Merge branch 'for-linus' of git://oss.sgi.com:8090/xfs/xfs-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
50 STATIC int      xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
51 STATIC int      xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
52 STATIC void     xlog_recover_insert_item_backq(xlog_recover_item_t **q,
53                                                xlog_recover_item_t *item);
54 #if defined(DEBUG)
55 STATIC void     xlog_recover_check_summary(xlog_t *);
56 STATIC void     xlog_recover_check_ail(xfs_mount_t *, xfs_log_item_t *, int);
57 #else
58 #define xlog_recover_check_summary(log)
59 #define xlog_recover_check_ail(mp, lip, gen)
60 #endif
61
62
63 /*
64  * Sector aligned buffer routines for buffer create/read/write/access
65  */
66
67 #define XLOG_SECTOR_ROUNDUP_BBCOUNT(log, bbs)   \
68         ( ((log)->l_sectbb_mask && (bbs & (log)->l_sectbb_mask)) ? \
69         ((bbs + (log)->l_sectbb_mask + 1) & ~(log)->l_sectbb_mask) : (bbs) )
70 #define XLOG_SECTOR_ROUNDDOWN_BLKNO(log, bno)   ((bno) & ~(log)->l_sectbb_mask)
71
72 xfs_buf_t *
73 xlog_get_bp(
74         xlog_t          *log,
75         int             num_bblks)
76 {
77         ASSERT(num_bblks > 0);
78
79         if (log->l_sectbb_log) {
80                 if (num_bblks > 1)
81                         num_bblks += XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
82                 num_bblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, num_bblks);
83         }
84         return xfs_buf_get_noaddr(BBTOB(num_bblks), log->l_mp->m_logdev_targp);
85 }
86
87 void
88 xlog_put_bp(
89         xfs_buf_t       *bp)
90 {
91         xfs_buf_free(bp);
92 }
93
94
95 /*
96  * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
97  */
98 int
99 xlog_bread(
100         xlog_t          *log,
101         xfs_daddr_t     blk_no,
102         int             nbblks,
103         xfs_buf_t       *bp)
104 {
105         int             error;
106
107         if (log->l_sectbb_log) {
108                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
109                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
110         }
111
112         ASSERT(nbblks > 0);
113         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
114         ASSERT(bp);
115
116         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
117         XFS_BUF_READ(bp);
118         XFS_BUF_BUSY(bp);
119         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
120         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
121
122         xfsbdstrat(log->l_mp, bp);
123         if ((error = xfs_iowait(bp)))
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 = ", __FUNCTION__);
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", INT_GET(head->h_fmt, ARCH_CONVERT));
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(INT_GET(head->h_magicno, ARCH_CONVERT) == 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(INT_GET(head->h_fmt, ARCH_CONVERT) != 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(INT_GET(head->h_magicno, ARCH_CONVERT) == 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 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 = GET_CYCLE(offset, ARCH_CONVERT);
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 = GET_CYCLE(buf, ARCH_CONVERT);
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 ==
451                     INT_GET(head->h_magicno, ARCH_CONVERT))
452                         break;
453
454                 if (!smallmem)
455                         offset -= BBSIZE;
456         }
457
458         /*
459          * We hit the beginning of the physical log & still no header.  Return
460          * to caller.  If caller can handle a return of -1, then this routine
461          * will be called again for the end of the physical log.
462          */
463         if (i == -1) {
464                 error = -1;
465                 goto out;
466         }
467
468         /*
469          * We have the final block of the good log (the first block
470          * of the log record _before_ the head. So we check the uuid.
471          */
472         if ((error = xlog_header_check_mount(log->l_mp, head)))
473                 goto out;
474
475         /*
476          * We may have found a log record header before we expected one.
477          * last_blk will be the 1st block # with a given cycle #.  We may end
478          * up reading an entire log record.  In this case, we don't want to
479          * reset last_blk.  Only when last_blk points in the middle of a log
480          * record do we update last_blk.
481          */
482         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
483                 uint    h_size = INT_GET(head->h_size, ARCH_CONVERT);
484
485                 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
486                 if (h_size % XLOG_HEADER_CYCLE_SIZE)
487                         xhdrs++;
488         } else {
489                 xhdrs = 1;
490         }
491
492         if (*last_blk - i + extra_bblks
493                         != BTOBB(INT_GET(head->h_len, ARCH_CONVERT)) + xhdrs)
494                 *last_blk = i;
495
496 out:
497         xlog_put_bp(bp);
498         return error;
499 }
500
501 /*
502  * Head is defined to be the point of the log where the next log write
503  * write could go.  This means that incomplete LR writes at the end are
504  * eliminated when calculating the head.  We aren't guaranteed that previous
505  * LR have complete transactions.  We only know that a cycle number of
506  * current cycle number -1 won't be present in the log if we start writing
507  * from our current block number.
508  *
509  * last_blk contains the block number of the first block with a given
510  * cycle number.
511  *
512  * Return: zero if normal, non-zero if error.
513  */
514 STATIC int
515 xlog_find_head(
516         xlog_t          *log,
517         xfs_daddr_t     *return_head_blk)
518 {
519         xfs_buf_t       *bp;
520         xfs_caddr_t     offset;
521         xfs_daddr_t     new_blk, first_blk, start_blk, last_blk, head_blk;
522         int             num_scan_bblks;
523         uint            first_half_cycle, last_half_cycle;
524         uint            stop_on_cycle;
525         int             error, log_bbnum = log->l_logBBsize;
526
527         /* Is the end of the log device zeroed? */
528         if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
529                 *return_head_blk = first_blk;
530
531                 /* Is the whole lot zeroed? */
532                 if (!first_blk) {
533                         /* Linux XFS shouldn't generate totally zeroed logs -
534                          * mkfs etc write a dummy unmount record to a fresh
535                          * log so we can store the uuid in there
536                          */
537                         xlog_warn("XFS: totally zeroed log");
538                 }
539
540                 return 0;
541         } else if (error) {
542                 xlog_warn("XFS: empty log check failed");
543                 return error;
544         }
545
546         first_blk = 0;                  /* get cycle # of 1st block */
547         bp = xlog_get_bp(log, 1);
548         if (!bp)
549                 return ENOMEM;
550         if ((error = xlog_bread(log, 0, 1, bp)))
551                 goto bp_err;
552         offset = xlog_align(log, 0, 1, bp);
553         first_half_cycle = GET_CYCLE(offset, ARCH_CONVERT);
554
555         last_blk = head_blk = log_bbnum - 1;    /* get cycle # of last block */
556         if ((error = xlog_bread(log, last_blk, 1, bp)))
557                 goto bp_err;
558         offset = xlog_align(log, last_blk, 1, bp);
559         last_half_cycle = GET_CYCLE(offset, ARCH_CONVERT);
560         ASSERT(last_half_cycle != 0);
561
562         /*
563          * If the 1st half cycle number is equal to the last half cycle number,
564          * then the entire log is stamped with the same cycle number.  In this
565          * case, head_blk can't be set to zero (which makes sense).  The below
566          * math doesn't work out properly with head_blk equal to zero.  Instead,
567          * we set it to log_bbnum which is an invalid block number, but this
568          * value makes the math correct.  If head_blk doesn't changed through
569          * all the tests below, *head_blk is set to zero at the very end rather
570          * than log_bbnum.  In a sense, log_bbnum and zero are the same block
571          * in a circular file.
572          */
573         if (first_half_cycle == last_half_cycle) {
574                 /*
575                  * In this case we believe that the entire log should have
576                  * cycle number last_half_cycle.  We need to scan backwards
577                  * from the end verifying that there are no holes still
578                  * containing last_half_cycle - 1.  If we find such a hole,
579                  * then the start of that hole will be the new head.  The
580                  * simple case looks like
581                  *        x | x ... | x - 1 | x
582                  * Another case that fits this picture would be
583                  *        x | x + 1 | x ... | x
584                  * In this case the head really is somewhere at the end of the
585                  * log, as one of the latest writes at the beginning was
586                  * incomplete.
587                  * One more case is
588                  *        x | x + 1 | x ... | x - 1 | x
589                  * This is really the combination of the above two cases, and
590                  * the head has to end up at the start of the x-1 hole at the
591                  * end of the log.
592                  *
593                  * In the 256k log case, we will read from the beginning to the
594                  * end of the log and search for cycle numbers equal to x-1.
595                  * We don't worry about the x+1 blocks that we encounter,
596                  * because we know that they cannot be the head since the log
597                  * started with x.
598                  */
599                 head_blk = log_bbnum;
600                 stop_on_cycle = last_half_cycle - 1;
601         } else {
602                 /*
603                  * In this case we want to find the first block with cycle
604                  * number matching last_half_cycle.  We expect the log to be
605                  * some variation on
606                  *        x + 1 ... | x ...
607                  * The first block with cycle number x (last_half_cycle) will
608                  * be where the new head belongs.  First we do a binary search
609                  * for the first occurrence of last_half_cycle.  The binary
610                  * search may not be totally accurate, so then we scan back
611                  * from there looking for occurrences of last_half_cycle before
612                  * us.  If that backwards scan wraps around the beginning of
613                  * the log, then we look for occurrences of last_half_cycle - 1
614                  * at the end of the log.  The cases we're looking for look
615                  * like
616                  *        x + 1 ... | x | x + 1 | x ...
617                  *                               ^ binary search stopped here
618                  * or
619                  *        x + 1 ... | x ... | x - 1 | x
620                  *        <---------> less than scan distance
621                  */
622                 stop_on_cycle = last_half_cycle;
623                 if ((error = xlog_find_cycle_start(log, bp, first_blk,
624                                                 &head_blk, last_half_cycle)))
625                         goto bp_err;
626         }
627
628         /*
629          * Now validate the answer.  Scan back some number of maximum possible
630          * blocks and make sure each one has the expected cycle number.  The
631          * maximum is determined by the total possible amount of buffering
632          * in the in-core log.  The following number can be made tighter if
633          * we actually look at the block size of the filesystem.
634          */
635         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
636         if (head_blk >= num_scan_bblks) {
637                 /*
638                  * We are guaranteed that the entire check can be performed
639                  * in one buffer.
640                  */
641                 start_blk = head_blk - num_scan_bblks;
642                 if ((error = xlog_find_verify_cycle(log,
643                                                 start_blk, num_scan_bblks,
644                                                 stop_on_cycle, &new_blk)))
645                         goto bp_err;
646                 if (new_blk != -1)
647                         head_blk = new_blk;
648         } else {                /* need to read 2 parts of log */
649                 /*
650                  * We are going to scan backwards in the log in two parts.
651                  * First we scan the physical end of the log.  In this part
652                  * of the log, we are looking for blocks with cycle number
653                  * last_half_cycle - 1.
654                  * If we find one, then we know that the log starts there, as
655                  * we've found a hole that didn't get written in going around
656                  * the end of the physical log.  The simple case for this is
657                  *        x + 1 ... | x ... | x - 1 | x
658                  *        <---------> less than scan distance
659                  * If all of the blocks at the end of the log have cycle number
660                  * last_half_cycle, then we check the blocks at the start of
661                  * the log looking for occurrences of last_half_cycle.  If we
662                  * find one, then our current estimate for the location of the
663                  * first occurrence of last_half_cycle is wrong and we move
664                  * back to the hole we've found.  This case looks like
665                  *        x + 1 ... | x | x + 1 | x ...
666                  *                               ^ binary search stopped here
667                  * Another case we need to handle that only occurs in 256k
668                  * logs is
669                  *        x + 1 ... | x ... | x+1 | x ...
670                  *                   ^ binary search stops here
671                  * In a 256k log, the scan at the end of the log will see the
672                  * x + 1 blocks.  We need to skip past those since that is
673                  * certainly not the head of the log.  By searching for
674                  * last_half_cycle-1 we accomplish that.
675                  */
676                 start_blk = log_bbnum - num_scan_bblks + head_blk;
677                 ASSERT(head_blk <= INT_MAX &&
678                         (xfs_daddr_t) num_scan_bblks - head_blk >= 0);
679                 if ((error = xlog_find_verify_cycle(log, start_blk,
680                                         num_scan_bblks - (int)head_blk,
681                                         (stop_on_cycle - 1), &new_blk)))
682                         goto bp_err;
683                 if (new_blk != -1) {
684                         head_blk = new_blk;
685                         goto bad_blk;
686                 }
687
688                 /*
689                  * Scan beginning of log now.  The last part of the physical
690                  * log is good.  This scan needs to verify that it doesn't find
691                  * the last_half_cycle.
692                  */
693                 start_blk = 0;
694                 ASSERT(head_blk <= INT_MAX);
695                 if ((error = xlog_find_verify_cycle(log,
696                                         start_blk, (int)head_blk,
697                                         stop_on_cycle, &new_blk)))
698                         goto bp_err;
699                 if (new_blk != -1)
700                         head_blk = new_blk;
701         }
702
703  bad_blk:
704         /*
705          * Now we need to make sure head_blk is not pointing to a block in
706          * the middle of a log record.
707          */
708         num_scan_bblks = XLOG_REC_SHIFT(log);
709         if (head_blk >= num_scan_bblks) {
710                 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
711
712                 /* start ptr at last block ptr before head_blk */
713                 if ((error = xlog_find_verify_log_record(log, start_blk,
714                                                         &head_blk, 0)) == -1) {
715                         error = XFS_ERROR(EIO);
716                         goto bp_err;
717                 } else if (error)
718                         goto bp_err;
719         } else {
720                 start_blk = 0;
721                 ASSERT(head_blk <= INT_MAX);
722                 if ((error = xlog_find_verify_log_record(log, start_blk,
723                                                         &head_blk, 0)) == -1) {
724                         /* We hit the beginning of the log during our search */
725                         start_blk = log_bbnum - num_scan_bblks + head_blk;
726                         new_blk = log_bbnum;
727                         ASSERT(start_blk <= INT_MAX &&
728                                 (xfs_daddr_t) log_bbnum-start_blk >= 0);
729                         ASSERT(head_blk <= INT_MAX);
730                         if ((error = xlog_find_verify_log_record(log,
731                                                         start_blk, &new_blk,
732                                                         (int)head_blk)) == -1) {
733                                 error = XFS_ERROR(EIO);
734                                 goto bp_err;
735                         } else if (error)
736                                 goto bp_err;
737                         if (new_blk != log_bbnum)
738                                 head_blk = new_blk;
739                 } else if (error)
740                         goto bp_err;
741         }
742
743         xlog_put_bp(bp);
744         if (head_blk == log_bbnum)
745                 *return_head_blk = 0;
746         else
747                 *return_head_blk = head_blk;
748         /*
749          * When returning here, we have a good block number.  Bad block
750          * means that during a previous crash, we didn't have a clean break
751          * from cycle number N to cycle number N-1.  In this case, we need
752          * to find the first block with cycle number N-1.
753          */
754         return 0;
755
756  bp_err:
757         xlog_put_bp(bp);
758
759         if (error)
760             xlog_warn("XFS: failed to find log head");
761         return error;
762 }
763
764 /*
765  * Find the sync block number or the tail of the log.
766  *
767  * This will be the block number of the last record to have its
768  * associated buffers synced to disk.  Every log record header has
769  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
770  * to get a sync block number.  The only concern is to figure out which
771  * log record header to believe.
772  *
773  * The following algorithm uses the log record header with the largest
774  * lsn.  The entire log record does not need to be valid.  We only care
775  * that the header is valid.
776  *
777  * We could speed up search by using current head_blk buffer, but it is not
778  * available.
