xfs: cleanup xlog_bread
[linux-2.6.git] / fs / xfs / xfs_log_recover.c
1 /*
2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_dir2_sf.h"
35 #include "xfs_attr_sf.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
38 #include "xfs_inode_item.h"
39 #include "xfs_alloc.h"
40 #include "xfs_ialloc.h"
41 #include "xfs_log_priv.h"
42 #include "xfs_buf_item.h"
43 #include "xfs_log_recover.h"
44 #include "xfs_extfree_item.h"
45 #include "xfs_trans_priv.h"
46 #include "xfs_quota.h"
47 #include "xfs_rw.h"
48 #include "xfs_utils.h"
49
50 STATIC int      xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
51 STATIC int      xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
52 STATIC void     xlog_recover_insert_item_backq(xlog_recover_item_t **q,
53                                                xlog_recover_item_t *item);
54 #if defined(DEBUG)
55 STATIC void     xlog_recover_check_summary(xlog_t *);
56 #else
57 #define xlog_recover_check_summary(log)
58 #endif
59
60
61 /*
62  * Sector aligned buffer routines for buffer create/read/write/access
63  */
64
65 #define XLOG_SECTOR_ROUNDUP_BBCOUNT(log, bbs)   \
66         ( ((log)->l_sectbb_mask && (bbs & (log)->l_sectbb_mask)) ? \
67         ((bbs + (log)->l_sectbb_mask + 1) & ~(log)->l_sectbb_mask) : (bbs) )
68 #define XLOG_SECTOR_ROUNDDOWN_BLKNO(log, bno)   ((bno) & ~(log)->l_sectbb_mask)
69
70 xfs_buf_t *
71 xlog_get_bp(
72         xlog_t          *log,
73         int             nbblks)
74 {
75         if (nbblks <= 0 || nbblks > log->l_logBBsize) {
76                 xlog_warn("XFS: Invalid block length (0x%x) given for buffer", nbblks);
77                 XFS_ERROR_REPORT("xlog_get_bp(1)",
78                                  XFS_ERRLEVEL_HIGH, log->l_mp);
79                 return NULL;
80         }
81
82         if (log->l_sectbb_log) {
83                 if (nbblks > 1)
84                         nbblks += XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
85                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
86         }
87         return xfs_buf_get_noaddr(BBTOB(nbblks), log->l_mp->m_logdev_targp);
88 }
89
90 void
91 xlog_put_bp(
92         xfs_buf_t       *bp)
93 {
94         xfs_buf_free(bp);
95 }
96
97 STATIC xfs_caddr_t
98 xlog_align(
99         xlog_t          *log,
100         xfs_daddr_t     blk_no,
101         int             nbblks,
102         xfs_buf_t       *bp)
103 {
104         xfs_caddr_t     ptr;
105
106         if (!log->l_sectbb_log)
107                 return XFS_BUF_PTR(bp);
108
109         ptr = XFS_BUF_PTR(bp) + BBTOB((int)blk_no & log->l_sectbb_mask);
110         ASSERT(XFS_BUF_SIZE(bp) >=
111                 BBTOB(nbblks + (blk_no & log->l_sectbb_mask)));
112         return ptr;
113 }
114
115
116 /*
117  * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
118  */
119 STATIC int
120 xlog_bread_noalign(
121         xlog_t          *log,
122         xfs_daddr_t     blk_no,
123         int             nbblks,
124         xfs_buf_t       *bp)
125 {
126         int             error;
127
128         if (nbblks <= 0 || nbblks > log->l_logBBsize) {
129                 xlog_warn("XFS: Invalid block length (0x%x) given for buffer", nbblks);
130                 XFS_ERROR_REPORT("xlog_bread(1)",
131                                  XFS_ERRLEVEL_HIGH, log->l_mp);
132                 return EFSCORRUPTED;
133         }
134
135         if (log->l_sectbb_log) {
136                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
137                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
138         }
139
140         ASSERT(nbblks > 0);
141         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
142         ASSERT(bp);
143
144         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
145         XFS_BUF_READ(bp);
146         XFS_BUF_BUSY(bp);
147         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
148         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
149
150         xfsbdstrat(log->l_mp, bp);
151         error = xfs_iowait(bp);
152         if (error)
153                 xfs_ioerror_alert("xlog_bread", log->l_mp,
154                                   bp, XFS_BUF_ADDR(bp));
155         return error;
156 }
157
158 STATIC int
159 xlog_bread(
160         xlog_t          *log,
161         xfs_daddr_t     blk_no,
162         int             nbblks,
163         xfs_buf_t       *bp,
164         xfs_caddr_t     *offset)
165 {
166         int             error;
167
168         error = xlog_bread_noalign(log, blk_no, nbblks, bp);
169         if (error)
170                 return error;
171
172         *offset = xlog_align(log, blk_no, nbblks, bp);
173         return 0;
174 }
175
176 /*
177  * Write out the buffer at the given block for the given number of blocks.
178  * The buffer is kept locked across the write and is returned locked.
179  * This can only be used for synchronous log writes.
180  */
181 STATIC int
182 xlog_bwrite(
183         xlog_t          *log,
184         xfs_daddr_t     blk_no,
185         int             nbblks,
186         xfs_buf_t       *bp)
187 {
188         int             error;
189
190         if (nbblks <= 0 || nbblks > log->l_logBBsize) {
191                 xlog_warn("XFS: Invalid block length (0x%x) given for buffer", nbblks);
192                 XFS_ERROR_REPORT("xlog_bwrite(1)",
193                                  XFS_ERRLEVEL_HIGH, log->l_mp);
194                 return EFSCORRUPTED;
195         }
196
197         if (log->l_sectbb_log) {
198                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
199                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
200         }
201
202         ASSERT(nbblks > 0);
203         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
204
205         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
206         XFS_BUF_ZEROFLAGS(bp);
207         XFS_BUF_BUSY(bp);
208         XFS_BUF_HOLD(bp);
209         XFS_BUF_PSEMA(bp, PRIBIO);
210         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
211         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
212
213         if ((error = xfs_bwrite(log->l_mp, bp)))
214                 xfs_ioerror_alert("xlog_bwrite", log->l_mp,
215                                   bp, XFS_BUF_ADDR(bp));
216         return error;
217 }
218
219 #ifdef DEBUG
220 /*
221  * dump debug superblock and log record information
222  */
223 STATIC void
224 xlog_header_check_dump(
225         xfs_mount_t             *mp,
226         xlog_rec_header_t       *head)
227 {
228         int                     b;
229
230         cmn_err(CE_DEBUG, "%s:  SB : uuid = ", __func__);
231         for (b = 0; b < 16; b++)
232                 cmn_err(CE_DEBUG, "%02x", ((__uint8_t *)&mp->m_sb.sb_uuid)[b]);
233         cmn_err(CE_DEBUG, ", fmt = %d\n", XLOG_FMT);
234         cmn_err(CE_DEBUG, "    log : uuid = ");
235         for (b = 0; b < 16; b++)
236                 cmn_err(CE_DEBUG, "%02x", ((__uint8_t *)&head->h_fs_uuid)[b]);
237         cmn_err(CE_DEBUG, ", fmt = %d\n", be32_to_cpu(head->h_fmt));
238 }
239 #else
240 #define xlog_header_check_dump(mp, head)
241 #endif
242
243 /*
244  * check log record header for recovery
245  */
246 STATIC int
247 xlog_header_check_recover(
248         xfs_mount_t             *mp,
249         xlog_rec_header_t       *head)
250 {
251         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
252
253         /*
254          * IRIX doesn't write the h_fmt field and leaves it zeroed
255          * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
256          * a dirty log created in IRIX.
257          */
258         if (unlikely(be32_to_cpu(head->h_fmt) != XLOG_FMT)) {
259                 xlog_warn(
260         "XFS: dirty log written in incompatible format - can't recover");
261                 xlog_header_check_dump(mp, head);
262                 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
263                                  XFS_ERRLEVEL_HIGH, mp);
264                 return XFS_ERROR(EFSCORRUPTED);
265         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
266                 xlog_warn(
267         "XFS: dirty log entry has mismatched uuid - can't recover");
268                 xlog_header_check_dump(mp, head);
269                 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
270                                  XFS_ERRLEVEL_HIGH, mp);
271                 return XFS_ERROR(EFSCORRUPTED);
272         }
273         return 0;
274 }
275
276 /*
277  * read the head block of the log and check the header
278  */
279 STATIC int
280 xlog_header_check_mount(
281         xfs_mount_t             *mp,
282         xlog_rec_header_t       *head)
283 {
284         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
285
286         if (uuid_is_nil(&head->h_fs_uuid)) {
287                 /*
288                  * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
289                  * h_fs_uuid is nil, we assume this log was last mounted
290                  * by IRIX and continue.
291                  */
292                 xlog_warn("XFS: nil uuid in log - IRIX style log");
293         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
294                 xlog_warn("XFS: log has mismatched uuid - can't recover");
295                 xlog_header_check_dump(mp, head);
296                 XFS_ERROR_REPORT("xlog_header_check_mount",
297                                  XFS_ERRLEVEL_HIGH, mp);
298                 return XFS_ERROR(EFSCORRUPTED);
299         }
300         return 0;
301 }
302
303 STATIC void
304 xlog_recover_iodone(
305         struct xfs_buf  *bp)
306 {
307         if (XFS_BUF_GETERROR(bp)) {
308                 /*
309                  * We're not going to bother about retrying
310                  * this during recovery. One strike!
311                  */
312                 xfs_ioerror_alert("xlog_recover_iodone",
313                                   bp->b_mount, bp, XFS_BUF_ADDR(bp));
314                 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
315         }
316         bp->b_mount = NULL;
317         XFS_BUF_CLR_IODONE_FUNC(bp);
318         xfs_biodone(bp);
319 }
320
321 /*
322  * This routine finds (to an approximation) the first block in the physical
323  * log which contains the given cycle.  It uses a binary search algorithm.
324  * Note that the algorithm can not be perfect because the disk will not
325  * necessarily be perfect.
326  */
327 STATIC int
328 xlog_find_cycle_start(
329         xlog_t          *log,
330         xfs_buf_t       *bp,
331         xfs_daddr_t     first_blk,
332         xfs_daddr_t     *last_blk,
333         uint            cycle)
334 {
335         xfs_caddr_t     offset;
336         xfs_daddr_t     mid_blk;
337         uint            mid_cycle;
338         int             error;
339
340         mid_blk = BLK_AVG(first_blk, *last_blk);
341         while (mid_blk != first_blk && mid_blk != *last_blk) {
342                 error = xlog_bread(log, mid_blk, 1, bp, &offset);
343                 if (error)
344                         return error;
345                 mid_cycle = xlog_get_cycle(offset);
346                 if (mid_cycle == cycle) {
347                         *last_blk = mid_blk;
348                         /* last_half_cycle == mid_cycle */
349                 } else {
350                         first_blk = mid_blk;
351                         /* first_half_cycle == mid_cycle */
352                 }
353                 mid_blk = BLK_AVG(first_blk, *last_blk);
354         }
355         ASSERT((mid_blk == first_blk && mid_blk+1 == *last_blk) ||
356                (mid_blk == *last_blk && mid_blk-1 == first_blk));
357
358         return 0;
359 }
360
361 /*
362  * Check that the range of blocks does not contain the cycle number
363  * given.  The scan needs to occur from front to back and the ptr into the
364  * region must be updated since a later routine will need to perform another
365  * test.  If the region is completely good, we end up returning the same
366  * last block number.
367  *
368  * Set blkno to -1 if we encounter no errors.  This is an invalid block number
369  * since we don't ever expect logs to get this large.
370  */
371 STATIC int
372 xlog_find_verify_cycle(
373         xlog_t          *log,
374         xfs_daddr_t     start_blk,
375         int             nbblks,
376         uint            stop_on_cycle_no,
377         xfs_daddr_t     *new_blk)
378 {
379         xfs_daddr_t     i, j;
380         uint            cycle;
381         xfs_buf_t       *bp;
382         xfs_daddr_t     bufblks;
383         xfs_caddr_t     buf = NULL;
384         int             error = 0;
385
386         bufblks = 1 << ffs(nbblks);
387
388         while (!(bp = xlog_get_bp(log, bufblks))) {
389                 /* can't get enough memory to do everything in one big buffer */
390                 bufblks >>= 1;
391                 if (bufblks <= log->l_sectbb_log)
392                         return ENOMEM;
393         }
394
395         for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
396                 int     bcount;
397
398                 bcount = min(bufblks, (start_blk + nbblks - i));
399
400                 error = xlog_bread(log, i, bcount, bp, &buf);
401                 if (error)
402                         goto out;
403
404                 for (j = 0; j < bcount; j++) {
405                         cycle = xlog_get_cycle(buf);
406                         if (cycle == stop_on_cycle_no) {
407                                 *new_blk = i+j;
408                                 goto out;
409                         }
410
411                         buf += BBSIZE;
412                 }
413         }
414
415         *new_blk = -1;
416
417 out:
418         xlog_put_bp(bp);
419         return error;
420 }
421
422 /*
423  * Potentially backup over partial log record write.
424  *
425  * In the typical case, last_blk is the number of the block directly after
426  * a good log record.  Therefore, we subtract one to get the block number
427  * of the last block in the given buffer.  extra_bblks contains the number
428  * of blocks we would have read on a previous read.  This happens when the
429  * last log record is split over the end of the physical log.
430  *
431  * extra_bblks is the number of blocks potentially verified on a previous
432  * call to this routine.
433  */
434 STATIC int
435 xlog_find_verify_log_record(
436         xlog_t                  *log,
437         xfs_daddr_t             start_blk,
438         xfs_daddr_t             *last_blk,
439         int                     extra_bblks)
440 {
441         xfs_daddr_t             i;
442         xfs_buf_t               *bp;
443         xfs_caddr_t             offset = NULL;
444         xlog_rec_header_t       *head = NULL;
445         int                     error = 0;
446         int                     smallmem = 0;
447         int                     num_blks = *last_blk - start_blk;
448         int                     xhdrs;
449
450         ASSERT(start_blk != 0 || *last_blk != start_blk);
451
452         if (!(bp = xlog_get_bp(log, num_blks))) {
453                 if (!(bp = xlog_get_bp(log, 1)))
454                         return ENOMEM;
455                 smallmem = 1;
456         } else {
457                 error = xlog_bread(log, start_blk, num_blks, bp, &offset);
458                 if (error)
459                         goto out;
460                 offset += ((num_blks - 1) << BBSHIFT);
461         }
462
463         for (i = (*last_blk) - 1; i >= 0; i--) {
464                 if (i < start_blk) {
465                         /* valid log record not found */
466                         xlog_warn(
467                 "XFS: Log inconsistent (didn't find previous header)");
468                         ASSERT(0);
469                         error = XFS_ERROR(EIO);
470                         goto out;
471                 }
472
473                 if (smallmem) {
474                         error = xlog_bread(log, i, 1, bp, &offset);
475                         if (error)
476                                 goto out;
477                 }
478
479                 head = (xlog_rec_header_t *)offset;
480
481                 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(head->h_magicno))
482                         break;
483
484                 if (!smallmem)
485                         offset -= BBSIZE;
486         }
487
488         /*
489          * We hit the beginning of the physical log & still no header.  Return
490          * to caller.  If caller can handle a return of -1, then this routine
491          * will be called again for the end of the physical log.
492          */
493         if (i == -1) {
494                 error = -1;
495                 goto out;
496         }
497
498         /*
499          * We have the final block of the good log (the first block
500          * of the log record _before_ the head. So we check the uuid.
501          */
502         if ((error = xlog_header_check_mount(log->l_mp, head)))
503                 goto out;
504
505         /*
506          * We may have found a log record header before we expected one.
507          * last_blk will be the 1st block # with a given cycle #.  We may end
508          * up reading an entire log record.  In this case, we don't want to
509          * reset last_blk.  Only when last_blk points in the middle of a log
510          * record do we update last_blk.
511          */
512         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
513                 uint    h_size = be32_to_cpu(head->h_size);
514
515                 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
516                 if (h_size % XLOG_HEADER_CYCLE_SIZE)
517                         xhdrs++;
518         } else {
519                 xhdrs = 1;
520         }
521
522         if (*last_blk - i + extra_bblks !=
523             BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
524                 *last_blk = i;
525
526 out:
527         xlog_put_bp(bp);
528         return error;
529 }
530
531 /*
532  * Head is defined to be the point of the log where the next log write
533  * write could go.  This means that incomplete LR writes at the end are
534  * eliminated when calculating the head.  We aren't guaranteed that previous
535  * LR have complete transactions.  We only know that a cycle number of
536  * current cycle number -1 won't be present in the log if we start writing
537  * from our current block number.
538  *
539  * last_blk contains the block number of the first block with a given
540  * cycle number.
541  *
542  * Return: zero if normal, non-zero if error.
