xfs: fix min bufsize bugs in two places
[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 #include "xfs_trace.h"
50
51 STATIC int      xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
52 STATIC int      xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
53 #if defined(DEBUG)
54 STATIC void     xlog_recover_check_summary(xlog_t *);
55 #else
56 #define xlog_recover_check_summary(log)
57 #endif
58
59
60 /*
61  * Sector aligned buffer routines for buffer create/read/write/access
62  */
63
64 #define XLOG_SECTOR_ROUNDUP_BBCOUNT(log, bbs)   \
65         ( ((log)->l_sectbb_mask && (bbs & (log)->l_sectbb_mask)) ? \
66         ((bbs + (log)->l_sectbb_mask + 1) & ~(log)->l_sectbb_mask) : (bbs) )
67 #define XLOG_SECTOR_ROUNDDOWN_BLKNO(log, bno)   ((bno) & ~(log)->l_sectbb_mask)
68
69 /* Number of basic blocks in a log sector */
70 #define xlog_sectbb(log) (1 << (log)->l_sectbb_log)
71
72 STATIC xfs_buf_t *
73 xlog_get_bp(
74         xlog_t          *log,
75         int             nbblks)
76 {
77         if (nbblks <= 0 || nbblks > log->l_logBBsize) {
78                 xlog_warn("XFS: Invalid block length (0x%x) given for buffer", nbblks);
79                 XFS_ERROR_REPORT("xlog_get_bp(1)",
80                                  XFS_ERRLEVEL_HIGH, log->l_mp);
81                 return NULL;
82         }
83
84         if (log->l_sectbb_log) {
85                 if (nbblks > 1)
86                         nbblks += XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
87                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
88         }
89         return xfs_buf_get_noaddr(BBTOB(nbblks), log->l_mp->m_logdev_targp);
90 }
91
92 STATIC void
93 xlog_put_bp(
94         xfs_buf_t       *bp)
95 {
96         xfs_buf_free(bp);
97 }
98
99 STATIC xfs_caddr_t
100 xlog_align(
101         xlog_t          *log,
102         xfs_daddr_t     blk_no,
103         int             nbblks,
104         xfs_buf_t       *bp)
105 {
106         xfs_caddr_t     ptr;
107
108         if (!log->l_sectbb_log)
109                 return XFS_BUF_PTR(bp);
110
111         ptr = XFS_BUF_PTR(bp) + BBTOB((int)blk_no & log->l_sectbb_mask);
112         ASSERT(XFS_BUF_SIZE(bp) >=
113                 BBTOB(nbblks + (blk_no & log->l_sectbb_mask)));
114         return ptr;
115 }
116
117
118 /*
119  * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
120  */
121 STATIC int
122 xlog_bread_noalign(
123         xlog_t          *log,
124         xfs_daddr_t     blk_no,
125         int             nbblks,
126         xfs_buf_t       *bp)
127 {
128         int             error;
129
130         if (nbblks <= 0 || nbblks > log->l_logBBsize) {
131                 xlog_warn("XFS: Invalid block length (0x%x) given for buffer", nbblks);
132                 XFS_ERROR_REPORT("xlog_bread(1)",
133                                  XFS_ERRLEVEL_HIGH, log->l_mp);
134                 return EFSCORRUPTED;
135         }
136
137         if (log->l_sectbb_log) {
138                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
139                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
140         }
141
142         ASSERT(nbblks > 0);
143         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
144         ASSERT(bp);
145
146         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
147         XFS_BUF_READ(bp);
148         XFS_BUF_BUSY(bp);
149         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
150         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
151
152         xfsbdstrat(log->l_mp, bp);
153         error = xfs_iowait(bp);
154         if (error)
155                 xfs_ioerror_alert("xlog_bread", log->l_mp,
156                                   bp, XFS_BUF_ADDR(bp));
157         return error;
158 }
159
160 STATIC int
161 xlog_bread(
162         xlog_t          *log,
163         xfs_daddr_t     blk_no,
164         int             nbblks,
165         xfs_buf_t       *bp,
166         xfs_caddr_t     *offset)
167 {
168         int             error;
169
170         error = xlog_bread_noalign(log, blk_no, nbblks, bp);
171         if (error)
172                 return error;
173
174         *offset = xlog_align(log, blk_no, nbblks, bp);
175         return 0;
176 }
177
178 /*
179  * Write out the buffer at the given block for the given number of blocks.
180  * The buffer is kept locked across the write and is returned locked.
181  * This can only be used for synchronous log writes.
182  */
183 STATIC int
184 xlog_bwrite(
185         xlog_t          *log,
186         xfs_daddr_t     blk_no,
187         int             nbblks,
188         xfs_buf_t       *bp)
189 {
190         int             error;
191
192         if (nbblks <= 0 || nbblks > log->l_logBBsize) {
193                 xlog_warn("XFS: Invalid block length (0x%x) given for buffer", nbblks);
194                 XFS_ERROR_REPORT("xlog_bwrite(1)",
195                                  XFS_ERRLEVEL_HIGH, log->l_mp);
196                 return EFSCORRUPTED;
197         }
198
199         if (log->l_sectbb_log) {
200                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
201                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
202         }
203
204         ASSERT(nbblks > 0);
205         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
206
207         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
208         XFS_BUF_ZEROFLAGS(bp);
209         XFS_BUF_BUSY(bp);
210         XFS_BUF_HOLD(bp);
211         XFS_BUF_PSEMA(bp, PRIBIO);
212         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
213         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
214
215         if ((error = xfs_bwrite(log->l_mp, bp)))
216                 xfs_ioerror_alert("xlog_bwrite", log->l_mp,
217                                   bp, XFS_BUF_ADDR(bp));
218         return error;
219 }
220
221 #ifdef DEBUG
222 /*
223  * dump debug superblock and log record information
224  */
225 STATIC void
226 xlog_header_check_dump(
227         xfs_mount_t             *mp,
228         xlog_rec_header_t       *head)
229 {
230         cmn_err(CE_DEBUG, "%s:  SB : uuid = %pU, fmt = %d\n",
231                 __func__, &mp->m_sb.sb_uuid, XLOG_FMT);
232         cmn_err(CE_DEBUG, "    log : uuid = %pU, fmt = %d\n",
233                 &head->h_fs_uuid, be32_to_cpu(head->h_fmt));
234 }
235 #else
236 #define xlog_header_check_dump(mp, head)
237 #endif
238
239 /*
240  * check log record header for recovery
241  */
242 STATIC int
243 xlog_header_check_recover(
244         xfs_mount_t             *mp,
245         xlog_rec_header_t       *head)
246 {
247         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
248
249         /*
250          * IRIX doesn't write the h_fmt field and leaves it zeroed
251          * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
252          * a dirty log created in IRIX.
253          */
254         if (unlikely(be32_to_cpu(head->h_fmt) != XLOG_FMT)) {
255                 xlog_warn(
256         "XFS: dirty log written in incompatible format - can't recover");
257                 xlog_header_check_dump(mp, head);
258                 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
259                                  XFS_ERRLEVEL_HIGH, mp);
260                 return XFS_ERROR(EFSCORRUPTED);
261         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
262                 xlog_warn(
263         "XFS: dirty log entry has mismatched uuid - can't recover");
264                 xlog_header_check_dump(mp, head);
265                 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
266                                  XFS_ERRLEVEL_HIGH, mp);
267                 return XFS_ERROR(EFSCORRUPTED);
268         }
269         return 0;
270 }
271
272 /*
273  * read the head block of the log and check the header
274  */
275 STATIC int
276 xlog_header_check_mount(
277         xfs_mount_t             *mp,
278         xlog_rec_header_t       *head)
279 {
280         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
281
282         if (uuid_is_nil(&head->h_fs_uuid)) {
283                 /*
284                  * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
285                  * h_fs_uuid is nil, we assume this log was last mounted
286                  * by IRIX and continue.
287                  */
288                 xlog_warn("XFS: nil uuid in log - IRIX style log");
289         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
290                 xlog_warn("XFS: log has mismatched uuid - can't recover");
291                 xlog_header_check_dump(mp, head);
292                 XFS_ERROR_REPORT("xlog_header_check_mount",
293                                  XFS_ERRLEVEL_HIGH, mp);
294                 return XFS_ERROR(EFSCORRUPTED);
295         }
296         return 0;
297 }
298
299 STATIC void
300 xlog_recover_iodone(
301         struct xfs_buf  *bp)
302 {
303         if (XFS_BUF_GETERROR(bp)) {
304                 /*
305                  * We're not going to bother about retrying
306                  * this during recovery. One strike!
307                  */
308                 xfs_ioerror_alert("xlog_recover_iodone",
309                                   bp->b_mount, bp, XFS_BUF_ADDR(bp));
310                 xfs_force_shutdown(bp->b_mount, SHUTDOWN_META_IO_ERROR);
311         }
312         bp->b_mount = NULL;
313         XFS_BUF_CLR_IODONE_FUNC(bp);
314         xfs_biodone(bp);
315 }
316
317 /*
318  * This routine finds (to an approximation) the first block in the physical
319  * log which contains the given cycle.  It uses a binary search algorithm.
320  * Note that the algorithm can not be perfect because the disk will not
321  * necessarily be perfect.
322  */
323 STATIC int
324 xlog_find_cycle_start(
325         xlog_t          *log,
326         xfs_buf_t       *bp,
327         xfs_daddr_t     first_blk,
328         xfs_daddr_t     *last_blk,
329         uint            cycle)
330 {
331         xfs_caddr_t     offset;
332         xfs_daddr_t     mid_blk;
333         uint            mid_cycle;
334         int             error;
335
336         mid_blk = BLK_AVG(first_blk, *last_blk);
337         while (mid_blk != first_blk && mid_blk != *last_blk) {
338                 error = xlog_bread(log, mid_blk, 1, bp, &offset);
339                 if (error)
340                         return error;
341                 mid_cycle = xlog_get_cycle(offset);
342                 if (mid_cycle == cycle) {
343                         *last_blk = mid_blk;
344                         /* last_half_cycle == mid_cycle */
345                 } else {
346                         first_blk = mid_blk;
347                         /* first_half_cycle == mid_cycle */
348                 }
349                 mid_blk = BLK_AVG(first_blk, *last_blk);
350         }
351         ASSERT((mid_blk == first_blk && mid_blk+1 == *last_blk) ||
352                (mid_blk == *last_blk && mid_blk-1 == first_blk));
353
354         return 0;
355 }
356
357 /*
358  * Check that the range of blocks does not contain the cycle number
359  * given.  The scan needs to occur from front to back and the ptr into the
360  * region must be updated since a later routine will need to perform another
361  * test.  If the region is completely good, we end up returning the same
362  * last block number.
363  *
364  * Set blkno to -1 if we encounter no errors.  This is an invalid block number
365  * since we don't ever expect logs to get this large.
366  */
367 STATIC int
368 xlog_find_verify_cycle(
369         xlog_t          *log,
370         xfs_daddr_t     start_blk,
371         int             nbblks,
372         uint            stop_on_cycle_no,
373         xfs_daddr_t     *new_blk)
374 {
375         xfs_daddr_t     i, j;
376         uint            cycle;
377         xfs_buf_t       *bp;
378         xfs_daddr_t     bufblks;
379         xfs_caddr_t     buf = NULL;
380         int             error = 0;
381
382         /*
383          * Greedily allocate a buffer big enough to handle the full
384          * range of basic blocks we'll be examining.  If that fails,
385          * try a smaller size.  We need to be able to read at least
386          * a log sector, or we're out of luck.
387          */
388         bufblks = 1 << ffs(nbblks);
389         while (!(bp = xlog_get_bp(log, bufblks))) {
390                 bufblks >>= 1;
391                 if (bufblks < xlog_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 STATIC 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         /*
1169          * Greedily allocate a buffer big enough to handle the full
1170          * range of basic blocks to be written.  If that fails, try
1171          * a smaller size.  We need to be able to write at least a
1172          * log sector, or we're out of luck.
1173          */
1174         bufblks = 1 << ffs(blocks);
1175         while (!(bp = xlog_get_bp(log, bufblks))) {
1176                 bufblks >>= 1;
1177                 if (bufblks < xlog_sectbb(log))
1178                         return ENOMEM;
1179         }
1180
1181         /* We may need to do a read at the start to fill in part of
1182          * the buffer in the starting sector not covered by the first
1183          * write below.