779  */
780 int
781 xlog_find_tail(
782         xlog_t                  *log,
783         xfs_daddr_t             *head_blk,
784         xfs_daddr_t             *tail_blk)
785 {
786         xlog_rec_header_t       *rhead;
787         xlog_op_header_t        *op_head;
788         xfs_caddr_t             offset = NULL;
789         xfs_buf_t               *bp;
790         int                     error, i, found;
791         xfs_daddr_t             umount_data_blk;
792         xfs_daddr_t             after_umount_blk;
793         xfs_lsn_t               tail_lsn;
794         int                     hblks;
795
796         found = 0;
797
798         /*
799          * Find previous log record
800          */
801         if ((error = xlog_find_head(log, head_blk)))
802                 return error;
803
804         bp = xlog_get_bp(log, 1);
805         if (!bp)
806                 return ENOMEM;
807         if (*head_blk == 0) {                           /* special case */
808                 if ((error = xlog_bread(log, 0, 1, bp)))
809                         goto bread_err;
810                 offset = xlog_align(log, 0, 1, bp);
811                 if (GET_CYCLE(offset, ARCH_CONVERT) == 0) {
812                         *tail_blk = 0;
813                         /* leave all other log inited values alone */
814                         goto exit;
815                 }
816         }
817
818         /*
819          * Search backwards looking for log record header block
820          */
821         ASSERT(*head_blk < INT_MAX);
822         for (i = (int)(*head_blk) - 1; i >= 0; i--) {
823                 if ((error = xlog_bread(log, i, 1, bp)))
824                         goto bread_err;
825                 offset = xlog_align(log, i, 1, bp);
826                 if (XLOG_HEADER_MAGIC_NUM ==
827                     INT_GET(*(uint *)offset, ARCH_CONVERT)) {
828                         found = 1;
829                         break;
830                 }
831         }
832         /*
833          * If we haven't found the log record header block, start looking
834          * again from the end of the physical log.  XXXmiken: There should be
835          * a check here to make sure we didn't search more than N blocks in
836          * the previous code.
837          */
838         if (!found) {
839                 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
840                         if ((error = xlog_bread(log, i, 1, bp)))
841                                 goto bread_err;
842                         offset = xlog_align(log, i, 1, bp);
843                         if (XLOG_HEADER_MAGIC_NUM ==
844                             INT_GET(*(uint*)offset, ARCH_CONVERT)) {
845                                 found = 2;
846                                 break;
847                         }
848                 }
849         }
850         if (!found) {
851                 xlog_warn("XFS: xlog_find_tail: couldn't find sync record");
852                 ASSERT(0);
853                 return XFS_ERROR(EIO);
854         }
855
856         /* find blk_no of tail of log */
857         rhead = (xlog_rec_header_t *)offset;
858         *tail_blk = BLOCK_LSN(INT_GET(rhead->h_tail_lsn, ARCH_CONVERT));
859
860         /*
861          * Reset log values according to the state of the log when we
862          * crashed.  In the case where head_blk == 0, we bump curr_cycle
863          * one because the next write starts a new cycle rather than
864          * continuing the cycle of the last good log record.  At this
865          * point we have guaranteed that all partial log records have been
866          * accounted for.  Therefore, we know that the last good log record
867          * written was complete and ended exactly on the end boundary
868          * of the physical log.
869          */
870         log->l_prev_block = i;
871         log->l_curr_block = (int)*head_blk;
872         log->l_curr_cycle = INT_GET(rhead->h_cycle, ARCH_CONVERT);
873         if (found == 2)
874                 log->l_curr_cycle++;
875         log->l_tail_lsn = INT_GET(rhead->h_tail_lsn, ARCH_CONVERT);
876         log->l_last_sync_lsn = INT_GET(rhead->h_lsn, ARCH_CONVERT);
877         log->l_grant_reserve_cycle = log->l_curr_cycle;
878         log->l_grant_reserve_bytes = BBTOB(log->l_curr_block);
879         log->l_grant_write_cycle = log->l_curr_cycle;
880         log->l_grant_write_bytes = BBTOB(log->l_curr_block);
881
882         /*
883          * Look for unmount record.  If we find it, then we know there
884          * was a clean unmount.  Since 'i' could be the last block in
885          * the physical log, we convert to a log block before comparing
886          * to the head_blk.
887          *
888          * Save the current tail lsn to use to pass to
889          * xlog_clear_stale_blocks() below.  We won't want to clear the
890          * unmount record if there is one, so we pass the lsn of the
891          * unmount record rather than the block after it.
892          */
893         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
894                 int     h_size = INT_GET(rhead->h_size, ARCH_CONVERT);
895                 int     h_version = INT_GET(rhead->h_version, ARCH_CONVERT);
896
897                 if ((h_version & XLOG_VERSION_2) &&
898                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
899                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
900                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
901                                 hblks++;
902                 } else {
903                         hblks = 1;
904                 }
905         } else {
906                 hblks = 1;
907         }
908         after_umount_blk = (i + hblks + (int)
909                 BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT))) % log->l_logBBsize;
910         tail_lsn = log->l_tail_lsn;
911         if (*head_blk == after_umount_blk &&
912             INT_GET(rhead->h_num_logops, ARCH_CONVERT) == 1) {
913                 umount_data_blk = (i + hblks) % log->l_logBBsize;
914                 if ((error = xlog_bread(log, umount_data_blk, 1, bp))) {
915                         goto bread_err;
916                 }
917                 offset = xlog_align(log, umount_data_blk, 1, bp);
918                 op_head = (xlog_op_header_t *)offset;
919                 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
920                         /*
921                          * Set tail and last sync so that newly written
922                          * log records will point recovery to after the
923                          * current unmount record.
924                          */
925                         ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, log->l_curr_cycle,
926                                         after_umount_blk);
927                         ASSIGN_ANY_LSN_HOST(log->l_last_sync_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 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 = GET_CYCLE(offset, ARCH_CONVERT);
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 = GET_CYCLE(offset, ARCH_CONVERT);
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         INT_SET(recp->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1102         INT_SET(recp->h_cycle, ARCH_CONVERT, cycle);
1103         INT_SET(recp->h_version, ARCH_CONVERT,
1104                         XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1105         ASSIGN_ANY_LSN_DISK(recp->h_lsn, cycle, block);
1106         ASSIGN_ANY_LSN_DISK(recp->h_tail_lsn, tail_cycle, tail_block);
1107         INT_SET(recp->h_fmt, ARCH_CONVERT, 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                         XFS_BUF_SET_PTR(bp, offset + balign, BBTOB(sectbb));
1164                         if ((error = xlog_bread(log, ealign, sectbb, bp)))
1165                                 break;
1166                         XFS_BUF_SET_PTR(bp, offset, bufblks);
1167                 }
1168
1169                 offset = xlog_align(log, start_block, endcount, bp);
1170                 for (; j < endcount; j++) {
1171                         xlog_add_record(log, offset, cycle, i+j,
1172                                         tail_cycle, tail_block);
1173                         offset += BBSIZE;
1174                 }
1175                 error = xlog_bwrite(log, start_block, endcount, bp);
1176                 if (error)
1177                         break;
1178                 start_block += endcount;
1179                 j = 0;
1180         }
1181         xlog_put_bp(bp);
1182         return error;
1183 }
1184
1185 /*
1186  * This routine is called to blow away any incomplete log writes out
1187  * in front of the log head.  We do this so that we won't become confused
1188  * if we come up, write only a little bit more, and then crash again.
1189  * If we leave the partial log records out there, this situation could
1190  * cause us to think those partial writes are valid blocks since they
1191  * have the current cycle number.  We get rid of them by overwriting them
1192  * with empty log records with the old cycle number rather than the
1193  * current one.
1194  *
1195  * The tail lsn is passed in rather than taken from
1196  * the log so that we will not write over the unmount record after a
1197  * clean unmount in a 512 block log.  Doing so would leave the log without
1198  * any valid log records in it until a new one was written.  If we crashed
1199  * during that time we would not be able to recover.
1200  */
1201 STATIC int
1202 xlog_clear_stale_blocks(
1203         xlog_t          *log,
1204         xfs_lsn_t       tail_lsn)
1205 {
1206         int             tail_cycle, head_cycle;
1207         int             tail_block, head_block;
1208         int             tail_distance, max_distance;
1209         int             distance;
1210         int             error;
1211
1212         tail_cycle = CYCLE_LSN(tail_lsn);
1213         tail_block = BLOCK_LSN(tail_lsn);
1214         head_cycle = log->l_curr_cycle;
1215         head_block = log->l_curr_block;
1216
1217         /*
1218          * Figure out the distance between the new head of the log
1219          * and the tail.  We want to write over any blocks beyond the
1220          * head that we may have written just before the crash, but
1221          * we don't want to overwrite the tail of the log.
1222          */
1223         if (head_cycle == tail_cycle) {
1224                 /*
1225                  * The tail is behind the head in the physical log,
1226                  * so the distance from the head to the tail is the
1227                  * distance from the head to the end of the log plus
1228                  * the distance from the beginning of the log to the
1229                  * tail.
1230                  */
1231                 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1232                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1233                                          XFS_ERRLEVEL_LOW, log->l_mp);
1234                         return XFS_ERROR(EFSCORRUPTED);
1235                 }
1236                 tail_distance = tail_block + (log->l_logBBsize - head_block);
1237         } else {
1238                 /*
1239                  * The head is behind the tail in the physical log,
1240                  * so the distance from the head to the tail is just
1241                  * the tail block minus the head block.
1242                  */
1243                 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1244                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1245                                          XFS_ERRLEVEL_LOW, log->l_mp);
1246                         return XFS_ERROR(EFSCORRUPTED);
1247                 }
1248                 tail_distance = tail_block - head_block;
1249         }
1250
1251         /*
1252          * If the head is right up against the tail, we can't clear
1253          * anything.
1254          */
1255         if (tail_distance <= 0) {
1256                 ASSERT(tail_distance == 0);
1257                 return 0;
1258         }
1259
1260         max_distance = XLOG_TOTAL_REC_SHIFT(log);
1261         /*
1262          * Take the smaller of the maximum amount of outstanding I/O
1263          * we could have and the distance to the tail to clear out.
1264          * We take the smaller so that we don't overwrite the tail and
1265          * we don't waste all day writing from the head to the tail
1266          * for no reason.
1267          */
1268         max_distance = MIN(max_distance, tail_distance);
1269
1270         if ((head_block + max_distance) <= log->l_logBBsize) {
1271                 /*
1272                  * We can stomp all the blocks we need to without
1273                  * wrapping around the end of the log.  Just do it
1274                  * in a single write.  Use the cycle number of the
1275                  * current cycle minus one so that the log will look like:
1276                  *     n ... | n - 1 ...
1277                  */
1278                 error = xlog_write_log_records(log, (head_cycle - 1),
1279                                 head_block, max_distance, tail_cycle,
1280                                 tail_block);
1281                 if (error)
1282                         return error;
1283         } else {
1284                 /*
1285                  * We need to wrap around the end of the physical log in
1286                  * order to clear all the blocks.  Do it in two separate
1287                  * I/Os.  The first write should be from the head to the
1288                  * end of the physical log, and it should use the current
1289                  * cycle number minus one just like above.
1290                  */
1291                 distance = log->l_logBBsize - head_block;
1292                 error = xlog_write_log_records(log, (head_cycle - 1),
1293                                 head_block, distance, tail_cycle,
1294                                 tail_block);
1295
1296                 if (error)
1297                         return error;
1298
1299                 /*
1300                  * Now write the blocks at the start of the physical log.
1301                  * This writes the remainder of the blocks we want to clear.
1302                  * It uses the current cycle number since we're now on the
1303                  * same cycle as the head so that we get:
1304                  *    n ... n ... | n - 1 ...
1305                  *    ^^^^^ blocks we're writing
1306                  */
1307                 distance = max_distance - (log->l_logBBsize - head_block);
1308                 error = xlog_write_log_records(log, head_cycle, 0, distance,
1309                                 tail_cycle, tail_block);
1310                 if (error)
1311                         return error;
1312         }
1313
1314         return 0;
1315 }
1316
1317 /******************************************************************************
1318  *
1319  *              Log recover routines
1320  *
1321  ******************************************************************************
1322  */
1323
1324 STATIC xlog_recover_t *
1325 xlog_recover_find_tid(
1326         xlog_recover_t          *q,
1327         xlog_tid_t              tid)
1328 {
1329         xlog_recover_t          *p = q;
1330
1331         while (p != NULL) {
1332                 if (p->r_log_tid == tid)
1333                     break;
1334                 p = p->r_next;
1335         }
1336         return p;
1337 }
1338
1339 STATIC void
1340 xlog_recover_put_hashq(
1341         xlog_recover_t          **q,
1342         xlog_recover_t          *trans)
1343 {
1344         trans->r_next = *q;
1345         *q = trans;
1346 }
1347
1348 STATIC void
1349 xlog_recover_add_item(
1350         xlog_recover_item_t     **itemq)
1351 {
1352         xlog_recover_item_t     *item;
1353
1354         item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1355         xlog_recover_insert_item_backq(itemq, item);
1356 }
1357
1358 STATIC int
1359 xlog_recover_add_to_cont_trans(
1360         xlog_recover_t          *trans,
1361         xfs_caddr_t             dp,
1362         int                     len)
1363 {
1364         xlog_recover_item_t     *item;
1365         xfs_caddr_t             ptr, old_ptr;
1366         int                     old_len;
1367
1368         item = trans->r_itemq;
1369         if (item == NULL) {
1370                 /* finish copying rest of trans header */
1371                 xlog_recover_add_item(&trans->r_itemq);
1372                 ptr = (xfs_caddr_t) &trans->r_theader +
1373                                 sizeof(xfs_trans_header_t) - len;
1374                 memcpy(ptr, dp, len); /* d, s, l */
1375                 return 0;
1376         }
1377         item = item->ri_prev;
1378
1379         old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1380         old_len = item->ri_buf[item->ri_cnt-1].i_len;
1381
1382         ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1383         memcpy(&ptr[old_len], dp, len); /* d, s, l */
1384         item->ri_buf[item->ri_cnt-1].i_len += len;
1385         item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1386         return 0;
1387 }
1388
1389 /*
1390  * The next region to add is the start of a new region.  It could be
1391  * a whole region or it could be the first part of a new region.  Because
1392  * of this, the assumption here is that the type and size fields of all
1393  * format structures fit into the first 32 bits of the structure.
1394  *
1395  * This works because all regions must be 32 bit aligned.  Therefore, we
1396  * either have both fields or we have neither field.  In the case we have
1397  * neither field, the data part of the region is zero length.  We only have
1398  * a log_op_header and can throw away the header since a new one will appear
1399  * later.  If we have at least 4 bytes, then we can determine how many regions
1400  * will appear in the current log item.