543  */
544 STATIC int
545 xlog_find_head(
546         xlog_t          *log,
547         xfs_daddr_t     *return_head_blk)
548 {
549         xfs_buf_t       *bp;
550         xfs_caddr_t     offset;
551         xfs_daddr_t     new_blk, first_blk, start_blk, last_blk, head_blk;
552         int             num_scan_bblks;
553         uint            first_half_cycle, last_half_cycle;
554         uint            stop_on_cycle;
555         int             error, log_bbnum = log->l_logBBsize;
556
557         /* Is the end of the log device zeroed? */
558         if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
559                 *return_head_blk = first_blk;
560
561                 /* Is the whole lot zeroed? */
562                 if (!first_blk) {
563                         /* Linux XFS shouldn't generate totally zeroed logs -
564                          * mkfs etc write a dummy unmount record to a fresh
565                          * log so we can store the uuid in there
566                          */
567                         xlog_warn("XFS: totally zeroed log");
568                 }
569
570                 return 0;
571         } else if (error) {
572                 xlog_warn("XFS: empty log check failed");
573                 return error;
574         }
575
576         first_blk = 0;                  /* get cycle # of 1st block */
577         bp = xlog_get_bp(log, 1);
578         if (!bp)
579                 return ENOMEM;
580
581         error = xlog_bread(log, 0, 1, bp, &offset);
582         if (error)
583                 goto bp_err;
584
585         first_half_cycle = xlog_get_cycle(offset);
586
587         last_blk = head_blk = log_bbnum - 1;    /* get cycle # of last block */
588         error = xlog_bread(log, last_blk, 1, bp, &offset);
589         if (error)
590                 goto bp_err;
591
592         last_half_cycle = xlog_get_cycle(offset);
593         ASSERT(last_half_cycle != 0);
594
595         /*
596          * If the 1st half cycle number is equal to the last half cycle number,
597          * then the entire log is stamped with the same cycle number.  In this
598          * case, head_blk can't be set to zero (which makes sense).  The below
599          * math doesn't work out properly with head_blk equal to zero.  Instead,
600          * we set it to log_bbnum which is an invalid block number, but this
601          * value makes the math correct.  If head_blk doesn't changed through
602          * all the tests below, *head_blk is set to zero at the very end rather
603          * than log_bbnum.  In a sense, log_bbnum and zero are the same block
604          * in a circular file.
605          */
606         if (first_half_cycle == last_half_cycle) {
607                 /*
608                  * In this case we believe that the entire log should have
609                  * cycle number last_half_cycle.  We need to scan backwards
610                  * from the end verifying that there are no holes still
611                  * containing last_half_cycle - 1.  If we find such a hole,
612                  * then the start of that hole will be the new head.  The
613                  * simple case looks like
614                  *        x | x ... | x - 1 | x
615                  * Another case that fits this picture would be
616                  *        x | x + 1 | x ... | x
617                  * In this case the head really is somewhere at the end of the
618                  * log, as one of the latest writes at the beginning was
619                  * incomplete.
620                  * One more case is
621                  *        x | x + 1 | x ... | x - 1 | x
622                  * This is really the combination of the above two cases, and
623                  * the head has to end up at the start of the x-1 hole at the
624                  * end of the log.
625                  *
626                  * In the 256k log case, we will read from the beginning to the
627                  * end of the log and search for cycle numbers equal to x-1.
628                  * We don't worry about the x+1 blocks that we encounter,
629                  * because we know that they cannot be the head since the log
630                  * started with x.
631                  */
632                 head_blk = log_bbnum;
633                 stop_on_cycle = last_half_cycle - 1;
634         } else {
635                 /*
636                  * In this case we want to find the first block with cycle
637                  * number matching last_half_cycle.  We expect the log to be
638                  * some variation on
639                  *        x + 1 ... | x ...
640                  * The first block with cycle number x (last_half_cycle) will
641                  * be where the new head belongs.  First we do a binary search
642                  * for the first occurrence of last_half_cycle.  The binary
643                  * search may not be totally accurate, so then we scan back
644                  * from there looking for occurrences of last_half_cycle before
645                  * us.  If that backwards scan wraps around the beginning of
646                  * the log, then we look for occurrences of last_half_cycle - 1
647                  * at the end of the log.  The cases we're looking for look
648                  * like
649                  *        x + 1 ... | x | x + 1 | x ...
650                  *                               ^ binary search stopped here
651                  * or
652                  *        x + 1 ... | x ... | x - 1 | x
653                  *        <---------> less than scan distance
654                  */
655                 stop_on_cycle = last_half_cycle;
656                 if ((error = xlog_find_cycle_start(log, bp, first_blk,
657                                                 &head_blk, last_half_cycle)))
658                         goto bp_err;
659         }
660
661         /*
662          * Now validate the answer.  Scan back some number of maximum possible
663          * blocks and make sure each one has the expected cycle number.  The
664          * maximum is determined by the total possible amount of buffering
665          * in the in-core log.  The following number can be made tighter if
666          * we actually look at the block size of the filesystem.
667          */
668         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
669         if (head_blk >= num_scan_bblks) {
670                 /*
671                  * We are guaranteed that the entire check can be performed
672                  * in one buffer.
673                  */
674                 start_blk = head_blk - num_scan_bblks;
675                 if ((error = xlog_find_verify_cycle(log,
676                                                 start_blk, num_scan_bblks,
677                                                 stop_on_cycle, &new_blk)))
678                         goto bp_err;
679                 if (new_blk != -1)
680                         head_blk = new_blk;
681         } else {                /* need to read 2 parts of log */
682                 /*
683                  * We are going to scan backwards in the log in two parts.
684                  * First we scan the physical end of the log.  In this part
685                  * of the log, we are looking for blocks with cycle number
686                  * last_half_cycle - 1.
687                  * If we find one, then we know that the log starts there, as
688                  * we've found a hole that didn't get written in going around
689                  * the end of the physical log.  The simple case for this is
690                  *        x + 1 ... | x ... | x - 1 | x
691                  *        <---------> less than scan distance
692                  * If all of the blocks at the end of the log have cycle number
693                  * last_half_cycle, then we check the blocks at the start of
694                  * the log looking for occurrences of last_half_cycle.  If we
695                  * find one, then our current estimate for the location of the
696                  * first occurrence of last_half_cycle is wrong and we move
697                  * back to the hole we've found.  This case looks like
698                  *        x + 1 ... | x | x + 1 | x ...
699                  *                               ^ binary search stopped here
700                  * Another case we need to handle that only occurs in 256k
701                  * logs is
702                  *        x + 1 ... | x ... | x+1 | x ...
703                  *                   ^ binary search stops here
704                  * In a 256k log, the scan at the end of the log will see the
705                  * x + 1 blocks.  We need to skip past those since that is
706                  * certainly not the head of the log.  By searching for
707                  * last_half_cycle-1 we accomplish that.
708                  */
709                 start_blk = log_bbnum - num_scan_bblks + head_blk;
710                 ASSERT(head_blk <= INT_MAX &&
711                         (xfs_daddr_t) num_scan_bblks - head_blk >= 0);
712                 if ((error = xlog_find_verify_cycle(log, start_blk,
713                                         num_scan_bblks - (int)head_blk,
714                                         (stop_on_cycle - 1), &new_blk)))
715                         goto bp_err;
716                 if (new_blk != -1) {
717                         head_blk = new_blk;
718                         goto bad_blk;
719                 }
720
721                 /*
722                  * Scan beginning of log now.  The last part of the physical
723                  * log is good.  This scan needs to verify that it doesn't find
724                  * the last_half_cycle.
725                  */
726                 start_blk = 0;
727                 ASSERT(head_blk <= INT_MAX);
728                 if ((error = xlog_find_verify_cycle(log,
729                                         start_blk, (int)head_blk,
730                                         stop_on_cycle, &new_blk)))
731                         goto bp_err;
732                 if (new_blk != -1)
733                         head_blk = new_blk;
734         }
735
736  bad_blk:
737         /*
738          * Now we need to make sure head_blk is not pointing to a block in
739          * the middle of a log record.
740          */
741         num_scan_bblks = XLOG_REC_SHIFT(log);
742         if (head_blk >= num_scan_bblks) {
743                 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
744
745                 /* start ptr at last block ptr before head_blk */
746                 if ((error = xlog_find_verify_log_record(log, start_blk,
747                                                         &head_blk, 0)) == -1) {
748                         error = XFS_ERROR(EIO);
749                         goto bp_err;
750                 } else if (error)
751                         goto bp_err;
752         } else {
753                 start_blk = 0;
754                 ASSERT(head_blk <= INT_MAX);
755                 if ((error = xlog_find_verify_log_record(log, start_blk,
756                                                         &head_blk, 0)) == -1) {
757                         /* We hit the beginning of the log during our search */
758                         start_blk = log_bbnum - num_scan_bblks + head_blk;
759                         new_blk = log_bbnum;
760                         ASSERT(start_blk <= INT_MAX &&
761                                 (xfs_daddr_t) log_bbnum-start_blk >= 0);
762                         ASSERT(head_blk <= INT_MAX);
763                         if ((error = xlog_find_verify_log_record(log,
764                                                         start_blk, &new_blk,
765                                                         (int)head_blk)) == -1) {
766                                 error = XFS_ERROR(EIO);
767                                 goto bp_err;
768                         } else if (error)
769                                 goto bp_err;
770                         if (new_blk != log_bbnum)
771                                 head_blk = new_blk;
772                 } else if (error)
773                         goto bp_err;
774         }
775
776         xlog_put_bp(bp);
777         if (head_blk == log_bbnum)
778                 *return_head_blk = 0;
779         else
780                 *return_head_blk = head_blk;
781         /*
782          * When returning here, we have a good block number.  Bad block
783          * means that during a previous crash, we didn't have a clean break
784          * from cycle number N to cycle number N-1.  In this case, we need
785          * to find the first block with cycle number N-1.
786          */
787         return 0;
788
789  bp_err:
790         xlog_put_bp(bp);
791
792         if (error)
793             xlog_warn("XFS: failed to find log head");
794         return error;
795 }
796
797 /*
798  * Find the sync block number or the tail of the log.
799  *
800  * This will be the block number of the last record to have its
801  * associated buffers synced to disk.  Every log record header has
802  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
803  * to get a sync block number.  The only concern is to figure out which
804  * log record header to believe.
805  *
806  * The following algorithm uses the log record header with the largest
807  * lsn.  The entire log record does not need to be valid.  We only care
808  * that the header is valid.
809  *
810  * We could speed up search by using current head_blk buffer, but it is not
811  * available.
812  */
813 int
814 xlog_find_tail(
815         xlog_t                  *log,
816         xfs_daddr_t             *head_blk,
817         xfs_daddr_t             *tail_blk)
818 {
819         xlog_rec_header_t       *rhead;
820         xlog_op_header_t        *op_head;
821         xfs_caddr_t             offset = NULL;
822         xfs_buf_t               *bp;
823         int                     error, i, found;
824         xfs_daddr_t             umount_data_blk;
825         xfs_daddr_t             after_umount_blk;
826         xfs_lsn_t               tail_lsn;
827         int                     hblks;
828
829         found = 0;
830
831         /*
832          * Find previous log record
833          */
834         if ((error = xlog_find_head(log, head_blk)))
835                 return error;
836
837         bp = xlog_get_bp(log, 1);
838         if (!bp)
839                 return ENOMEM;
840         if (*head_blk == 0) {                           /* special case */
841                 error = xlog_bread(log, 0, 1, bp, &offset);
842                 if (error)
843                         goto bread_err;
844
845                 if (xlog_get_cycle(offset) == 0) {
846                         *tail_blk = 0;
847                         /* leave all other log inited values alone */
848                         goto exit;
849                 }
850         }
851
852         /*
853          * Search backwards looking for log record header block
854          */
855         ASSERT(*head_blk < INT_MAX);
856         for (i = (int)(*head_blk) - 1; i >= 0; i--) {
857                 error = xlog_bread(log, i, 1, bp, &offset);
858                 if (error)
859                         goto bread_err;
860
861                 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(*(__be32 *)offset)) {
862                         found = 1;
863                         break;
864                 }
865         }
866         /*
867          * If we haven't found the log record header block, start looking
868          * again from the end of the physical log.  XXXmiken: There should be
869          * a check here to make sure we didn't search more than N blocks in
870          * the previous code.
871          */
872         if (!found) {
873                 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
874                         error = xlog_bread(log, i, 1, bp, &offset);
875                         if (error)
876                                 goto bread_err;
877
878                         if (XLOG_HEADER_MAGIC_NUM ==
879                             be32_to_cpu(*(__be32 *)offset)) {
880                                 found = 2;
881                                 break;
882                         }
883                 }
884         }
885         if (!found) {
886                 xlog_warn("XFS: xlog_find_tail: couldn't find sync record");
887                 ASSERT(0);
888                 return XFS_ERROR(EIO);
889         }
890
891         /* find blk_no of tail of log */
892         rhead = (xlog_rec_header_t *)offset;
893         *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
894
895         /*
896          * Reset log values according to the state of the log when we
897          * crashed.  In the case where head_blk == 0, we bump curr_cycle
898          * one because the next write starts a new cycle rather than
899          * continuing the cycle of the last good log record.  At this
900          * point we have guaranteed that all partial log records have been
901          * accounted for.  Therefore, we know that the last good log record
902          * written was complete and ended exactly on the end boundary
903          * of the physical log.
904          */
905         log->l_prev_block = i;
906         log->l_curr_block = (int)*head_blk;
907         log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
908         if (found == 2)
909                 log->l_curr_cycle++;
910         log->l_tail_lsn = be64_to_cpu(rhead->h_tail_lsn);
911         log->l_last_sync_lsn = be64_to_cpu(rhead->h_lsn);
912         log->l_grant_reserve_cycle = log->l_curr_cycle;
913         log->l_grant_reserve_bytes = BBTOB(log->l_curr_block);
914         log->l_grant_write_cycle = log->l_curr_cycle;
915         log->l_grant_write_bytes = BBTOB(log->l_curr_block);
916
917         /*
918          * Look for unmount record.  If we find it, then we know there
919          * was a clean unmount.  Since 'i' could be the last block in
920          * the physical log, we convert to a log block before comparing
921          * to the head_blk.
922          *
923          * Save the current tail lsn to use to pass to
924          * xlog_clear_stale_blocks() below.  We won't want to clear the
925          * unmount record if there is one, so we pass the lsn of the
926          * unmount record rather than the block after it.
927          */
928         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
929                 int     h_size = be32_to_cpu(rhead->h_size);
930                 int     h_version = be32_to_cpu(rhead->h_version);
931
932                 if ((h_version & XLOG_VERSION_2) &&
933                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
934                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
935                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
936                                 hblks++;
937                 } else {
938                         hblks = 1;
939                 }
940         } else {
941                 hblks = 1;
942         }
943         after_umount_blk = (i + hblks + (int)
944                 BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
945         tail_lsn = log->l_tail_lsn;
946         if (*head_blk == after_umount_blk &&
947             be32_to_cpu(rhead->h_num_logops) == 1) {
948                 umount_data_blk = (i + hblks) % log->l_logBBsize;
949                 error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
950                 if (error)
951                         goto bread_err;
952
953                 op_head = (xlog_op_header_t *)offset;
954                 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
955                         /*
956                          * Set tail and last sync so that newly written
957                          * log records will point recovery to after the
958                          * current unmount record.
959                          */
960                         log->l_tail_lsn =
961                                 xlog_assign_lsn(log->l_curr_cycle,
962                                                 after_umount_blk);
963                         log->l_last_sync_lsn =
964                                 xlog_assign_lsn(log->l_curr_cycle,
965                                                 after_umount_blk);
966                         *tail_blk = after_umount_blk;
967
968                         /*
969                          * Note that the unmount was clean. If the unmount
970                          * was not clean, we need to know this to rebuild the
971                          * superblock counters from the perag headers if we
972                          * have a filesystem using non-persistent counters.
973                          */
974                         log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
975                 }
976         }
977
978         /*
979          * Make sure that there are no blocks in front of the head
980          * with the same cycle number as the head.  This can happen
981          * because we allow multiple outstanding log writes concurrently,
982          * and the later writes might make it out before earlier ones.
983          *
984          * We use the lsn from before modifying it so that we'll never
985          * overwrite the unmount record after a clean unmount.
986          *
987          * Do this only if we are going to recover the filesystem
988          *
989          * NOTE: This used to say "if (!readonly)"
990          * However on Linux, we can & do recover a read-only filesystem.
991          * We only skip recovery if NORECOVERY is specified on mount,
992          * in which case we would not be here.
993          *
994          * But... if the -device- itself is readonly, just skip this.
995          * We can't recover this device anyway, so it won't matter.
996          */
997         if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
998                 error = xlog_clear_stale_blocks(log, tail_lsn);
999         }
1000
1001 bread_err:
1002 exit:
1003         xlog_put_bp(bp);
1004
1005         if (error)
1006                 xlog_warn("XFS: failed to locate log tail");
1007         return error;
1008 }
1009
1010 /*
1011  * Is the log zeroed at all?
1012  *
1013  * The last binary search should be changed to perform an X block read
1014  * once X becomes small enough.  You can then search linearly through
1015  * the X blocks.  This will cut down on the number of reads we need to do.
1016  *
1017  * If the log is partially zeroed, this routine will pass back the blkno
1018  * of the first block with cycle number 0.  It won't have a complete LR
1019  * preceding it.
1020  *
1021  * Return:
1022  *      0  => the log is completely written to
1023  *      -1 => use *blk_no as the first block of the log
1024  *      >0 => error has occurred
1025  */
1026 STATIC int
1027 xlog_find_zeroed(
1028         xlog_t          *log,
1029         xfs_daddr_t     *blk_no)
1030 {
1031         xfs_buf_t       *bp;
1032         xfs_caddr_t     offset;
1033         uint            first_cycle, last_cycle;
1034         xfs_daddr_t     new_blk, last_blk, start_blk;
1035         xfs_daddr_t     num_scan_bblks;
1036         int             error, log_bbnum = log->l_logBBsize;
1037
1038         *blk_no = 0;
1039
1040         /* check totally zeroed log */
1041         bp = xlog_get_bp(log, 1);
1042         if (!bp)
1043                 return ENOMEM;
1044         error = xlog_bread(log, 0, 1, bp, &offset);
1045         if (error)
1046                 goto bp_err;
1047
1048         first_cycle = xlog_get_cycle(offset);
1049         if (first_cycle == 0) {         /* completely zeroed log */
1050                 *blk_no = 0;
1051                 xlog_put_bp(bp);
1052                 return -1;
1053         }
1054
1055         /* check partially zeroed log */
1056         error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1057         if (error)
1058                 goto bp_err;
1059
1060         last_cycle = xlog_get_cycle(offset);
1061         if (last_cycle != 0) {          /* log completely written to */
1062                 xlog_put_bp(bp);
1063                 return 0;
1064         } else if (first_cycle != 1) {
1065                 /*
1066                  * If the cycle of the last block is zero, the cycle of
1067                  * the first block must be 1. If it's not, maybe we're
1068                  * not looking at a log... Bail out.