1184          */
1185         balign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, start_block);
1186         if (balign != start_block) {
1187                 error = xlog_bread_noalign(log, start_block, 1, bp);
1188                 if (error)
1189                         goto out_put_bp;
1190
1191                 j = start_block - balign;
1192         }
1193
1194         for (i = start_block; i < end_block; i += bufblks) {
1195                 int             bcount, endcount;
1196
1197                 bcount = min(bufblks, end_block - start_block);
1198                 endcount = bcount - j;
1199
1200                 /* We may need to do a read at the end to fill in part of
1201                  * the buffer in the final sector not covered by the write.
1202                  * If this is the same sector as the above read, skip it.
1203                  */
1204                 ealign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, end_block);
1205                 if (j == 0 && (start_block + endcount > ealign)) {
1206                         offset = XFS_BUF_PTR(bp);
1207                         balign = BBTOB(ealign - start_block);
1208                         error = XFS_BUF_SET_PTR(bp, offset + balign,
1209                                                 BBTOB(sectbb));
1210                         if (error)
1211                                 break;
1212
1213                         error = xlog_bread_noalign(log, ealign, sectbb, bp);
1214                         if (error)
1215                                 break;
1216
1217                         error = XFS_BUF_SET_PTR(bp, offset, bufblks);
1218                         if (error)
1219                                 break;
1220                 }
1221
1222                 offset = xlog_align(log, start_block, endcount, bp);
1223                 for (; j < endcount; j++) {
1224                         xlog_add_record(log, offset, cycle, i+j,
1225                                         tail_cycle, tail_block);
1226                         offset += BBSIZE;
1227                 }
1228                 error = xlog_bwrite(log, start_block, endcount, bp);
1229                 if (error)
1230                         break;
1231                 start_block += endcount;
1232                 j = 0;
1233         }
1234
1235  out_put_bp:
1236         xlog_put_bp(bp);
1237         return error;
1238 }
1239
1240 /*
1241  * This routine is called to blow away any incomplete log writes out
1242  * in front of the log head.  We do this so that we won't become confused
1243  * if we come up, write only a little bit more, and then crash again.
1244  * If we leave the partial log records out there, this situation could
1245  * cause us to think those partial writes are valid blocks since they
1246  * have the current cycle number.  We get rid of them by overwriting them
1247  * with empty log records with the old cycle number rather than the
1248  * current one.
1249  *
1250  * The tail lsn is passed in rather than taken from
1251  * the log so that we will not write over the unmount record after a
1252  * clean unmount in a 512 block log.  Doing so would leave the log without
1253  * any valid log records in it until a new one was written.  If we crashed
1254  * during that time we would not be able to recover.
1255  */
1256 STATIC int
1257 xlog_clear_stale_blocks(
1258         xlog_t          *log,
1259         xfs_lsn_t       tail_lsn)
1260 {
1261         int             tail_cycle, head_cycle;
1262         int             tail_block, head_block;
1263         int             tail_distance, max_distance;
1264         int             distance;
1265         int             error;
1266
1267         tail_cycle = CYCLE_LSN(tail_lsn);
1268         tail_block = BLOCK_LSN(tail_lsn);
1269         head_cycle = log->l_curr_cycle;
1270         head_block = log->l_curr_block;
1271
1272         /*
1273          * Figure out the distance between the new head of the log
1274          * and the tail.  We want to write over any blocks beyond the
1275          * head that we may have written just before the crash, but
1276          * we don't want to overwrite the tail of the log.
1277          */
1278         if (head_cycle == tail_cycle) {
1279                 /*
1280                  * The tail is behind the head in the physical log,
1281                  * so the distance from the head to the tail is the
1282                  * distance from the head to the end of the log plus
1283                  * the distance from the beginning of the log to the
1284                  * tail.
1285                  */
1286                 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1287                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1288                                          XFS_ERRLEVEL_LOW, log->l_mp);
1289                         return XFS_ERROR(EFSCORRUPTED);
1290                 }
1291                 tail_distance = tail_block + (log->l_logBBsize - head_block);
1292         } else {
1293                 /*
1294                  * The head is behind the tail in the physical log,
1295                  * so the distance from the head to the tail is just
1296                  * the tail block minus the head block.
1297                  */
1298                 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1299                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1300                                          XFS_ERRLEVEL_LOW, log->l_mp);
1301                         return XFS_ERROR(EFSCORRUPTED);
1302                 }
1303                 tail_distance = tail_block - head_block;
1304         }
1305
1306         /*
1307          * If the head is right up against the tail, we can't clear
1308          * anything.
1309          */
1310         if (tail_distance <= 0) {
1311                 ASSERT(tail_distance == 0);
1312                 return 0;
1313         }
1314
1315         max_distance = XLOG_TOTAL_REC_SHIFT(log);
1316         /*
1317          * Take the smaller of the maximum amount of outstanding I/O
1318          * we could have and the distance to the tail to clear out.
1319          * We take the smaller so that we don't overwrite the tail and
1320          * we don't waste all day writing from the head to the tail
1321          * for no reason.
1322          */
1323         max_distance = MIN(max_distance, tail_distance);
1324
1325         if ((head_block + max_distance) <= log->l_logBBsize) {
1326                 /*
1327                  * We can stomp all the blocks we need to without
1328                  * wrapping around the end of the log.  Just do it
1329                  * in a single write.  Use the cycle number of the
1330                  * current cycle minus one so that the log will look like:
1331                  *     n ... | n - 1 ...
1332                  */
1333                 error = xlog_write_log_records(log, (head_cycle - 1),
1334                                 head_block, max_distance, tail_cycle,
1335                                 tail_block);
1336                 if (error)
1337                         return error;
1338         } else {
1339                 /*
1340                  * We need to wrap around the end of the physical log in
1341                  * order to clear all the blocks.  Do it in two separate
1342                  * I/Os.  The first write should be from the head to the
1343                  * end of the physical log, and it should use the current
1344                  * cycle number minus one just like above.
1345                  */
1346                 distance = log->l_logBBsize - head_block;
1347                 error = xlog_write_log_records(log, (head_cycle - 1),
1348                                 head_block, distance, tail_cycle,
1349                                 tail_block);
1350
1351                 if (error)
1352                         return error;
1353
1354                 /*
1355                  * Now write the blocks at the start of the physical log.
1356                  * This writes the remainder of the blocks we want to clear.
1357                  * It uses the current cycle number since we're now on the
1358                  * same cycle as the head so that we get:
1359                  *    n ... n ... | n - 1 ...
1360                  *    ^^^^^ blocks we're writing
1361                  */
1362                 distance = max_distance - (log->l_logBBsize - head_block);
1363                 error = xlog_write_log_records(log, head_cycle, 0, distance,
1364                                 tail_cycle, tail_block);
1365                 if (error)
1366                         return error;
1367         }
1368
1369         return 0;
1370 }
1371
1372 /******************************************************************************
1373  *
1374  *              Log recover routines
1375  *
1376  ******************************************************************************
1377  */
1378
1379 STATIC xlog_recover_t *
1380 xlog_recover_find_tid(
1381         struct hlist_head       *head,
1382         xlog_tid_t              tid)
1383 {
1384         xlog_recover_t          *trans;
1385         struct hlist_node       *n;
1386
1387         hlist_for_each_entry(trans, n, head, r_list) {
1388                 if (trans->r_log_tid == tid)
1389                         return trans;
1390         }
1391         return NULL;
1392 }
1393
1394 STATIC void
1395 xlog_recover_new_tid(
1396         struct hlist_head       *head,
1397         xlog_tid_t              tid,
1398         xfs_lsn_t               lsn)
1399 {
1400         xlog_recover_t          *trans;
1401
1402         trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1403         trans->r_log_tid   = tid;
1404         trans->r_lsn       = lsn;
1405         INIT_LIST_HEAD(&trans->r_itemq);
1406
1407         INIT_HLIST_NODE(&trans->r_list);
1408         hlist_add_head(&trans->r_list, head);
1409 }
1410
1411 STATIC void
1412 xlog_recover_add_item(
1413         struct list_head        *head)
1414 {
1415         xlog_recover_item_t     *item;
1416
1417         item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1418         INIT_LIST_HEAD(&item->ri_list);
1419         list_add_tail(&item->ri_list, head);
1420 }
1421
1422 STATIC int
1423 xlog_recover_add_to_cont_trans(
1424         struct log              *log,
1425         xlog_recover_t          *trans,
1426         xfs_caddr_t             dp,
1427         int                     len)
1428 {
1429         xlog_recover_item_t     *item;
1430         xfs_caddr_t             ptr, old_ptr;
1431         int                     old_len;
1432
1433         if (list_empty(&trans->r_itemq)) {
1434                 /* finish copying rest of trans header */
1435                 xlog_recover_add_item(&trans->r_itemq);
1436                 ptr = (xfs_caddr_t) &trans->r_theader +
1437                                 sizeof(xfs_trans_header_t) - len;
1438                 memcpy(ptr, dp, len); /* d, s, l */
1439                 return 0;
1440         }
1441         /* take the tail entry */
1442         item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1443
1444         old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1445         old_len = item->ri_buf[item->ri_cnt-1].i_len;
1446
1447         ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1448         memcpy(&ptr[old_len], dp, len); /* d, s, l */
1449         item->ri_buf[item->ri_cnt-1].i_len += len;
1450         item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1451         trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
1452         return 0;
1453 }
1454
1455 /*
1456  * The next region to add is the start of a new region.  It could be
1457  * a whole region or it could be the first part of a new region.  Because
1458  * of this, the assumption here is that the type and size fields of all
1459  * format structures fit into the first 32 bits of the structure.
1460  *
1461  * This works because all regions must be 32 bit aligned.  Therefore, we
1462  * either have both fields or we have neither field.  In the case we have
1463  * neither field, the data part of the region is zero length.  We only have
1464  * a log_op_header and can throw away the header since a new one will appear
1465  * later.  If we have at least 4 bytes, then we can determine how many regions
1466  * will appear in the current log item.
1467  */
1468 STATIC int
1469 xlog_recover_add_to_trans(
1470         struct log              *log,
1471         xlog_recover_t          *trans,
1472         xfs_caddr_t             dp,
1473         int                     len)
1474 {
1475         xfs_inode_log_format_t  *in_f;                  /* any will do */
1476         xlog_recover_item_t     *item;
1477         xfs_caddr_t             ptr;
1478
1479         if (!len)
1480                 return 0;
1481         if (list_empty(&trans->r_itemq)) {
1482                 /* we need to catch log corruptions here */
1483                 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1484                         xlog_warn("XFS: xlog_recover_add_to_trans: "
1485                                   "bad header magic number");
1486                         ASSERT(0);
1487                         return XFS_ERROR(EIO);
1488                 }
1489                 if (len == sizeof(xfs_trans_header_t))
1490                         xlog_recover_add_item(&trans->r_itemq);
1491                 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1492                 return 0;
1493         }
1494
1495         ptr = kmem_alloc(len, KM_SLEEP);
1496         memcpy(ptr, dp, len);
1497         in_f = (xfs_inode_log_format_t *)ptr;
1498
1499         /* take the tail entry */
1500         item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
1501         if (item->ri_total != 0 &&
1502              item->ri_total == item->ri_cnt) {
1503                 /* tail item is in use, get a new one */
1504                 xlog_recover_add_item(&trans->r_itemq);
1505                 item = list_entry(trans->r_itemq.prev,
1506                                         xlog_recover_item_t, ri_list);
1507         }
1508
1509         if (item->ri_total == 0) {              /* first region to be added */
1510                 if (in_f->ilf_size == 0 ||
1511                     in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
1512                         xlog_warn(
1513         "XFS: bad number of regions (%d) in inode log format",
1514                                   in_f->ilf_size);
1515                         ASSERT(0);
1516                         return XFS_ERROR(EIO);
1517                 }
1518
1519                 item->ri_total = in_f->ilf_size;
1520                 item->ri_buf =
1521                         kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
1522                                     KM_SLEEP);
1523         }
1524         ASSERT(item->ri_total > item->ri_cnt);
1525         /* Description region is ri_buf[0] */
1526         item->ri_buf[item->ri_cnt].i_addr = ptr;
1527         item->ri_buf[item->ri_cnt].i_len  = len;
1528         item->ri_cnt++;
1529         trace_xfs_log_recover_item_add(log, trans, item, 0);
1530         return 0;
1531 }
1532
1533 /*
1534  * Sort the log items in the transaction. Cancelled buffers need
1535  * to be put first so they are processed before any items that might
1536  * modify the buffers. If they are cancelled, then the modifications
1537  * don't need to be replayed.