1401  */
1402 STATIC int
1403 xlog_recover_add_to_trans(
1404         xlog_recover_t          *trans,
1405         xfs_caddr_t             dp,
1406         int                     len)
1407 {
1408         xfs_inode_log_format_t  *in_f;                  /* any will do */
1409         xlog_recover_item_t     *item;
1410         xfs_caddr_t             ptr;
1411
1412         if (!len)
1413                 return 0;
1414         item = trans->r_itemq;
1415         if (item == NULL) {
1416                 ASSERT(*(uint *)dp == XFS_TRANS_HEADER_MAGIC);
1417                 if (len == sizeof(xfs_trans_header_t))
1418                         xlog_recover_add_item(&trans->r_itemq);
1419                 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1420                 return 0;
1421         }
1422
1423         ptr = kmem_alloc(len, KM_SLEEP);
1424         memcpy(ptr, dp, len);
1425         in_f = (xfs_inode_log_format_t *)ptr;
1426
1427         if (item->ri_prev->ri_total != 0 &&
1428              item->ri_prev->ri_total == item->ri_prev->ri_cnt) {
1429                 xlog_recover_add_item(&trans->r_itemq);
1430         }
1431         item = trans->r_itemq;
1432         item = item->ri_prev;
1433
1434         if (item->ri_total == 0) {              /* first region to be added */
1435                 item->ri_total  = in_f->ilf_size;
1436                 ASSERT(item->ri_total <= XLOG_MAX_REGIONS_IN_ITEM);
1437                 item->ri_buf = kmem_zalloc((item->ri_total *
1438                                             sizeof(xfs_log_iovec_t)), KM_SLEEP);
1439         }
1440         ASSERT(item->ri_total > item->ri_cnt);
1441         /* Description region is ri_buf[0] */
1442         item->ri_buf[item->ri_cnt].i_addr = ptr;
1443         item->ri_buf[item->ri_cnt].i_len  = len;
1444         item->ri_cnt++;
1445         return 0;
1446 }
1447
1448 STATIC void
1449 xlog_recover_new_tid(
1450         xlog_recover_t          **q,
1451         xlog_tid_t              tid,
1452         xfs_lsn_t               lsn)
1453 {
1454         xlog_recover_t          *trans;
1455
1456         trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1457         trans->r_log_tid   = tid;
1458         trans->r_lsn       = lsn;
1459         xlog_recover_put_hashq(q, trans);
1460 }
1461
1462 STATIC int
1463 xlog_recover_unlink_tid(
1464         xlog_recover_t          **q,
1465         xlog_recover_t          *trans)
1466 {
1467         xlog_recover_t          *tp;
1468         int                     found = 0;
1469
1470         ASSERT(trans != NULL);
1471         if (trans == *q) {
1472                 *q = (*q)->r_next;
1473         } else {
1474                 tp = *q;
1475                 while (tp) {
1476                         if (tp->r_next == trans) {
1477                                 found = 1;
1478                                 break;
1479                         }
1480                         tp = tp->r_next;
1481                 }
1482                 if (!found) {
1483                         xlog_warn(
1484                              "XFS: xlog_recover_unlink_tid: trans not found");
1485                         ASSERT(0);
1486                         return XFS_ERROR(EIO);
1487                 }
1488                 tp->r_next = tp->r_next->r_next;
1489         }
1490         return 0;
1491 }
1492
1493 STATIC void
1494 xlog_recover_insert_item_backq(
1495         xlog_recover_item_t     **q,
1496         xlog_recover_item_t     *item)
1497 {
1498         if (*q == NULL) {
1499                 item->ri_prev = item->ri_next = item;
1500                 *q = item;
1501         } else {
1502                 item->ri_next           = *q;
1503                 item->ri_prev           = (*q)->ri_prev;
1504                 (*q)->ri_prev           = item;
1505                 item->ri_prev->ri_next  = item;
1506         }
1507 }
1508
1509 STATIC void
1510 xlog_recover_insert_item_frontq(
1511         xlog_recover_item_t     **q,
1512         xlog_recover_item_t     *item)
1513 {
1514         xlog_recover_insert_item_backq(q, item);
1515         *q = item;
1516 }
1517
1518 STATIC int
1519 xlog_recover_reorder_trans(
1520         xlog_recover_t          *trans)
1521 {
1522         xlog_recover_item_t     *first_item, *itemq, *itemq_next;
1523         xfs_buf_log_format_t    *buf_f;
1524         ushort                  flags = 0;
1525
1526         first_item = itemq = trans->r_itemq;
1527         trans->r_itemq = NULL;
1528         do {
1529                 itemq_next = itemq->ri_next;
1530                 buf_f = (xfs_buf_log_format_t *)itemq->ri_buf[0].i_addr;
1531
1532                 switch (ITEM_TYPE(itemq)) {
1533                 case XFS_LI_BUF:
1534                         flags = buf_f->blf_flags;
1535                         if (!(flags & XFS_BLI_CANCEL)) {
1536                                 xlog_recover_insert_item_frontq(&trans->r_itemq,
1537                                                                 itemq);
1538                                 break;
1539                         }
1540                 case XFS_LI_INODE:
1541                 case XFS_LI_DQUOT:
1542                 case XFS_LI_QUOTAOFF:
1543                 case XFS_LI_EFD:
1544                 case XFS_LI_EFI:
1545                         xlog_recover_insert_item_backq(&trans->r_itemq, itemq);
1546                         break;
1547                 default:
1548                         xlog_warn(
1549         "XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1550                         ASSERT(0);
1551                         return XFS_ERROR(EIO);
1552                 }
1553                 itemq = itemq_next;
1554         } while (first_item != itemq);
1555         return 0;
1556 }
1557
1558 /*
1559  * Build up the table of buf cancel records so that we don't replay
1560  * cancelled data in the second pass.  For buffer records that are
1561  * not cancel records, there is nothing to do here so we just return.
1562  *
1563  * If we get a cancel record which is already in the table, this indicates
1564  * that the buffer was cancelled multiple times.  In order to ensure
1565  * that during pass 2 we keep the record in the table until we reach its
1566  * last occurrence in the log, we keep a reference count in the cancel
1567  * record in the table to tell us how many times we expect to see this
1568  * record during the second pass.
1569  */
1570 STATIC void
1571 xlog_recover_do_buffer_pass1(
1572         xlog_t                  *log,
1573         xfs_buf_log_format_t    *buf_f)
1574 {
1575         xfs_buf_cancel_t        *bcp;
1576         xfs_buf_cancel_t        *nextp;
1577         xfs_buf_cancel_t        *prevp;
1578         xfs_buf_cancel_t        **bucket;
1579         xfs_daddr_t             blkno = 0;
1580         uint                    len = 0;
1581         ushort                  flags = 0;
1582
1583         switch (buf_f->blf_type) {
1584         case XFS_LI_BUF:
1585                 blkno = buf_f->blf_blkno;
1586                 len = buf_f->blf_len;
1587                 flags = buf_f->blf_flags;
1588                 break;
1589         }
1590
1591         /*
1592          * If this isn't a cancel buffer item, then just return.
1593          */
1594         if (!(flags & XFS_BLI_CANCEL))
1595                 return;
1596
1597         /*
1598          * Insert an xfs_buf_cancel record into the hash table of
1599          * them.  If there is already an identical record, bump
1600          * its reference count.
1601          */
1602         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1603                                           XLOG_BC_TABLE_SIZE];
1604         /*
1605          * If the hash bucket is empty then just insert a new record into
1606          * the bucket.
1607          */
1608         if (*bucket == NULL) {
1609                 bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1610                                                      KM_SLEEP);
1611                 bcp->bc_blkno = blkno;
1612                 bcp->bc_len = len;
1613                 bcp->bc_refcount = 1;
1614                 bcp->bc_next = NULL;
1615                 *bucket = bcp;
1616                 return;
1617         }
1618
1619         /*
1620          * The hash bucket is not empty, so search for duplicates of our
1621          * record.  If we find one them just bump its refcount.  If not
1622          * then add us at the end of the list.
1623          */
1624         prevp = NULL;
1625         nextp = *bucket;
1626         while (nextp != NULL) {
1627                 if (nextp->bc_blkno == blkno && nextp->bc_len == len) {
1628                         nextp->bc_refcount++;
1629                         return;
1630                 }
1631                 prevp = nextp;
1632                 nextp = nextp->bc_next;
1633         }
1634         ASSERT(prevp != NULL);
1635         bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1636                                              KM_SLEEP);
1637         bcp->bc_blkno = blkno;
1638         bcp->bc_len = len;
1639         bcp->bc_refcount = 1;
1640         bcp->bc_next = NULL;
1641         prevp->bc_next = bcp;
1642 }
1643
1644 /*
1645  * Check to see whether the buffer being recovered has a corresponding
1646  * entry in the buffer cancel record table.  If it does then return 1
1647  * so that it will be cancelled, otherwise return 0.  If the buffer is
1648  * actually a buffer cancel item (XFS_BLI_CANCEL is set), then decrement
1649  * the refcount on the entry in the table and remove it from the table
1650  * if this is the last reference.
1651  *
1652  * We remove the cancel record from the table when we encounter its
1653  * last occurrence in the log so that if the same buffer is re-used
1654  * again after its last cancellation we actually replay the changes
1655  * made at that point.
1656  */
1657 STATIC int
1658 xlog_check_buffer_cancelled(
1659         xlog_t                  *log,
1660         xfs_daddr_t             blkno,
1661         uint                    len,
1662         ushort                  flags)
1663 {
1664         xfs_buf_cancel_t        *bcp;
1665         xfs_buf_cancel_t        *prevp;
1666         xfs_buf_cancel_t        **bucket;
1667
1668         if (log->l_buf_cancel_table == NULL) {
1669                 /*
1670                  * There is nothing in the table built in pass one,
1671                  * so this buffer must not be cancelled.
1672                  */
1673                 ASSERT(!(flags & XFS_BLI_CANCEL));
1674                 return 0;
1675         }
1676
1677         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1678                                           XLOG_BC_TABLE_SIZE];
1679         bcp = *bucket;
1680         if (bcp == NULL) {
1681                 /*
1682                  * There is no corresponding entry in the table built
1683                  * in pass one, so this buffer has not been cancelled.
1684                  */
1685                 ASSERT(!(flags & XFS_BLI_CANCEL));
1686                 return 0;
1687         }
1688
1689         /*
1690          * Search for an entry in the buffer cancel table that
1691          * matches our buffer.
1692          */
1693         prevp = NULL;
1694         while (bcp != NULL) {
1695                 if (bcp->bc_blkno == blkno && bcp->bc_len == len) {
1696                         /*
1697                          * We've go a match, so return 1 so that the
1698                          * recovery of this buffer is cancelled.
1699                          * If this buffer is actually a buffer cancel
1700                          * log item, then decrement the refcount on the
1701                          * one in the table and remove it if this is the
1702                          * last reference.
1703                          */
1704                         if (flags & XFS_BLI_CANCEL) {
1705                                 bcp->bc_refcount--;
1706                                 if (bcp->bc_refcount == 0) {
1707                                         if (prevp == NULL) {
1708                                                 *bucket = bcp->bc_next;
1709                                         } else {
1710                                                 prevp->bc_next = bcp->bc_next;
1711                                         }
1712                                         kmem_free(bcp,
1713                                                   sizeof(xfs_buf_cancel_t));
1714                                 }
1715                         }
1716                         return 1;
1717                 }
1718                 prevp = bcp;
1719                 bcp = bcp->bc_next;
1720         }
1721         /*
1722          * We didn't find a corresponding entry in the table, so
1723          * return 0 so that the buffer is NOT cancelled.
1724          */
1725         ASSERT(!(flags & XFS_BLI_CANCEL));
1726         return 0;
1727 }
1728
1729 STATIC int
1730 xlog_recover_do_buffer_pass2(
1731         xlog_t                  *log,
1732         xfs_buf_log_format_t    *buf_f)
1733 {
1734         xfs_daddr_t             blkno = 0;
1735         ushort                  flags = 0;
1736         uint                    len = 0;
1737
1738         switch (buf_f->blf_type) {
1739         case XFS_LI_BUF:
1740                 blkno = buf_f->blf_blkno;
1741                 flags = buf_f->blf_flags;
1742                 len = buf_f->blf_len;
1743                 break;
1744         }
1745
1746         return xlog_check_buffer_cancelled(log, blkno, len, flags);
1747 }
1748
1749 /*
1750  * Perform recovery for a buffer full of inodes.  In these buffers,
1751  * the only data which should be recovered is that which corresponds
1752  * to the di_next_unlinked pointers in the on disk inode structures.
1753  * The rest of the data for the inodes is always logged through the
1754  * inodes themselves rather than the inode buffer and is recovered
1755  * in xlog_recover_do_inode_trans().
1756  *
1757  * The only time when buffers full of inodes are fully recovered is
1758  * when the buffer is full of newly allocated inodes.  In this case
1759  * the buffer will not be marked as an inode buffer and so will be
1760  * sent to xlog_recover_do_reg_buffer() below during recovery.
1761  */
1762 STATIC int
1763 xlog_recover_do_inode_buffer(
1764         xfs_mount_t             *mp,
1765         xlog_recover_item_t     *item,
1766         xfs_buf_t               *bp,
1767         xfs_buf_log_format_t    *buf_f)
1768 {
1769         int                     i;
1770         int                     item_index;
1771         int                     bit;
1772         int                     nbits;
1773         int                     reg_buf_offset;
1774         int                     reg_buf_bytes;
1775         int                     next_unlinked_offset;
1776         int                     inodes_per_buf;
1777         xfs_agino_t             *logged_nextp;
1778         xfs_agino_t             *buffer_nextp;
1779         unsigned int            *data_map = NULL;
1780         unsigned int            map_size = 0;
1781
1782         switch (buf_f->blf_type) {
1783         case XFS_LI_BUF:
1784                 data_map = buf_f->blf_data_map;
1785                 map_size = buf_f->blf_map_size;
1786                 break;
1787         }
1788         /*
1789          * Set the variables corresponding to the current region to
1790          * 0 so that we'll initialize them on the first pass through
1791          * the loop.
1792          */
1793         reg_buf_offset = 0;
1794         reg_buf_bytes = 0;
1795         bit = 0;
1796         nbits = 0;
1797         item_index = 0;
1798         inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1799         for (i = 0; i < inodes_per_buf; i++) {
1800                 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1801                         offsetof(xfs_dinode_t, di_next_unlinked);
1802
1803                 while (next_unlinked_offset >=
1804                        (reg_buf_offset + reg_buf_bytes)) {
1805                         /*
1806                          * The next di_next_unlinked field is beyond
1807                          * the current logged region.  Find the next
1808                          * logged region that contains or is beyond
1809                          * the current di_next_unlinked field.
1810                          */
1811                         bit += nbits;
1812                         bit = xfs_next_bit(data_map, map_size, bit);
1813
1814                         /*
1815                          * If there are no more logged regions in the
1816                          * buffer, then we're done.
1817                          */
1818                         if (bit == -1) {
1819                                 return 0;
1820                         }
1821
1822                         nbits = xfs_contig_bits(data_map, map_size,
1823                                                          bit);
1824                         ASSERT(nbits > 0);
1825                         reg_buf_offset = bit << XFS_BLI_SHIFT;
1826                         reg_buf_bytes = nbits << XFS_BLI_SHIFT;
1827                         item_index++;
1828                 }
1829
1830                 /*
1831                  * If the current logged region starts after the current
1832                  * di_next_unlinked field, then move on to the next
1833                  * di_next_unlinked field.
1834                  */
1835                 if (next_unlinked_offset < reg_buf_offset) {
1836                         continue;
1837                 }
1838
1839                 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1840                 ASSERT((item->ri_buf[item_index].i_len % XFS_BLI_CHUNK) == 0);
1841                 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1842
1843                 /*
1844                  * The current logged region contains a copy of the
1845                  * current di_next_unlinked field.  Extract its value
1846                  * and copy it to the buffer copy.
1847                  */
1848                 logged_nextp = (xfs_agino_t *)
1849                                ((char *)(item->ri_buf[item_index].i_addr) +
1850                                 (next_unlinked_offset - reg_buf_offset));
1851                 if (unlikely(*logged_nextp == 0)) {
1852                         xfs_fs_cmn_err(CE_ALERT, mp,
1853                                 "bad inode buffer log record (ptr = 0x%p, bp = 0x%p).  XFS trying to replay bad (0) inode di_next_unlinked field",
1854                                 item, bp);
1855                         XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1856                                          XFS_ERRLEVEL_LOW, mp);
1857                         return XFS_ERROR(EFSCORRUPTED);
1858                 }
1859
1860                 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1861                                               next_unlinked_offset);
1862                 *buffer_nextp = *logged_nextp;
1863         }
1864
1865         return 0;
1866 }
1867
1868 /*
1869  * Perform a 'normal' buffer recovery.  Each logged region of the
1870  * buffer should be copied over the corresponding region in the
1871  * given buffer.  The bitmap in the buf log format structure indicates
1872  * where to place the logged data.
1873  */
1874 /*ARGSUSED*/
1875 STATIC void
1876 xlog_recover_do_reg_buffer(
1877         xfs_mount_t             *mp,
1878         xlog_recover_item_t     *item,
1879         xfs_buf_t               *bp,
1880         xfs_buf_log_format_t    *buf_f)
1881 {
1882         int                     i;
1883         int                     bit;
1884         int                     nbits;
1885         unsigned int            *data_map = NULL;
1886         unsigned int            map_size = 0;
1887         int                     error;
1888         int                     stale_buf = 1;
1889
1890         /*
1891          * Scan through the on-disk inode buffer and attempt to
1892          * determine if it has been written to since it was logged.
1893          *
1894          * - If any of the magic numbers are incorrect then the buffer is stale
1895          * - If any of the modes are non-zero then the buffer is not stale
1896          * - If all of the modes are zero and at least one of the generation
1897          *   counts is non-zero then the buffer is stale
1898          *
1899          * If the end result is a stale buffer then the log buffer is replayed
1900          * otherwise it is skipped.
1901          *
1902          * This heuristic is not perfect.  It can be improved by scanning the
1903          * entire inode chunk for evidence that any of the inode clusters have
1904          * been updated.  To fix this problem completely we will need a major
1905          * architectural change to the logging system.