1069                  */
1070                 xlog_warn("XFS: Log inconsistent or not a log (last==0, first!=1)");
1071                 return XFS_ERROR(EINVAL);
1072         }
1073
1074         /* we have a partially zeroed log */
1075         last_blk = log_bbnum-1;
1076         if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1077                 goto bp_err;
1078
1079         /*
1080          * Validate the answer.  Because there is no way to guarantee that
1081          * the entire log is made up of log records which are the same size,
1082          * we scan over the defined maximum blocks.  At this point, the maximum
1083          * is not chosen to mean anything special.   XXXmiken
1084          */
1085         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1086         ASSERT(num_scan_bblks <= INT_MAX);
1087
1088         if (last_blk < num_scan_bblks)
1089                 num_scan_bblks = last_blk;
1090         start_blk = last_blk - num_scan_bblks;
1091
1092         /*
1093          * We search for any instances of cycle number 0 that occur before
1094          * our current estimate of the head.  What we're trying to detect is
1095          *        1 ... | 0 | 1 | 0...
1096          *                       ^ binary search ends here
1097          */
1098         if ((error = xlog_find_verify_cycle(log, start_blk,
1099                                          (int)num_scan_bblks, 0, &new_blk)))
1100                 goto bp_err;
1101         if (new_blk != -1)
1102                 last_blk = new_blk;
1103
1104         /*
1105          * Potentially backup over partial log record write.  We don't need
1106          * to search the end of the log because we know it is zero.
1107          */
1108         if ((error = xlog_find_verify_log_record(log, start_blk,
1109                                 &last_blk, 0)) == -1) {
1110             error = XFS_ERROR(EIO);
1111             goto bp_err;
1112         } else if (error)
1113             goto bp_err;
1114
1115         *blk_no = last_blk;
1116 bp_err:
1117         xlog_put_bp(bp);
1118         if (error)
1119                 return error;
1120         return -1;
1121 }
1122
1123 /*
1124  * These are simple subroutines used by xlog_clear_stale_blocks() below
1125  * to initialize a buffer full of empty log record headers and write
1126  * them into the log.
1127  */
1128 STATIC void
1129 xlog_add_record(
1130         xlog_t                  *log,
1131         xfs_caddr_t             buf,
1132         int                     cycle,
1133         int                     block,
1134         int                     tail_cycle,
1135         int                     tail_block)
1136 {
1137         xlog_rec_header_t       *recp = (xlog_rec_header_t *)buf;
1138
1139         memset(buf, 0, BBSIZE);
1140         recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1141         recp->h_cycle = cpu_to_be32(cycle);
1142         recp->h_version = cpu_to_be32(
1143                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1144         recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1145         recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1146         recp->h_fmt = cpu_to_be32(XLOG_FMT);
1147         memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1148 }
1149
1150 STATIC int
1151 xlog_write_log_records(
1152         xlog_t          *log,
1153         int             cycle,
1154         int             start_block,
1155         int             blocks,
1156         int             tail_cycle,
1157         int             tail_block)
1158 {
1159         xfs_caddr_t     offset;
1160         xfs_buf_t       *bp;
1161         int             balign, ealign;
1162         int             sectbb = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
1163         int             end_block = start_block + blocks;
1164         int             bufblks;
1165         int             error = 0;
1166         int             i, j = 0;
1167
1168         bufblks = 1 << ffs(blocks);
1169         while (!(bp = xlog_get_bp(log, bufblks))) {
1170                 bufblks >>= 1;
1171                 if (bufblks <= log->l_sectbb_log)
1172                         return ENOMEM;
1173         }
1174
1175         /* We may need to do a read at the start to fill in part of
1176          * the buffer in the starting sector not covered by the first
1177          * write below.
1178          */
1179         balign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, start_block);
1180         if (balign != start_block) {
1181                 error = xlog_bread_noalign(log, start_block, 1, bp);
1182                 if (error)
1183                         goto out_put_bp;
1184
1185                 j = start_block - balign;
1186         }
1187
1188         for (i = start_block; i < end_block; i += bufblks) {
1189                 int             bcount, endcount;
1190
1191                 bcount = min(bufblks, end_block - start_block);
1192                 endcount = bcount - j;
1193
1194                 /* We may need to do a read at the end to fill in part of
1195                  * the buffer in the final sector not covered by the write.
1196                  * If this is the same sector as the above read, skip it.
1197                  */
1198                 ealign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, end_block);
1199                 if (j == 0 && (start_block + endcount > ealign)) {
1200                         offset = XFS_BUF_PTR(bp);
1201                         balign = BBTOB(ealign - start_block);
1202                         error = XFS_BUF_SET_PTR(bp, offset + balign,
1203                                                 BBTOB(sectbb));
1204                         if (error)
1205                                 break;
1206
1207                         error = xlog_bread_noalign(log, ealign, sectbb, bp);
1208                         if (error)
1209                                 break;
1210
1211                         error = XFS_BUF_SET_PTR(bp, offset, bufblks);
1212                         if (error)
1213                                 break;
1214                 }
1215
1216                 offset = xlog_align(log, start_block, endcount, bp);
1217                 for (; j < endcount; j++) {
1218                         xlog_add_record(log, offset, cycle, i+j,
1219                                         tail_cycle, tail_block);
1220                         offset += BBSIZE;
1221                 }
1222                 error = xlog_bwrite(log, start_block, endcount, bp);
1223                 if (error)
1224                         break;
1225                 start_block += endcount;
1226                 j = 0;
1227         }
1228
1229  out_put_bp:
1230         xlog_put_bp(bp);
1231         return error;
1232 }
1233
1234 /*
1235  * This routine is called to blow away any incomplete log writes out
1236  * in front of the log head.  We do this so that we won't become confused
1237  * if we come up, write only a little bit more, and then crash again.
1238  * If we leave the partial log records out there, this situation could
1239  * cause us to think those partial writes are valid blocks since they
1240  * have the current cycle number.  We get rid of them by overwriting them
1241  * with empty log records with the old cycle number rather than the
1242  * current one.
1243  *
1244  * The tail lsn is passed in rather than taken from
1245  * the log so that we will not write over the unmount record after a
1246  * clean unmount in a 512 block log.  Doing so would leave the log without
1247  * any valid log records in it until a new one was written.  If we crashed
1248  * during that time we would not be able to recover.
1249  */
1250 STATIC int
1251 xlog_clear_stale_blocks(
1252         xlog_t          *log,
1253         xfs_lsn_t       tail_lsn)
1254 {
1255         int             tail_cycle, head_cycle;
1256         int             tail_block, head_block;
1257         int             tail_distance, max_distance;
1258         int             distance;
1259         int             error;
1260
1261         tail_cycle = CYCLE_LSN(tail_lsn);
1262         tail_block = BLOCK_LSN(tail_lsn);
1263         head_cycle = log->l_curr_cycle;
1264         head_block = log->l_curr_block;
1265
1266         /*
1267          * Figure out the distance between the new head of the log
1268          * and the tail.  We want to write over any blocks beyond the
1269          * head that we may have written just before the crash, but
1270          * we don't want to overwrite the tail of the log.
1271          */
1272         if (head_cycle == tail_cycle) {
1273                 /*
1274                  * The tail is behind the head in the physical log,
1275                  * so the distance from the head to the tail is the
1276                  * distance from the head to the end of the log plus
1277                  * the distance from the beginning of the log to the
1278                  * tail.
1279                  */
1280                 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1281                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1282                                          XFS_ERRLEVEL_LOW, log->l_mp);
1283                         return XFS_ERROR(EFSCORRUPTED);
1284                 }
1285                 tail_distance = tail_block + (log->l_logBBsize - head_block);
1286         } else {
1287                 /*
1288                  * The head is behind the tail in the physical log,
1289                  * so the distance from the head to the tail is just
1290                  * the tail block minus the head block.
1291                  */
1292                 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1293                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1294                                          XFS_ERRLEVEL_LOW, log->l_mp);
1295                         return XFS_ERROR(EFSCORRUPTED);
1296                 }
1297                 tail_distance = tail_block - head_block;
1298         }
1299
1300         /*
1301          * If the head is right up against the tail, we can't clear
1302          * anything.
1303          */
1304         if (tail_distance <= 0) {
1305                 ASSERT(tail_distance == 0);
1306                 return 0;
1307         }
1308
1309         max_distance = XLOG_TOTAL_REC_SHIFT(log);
1310         /*
1311          * Take the smaller of the maximum amount of outstanding I/O
1312          * we could have and the distance to the tail to clear out.
1313          * We take the smaller so that we don't overwrite the tail and
1314          * we don't waste all day writing from the head to the tail
1315          * for no reason.
1316          */
1317         max_distance = MIN(max_distance, tail_distance);
1318
1319         if ((head_block + max_distance) <= log->l_logBBsize) {
1320                 /*
1321                  * We can stomp all the blocks we need to without
1322                  * wrapping around the end of the log.  Just do it
1323                  * in a single write.  Use the cycle number of the
1324                  * current cycle minus one so that the log will look like:
1325                  *     n ... | n - 1 ...
1326                  */
1327                 error = xlog_write_log_records(log, (head_cycle - 1),
1328                                 head_block, max_distance, tail_cycle,
1329                                 tail_block);
1330                 if (error)
1331                         return error;
1332         } else {
1333                 /*
1334                  * We need to wrap around the end of the physical log in
1335                  * order to clear all the blocks.  Do it in two separate
1336                  * I/Os.  The first write should be from the head to the
1337                  * end of the physical log, and it should use the current
1338                  * cycle number minus one just like above.
1339                  */
1340                 distance = log->l_logBBsize - head_block;
1341                 error = xlog_write_log_records(log, (head_cycle - 1),
1342                                 head_block, distance, tail_cycle,
1343                                 tail_block);
1344
1345                 if (error)
1346                         return error;
1347
1348                 /*
1349                  * Now write the blocks at the start of the physical log.
1350                  * This writes the remainder of the blocks we want to clear.
1351                  * It uses the current cycle number since we're now on the
1352                  * same cycle as the head so that we get:
1353                  *    n ... n ... | n - 1 ...
1354                  *    ^^^^^ blocks we're writing
1355                  */
1356                 distance = max_distance - (log->l_logBBsize - head_block);
1357                 error = xlog_write_log_records(log, head_cycle, 0, distance,
1358                                 tail_cycle, tail_block);
1359                 if (error)
1360                         return error;
1361         }
1362
1363         return 0;
1364 }
1365
1366 /******************************************************************************
1367  *
1368  *              Log recover routines
1369  *
1370  ******************************************************************************
1371  */
1372
1373 STATIC xlog_recover_t *
1374 xlog_recover_find_tid(
1375         xlog_recover_t          *q,
1376         xlog_tid_t              tid)
1377 {
1378         xlog_recover_t          *p = q;
1379
1380         while (p != NULL) {
1381                 if (p->r_log_tid == tid)
1382                     break;
1383                 p = p->r_next;
1384         }
1385         return p;
1386 }
1387
1388 STATIC void
1389 xlog_recover_put_hashq(
1390         xlog_recover_t          **q,
1391         xlog_recover_t          *trans)
1392 {
1393         trans->r_next = *q;
1394         *q = trans;
1395 }
1396
1397 STATIC void
1398 xlog_recover_add_item(
1399         xlog_recover_item_t     **itemq)
1400 {
1401         xlog_recover_item_t     *item;
1402
1403         item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1404         xlog_recover_insert_item_backq(itemq, item);
1405 }
1406
1407 STATIC int
1408 xlog_recover_add_to_cont_trans(
1409         xlog_recover_t          *trans,
1410         xfs_caddr_t             dp,
1411         int                     len)
1412 {
1413         xlog_recover_item_t     *item;
1414         xfs_caddr_t             ptr, old_ptr;
1415         int                     old_len;
1416
1417         item = trans->r_itemq;
1418         if (item == NULL) {
1419                 /* finish copying rest of trans header */
1420                 xlog_recover_add_item(&trans->r_itemq);
1421                 ptr = (xfs_caddr_t) &trans->r_theader +
1422                                 sizeof(xfs_trans_header_t) - len;
1423                 memcpy(ptr, dp, len); /* d, s, l */
1424                 return 0;
1425         }
1426         item = item->ri_prev;
1427
1428         old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1429         old_len = item->ri_buf[item->ri_cnt-1].i_len;
1430
1431         ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1432         memcpy(&ptr[old_len], dp, len); /* d, s, l */
1433         item->ri_buf[item->ri_cnt-1].i_len += len;
1434         item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1435         return 0;
1436 }
1437
1438 /*
1439  * The next region to add is the start of a new region.  It could be
1440  * a whole region or it could be the first part of a new region.  Because
1441  * of this, the assumption here is that the type and size fields of all
1442  * format structures fit into the first 32 bits of the structure.
1443  *
1444  * This works because all regions must be 32 bit aligned.  Therefore, we
1445  * either have both fields or we have neither field.  In the case we have
1446  * neither field, the data part of the region is zero length.  We only have
1447  * a log_op_header and can throw away the header since a new one will appear
1448  * later.  If we have at least 4 bytes, then we can determine how many regions
1449  * will appear in the current log item.
1450  */
1451 STATIC int
1452 xlog_recover_add_to_trans(
1453         xlog_recover_t          *trans,
1454         xfs_caddr_t             dp,
1455         int                     len)
1456 {
1457         xfs_inode_log_format_t  *in_f;                  /* any will do */
1458         xlog_recover_item_t     *item;
1459         xfs_caddr_t             ptr;
1460
1461         if (!len)
1462                 return 0;
1463         item = trans->r_itemq;
1464         if (item == NULL) {
1465                 /* we need to catch log corruptions here */
1466                 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1467                         xlog_warn("XFS: xlog_recover_add_to_trans: "
1468                                   "bad header magic number");
1469                         ASSERT(0);
1470                         return XFS_ERROR(EIO);
1471                 }
1472                 if (len == sizeof(xfs_trans_header_t))
1473                         xlog_recover_add_item(&trans->r_itemq);
1474                 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1475                 return 0;
1476         }
1477
1478         ptr = kmem_alloc(len, KM_SLEEP);
1479         memcpy(ptr, dp, len);
1480         in_f = (xfs_inode_log_format_t *)ptr;
1481
1482         if (item->ri_prev->ri_total != 0 &&
1483              item->ri_prev->ri_total == item->ri_prev->ri_cnt) {
1484                 xlog_recover_add_item(&trans->r_itemq);
1485         }
1486         item = trans->r_itemq;
1487         item = item->ri_prev;
1488
1489         if (item->ri_total == 0) {              /* first region to be added */
1490                 if (in_f->ilf_size == 0 ||
1491                     in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
1492                         xlog_warn(
1493         "XFS: bad number of regions (%d) in inode log format",
1494                                   in_f->ilf_size);
1495                         ASSERT(0);
1496                         return XFS_ERROR(EIO);
1497                 }
1498
1499                 item->ri_total = in_f->ilf_size;
1500                 item->ri_buf =
1501                         kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
1502                                     KM_SLEEP);
1503         }
1504         ASSERT(item->ri_total > item->ri_cnt);
1505         /* Description region is ri_buf[0] */
1506         item->ri_buf[item->ri_cnt].i_addr = ptr;
1507         item->ri_buf[item->ri_cnt].i_len  = len;
1508         item->ri_cnt++;
1509         return 0;
1510 }
1511
1512 STATIC void
1513 xlog_recover_new_tid(
1514         xlog_recover_t          **q,
1515         xlog_tid_t              tid,
1516         xfs_lsn_t               lsn)
1517 {
1518         xlog_recover_t          *trans;
1519
1520         trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1521         trans->r_log_tid   = tid;
1522         trans->r_lsn       = lsn;
1523         xlog_recover_put_hashq(q, trans);
1524 }
1525
1526 STATIC int
1527 xlog_recover_unlink_tid(
1528         xlog_recover_t          **q,
1529         xlog_recover_t          *trans)
1530 {
1531         xlog_recover_t          *tp;
1532         int                     found = 0;
1533
1534         ASSERT(trans != NULL);
1535         if (trans == *q) {
1536                 *q = (*q)->r_next;
1537         } else {
1538                 tp = *q;
1539                 while (tp) {
1540                         if (tp->r_next == trans) {
1541                                 found = 1;
1542                                 break;
1543                         }
1544                         tp = tp->r_next;
1545                 }
1546                 if (!found) {
1547                         xlog_warn(
1548                              "XFS: xlog_recover_unlink_tid: trans not found");
1549                         ASSERT(0);
1550                         return XFS_ERROR(EIO);
1551                 }
1552                 tp->r_next = tp->r_next->r_next;
1553         }
1554         return 0;
1555 }
1556
1557 STATIC void
1558 xlog_recover_insert_item_backq(
1559         xlog_recover_item_t     **q,
1560         xlog_recover_item_t     *item)
1561 {
1562         if (*q == NULL) {
1563                 item->ri_prev = item->ri_next = item;
1564                 *q = item;
1565         } else {
1566                 item->ri_next           = *q;
1567                 item->ri_prev           = (*q)->ri_prev;
1568                 (*q)->ri_prev           = item;
1569                 item->ri_prev->ri_next  = item;
1570         }
1571 }
1572
1573 STATIC void
1574 xlog_recover_insert_item_frontq(
1575         xlog_recover_item_t     **q,
1576         xlog_recover_item_t     *item)
1577 {
1578         xlog_recover_insert_item_backq(q, item);
1579         *q = item;
1580 }
1581
1582 STATIC int
1583 xlog_recover_reorder_trans(
1584         xlog_recover_t          *trans)
1585 {
1586         xlog_recover_item_t     *first_item, *itemq, *itemq_next;
1587         xfs_buf_log_format_t    *buf_f;
1588         ushort                  flags = 0;
1589
1590         first_item = itemq = trans->r_itemq;
1591         trans->r_itemq = NULL;
1592         do {
1593                 itemq_next = itemq->ri_next;
1594                 buf_f = (xfs_buf_log_format_t *)itemq->ri_buf[0].i_addr;
1595
1596                 switch (ITEM_TYPE(itemq)) {
1597                 case XFS_LI_BUF:
1598                         flags = buf_f->blf_flags;
1599                         if (!(flags & XFS_BLI_CANCEL)) {
1600                                 xlog_recover_insert_item_frontq(&trans->r_itemq,
1601                                                                 itemq);
1602                                 break;
1603                         }
1604                 case XFS_LI_INODE:
1605                 case XFS_LI_DQUOT:
1606                 case XFS_LI_QUOTAOFF:
1607                 case XFS_LI_EFD:
1608                 case XFS_LI_EFI:
1609                         xlog_recover_insert_item_backq(&trans->r_itemq, itemq);
1610                         break;
1611                 default:
1612                         xlog_warn(
1613         "XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1614                         ASSERT(0);
1615                         return XFS_ERROR(EIO);
1616                 }
1617                 itemq = itemq_next;
1618         } while (first_item != itemq);
1619         return 0;
1620 }
1621
1622 /*
1623  * Build up the table of buf cancel records so that we don't replay
1624  * cancelled data in the second pass.  For buffer records that are
1625  * not cancel records, there is nothing to do here so we just return.