1538  */
1539 STATIC int
1540 xlog_recover_reorder_trans(
1541         struct log              *log,
1542         xlog_recover_t          *trans,
1543         int                     pass)
1544 {
1545         xlog_recover_item_t     *item, *n;
1546         LIST_HEAD(sort_list);
1547
1548         list_splice_init(&trans->r_itemq, &sort_list);
1549         list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1550                 xfs_buf_log_format_t    *buf_f;
1551
1552                 buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
1553
1554                 switch (ITEM_TYPE(item)) {
1555                 case XFS_LI_BUF:
1556                         if (!(buf_f->blf_flags & XFS_BLI_CANCEL)) {
1557                                 trace_xfs_log_recover_item_reorder_head(log,
1558                                                         trans, item, pass);
1559                                 list_move(&item->ri_list, &trans->r_itemq);
1560                                 break;
1561                         }
1562                 case XFS_LI_INODE:
1563                 case XFS_LI_DQUOT:
1564                 case XFS_LI_QUOTAOFF:
1565                 case XFS_LI_EFD:
1566                 case XFS_LI_EFI:
1567                         trace_xfs_log_recover_item_reorder_tail(log,
1568                                                         trans, item, pass);
1569                         list_move_tail(&item->ri_list, &trans->r_itemq);
1570                         break;
1571                 default:
1572                         xlog_warn(
1573         "XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1574                         ASSERT(0);
1575                         return XFS_ERROR(EIO);
1576                 }
1577         }
1578         ASSERT(list_empty(&sort_list));
1579         return 0;
1580 }
1581
1582 /*
1583  * Build up the table of buf cancel records so that we don't replay
1584  * cancelled data in the second pass.  For buffer records that are
1585  * not cancel records, there is nothing to do here so we just return.
1586  *
1587  * If we get a cancel record which is already in the table, this indicates
1588  * that the buffer was cancelled multiple times.  In order to ensure
1589  * that during pass 2 we keep the record in the table until we reach its
1590  * last occurrence in the log, we keep a reference count in the cancel
1591  * record in the table to tell us how many times we expect to see this
1592  * record during the second pass.
1593  */
1594 STATIC void
1595 xlog_recover_do_buffer_pass1(
1596         xlog_t                  *log,
1597         xfs_buf_log_format_t    *buf_f)
1598 {
1599         xfs_buf_cancel_t        *bcp;
1600         xfs_buf_cancel_t        *nextp;
1601         xfs_buf_cancel_t        *prevp;
1602         xfs_buf_cancel_t        **bucket;
1603         xfs_daddr_t             blkno = 0;
1604         uint                    len = 0;
1605         ushort                  flags = 0;
1606
1607         switch (buf_f->blf_type) {
1608         case XFS_LI_BUF:
1609                 blkno = buf_f->blf_blkno;
1610                 len = buf_f->blf_len;
1611                 flags = buf_f->blf_flags;
1612                 break;
1613         }
1614
1615         /*
1616          * If this isn't a cancel buffer item, then just return.
1617          */
1618         if (!(flags & XFS_BLI_CANCEL)) {
1619                 trace_xfs_log_recover_buf_not_cancel(log, buf_f);
1620                 return;
1621         }
1622
1623         /*
1624          * Insert an xfs_buf_cancel record into the hash table of
1625          * them.  If there is already an identical record, bump
1626          * its reference count.
1627          */
1628         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1629                                           XLOG_BC_TABLE_SIZE];
1630         /*
1631          * If the hash bucket is empty then just insert a new record into
1632          * the bucket.
1633          */
1634         if (*bucket == NULL) {
1635                 bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1636                                                      KM_SLEEP);
1637                 bcp->bc_blkno = blkno;
1638                 bcp->bc_len = len;
1639                 bcp->bc_refcount = 1;
1640                 bcp->bc_next = NULL;
1641                 *bucket = bcp;
1642                 return;
1643         }
1644
1645         /*
1646          * The hash bucket is not empty, so search for duplicates of our
1647          * record.  If we find one them just bump its refcount.  If not
1648          * then add us at the end of the list.
1649          */
1650         prevp = NULL;
1651         nextp = *bucket;
1652         while (nextp != NULL) {
1653                 if (nextp->bc_blkno == blkno && nextp->bc_len == len) {
1654                         nextp->bc_refcount++;
1655                         trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
1656                         return;
1657                 }
1658                 prevp = nextp;
1659                 nextp = nextp->bc_next;
1660         }
1661         ASSERT(prevp != NULL);
1662         bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1663                                              KM_SLEEP);
1664         bcp->bc_blkno = blkno;
1665         bcp->bc_len = len;
1666         bcp->bc_refcount = 1;
1667         bcp->bc_next = NULL;
1668         prevp->bc_next = bcp;
1669         trace_xfs_log_recover_buf_cancel_add(log, buf_f);
1670 }
1671
1672 /*
1673  * Check to see whether the buffer being recovered has a corresponding
1674  * entry in the buffer cancel record table.  If it does then return 1
1675  * so that it will be cancelled, otherwise return 0.  If the buffer is
1676  * actually a buffer cancel item (XFS_BLI_CANCEL is set), then decrement
1677  * the refcount on the entry in the table and remove it from the table
1678  * if this is the last reference.
1679  *
1680  * We remove the cancel record from the table when we encounter its
1681  * last occurrence in the log so that if the same buffer is re-used
1682  * again after its last cancellation we actually replay the changes
1683  * made at that point.
1684  */
1685 STATIC int
1686 xlog_check_buffer_cancelled(
1687         xlog_t                  *log,
1688         xfs_daddr_t             blkno,
1689         uint                    len,
1690         ushort                  flags)
1691 {
1692         xfs_buf_cancel_t        *bcp;
1693         xfs_buf_cancel_t        *prevp;
1694         xfs_buf_cancel_t        **bucket;
1695
1696         if (log->l_buf_cancel_table == NULL) {
1697                 /*
1698                  * There is nothing in the table built in pass one,
1699                  * so this buffer must not be cancelled.
1700                  */
1701                 ASSERT(!(flags & XFS_BLI_CANCEL));
1702                 return 0;
1703         }
1704
1705         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1706                                           XLOG_BC_TABLE_SIZE];
1707         bcp = *bucket;
1708         if (bcp == NULL) {
1709                 /*
1710                  * There is no corresponding entry in the table built
1711                  * in pass one, so this buffer has not been cancelled.
1712                  */
1713                 ASSERT(!(flags & XFS_BLI_CANCEL));
1714                 return 0;
1715         }
1716
1717         /*
1718          * Search for an entry in the buffer cancel table that
1719          * matches our buffer.
1720          */
1721         prevp = NULL;
1722         while (bcp != NULL) {
1723                 if (bcp->bc_blkno == blkno && bcp->bc_len == len) {
1724                         /*
1725                          * We've go a match, so return 1 so that the
1726                          * recovery of this buffer is cancelled.
1727                          * If this buffer is actually a buffer cancel
1728                          * log item, then decrement the refcount on the
1729                          * one in the table and remove it if this is the
1730                          * last reference.
1731                          */
1732                         if (flags & XFS_BLI_CANCEL) {
1733                                 bcp->bc_refcount--;
1734                                 if (bcp->bc_refcount == 0) {
1735                                         if (prevp == NULL) {
1736                                                 *bucket = bcp->bc_next;
1737                                         } else {
1738                                                 prevp->bc_next = bcp->bc_next;
1739                                         }
1740                                         kmem_free(bcp);
1741                                 }
1742                         }
1743                         return 1;
1744                 }
1745                 prevp = bcp;
1746                 bcp = bcp->bc_next;
1747         }
1748         /*
1749          * We didn't find a corresponding entry in the table, so
1750          * return 0 so that the buffer is NOT cancelled.
1751          */
1752         ASSERT(!(flags & XFS_BLI_CANCEL));
1753         return 0;
1754 }
1755
1756 STATIC int
1757 xlog_recover_do_buffer_pass2(
1758         xlog_t                  *log,
1759         xfs_buf_log_format_t    *buf_f)
1760 {
1761         xfs_daddr_t             blkno = 0;
1762         ushort                  flags = 0;
1763         uint                    len = 0;
1764
1765         switch (buf_f->blf_type) {
1766         case XFS_LI_BUF:
1767                 blkno = buf_f->blf_blkno;
1768                 flags = buf_f->blf_flags;
1769                 len = buf_f->blf_len;
1770                 break;
1771         }
1772
1773         return xlog_check_buffer_cancelled(log, blkno, len, flags);
1774 }
1775
1776 /*
1777  * Perform recovery for a buffer full of inodes.  In these buffers,
1778  * the only data which should be recovered is that which corresponds
1779  * to the di_next_unlinked pointers in the on disk inode structures.
1780  * The rest of the data for the inodes is always logged through the
1781  * inodes themselves rather than the inode buffer and is recovered
1782  * in xlog_recover_do_inode_trans().
1783  *
1784  * The only time when buffers full of inodes are fully recovered is
1785  * when the buffer is full of newly allocated inodes.  In this case
1786  * the buffer will not be marked as an inode buffer and so will be
1787  * sent to xlog_recover_do_reg_buffer() below during recovery.
1788  */
1789 STATIC int
1790 xlog_recover_do_inode_buffer(
1791         xfs_mount_t             *mp,
1792         xlog_recover_item_t     *item,
1793         xfs_buf_t               *bp,
1794         xfs_buf_log_format_t    *buf_f)
1795 {
1796         int                     i;
1797         int                     item_index;
1798         int                     bit;
1799         int                     nbits;
1800         int                     reg_buf_offset;
1801         int                     reg_buf_bytes;
1802         int                     next_unlinked_offset;
1803         int                     inodes_per_buf;
1804         xfs_agino_t             *logged_nextp;
1805         xfs_agino_t             *buffer_nextp;
1806         unsigned int            *data_map = NULL;
1807         unsigned int            map_size = 0;
1808
1809         trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
1810
1811         switch (buf_f->blf_type) {
1812         case XFS_LI_BUF:
1813                 data_map = buf_f->blf_data_map;
1814                 map_size = buf_f->blf_map_size;
1815                 break;
1816         }
1817         /*
1818          * Set the variables corresponding to the current region to
1819          * 0 so that we'll initialize them on the first pass through
1820          * the loop.
1821          */
1822         reg_buf_offset = 0;
1823         reg_buf_bytes = 0;
1824         bit = 0;
1825         nbits = 0;
1826         item_index = 0;
1827         inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1828         for (i = 0; i < inodes_per_buf; i++) {
1829                 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1830                         offsetof(xfs_dinode_t, di_next_unlinked);
1831
1832                 while (next_unlinked_offset >=
1833                        (reg_buf_offset + reg_buf_bytes)) {
1834                         /*
1835                          * The next di_next_unlinked field is beyond
1836                          * the current logged region.  Find the next
1837                          * logged region that contains or is beyond
1838                          * the current di_next_unlinked field.
1839                          */
1840                         bit += nbits;
1841                         bit = xfs_next_bit(data_map, map_size, bit);
1842
1843                         /*
1844                          * If there are no more logged regions in the
1845                          * buffer, then we're done.
1846                          */
1847                         if (bit == -1) {
1848                                 return 0;
1849                         }
1850
1851                         nbits = xfs_contig_bits(data_map, map_size,
1852                                                          bit);
1853                         ASSERT(nbits > 0);
1854                         reg_buf_offset = bit << XFS_BLI_SHIFT;
1855                         reg_buf_bytes = nbits << XFS_BLI_SHIFT;
1856                         item_index++;
1857                 }
1858
1859                 /*
1860                  * If the current logged region starts after the current
1861                  * di_next_unlinked field, then move on to the next
1862                  * di_next_unlinked field.
1863                  */
1864                 if (next_unlinked_offset < reg_buf_offset) {
1865                         continue;
1866                 }
1867
1868                 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1869                 ASSERT((item->ri_buf[item_index].i_len % XFS_BLI_CHUNK) == 0);
1870                 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1871
1872                 /*
1873                  * The current logged region contains a copy of the
1874                  * current di_next_unlinked field.  Extract its value
1875                  * and copy it to the buffer copy.
1876                  */
1877                 logged_nextp = (xfs_agino_t *)
1878                                ((char *)(item->ri_buf[item_index].i_addr) +
1879                                 (next_unlinked_offset - reg_buf_offset));
1880                 if (unlikely(*logged_nextp == 0)) {
1881                         xfs_fs_cmn_err(CE_ALERT, mp,
1882                                 "bad inode buffer log record (ptr = 0x%p, bp = 0x%p).  XFS trying to replay bad (0) inode di_next_unlinked field",
1883                                 item, bp);
1884                         XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1885                                          XFS_ERRLEVEL_LOW, mp);
1886                         return XFS_ERROR(EFSCORRUPTED);
1887                 }
1888
1889                 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1890                                               next_unlinked_offset);
1891                 *buffer_nextp = *logged_nextp;
1892         }
1893
1894         return 0;
1895 }
1896
1897 /*
1898  * Perform a 'normal' buffer recovery.  Each logged region of the
1899  * buffer should be copied over the corresponding region in the
1900  * given buffer.  The bitmap in the buf log format structure indicates
1901  * where to place the logged data.