1906          */
1907         if (buf_f->blf_flags & XFS_BLI_INODE_NEW_BUF) {
1908                 xfs_dinode_t    *dip;
1909                 int             inodes_per_buf;
1910                 int             mode_count = 0;
1911                 int             gen_count = 0;
1912
1913                 stale_buf = 0;
1914                 inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1915                 for (i = 0; i < inodes_per_buf; i++) {
1916                         dip = (xfs_dinode_t *)xfs_buf_offset(bp,
1917                                 i * mp->m_sb.sb_inodesize);
1918                         if (be16_to_cpu(dip->di_core.di_magic) !=
1919                                         XFS_DINODE_MAGIC) {
1920                                 stale_buf = 1;
1921                                 break;
1922                         }
1923                         if (be16_to_cpu(dip->di_core.di_mode))
1924                                 mode_count++;
1925                         if (be16_to_cpu(dip->di_core.di_gen))
1926                                 gen_count++;
1927                 }
1928
1929                 if (!mode_count && gen_count)
1930                         stale_buf = 1;
1931         }
1932
1933         switch (buf_f->blf_type) {
1934         case XFS_LI_BUF:
1935                 data_map = buf_f->blf_data_map;
1936                 map_size = buf_f->blf_map_size;
1937                 break;
1938         }
1939         bit = 0;
1940         i = 1;  /* 0 is the buf format structure */
1941         while (1) {
1942                 bit = xfs_next_bit(data_map, map_size, bit);
1943                 if (bit == -1)
1944                         break;
1945                 nbits = xfs_contig_bits(data_map, map_size, bit);
1946                 ASSERT(nbits > 0);
1947                 ASSERT(item->ri_buf[i].i_addr != NULL);
1948                 ASSERT(item->ri_buf[i].i_len % XFS_BLI_CHUNK == 0);
1949                 ASSERT(XFS_BUF_COUNT(bp) >=
1950                        ((uint)bit << XFS_BLI_SHIFT)+(nbits<<XFS_BLI_SHIFT));
1951
1952                 /*
1953                  * Do a sanity check if this is a dquot buffer. Just checking
1954                  * the first dquot in the buffer should do. XXXThis is
1955                  * probably a good thing to do for other buf types also.
1956                  */
1957                 error = 0;
1958                 if (buf_f->blf_flags &
1959                    (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
1960                         error = xfs_qm_dqcheck((xfs_disk_dquot_t *)
1961                                                item->ri_buf[i].i_addr,
1962                                                -1, 0, XFS_QMOPT_DOWARN,
1963                                                "dquot_buf_recover");
1964                 }
1965                 if (!error && stale_buf)
1966                         memcpy(xfs_buf_offset(bp,
1967                                 (uint)bit << XFS_BLI_SHIFT),    /* dest */
1968                                 item->ri_buf[i].i_addr,         /* source */
1969                                 nbits<<XFS_BLI_SHIFT);          /* length */
1970                 i++;
1971                 bit += nbits;
1972         }
1973
1974         /* Shouldn't be any more regions */
1975         ASSERT(i == item->ri_total);
1976 }
1977
1978 /*
1979  * Do some primitive error checking on ondisk dquot data structures.
1980  */
1981 int
1982 xfs_qm_dqcheck(
1983         xfs_disk_dquot_t *ddq,
1984         xfs_dqid_t       id,
1985         uint             type,    /* used only when IO_dorepair is true */
1986         uint             flags,
1987         char             *str)
1988 {
1989         xfs_dqblk_t      *d = (xfs_dqblk_t *)ddq;
1990         int             errs = 0;
1991
1992         /*
1993          * We can encounter an uninitialized dquot buffer for 2 reasons:
1994          * 1. If we crash while deleting the quotainode(s), and those blks got
1995          *    used for user data. This is because we take the path of regular
1996          *    file deletion; however, the size field of quotainodes is never
1997          *    updated, so all the tricks that we play in itruncate_finish
1998          *    don't quite matter.
1999          *
2000          * 2. We don't play the quota buffers when there's a quotaoff logitem.
2001          *    But the allocation will be replayed so we'll end up with an
2002          *    uninitialized quota block.
2003          *
2004          * This is all fine; things are still consistent, and we haven't lost
2005          * any quota information. Just don't complain about bad dquot blks.
2006          */
2007         if (be16_to_cpu(ddq->d_magic) != XFS_DQUOT_MAGIC) {
2008                 if (flags & XFS_QMOPT_DOWARN)
2009                         cmn_err(CE_ALERT,
2010                         "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
2011                         str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
2012                 errs++;
2013         }
2014         if (ddq->d_version != XFS_DQUOT_VERSION) {
2015                 if (flags & XFS_QMOPT_DOWARN)
2016                         cmn_err(CE_ALERT,
2017                         "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
2018                         str, id, ddq->d_version, XFS_DQUOT_VERSION);
2019                 errs++;
2020         }
2021
2022         if (ddq->d_flags != XFS_DQ_USER &&
2023             ddq->d_flags != XFS_DQ_PROJ &&
2024             ddq->d_flags != XFS_DQ_GROUP) {
2025                 if (flags & XFS_QMOPT_DOWARN)
2026                         cmn_err(CE_ALERT,
2027                         "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
2028                         str, id, ddq->d_flags);
2029                 errs++;
2030         }
2031
2032         if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
2033                 if (flags & XFS_QMOPT_DOWARN)
2034                         cmn_err(CE_ALERT,
2035                         "%s : ondisk-dquot 0x%p, ID mismatch: "
2036                         "0x%x expected, found id 0x%x",
2037                         str, ddq, id, be32_to_cpu(ddq->d_id));
2038                 errs++;
2039         }
2040
2041         if (!errs && ddq->d_id) {
2042                 if (ddq->d_blk_softlimit &&
2043                     be64_to_cpu(ddq->d_bcount) >=
2044                                 be64_to_cpu(ddq->d_blk_softlimit)) {
2045                         if (!ddq->d_btimer) {
2046                                 if (flags & XFS_QMOPT_DOWARN)
2047                                         cmn_err(CE_ALERT,
2048                                         "%s : Dquot ID 0x%x (0x%p) "
2049                                         "BLK TIMER NOT STARTED",
2050                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2051                                 errs++;
2052                         }
2053                 }
2054                 if (ddq->d_ino_softlimit &&
2055                     be64_to_cpu(ddq->d_icount) >=
2056                                 be64_to_cpu(ddq->d_ino_softlimit)) {
2057                         if (!ddq->d_itimer) {
2058                                 if (flags & XFS_QMOPT_DOWARN)
2059                                         cmn_err(CE_ALERT,
2060                                         "%s : Dquot ID 0x%x (0x%p) "
2061                                         "INODE TIMER NOT STARTED",
2062                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2063                                 errs++;
2064                         }
2065                 }
2066                 if (ddq->d_rtb_softlimit &&
2067                     be64_to_cpu(ddq->d_rtbcount) >=
2068                                 be64_to_cpu(ddq->d_rtb_softlimit)) {
2069                         if (!ddq->d_rtbtimer) {
2070                                 if (flags & XFS_QMOPT_DOWARN)
2071                                         cmn_err(CE_ALERT,
2072                                         "%s : Dquot ID 0x%x (0x%p) "
2073                                         "RTBLK TIMER NOT STARTED",
2074                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2075                                 errs++;
2076                         }
2077                 }
2078         }
2079
2080         if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2081                 return errs;
2082
2083         if (flags & XFS_QMOPT_DOWARN)
2084                 cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2085
2086         /*
2087          * Typically, a repair is only requested by quotacheck.
2088          */
2089         ASSERT(id != -1);
2090         ASSERT(flags & XFS_QMOPT_DQREPAIR);
2091         memset(d, 0, sizeof(xfs_dqblk_t));
2092
2093         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2094         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2095         d->dd_diskdq.d_flags = type;
2096         d->dd_diskdq.d_id = cpu_to_be32(id);
2097
2098         return errs;
2099 }
2100
2101 /*
2102  * Perform a dquot buffer recovery.
2103  * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2104  * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2105  * Else, treat it as a regular buffer and do recovery.
2106  */
2107 STATIC void
2108 xlog_recover_do_dquot_buffer(
2109         xfs_mount_t             *mp,
2110         xlog_t                  *log,
2111         xlog_recover_item_t     *item,
2112         xfs_buf_t               *bp,
2113         xfs_buf_log_format_t    *buf_f)
2114 {
2115         uint                    type;
2116
2117         /*
2118          * Filesystems are required to send in quota flags at mount time.
2119          */
2120         if (mp->m_qflags == 0) {
2121                 return;
2122         }
2123
2124         type = 0;
2125         if (buf_f->blf_flags & XFS_BLI_UDQUOT_BUF)
2126                 type |= XFS_DQ_USER;
2127         if (buf_f->blf_flags & XFS_BLI_PDQUOT_BUF)
2128                 type |= XFS_DQ_PROJ;
2129         if (buf_f->blf_flags & XFS_BLI_GDQUOT_BUF)
2130                 type |= XFS_DQ_GROUP;
2131         /*
2132          * This type of quotas was turned off, so ignore this buffer
2133          */
2134         if (log->l_quotaoffs_flag & type)
2135                 return;
2136
2137         xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2138 }
2139
2140 /*
2141  * This routine replays a modification made to a buffer at runtime.
2142  * There are actually two types of buffer, regular and inode, which
2143  * are handled differently.  Inode buffers are handled differently
2144  * in that we only recover a specific set of data from them, namely
2145  * the inode di_next_unlinked fields.  This is because all other inode
2146  * data is actually logged via inode records and any data we replay
2147  * here which overlaps that may be stale.
2148  *
2149  * When meta-data buffers are freed at run time we log a buffer item
2150  * with the XFS_BLI_CANCEL bit set to indicate that previous copies
2151  * of the buffer in the log should not be replayed at recovery time.
2152  * This is so that if the blocks covered by the buffer are reused for
2153  * file data before we crash we don't end up replaying old, freed
2154  * meta-data into a user's file.
2155  *
2156  * To handle the cancellation of buffer log items, we make two passes
2157  * over the log during recovery.  During the first we build a table of
2158  * those buffers which have been cancelled, and during the second we
2159  * only replay those buffers which do not have corresponding cancel
2160  * records in the table.  See xlog_recover_do_buffer_pass[1,2] above
2161  * for more details on the implementation of the table of cancel records.
2162  */
2163 STATIC int
2164 xlog_recover_do_buffer_trans(
2165         xlog_t                  *log,
2166         xlog_recover_item_t     *item,
2167         int                     pass)
2168 {
2169         xfs_buf_log_format_t    *buf_f;
2170         xfs_mount_t             *mp;
2171         xfs_buf_t               *bp;
2172         int                     error;
2173         int                     cancel;
2174         xfs_daddr_t             blkno;
2175         int                     len;
2176         ushort                  flags;
2177
2178         buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
2179
2180         if (pass == XLOG_RECOVER_PASS1) {
2181                 /*
2182                  * In this pass we're only looking for buf items
2183                  * with the XFS_BLI_CANCEL bit set.
2184                  */
2185                 xlog_recover_do_buffer_pass1(log, buf_f);
2186                 return 0;
2187         } else {
2188                 /*
2189                  * In this pass we want to recover all the buffers
2190                  * which have not been cancelled and are not
2191                  * cancellation buffers themselves.  The routine
2192                  * we call here will tell us whether or not to
2193                  * continue with the replay of this buffer.
2194                  */
2195                 cancel = xlog_recover_do_buffer_pass2(log, buf_f);
2196                 if (cancel) {
2197                         return 0;
2198                 }
2199         }
2200         switch (buf_f->blf_type) {
2201         case XFS_LI_BUF:
2202                 blkno = buf_f->blf_blkno;
2203                 len = buf_f->blf_len;
2204                 flags = buf_f->blf_flags;
2205                 break;
2206         default:
2207                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
2208                         "xfs_log_recover: unknown buffer type 0x%x, logdev %s",
2209                         buf_f->blf_type, log->l_mp->m_logname ?
2210                         log->l_mp->m_logname : "internal");
2211                 XFS_ERROR_REPORT("xlog_recover_do_buffer_trans",
2212                                  XFS_ERRLEVEL_LOW, log->l_mp);
2213                 return XFS_ERROR(EFSCORRUPTED);
2214         }
2215
2216         mp = log->l_mp;
2217         if (flags & XFS_BLI_INODE_BUF) {
2218                 bp = xfs_buf_read_flags(mp->m_ddev_targp, blkno, len,
2219                                                                 XFS_BUF_LOCK);
2220         } else {
2221                 bp = xfs_buf_read(mp->m_ddev_targp, blkno, len, 0);
2222         }
2223         if (XFS_BUF_ISERROR(bp)) {
2224                 xfs_ioerror_alert("xlog_recover_do..(read#1)", log->l_mp,
2225                                   bp, blkno);
2226                 error = XFS_BUF_GETERROR(bp);
2227                 xfs_buf_relse(bp);
2228                 return error;
2229         }
2230
2231         error = 0;
2232         if (flags & XFS_BLI_INODE_BUF) {
2233                 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2234         } else if (flags &
2235                   (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
2236                 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2237         } else {
2238                 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2239         }
2240         if (error)
2241                 return XFS_ERROR(error);
2242
2243         /*
2244          * Perform delayed write on the buffer.  Asynchronous writes will be
2245          * slower when taking into account all the buffers to be flushed.
2246          *
2247          * Also make sure that only inode buffers with good sizes stay in
2248          * the buffer cache.  The kernel moves inodes in buffers of 1 block
2249          * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger.  The inode
2250          * buffers in the log can be a different size if the log was generated
2251          * by an older kernel using unclustered inode buffers or a newer kernel
2252          * running with a different inode cluster size.  Regardless, if the
2253          * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2254          * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2255          * the buffer out of the buffer cache so that the buffer won't
2256          * overlap with future reads of those inodes.
2257          */
2258         if (XFS_DINODE_MAGIC ==
2259             INT_GET(*((__uint16_t *)(xfs_buf_offset(bp, 0))), ARCH_CONVERT) &&
2260             (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2261                         (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2262                 XFS_BUF_STALE(bp);
2263                 error = xfs_bwrite(mp, bp);
2264         } else {
2265                 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2266                        XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2267                 XFS_BUF_SET_FSPRIVATE(bp, mp);
2268                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2269                 xfs_bdwrite(mp, bp);
2270         }
2271
2272         return (error);
2273 }
2274
2275 STATIC int
2276 xlog_recover_do_inode_trans(
2277         xlog_t                  *log,
2278         xlog_recover_item_t     *item,
2279         int                     pass)
2280 {
2281         xfs_inode_log_format_t  *in_f;
2282         xfs_mount_t             *mp;
2283         xfs_buf_t               *bp;
2284         xfs_imap_t              imap;
2285         xfs_dinode_t            *dip;
2286         xfs_ino_t               ino;
2287         int                     len;
2288         xfs_caddr_t             src;
2289         xfs_caddr_t             dest;
2290         int                     error;
2291         int                     attr_index;
2292         uint                    fields;
2293         xfs_dinode_core_t       *dicp;
2294         int                     need_free = 0;
2295
2296         if (pass == XLOG_RECOVER_PASS1) {
2297                 return 0;
2298         }
2299
2300         if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2301                 in_f = (xfs_inode_log_format_t *)item->ri_buf[0].i_addr;
2302         } else {
2303                 in_f = (xfs_inode_log_format_t *)kmem_alloc(
2304                         sizeof(xfs_inode_log_format_t), KM_SLEEP);
2305                 need_free = 1;
2306                 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2307                 if (error)
2308                         goto error;
2309         }
2310         ino = in_f->ilf_ino;
2311         mp = log->l_mp;
2312         if (ITEM_TYPE(item) == XFS_LI_INODE) {
2313                 imap.im_blkno = (xfs_daddr_t)in_f->ilf_blkno;
2314                 imap.im_len = in_f->ilf_len;
2315                 imap.im_boffset = in_f->ilf_boffset;
2316         } else {
2317                 /*
2318                  * It's an old inode format record.  We don't know where
2319                  * its cluster is located on disk, and we can't allow
2320                  * xfs_imap() to figure it out because the inode btrees
2321                  * are not ready to be used.  Therefore do not pass the
2322                  * XFS_IMAP_LOOKUP flag to xfs_imap().  This will give
2323                  * us only the single block in which the inode lives
2324                  * rather than its cluster, so we must make sure to
2325                  * invalidate the buffer when we write it out below.
2326                  */
2327                 imap.im_blkno = 0;
2328                 xfs_imap(log->l_mp, NULL, ino, &imap, 0);
2329         }
2330
2331         /*
2332          * Inode buffers can be freed, look out for it,
2333          * and do not replay the inode.