1626  *
1627  * If we get a cancel record which is already in the table, this indicates
1628  * that the buffer was cancelled multiple times.  In order to ensure
1629  * that during pass 2 we keep the record in the table until we reach its
1630  * last occurrence in the log, we keep a reference count in the cancel
1631  * record in the table to tell us how many times we expect to see this
1632  * record during the second pass.
1633  */
1634 STATIC void
1635 xlog_recover_do_buffer_pass1(
1636         xlog_t                  *log,
1637         xfs_buf_log_format_t    *buf_f)
1638 {
1639         xfs_buf_cancel_t        *bcp;
1640         xfs_buf_cancel_t        *nextp;
1641         xfs_buf_cancel_t        *prevp;
1642         xfs_buf_cancel_t        **bucket;
1643         xfs_daddr_t             blkno = 0;
1644         uint                    len = 0;
1645         ushort                  flags = 0;
1646
1647         switch (buf_f->blf_type) {
1648         case XFS_LI_BUF:
1649                 blkno = buf_f->blf_blkno;
1650                 len = buf_f->blf_len;
1651                 flags = buf_f->blf_flags;
1652                 break;
1653         }
1654
1655         /*
1656          * If this isn't a cancel buffer item, then just return.
1657          */
1658         if (!(flags & XFS_BLI_CANCEL))
1659                 return;
1660
1661         /*
1662          * Insert an xfs_buf_cancel record into the hash table of
1663          * them.  If there is already an identical record, bump
1664          * its reference count.
1665          */
1666         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1667                                           XLOG_BC_TABLE_SIZE];
1668         /*
1669          * If the hash bucket is empty then just insert a new record into
1670          * the bucket.
1671          */
1672         if (*bucket == NULL) {
1673                 bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1674                                                      KM_SLEEP);
1675                 bcp->bc_blkno = blkno;
1676                 bcp->bc_len = len;
1677                 bcp->bc_refcount = 1;
1678                 bcp->bc_next = NULL;
1679                 *bucket = bcp;
1680                 return;
1681         }
1682
1683         /*
1684          * The hash bucket is not empty, so search for duplicates of our
1685          * record.  If we find one them just bump its refcount.  If not
1686          * then add us at the end of the list.
1687          */
1688         prevp = NULL;
1689         nextp = *bucket;
1690         while (nextp != NULL) {
1691                 if (nextp->bc_blkno == blkno && nextp->bc_len == len) {
1692                         nextp->bc_refcount++;
1693                         return;
1694                 }
1695                 prevp = nextp;
1696                 nextp = nextp->bc_next;
1697         }
1698         ASSERT(prevp != NULL);
1699         bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1700                                              KM_SLEEP);
1701         bcp->bc_blkno = blkno;
1702         bcp->bc_len = len;
1703         bcp->bc_refcount = 1;
1704         bcp->bc_next = NULL;
1705         prevp->bc_next = bcp;
1706 }
1707
1708 /*
1709  * Check to see whether the buffer being recovered has a corresponding
1710  * entry in the buffer cancel record table.  If it does then return 1
1711  * so that it will be cancelled, otherwise return 0.  If the buffer is
1712  * actually a buffer cancel item (XFS_BLI_CANCEL is set), then decrement
1713  * the refcount on the entry in the table and remove it from the table
1714  * if this is the last reference.
1715  *
1716  * We remove the cancel record from the table when we encounter its
1717  * last occurrence in the log so that if the same buffer is re-used
1718  * again after its last cancellation we actually replay the changes
1719  * made at that point.
1720  */
1721 STATIC int
1722 xlog_check_buffer_cancelled(
1723         xlog_t                  *log,
1724         xfs_daddr_t             blkno,
1725         uint                    len,
1726         ushort                  flags)
1727 {
1728         xfs_buf_cancel_t        *bcp;
1729         xfs_buf_cancel_t        *prevp;
1730         xfs_buf_cancel_t        **bucket;
1731
1732         if (log->l_buf_cancel_table == NULL) {
1733                 /*
1734                  * There is nothing in the table built in pass one,
1735                  * so this buffer must not be cancelled.
1736                  */
1737                 ASSERT(!(flags & XFS_BLI_CANCEL));
1738                 return 0;
1739         }
1740
1741         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1742                                           XLOG_BC_TABLE_SIZE];
1743         bcp = *bucket;
1744         if (bcp == NULL) {
1745                 /*
1746                  * There is no corresponding entry in the table built
1747                  * in pass one, so this buffer has not been cancelled.
1748                  */
1749                 ASSERT(!(flags & XFS_BLI_CANCEL));
1750                 return 0;
1751         }
1752
1753         /*
1754          * Search for an entry in the buffer cancel table that
1755          * matches our buffer.
1756          */
1757         prevp = NULL;
1758         while (bcp != NULL) {
1759                 if (bcp->bc_blkno == blkno && bcp->bc_len == len) {
1760                         /*
1761                          * We've go a match, so return 1 so that the
1762                          * recovery of this buffer is cancelled.
1763                          * If this buffer is actually a buffer cancel
1764                          * log item, then decrement the refcount on the
1765                          * one in the table and remove it if this is the
1766                          * last reference.
1767                          */
1768                         if (flags & XFS_BLI_CANCEL) {
1769                                 bcp->bc_refcount--;
1770                                 if (bcp->bc_refcount == 0) {
1771                                         if (prevp == NULL) {
1772                                                 *bucket = bcp->bc_next;
1773                                         } else {
1774                                                 prevp->bc_next = bcp->bc_next;
1775                                         }
1776                                         kmem_free(bcp);
1777                                 }
1778                         }
1779                         return 1;
1780                 }
1781                 prevp = bcp;
1782                 bcp = bcp->bc_next;
1783         }
1784         /*
1785          * We didn't find a corresponding entry in the table, so
1786          * return 0 so that the buffer is NOT cancelled.
1787          */
1788         ASSERT(!(flags & XFS_BLI_CANCEL));
1789         return 0;
1790 }
1791
1792 STATIC int
1793 xlog_recover_do_buffer_pass2(
1794         xlog_t                  *log,
1795         xfs_buf_log_format_t    *buf_f)
1796 {
1797         xfs_daddr_t             blkno = 0;
1798         ushort                  flags = 0;
1799         uint                    len = 0;
1800
1801         switch (buf_f->blf_type) {
1802         case XFS_LI_BUF:
1803                 blkno = buf_f->blf_blkno;
1804                 flags = buf_f->blf_flags;
1805                 len = buf_f->blf_len;
1806                 break;
1807         }
1808
1809         return xlog_check_buffer_cancelled(log, blkno, len, flags);
1810 }
1811
1812 /*
1813  * Perform recovery for a buffer full of inodes.  In these buffers,
1814  * the only data which should be recovered is that which corresponds
1815  * to the di_next_unlinked pointers in the on disk inode structures.
1816  * The rest of the data for the inodes is always logged through the
1817  * inodes themselves rather than the inode buffer and is recovered
1818  * in xlog_recover_do_inode_trans().
1819  *
1820  * The only time when buffers full of inodes are fully recovered is
1821  * when the buffer is full of newly allocated inodes.  In this case
1822  * the buffer will not be marked as an inode buffer and so will be
1823  * sent to xlog_recover_do_reg_buffer() below during recovery.
1824  */
1825 STATIC int
1826 xlog_recover_do_inode_buffer(
1827         xfs_mount_t             *mp,
1828         xlog_recover_item_t     *item,
1829         xfs_buf_t               *bp,
1830         xfs_buf_log_format_t    *buf_f)
1831 {
1832         int                     i;
1833         int                     item_index;
1834         int                     bit;
1835         int                     nbits;
1836         int                     reg_buf_offset;
1837         int                     reg_buf_bytes;
1838         int                     next_unlinked_offset;
1839         int                     inodes_per_buf;
1840         xfs_agino_t             *logged_nextp;
1841         xfs_agino_t             *buffer_nextp;
1842         unsigned int            *data_map = NULL;
1843         unsigned int            map_size = 0;
1844
1845         switch (buf_f->blf_type) {
1846         case XFS_LI_BUF:
1847                 data_map = buf_f->blf_data_map;
1848                 map_size = buf_f->blf_map_size;
1849                 break;
1850         }
1851         /*
1852          * Set the variables corresponding to the current region to
1853          * 0 so that we'll initialize them on the first pass through
1854          * the loop.
1855          */
1856         reg_buf_offset = 0;
1857         reg_buf_bytes = 0;
1858         bit = 0;
1859         nbits = 0;
1860         item_index = 0;
1861         inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1862         for (i = 0; i < inodes_per_buf; i++) {
1863                 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1864                         offsetof(xfs_dinode_t, di_next_unlinked);
1865
1866                 while (next_unlinked_offset >=
1867                        (reg_buf_offset + reg_buf_bytes)) {
1868                         /*
1869                          * The next di_next_unlinked field is beyond
1870                          * the current logged region.  Find the next
1871                          * logged region that contains or is beyond
1872                          * the current di_next_unlinked field.
1873                          */
1874                         bit += nbits;
1875                         bit = xfs_next_bit(data_map, map_size, bit);
1876
1877                         /*
1878                          * If there are no more logged regions in the
1879                          * buffer, then we're done.
1880                          */
1881                         if (bit == -1) {
1882                                 return 0;
1883                         }
1884
1885                         nbits = xfs_contig_bits(data_map, map_size,
1886                                                          bit);
1887                         ASSERT(nbits > 0);
1888                         reg_buf_offset = bit << XFS_BLI_SHIFT;
1889                         reg_buf_bytes = nbits << XFS_BLI_SHIFT;
1890                         item_index++;
1891                 }
1892
1893                 /*
1894                  * If the current logged region starts after the current
1895                  * di_next_unlinked field, then move on to the next
1896                  * di_next_unlinked field.
1897                  */
1898                 if (next_unlinked_offset < reg_buf_offset) {
1899                         continue;
1900                 }
1901
1902                 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1903                 ASSERT((item->ri_buf[item_index].i_len % XFS_BLI_CHUNK) == 0);
1904                 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1905
1906                 /*
1907                  * The current logged region contains a copy of the
1908                  * current di_next_unlinked field.  Extract its value
1909                  * and copy it to the buffer copy.
1910                  */
1911                 logged_nextp = (xfs_agino_t *)
1912                                ((char *)(item->ri_buf[item_index].i_addr) +
1913                                 (next_unlinked_offset - reg_buf_offset));
1914                 if (unlikely(*logged_nextp == 0)) {
1915                         xfs_fs_cmn_err(CE_ALERT, mp,
1916                                 "bad inode buffer log record (ptr = 0x%p, bp = 0x%p).  XFS trying to replay bad (0) inode di_next_unlinked field",
1917                                 item, bp);
1918                         XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1919                                          XFS_ERRLEVEL_LOW, mp);
1920                         return XFS_ERROR(EFSCORRUPTED);
1921                 }
1922
1923                 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1924                                               next_unlinked_offset);
1925                 *buffer_nextp = *logged_nextp;
1926         }
1927
1928         return 0;
1929 }
1930
1931 /*
1932  * Perform a 'normal' buffer recovery.  Each logged region of the
1933  * buffer should be copied over the corresponding region in the
1934  * given buffer.  The bitmap in the buf log format structure indicates
1935  * where to place the logged data.
1936  */
1937 /*ARGSUSED*/
1938 STATIC void
1939 xlog_recover_do_reg_buffer(
1940         xlog_recover_item_t     *item,
1941         xfs_buf_t               *bp,
1942         xfs_buf_log_format_t    *buf_f)
1943 {
1944         int                     i;
1945         int                     bit;
1946         int                     nbits;
1947         unsigned int            *data_map = NULL;
1948         unsigned int            map_size = 0;
1949         int                     error;
1950
1951         switch (buf_f->blf_type) {
1952         case XFS_LI_BUF:
1953                 data_map = buf_f->blf_data_map;
1954                 map_size = buf_f->blf_map_size;
1955                 break;
1956         }
1957         bit = 0;
1958         i = 1;  /* 0 is the buf format structure */
1959         while (1) {
1960                 bit = xfs_next_bit(data_map, map_size, bit);
1961                 if (bit == -1)
1962                         break;
1963                 nbits = xfs_contig_bits(data_map, map_size, bit);
1964                 ASSERT(nbits > 0);
1965                 ASSERT(item->ri_buf[i].i_addr != NULL);
1966                 ASSERT(item->ri_buf[i].i_len % XFS_BLI_CHUNK == 0);
1967                 ASSERT(XFS_BUF_COUNT(bp) >=
1968                        ((uint)bit << XFS_BLI_SHIFT)+(nbits<<XFS_BLI_SHIFT));
1969
1970                 /*
1971                  * Do a sanity check if this is a dquot buffer. Just checking
1972                  * the first dquot in the buffer should do. XXXThis is
1973                  * probably a good thing to do for other buf types also.
1974                  */
1975                 error = 0;
1976                 if (buf_f->blf_flags &
1977                    (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
1978                         error = xfs_qm_dqcheck((xfs_disk_dquot_t *)
1979                                                item->ri_buf[i].i_addr,
1980                                                -1, 0, XFS_QMOPT_DOWARN,
1981                                                "dquot_buf_recover");
1982                 }
1983                 if (!error)
1984                         memcpy(xfs_buf_offset(bp,
1985                                 (uint)bit << XFS_BLI_SHIFT),    /* dest */
1986                                 item->ri_buf[i].i_addr,         /* source */
1987                                 nbits<<XFS_BLI_SHIFT);          /* length */
1988                 i++;
1989                 bit += nbits;
1990         }
1991
1992         /* Shouldn't be any more regions */
1993         ASSERT(i == item->ri_total);
1994 }
1995
1996 /*
1997  * Do some primitive error checking on ondisk dquot data structures.
1998  */
1999 int
2000 xfs_qm_dqcheck(
2001         xfs_disk_dquot_t *ddq,
2002         xfs_dqid_t       id,
2003         uint             type,    /* used only when IO_dorepair is true */
2004         uint             flags,
2005         char             *str)
2006 {
2007         xfs_dqblk_t      *d = (xfs_dqblk_t *)ddq;
2008         int             errs = 0;
2009
2010         /*
2011          * We can encounter an uninitialized dquot buffer for 2 reasons:
2012          * 1. If we crash while deleting the quotainode(s), and those blks got
2013          *    used for user data. This is because we take the path of regular
2014          *    file deletion; however, the size field of quotainodes is never
2015          *    updated, so all the tricks that we play in itruncate_finish
2016          *    don't quite matter.
2017          *
2018          * 2. We don't play the quota buffers when there's a quotaoff logitem.
2019          *    But the allocation will be replayed so we'll end up with an
2020          *    uninitialized quota block.
2021          *
2022          * This is all fine; things are still consistent, and we haven't lost
2023          * any quota information. Just don't complain about bad dquot blks.