1902  */
1903 /*ARGSUSED*/
1904 STATIC void
1905 xlog_recover_do_reg_buffer(
1906         struct xfs_mount        *mp,
1907         xlog_recover_item_t     *item,
1908         xfs_buf_t               *bp,
1909         xfs_buf_log_format_t    *buf_f)
1910 {
1911         int                     i;
1912         int                     bit;
1913         int                     nbits;
1914         unsigned int            *data_map = NULL;
1915         unsigned int            map_size = 0;
1916         int                     error;
1917
1918         trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
1919
1920         switch (buf_f->blf_type) {
1921         case XFS_LI_BUF:
1922                 data_map = buf_f->blf_data_map;
1923                 map_size = buf_f->blf_map_size;
1924                 break;
1925         }
1926         bit = 0;
1927         i = 1;  /* 0 is the buf format structure */
1928         while (1) {
1929                 bit = xfs_next_bit(data_map, map_size, bit);
1930                 if (bit == -1)
1931                         break;
1932                 nbits = xfs_contig_bits(data_map, map_size, bit);
1933                 ASSERT(nbits > 0);
1934                 ASSERT(item->ri_buf[i].i_addr != NULL);
1935                 ASSERT(item->ri_buf[i].i_len % XFS_BLI_CHUNK == 0);
1936                 ASSERT(XFS_BUF_COUNT(bp) >=
1937                        ((uint)bit << XFS_BLI_SHIFT)+(nbits<<XFS_BLI_SHIFT));
1938
1939                 /*
1940                  * Do a sanity check if this is a dquot buffer. Just checking
1941                  * the first dquot in the buffer should do. XXXThis is
1942                  * probably a good thing to do for other buf types also.
1943                  */
1944                 error = 0;
1945                 if (buf_f->blf_flags &
1946                    (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
1947                         if (item->ri_buf[i].i_addr == NULL) {
1948                                 cmn_err(CE_ALERT,
1949                                         "XFS: NULL dquot in %s.", __func__);
1950                                 goto next;
1951                         }
1952                         if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
1953                                 cmn_err(CE_ALERT,
1954                                         "XFS: dquot too small (%d) in %s.",
1955                                         item->ri_buf[i].i_len, __func__);
1956                                 goto next;
1957                         }
1958                         error = xfs_qm_dqcheck((xfs_disk_dquot_t *)
1959                                                item->ri_buf[i].i_addr,
1960                                                -1, 0, XFS_QMOPT_DOWARN,
1961                                                "dquot_buf_recover");
1962                         if (error)
1963                                 goto next;
1964                 }
1965
1966                 memcpy(xfs_buf_offset(bp,
1967                         (uint)bit << XFS_BLI_SHIFT),    /* dest */
1968                         item->ri_buf[i].i_addr,         /* source */
1969                         nbits<<XFS_BLI_SHIFT);          /* length */
1970  next:
1971                 i++;
1972                 bit += nbits;
1973         }
1974
1975         /* Shouldn't be any more regions */
1976         ASSERT(i == item->ri_total);
1977 }
1978
1979 /*
1980  * Do some primitive error checking on ondisk dquot data structures.
1981  */
1982 int
1983 xfs_qm_dqcheck(
1984         xfs_disk_dquot_t *ddq,
1985         xfs_dqid_t       id,
1986         uint             type,    /* used only when IO_dorepair is true */
1987         uint             flags,
1988         char             *str)
1989 {
1990         xfs_dqblk_t      *d = (xfs_dqblk_t *)ddq;
1991         int             errs = 0;
1992
1993         /*
1994          * We can encounter an uninitialized dquot buffer for 2 reasons:
1995          * 1. If we crash while deleting the quotainode(s), and those blks got
1996          *    used for user data. This is because we take the path of regular
1997          *    file deletion; however, the size field of quotainodes is never
1998          *    updated, so all the tricks that we play in itruncate_finish
1999          *    don't quite matter.
2000          *
2001          * 2. We don't play the quota buffers when there's a quotaoff logitem.
2002          *    But the allocation will be replayed so we'll end up with an
2003          *    uninitialized quota block.
2004          *
2005          * This is all fine; things are still consistent, and we haven't lost
2006          * any quota information. Just don't complain about bad dquot blks.
2007          */
2008         if (be16_to_cpu(ddq->d_magic) != XFS_DQUOT_MAGIC) {
2009                 if (flags & XFS_QMOPT_DOWARN)
2010                         cmn_err(CE_ALERT,
2011                         "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
2012                         str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
2013                 errs++;
2014         }
2015         if (ddq->d_version != XFS_DQUOT_VERSION) {
2016                 if (flags & XFS_QMOPT_DOWARN)
2017                         cmn_err(CE_ALERT,
2018                         "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
2019                         str, id, ddq->d_version, XFS_DQUOT_VERSION);
2020                 errs++;
2021         }
2022
2023         if (ddq->d_flags != XFS_DQ_USER &&
2024             ddq->d_flags != XFS_DQ_PROJ &&
2025             ddq->d_flags != XFS_DQ_GROUP) {
2026                 if (flags & XFS_QMOPT_DOWARN)
2027                         cmn_err(CE_ALERT,
2028                         "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
2029                         str, id, ddq->d_flags);
2030                 errs++;
2031         }
2032
2033         if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
2034                 if (flags & XFS_QMOPT_DOWARN)
2035                         cmn_err(CE_ALERT,
2036                         "%s : ondisk-dquot 0x%p, ID mismatch: "
2037                         "0x%x expected, found id 0x%x",
2038                         str, ddq, id, be32_to_cpu(ddq->d_id));
2039                 errs++;
2040         }
2041
2042         if (!errs && ddq->d_id) {
2043                 if (ddq->d_blk_softlimit &&
2044                     be64_to_cpu(ddq->d_bcount) >=
2045                                 be64_to_cpu(ddq->d_blk_softlimit)) {
2046                         if (!ddq->d_btimer) {
2047                                 if (flags & XFS_QMOPT_DOWARN)
2048                                         cmn_err(CE_ALERT,
2049                                         "%s : Dquot ID 0x%x (0x%p) "
2050                                         "BLK TIMER NOT STARTED",
2051                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2052                                 errs++;
2053                         }
2054                 }
2055                 if (ddq->d_ino_softlimit &&
2056                     be64_to_cpu(ddq->d_icount) >=
2057                                 be64_to_cpu(ddq->d_ino_softlimit)) {
2058                         if (!ddq->d_itimer) {
2059                                 if (flags & XFS_QMOPT_DOWARN)
2060                                         cmn_err(CE_ALERT,
2061                                         "%s : Dquot ID 0x%x (0x%p) "
2062                                         "INODE TIMER NOT STARTED",
2063                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2064                                 errs++;
2065                         }
2066                 }
2067                 if (ddq->d_rtb_softlimit &&
2068                     be64_to_cpu(ddq->d_rtbcount) >=
2069                                 be64_to_cpu(ddq->d_rtb_softlimit)) {
2070                         if (!ddq->d_rtbtimer) {
2071                                 if (flags & XFS_QMOPT_DOWARN)
2072                                         cmn_err(CE_ALERT,
2073                                         "%s : Dquot ID 0x%x (0x%p) "
2074                                         "RTBLK TIMER NOT STARTED",
2075                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2076                                 errs++;
2077                         }
2078                 }
2079         }
2080
2081         if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2082                 return errs;
2083
2084         if (flags & XFS_QMOPT_DOWARN)
2085                 cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2086
2087         /*
2088          * Typically, a repair is only requested by quotacheck.
2089          */
2090         ASSERT(id != -1);
2091         ASSERT(flags & XFS_QMOPT_DQREPAIR);
2092         memset(d, 0, sizeof(xfs_dqblk_t));
2093
2094         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2095         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2096         d->dd_diskdq.d_flags = type;
2097         d->dd_diskdq.d_id = cpu_to_be32(id);
2098
2099         return errs;
2100 }
2101
2102 /*
2103  * Perform a dquot buffer recovery.
2104  * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2105  * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2106  * Else, treat it as a regular buffer and do recovery.
2107  */
2108 STATIC void
2109 xlog_recover_do_dquot_buffer(
2110         xfs_mount_t             *mp,
2111         xlog_t                  *log,
2112         xlog_recover_item_t     *item,
2113         xfs_buf_t               *bp,
2114         xfs_buf_log_format_t    *buf_f)
2115 {
2116         uint                    type;
2117
2118         trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2119
2120         /*
2121          * Filesystems are required to send in quota flags at mount time.
2122          */
2123         if (mp->m_qflags == 0) {
2124                 return;
2125         }
2126
2127         type = 0;
2128         if (buf_f->blf_flags & XFS_BLI_UDQUOT_BUF)
2129                 type |= XFS_DQ_USER;
2130         if (buf_f->blf_flags & XFS_BLI_PDQUOT_BUF)
2131                 type |= XFS_DQ_PROJ;
2132         if (buf_f->blf_flags & XFS_BLI_GDQUOT_BUF)
2133                 type |= XFS_DQ_GROUP;
2134         /*
2135          * This type of quotas was turned off, so ignore this buffer
2136          */
2137         if (log->l_quotaoffs_flag & type)
2138                 return;
2139
2140         xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2141 }
2142
2143 /*
2144  * This routine replays a modification made to a buffer at runtime.
2145  * There are actually two types of buffer, regular and inode, which
2146  * are handled differently.  Inode buffers are handled differently
2147  * in that we only recover a specific set of data from them, namely
2148  * the inode di_next_unlinked fields.  This is because all other inode
2149  * data is actually logged via inode records and any data we replay
2150  * here which overlaps that may be stale.
2151  *
2152  * When meta-data buffers are freed at run time we log a buffer item
2153  * with the XFS_BLI_CANCEL bit set to indicate that previous copies
2154  * of the buffer in the log should not be replayed at recovery time.
2155  * This is so that if the blocks covered by the buffer are reused for
2156  * file data before we crash we don't end up replaying old, freed
2157  * meta-data into a user's file.
2158  *
2159  * To handle the cancellation of buffer log items, we make two passes
2160  * over the log during recovery.  During the first we build a table of
2161  * those buffers which have been cancelled, and during the second we
2162  * only replay those buffers which do not have corresponding cancel
2163  * records in the table.  See xlog_recover_do_buffer_pass[1,2] above
2164  * for more details on the implementation of the table of cancel records.
2165  */
2166 STATIC int
2167 xlog_recover_do_buffer_trans(
2168         xlog_t                  *log,
2169         xlog_recover_item_t     *item,
2170         int                     pass)
2171 {
2172         xfs_buf_log_format_t    *buf_f;
2173         xfs_mount_t             *mp;
2174         xfs_buf_t               *bp;
2175         int                     error;
2176         int                     cancel;
2177         xfs_daddr_t             blkno;
2178         int                     len;
2179         ushort                  flags;
2180         uint                    buf_flags;
2181
2182         buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
2183
2184         if (pass == XLOG_RECOVER_PASS1) {
2185                 /*
2186                  * In this pass we're only looking for buf items
2187                  * with the XFS_BLI_CANCEL bit set.
2188                  */
2189                 xlog_recover_do_buffer_pass1(log, buf_f);
2190                 return 0;
2191         } else {
2192                 /*
2193                  * In this pass we want to recover all the buffers
2194                  * which have not been cancelled and are not
2195                  * cancellation buffers themselves.  The routine
2196                  * we call here will tell us whether or not to
2197                  * continue with the replay of this buffer.
2198                  */
2199                 cancel = xlog_recover_do_buffer_pass2(log, buf_f);
2200                 if (cancel) {
2201                         trace_xfs_log_recover_buf_cancel(log, buf_f);
2202                         return 0;
2203                 }
2204         }
2205         trace_xfs_log_recover_buf_recover(log, buf_f);
2206         switch (buf_f->blf_type) {
2207         case XFS_LI_BUF:
2208                 blkno = buf_f->blf_blkno;
2209                 len = buf_f->blf_len;
2210                 flags = buf_f->blf_flags;
2211                 break;
2212         default:
2213                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
2214                         "xfs_log_recover: unknown buffer type 0x%x, logdev %s",
2215                         buf_f->blf_type, log->l_mp->m_logname ?