2334          */
2335         if (xlog_check_buffer_cancelled(log, imap.im_blkno, imap.im_len, 0)) {
2336                 error = 0;
2337                 goto error;
2338         }
2339
2340         bp = xfs_buf_read_flags(mp->m_ddev_targp, imap.im_blkno, imap.im_len,
2341                                                                 XFS_BUF_LOCK);
2342         if (XFS_BUF_ISERROR(bp)) {
2343                 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2344                                   bp, imap.im_blkno);
2345                 error = XFS_BUF_GETERROR(bp);
2346                 xfs_buf_relse(bp);
2347                 goto error;
2348         }
2349         error = 0;
2350         ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2351         dip = (xfs_dinode_t *)xfs_buf_offset(bp, imap.im_boffset);
2352
2353         /*
2354          * Make sure the place we're flushing out to really looks
2355          * like an inode!
2356          */
2357         if (unlikely(INT_GET(dip->di_core.di_magic, ARCH_CONVERT) != XFS_DINODE_MAGIC)) {
2358                 xfs_buf_relse(bp);
2359                 xfs_fs_cmn_err(CE_ALERT, mp,
2360                         "xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2361                         dip, bp, ino);
2362                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(1)",
2363                                  XFS_ERRLEVEL_LOW, mp);
2364                 error = EFSCORRUPTED;
2365                 goto error;
2366         }
2367         dicp = (xfs_dinode_core_t*)(item->ri_buf[1].i_addr);
2368         if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2369                 xfs_buf_relse(bp);
2370                 xfs_fs_cmn_err(CE_ALERT, mp,
2371                         "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2372                         item, ino);
2373                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(2)",
2374                                  XFS_ERRLEVEL_LOW, mp);
2375                 error = EFSCORRUPTED;
2376                 goto error;
2377         }
2378
2379         /* Skip replay when the on disk inode is newer than the log one */
2380         if (dicp->di_flushiter <
2381             INT_GET(dip->di_core.di_flushiter, ARCH_CONVERT)) {
2382                 /*
2383                  * Deal with the wrap case, DI_MAX_FLUSH is less
2384                  * than smaller numbers
2385                  */
2386                 if ((INT_GET(dip->di_core.di_flushiter, ARCH_CONVERT)
2387                                                         == DI_MAX_FLUSH) &&
2388                     (dicp->di_flushiter < (DI_MAX_FLUSH>>1))) {
2389                         /* do nothing */
2390                 } else {
2391                         xfs_buf_relse(bp);
2392                         error = 0;
2393                         goto error;
2394                 }
2395         }
2396         /* Take the opportunity to reset the flush iteration count */
2397         dicp->di_flushiter = 0;
2398
2399         if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2400                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2401                     (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2402                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(3)",
2403                                          XFS_ERRLEVEL_LOW, mp, dicp);
2404                         xfs_buf_relse(bp);
2405                         xfs_fs_cmn_err(CE_ALERT, mp,
2406                                 "xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2407                                 item, dip, bp, ino);
2408                         error = EFSCORRUPTED;
2409                         goto error;
2410                 }
2411         } else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2412                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2413                     (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2414                     (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2415                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(4)",
2416                                              XFS_ERRLEVEL_LOW, mp, dicp);
2417                         xfs_buf_relse(bp);
2418                         xfs_fs_cmn_err(CE_ALERT, mp,
2419                                 "xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2420                                 item, dip, bp, ino);
2421                         error = EFSCORRUPTED;
2422                         goto error;
2423                 }
2424         }
2425         if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2426                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(5)",
2427                                      XFS_ERRLEVEL_LOW, mp, dicp);
2428                 xfs_buf_relse(bp);
2429                 xfs_fs_cmn_err(CE_ALERT, mp,
2430                         "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",
2431                         item, dip, bp, ino,
2432                         dicp->di_nextents + dicp->di_anextents,
2433                         dicp->di_nblocks);
2434                 error = EFSCORRUPTED;
2435                 goto error;
2436         }
2437         if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2438                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(6)",
2439                                      XFS_ERRLEVEL_LOW, mp, dicp);
2440                 xfs_buf_relse(bp);
2441                 xfs_fs_cmn_err(CE_ALERT, mp,
2442                         "xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2443                         item, dip, bp, ino, dicp->di_forkoff);
2444                 error = EFSCORRUPTED;
2445                 goto error;
2446         }
2447         if (unlikely(item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t))) {
2448                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(7)",
2449                                      XFS_ERRLEVEL_LOW, mp, dicp);
2450                 xfs_buf_relse(bp);
2451                 xfs_fs_cmn_err(CE_ALERT, mp,
2452                         "xfs_inode_recover: Bad inode log record length %d, rec ptr 0x%p",
2453                         item->ri_buf[1].i_len, item);
2454                 error = EFSCORRUPTED;
2455                 goto error;
2456         }
2457
2458         /* The core is in in-core format */
2459         xfs_xlate_dinode_core((xfs_caddr_t)&dip->di_core,
2460                               (xfs_dinode_core_t*)item->ri_buf[1].i_addr, -1);
2461
2462         /* the rest is in on-disk format */
2463         if (item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t)) {
2464                 memcpy((xfs_caddr_t) dip + sizeof(xfs_dinode_core_t),
2465                         item->ri_buf[1].i_addr + sizeof(xfs_dinode_core_t),
2466                         item->ri_buf[1].i_len  - sizeof(xfs_dinode_core_t));
2467         }
2468
2469         fields = in_f->ilf_fields;
2470         switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2471         case XFS_ILOG_DEV:
2472                 INT_SET(dip->di_u.di_dev, ARCH_CONVERT, in_f->ilf_u.ilfu_rdev);
2473
2474                 break;
2475         case XFS_ILOG_UUID:
2476                 dip->di_u.di_muuid = in_f->ilf_u.ilfu_uuid;
2477                 break;
2478         }
2479
2480         if (in_f->ilf_size == 2)
2481                 goto write_inode_buffer;
2482         len = item->ri_buf[2].i_len;
2483         src = item->ri_buf[2].i_addr;
2484         ASSERT(in_f->ilf_size <= 4);
2485         ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2486         ASSERT(!(fields & XFS_ILOG_DFORK) ||
2487                (len == in_f->ilf_dsize));
2488
2489         switch (fields & XFS_ILOG_DFORK) {
2490         case XFS_ILOG_DDATA:
2491         case XFS_ILOG_DEXT:
2492                 memcpy(&dip->di_u, src, len);
2493                 break;
2494
2495         case XFS_ILOG_DBROOT:
2496                 xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2497                                  &(dip->di_u.di_bmbt),
2498                                  XFS_DFORK_DSIZE(dip, mp));
2499                 break;
2500
2501         default:
2502                 /*
2503                  * There are no data fork flags set.
2504                  */
2505                 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2506                 break;
2507         }
2508
2509         /*
2510          * If we logged any attribute data, recover it.  There may or
2511          * may not have been any other non-core data logged in this
2512          * transaction.
2513          */
2514         if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2515                 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2516                         attr_index = 3;
2517                 } else {
2518                         attr_index = 2;
2519                 }
2520                 len = item->ri_buf[attr_index].i_len;
2521                 src = item->ri_buf[attr_index].i_addr;
2522                 ASSERT(len == in_f->ilf_asize);
2523
2524                 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2525                 case XFS_ILOG_ADATA:
2526                 case XFS_ILOG_AEXT:
2527                         dest = XFS_DFORK_APTR(dip);
2528                         ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2529                         memcpy(dest, src, len);
2530                         break;
2531
2532                 case XFS_ILOG_ABROOT:
2533                         dest = XFS_DFORK_APTR(dip);
2534                         xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2535                                          (xfs_bmdr_block_t*)dest,
2536                                          XFS_DFORK_ASIZE(dip, mp));
2537                         break;
2538
2539                 default:
2540                         xlog_warn("XFS: xlog_recover_do_inode_trans: Invalid flag");
2541                         ASSERT(0);
2542                         xfs_buf_relse(bp);
2543                         error = EIO;
2544                         goto error;
2545                 }
2546         }
2547
2548 write_inode_buffer:
2549         if (ITEM_TYPE(item) == XFS_LI_INODE) {
2550                 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2551                        XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2552                 XFS_BUF_SET_FSPRIVATE(bp, mp);
2553                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2554                 xfs_bdwrite(mp, bp);
2555         } else {
2556                 XFS_BUF_STALE(bp);
2557                 error = xfs_bwrite(mp, bp);
2558         }
2559
2560 error:
2561         if (need_free)
2562                 kmem_free(in_f, sizeof(*in_f));
2563         return XFS_ERROR(error);
2564 }
2565
2566 /*
2567  * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2568  * structure, so that we know not to do any dquot item or dquot buffer recovery,
2569  * of that type.
2570  */
2571 STATIC int
2572 xlog_recover_do_quotaoff_trans(
2573         xlog_t                  *log,
2574         xlog_recover_item_t     *item,
2575         int                     pass)
2576 {
2577         xfs_qoff_logformat_t    *qoff_f;
2578
2579         if (pass == XLOG_RECOVER_PASS2) {
2580                 return (0);
2581         }
2582
2583         qoff_f = (xfs_qoff_logformat_t *)item->ri_buf[0].i_addr;
2584         ASSERT(qoff_f);
2585
2586         /*
2587          * The logitem format's flag tells us if this was user quotaoff,
2588          * group/project quotaoff or both.
2589          */
2590         if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2591                 log->l_quotaoffs_flag |= XFS_DQ_USER;
2592         if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2593                 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2594         if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2595                 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2596
2597         return (0);
2598 }
2599
2600 /*
2601  * Recover a dquot record
2602  */
2603 STATIC int
2604 xlog_recover_do_dquot_trans(
2605         xlog_t                  *log,
2606         xlog_recover_item_t     *item,
2607         int                     pass)
2608 {
2609         xfs_mount_t             *mp;
2610         xfs_buf_t               *bp;
2611         struct xfs_disk_dquot   *ddq, *recddq;
2612         int                     error;
2613         xfs_dq_logformat_t      *dq_f;
2614         uint                    type;
2615
2616         if (pass == XLOG_RECOVER_PASS1) {
2617                 return 0;
2618         }
2619         mp = log->l_mp;
2620
2621         /*
2622          * Filesystems are required to send in quota flags at mount time.
2623          */
2624         if (mp->m_qflags == 0)
2625                 return (0);
2626
2627         recddq = (xfs_disk_dquot_t *)item->ri_buf[1].i_addr;
2628         ASSERT(recddq);
2629         /*
2630          * This type of quotas was turned off, so ignore this record.
2631          */
2632         type = INT_GET(recddq->d_flags, ARCH_CONVERT) &
2633                         (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2634         ASSERT(type);
2635         if (log->l_quotaoffs_flag & type)
2636                 return (0);
2637
2638         /*
2639          * At this point we know that quota was _not_ turned off.
2640          * Since the mount flags are not indicating to us otherwise, this
2641          * must mean that quota is on, and the dquot needs to be replayed.
2642          * Remember that we may not have fully recovered the superblock yet,
2643          * so we can't do the usual trick of looking at the SB quota bits.
2644          *
2645          * The other possibility, of course, is that the quota subsystem was
2646          * removed since the last mount - ENOSYS.
2647          */
2648         dq_f = (xfs_dq_logformat_t *)item->ri_buf[0].i_addr;
2649         ASSERT(dq_f);
2650         if ((error = xfs_qm_dqcheck(recddq,
2651                            dq_f->qlf_id,
2652                            0, XFS_QMOPT_DOWARN,
2653                            "xlog_recover_do_dquot_trans (log copy)"))) {
2654                 return XFS_ERROR(EIO);
2655         }
2656         ASSERT(dq_f->qlf_len == 1);
2657
2658         error = xfs_read_buf(mp, mp->m_ddev_targp,
2659                              dq_f->qlf_blkno,
2660                              XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2661                              0, &bp);
2662         if (error) {
2663                 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2664                                   bp, dq_f->qlf_blkno);
2665                 return error;
2666         }
2667         ASSERT(bp);
2668         ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2669
2670         /*
2671          * At least the magic num portion should be on disk because this
2672          * was among a chunk of dquots created earlier, and we did some
2673          * minimal initialization then.
2674          */
2675         if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2676                            "xlog_recover_do_dquot_trans")) {
2677                 xfs_buf_relse(bp);
2678                 return XFS_ERROR(EIO);
2679         }
2680
2681         memcpy(ddq, recddq, item->ri_buf[1].i_len);
2682
2683         ASSERT(dq_f->qlf_size == 2);
2684         ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2685                XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2686         XFS_BUF_SET_FSPRIVATE(bp, mp);
2687         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2688         xfs_bdwrite(mp, bp);
2689
2690         return (0);
2691 }
2692
2693 /*
2694  * This routine is called to create an in-core extent free intent
2695  * item from the efi format structure which was logged on disk.
2696  * It allocates an in-core efi, copies the extents from the format
2697  * structure into it, and adds the efi to the AIL with the given
2698  * LSN.
2699  */
2700 STATIC int
2701 xlog_recover_do_efi_trans(
2702         xlog_t                  *log,
2703         xlog_recover_item_t     *item,
2704         xfs_lsn_t               lsn,
2705         int                     pass)
2706 {
2707         int                     error;
2708         xfs_mount_t             *mp;
2709         xfs_efi_log_item_t      *efip;
2710         xfs_efi_log_format_t    *efi_formatp;
2711         SPLDECL(s);
2712
2713         if (pass == XLOG_RECOVER_PASS1) {
2714                 return 0;
2715         }
2716
2717         efi_formatp = (xfs_efi_log_format_t *)item->ri_buf[0].i_addr;
2718
2719         mp = log->l_mp;
2720         efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2721         if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2722                                          &(efip->efi_format)))) {
2723                 xfs_efi_item_free(efip);
2724                 return error;
2725         }
2726         efip->efi_next_extent = efi_formatp->efi_nextents;
2727         efip->efi_flags |= XFS_EFI_COMMITTED;
2728
2729         AIL_LOCK(mp,s);
2730         /*
2731          * xfs_trans_update_ail() drops the AIL lock.
2732          */
2733         xfs_trans_update_ail(mp, (xfs_log_item_t *)efip, lsn, s);
2734         return 0;
2735 }
2736
2737
2738 /*
2739  * This routine is called when an efd format structure is found in
2740  * a committed transaction in the log.  It's purpose is to cancel
2741  * the corresponding efi if it was still in the log.  To do this
2742  * it searches the AIL for the efi with an id equal to that in the
2743  * efd format structure.  If we find it, we remove the efi from the
2744  * AIL and free it.
2745  */
2746 STATIC void
2747 xlog_recover_do_efd_trans(
2748         xlog_t                  *log,
2749         xlog_recover_item_t     *item,
2750         int                     pass)
2751 {
2752         xfs_mount_t             *mp;
2753         xfs_efd_log_format_t    *efd_formatp;
2754         xfs_efi_log_item_t      *efip = NULL;
2755         xfs_log_item_t          *lip;
2756         int                     gen;
2757         __uint64_t              efi_id;
2758         SPLDECL(s);
2759
2760         if (pass == XLOG_RECOVER_PASS1) {
2761                 return;
2762         }
2763
2764         efd_formatp = (xfs_efd_log_format_t *)item->ri_buf[0].i_addr;
2765         ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2766                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2767                (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2768                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2769         efi_id = efd_formatp->efd_efi_id;
2770
2771         /*
2772          * Search for the efi with the id in the efd format structure
2773          * in the AIL.
2774          */
2775         mp = log->l_mp;
2776         AIL_LOCK(mp,s);
2777         lip = xfs_trans_first_ail(mp, &gen);
2778         while (lip != NULL) {
2779                 if (lip->li_type == XFS_LI_EFI) {
2780                         efip = (xfs_efi_log_item_t *)lip;
2781                         if (efip->efi_format.efi_id == efi_id) {
2782                                 /*
2783                                  * xfs_trans_delete_ail() drops the
2784                                  * AIL lock.
2785                                  */
2786                                 xfs_trans_delete_ail(mp, lip, s);
2787                                 break;
2788                         }
2789                 }
2790                 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
2791         }
2792
2793         /*
2794          * If we found it, then free it up.  If it wasn't there, it
2795          * must have been overwritten in the log.  Oh well.
2796          */
2797         if (lip != NULL) {
2798                 xfs_efi_item_free(efip);
2799         } else {
2800                 AIL_UNLOCK(mp, s);
2801         }
2802 }
2803
2804 /*
2805  * Perform the transaction
2806  *
2807  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2808  * EFIs and EFDs get queued up by adding entries into the AIL for them.