2024          */
2025         if (be16_to_cpu(ddq->d_magic) != XFS_DQUOT_MAGIC) {
2026                 if (flags & XFS_QMOPT_DOWARN)
2027                         cmn_err(CE_ALERT,
2028                         "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
2029                         str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
2030                 errs++;
2031         }
2032         if (ddq->d_version != XFS_DQUOT_VERSION) {
2033                 if (flags & XFS_QMOPT_DOWARN)
2034                         cmn_err(CE_ALERT,
2035                         "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
2036                         str, id, ddq->d_version, XFS_DQUOT_VERSION);
2037                 errs++;
2038         }
2039
2040         if (ddq->d_flags != XFS_DQ_USER &&
2041             ddq->d_flags != XFS_DQ_PROJ &&
2042             ddq->d_flags != XFS_DQ_GROUP) {
2043                 if (flags & XFS_QMOPT_DOWARN)
2044                         cmn_err(CE_ALERT,
2045                         "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
2046                         str, id, ddq->d_flags);
2047                 errs++;
2048         }
2049
2050         if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
2051                 if (flags & XFS_QMOPT_DOWARN)
2052                         cmn_err(CE_ALERT,
2053                         "%s : ondisk-dquot 0x%p, ID mismatch: "
2054                         "0x%x expected, found id 0x%x",
2055                         str, ddq, id, be32_to_cpu(ddq->d_id));
2056                 errs++;
2057         }
2058
2059         if (!errs && ddq->d_id) {
2060                 if (ddq->d_blk_softlimit &&
2061                     be64_to_cpu(ddq->d_bcount) >=
2062                                 be64_to_cpu(ddq->d_blk_softlimit)) {
2063                         if (!ddq->d_btimer) {
2064                                 if (flags & XFS_QMOPT_DOWARN)
2065                                         cmn_err(CE_ALERT,
2066                                         "%s : Dquot ID 0x%x (0x%p) "
2067                                         "BLK TIMER NOT STARTED",
2068                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2069                                 errs++;
2070                         }
2071                 }
2072                 if (ddq->d_ino_softlimit &&
2073                     be64_to_cpu(ddq->d_icount) >=
2074                                 be64_to_cpu(ddq->d_ino_softlimit)) {
2075                         if (!ddq->d_itimer) {
2076                                 if (flags & XFS_QMOPT_DOWARN)
2077                                         cmn_err(CE_ALERT,
2078                                         "%s : Dquot ID 0x%x (0x%p) "
2079                                         "INODE TIMER NOT STARTED",
2080                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2081                                 errs++;
2082                         }
2083                 }
2084                 if (ddq->d_rtb_softlimit &&
2085                     be64_to_cpu(ddq->d_rtbcount) >=
2086                                 be64_to_cpu(ddq->d_rtb_softlimit)) {
2087                         if (!ddq->d_rtbtimer) {
2088                                 if (flags & XFS_QMOPT_DOWARN)
2089                                         cmn_err(CE_ALERT,
2090                                         "%s : Dquot ID 0x%x (0x%p) "
2091                                         "RTBLK TIMER NOT STARTED",
2092                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2093                                 errs++;
2094                         }
2095                 }
2096         }
2097
2098         if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2099                 return errs;
2100
2101         if (flags & XFS_QMOPT_DOWARN)
2102                 cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2103
2104         /*
2105          * Typically, a repair is only requested by quotacheck.
2106          */
2107         ASSERT(id != -1);
2108         ASSERT(flags & XFS_QMOPT_DQREPAIR);
2109         memset(d, 0, sizeof(xfs_dqblk_t));
2110
2111         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2112         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2113         d->dd_diskdq.d_flags = type;
2114         d->dd_diskdq.d_id = cpu_to_be32(id);
2115
2116         return errs;
2117 }
2118
2119 /*
2120  * Perform a dquot buffer recovery.
2121  * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2122  * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2123  * Else, treat it as a regular buffer and do recovery.
2124  */
2125 STATIC void
2126 xlog_recover_do_dquot_buffer(
2127         xfs_mount_t             *mp,
2128         xlog_t                  *log,
2129         xlog_recover_item_t     *item,
2130         xfs_buf_t               *bp,
2131         xfs_buf_log_format_t    *buf_f)
2132 {
2133         uint                    type;
2134
2135         /*
2136          * Filesystems are required to send in quota flags at mount time.
2137          */
2138         if (mp->m_qflags == 0) {
2139                 return;
2140         }
2141
2142         type = 0;
2143         if (buf_f->blf_flags & XFS_BLI_UDQUOT_BUF)
2144                 type |= XFS_DQ_USER;
2145         if (buf_f->blf_flags & XFS_BLI_PDQUOT_BUF)
2146                 type |= XFS_DQ_PROJ;
2147         if (buf_f->blf_flags & XFS_BLI_GDQUOT_BUF)
2148                 type |= XFS_DQ_GROUP;
2149         /*
2150          * This type of quotas was turned off, so ignore this buffer
2151          */
2152         if (log->l_quotaoffs_flag & type)
2153                 return;
2154
2155         xlog_recover_do_reg_buffer(item, bp, buf_f);
2156 }
2157
2158 /*
2159  * This routine replays a modification made to a buffer at runtime.
2160  * There are actually two types of buffer, regular and inode, which
2161  * are handled differently.  Inode buffers are handled differently
2162  * in that we only recover a specific set of data from them, namely
2163  * the inode di_next_unlinked fields.  This is because all other inode
2164  * data is actually logged via inode records and any data we replay
2165  * here which overlaps that may be stale.
2166  *
2167  * When meta-data buffers are freed at run time we log a buffer item
2168  * with the XFS_BLI_CANCEL bit set to indicate that previous copies
2169  * of the buffer in the log should not be replayed at recovery time.
2170  * This is so that if the blocks covered by the buffer are reused for
2171  * file data before we crash we don't end up replaying old, freed
2172  * meta-data into a user's file.
2173  *
2174  * To handle the cancellation of buffer log items, we make two passes
2175  * over the log during recovery.  During the first we build a table of
2176  * those buffers which have been cancelled, and during the second we
2177  * only replay those buffers which do not have corresponding cancel
2178  * records in the table.  See xlog_recover_do_buffer_pass[1,2] above
2179  * for more details on the implementation of the table of cancel records.
2180  */
2181 STATIC int
2182 xlog_recover_do_buffer_trans(
2183         xlog_t                  *log,
2184         xlog_recover_item_t     *item,
2185         int                     pass)
2186 {
2187         xfs_buf_log_format_t    *buf_f;
2188         xfs_mount_t             *mp;
2189         xfs_buf_t               *bp;
2190         int                     error;
2191         int                     cancel;
2192         xfs_daddr_t             blkno;
2193         int                     len;
2194         ushort                  flags;
2195
2196         buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
2197
2198         if (pass == XLOG_RECOVER_PASS1) {
2199                 /*
2200                  * In this pass we're only looking for buf items
2201                  * with the XFS_BLI_CANCEL bit set.
2202                  */
2203                 xlog_recover_do_buffer_pass1(log, buf_f);
2204                 return 0;
2205         } else {
2206                 /*
2207                  * In this pass we want to recover all the buffers
2208                  * which have not been cancelled and are not
2209                  * cancellation buffers themselves.  The routine
2210                  * we call here will tell us whether or not to
2211                  * continue with the replay of this buffer.
2212                  */
2213                 cancel = xlog_recover_do_buffer_pass2(log, buf_f);
2214                 if (cancel) {
2215                         return 0;
2216                 }
2217         }
2218         switch (buf_f->blf_type) {
2219         case XFS_LI_BUF:
2220                 blkno = buf_f->blf_blkno;
2221                 len = buf_f->blf_len;
2222                 flags = buf_f->blf_flags;
2223                 break;
2224         default:
2225                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
2226                         "xfs_log_recover: unknown buffer type 0x%x, logdev %s",
2227                         buf_f->blf_type, log->l_mp->m_logname ?
2228                         log->l_mp->m_logname : "internal");
2229                 XFS_ERROR_REPORT("xlog_recover_do_buffer_trans",
2230                                  XFS_ERRLEVEL_LOW, log->l_mp);
2231                 return XFS_ERROR(EFSCORRUPTED);
2232         }
2233
2234         mp = log->l_mp;
2235         if (flags & XFS_BLI_INODE_BUF) {
2236                 bp = xfs_buf_read_flags(mp->m_ddev_targp, blkno, len,
2237                                                                 XFS_BUF_LOCK);
2238         } else {
2239                 bp = xfs_buf_read(mp->m_ddev_targp, blkno, len, 0);
2240         }
2241         if (XFS_BUF_ISERROR(bp)) {
2242                 xfs_ioerror_alert("xlog_recover_do..(read#1)", log->l_mp,
2243                                   bp, blkno);
2244                 error = XFS_BUF_GETERROR(bp);
2245                 xfs_buf_relse(bp);
2246                 return error;
2247         }
2248
2249         error = 0;
2250         if (flags & XFS_BLI_INODE_BUF) {
2251                 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2252         } else if (flags &
2253                   (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
2254                 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2255         } else {
2256                 xlog_recover_do_reg_buffer(item, bp, buf_f);
2257         }
2258         if (error)
2259                 return XFS_ERROR(error);
2260
2261         /*
2262          * Perform delayed write on the buffer.  Asynchronous writes will be
2263          * slower when taking into account all the buffers to be flushed.
2264          *
2265          * Also make sure that only inode buffers with good sizes stay in
2266          * the buffer cache.  The kernel moves inodes in buffers of 1 block
2267          * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger.  The inode
2268          * buffers in the log can be a different size if the log was generated
2269          * by an older kernel using unclustered inode buffers or a newer kernel
2270          * running with a different inode cluster size.  Regardless, if the
2271          * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2272          * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2273          * the buffer out of the buffer cache so that the buffer won't
2274          * overlap with future reads of those inodes.
2275          */
2276         if (XFS_DINODE_MAGIC ==
2277             be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2278             (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2279                         (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2280                 XFS_BUF_STALE(bp);
2281                 error = xfs_bwrite(mp, bp);
2282         } else {
2283                 ASSERT(bp->b_mount == NULL || bp->b_mount == mp);
2284                 bp->b_mount = mp;
2285                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2286                 xfs_bdwrite(mp, bp);
2287         }
2288
2289         return (error);
2290 }
2291
2292 STATIC int
2293 xlog_recover_do_inode_trans(
2294         xlog_t                  *log,
2295         xlog_recover_item_t     *item,
2296         int                     pass)
2297 {
2298         xfs_inode_log_format_t  *in_f;
2299         xfs_mount_t             *mp;
2300         xfs_buf_t               *bp;
2301         xfs_dinode_t            *dip;
2302         xfs_ino_t               ino;
2303         int                     len;
2304         xfs_caddr_t             src;
2305         xfs_caddr_t             dest;
2306         int                     error;
2307         int                     attr_index;
2308         uint                    fields;
2309         xfs_icdinode_t          *dicp;
2310         int                     need_free = 0;
2311
2312         if (pass == XLOG_RECOVER_PASS1) {
2313                 return 0;
2314         }
2315
2316         if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2317                 in_f = (xfs_inode_log_format_t *)item->ri_buf[0].i_addr;
2318         } else {
2319                 in_f = (xfs_inode_log_format_t *)kmem_alloc(
2320                         sizeof(xfs_inode_log_format_t), KM_SLEEP);
2321                 need_free = 1;
2322                 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2323                 if (error)
2324                         goto error;
2325         }
2326         ino = in_f->ilf_ino;
2327         mp = log->l_mp;
2328
2329         /*
2330          * Inode buffers can be freed, look out for it,
2331          * and do not replay the inode.
2332          */
2333         if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2334                                         in_f->ilf_len, 0)) {
2335                 error = 0;
2336                 goto error;
2337         }
2338
2339         bp = xfs_buf_read_flags(mp->m_ddev_targp, in_f->ilf_blkno,
2340                                 in_f->ilf_len, XFS_BUF_LOCK);
2341         if (XFS_BUF_ISERROR(bp)) {
2342                 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2343                                   bp, in_f->ilf_blkno);
2344                 error = XFS_BUF_GETERROR(bp);
2345                 xfs_buf_relse(bp);
2346                 goto error;
2347         }
2348         error = 0;
2349         ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2350         dip = (xfs_dinode_t *)xfs_buf_offset(bp, in_f->ilf_boffset);
2351
2352         /*
2353          * Make sure the place we're flushing out to really looks
2354          * like an inode!
2355          */
2356         if (unlikely(be16_to_cpu(dip->di_magic) != XFS_DINODE_MAGIC)) {
2357                 xfs_buf_relse(bp);
2358                 xfs_fs_cmn_err(CE_ALERT, mp,
2359                         "xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2360                         dip, bp, ino);
2361                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(1)",
2362                                  XFS_ERRLEVEL_LOW, mp);
2363                 error = EFSCORRUPTED;
2364                 goto error;
2365         }
2366         dicp = (xfs_icdinode_t *)(item->ri_buf[1].i_addr);
2367         if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2368                 xfs_buf_relse(bp);
2369                 xfs_fs_cmn_err(CE_ALERT, mp,
2370                         "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2371                         item, ino);
2372                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(2)",
2373                                  XFS_ERRLEVEL_LOW, mp);
2374                 error = EFSCORRUPTED;
2375                 goto error;
2376         }
2377
2378         /* Skip replay when the on disk inode is newer than the log one */
2379         if (dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
2380                 /*
2381                  * Deal with the wrap case, DI_MAX_FLUSH is less
2382                  * than smaller numbers
2383                  */
2384                 if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
2385                     dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2386                         /* do nothing */
2387                 } else {
2388                         xfs_buf_relse(bp);
2389                         error = 0;
2390                         goto error;
2391                 }
2392         }
2393         /* Take the opportunity to reset the flush iteration count */
2394         dicp->di_flushiter = 0;
2395
2396         if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2397                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2398                     (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2399                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(3)",
2400                                          XFS_ERRLEVEL_LOW, mp, dicp);
2401                         xfs_buf_relse(bp);
2402                         xfs_fs_cmn_err(CE_ALERT, mp,
2403                                 "xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2404                                 item, dip, bp, ino);
2405                         error = EFSCORRUPTED;
2406                         goto error;
2407                 }
2408         } else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2409                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2410                     (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2411                     (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2412                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(4)",
2413                                              XFS_ERRLEVEL_LOW, mp, dicp);
2414                         xfs_buf_relse(bp);
2415                         xfs_fs_cmn_err(CE_ALERT, mp,
2416                                 "xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2417                                 item, dip, bp, ino);
2418                         error = EFSCORRUPTED;
2419                         goto error;
2420                 }
2421         }
2422         if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2423                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(5)",
2424                                      XFS_ERRLEVEL_LOW, mp, dicp);
2425                 xfs_buf_relse(bp);
2426                 xfs_fs_cmn_err(CE_ALERT, mp,
2427                         "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",
2428                         item, dip, bp, ino,
2429                         dicp->di_nextents + dicp->di_anextents,
2430                         dicp->di_nblocks);
2431                 error = EFSCORRUPTED;
2432                 goto error;
2433         }
2434         if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2435                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(6)",
2436                                      XFS_ERRLEVEL_LOW, mp, dicp);
2437                 xfs_buf_relse(bp);
2438                 xfs_fs_cmn_err(CE_ALERT, mp,
2439                         "xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2440                         item, dip, bp, ino, dicp->di_forkoff);
2441                 error = EFSCORRUPTED;
2442                 goto error;
2443         }
2444         if (unlikely(item->ri_buf[1].i_len > sizeof(struct xfs_icdinode))) {
2445                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(7)",
2446                                      XFS_ERRLEVEL_LOW, mp, dicp);
2447                 xfs_buf_relse(bp);
2448                 xfs_fs_cmn_err(CE_ALERT, mp,
2449                         "xfs_inode_recover: Bad inode log record length %d, rec ptr 0x%p",
2450                         item->ri_buf[1].i_len, item);
2451                 error = EFSCORRUPTED;
2452                 goto error;
2453         }
2454
2455         /* The core is in in-core format */
2456         xfs_dinode_to_disk(dip, (xfs_icdinode_t *)item->ri_buf[1].i_addr);
2457
2458         /* the rest is in on-disk format */
2459         if (item->ri_buf[1].i_len > sizeof(struct xfs_icdinode)) {
2460                 memcpy((xfs_caddr_t) dip + sizeof(struct xfs_icdinode),
2461                         item->ri_buf[1].i_addr + sizeof(struct xfs_icdinode),
2462                         item->ri_buf[1].i_len  - sizeof(struct xfs_icdinode));
2463         }
2464
2465         fields = in_f->ilf_fields;
2466         switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2467         case XFS_ILOG_DEV:
2468                 xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
2469                 break;
2470         case XFS_ILOG_UUID:
2471                 memcpy(XFS_DFORK_DPTR(dip),
2472                        &in_f->ilf_u.ilfu_uuid,
2473                        sizeof(uuid_t));
2474                 break;
2475         }
2476
2477         if (in_f->ilf_size == 2)
2478                 goto write_inode_buffer;
2479         len = item->ri_buf[2].i_len;
2480         src = item->ri_buf[2].i_addr;
2481         ASSERT(in_f->ilf_size <= 4);
2482         ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2483         ASSERT(!(fields & XFS_ILOG_DFORK) ||
2484                (len == in_f->ilf_dsize));
2485
2486         switch (fields & XFS_ILOG_DFORK) {
2487         case XFS_ILOG_DDATA:
2488         case XFS_ILOG_DEXT:
2489                 memcpy(XFS_DFORK_DPTR(dip), src, len);
2490                 break;
2491
2492         case XFS_ILOG_DBROOT:
2493                 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
2494                                  (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
2495                                  XFS_DFORK_DSIZE(dip, mp));
2496                 break;
2497
2498         default:
2499                 /*
2500                  * There are no data fork flags set.
2501                  */
2502                 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2503                 break;
2504         }
2505
2506         /*
2507          * If we logged any attribute data, recover it.  There may or
2508          * may not have been any other non-core data logged in this
2509          * transaction.