2216                         log->l_mp->m_logname : "internal");
2217                 XFS_ERROR_REPORT("xlog_recover_do_buffer_trans",
2218                                  XFS_ERRLEVEL_LOW, log->l_mp);
2219                 return XFS_ERROR(EFSCORRUPTED);
2220         }
2221
2222         mp = log->l_mp;
2223         buf_flags = XBF_LOCK;
2224         if (!(flags & XFS_BLI_INODE_BUF))
2225                 buf_flags |= XBF_MAPPED;
2226
2227         bp = xfs_buf_read(mp->m_ddev_targp, blkno, len, buf_flags);
2228         if (XFS_BUF_ISERROR(bp)) {
2229                 xfs_ioerror_alert("xlog_recover_do..(read#1)", log->l_mp,
2230                                   bp, blkno);
2231                 error = XFS_BUF_GETERROR(bp);
2232                 xfs_buf_relse(bp);
2233                 return error;
2234         }
2235
2236         error = 0;
2237         if (flags & XFS_BLI_INODE_BUF) {
2238                 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2239         } else if (flags &
2240                   (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
2241                 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2242         } else {
2243                 xlog_recover_do_reg_buffer(mp, item, bp, buf_f);
2244         }
2245         if (error)
2246                 return XFS_ERROR(error);
2247
2248         /*
2249          * Perform delayed write on the buffer.  Asynchronous writes will be
2250          * slower when taking into account all the buffers to be flushed.
2251          *
2252          * Also make sure that only inode buffers with good sizes stay in
2253          * the buffer cache.  The kernel moves inodes in buffers of 1 block
2254          * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger.  The inode
2255          * buffers in the log can be a different size if the log was generated
2256          * by an older kernel using unclustered inode buffers or a newer kernel
2257          * running with a different inode cluster size.  Regardless, if the
2258          * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2259          * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2260          * the buffer out of the buffer cache so that the buffer won't
2261          * overlap with future reads of those inodes.
2262          */
2263         if (XFS_DINODE_MAGIC ==
2264             be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2265             (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2266                         (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2267                 XFS_BUF_STALE(bp);
2268                 error = xfs_bwrite(mp, bp);
2269         } else {
2270                 ASSERT(bp->b_mount == NULL || bp->b_mount == mp);
2271                 bp->b_mount = mp;
2272                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2273                 xfs_bdwrite(mp, bp);
2274         }
2275
2276         return (error);
2277 }
2278
2279 STATIC int
2280 xlog_recover_do_inode_trans(
2281         xlog_t                  *log,
2282         xlog_recover_item_t     *item,
2283         int                     pass)
2284 {
2285         xfs_inode_log_format_t  *in_f;
2286         xfs_mount_t             *mp;
2287         xfs_buf_t               *bp;
2288         xfs_dinode_t            *dip;
2289         xfs_ino_t               ino;
2290         int                     len;
2291         xfs_caddr_t             src;
2292         xfs_caddr_t             dest;
2293         int                     error;
2294         int                     attr_index;
2295         uint                    fields;
2296         xfs_icdinode_t          *dicp;
2297         int                     need_free = 0;
2298
2299         if (pass == XLOG_RECOVER_PASS1) {
2300                 return 0;
2301         }
2302
2303         if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2304                 in_f = (xfs_inode_log_format_t *)item->ri_buf[0].i_addr;
2305         } else {
2306                 in_f = (xfs_inode_log_format_t *)kmem_alloc(
2307                         sizeof(xfs_inode_log_format_t), KM_SLEEP);
2308                 need_free = 1;
2309                 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2310                 if (error)
2311                         goto error;
2312         }
2313         ino = in_f->ilf_ino;
2314         mp = log->l_mp;
2315
2316         /*
2317          * Inode buffers can be freed, look out for it,
2318          * and do not replay the inode.
2319          */
2320         if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
2321                                         in_f->ilf_len, 0)) {
2322                 error = 0;
2323                 trace_xfs_log_recover_inode_cancel(log, in_f);
2324                 goto error;
2325         }
2326         trace_xfs_log_recover_inode_recover(log, in_f);
2327
2328         bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len,
2329                           XBF_LOCK);
2330         if (XFS_BUF_ISERROR(bp)) {
2331                 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2332                                   bp, in_f->ilf_blkno);
2333                 error = XFS_BUF_GETERROR(bp);
2334                 xfs_buf_relse(bp);
2335                 goto error;
2336         }
2337         error = 0;
2338         ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2339         dip = (xfs_dinode_t *)xfs_buf_offset(bp, in_f->ilf_boffset);
2340
2341         /*
2342          * Make sure the place we're flushing out to really looks
2343          * like an inode!
2344          */
2345         if (unlikely(be16_to_cpu(dip->di_magic) != XFS_DINODE_MAGIC)) {
2346                 xfs_buf_relse(bp);
2347                 xfs_fs_cmn_err(CE_ALERT, mp,
2348                         "xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2349                         dip, bp, ino);
2350                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(1)",
2351                                  XFS_ERRLEVEL_LOW, mp);
2352                 error = EFSCORRUPTED;
2353                 goto error;
2354         }
2355         dicp = (xfs_icdinode_t *)(item->ri_buf[1].i_addr);
2356         if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2357                 xfs_buf_relse(bp);
2358                 xfs_fs_cmn_err(CE_ALERT, mp,
2359                         "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2360                         item, ino);
2361                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(2)",
2362                                  XFS_ERRLEVEL_LOW, mp);
2363                 error = EFSCORRUPTED;
2364                 goto error;
2365         }
2366
2367         /* Skip replay when the on disk inode is newer than the log one */
2368         if (dicp->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
2369                 /*
2370                  * Deal with the wrap case, DI_MAX_FLUSH is less
2371                  * than smaller numbers
2372                  */
2373                 if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
2374                     dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2375                         /* do nothing */
2376                 } else {
2377                         xfs_buf_relse(bp);
2378                         trace_xfs_log_recover_inode_skip(log, in_f);
2379                         error = 0;
2380                         goto error;
2381                 }
2382         }
2383         /* Take the opportunity to reset the flush iteration count */
2384         dicp->di_flushiter = 0;
2385
2386         if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2387                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2388                     (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2389                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(3)",
2390                                          XFS_ERRLEVEL_LOW, mp, dicp);
2391                         xfs_buf_relse(bp);
2392                         xfs_fs_cmn_err(CE_ALERT, mp,
2393                                 "xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2394                                 item, dip, bp, ino);
2395                         error = EFSCORRUPTED;
2396                         goto error;
2397                 }
2398         } else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2399                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2400                     (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2401                     (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2402                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(4)",
2403                                              XFS_ERRLEVEL_LOW, mp, dicp);
2404                         xfs_buf_relse(bp);
2405                         xfs_fs_cmn_err(CE_ALERT, mp,
2406                                 "xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2407                                 item, dip, bp, ino);
2408                         error = EFSCORRUPTED;
2409                         goto error;
2410                 }
2411         }
2412         if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2413                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(5)",
2414                                      XFS_ERRLEVEL_LOW, mp, dicp);
2415                 xfs_buf_relse(bp);
2416                 xfs_fs_cmn_err(CE_ALERT, mp,
2417                         "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",
2418                         item, dip, bp, ino,
2419                         dicp->di_nextents + dicp->di_anextents,
2420                         dicp->di_nblocks);
2421                 error = EFSCORRUPTED;
2422                 goto error;
2423         }
2424         if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2425                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(6)",
2426                                      XFS_ERRLEVEL_LOW, mp, dicp);
2427                 xfs_buf_relse(bp);
2428                 xfs_fs_cmn_err(CE_ALERT, mp,
2429                         "xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2430                         item, dip, bp, ino, dicp->di_forkoff);
2431                 error = EFSCORRUPTED;
2432                 goto error;
2433         }
2434         if (unlikely(item->ri_buf[1].i_len > sizeof(struct xfs_icdinode))) {
2435                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(7)",
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 record length %d, rec ptr 0x%p",
2440                         item->ri_buf[1].i_len, item);
2441                 error = EFSCORRUPTED;
2442                 goto error;
2443         }
2444
2445         /* The core is in in-core format */
2446         xfs_dinode_to_disk(dip, (xfs_icdinode_t *)item->ri_buf[1].i_addr);
2447
2448         /* the rest is in on-disk format */
2449         if (item->ri_buf[1].i_len > sizeof(struct xfs_icdinode)) {
2450                 memcpy((xfs_caddr_t) dip + sizeof(struct xfs_icdinode),
2451                         item->ri_buf[1].i_addr + sizeof(struct xfs_icdinode),
2452                         item->ri_buf[1].i_len  - sizeof(struct xfs_icdinode));
2453         }
2454
2455         fields = in_f->ilf_fields;
2456         switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2457         case XFS_ILOG_DEV:
2458                 xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
2459                 break;
2460         case XFS_ILOG_UUID:
2461                 memcpy(XFS_DFORK_DPTR(dip),
2462                        &in_f->ilf_u.ilfu_uuid,
2463                        sizeof(uuid_t));
2464                 break;
2465         }
2466
2467         if (in_f->ilf_size == 2)
2468                 goto write_inode_buffer;
2469         len = item->ri_buf[2].i_len;
2470         src = item->ri_buf[2].i_addr;
2471         ASSERT(in_f->ilf_size <= 4);
2472         ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2473         ASSERT(!(fields & XFS_ILOG_DFORK) ||
2474                (len == in_f->ilf_dsize));
2475
2476         switch (fields & XFS_ILOG_DFORK) {
2477         case XFS_ILOG_DDATA:
2478         case XFS_ILOG_DEXT:
2479                 memcpy(XFS_DFORK_DPTR(dip), src, len);
2480                 break;
2481
2482         case XFS_ILOG_DBROOT:
2483                 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
2484                                  (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
2485                                  XFS_DFORK_DSIZE(dip, mp));
2486                 break;
2487
2488         default:
2489                 /*
2490                  * There are no data fork flags set.
2491                  */
2492                 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2493                 break;
2494         }
2495
2496         /*
2497          * If we logged any attribute data, recover it.  There may or
2498          * may not have been any other non-core data logged in this
2499          * transaction.
2500          */
2501         if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2502                 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2503                         attr_index = 3;
2504                 } else {
2505                         attr_index = 2;
2506                 }
2507                 len = item->ri_buf[attr_index].i_len;
2508                 src = item->ri_buf[attr_index].i_addr;
2509                 ASSERT(len == in_f->ilf_asize);
2510
2511                 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2512                 case XFS_ILOG_ADATA:
2513                 case XFS_ILOG_AEXT:
2514                         dest = XFS_DFORK_APTR(dip);
2515                         ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2516                         memcpy(dest, src, len);
2517                         break;
2518
2519                 case XFS_ILOG_ABROOT:
2520                         dest = XFS_DFORK_APTR(dip);
2521                         xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
2522                                          len, (xfs_bmdr_block_t*)dest,
2523                                          XFS_DFORK_ASIZE(dip, mp));
2524                         break;
2525
2526                 default:
2527                         xlog_warn("XFS: xlog_recover_do_inode_trans: Invalid flag");
2528                         ASSERT(0);
2529                         xfs_buf_relse(bp);
2530                         error = EIO;
2531                         goto error;
2532                 }
2533         }
2534
2535 write_inode_buffer:
2536         ASSERT(bp->b_mount == NULL || bp->b_mount == mp);
2537         bp->b_mount = mp;
2538         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2539         xfs_bdwrite(mp, bp);
2540 error:
2541         if (need_free)
2542                 kmem_free(in_f);
2543         return XFS_ERROR(error);
2544 }
2545
2546 /*
2547  * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2548  * structure, so that we know not to do any dquot item or dquot buffer recovery,
2549  * of that type.
2550  */
2551 STATIC int
2552 xlog_recover_do_quotaoff_trans(
2553         xlog_t                  *log,
2554         xlog_recover_item_t     *item,
2555         int                     pass)
2556 {
2557         xfs_qoff_logformat_t    *qoff_f;
2558
2559         if (pass == XLOG_RECOVER_PASS2) {
2560                 return (0);
2561         }
2562
2563         qoff_f = (xfs_qoff_logformat_t *)item->ri_buf[0].i_addr;
2564         ASSERT(qoff_f);
2565
2566         /*
2567          * The logitem format's flag tells us if this was user quotaoff,
2568          * group/project quotaoff or both.