2809  */
2810 STATIC int
2811 xlog_recover_do_trans(
2812         xlog_t                  *log,
2813         xlog_recover_t          *trans,
2814         int                     pass)
2815 {
2816         int                     error = 0;
2817         xlog_recover_item_t     *item, *first_item;
2818
2819         if ((error = xlog_recover_reorder_trans(trans)))
2820                 return error;
2821         first_item = item = trans->r_itemq;
2822         do {
2823                 /*
2824                  * we don't need to worry about the block number being
2825                  * truncated in > 1 TB buffers because in user-land,
2826                  * we're now n32 or 64-bit so xfs_daddr_t is 64-bits so
2827                  * the blknos will get through the user-mode buffer
2828                  * cache properly.  The only bad case is o32 kernels
2829                  * where xfs_daddr_t is 32-bits but mount will warn us
2830                  * off a > 1 TB filesystem before we get here.
2831                  */
2832                 if ((ITEM_TYPE(item) == XFS_LI_BUF)) {
2833                         if  ((error = xlog_recover_do_buffer_trans(log, item,
2834                                                                  pass)))
2835                                 break;
2836                 } else if ((ITEM_TYPE(item) == XFS_LI_INODE)) {
2837                         if ((error = xlog_recover_do_inode_trans(log, item,
2838                                                                 pass)))
2839                                 break;
2840                 } else if (ITEM_TYPE(item) == XFS_LI_EFI) {
2841                         if ((error = xlog_recover_do_efi_trans(log, item, trans->r_lsn,
2842                                                   pass)))
2843                                 break;
2844                 } else if (ITEM_TYPE(item) == XFS_LI_EFD) {
2845                         xlog_recover_do_efd_trans(log, item, pass);
2846                 } else if (ITEM_TYPE(item) == XFS_LI_DQUOT) {
2847                         if ((error = xlog_recover_do_dquot_trans(log, item,
2848                                                                    pass)))
2849                                         break;
2850                 } else if ((ITEM_TYPE(item) == XFS_LI_QUOTAOFF)) {
2851                         if ((error = xlog_recover_do_quotaoff_trans(log, item,
2852                                                                    pass)))
2853                                         break;
2854                 } else {
2855                         xlog_warn("XFS: xlog_recover_do_trans");
2856                         ASSERT(0);
2857                         error = XFS_ERROR(EIO);
2858                         break;
2859                 }
2860                 item = item->ri_next;
2861         } while (first_item != item);
2862
2863         return error;
2864 }
2865
2866 /*
2867  * Free up any resources allocated by the transaction
2868  *
2869  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2870  */
2871 STATIC void
2872 xlog_recover_free_trans(
2873         xlog_recover_t          *trans)
2874 {
2875         xlog_recover_item_t     *first_item, *item, *free_item;
2876         int                     i;
2877
2878         item = first_item = trans->r_itemq;
2879         do {
2880                 free_item = item;
2881                 item = item->ri_next;
2882                  /* Free the regions in the item. */
2883                 for (i = 0; i < free_item->ri_cnt; i++) {
2884                         kmem_free(free_item->ri_buf[i].i_addr,
2885                                   free_item->ri_buf[i].i_len);
2886                 }
2887                 /* Free the item itself */
2888                 kmem_free(free_item->ri_buf,
2889                           (free_item->ri_total * sizeof(xfs_log_iovec_t)));
2890                 kmem_free(free_item, sizeof(xlog_recover_item_t));
2891         } while (first_item != item);
2892         /* Free the transaction recover structure */
2893         kmem_free(trans, sizeof(xlog_recover_t));
2894 }
2895
2896 STATIC int
2897 xlog_recover_commit_trans(
2898         xlog_t                  *log,
2899         xlog_recover_t          **q,
2900         xlog_recover_t          *trans,
2901         int                     pass)
2902 {
2903         int                     error;
2904
2905         if ((error = xlog_recover_unlink_tid(q, trans)))
2906                 return error;
2907         if ((error = xlog_recover_do_trans(log, trans, pass)))
2908                 return error;
2909         xlog_recover_free_trans(trans);                 /* no error */
2910         return 0;
2911 }
2912
2913 STATIC int
2914 xlog_recover_unmount_trans(
2915         xlog_recover_t          *trans)
2916 {
2917         /* Do nothing now */
2918         xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2919         return 0;
2920 }
2921
2922 /*
2923  * There are two valid states of the r_state field.  0 indicates that the
2924  * transaction structure is in a normal state.  We have either seen the
2925  * start of the transaction or the last operation we added was not a partial
2926  * operation.  If the last operation we added to the transaction was a
2927  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2928  *
2929  * NOTE: skip LRs with 0 data length.
2930  */
2931 STATIC int
2932 xlog_recover_process_data(
2933         xlog_t                  *log,
2934         xlog_recover_t          *rhash[],
2935         xlog_rec_header_t       *rhead,
2936         xfs_caddr_t             dp,
2937         int                     pass)
2938 {
2939         xfs_caddr_t             lp;
2940         int                     num_logops;
2941         xlog_op_header_t        *ohead;
2942         xlog_recover_t          *trans;
2943         xlog_tid_t              tid;
2944         int                     error;
2945         unsigned long           hash;
2946         uint                    flags;
2947
2948         lp = dp + INT_GET(rhead->h_len, ARCH_CONVERT);
2949         num_logops = INT_GET(rhead->h_num_logops, ARCH_CONVERT);
2950
2951         /* check the log format matches our own - else we can't recover */
2952         if (xlog_header_check_recover(log->l_mp, rhead))
2953                 return (XFS_ERROR(EIO));
2954
2955         while ((dp < lp) && num_logops) {
2956                 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2957                 ohead = (xlog_op_header_t *)dp;
2958                 dp += sizeof(xlog_op_header_t);
2959                 if (ohead->oh_clientid != XFS_TRANSACTION &&
2960                     ohead->oh_clientid != XFS_LOG) {
2961                         xlog_warn(
2962                 "XFS: xlog_recover_process_data: bad clientid");
2963                         ASSERT(0);
2964                         return (XFS_ERROR(EIO));
2965                 }
2966                 tid = INT_GET(ohead->oh_tid, ARCH_CONVERT);
2967                 hash = XLOG_RHASH(tid);
2968                 trans = xlog_recover_find_tid(rhash[hash], tid);
2969                 if (trans == NULL) {               /* not found; add new tid */
2970                         if (ohead->oh_flags & XLOG_START_TRANS)
2971                                 xlog_recover_new_tid(&rhash[hash], tid,
2972                                         INT_GET(rhead->h_lsn, ARCH_CONVERT));
2973                 } else {
2974                         ASSERT(dp+INT_GET(ohead->oh_len, ARCH_CONVERT) <= lp);
2975                         flags = ohead->oh_flags & ~XLOG_END_TRANS;
2976                         if (flags & XLOG_WAS_CONT_TRANS)
2977                                 flags &= ~XLOG_CONTINUE_TRANS;
2978                         switch (flags) {
2979                         case XLOG_COMMIT_TRANS:
2980                                 error = xlog_recover_commit_trans(log,
2981                                                 &rhash[hash], trans, pass);
2982                                 break;
2983                         case XLOG_UNMOUNT_TRANS:
2984                                 error = xlog_recover_unmount_trans(trans);
2985                                 break;
2986                         case XLOG_WAS_CONT_TRANS:
2987                                 error = xlog_recover_add_to_cont_trans(trans,
2988                                                 dp, INT_GET(ohead->oh_len,
2989                                                         ARCH_CONVERT));
2990                                 break;
2991                         case XLOG_START_TRANS:
2992                                 xlog_warn(
2993                         "XFS: xlog_recover_process_data: bad transaction");
2994                                 ASSERT(0);
2995                                 error = XFS_ERROR(EIO);
2996                                 break;
2997                         case 0:
2998                         case XLOG_CONTINUE_TRANS:
2999                                 error = xlog_recover_add_to_trans(trans,
3000                                                 dp, INT_GET(ohead->oh_len,
3001                                                         ARCH_CONVERT));
3002                                 break;
3003                         default:
3004                                 xlog_warn(
3005                         "XFS: xlog_recover_process_data: bad flag");
3006                                 ASSERT(0);
3007                                 error = XFS_ERROR(EIO);
3008                                 break;
3009                         }
3010                         if (error)
3011                                 return error;
3012                 }
3013                 dp += INT_GET(ohead->oh_len, ARCH_CONVERT);
3014                 num_logops--;
3015         }
3016         return 0;
3017 }
3018
3019 /*
3020  * Process an extent free intent item that was recovered from
3021  * the log.  We need to free the extents that it describes.
3022  */
3023 STATIC void
3024 xlog_recover_process_efi(
3025         xfs_mount_t             *mp,
3026         xfs_efi_log_item_t      *efip)
3027 {
3028         xfs_efd_log_item_t      *efdp;
3029         xfs_trans_t             *tp;
3030         int                     i;
3031         xfs_extent_t            *extp;
3032         xfs_fsblock_t           startblock_fsb;
3033
3034         ASSERT(!(efip->efi_flags & XFS_EFI_RECOVERED));
3035
3036         /*
3037          * First check the validity of the extents described by the
3038          * EFI.  If any are bad, then assume that all are bad and
3039          * just toss the EFI.
3040          */
3041         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3042                 extp = &(efip->efi_format.efi_extents[i]);
3043                 startblock_fsb = XFS_BB_TO_FSB(mp,
3044                                    XFS_FSB_TO_DADDR(mp, extp->ext_start));
3045                 if ((startblock_fsb == 0) ||
3046                     (extp->ext_len == 0) ||
3047                     (startblock_fsb >= mp->m_sb.sb_dblocks) ||
3048                     (extp->ext_len >= mp->m_sb.sb_agblocks)) {
3049                         /*
3050                          * This will pull the EFI from the AIL and
3051                          * free the memory associated with it.
3052                          */
3053                         xfs_efi_release(efip, efip->efi_format.efi_nextents);
3054                         return;
3055                 }
3056         }
3057
3058         tp = xfs_trans_alloc(mp, 0);
3059         xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
3060         efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3061
3062         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3063                 extp = &(efip->efi_format.efi_extents[i]);
3064                 xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3065                 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3066                                          extp->ext_len);
3067         }
3068
3069         efip->efi_flags |= XFS_EFI_RECOVERED;
3070         xfs_trans_commit(tp, 0);
3071 }
3072
3073 /*
3074  * Verify that once we've encountered something other than an EFI
3075  * in the AIL that there are no more EFIs in the AIL.
3076  */
3077 #if defined(DEBUG)
3078 STATIC void
3079 xlog_recover_check_ail(
3080         xfs_mount_t             *mp,
3081         xfs_log_item_t          *lip,
3082         int                     gen)
3083 {
3084         int                     orig_gen = gen;
3085
3086         do {
3087                 ASSERT(lip->li_type != XFS_LI_EFI);
3088                 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3089                 /*
3090                  * The check will be bogus if we restart from the
3091                  * beginning of the AIL, so ASSERT that we don't.
3092                  * We never should since we're holding the AIL lock
3093                  * the entire time.
3094                  */
3095                 ASSERT(gen == orig_gen);
3096         } while (lip != NULL);
3097 }
3098 #endif  /* DEBUG */
3099
3100 /*
3101  * When this is called, all of the EFIs which did not have
3102  * corresponding EFDs should be in the AIL.  What we do now
3103  * is free the extents associated with each one.
3104  *
3105  * Since we process the EFIs in normal transactions, they
3106  * will be removed at some point after the commit.  This prevents
3107  * us from just walking down the list processing each one.
3108  * We'll use a flag in the EFI to skip those that we've already
3109  * processed and use the AIL iteration mechanism's generation
3110  * count to try to speed this up at least a bit.
3111  *
3112  * When we start, we know that the EFIs are the only things in
3113  * the AIL.  As we process them, however, other items are added
3114  * to the AIL.  Since everything added to the AIL must come after
3115  * everything already in the AIL, we stop processing as soon as
3116  * we see something other than an EFI in the AIL.
3117  */
3118 STATIC void
3119 xlog_recover_process_efis(
3120         xlog_t                  *log)
3121 {
3122         xfs_log_item_t          *lip;
3123         xfs_efi_log_item_t      *efip;
3124         int                     gen;
3125         xfs_mount_t             *mp;
3126         SPLDECL(s);
3127
3128         mp = log->l_mp;
3129         AIL_LOCK(mp,s);
3130
3131         lip = xfs_trans_first_ail(mp, &gen);
3132         while (lip != NULL) {
3133                 /*
3134                  * We're done when we see something other than an EFI.
3135                  */
3136                 if (lip->li_type != XFS_LI_EFI) {
3137                         xlog_recover_check_ail(mp, lip, gen);
3138                         break;
3139                 }
3140
3141                 /*
3142                  * Skip EFIs that we've already processed.
3143                  */
3144                 efip = (xfs_efi_log_item_t *)lip;
3145                 if (efip->efi_flags & XFS_EFI_RECOVERED) {
3146                         lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3147                         continue;
3148                 }
3149
3150                 AIL_UNLOCK(mp, s);
3151                 xlog_recover_process_efi(mp, efip);
3152                 AIL_LOCK(mp,s);
3153                 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3154         }
3155         AIL_UNLOCK(mp, s);
3156 }
3157
3158 /*
3159  * This routine performs a transaction to null out a bad inode pointer
3160  * in an agi unlinked inode hash bucket.
3161  */
3162 STATIC void
3163 xlog_recover_clear_agi_bucket(
3164         xfs_mount_t     *mp,
3165         xfs_agnumber_t  agno,
3166         int             bucket)
3167 {
3168         xfs_trans_t     *tp;
3169         xfs_agi_t       *agi;
3170         xfs_buf_t       *agibp;
3171         int             offset;
3172         int             error;
3173
3174         tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3175         xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp), 0, 0, 0);
3176
3177         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
3178                                    XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3179                                    XFS_FSS_TO_BB(mp, 1), 0, &agibp);
3180         if (error) {
3181                 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3182                 return;
3183         }
3184
3185         agi = XFS_BUF_TO_AGI(agibp);
3186         if (be32_to_cpu(agi->agi_magicnum) != XFS_AGI_MAGIC) {
3187                 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3188                 return;
3189         }
3190
3191         agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3192         offset = offsetof(xfs_agi_t, agi_unlinked) +
3193                  (sizeof(xfs_agino_t) * bucket);
3194         xfs_trans_log_buf(tp, agibp, offset,
3195                           (offset + sizeof(xfs_agino_t) - 1));
3196
3197         (void) xfs_trans_commit(tp, 0);
3198 }
3199
3200 /*
3201  * xlog_iunlink_recover
3202  *
3203  * This is called during recovery to process any inodes which
3204  * we unlinked but not freed when the system crashed.  These
3205  * inodes will be on the lists in the AGI blocks.  What we do
3206  * here is scan all the AGIs and fully truncate and free any
3207  * inodes found on the lists.  Each inode is removed from the
3208  * lists when it has been fully truncated and is freed.  The
3209  * freeing of the inode and its removal from the list must be
3210  * atomic.
3211  */
3212 void
3213 xlog_recover_process_iunlinks(
3214         xlog_t          *log)
3215 {
3216         xfs_mount_t     *mp;
3217         xfs_agnumber_t  agno;
3218         xfs_agi_t       *agi;
3219         xfs_buf_t       *agibp;
3220         xfs_buf_t       *ibp;
3221         xfs_dinode_t    *dip;
3222         xfs_inode_t     *ip;
3223         xfs_agino_t     agino;
3224         xfs_ino_t       ino;
3225         int             bucket;
3226         int             error;
3227         uint            mp_dmevmask;
3228
3229         mp = log->l_mp;
3230
3231         /*
3232          * Prevent any DMAPI event from being sent while in this function.
3233          */
3234         mp_dmevmask = mp->m_dmevmask;
3235         mp->m_dmevmask = 0;
3236
3237         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3238                 /*
3239                  * Find the agi for this ag.
3240                  */
3241                 agibp = xfs_buf_read(mp->m_ddev_targp,
3242                                 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3243                                 XFS_FSS_TO_BB(mp, 1), 0);
3244                 if (XFS_BUF_ISERROR(agibp)) {
3245                         xfs_ioerror_alert("xlog_recover_process_iunlinks(#1)",
3246                                 log->l_mp, agibp,
3247                                 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)));
3248                 }
3249                 agi = XFS_BUF_TO_AGI(agibp);
3250                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agi->agi_magicnum));
3251
3252                 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3253
3254                         agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3255                         while (agino != NULLAGINO) {
3256
3257                                 /*
3258                                  * Release the agi buffer so that it can
3259                                  * be acquired in the normal course of the
3260                                  * transaction to truncate and free the inode.