2510          */
2511         if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2512                 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2513                         attr_index = 3;
2514                 } else {
2515                         attr_index = 2;
2516                 }
2517                 len = item->ri_buf[attr_index].i_len;
2518                 src = item->ri_buf[attr_index].i_addr;
2519                 ASSERT(len == in_f->ilf_asize);
2520
2521                 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2522                 case XFS_ILOG_ADATA:
2523                 case XFS_ILOG_AEXT:
2524                         dest = XFS_DFORK_APTR(dip);
2525                         ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2526                         memcpy(dest, src, len);
2527                         break;
2528
2529                 case XFS_ILOG_ABROOT:
2530                         dest = XFS_DFORK_APTR(dip);
2531                         xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
2532                                          len, (xfs_bmdr_block_t*)dest,
2533                                          XFS_DFORK_ASIZE(dip, mp));
2534                         break;
2535
2536                 default:
2537                         xlog_warn("XFS: xlog_recover_do_inode_trans: Invalid flag");
2538                         ASSERT(0);
2539                         xfs_buf_relse(bp);
2540                         error = EIO;
2541                         goto error;
2542                 }
2543         }
2544
2545 write_inode_buffer:
2546         ASSERT(bp->b_mount == NULL || bp->b_mount == mp);
2547         bp->b_mount = mp;
2548         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2549         xfs_bdwrite(mp, bp);
2550 error:
2551         if (need_free)
2552                 kmem_free(in_f);
2553         return XFS_ERROR(error);
2554 }
2555
2556 /*
2557  * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2558  * structure, so that we know not to do any dquot item or dquot buffer recovery,
2559  * of that type.
2560  */
2561 STATIC int
2562 xlog_recover_do_quotaoff_trans(
2563         xlog_t                  *log,
2564         xlog_recover_item_t     *item,
2565         int                     pass)
2566 {
2567         xfs_qoff_logformat_t    *qoff_f;
2568
2569         if (pass == XLOG_RECOVER_PASS2) {
2570                 return (0);
2571         }
2572
2573         qoff_f = (xfs_qoff_logformat_t *)item->ri_buf[0].i_addr;
2574         ASSERT(qoff_f);
2575
2576         /*
2577          * The logitem format's flag tells us if this was user quotaoff,
2578          * group/project quotaoff or both.
2579          */
2580         if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2581                 log->l_quotaoffs_flag |= XFS_DQ_USER;
2582         if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2583                 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2584         if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2585                 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2586
2587         return (0);
2588 }
2589
2590 /*
2591  * Recover a dquot record
2592  */
2593 STATIC int
2594 xlog_recover_do_dquot_trans(
2595         xlog_t                  *log,
2596         xlog_recover_item_t     *item,
2597         int                     pass)
2598 {
2599         xfs_mount_t             *mp;
2600         xfs_buf_t               *bp;
2601         struct xfs_disk_dquot   *ddq, *recddq;
2602         int                     error;
2603         xfs_dq_logformat_t      *dq_f;
2604         uint                    type;
2605
2606         if (pass == XLOG_RECOVER_PASS1) {
2607                 return 0;
2608         }
2609         mp = log->l_mp;
2610
2611         /*
2612          * Filesystems are required to send in quota flags at mount time.
2613          */
2614         if (mp->m_qflags == 0)
2615                 return (0);
2616
2617         recddq = (xfs_disk_dquot_t *)item->ri_buf[1].i_addr;
2618         ASSERT(recddq);
2619         /*
2620          * This type of quotas was turned off, so ignore this record.
2621          */
2622         type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2623         ASSERT(type);
2624         if (log->l_quotaoffs_flag & type)
2625                 return (0);
2626
2627         /*
2628          * At this point we know that quota was _not_ turned off.
2629          * Since the mount flags are not indicating to us otherwise, this
2630          * must mean that quota is on, and the dquot needs to be replayed.
2631          * Remember that we may not have fully recovered the superblock yet,
2632          * so we can't do the usual trick of looking at the SB quota bits.
2633          *
2634          * The other possibility, of course, is that the quota subsystem was
2635          * removed since the last mount - ENOSYS.
2636          */
2637         dq_f = (xfs_dq_logformat_t *)item->ri_buf[0].i_addr;
2638         ASSERT(dq_f);
2639         if ((error = xfs_qm_dqcheck(recddq,
2640                            dq_f->qlf_id,
2641                            0, XFS_QMOPT_DOWARN,
2642                            "xlog_recover_do_dquot_trans (log copy)"))) {
2643                 return XFS_ERROR(EIO);
2644         }
2645         ASSERT(dq_f->qlf_len == 1);
2646
2647         error = xfs_read_buf(mp, mp->m_ddev_targp,
2648                              dq_f->qlf_blkno,
2649                              XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2650                              0, &bp);
2651         if (error) {
2652                 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2653                                   bp, dq_f->qlf_blkno);
2654                 return error;
2655         }
2656         ASSERT(bp);
2657         ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2658
2659         /*
2660          * At least the magic num portion should be on disk because this
2661          * was among a chunk of dquots created earlier, and we did some
2662          * minimal initialization then.
2663          */
2664         if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2665                            "xlog_recover_do_dquot_trans")) {
2666                 xfs_buf_relse(bp);
2667                 return XFS_ERROR(EIO);
2668         }
2669
2670         memcpy(ddq, recddq, item->ri_buf[1].i_len);
2671
2672         ASSERT(dq_f->qlf_size == 2);
2673         ASSERT(bp->b_mount == NULL || bp->b_mount == mp);
2674         bp->b_mount = mp;
2675         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2676         xfs_bdwrite(mp, bp);
2677
2678         return (0);
2679 }
2680
2681 /*
2682  * This routine is called to create an in-core extent free intent
2683  * item from the efi format structure which was logged on disk.
2684  * It allocates an in-core efi, copies the extents from the format
2685  * structure into it, and adds the efi to the AIL with the given
2686  * LSN.
2687  */
2688 STATIC int
2689 xlog_recover_do_efi_trans(
2690         xlog_t                  *log,
2691         xlog_recover_item_t     *item,
2692         xfs_lsn_t               lsn,
2693         int                     pass)
2694 {
2695         int                     error;
2696         xfs_mount_t             *mp;
2697         xfs_efi_log_item_t      *efip;
2698         xfs_efi_log_format_t    *efi_formatp;
2699
2700         if (pass == XLOG_RECOVER_PASS1) {
2701                 return 0;
2702         }
2703
2704         efi_formatp = (xfs_efi_log_format_t *)item->ri_buf[0].i_addr;
2705
2706         mp = log->l_mp;
2707         efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2708         if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2709                                          &(efip->efi_format)))) {
2710                 xfs_efi_item_free(efip);
2711                 return error;
2712         }
2713         efip->efi_next_extent = efi_formatp->efi_nextents;
2714         efip->efi_flags |= XFS_EFI_COMMITTED;
2715
2716         spin_lock(&log->l_ailp->xa_lock);
2717         /*
2718          * xfs_trans_ail_update() drops the AIL lock.
2719          */
2720         xfs_trans_ail_update(log->l_ailp, (xfs_log_item_t *)efip, lsn);
2721         return 0;
2722 }
2723
2724
2725 /*
2726  * This routine is called when an efd format structure is found in
2727  * a committed transaction in the log.  It's purpose is to cancel
2728  * the corresponding efi if it was still in the log.  To do this
2729  * it searches the AIL for the efi with an id equal to that in the
2730  * efd format structure.  If we find it, we remove the efi from the
2731  * AIL and free it.
2732  */
2733 STATIC void
2734 xlog_recover_do_efd_trans(
2735         xlog_t                  *log,
2736         xlog_recover_item_t     *item,
2737         int                     pass)
2738 {
2739         xfs_efd_log_format_t    *efd_formatp;
2740         xfs_efi_log_item_t      *efip = NULL;
2741         xfs_log_item_t          *lip;
2742         __uint64_t              efi_id;
2743         struct xfs_ail_cursor   cur;
2744         struct xfs_ail          *ailp = log->l_ailp;
2745
2746         if (pass == XLOG_RECOVER_PASS1) {
2747                 return;
2748         }
2749
2750         efd_formatp = (xfs_efd_log_format_t *)item->ri_buf[0].i_addr;
2751         ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2752                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2753                (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2754                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2755         efi_id = efd_formatp->efd_efi_id;
2756
2757         /*
2758          * Search for the efi with the id in the efd format structure
2759          * in the AIL.
2760          */
2761         spin_lock(&ailp->xa_lock);
2762         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2763         while (lip != NULL) {
2764                 if (lip->li_type == XFS_LI_EFI) {
2765                         efip = (xfs_efi_log_item_t *)lip;
2766                         if (efip->efi_format.efi_id == efi_id) {
2767                                 /*
2768                                  * xfs_trans_ail_delete() drops the
2769                                  * AIL lock.
2770                                  */
2771                                 xfs_trans_ail_delete(ailp, lip);
2772                                 xfs_efi_item_free(efip);
2773                                 spin_lock(&ailp->xa_lock);
2774                                 break;
2775                         }
2776                 }
2777                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2778         }
2779         xfs_trans_ail_cursor_done(ailp, &cur);
2780         spin_unlock(&ailp->xa_lock);
2781 }
2782
2783 /*
2784  * Perform the transaction
2785  *
2786  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2787  * EFIs and EFDs get queued up by adding entries into the AIL for them.
2788  */
2789 STATIC int
2790 xlog_recover_do_trans(
2791         xlog_t                  *log,
2792         xlog_recover_t          *trans,
2793         int                     pass)
2794 {
2795         int                     error = 0;
2796         xlog_recover_item_t     *item, *first_item;
2797
2798         error = xlog_recover_reorder_trans(trans);
2799         if (error)
2800                 return error;
2801
2802         first_item = item = trans->r_itemq;
2803         do {
2804                 switch (ITEM_TYPE(item)) {
2805                 case XFS_LI_BUF:
2806                         error = xlog_recover_do_buffer_trans(log, item, pass);
2807                         break;
2808                 case XFS_LI_INODE:
2809                         error = xlog_recover_do_inode_trans(log, item, pass);
2810                         break;
2811                 case XFS_LI_EFI:
2812                         error = xlog_recover_do_efi_trans(log, item,
2813                                                           trans->r_lsn, pass);
2814                         break;
2815                 case XFS_LI_EFD:
2816                         xlog_recover_do_efd_trans(log, item, pass);
2817                         error = 0;
2818                         break;
2819                 case XFS_LI_DQUOT:
2820                         error = xlog_recover_do_dquot_trans(log, item, pass);
2821                         break;
2822                 case XFS_LI_QUOTAOFF:
2823                         error = xlog_recover_do_quotaoff_trans(log, item,
2824                                                                pass);
2825                         break;
2826                 default:
2827                         xlog_warn(
2828         "XFS: invalid item type (%d) xlog_recover_do_trans", ITEM_TYPE(item));
2829                         ASSERT(0);
2830                         error = XFS_ERROR(EIO);
2831                         break;
2832                 }
2833
2834                 if (error)
2835                         return error;
2836                 item = item->ri_next;
2837         } while (first_item != item);
2838
2839         return 0;
2840 }
2841
2842 /*
2843  * Free up any resources allocated by the transaction
2844  *
2845  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2846  */
2847 STATIC void
2848 xlog_recover_free_trans(
2849         xlog_recover_t          *trans)
2850 {
2851         xlog_recover_item_t     *first_item, *item, *free_item;
2852         int                     i;
2853
2854         item = first_item = trans->r_itemq;
2855         do {
2856                 free_item = item;
2857                 item = item->ri_next;
2858                  /* Free the regions in the item. */
2859                 for (i = 0; i < free_item->ri_cnt; i++) {
2860                         kmem_free(free_item->ri_buf[i].i_addr);
2861                 }
2862                 /* Free the item itself */
2863                 kmem_free(free_item->ri_buf);
2864                 kmem_free(free_item);
2865         } while (first_item != item);
2866         /* Free the transaction recover structure */
2867         kmem_free(trans);
2868 }
2869
2870 STATIC int
2871 xlog_recover_commit_trans(
2872         xlog_t                  *log,
2873         xlog_recover_t          **q,
2874         xlog_recover_t          *trans,
2875         int                     pass)
2876 {
2877         int                     error;
2878
2879         if ((error = xlog_recover_unlink_tid(q, trans)))
2880                 return error;
2881         if ((error = xlog_recover_do_trans(log, trans, pass)))
2882                 return error;
2883         xlog_recover_free_trans(trans);                 /* no error */
2884         return 0;
2885 }
2886
2887 STATIC int
2888 xlog_recover_unmount_trans(
2889         xlog_recover_t          *trans)
2890 {
2891         /* Do nothing now */
2892         xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2893         return 0;
2894 }
2895
2896 /*
2897  * There are two valid states of the r_state field.  0 indicates that the
2898  * transaction structure is in a normal state.  We have either seen the
2899  * start of the transaction or the last operation we added was not a partial
2900  * operation.  If the last operation we added to the transaction was a
2901  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2902  *
2903  * NOTE: skip LRs with 0 data length.
2904  */
2905 STATIC int
2906 xlog_recover_process_data(
2907         xlog_t                  *log,
2908         xlog_recover_t          *rhash[],
2909         xlog_rec_header_t       *rhead,
2910         xfs_caddr_t             dp,
2911         int                     pass)
2912 {
2913         xfs_caddr_t             lp;
2914         int                     num_logops;
2915         xlog_op_header_t        *ohead;
2916         xlog_recover_t          *trans;
2917         xlog_tid_t              tid;
2918         int                     error;
2919         unsigned long           hash;
2920         uint                    flags;
2921
2922         lp = dp + be32_to_cpu(rhead->h_len);
2923         num_logops = be32_to_cpu(rhead->h_num_logops);
2924
2925         /* check the log format matches our own - else we can't recover */
2926         if (xlog_header_check_recover(log->l_mp, rhead))
2927                 return (XFS_ERROR(EIO));
2928
2929         while ((dp < lp) && num_logops) {
2930                 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2931                 ohead = (xlog_op_header_t *)dp;
2932                 dp += sizeof(xlog_op_header_t);
2933                 if (ohead->oh_clientid != XFS_TRANSACTION &&
2934                     ohead->oh_clientid != XFS_LOG) {
2935                         xlog_warn(
2936                 "XFS: xlog_recover_process_data: bad clientid");
2937                         ASSERT(0);
2938                         return (XFS_ERROR(EIO));
2939                 }
2940                 tid = be32_to_cpu(ohead->oh_tid);
2941                 hash = XLOG_RHASH(tid);
2942                 trans = xlog_recover_find_tid(rhash[hash], tid);
2943                 if (trans == NULL) {               /* not found; add new tid */
2944                         if (ohead->oh_flags & XLOG_START_TRANS)
2945                                 xlog_recover_new_tid(&rhash[hash], tid,
2946                                         be64_to_cpu(rhead->h_lsn));
2947                 } else {
2948                         if (dp + be32_to_cpu(ohead->oh_len) > lp) {
2949                                 xlog_warn(
2950                         "XFS: xlog_recover_process_data: bad length");
2951                                 WARN_ON(1);
2952                                 return (XFS_ERROR(EIO));
2953                         }
2954                         flags = ohead->oh_flags & ~XLOG_END_TRANS;
2955                         if (flags & XLOG_WAS_CONT_TRANS)
2956                                 flags &= ~XLOG_CONTINUE_TRANS;
2957                         switch (flags) {
2958                         case XLOG_COMMIT_TRANS:
2959                                 error = xlog_recover_commit_trans(log,
2960                                                 &rhash[hash], trans, pass);
2961                                 break;
2962                         case XLOG_UNMOUNT_TRANS:
2963                                 error = xlog_recover_unmount_trans(trans);
2964                                 break;
2965                         case XLOG_WAS_CONT_TRANS:
2966                                 error = xlog_recover_add_to_cont_trans(trans,
2967                                                 dp, be32_to_cpu(ohead->oh_len));
2968                                 break;
2969                         case XLOG_START_TRANS:
2970                                 xlog_warn(
2971                         "XFS: xlog_recover_process_data: bad transaction");
2972                                 ASSERT(0);
2973                                 error = XFS_ERROR(EIO);
2974                                 break;
2975                         case 0:
2976                         case XLOG_CONTINUE_TRANS:
2977                                 error = xlog_recover_add_to_trans(trans,
2978                                                 dp, be32_to_cpu(ohead->oh_len));
2979                                 break;
2980                         default:
2981                                 xlog_warn(
2982                         "XFS: xlog_recover_process_data: bad flag");
2983                                 ASSERT(0);
2984                                 error = XFS_ERROR(EIO);
2985                                 break;
2986                         }
2987                         if (error)
2988                                 return error;
2989                 }
2990                 dp += be32_to_cpu(ohead->oh_len);
2991                 num_logops--;
2992         }
2993         return 0;
2994 }
2995
2996 /*
2997  * Process an extent free intent item that was recovered from
2998  * the log.  We need to free the extents that it describes.
2999  */
3000 STATIC int
3001 xlog_recover_process_efi(
3002         xfs_mount_t             *mp,
3003         xfs_efi_log_item_t      *efip)
3004 {
3005         xfs_efd_log_item_t      *efdp;
3006         xfs_trans_t             *tp;
3007         int                     i;
3008         int                     error = 0;
3009         xfs_extent_t            *extp;
3010         xfs_fsblock_t           startblock_fsb;
3011
3012         ASSERT(!(efip->efi_flags & XFS_EFI_RECOVERED));
3013
3014         /*
3015          * First check the validity of the extents described by the
3016          * EFI.  If any are bad, then assume that all are bad and
3017          * just toss the EFI.
3018          */
3019         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3020                 extp = &(efip->efi_format.efi_extents[i]);
3021                 startblock_fsb = XFS_BB_TO_FSB(mp,
3022                                    XFS_FSB_TO_DADDR(mp, extp->ext_start));
3023                 if ((startblock_fsb == 0) ||
3024                     (extp->ext_len == 0) ||
3025                     (startblock_fsb >= mp->m_sb.sb_dblocks) ||
3026                     (extp->ext_len >= mp->m_sb.sb_agblocks)) {
3027                         /*
3028                          * This will pull the EFI from the AIL and
3029                          * free the memory associated with it.