2569          */
2570         if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2571                 log->l_quotaoffs_flag |= XFS_DQ_USER;
2572         if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2573                 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2574         if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2575                 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2576
2577         return (0);
2578 }
2579
2580 /*
2581  * Recover a dquot record
2582  */
2583 STATIC int
2584 xlog_recover_do_dquot_trans(
2585         xlog_t                  *log,
2586         xlog_recover_item_t     *item,
2587         int                     pass)
2588 {
2589         xfs_mount_t             *mp;
2590         xfs_buf_t               *bp;
2591         struct xfs_disk_dquot   *ddq, *recddq;
2592         int                     error;
2593         xfs_dq_logformat_t      *dq_f;
2594         uint                    type;
2595
2596         if (pass == XLOG_RECOVER_PASS1) {
2597                 return 0;
2598         }
2599         mp = log->l_mp;
2600
2601         /*
2602          * Filesystems are required to send in quota flags at mount time.
2603          */
2604         if (mp->m_qflags == 0)
2605                 return (0);
2606
2607         recddq = (xfs_disk_dquot_t *)item->ri_buf[1].i_addr;
2608
2609         if (item->ri_buf[1].i_addr == NULL) {
2610                 cmn_err(CE_ALERT,
2611                         "XFS: NULL dquot in %s.", __func__);
2612                 return XFS_ERROR(EIO);
2613         }
2614         if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
2615                 cmn_err(CE_ALERT,
2616                         "XFS: dquot too small (%d) in %s.",
2617                         item->ri_buf[1].i_len, __func__);
2618                 return XFS_ERROR(EIO);
2619         }
2620
2621         /*
2622          * This type of quotas was turned off, so ignore this record.
2623          */
2624         type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2625         ASSERT(type);
2626         if (log->l_quotaoffs_flag & type)
2627                 return (0);
2628
2629         /*
2630          * At this point we know that quota was _not_ turned off.
2631          * Since the mount flags are not indicating to us otherwise, this
2632          * must mean that quota is on, and the dquot needs to be replayed.
2633          * Remember that we may not have fully recovered the superblock yet,
2634          * so we can't do the usual trick of looking at the SB quota bits.
2635          *
2636          * The other possibility, of course, is that the quota subsystem was
2637          * removed since the last mount - ENOSYS.
2638          */
2639         dq_f = (xfs_dq_logformat_t *)item->ri_buf[0].i_addr;
2640         ASSERT(dq_f);
2641         if ((error = xfs_qm_dqcheck(recddq,
2642                            dq_f->qlf_id,
2643                            0, XFS_QMOPT_DOWARN,
2644                            "xlog_recover_do_dquot_trans (log copy)"))) {
2645                 return XFS_ERROR(EIO);
2646         }
2647         ASSERT(dq_f->qlf_len == 1);
2648
2649         error = xfs_read_buf(mp, mp->m_ddev_targp,
2650                              dq_f->qlf_blkno,
2651                              XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2652                              0, &bp);
2653         if (error) {
2654                 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2655                                   bp, dq_f->qlf_blkno);
2656                 return error;
2657         }
2658         ASSERT(bp);
2659         ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2660
2661         /*
2662          * At least the magic num portion should be on disk because this
2663          * was among a chunk of dquots created earlier, and we did some
2664          * minimal initialization then.
2665          */
2666         if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2667                            "xlog_recover_do_dquot_trans")) {
2668                 xfs_buf_relse(bp);
2669                 return XFS_ERROR(EIO);
2670         }
2671
2672         memcpy(ddq, recddq, item->ri_buf[1].i_len);
2673
2674         ASSERT(dq_f->qlf_size == 2);
2675         ASSERT(bp->b_mount == NULL || bp->b_mount == mp);
2676         bp->b_mount = mp;
2677         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2678         xfs_bdwrite(mp, bp);
2679
2680         return (0);
2681 }
2682
2683 /*
2684  * This routine is called to create an in-core extent free intent
2685  * item from the efi format structure which was logged on disk.
2686  * It allocates an in-core efi, copies the extents from the format
2687  * structure into it, and adds the efi to the AIL with the given
2688  * LSN.
2689  */
2690 STATIC int
2691 xlog_recover_do_efi_trans(
2692         xlog_t                  *log,
2693         xlog_recover_item_t     *item,
2694         xfs_lsn_t               lsn,
2695         int                     pass)
2696 {
2697         int                     error;
2698         xfs_mount_t             *mp;
2699         xfs_efi_log_item_t      *efip;
2700         xfs_efi_log_format_t    *efi_formatp;
2701
2702         if (pass == XLOG_RECOVER_PASS1) {
2703                 return 0;
2704         }
2705
2706         efi_formatp = (xfs_efi_log_format_t *)item->ri_buf[0].i_addr;
2707
2708         mp = log->l_mp;
2709         efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2710         if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2711                                          &(efip->efi_format)))) {
2712                 xfs_efi_item_free(efip);
2713                 return error;
2714         }
2715         efip->efi_next_extent = efi_formatp->efi_nextents;
2716         efip->efi_flags |= XFS_EFI_COMMITTED;
2717
2718         spin_lock(&log->l_ailp->xa_lock);
2719         /*
2720          * xfs_trans_ail_update() drops the AIL lock.
2721          */
2722         xfs_trans_ail_update(log->l_ailp, (xfs_log_item_t *)efip, lsn);
2723         return 0;
2724 }
2725
2726
2727 /*
2728  * This routine is called when an efd format structure is found in
2729  * a committed transaction in the log.  It's purpose is to cancel
2730  * the corresponding efi if it was still in the log.  To do this
2731  * it searches the AIL for the efi with an id equal to that in the
2732  * efd format structure.  If we find it, we remove the efi from the
2733  * AIL and free it.
2734  */
2735 STATIC void
2736 xlog_recover_do_efd_trans(
2737         xlog_t                  *log,
2738         xlog_recover_item_t     *item,
2739         int                     pass)
2740 {
2741         xfs_efd_log_format_t    *efd_formatp;
2742         xfs_efi_log_item_t      *efip = NULL;
2743         xfs_log_item_t          *lip;
2744         __uint64_t              efi_id;
2745         struct xfs_ail_cursor   cur;
2746         struct xfs_ail          *ailp = log->l_ailp;
2747
2748         if (pass == XLOG_RECOVER_PASS1) {
2749                 return;
2750         }
2751
2752         efd_formatp = (xfs_efd_log_format_t *)item->ri_buf[0].i_addr;
2753         ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2754                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2755                (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2756                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2757         efi_id = efd_formatp->efd_efi_id;
2758
2759         /*
2760          * Search for the efi with the id in the efd format structure
2761          * in the AIL.
2762          */
2763         spin_lock(&ailp->xa_lock);
2764         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
2765         while (lip != NULL) {
2766                 if (lip->li_type == XFS_LI_EFI) {
2767                         efip = (xfs_efi_log_item_t *)lip;
2768                         if (efip->efi_format.efi_id == efi_id) {
2769                                 /*
2770                                  * xfs_trans_ail_delete() drops the
2771                                  * AIL lock.
2772                                  */
2773                                 xfs_trans_ail_delete(ailp, lip);
2774                                 xfs_efi_item_free(efip);
2775                                 spin_lock(&ailp->xa_lock);
2776                                 break;
2777                         }
2778                 }
2779                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
2780         }
2781         xfs_trans_ail_cursor_done(ailp, &cur);
2782         spin_unlock(&ailp->xa_lock);
2783 }
2784
2785 /*
2786  * Perform the transaction
2787  *
2788  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2789  * EFIs and EFDs get queued up by adding entries into the AIL for them.
2790  */
2791 STATIC int
2792 xlog_recover_do_trans(
2793         xlog_t                  *log,
2794         xlog_recover_t          *trans,
2795         int                     pass)
2796 {
2797         int                     error = 0;
2798         xlog_recover_item_t     *item;
2799
2800         error = xlog_recover_reorder_trans(log, trans, pass);
2801         if (error)
2802                 return error;
2803
2804         list_for_each_entry(item, &trans->r_itemq, ri_list) {
2805                 trace_xfs_log_recover_item_recover(log, trans, item, pass);
2806                 switch (ITEM_TYPE(item)) {
2807                 case XFS_LI_BUF:
2808                         error = xlog_recover_do_buffer_trans(log, item, pass);
2809                         break;
2810                 case XFS_LI_INODE:
2811                         error = xlog_recover_do_inode_trans(log, item, pass);
2812                         break;
2813                 case XFS_LI_EFI:
2814                         error = xlog_recover_do_efi_trans(log, item,
2815                                                           trans->r_lsn, pass);
2816                         break;
2817                 case XFS_LI_EFD:
2818                         xlog_recover_do_efd_trans(log, item, pass);
2819                         error = 0;
2820                         break;
2821                 case XFS_LI_DQUOT:
2822                         error = xlog_recover_do_dquot_trans(log, item, pass);
2823                         break;
2824                 case XFS_LI_QUOTAOFF:
2825                         error = xlog_recover_do_quotaoff_trans(log, item,
2826                                                                pass);
2827                         break;
2828                 default:
2829                         xlog_warn(
2830         "XFS: invalid item type (%d) xlog_recover_do_trans", ITEM_TYPE(item));
2831                         ASSERT(0);
2832                         error = XFS_ERROR(EIO);
2833                         break;
2834                 }
2835
2836                 if (error)
2837                         return error;
2838         }
2839
2840         return 0;
2841 }
2842
2843 /*
2844  * Free up any resources allocated by the transaction
2845  *
2846  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2847  */
2848 STATIC void
2849 xlog_recover_free_trans(
2850         xlog_recover_t          *trans)
2851 {
2852         xlog_recover_item_t     *item, *n;
2853         int                     i;
2854
2855         list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
2856                 /* Free the regions in the item. */
2857                 list_del(&item->ri_list);
2858                 for (i = 0; i < item->ri_cnt; i++)
2859                         kmem_free(item->ri_buf[i].i_addr);
2860                 /* Free the item itself */
2861                 kmem_free(item->ri_buf);
2862                 kmem_free(item);
2863         }
2864         /* Free the transaction recover structure */
2865         kmem_free(trans);
2866 }
2867
2868 STATIC int
2869 xlog_recover_commit_trans(
2870         xlog_t                  *log,
2871         xlog_recover_t          *trans,
2872         int                     pass)
2873 {
2874         int                     error;
2875
2876         hlist_del(&trans->r_list);
2877         if ((error = xlog_recover_do_trans(log, trans, pass)))
2878                 return error;
2879         xlog_recover_free_trans(trans);                 /* no error */
2880         return 0;
2881 }
2882
2883 STATIC int
2884 xlog_recover_unmount_trans(
2885         xlog_recover_t          *trans)
2886 {
2887         /* Do nothing now */
2888         xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2889         return 0;
2890 }
2891
2892 /*
2893  * There are two valid states of the r_state field.  0 indicates that the
2894  * transaction structure is in a normal state.  We have either seen the
2895  * start of the transaction or the last operation we added was not a partial
2896  * operation.  If the last operation we added to the transaction was a
2897  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2898  *
2899  * NOTE: skip LRs with 0 data length.