3261                                  */
3262                                 xfs_buf_relse(agibp);
3263
3264                                 ino = XFS_AGINO_TO_INO(mp, agno, agino);
3265                                 error = xfs_iget(mp, NULL, ino, 0, 0, &ip, 0);
3266                                 ASSERT(error || (ip != NULL));
3267
3268                                 if (!error) {
3269                                         /*
3270                                          * Get the on disk inode to find the
3271                                          * next inode in the bucket.
3272                                          */
3273                                         error = xfs_itobp(mp, NULL, ip, &dip,
3274                                                         &ibp, 0, 0);
3275                                         ASSERT(error || (dip != NULL));
3276                                 }
3277
3278                                 if (!error) {
3279                                         ASSERT(ip->i_d.di_nlink == 0);
3280
3281                                         /* setup for the next pass */
3282                                         agino = INT_GET(dip->di_next_unlinked,
3283                                                         ARCH_CONVERT);
3284                                         xfs_buf_relse(ibp);
3285                                         /*
3286                                          * Prevent any DMAPI event from
3287                                          * being sent when the
3288                                          * reference on the inode is
3289                                          * dropped.
3290                                          */
3291                                         ip->i_d.di_dmevmask = 0;
3292
3293                                         /*
3294                                          * If this is a new inode, handle
3295                                          * it specially.  Otherwise,
3296                                          * just drop our reference to the
3297                                          * inode.  If there are no
3298                                          * other references, this will
3299                                          * send the inode to
3300                                          * xfs_inactive() which will
3301                                          * truncate the file and free
3302                                          * the inode.
3303                                          */
3304                                         if (ip->i_d.di_mode == 0)
3305                                                 xfs_iput_new(ip, 0);
3306                                         else
3307                                                 VN_RELE(XFS_ITOV(ip));
3308                                 } else {
3309                                         /*
3310                                          * We can't read in the inode
3311                                          * this bucket points to, or
3312                                          * this inode is messed up.  Just
3313                                          * ditch this bucket of inodes.  We
3314                                          * will lose some inodes and space,
3315                                          * but at least we won't hang.  Call
3316                                          * xlog_recover_clear_agi_bucket()
3317                                          * to perform a transaction to clear
3318                                          * the inode pointer in the bucket.
3319                                          */
3320                                         xlog_recover_clear_agi_bucket(mp, agno,
3321                                                         bucket);
3322
3323                                         agino = NULLAGINO;
3324                                 }
3325
3326                                 /*
3327                                  * Reacquire the agibuffer and continue around
3328                                  * the loop.
3329                                  */
3330                                 agibp = xfs_buf_read(mp->m_ddev_targp,
3331                                                 XFS_AG_DADDR(mp, agno,
3332                                                         XFS_AGI_DADDR(mp)),
3333                                                 XFS_FSS_TO_BB(mp, 1), 0);
3334                                 if (XFS_BUF_ISERROR(agibp)) {
3335                                         xfs_ioerror_alert(
3336                                 "xlog_recover_process_iunlinks(#2)",
3337                                                 log->l_mp, agibp,
3338                                                 XFS_AG_DADDR(mp, agno,
3339                                                         XFS_AGI_DADDR(mp)));
3340                                 }
3341                                 agi = XFS_BUF_TO_AGI(agibp);
3342                                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(
3343                                         agi->agi_magicnum));
3344                         }
3345                 }
3346
3347                 /*
3348                  * Release the buffer for the current agi so we can
3349                  * go on to the next one.
3350                  */
3351                 xfs_buf_relse(agibp);
3352         }
3353
3354         mp->m_dmevmask = mp_dmevmask;
3355 }
3356
3357
3358 #ifdef DEBUG
3359 STATIC void
3360 xlog_pack_data_checksum(
3361         xlog_t          *log,
3362         xlog_in_core_t  *iclog,
3363         int             size)
3364 {
3365         int             i;
3366         uint            *up;
3367         uint            chksum = 0;
3368
3369         up = (uint *)iclog->ic_datap;
3370         /* divide length by 4 to get # words */
3371         for (i = 0; i < (size >> 2); i++) {
3372                 chksum ^= INT_GET(*up, ARCH_CONVERT);
3373                 up++;
3374         }
3375         INT_SET(iclog->ic_header.h_chksum, ARCH_CONVERT, chksum);
3376 }
3377 #else
3378 #define xlog_pack_data_checksum(log, iclog, size)
3379 #endif
3380
3381 /*
3382  * Stamp cycle number in every block
3383  */
3384 void
3385 xlog_pack_data(
3386         xlog_t                  *log,
3387         xlog_in_core_t          *iclog,
3388         int                     roundoff)
3389 {
3390         int                     i, j, k;
3391         int                     size = iclog->ic_offset + roundoff;
3392         uint                    cycle_lsn;
3393         xfs_caddr_t             dp;
3394         xlog_in_core_2_t        *xhdr;
3395
3396         xlog_pack_data_checksum(log, iclog, size);
3397
3398         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3399
3400         dp = iclog->ic_datap;
3401         for (i = 0; i < BTOBB(size) &&
3402                 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3403                 iclog->ic_header.h_cycle_data[i] = *(uint *)dp;
3404                 *(uint *)dp = cycle_lsn;
3405                 dp += BBSIZE;
3406         }
3407
3408         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3409                 xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3410                 for ( ; i < BTOBB(size); i++) {
3411                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3412                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3413                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(uint *)dp;
3414                         *(uint *)dp = cycle_lsn;
3415                         dp += BBSIZE;
3416                 }
3417
3418                 for (i = 1; i < log->l_iclog_heads; i++) {
3419                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3420                 }
3421         }
3422 }
3423
3424 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
3425 STATIC void
3426 xlog_unpack_data_checksum(
3427         xlog_rec_header_t       *rhead,
3428         xfs_caddr_t             dp,
3429         xlog_t                  *log)
3430 {
3431         uint                    *up = (uint *)dp;
3432         uint                    chksum = 0;
3433         int                     i;
3434
3435         /* divide length by 4 to get # words */
3436         for (i=0; i < INT_GET(rhead->h_len, ARCH_CONVERT) >> 2; i++) {
3437                 chksum ^= INT_GET(*up, ARCH_CONVERT);
3438                 up++;
3439         }
3440         if (chksum != INT_GET(rhead->h_chksum, ARCH_CONVERT)) {
3441             if (rhead->h_chksum ||
3442                 ((log->l_flags & XLOG_CHKSUM_MISMATCH) == 0)) {
3443                     cmn_err(CE_DEBUG,
3444                         "XFS: LogR chksum mismatch: was (0x%x) is (0x%x)\n",
3445                             INT_GET(rhead->h_chksum, ARCH_CONVERT), chksum);
3446                     cmn_err(CE_DEBUG,
3447 "XFS: Disregard message if filesystem was created with non-DEBUG kernel");
3448                     if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3449                             cmn_err(CE_DEBUG,
3450                                 "XFS: LogR this is a LogV2 filesystem\n");
3451                     }
3452                     log->l_flags |= XLOG_CHKSUM_MISMATCH;
3453             }
3454         }
3455 }
3456 #else
3457 #define xlog_unpack_data_checksum(rhead, dp, log)
3458 #endif
3459
3460 STATIC void
3461 xlog_unpack_data(
3462         xlog_rec_header_t       *rhead,
3463         xfs_caddr_t             dp,
3464         xlog_t                  *log)
3465 {
3466         int                     i, j, k;
3467         xlog_in_core_2_t        *xhdr;
3468
3469         for (i = 0; i < BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT)) &&
3470                   i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3471                 *(uint *)dp = *(uint *)&rhead->h_cycle_data[i];
3472                 dp += BBSIZE;
3473         }
3474
3475         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3476                 xhdr = (xlog_in_core_2_t *)rhead;
3477                 for ( ; i < BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT)); i++) {
3478                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3479                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3480                         *(uint *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3481                         dp += BBSIZE;
3482                 }
3483         }
3484
3485         xlog_unpack_data_checksum(rhead, dp, log);
3486 }
3487
3488 STATIC int
3489 xlog_valid_rec_header(
3490         xlog_t                  *log,
3491         xlog_rec_header_t       *rhead,
3492         xfs_daddr_t             blkno)
3493 {
3494         int                     hlen;
3495
3496         if (unlikely(
3497             (INT_GET(rhead->h_magicno, ARCH_CONVERT) !=
3498                         XLOG_HEADER_MAGIC_NUM))) {
3499                 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3500                                 XFS_ERRLEVEL_LOW, log->l_mp);
3501                 return XFS_ERROR(EFSCORRUPTED);
3502         }
3503         if (unlikely(
3504             (!rhead->h_version ||
3505             (INT_GET(rhead->h_version, ARCH_CONVERT) &
3506                         (~XLOG_VERSION_OKBITS)) != 0))) {
3507                 xlog_warn("XFS: %s: unrecognised log version (%d).",
3508                         __FUNCTION__, INT_GET(rhead->h_version, ARCH_CONVERT));
3509                 return XFS_ERROR(EIO);
3510         }
3511
3512         /* LR body must have data or it wouldn't have been written */
3513         hlen = INT_GET(rhead->h_len, ARCH_CONVERT);
3514         if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3515                 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3516                                 XFS_ERRLEVEL_LOW, log->l_mp);
3517                 return XFS_ERROR(EFSCORRUPTED);
3518         }
3519         if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3520                 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3521                                 XFS_ERRLEVEL_LOW, log->l_mp);
3522                 return XFS_ERROR(EFSCORRUPTED);
3523         }
3524         return 0;
3525 }
3526
3527 /*
3528  * Read the log from tail to head and process the log records found.
3529  * Handle the two cases where the tail and head are in the same cycle
3530  * and where the active portion of the log wraps around the end of
3531  * the physical log separately.  The pass parameter is passed through
3532  * to the routines called to process the data and is not looked at
3533  * here.
3534  */
3535 STATIC int
3536 xlog_do_recovery_pass(
3537         xlog_t                  *log,
3538         xfs_daddr_t             head_blk,
3539         xfs_daddr_t             tail_blk,
3540         int                     pass)
3541 {
3542         xlog_rec_header_t       *rhead;
3543         xfs_daddr_t             blk_no;
3544         xfs_caddr_t             bufaddr, offset;
3545         xfs_buf_t               *hbp, *dbp;
3546         int                     error = 0, h_size;
3547         int                     bblks, split_bblks;
3548         int                     hblks, split_hblks, wrapped_hblks;
3549         xlog_recover_t          *rhash[XLOG_RHASH_SIZE];
3550
3551         ASSERT(head_blk != tail_blk);
3552
3553         /*
3554          * Read the header of the tail block and get the iclog buffer size from
3555          * h_size.  Use this to tell how many sectors make up the log header.
3556          */
3557         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3558                 /*
3559                  * When using variable length iclogs, read first sector of
3560                  * iclog header and extract the header size from it.  Get a
3561                  * new hbp that is the correct size.
3562                  */
3563                 hbp = xlog_get_bp(log, 1);
3564                 if (!hbp)
3565                         return ENOMEM;
3566                 if ((error = xlog_bread(log, tail_blk, 1, hbp)))
3567                         goto bread_err1;
3568                 offset = xlog_align(log, tail_blk, 1, hbp);
3569                 rhead = (xlog_rec_header_t *)offset;
3570                 error = xlog_valid_rec_header(log, rhead, tail_blk);
3571                 if (error)
3572                         goto bread_err1;
3573                 h_size = INT_GET(rhead->h_size, ARCH_CONVERT);
3574                 if ((INT_GET(rhead->h_version, ARCH_CONVERT)
3575                                 & XLOG_VERSION_2) &&
3576                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3577                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3578                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
3579                                 hblks++;
3580                         xlog_put_bp(hbp);
3581                         hbp = xlog_get_bp(log, hblks);
3582                 } else {
3583                         hblks = 1;
3584                 }
3585         } else {
3586                 ASSERT(log->l_sectbb_log == 0);
3587                 hblks = 1;
3588                 hbp = xlog_get_bp(log, 1);
3589                 h_size = XLOG_BIG_RECORD_BSIZE;
3590         }
3591
3592         if (!hbp)
3593                 return ENOMEM;
3594         dbp = xlog_get_bp(log, BTOBB(h_size));
3595         if (!dbp) {
3596                 xlog_put_bp(hbp);
3597                 return ENOMEM;
3598         }
3599
3600         memset(rhash, 0, sizeof(rhash));
3601         if (tail_blk <= head_blk) {
3602                 for (blk_no = tail_blk; blk_no < head_blk; ) {
3603                         if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3604                                 goto bread_err2;
3605                         offset = xlog_align(log, blk_no, hblks, hbp);
3606                         rhead = (xlog_rec_header_t *)offset;
3607                         error = xlog_valid_rec_header(log, rhead, blk_no);
3608                         if (error)
3609                                 goto bread_err2;
3610
3611                         /* blocks in data section */
3612                         bblks = (int)BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT));
3613                         error = xlog_bread(log, blk_no + hblks, bblks, dbp);
3614                         if (error)
3615                                 goto bread_err2;
3616                         offset = xlog_align(log, blk_no + hblks, bblks, dbp);
3617                         xlog_unpack_data(rhead, offset, log);
3618                         if ((error = xlog_recover_process_data(log,
3619                                                 rhash, rhead, offset, pass)))
3620                                 goto bread_err2;
3621                         blk_no += bblks + hblks;
3622                 }
3623         } else {
3624                 /*
3625                  * Perform recovery around the end of the physical log.
3626                  * When the head is not on the same cycle number as the tail,
3627                  * we can't do a sequential recovery as above.
3628                  */
3629                 blk_no = tail_blk;
3630                 while (blk_no < log->l_logBBsize) {
3631                         /*
3632                          * Check for header wrapping around physical end-of-log
3633                          */
3634                         offset = NULL;
3635                         split_hblks = 0;
3636                         wrapped_hblks = 0;
3637                         if (blk_no + hblks <= log->l_logBBsize) {
3638                                 /* Read header in one read */
3639                                 error = xlog_bread(log, blk_no, hblks, hbp);
3640                                 if (error)
3641                                         goto bread_err2;
3642                                 offset = xlog_align(log, blk_no, hblks, hbp);
3643                         } else {
3644                                 /* This LR is split across physical log end */
3645                                 if (blk_no != log->l_logBBsize) {
3646                                         /* some data before physical log end */
3647                                         ASSERT(blk_no <= INT_MAX);
3648                                         split_hblks = log->l_logBBsize - (int)blk_no;
3649                                         ASSERT(split_hblks > 0);
3650                                         if ((error = xlog_bread(log, blk_no,
3651                                                         split_hblks, hbp)))
3652                                                 goto bread_err2;
3653                                         offset = xlog_align(log, blk_no,
3654                                                         split_hblks, hbp);
3655                                 }
3656                                 /*
3657                                  * Note: this black magic still works with
3658                                  * large sector sizes (non-512) only because:
3659                                  * - we increased the buffer size originally
3660                                  *   by 1 sector giving us enough extra space
3661                                  *   for the second read;
3662                                  * - the log start is guaranteed to be sector
3663                                  *   aligned;
3664                                  * - we read the log end (LR header start)
3665                                  *   _first_, then the log start (LR header end)
3666                                  *   - order is important.
