
4049 #if 0
4050         /*
4051          * This is turned off until I account for the allocation
4052          * btree blocks which live in free space.
4053          */
4054         ASSERT(sbp->sb_icount == itotal);
4055         ASSERT(sbp->sb_ifree == ifree);
4056         ASSERT(sbp->sb_fdblocks == freeblks);
4057 #endif
4058 #endif
4059         xfs_buf_relse(sbbp);
4060 }
4061 #endif /* DEBUG */