2900  */
2901 STATIC int
2902 xlog_recover_process_data(
2903         xlog_t                  *log,
2904         struct hlist_head       rhash[],
2905         xlog_rec_header_t       *rhead,
2906         xfs_caddr_t             dp,
2907         int                     pass)
2908 {
2909         xfs_caddr_t             lp;
2910         int                     num_logops;
2911         xlog_op_header_t        *ohead;
2912         xlog_recover_t          *trans;
2913         xlog_tid_t              tid;
2914         int                     error;
2915         unsigned long           hash;
2916         uint                    flags;
2917
2918         lp = dp + be32_to_cpu(rhead->h_len);
2919         num_logops = be32_to_cpu(rhead->h_num_logops);
2920
2921         /* check the log format matches our own - else we can't recover */
2922         if (xlog_header_check_recover(log->l_mp, rhead))
2923                 return (XFS_ERROR(EIO));
2924
2925         while ((dp < lp) && num_logops) {
2926                 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2927                 ohead = (xlog_op_header_t *)dp;
2928                 dp += sizeof(xlog_op_header_t);
2929                 if (ohead->oh_clientid != XFS_TRANSACTION &&
2930                     ohead->oh_clientid != XFS_LOG) {
2931                         xlog_warn(
2932                 "XFS: xlog_recover_process_data: bad clientid");
2933                         ASSERT(0);
2934                         return (XFS_ERROR(EIO));
2935                 }
2936                 tid = be32_to_cpu(ohead->oh_tid);
2937                 hash = XLOG_RHASH(tid);
2938                 trans = xlog_recover_find_tid(&rhash[hash], tid);
2939                 if (trans == NULL) {               /* not found; add new tid */
2940                         if (ohead->oh_flags & XLOG_START_TRANS)
2941                                 xlog_recover_new_tid(&rhash[hash], tid,
2942                                         be64_to_cpu(rhead->h_lsn));
2943                 } else {
2944                         if (dp + be32_to_cpu(ohead->oh_len) > lp) {
2945                                 xlog_warn(
2946                         "XFS: xlog_recover_process_data: bad length");
2947                                 WARN_ON(1);
2948                                 return (XFS_ERROR(EIO));
2949                         }
2950                         flags = ohead->oh_flags & ~XLOG_END_TRANS;
2951                         if (flags & XLOG_WAS_CONT_TRANS)
2952                                 flags &= ~XLOG_CONTINUE_TRANS;
2953                         switch (flags) {
2954                         case XLOG_COMMIT_TRANS:
2955                                 error = xlog_recover_commit_trans(log,
2956                                                                 trans, pass);
2957                                 break;
2958                         case XLOG_UNMOUNT_TRANS:
2959                                 error = xlog_recover_unmount_trans(trans);
2960                                 break;
2961                         case XLOG_WAS_CONT_TRANS:
2962                                 error = xlog_recover_add_to_cont_trans(log,
2963                                                 trans, dp,
2964                                                 be32_to_cpu(ohead->oh_len));
2965                                 break;
2966                         case XLOG_START_TRANS:
2967                                 xlog_warn(
2968                         "XFS: xlog_recover_process_data: bad transaction");
2969                                 ASSERT(0);
2970                                 error = XFS_ERROR(EIO);
2971                                 break;
2972                         case 0:
2973                         case XLOG_CONTINUE_TRANS:
2974                                 error = xlog_recover_add_to_trans(log, trans,
2975                                                 dp, be32_to_cpu(ohead->oh_len));
2976                                 break;
2977                         default:
2978                                 xlog_warn(
2979                         "XFS: xlog_recover_process_data: bad flag");
2980                                 ASSERT(0);
2981                                 error = XFS_ERROR(EIO);
2982                                 break;
2983                         }
2984                         if (error)
2985                                 return error;
2986                 }
2987                 dp += be32_to_cpu(ohead->oh_len);
2988                 num_logops--;
2989         }
2990         return 0;
2991 }
2992
2993 /*
2994  * Process an extent free intent item that was recovered from
2995  * the log.  We need to free the extents that it describes.
2996  */
2997 STATIC int
2998 xlog_recover_process_efi(
2999         xfs_mount_t             *mp,
3000         xfs_efi_log_item_t      *efip)
3001 {
3002         xfs_efd_log_item_t      *efdp;
3003         xfs_trans_t             *tp;
3004         int                     i;
3005         int                     error = 0;
3006         xfs_extent_t            *extp;
3007         xfs_fsblock_t           startblock_fsb;
3008
3009         ASSERT(!(efip->efi_flags & XFS_EFI_RECOVERED));
3010
3011         /*
3012          * First check the validity of the extents described by the
3013          * EFI.  If any are bad, then assume that all are bad and
3014          * just toss the EFI.
3015          */
3016         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3017                 extp = &(efip->efi_format.efi_extents[i]);
3018                 startblock_fsb = XFS_BB_TO_FSB(mp,
3019                                    XFS_FSB_TO_DADDR(mp, extp->ext_start));
3020                 if ((startblock_fsb == 0) ||
3021                     (extp->ext_len == 0) ||
3022                     (startblock_fsb >= mp->m_sb.sb_dblocks) ||
3023                     (extp->ext_len >= mp->m_sb.sb_agblocks)) {
3024                         /*
3025                          * This will pull the EFI from the AIL and
3026                          * free the memory associated with it.
3027                          */
3028                         xfs_efi_release(efip, efip->efi_format.efi_nextents);
3029                         return XFS_ERROR(EIO);
3030                 }
3031         }
3032
3033         tp = xfs_trans_alloc(mp, 0);
3034         error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
3035         if (error)
3036                 goto abort_error;
3037         efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3038
3039         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3040                 extp = &(efip->efi_format.efi_extents[i]);
3041                 error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3042                 if (error)
3043                         goto abort_error;
3044                 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3045                                          extp->ext_len);
3046         }
3047
3048         efip->efi_flags |= XFS_EFI_RECOVERED;
3049         error = xfs_trans_commit(tp, 0);
3050         return error;
3051
3052 abort_error:
3053         xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3054         return error;
3055 }
3056
3057 /*
3058  * When this is called, all of the EFIs which did not have
3059  * corresponding EFDs should be in the AIL.  What we do now
3060  * is free the extents associated with each one.
3061  *
3062  * Since we process the EFIs in normal transactions, they
3063  * will be removed at some point after the commit.  This prevents
3064  * us from just walking down the list processing each one.
3065  * We'll use a flag in the EFI to skip those that we've already
3066  * processed and use the AIL iteration mechanism's generation
3067  * count to try to speed this up at least a bit.
3068  *
3069  * When we start, we know that the EFIs are the only things in
3070  * the AIL.  As we process them, however, other items are added
3071  * to the AIL.  Since everything added to the AIL must come after
3072  * everything already in the AIL, we stop processing as soon as
3073  * we see something other than an EFI in the AIL.
3074  */
3075 STATIC int
3076 xlog_recover_process_efis(
3077         xlog_t                  *log)
3078 {
3079         xfs_log_item_t          *lip;
3080         xfs_efi_log_item_t      *efip;
3081         int                     error = 0;
3082         struct xfs_ail_cursor   cur;
3083         struct xfs_ail          *ailp;
3084
3085         ailp = log->l_ailp;
3086         spin_lock(&ailp->xa_lock);
3087         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3088         while (lip != NULL) {
3089                 /*
3090                  * We're done when we see something other than an EFI.
3091                  * There should be no EFIs left in the AIL now.
3092                  */
3093                 if (lip->li_type != XFS_LI_EFI) {
3094 #ifdef DEBUG
3095                         for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
3096                                 ASSERT(lip->li_type != XFS_LI_EFI);
3097 #endif
3098                         break;
3099                 }
3100
3101                 /*
3102                  * Skip EFIs that we've already processed.
3103                  */
3104                 efip = (xfs_efi_log_item_t *)lip;
3105                 if (efip->efi_flags & XFS_EFI_RECOVERED) {
3106                         lip = xfs_trans_ail_cursor_next(ailp, &cur);
3107                         continue;
3108                 }
3109
3110                 spin_unlock(&ailp->xa_lock);
3111                 error = xlog_recover_process_efi(log->l_mp, efip);
3112                 spin_lock(&ailp->xa_lock);
3113                 if (error)
3114                         goto out;
3115                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3116         }
3117 out:
3118         xfs_trans_ail_cursor_done(ailp, &cur);
3119         spin_unlock(&ailp->xa_lock);
3120         return error;
3121 }
3122
3123 /*
3124  * This routine performs a transaction to null out a bad inode pointer
3125  * in an agi unlinked inode hash bucket.
3126  */
3127 STATIC void
3128 xlog_recover_clear_agi_bucket(
3129         xfs_mount_t     *mp,
3130         xfs_agnumber_t  agno,
3131         int             bucket)
3132 {
3133         xfs_trans_t     *tp;
3134         xfs_agi_t       *agi;
3135         xfs_buf_t       *agibp;
3136         int             offset;
3137         int             error;
3138
3139         tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3140         error = xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp),
3141                                   0, 0, 0);
3142         if (error)
3143                 goto out_abort;
3144
3145         error = xfs_read_agi(mp, tp, agno, &agibp);
3146         if (error)
3147                 goto out_abort;
3148
3149         agi = XFS_BUF_TO_AGI(agibp);
3150         agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3151         offset = offsetof(xfs_agi_t, agi_unlinked) +
3152                  (sizeof(xfs_agino_t) * bucket);
3153         xfs_trans_log_buf(tp, agibp, offset,
3154                           (offset + sizeof(xfs_agino_t) - 1));
3155
3156         error = xfs_trans_commit(tp, 0);
3157         if (error)
3158                 goto out_error;
3159         return;
3160
3161 out_abort:
3162         xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3163 out_error:
3164         xfs_fs_cmn_err(CE_WARN, mp, "xlog_recover_clear_agi_bucket: "
3165                         "failed to clear agi %d. Continuing.", agno);
3166         return;
3167 }
3168
3169 STATIC xfs_agino_t
3170 xlog_recover_process_one_iunlink(
3171         struct xfs_mount                *mp,
3172         xfs_agnumber_t                  agno,
3173         xfs_agino_t                     agino,
3174         int                             bucket)
3175 {
3176         struct xfs_buf                  *ibp;
3177         struct xfs_dinode               *dip;
3178         struct xfs_inode                *ip;
3179         xfs_ino_t                       ino;
3180         int                             error;
3181
3182         ino = XFS_AGINO_TO_INO(mp, agno, agino);
3183         error = xfs_iget(mp, NULL, ino, 0, 0, &ip, 0);
3184         if (error)
3185                 goto fail;
3186
3187         /*
3188          * Get the on disk inode to find the next inode in the bucket.
3189          */
3190         error = xfs_itobp(mp, NULL, ip, &dip, &ibp, XBF_LOCK);
3191         if (error)
3192                 goto fail_iput;
3193
3194         ASSERT(ip->i_d.di_nlink == 0);
3195         ASSERT(ip->i_d.di_mode != 0);
3196
3197         /* setup for the next pass */
3198         agino = be32_to_cpu(dip->di_next_unlinked);
3199         xfs_buf_relse(ibp);
3200
3201         /*
3202          * Prevent any DMAPI event from being sent when the reference on
3203          * the inode is dropped.
3204          */
3205         ip->i_d.di_dmevmask = 0;
3206
3207         IRELE(ip);
3208         return agino;
3209
3210  fail_iput:
3211         IRELE(ip);
3212  fail:
3213         /*
3214          * We can't read in the inode this bucket points to, or this inode
3215          * is messed up.  Just ditch this bucket of inodes.  We will lose
3216          * some inodes and space, but at least we won't hang.
3217          *
3218          * Call xlog_recover_clear_agi_bucket() to perform a transaction to
3219          * clear the inode pointer in the bucket.
3220          */
3221         xlog_recover_clear_agi_bucket(mp, agno, bucket);
3222         return NULLAGINO;
3223 }
3224
3225 /*
3226  * xlog_iunlink_recover
3227  *
3228  * This is called during recovery to process any inodes which
3229  * we unlinked but not freed when the system crashed.  These
3230  * inodes will be on the lists in the AGI blocks.  What we do
3231  * here is scan all the AGIs and fully truncate and free any
3232  * inodes found on the lists.  Each inode is removed from the
3233  * lists when it has been fully truncated and is freed.  The
3234  * freeing of the inode and its removal from the list must be
3235  * atomic.
3236  */
3237 STATIC void
3238 xlog_recover_process_iunlinks(
3239         xlog_t          *log)
3240 {
3241         xfs_mount_t     *mp;
3242         xfs_agnumber_t  agno;
3243         xfs_agi_t       *agi;
3244         xfs_buf_t       *agibp;
3245         xfs_agino_t     agino;
3246         int             bucket;
3247         int             error;
3248         uint            mp_dmevmask;
3249
3250         mp = log->l_mp;
3251
3252         /*
3253          * Prevent any DMAPI event from being sent while in this function.
3254          */
3255         mp_dmevmask = mp->m_dmevmask;
3256         mp->m_dmevmask = 0;
3257
3258         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3259                 /*
3260                  * Find the agi for this ag.
3261                  */
3262                 error = xfs_read_agi(mp, NULL, agno, &agibp);
3263                 if (error) {
3264                         /*
3265                          * AGI is b0rked. Don't process it.
3266                          *
3267                          * We should probably mark the filesystem as corrupt
3268                          * after we've recovered all the ag's we can....
3269                          */
3270                         continue;
3271                 }
3272                 agi = XFS_BUF_TO_AGI(agibp);
3273
3274                 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3275                         agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3276                         while (agino != NULLAGINO) {
3277                                 /*
3278                                  * Release the agi buffer so that it can
3279                                  * be acquired in the normal course of the
3280                                  * transaction to truncate and free the inode.
3281                                  */
3282                                 xfs_buf_relse(agibp);
3283
3284                                 agino = xlog_recover_process_one_iunlink(mp,
3285                                                         agno, agino, bucket);
3286
3287                                 /*
3288                                  * Reacquire the agibuffer and continue around
3289                                  * the loop. This should never fail as we know
3290                                  * the buffer was good earlier on.
3291                                  */
3292                                 error = xfs_read_agi(mp, NULL, agno, &agibp);
3293                                 ASSERT(error == 0);
3294                                 agi = XFS_BUF_TO_AGI(agibp);
3295                         }
3296                 }
3297
3298                 /*
3299                  * Release the buffer for the current agi so we can
3300                  * go on to the next one.