3667                                  */
3668                                 bufaddr = XFS_BUF_PTR(hbp);
3669                                 XFS_BUF_SET_PTR(hbp,
3670                                                 bufaddr + BBTOB(split_hblks),
3671                                                 BBTOB(hblks - split_hblks));
3672                                 wrapped_hblks = hblks - split_hblks;
3673                                 error = xlog_bread(log, 0, wrapped_hblks, hbp);
3674                                 if (error)
3675                                         goto bread_err2;
3676                                 XFS_BUF_SET_PTR(hbp, bufaddr, BBTOB(hblks));
3677                                 if (!offset)
3678                                         offset = xlog_align(log, 0,
3679                                                         wrapped_hblks, hbp);
3680                         }
3681                         rhead = (xlog_rec_header_t *)offset;
3682                         error = xlog_valid_rec_header(log, rhead,
3683                                                 split_hblks ? blk_no : 0);
3684                         if (error)
3685                                 goto bread_err2;
3686
3687                         bblks = (int)BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT));
3688                         blk_no += hblks;
3689
3690                         /* Read in data for log record */
3691                         if (blk_no + bblks <= log->l_logBBsize) {
3692                                 error = xlog_bread(log, blk_no, bblks, dbp);
3693                                 if (error)
3694                                         goto bread_err2;
3695                                 offset = xlog_align(log, blk_no, bblks, dbp);
3696                         } else {
3697                                 /* This log record is split across the
3698                                  * physical end of log */
3699                                 offset = NULL;
3700                                 split_bblks = 0;
3701                                 if (blk_no != log->l_logBBsize) {
3702                                         /* some data is before the physical
3703                                          * end of log */
3704                                         ASSERT(!wrapped_hblks);
3705                                         ASSERT(blk_no <= INT_MAX);
3706                                         split_bblks =
3707                                                 log->l_logBBsize - (int)blk_no;
3708                                         ASSERT(split_bblks > 0);
3709                                         if ((error = xlog_bread(log, blk_no,
3710                                                         split_bblks, dbp)))
3711                                                 goto bread_err2;
3712                                         offset = xlog_align(log, blk_no,
3713                                                         split_bblks, dbp);
3714                                 }
3715                                 /*
3716                                  * Note: this black magic still works with
3717                                  * large sector sizes (non-512) only because:
3718                                  * - we increased the buffer size originally
3719                                  *   by 1 sector giving us enough extra space
3720                                  *   for the second read;
3721                                  * - the log start is guaranteed to be sector
3722                                  *   aligned;
3723                                  * - we read the log end (LR header start)
3724                                  *   _first_, then the log start (LR header end)
3725                                  *   - order is important.
3726                                  */
3727                                 bufaddr = XFS_BUF_PTR(dbp);
3728                                 XFS_BUF_SET_PTR(dbp,
3729                                                 bufaddr + BBTOB(split_bblks),
3730                                                 BBTOB(bblks - split_bblks));
3731                                 if ((error = xlog_bread(log, wrapped_hblks,
3732                                                 bblks - split_bblks, dbp)))
3733                                         goto bread_err2;
3734                                 XFS_BUF_SET_PTR(dbp, bufaddr, h_size);
3735                                 if (!offset)
3736                                         offset = xlog_align(log, wrapped_hblks,
3737                                                 bblks - split_bblks, dbp);
3738                         }
3739                         xlog_unpack_data(rhead, offset, log);
3740                         if ((error = xlog_recover_process_data(log, rhash,
3741                                                         rhead, offset, pass)))
3742                                 goto bread_err2;
3743                         blk_no += bblks;
3744                 }
3745
3746                 ASSERT(blk_no >= log->l_logBBsize);
3747                 blk_no -= log->l_logBBsize;
3748
3749                 /* read first part of physical log */
3750                 while (blk_no < head_blk) {
3751                         if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3752                                 goto bread_err2;
3753                         offset = xlog_align(log, blk_no, hblks, hbp);
3754                         rhead = (xlog_rec_header_t *)offset;
3755                         error = xlog_valid_rec_header(log, rhead, blk_no);
3756                         if (error)
3757                                 goto bread_err2;
3758                         bblks = (int)BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT));
3759                         if ((error = xlog_bread(log, blk_no+hblks, bblks, dbp)))
3760                                 goto bread_err2;
3761                         offset = xlog_align(log, blk_no+hblks, bblks, dbp);
3762                         xlog_unpack_data(rhead, offset, log);
3763                         if ((error = xlog_recover_process_data(log, rhash,
3764                                                         rhead, offset, pass)))
3765                                 goto bread_err2;
3766                         blk_no += bblks + hblks;
3767                 }
3768         }
3769
3770  bread_err2:
3771         xlog_put_bp(dbp);
3772  bread_err1:
3773         xlog_put_bp(hbp);
3774         return error;
3775 }
3776
3777 /*
3778  * Do the recovery of the log.  We actually do this in two phases.
3779  * The two passes are necessary in order to implement the function
3780  * of cancelling a record written into the log.  The first pass
3781  * determines those things which have been cancelled, and the
3782  * second pass replays log items normally except for those which
3783  * have been cancelled.  The handling of the replay and cancellations
3784  * takes place in the log item type specific routines.
3785  *
3786  * The table of items which have cancel records in the log is allocated
3787  * and freed at this level, since only here do we know when all of
3788  * the log recovery has been completed.
3789  */
3790 STATIC int
3791 xlog_do_log_recovery(
3792         xlog_t          *log,
3793         xfs_daddr_t     head_blk,
3794         xfs_daddr_t     tail_blk)
3795 {
3796         int             error;
3797
3798         ASSERT(head_blk != tail_blk);
3799
3800         /*
3801          * First do a pass to find all of the cancelled buf log items.
3802          * Store them in the buf_cancel_table for use in the second pass.
3803          */
3804         log->l_buf_cancel_table =
3805                 (xfs_buf_cancel_t **)kmem_zalloc(XLOG_BC_TABLE_SIZE *
3806                                                  sizeof(xfs_buf_cancel_t*),
3807                                                  KM_SLEEP);
3808         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3809                                       XLOG_RECOVER_PASS1);
3810         if (error != 0) {
3811                 kmem_free(log->l_buf_cancel_table,
3812                           XLOG_BC_TABLE_SIZE * sizeof(xfs_buf_cancel_t*));
3813                 log->l_buf_cancel_table = NULL;
3814                 return error;
3815         }
3816         /*
3817          * Then do a second pass to actually recover the items in the log.
3818          * When it is complete free the table of buf cancel items.
3819          */
3820         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3821                                       XLOG_RECOVER_PASS2);
3822 #ifdef DEBUG
3823         if (!error) {
3824                 int     i;
3825
3826                 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3827                         ASSERT(log->l_buf_cancel_table[i] == NULL);
3828         }
3829 #endif  /* DEBUG */
3830
3831         kmem_free(log->l_buf_cancel_table,
3832                   XLOG_BC_TABLE_SIZE * sizeof(xfs_buf_cancel_t*));
3833         log->l_buf_cancel_table = NULL;
3834
3835         return error;
3836 }
3837
3838 /*
3839  * Do the actual recovery
3840  */
3841 STATIC int
3842 xlog_do_recover(
3843         xlog_t          *log,
3844         xfs_daddr_t     head_blk,
3845         xfs_daddr_t     tail_blk)
3846 {
3847         int             error;
3848         xfs_buf_t       *bp;
3849         xfs_sb_t        *sbp;
3850
3851         /*
3852          * First replay the images in the log.
3853          */
3854         error = xlog_do_log_recovery(log, head_blk, tail_blk);
3855         if (error) {
3856                 return error;
3857         }
3858
3859         XFS_bflush(log->l_mp->m_ddev_targp);
3860
3861         /*
3862          * If IO errors happened during recovery, bail out.
3863          */
3864         if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3865                 return (EIO);
3866         }
3867
3868         /*
3869          * We now update the tail_lsn since much of the recovery has completed
3870          * and there may be space available to use.  If there were no extent
3871          * or iunlinks, we can free up the entire log and set the tail_lsn to
3872          * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3873          * lsn of the last known good LR on disk.  If there are extent frees
3874          * or iunlinks they will have some entries in the AIL; so we look at
3875          * the AIL to determine how to set the tail_lsn.
3876          */
3877         xlog_assign_tail_lsn(log->l_mp);
3878
3879         /*
3880          * Now that we've finished replaying all buffer and inode
3881          * updates, re-read in the superblock.
3882          */
3883         bp = xfs_getsb(log->l_mp, 0);
3884         XFS_BUF_UNDONE(bp);
3885         XFS_BUF_READ(bp);
3886         xfsbdstrat(log->l_mp, bp);
3887         if ((error = xfs_iowait(bp))) {
3888                 xfs_ioerror_alert("xlog_do_recover",
3889                                   log->l_mp, bp, XFS_BUF_ADDR(bp));
3890                 ASSERT(0);
3891                 xfs_buf_relse(bp);
3892                 return error;
3893         }
3894
3895         /* Convert superblock from on-disk format */
3896         sbp = &log->l_mp->m_sb;
3897         xfs_xlatesb(XFS_BUF_TO_SBP(bp), sbp, 1, XFS_SB_ALL_BITS);
3898         ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3899         ASSERT(XFS_SB_GOOD_VERSION(sbp));
3900         xfs_buf_relse(bp);
3901
3902         /* We've re-read the superblock so re-initialize per-cpu counters */
3903         xfs_icsb_reinit_counters(log->l_mp);
3904
3905         xlog_recover_check_summary(log);
3906
3907         /* Normal transactions can now occur */
3908         log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3909         return 0;
3910 }
3911
3912 /*
3913  * Perform recovery and re-initialize some log variables in xlog_find_tail.
3914  *
3915  * Return error or zero.
3916  */
3917 int
3918 xlog_recover(
3919         xlog_t          *log)
3920 {
3921         xfs_daddr_t     head_blk, tail_blk;
3922         int             error;
3923
3924         /* find the tail of the log */
3925         if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3926                 return error;
3927
3928         if (tail_blk != head_blk) {
3929                 /* There used to be a comment here:
3930                  *
3931                  * disallow recovery on read-only mounts.  note -- mount
3932                  * checks for ENOSPC and turns it into an intelligent
3933                  * error message.
3934                  * ...but this is no longer true.  Now, unless you specify
3935                  * NORECOVERY (in which case this function would never be
3936                  * called), we just go ahead and recover.  We do this all
3937                  * under the vfs layer, so we can get away with it unless
3938                  * the device itself is read-only, in which case we fail.
3939                  */
3940                 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3941                         return error;
3942                 }
3943
3944                 cmn_err(CE_NOTE,
3945                         "Starting XFS recovery on filesystem: %s (logdev: %s)",
3946                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3947                         log->l_mp->m_logname : "internal");
3948
3949                 error = xlog_do_recover(log, head_blk, tail_blk);
3950                 log->l_flags |= XLOG_RECOVERY_NEEDED;
3951         }
3952         return error;
3953 }
3954
3955 /*
3956  * In the first part of recovery we replay inodes and buffers and build
3957  * up the list of extent free items which need to be processed.  Here
3958  * we process the extent free items and clean up the on disk unlinked
3959  * inode lists.  This is separated from the first part of recovery so
3960  * that the root and real-time bitmap inodes can be read in from disk in
3961  * between the two stages.  This is necessary so that we can free space
3962  * in the real-time portion of the file system.
3963  */
3964 int
3965 xlog_recover_finish(
3966         xlog_t          *log,
3967         int             mfsi_flags)
3968 {
3969         /*
3970          * Now we're ready to do the transactions needed for the
3971          * rest of recovery.  Start with completing all the extent
3972          * free intent records and then process the unlinked inode
3973          * lists.  At this point, we essentially run in normal mode
3974          * except that we're still performing recovery actions
3975          * rather than accepting new requests.
3976          */
3977         if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3978                 xlog_recover_process_efis(log);
3979                 /*
3980                  * Sync the log to get all the EFIs out of the AIL.
3981                  * This isn't absolutely necessary, but it helps in
3982                  * case the unlink transactions would have problems
3983                  * pushing the EFIs out of the way.
3984                  */
3985                 xfs_log_force(log->l_mp, (xfs_lsn_t)0,
3986                               (XFS_LOG_FORCE | XFS_LOG_SYNC));
3987
3988                 if ( (mfsi_flags & XFS_MFSI_NOUNLINK) == 0 ) {
3989                         xlog_recover_process_iunlinks(log);
3990                 }
3991
3992                 xlog_recover_check_summary(log);
3993
3994                 cmn_err(CE_NOTE,
3995                         "Ending XFS recovery on filesystem: %s (logdev: %s)",
3996                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3997                         log->l_mp->m_logname : "internal");
3998                 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3999         } else {
4000                 cmn_err(CE_DEBUG,
4001                         "!Ending clean XFS mount for filesystem: %s\n",
4002                         log->l_mp->m_fsname);
4003         }
4004         return 0;
4005 }
4006
4007
4008 #if defined(DEBUG)
4009 /*
4010  * Read all of the agf and agi counters and check that they
4011  * are consistent with the superblock counters.
4012  */
4013 void
4014 xlog_recover_check_summary(
4015         xlog_t          *log)
4016 {
4017         xfs_mount_t     *mp;
4018         xfs_agf_t       *agfp;
4019         xfs_agi_t       *agip;
4020         xfs_buf_t       *agfbp;
4021         xfs_buf_t       *agibp;
4022         xfs_daddr_t     agfdaddr;
4023         xfs_daddr_t     agidaddr;
4024         xfs_buf_t       *sbbp;
4025 #ifdef XFS_LOUD_RECOVERY
4026         xfs_sb_t        *sbp;
4027 #endif
4028         xfs_agnumber_t  agno;
4029         __uint64_t      freeblks;
4030         __uint64_t      itotal;
4031         __uint64_t      ifree;
4032
4033         mp = log->l_mp;
4034
4035         freeblks = 0LL;
4036         itotal = 0LL;
4037         ifree = 0LL;
4038         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4039                 agfdaddr = XFS_AG_DADDR(mp, agno, XFS_AGF_DADDR(mp));
4040                 agfbp = xfs_buf_read(mp->m_ddev_targp, agfdaddr,
4041                                 XFS_FSS_TO_BB(mp, 1), 0);
4042                 if (XFS_BUF_ISERROR(agfbp)) {
4043                         xfs_ioerror_alert("xlog_recover_check_summary(agf)",
4044                                                 mp, agfbp, agfdaddr);
4045                 }
4046                 agfp = XFS_BUF_TO_AGF(agfbp);
4047                 ASSERT(XFS_AGF_MAGIC == be32_to_cpu(agfp->agf_magicnum));
4048                 ASSERT(XFS_AGF_GOOD_VERSION(be32_to_cpu(agfp->agf_versionnum)));
4049                 ASSERT(be32_to_cpu(agfp->agf_seqno) == agno);
4050
4051                 freeblks += be32_to_cpu(agfp->agf_freeblks) +
4052                             be32_to_cpu(agfp->agf_flcount);
4053                 xfs_buf_relse(agfbp);
4054
4055                 agidaddr = XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp));
4056                 agibp = xfs_buf_read(mp->m_ddev_targp, agidaddr,
4057                                 XFS_FSS_TO_BB(mp, 1), 0);
4058                 if (XFS_BUF_ISERROR(agibp)) {
4059                         xfs_ioerror_alert("xlog_recover_check_summary(agi)",
4060                                           mp, agibp, agidaddr);
4061                 }
4062                 agip = XFS_BUF_TO_AGI(agibp);
4063                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agip->agi_magicnum));
4064                 ASSERT(XFS_AGI_GOOD_VERSION(be32_to_cpu(agip->agi_versionnum)));
4065                 ASSERT(be32_to_cpu(agip->agi_seqno) == agno);
4066
4067                 itotal += be32_to_cpu(agip->agi_count);
4068                 ifree += be32_to_cpu(agip->agi_freecount);
4069                 xfs_buf_relse(agibp);
4070         }
4071
4072         sbbp = xfs_getsb(mp, 0);
4073 #ifdef XFS_LOUD_RECOVERY
4074         sbp = &mp->m_sb;
4075         xfs_xlatesb(XFS_BUF_TO_SBP(sbbp), sbp, 1, XFS_SB_ALL_BITS);
4076         cmn_err(CE_NOTE,
4077                 "xlog_recover_check_summary: sb_icount %Lu itotal %Lu",
4078                 sbp->sb_icount, itotal);
4079         cmn_err(CE_NOTE,
4080                 "xlog_recover_check_summary: sb_ifree %Lu itotal %Lu",
4081                 sbp->sb_ifree, ifree);
4082         cmn_err(CE_NOTE,
4083                 "xlog_recover_check_summary: sb_fdblocks %Lu freeblks %Lu",
4084                 sbp->sb_fdblocks, freeblks);
4085 #if 0
4086         /*
4087          * This is turned off until I account for the allocation
4088          * btree blocks which live in free space.
4089          */
4090         ASSERT(sbp->sb_icount == itotal);
4091         ASSERT(sbp->sb_ifree == ifree);
4092         ASSERT(sbp->sb_fdblocks == freeblks);
4093 #endif
4094 #endif
4095         xfs_buf_relse(sbbp);
4096 }
4097 #endif /* DEBUG */