3030                          */
3031                         xfs_efi_release(efip, efip->efi_format.efi_nextents);
3032                         return XFS_ERROR(EIO);
3033                 }
3034         }
3035
3036         tp = xfs_trans_alloc(mp, 0);
3037         error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
3038         if (error)
3039                 goto abort_error;
3040         efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3041
3042         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3043                 extp = &(efip->efi_format.efi_extents[i]);
3044                 error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3045                 if (error)
3046                         goto abort_error;
3047                 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3048                                          extp->ext_len);
3049         }
3050
3051         efip->efi_flags |= XFS_EFI_RECOVERED;
3052         error = xfs_trans_commit(tp, 0);
3053         return error;
3054
3055 abort_error:
3056         xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3057         return error;
3058 }
3059
3060 /*
3061  * When this is called, all of the EFIs which did not have
3062  * corresponding EFDs should be in the AIL.  What we do now
3063  * is free the extents associated with each one.
3064  *
3065  * Since we process the EFIs in normal transactions, they
3066  * will be removed at some point after the commit.  This prevents
3067  * us from just walking down the list processing each one.
3068  * We'll use a flag in the EFI to skip those that we've already
3069  * processed and use the AIL iteration mechanism's generation
3070  * count to try to speed this up at least a bit.
3071  *
3072  * When we start, we know that the EFIs are the only things in
3073  * the AIL.  As we process them, however, other items are added
3074  * to the AIL.  Since everything added to the AIL must come after
3075  * everything already in the AIL, we stop processing as soon as
3076  * we see something other than an EFI in the AIL.
3077  */
3078 STATIC int
3079 xlog_recover_process_efis(
3080         xlog_t                  *log)
3081 {
3082         xfs_log_item_t          *lip;
3083         xfs_efi_log_item_t      *efip;
3084         int                     error = 0;
3085         struct xfs_ail_cursor   cur;
3086         struct xfs_ail          *ailp;
3087
3088         ailp = log->l_ailp;
3089         spin_lock(&ailp->xa_lock);
3090         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3091         while (lip != NULL) {
3092                 /*
3093                  * We're done when we see something other than an EFI.
3094                  * There should be no EFIs left in the AIL now.
3095                  */
3096                 if (lip->li_type != XFS_LI_EFI) {
3097 #ifdef DEBUG
3098                         for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
3099                                 ASSERT(lip->li_type != XFS_LI_EFI);
3100 #endif
3101                         break;
3102                 }
3103
3104                 /*
3105                  * Skip EFIs that we've already processed.
3106                  */
3107                 efip = (xfs_efi_log_item_t *)lip;
3108                 if (efip->efi_flags & XFS_EFI_RECOVERED) {
3109                         lip = xfs_trans_ail_cursor_next(ailp, &cur);
3110                         continue;
3111                 }
3112
3113                 spin_unlock(&ailp->xa_lock);
3114                 error = xlog_recover_process_efi(log->l_mp, efip);
3115                 spin_lock(&ailp->xa_lock);
3116                 if (error)
3117                         goto out;
3118                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3119         }
3120 out:
3121         xfs_trans_ail_cursor_done(ailp, &cur);
3122         spin_unlock(&ailp->xa_lock);
3123         return error;
3124 }
3125
3126 /*
3127  * This routine performs a transaction to null out a bad inode pointer
3128  * in an agi unlinked inode hash bucket.
3129  */
3130 STATIC void
3131 xlog_recover_clear_agi_bucket(
3132         xfs_mount_t     *mp,
3133         xfs_agnumber_t  agno,
3134         int             bucket)
3135 {
3136         xfs_trans_t     *tp;
3137         xfs_agi_t       *agi;
3138         xfs_buf_t       *agibp;
3139         int             offset;
3140         int             error;
3141
3142         tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3143         error = xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp),
3144                                   0, 0, 0);
3145         if (error)
3146                 goto out_abort;
3147
3148         error = xfs_read_agi(mp, tp, agno, &agibp);
3149         if (error)
3150                 goto out_abort;
3151
3152         agi = XFS_BUF_TO_AGI(agibp);
3153         agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3154         offset = offsetof(xfs_agi_t, agi_unlinked) +
3155                  (sizeof(xfs_agino_t) * bucket);
3156         xfs_trans_log_buf(tp, agibp, offset,
3157                           (offset + sizeof(xfs_agino_t) - 1));
3158
3159         error = xfs_trans_commit(tp, 0);
3160         if (error)
3161                 goto out_error;
3162         return;
3163
3164 out_abort:
3165         xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3166 out_error:
3167         xfs_fs_cmn_err(CE_WARN, mp, "xlog_recover_clear_agi_bucket: "
3168                         "failed to clear agi %d. Continuing.", agno);
3169         return;
3170 }
3171
3172 STATIC xfs_agino_t
3173 xlog_recover_process_one_iunlink(
3174         struct xfs_mount                *mp,
3175         xfs_agnumber_t                  agno,
3176         xfs_agino_t                     agino,
3177         int                             bucket)
3178 {
3179         struct xfs_buf                  *ibp;
3180         struct xfs_dinode               *dip;
3181         struct xfs_inode                *ip;
3182         xfs_ino_t                       ino;
3183         int                             error;
3184
3185         ino = XFS_AGINO_TO_INO(mp, agno, agino);
3186         error = xfs_iget(mp, NULL, ino, 0, 0, &ip, 0);
3187         if (error)
3188                 goto fail;
3189
3190         /*
3191          * Get the on disk inode to find the next inode in the bucket.
3192          */
3193         error = xfs_itobp(mp, NULL, ip, &dip, &ibp, XFS_BUF_LOCK);
3194         if (error)
3195                 goto fail_iput;
3196
3197         ASSERT(ip->i_d.di_nlink == 0);
3198         ASSERT(ip->i_d.di_mode != 0);
3199
3200         /* setup for the next pass */
3201         agino = be32_to_cpu(dip->di_next_unlinked);
3202         xfs_buf_relse(ibp);
3203
3204         /*
3205          * Prevent any DMAPI event from being sent when the reference on
3206          * the inode is dropped.
3207          */
3208         ip->i_d.di_dmevmask = 0;
3209
3210         IRELE(ip);
3211         return agino;
3212
3213  fail_iput:
3214         IRELE(ip);
3215  fail:
3216         /*
3217          * We can't read in the inode this bucket points to, or this inode
3218          * is messed up.  Just ditch this bucket of inodes.  We will lose
3219          * some inodes and space, but at least we won't hang.
3220          *
3221          * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3222          * clear the inode pointer in the bucket.
3223          */
3224         xlog_recover_clear_agi_bucket(mp, agno, bucket);
3225         return NULLAGINO;
3226 }
3227
3228 /*
3229  * xlog_iunlink_recover
3230  *
3231  * This is called during recovery to process any inodes which
3232  * we unlinked but not freed when the system crashed.  These
3233  * inodes will be on the lists in the AGI blocks.  What we do
3234  * here is scan all the AGIs and fully truncate and free any
3235  * inodes found on the lists.  Each inode is removed from the
3236  * lists when it has been fully truncated and is freed.  The
3237  * freeing of the inode and its removal from the list must be
3238  * atomic.
3239  */
3240 void
3241 xlog_recover_process_iunlinks(
3242         xlog_t          *log)
3243 {
3244         xfs_mount_t     *mp;
3245         xfs_agnumber_t  agno;
3246         xfs_agi_t       *agi;
3247         xfs_buf_t       *agibp;
3248         xfs_agino_t     agino;
3249         int             bucket;
3250         int             error;
3251         uint            mp_dmevmask;
3252
3253         mp = log->l_mp;
3254
3255         /*
3256          * Prevent any DMAPI event from being sent while in this function.
3257          */
3258         mp_dmevmask = mp->m_dmevmask;
3259         mp->m_dmevmask = 0;
3260
3261         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3262                 /*
3263                  * Find the agi for this ag.
3264                  */
3265                 error = xfs_read_agi(mp, NULL, agno, &agibp);
3266                 if (error) {
3267                         /*
3268                          * AGI is b0rked. Don't process it.
3269                          *
3270                          * We should probably mark the filesystem as corrupt
3271                          * after we've recovered all the ag's we can....
3272                          */
3273                         continue;
3274                 }
3275                 agi = XFS_BUF_TO_AGI(agibp);
3276
3277                 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3278                         agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3279                         while (agino != NULLAGINO) {
3280                                 /*
3281                                  * Release the agi buffer so that it can
3282                                  * be acquired in the normal course of the
3283                                  * transaction to truncate and free the inode.
3284                                  */
3285                                 xfs_buf_relse(agibp);
3286
3287                                 agino = xlog_recover_process_one_iunlink(mp,
3288                                                         agno, agino, bucket);
3289
3290                                 /*
3291                                  * Reacquire the agibuffer and continue around
3292                                  * the loop. This should never fail as we know
3293                                  * the buffer was good earlier on.
3294                                  */
3295                                 error = xfs_read_agi(mp, NULL, agno, &agibp);
3296                                 ASSERT(error == 0);
3297                                 agi = XFS_BUF_TO_AGI(agibp);
3298                         }
3299                 }
3300
3301                 /*
3302                  * Release the buffer for the current agi so we can
3303                  * go on to the next one.
3304                  */
3305                 xfs_buf_relse(agibp);
3306         }
3307
3308         mp->m_dmevmask = mp_dmevmask;
3309 }
3310
3311
3312 #ifdef DEBUG
3313 STATIC void
3314 xlog_pack_data_checksum(
3315         xlog_t          *log,
3316         xlog_in_core_t  *iclog,
3317         int             size)
3318 {
3319         int             i;
3320         __be32          *up;
3321         uint            chksum = 0;
3322
3323         up = (__be32 *)iclog->ic_datap;
3324         /* divide length by 4 to get # words */
3325         for (i = 0; i < (size >> 2); i++) {
3326                 chksum ^= be32_to_cpu(*up);
3327                 up++;
3328         }
3329         iclog->ic_header.h_chksum = cpu_to_be32(chksum);
3330 }
3331 #else
3332 #define xlog_pack_data_checksum(log, iclog, size)
3333 #endif
3334
3335 /*
3336  * Stamp cycle number in every block
3337  */
3338 void
3339 xlog_pack_data(
3340         xlog_t                  *log,
3341         xlog_in_core_t          *iclog,
3342         int                     roundoff)
3343 {
3344         int                     i, j, k;
3345         int                     size = iclog->ic_offset + roundoff;
3346         __be32                  cycle_lsn;
3347         xfs_caddr_t             dp;
3348
3349         xlog_pack_data_checksum(log, iclog, size);
3350
3351         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3352
3353         dp = iclog->ic_datap;
3354         for (i = 0; i < BTOBB(size) &&
3355                 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3356                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
3357                 *(__be32 *)dp = cycle_lsn;
3358                 dp += BBSIZE;
3359         }
3360
3361         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3362                 xlog_in_core_2_t *xhdr = iclog->ic_data;
3363
3364                 for ( ; i < BTOBB(size); i++) {
3365                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3366                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3367                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
3368                         *(__be32 *)dp = cycle_lsn;
3369                         dp += BBSIZE;
3370                 }
3371
3372                 for (i = 1; i < log->l_iclog_heads; i++) {
3373                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3374                 }
3375         }
3376 }
3377
3378 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
3379 STATIC void
3380 xlog_unpack_data_checksum(
3381         xlog_rec_header_t       *rhead,
3382         xfs_caddr_t             dp,
3383         xlog_t                  *log)
3384 {
3385         __be32                  *up = (__be32 *)dp;
3386         uint                    chksum = 0;
3387         int                     i;
3388
3389         /* divide length by 4 to get # words */
3390         for (i=0; i < be32_to_cpu(rhead->h_len) >> 2; i++) {
3391                 chksum ^= be32_to_cpu(*up);
3392                 up++;
3393         }
3394         if (chksum != be32_to_cpu(rhead->h_chksum)) {
3395             if (rhead->h_chksum ||
3396                 ((log->l_flags & XLOG_CHKSUM_MISMATCH) == 0)) {
3397                     cmn_err(CE_DEBUG,
3398                         "XFS: LogR chksum mismatch: was (0x%x) is (0x%x)\n",
3399                             be32_to_cpu(rhead->h_chksum), chksum);
3400                     cmn_err(CE_DEBUG,
3401 "XFS: Disregard message if filesystem was created with non-DEBUG kernel");
3402                     if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3403                             cmn_err(CE_DEBUG,
3404                                 "XFS: LogR this is a LogV2 filesystem\n");
3405                     }
3406                     log->l_flags |= XLOG_CHKSUM_MISMATCH;
3407             }
3408         }
3409 }
3410 #else
3411 #define xlog_unpack_data_checksum(rhead, dp, log)
3412 #endif
3413
3414 STATIC void
3415 xlog_unpack_data(
3416         xlog_rec_header_t       *rhead,
3417         xfs_caddr_t             dp,
3418         xlog_t                  *log)
3419 {
3420         int                     i, j, k;
3421
3422         for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3423                   i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3424                 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3425                 dp += BBSIZE;
3426         }
3427
3428         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3429                 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
3430                 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3431                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3432                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3433                         *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3434                         dp += BBSIZE;
3435                 }
3436         }
3437
3438         xlog_unpack_data_checksum(rhead, dp, log);
3439 }
3440
3441 STATIC int
3442 xlog_valid_rec_header(
3443         xlog_t                  *log,
3444         xlog_rec_header_t       *rhead,
3445         xfs_daddr_t             blkno)
3446 {
3447         int                     hlen;
3448
3449         if (unlikely(be32_to_cpu(rhead->h_magicno) != XLOG_HEADER_MAGIC_NUM)) {
3450                 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3451                                 XFS_ERRLEVEL_LOW, log->l_mp);
3452                 return XFS_ERROR(EFSCORRUPTED);
3453         }
3454         if (unlikely(
3455             (!rhead->h_version ||
3456             (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
3457                 xlog_warn("XFS: %s: unrecognised log version (%d).",
3458                         __func__, be32_to_cpu(rhead->h_version));
3459                 return XFS_ERROR(EIO);
3460         }
3461
3462         /* LR body must have data or it wouldn't have been written */
3463         hlen = be32_to_cpu(rhead->h_len);
3464         if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3465                 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3466                                 XFS_ERRLEVEL_LOW, log->l_mp);
3467                 return XFS_ERROR(EFSCORRUPTED);
3468         }
3469         if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3470                 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3471                                 XFS_ERRLEVEL_LOW, log->l_mp);
3472                 return XFS_ERROR(EFSCORRUPTED);
3473         }
3474         return 0;
3475 }
3476
3477 /*
3478  * Read the log from tail to head and process the log records found.
3479  * Handle the two cases where the tail and head are in the same cycle
3480  * and where the active portion of the log wraps around the end of
3481  * the physical log separately.  The pass parameter is passed through
3482  * to the routines called to process the data and is not looked at
3483  * here.
3484  */
3485 STATIC int
3486 xlog_do_recovery_pass(
3487         xlog_t                  *log,
3488         xfs_daddr_t             head_blk,
3489         xfs_daddr_t             tail_blk,
3490         int                     pass)
3491 {
3492         xlog_rec_header_t       *rhead;
3493         xfs_daddr_t             blk_no;
3494         xfs_caddr_t             bufaddr, offset;
3495         xfs_buf_t               *hbp, *dbp;
3496         int                     error = 0, h_size;
3497         int                     bblks, split_bblks;
3498         int                     hblks, split_hblks, wrapped_hblks;
3499         xlog_recover_t          *rhash[XLOG_RHASH_SIZE];
3500
3501         ASSERT(head_blk != tail_blk);
3502
3503         /*
3504          * Read the header of the tail block and get the iclog buffer size from
3505          * h_size.  Use this to tell how many sectors make up the log header.
3506          */
3507         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3508                 /*
3509                  * When using variable length iclogs, read first sector of
3510                  * iclog header and extract the header size from it.  Get a
3511                  * new hbp that is the correct size.
3512                  */
3513                 hbp = xlog_get_bp(log, 1);
3514                 if (!hbp)
3515                         return ENOMEM;
3516
3517                 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
3518                 if (error)
3519                         goto bread_err1;
3520
3521                 rhead = (xlog_rec_header_t *)offset;
3522                 error = xlog_valid_rec_header(log, rhead, tail_blk);
3523                 if (error)
3524                         goto bread_err1;
3525                 h_size = be32_to_cpu(rhead->h_size);
3526                 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3527                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3528                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3529                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
3530                                 hblks++;
3531                         xlog_put_bp(hbp);
3532                         hbp = xlog_get_bp(log, hblks);
3533                 } else {
3534                         hblks = 1;
3535                 }
3536         } else {
3537                 ASSERT(log->l_sectbb_log == 0);
3538                 hblks = 1;
3539                 hbp = xlog_get_bp(log, 1);
3540                 h_size = XLOG_BIG_RECORD_BSIZE;
3541         }
3542
3543         if (!hbp)
3544                 return ENOMEM;
3545         dbp = xlog_get_bp(log, BTOBB(h_size));
3546         if (!dbp) {
3547                 xlog_put_bp(hbp);
3548                 return ENOMEM;
3549         }
3550
3551         memset(rhash, 0, sizeof(rhash));
3552         if (tail_blk <= head_blk) {
3553                 for (blk_no = tail_blk; blk_no < head_blk; ) {
3554                         error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3555                         if (error)
3556                                 goto bread_err2;
3557
3558                         rhead = (xlog_rec_header_t *)offset;
3559                         error = xlog_valid_rec_header(log, rhead, blk_no);
3560                         if (error)
3561                                 goto bread_err2;
3562
3563                         /* blocks in data section */
3564                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3565                         error = xlog_bread(log, blk_no + hblks, bblks, dbp,
3566                                            &offset);
3567                         if (error)
3568                                 goto bread_err2;
3569
3570                         xlog_unpack_data(rhead, offset, log);
3571                         if ((error = xlog_recover_process_data(log,
3572                                                 rhash, rhead, offset, pass)))
3573                                 goto bread_err2;
3574                         blk_no += bblks + hblks;
3575                 }
3576         } else {
3577                 /*
3578                  * Perform recovery around the end of the physical log.
3579                  * When the head is not on the same cycle number as the tail,
3580                  * we can't do a sequential recovery as above.
3581                  */
3582                 blk_no = tail_blk;
3583                 while (blk_no < log->l_logBBsize) {
3584                         /*
3585                          * Check for header wrapping around physical end-of-log
3586                          */
3587                         offset = NULL;
3588                         split_hblks = 0;
3589                         wrapped_hblks = 0;
3590                         if (blk_no + hblks <= log->l_logBBsize) {
3591                                 /* Read header in one read */
3592                                 error = xlog_bread(log, blk_no, hblks, hbp,
3593                                                    &offset);
3594                                 if (error)
3595                                         goto bread_err2;
3596                         } else {
3597                                 /* This LR is split across physical log end */
3598                                 if (blk_no != log->l_logBBsize) {
3599                                         /* some data before physical log end */
3600                                         ASSERT(blk_no <= INT_MAX);
3601                                         split_hblks = log->l_logBBsize - (int)blk_no;
3602                                         ASSERT(split_hblks > 0);
3603                                         error = xlog_bread(log, blk_no,
3604                                                            split_hblks, hbp,
3605                                                            &offset);
3606                                         if (error)
3607                                                 goto bread_err2;
3608                                 }
3609
3610                                 /*
3611                                  * Note: this black magic still works with
3612                                  * large sector sizes (non-512) only because:
3613                                  * - we increased the buffer size originally
3614                                  *   by 1 sector giving us enough extra space
3615                                  *   for the second read;
3616                                  * - the log start is guaranteed to be sector
3617                                  *   aligned;
3618                                  * - we read the log end (LR header start)
3619                                  *   _first_, then the log start (LR header end)
3620                                  *   - order is important.
3621                                  */
3622                                 wrapped_hblks = hblks - split_hblks;
3623                                 bufaddr = XFS_BUF_PTR(hbp);
3624                                 error = XFS_BUF_SET_PTR(hbp,
3625                                                 bufaddr + BBTOB(split_hblks),
3626                                                 BBTOB(hblks - split_hblks));
3627                                 if (error)
3628                                         goto bread_err2;
3629
3630                                 error = xlog_bread_noalign(log, 0,
3631                                                            wrapped_hblks, hbp);
3632                                 if (error)
3633                                         goto bread_err2;
3634
3635                                 error = XFS_BUF_SET_PTR(hbp, bufaddr,
3636                                                         BBTOB(hblks));
3637                                 if (error)
3638                                         goto bread_err2;
3639
3640                                 if (!offset)
3641                                         offset = xlog_align(log, 0,
3642                                                         wrapped_hblks, hbp);
3643                         }
3644                         rhead = (xlog_rec_header_t *)offset;
3645                         error = xlog_valid_rec_header(log, rhead,
3646                                                 split_hblks ? blk_no : 0);
3647                         if (error)
3648                                 goto bread_err2;
3649
3650                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3651                         blk_no += hblks;
3652
3653                         /* Read in data for log record */
3654                         if (blk_no + bblks <= log->l_logBBsize) {
3655                                 error = xlog_bread(log, blk_no, bblks, dbp,
3656                                                    &offset);
3657                                 if (error)
3658                                         goto bread_err2;
3659                         } else {
3660                                 /* This log record is split across the
3661                                  * physical end of log */
3662                                 offset = NULL;
3663                                 split_bblks = 0;
3664                                 if (blk_no != log->l_logBBsize) {
3665                                         /* some data is before the physical
3666                                          * end of log */
3667                                         ASSERT(!wrapped_hblks);
3668                                         ASSERT(blk_no <= INT_MAX);
3669                                         split_bblks =
3670                                                 log->l_logBBsize - (int)blk_no;
3671                                         ASSERT(split_bblks > 0);
3672                                         error = xlog_bread(log, blk_no,
3673                                                         split_bblks, dbp,
3674                                                         &offset);
3675                                         if (error)
3676                                                 goto bread_err2;
3677                                 }
3678
3679                                 /*
3680                                  * Note: this black magic still works with
3681                                  * large sector sizes (non-512) only because:
3682                                  * - we increased the buffer size originally
3683                                  *   by 1 sector giving us enough extra space
3684                                  *   for the second read;
3685                                  * - the log start is guaranteed to be sector
3686                                  *   aligned;
3687                                  * - we read the log end (LR header start)
3688                                  *   _first_, then the log start (LR header end)
3689                                  *   - order is important.
3690                                  */
3691                                 bufaddr = XFS_BUF_PTR(dbp);
3692                                 error = XFS_BUF_SET_PTR(dbp,
3693                                                 bufaddr + BBTOB(split_bblks),
3694                                                 BBTOB(bblks - split_bblks));
3695                                 if (error)
3696                                         goto bread_err2;
3697
3698                                 error = xlog_bread_noalign(log, wrapped_hblks,
3699                                                 bblks - split_bblks,
3700                                                 dbp);
3701                                 if (error)
3702                                         goto bread_err2;
3703
3704                                 error = XFS_BUF_SET_PTR(dbp, bufaddr, h_size);
3705                                 if (error)
3706                                         goto bread_err2;
3707
3708                                 if (!offset)
3709                                         offset = xlog_align(log, wrapped_hblks,
3710                                                 bblks - split_bblks, dbp);
3711                         }
3712                         xlog_unpack_data(rhead, offset, log);
3713                         if ((error = xlog_recover_process_data(log, rhash,
3714                                                         rhead, offset, pass)))
3715                                 goto bread_err2;
3716                         blk_no += bblks;
3717                 }
3718
3719                 ASSERT(blk_no >= log->l_logBBsize);
3720                 blk_no -= log->l_logBBsize;
3721
3722                 /* read first part of physical log */
3723                 while (blk_no < head_blk) {
3724                         error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3725                         if (error)
3726                                 goto bread_err2;
3727
3728                         rhead = (xlog_rec_header_t *)offset;
3729                         error = xlog_valid_rec_header(log, rhead, blk_no);
3730                         if (error)
3731                                 goto bread_err2;
3732
3733                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3734                         error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3735                                            &offset);
3736                         if (error)
3737                                 goto bread_err2;
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 + hblks;
3744                 }
3745         }
3746
3747  bread_err2:
3748         xlog_put_bp(dbp);
3749  bread_err1:
3750         xlog_put_bp(hbp);
3751         return error;
3752 }
3753
3754 /*
3755  * Do the recovery of the log.  We actually do this in two phases.
3756  * The two passes are necessary in order to implement the function
3757  * of cancelling a record written into the log.  The first pass
3758  * determines those things which have been cancelled, and the
3759  * second pass replays log items normally except for those which
3760  * have been cancelled.  The handling of the replay and cancellations
3761  * takes place in the log item type specific routines.
3762  *
3763  * The table of items which have cancel records in the log is allocated
3764  * and freed at this level, since only here do we know when all of
3765  * the log recovery has been completed.
3766  */
3767 STATIC int
3768 xlog_do_log_recovery(
3769         xlog_t          *log,
3770         xfs_daddr_t     head_blk,
3771         xfs_daddr_t     tail_blk)
3772 {
3773         int             error;
3774
3775         ASSERT(head_blk != tail_blk);
3776
3777         /*
3778          * First do a pass to find all of the cancelled buf log items.
3779          * Store them in the buf_cancel_table for use in the second pass.
3780          */
3781         log->l_buf_cancel_table =
3782                 (xfs_buf_cancel_t **)kmem_zalloc(XLOG_BC_TABLE_SIZE *
3783                                                  sizeof(xfs_buf_cancel_t*),
3784                                                  KM_SLEEP);
3785         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3786                                       XLOG_RECOVER_PASS1);
3787         if (error != 0) {
3788                 kmem_free(log->l_buf_cancel_table);
3789                 log->l_buf_cancel_table = NULL;
3790                 return error;
3791         }
3792         /*
3793          * Then do a second pass to actually recover the items in the log.
3794          * When it is complete free the table of buf cancel items.
3795          */
3796         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3797                                       XLOG_RECOVER_PASS2);
3798 #ifdef DEBUG
3799         if (!error) {
3800                 int     i;
3801
3802                 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3803                         ASSERT(log->l_buf_cancel_table[i] == NULL);
3804         }
3805 #endif  /* DEBUG */
3806
3807         kmem_free(log->l_buf_cancel_table);
3808         log->l_buf_cancel_table = NULL;
3809
3810         return error;
3811 }
3812
3813 /*
3814  * Do the actual recovery
3815  */
3816 STATIC int
3817 xlog_do_recover(
3818         xlog_t          *log,
3819         xfs_daddr_t     head_blk,
3820         xfs_daddr_t     tail_blk)
3821 {
3822         int             error;
3823         xfs_buf_t       *bp;
3824         xfs_sb_t        *sbp;
3825
3826         /*
3827          * First replay the images in the log.
3828          */
3829         error = xlog_do_log_recovery(log, head_blk, tail_blk);
3830         if (error) {
3831                 return error;
3832         }
3833
3834         XFS_bflush(log->l_mp->m_ddev_targp);
3835
3836         /*
3837          * If IO errors happened during recovery, bail out.
3838          */
3839         if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3840                 return (EIO);
3841         }
3842
3843         /*
3844          * We now update the tail_lsn since much of the recovery has completed
3845          * and there may be space available to use.  If there were no extent
3846          * or iunlinks, we can free up the entire log and set the tail_lsn to
3847          * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3848          * lsn of the last known good LR on disk.  If there are extent frees
3849          * or iunlinks they will have some entries in the AIL; so we look at
3850          * the AIL to determine how to set the tail_lsn.
3851          */
3852         xlog_assign_tail_lsn(log->l_mp);
3853
3854         /*
3855          * Now that we've finished replaying all buffer and inode
3856          * updates, re-read in the superblock.
3857          */
3858         bp = xfs_getsb(log->l_mp, 0);
3859         XFS_BUF_UNDONE(bp);
3860         ASSERT(!(XFS_BUF_ISWRITE(bp)));
3861         ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
3862         XFS_BUF_READ(bp);
3863         XFS_BUF_UNASYNC(bp);
3864         xfsbdstrat(log->l_mp, bp);
3865         error = xfs_iowait(bp);
3866         if (error) {
3867                 xfs_ioerror_alert("xlog_do_recover",
3868                                   log->l_mp, bp, XFS_BUF_ADDR(bp));
3869                 ASSERT(0);
3870                 xfs_buf_relse(bp);
3871                 return error;
3872         }
3873
3874         /* Convert superblock from on-disk format */
3875         sbp = &log->l_mp->m_sb;
3876         xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
3877         ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3878         ASSERT(xfs_sb_good_version(sbp));
3879         xfs_buf_relse(bp);
3880
3881         /* We've re-read the superblock so re-initialize per-cpu counters */
3882         xfs_icsb_reinit_counters(log->l_mp);
3883
3884         xlog_recover_check_summary(log);
3885
3886         /* Normal transactions can now occur */
3887         log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3888         return 0;
3889 }
3890
3891 /*
3892  * Perform recovery and re-initialize some log variables in xlog_find_tail.
3893  *
3894  * Return error or zero.
3895  */
3896 int
3897 xlog_recover(
3898         xlog_t          *log)
3899 {
3900         xfs_daddr_t     head_blk, tail_blk;
3901         int             error;
3902
3903         /* find the tail of the log */
3904         if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3905                 return error;
3906
3907         if (tail_blk != head_blk) {
3908                 /* There used to be a comment here:
3909                  *
3910                  * disallow recovery on read-only mounts.  note -- mount
3911                  * checks for ENOSPC and turns it into an intelligent
3912                  * error message.
3913                  * ...but this is no longer true.  Now, unless you specify
3914                  * NORECOVERY (in which case this function would never be
3915                  * called), we just go ahead and recover.  We do this all
3916                  * under the vfs layer, so we can get away with it unless
3917                  * the device itself is read-only, in which case we fail.
3918                  */
3919                 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3920                         return error;
3921                 }
3922
3923                 cmn_err(CE_NOTE,
3924                         "Starting XFS recovery on filesystem: %s (logdev: %s)",
3925                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3926                         log->l_mp->m_logname : "internal");
3927
3928                 error = xlog_do_recover(log, head_blk, tail_blk);
3929                 log->l_flags |= XLOG_RECOVERY_NEEDED;
3930         }
3931         return error;
3932 }
3933
3934 /*
3935  * In the first part of recovery we replay inodes and buffers and build
3936  * up the list of extent free items which need to be processed.  Here
3937  * we process the extent free items and clean up the on disk unlinked
3938  * inode lists.  This is separated from the first part of recovery so
3939  * that the root and real-time bitmap inodes can be read in from disk in
3940  * between the two stages.  This is necessary so that we can free space
3941  * in the real-time portion of the file system.
3942  */
3943 int
3944 xlog_recover_finish(
3945         xlog_t          *log)
3946 {
3947         /*
3948          * Now we're ready to do the transactions needed for the
3949          * rest of recovery.  Start with completing all the extent
3950          * free intent records and then process the unlinked inode
3951          * lists.  At this point, we essentially run in normal mode
3952          * except that we're still performing recovery actions
3953          * rather than accepting new requests.
3954          */
3955         if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3956                 int     error;
3957                 error = xlog_recover_process_efis(log);
3958                 if (error) {
3959                         cmn_err(CE_ALERT,
3960                                 "Failed to recover EFIs on filesystem: %s",
3961                                 log->l_mp->m_fsname);
3962                         return error;
3963                 }
3964                 /*
3965                  * Sync the log to get all the EFIs out of the AIL.
3966                  * This isn't absolutely necessary, but it helps in
3967                  * case the unlink transactions would have problems
3968                  * pushing the EFIs out of the way.
3969                  */
3970                 xfs_log_force(log->l_mp, (xfs_lsn_t)0,
3971                               (XFS_LOG_FORCE | XFS_LOG_SYNC));
3972
3973                 xlog_recover_process_iunlinks(log);
3974
3975                 xlog_recover_check_summary(log);
3976
3977                 cmn_err(CE_NOTE,
3978                         "Ending XFS recovery on filesystem: %s (logdev: %s)",
3979                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3980                         log->l_mp->m_logname : "internal");
3981                 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3982         } else {
3983                 cmn_err(CE_DEBUG,
3984                         "!Ending clean XFS mount for filesystem: %s\n",
3985                         log->l_mp->m_fsname);
3986         }
3987         return 0;
3988 }
3989
3990
3991 #if defined(DEBUG)
3992 /*
3993  * Read all of the agf and agi counters and check that they
3994  * are consistent with the superblock counters.
3995  */
3996 void
3997 xlog_recover_check_summary(
3998         xlog_t          *log)
3999 {
4000         xfs_mount_t     *mp;
4001         xfs_agf_t       *agfp;
4002         xfs_buf_t       *agfbp;
4003         xfs_buf_t       *agibp;
4004         xfs_buf_t       *sbbp;
4005 #ifdef XFS_LOUD_RECOVERY
4006         xfs_sb_t        *sbp;
4007 #endif
4008         xfs_agnumber_t  agno;
4009         __uint64_t      freeblks;
4010         __uint64_t      itotal;
4011         __uint64_t      ifree;
4012         int             error;
4013
4014         mp = log->l_mp;
4015
4016         freeblks = 0LL;
4017         itotal = 0LL;
4018         ifree = 0LL;
4019         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4020                 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
4021                 if (error) {
4022                         xfs_fs_cmn_err(CE_ALERT, mp,
4023                                         "xlog_recover_check_summary(agf)"
4024                                         "agf read failed agno %d error %d",
4025                                                         agno, error);
4026                 } else {
4027                         agfp = XFS_BUF_TO_AGF(agfbp);
4028                         freeblks += be32_to_cpu(agfp->agf_freeblks) +
4029                                     be32_to_cpu(agfp->agf_flcount);
4030                         xfs_buf_relse(agfbp);
4031                 }
4032
4033                 error = xfs_read_agi(mp, NULL, agno, &agibp);
4034                 if (!error) {
4035                         struct xfs_agi  *agi = XFS_BUF_TO_AGI(agibp);
4036
4037                         itotal += be32_to_cpu(agi->agi_count);
4038                         ifree += be32_to_cpu(agi->agi_freecount);
4039                         xfs_buf_relse(agibp);
4040                 }
4041         }
4042
4043         sbbp = xfs_getsb(mp, 0);
4044 #ifdef XFS_LOUD_RECOVERY
4045         sbp = &mp->m_sb;
4046         xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(sbbp));
4047         cmn_err(CE_NOTE,
4048                 "xlog_recover_check_summary: sb_icount %Lu itotal %Lu",
4049                 sbp->sb_icount, itotal);
4050         cmn_err(CE_NOTE,
4051                 "xlog_recover_check_summary: sb_ifree %Lu itotal %Lu",
4052                 sbp->sb_ifree, ifree);
4053         cmn_err(CE_NOTE,
4054                 "xlog_recover_check_summary: sb_fdblocks %Lu freeblks %Lu",
4055                 sbp->sb_fdblocks, freeblks);
4056 #if 0
4057         /*
4058          * This is turned off until I account for the allocation
4059          * btree blocks which live in free space.
4060          */
4061         ASSERT(sbp->sb_icount == itotal);
4062         ASSERT(sbp->sb_ifree == ifree);
4063         ASSERT(sbp->sb_fdblocks == freeblks);
4064 #endif
4065 #endif
4066         xfs_buf_relse(sbbp);
4067 }
4068 #endif /* DEBUG */