3301                  */
3302                 xfs_buf_relse(agibp);
3303         }
3304
3305         mp->m_dmevmask = mp_dmevmask;
3306 }
3307
3308
3309 #ifdef DEBUG
3310 STATIC void
3311 xlog_pack_data_checksum(
3312         xlog_t          *log,
3313         xlog_in_core_t  *iclog,
3314         int             size)
3315 {
3316         int             i;
3317         __be32          *up;
3318         uint            chksum = 0;
3319
3320         up = (__be32 *)iclog->ic_datap;
3321         /* divide length by 4 to get # words */
3322         for (i = 0; i < (size >> 2); i++) {
3323                 chksum ^= be32_to_cpu(*up);
3324                 up++;
3325         }
3326         iclog->ic_header.h_chksum = cpu_to_be32(chksum);
3327 }
3328 #else
3329 #define xlog_pack_data_checksum(log, iclog, size)
3330 #endif
3331
3332 /*
3333  * Stamp cycle number in every block
3334  */
3335 void
3336 xlog_pack_data(
3337         xlog_t                  *log,
3338         xlog_in_core_t          *iclog,
3339         int                     roundoff)
3340 {
3341         int                     i, j, k;
3342         int                     size = iclog->ic_offset + roundoff;
3343         __be32                  cycle_lsn;
3344         xfs_caddr_t             dp;
3345
3346         xlog_pack_data_checksum(log, iclog, size);
3347
3348         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3349
3350         dp = iclog->ic_datap;
3351         for (i = 0; i < BTOBB(size) &&
3352                 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3353                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
3354                 *(__be32 *)dp = cycle_lsn;
3355                 dp += BBSIZE;
3356         }
3357
3358         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3359                 xlog_in_core_2_t *xhdr = iclog->ic_data;
3360
3361                 for ( ; i < BTOBB(size); i++) {
3362                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3363                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3364                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
3365                         *(__be32 *)dp = cycle_lsn;
3366                         dp += BBSIZE;
3367                 }
3368
3369                 for (i = 1; i < log->l_iclog_heads; i++) {
3370                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3371                 }
3372         }
3373 }
3374
3375 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
3376 STATIC void
3377 xlog_unpack_data_checksum(
3378         xlog_rec_header_t       *rhead,
3379         xfs_caddr_t             dp,
3380         xlog_t                  *log)
3381 {
3382         __be32                  *up = (__be32 *)dp;
3383         uint                    chksum = 0;
3384         int                     i;
3385
3386         /* divide length by 4 to get # words */
3387         for (i=0; i < be32_to_cpu(rhead->h_len) >> 2; i++) {
3388                 chksum ^= be32_to_cpu(*up);
3389                 up++;
3390         }
3391         if (chksum != be32_to_cpu(rhead->h_chksum)) {
3392             if (rhead->h_chksum ||
3393                 ((log->l_flags & XLOG_CHKSUM_MISMATCH) == 0)) {
3394                     cmn_err(CE_DEBUG,
3395                         "XFS: LogR chksum mismatch: was (0x%x) is (0x%x)\n",
3396                             be32_to_cpu(rhead->h_chksum), chksum);
3397                     cmn_err(CE_DEBUG,
3398 "XFS: Disregard message if filesystem was created with non-DEBUG kernel");
3399                     if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3400                             cmn_err(CE_DEBUG,
3401                                 "XFS: LogR this is a LogV2 filesystem\n");
3402                     }
3403                     log->l_flags |= XLOG_CHKSUM_MISMATCH;
3404             }
3405         }
3406 }
3407 #else
3408 #define xlog_unpack_data_checksum(rhead, dp, log)
3409 #endif
3410
3411 STATIC void
3412 xlog_unpack_data(
3413         xlog_rec_header_t       *rhead,
3414         xfs_caddr_t             dp,
3415         xlog_t                  *log)
3416 {
3417         int                     i, j, k;
3418
3419         for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3420                   i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3421                 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3422                 dp += BBSIZE;
3423         }
3424
3425         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3426                 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
3427                 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3428                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3429                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3430                         *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3431                         dp += BBSIZE;
3432                 }
3433         }
3434
3435         xlog_unpack_data_checksum(rhead, dp, log);
3436 }
3437
3438 STATIC int
3439 xlog_valid_rec_header(
3440         xlog_t                  *log,
3441         xlog_rec_header_t       *rhead,
3442         xfs_daddr_t             blkno)
3443 {
3444         int                     hlen;
3445
3446         if (unlikely(be32_to_cpu(rhead->h_magicno) != XLOG_HEADER_MAGIC_NUM)) {
3447                 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3448                                 XFS_ERRLEVEL_LOW, log->l_mp);
3449                 return XFS_ERROR(EFSCORRUPTED);
3450         }
3451         if (unlikely(
3452             (!rhead->h_version ||
3453             (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
3454                 xlog_warn("XFS: %s: unrecognised log version (%d).",
3455                         __func__, be32_to_cpu(rhead->h_version));
3456                 return XFS_ERROR(EIO);
3457         }
3458
3459         /* LR body must have data or it wouldn't have been written */
3460         hlen = be32_to_cpu(rhead->h_len);
3461         if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3462                 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3463                                 XFS_ERRLEVEL_LOW, log->l_mp);
3464                 return XFS_ERROR(EFSCORRUPTED);
3465         }
3466         if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3467                 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3468                                 XFS_ERRLEVEL_LOW, log->l_mp);
3469                 return XFS_ERROR(EFSCORRUPTED);
3470         }
3471         return 0;
3472 }
3473
3474 /*
3475  * Read the log from tail to head and process the log records found.
3476  * Handle the two cases where the tail and head are in the same cycle
3477  * and where the active portion of the log wraps around the end of
3478  * the physical log separately.  The pass parameter is passed through
3479  * to the routines called to process the data and is not looked at
3480  * here.
3481  */
3482 STATIC int
3483 xlog_do_recovery_pass(
3484         xlog_t                  *log,
3485         xfs_daddr_t             head_blk,
3486         xfs_daddr_t             tail_blk,
3487         int                     pass)
3488 {
3489         xlog_rec_header_t       *rhead;
3490         xfs_daddr_t             blk_no;
3491         xfs_caddr_t             offset;
3492         xfs_buf_t               *hbp, *dbp;
3493         int                     error = 0, h_size;
3494         int                     bblks, split_bblks;
3495         int                     hblks, split_hblks, wrapped_hblks;
3496         struct hlist_head       rhash[XLOG_RHASH_SIZE];
3497
3498         ASSERT(head_blk != tail_blk);
3499
3500         /*
3501          * Read the header of the tail block and get the iclog buffer size from
3502          * h_size.  Use this to tell how many sectors make up the log header.
3503          */
3504         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3505                 /*
3506                  * When using variable length iclogs, read first sector of
3507                  * iclog header and extract the header size from it.  Get a
3508                  * new hbp that is the correct size.
3509                  */
3510                 hbp = xlog_get_bp(log, 1);
3511                 if (!hbp)
3512                         return ENOMEM;
3513
3514                 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
3515                 if (error)
3516                         goto bread_err1;
3517
3518                 rhead = (xlog_rec_header_t *)offset;
3519                 error = xlog_valid_rec_header(log, rhead, tail_blk);
3520                 if (error)
3521                         goto bread_err1;
3522                 h_size = be32_to_cpu(rhead->h_size);
3523                 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3524                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3525                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3526                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
3527                                 hblks++;
3528                         xlog_put_bp(hbp);
3529                         hbp = xlog_get_bp(log, hblks);
3530                 } else {
3531                         hblks = 1;
3532                 }
3533         } else {
3534                 ASSERT(log->l_sectbb_log == 0);
3535                 hblks = 1;
3536                 hbp = xlog_get_bp(log, 1);
3537                 h_size = XLOG_BIG_RECORD_BSIZE;
3538         }
3539
3540         if (!hbp)
3541                 return ENOMEM;
3542         dbp = xlog_get_bp(log, BTOBB(h_size));
3543         if (!dbp) {
3544                 xlog_put_bp(hbp);
3545                 return ENOMEM;
3546         }
3547
3548         memset(rhash, 0, sizeof(rhash));
3549         if (tail_blk <= head_blk) {
3550                 for (blk_no = tail_blk; blk_no < head_blk; ) {
3551                         error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3552                         if (error)
3553                                 goto bread_err2;
3554
3555                         rhead = (xlog_rec_header_t *)offset;
3556                         error = xlog_valid_rec_header(log, rhead, blk_no);
3557                         if (error)
3558                                 goto bread_err2;
3559
3560                         /* blocks in data section */
3561                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3562                         error = xlog_bread(log, blk_no + hblks, bblks, dbp,
3563                                            &offset);
3564                         if (error)
3565                                 goto bread_err2;
3566
3567                         xlog_unpack_data(rhead, offset, log);
3568                         if ((error = xlog_recover_process_data(log,
3569                                                 rhash, rhead, offset, pass)))
3570                                 goto bread_err2;
3571                         blk_no += bblks + hblks;
3572                 }
3573         } else {
3574                 /*
3575                  * Perform recovery around the end of the physical log.
3576                  * When the head is not on the same cycle number as the tail,
3577                  * we can't do a sequential recovery as above.
3578                  */
3579                 blk_no = tail_blk;
3580                 while (blk_no < log->l_logBBsize) {
3581                         /*
3582                          * Check for header wrapping around physical end-of-log
3583                          */
3584                         offset = XFS_BUF_PTR(hbp);
3585                         split_hblks = 0;
3586                         wrapped_hblks = 0;
3587                         if (blk_no + hblks <= log->l_logBBsize) {
3588                                 /* Read header in one read */
3589                                 error = xlog_bread(log, blk_no, hblks, hbp,
3590                                                    &offset);
3591                                 if (error)
3592                                         goto bread_err2;
3593                         } else {
3594                                 /* This LR is split across physical log end */
3595                                 if (blk_no != log->l_logBBsize) {
3596                                         /* some data before physical log end */
3597                                         ASSERT(blk_no <= INT_MAX);
3598                                         split_hblks = log->l_logBBsize - (int)blk_no;
3599                                         ASSERT(split_hblks > 0);
3600                                         error = xlog_bread(log, blk_no,
3601                                                            split_hblks, hbp,
3602                                                            &offset);
3603                                         if (error)
3604                                                 goto bread_err2;
3605                                 }
3606
3607                                 /*
3608                                  * Note: this black magic still works with
3609                                  * large sector sizes (non-512) only because:
3610                                  * - we increased the buffer size originally
3611                                  *   by 1 sector giving us enough extra space
3612                                  *   for the second read;
3613                                  * - the log start is guaranteed to be sector
3614                                  *   aligned;
3615                                  * - we read the log end (LR header start)
3616                                  *   _first_, then the log start (LR header end)
3617                                  *   - order is important.
3618                                  */
3619                                 wrapped_hblks = hblks - split_hblks;
3620                                 error = XFS_BUF_SET_PTR(hbp,
3621                                                 offset + BBTOB(split_hblks),
3622                                                 BBTOB(hblks - split_hblks));
3623                                 if (error)
3624                                         goto bread_err2;
3625
3626                                 error = xlog_bread_noalign(log, 0,
3627                                                            wrapped_hblks, hbp);
3628                                 if (error)
3629                                         goto bread_err2;
3630
3631                                 error = XFS_BUF_SET_PTR(hbp, offset,
3632                                                         BBTOB(hblks));
3633                                 if (error)
3634                                         goto bread_err2;
3635                         }
3636                         rhead = (xlog_rec_header_t *)offset;
3637                         error = xlog_valid_rec_header(log, rhead,
3638                                                 split_hblks ? blk_no : 0);
3639                         if (error)
3640                                 goto bread_err2;
3641
3642                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3643                         blk_no += hblks;
3644
3645                         /* Read in data for log record */
3646                         if (blk_no + bblks <= log->l_logBBsize) {
3647                                 error = xlog_bread(log, blk_no, bblks, dbp,
3648                                                    &offset);
3649                                 if (error)
3650                                         goto bread_err2;
3651                         } else {
3652                                 /* This log record is split across the
3653                                  * physical end of log */
3654                                 offset = XFS_BUF_PTR(dbp);
3655                                 split_bblks = 0;
3656                                 if (blk_no != log->l_logBBsize) {
3657                                         /* some data is before the physical
3658                                          * end of log */
3659                                         ASSERT(!wrapped_hblks);
3660                                         ASSERT(blk_no <= INT_MAX);
3661                                         split_bblks =
3662                                                 log->l_logBBsize - (int)blk_no;
3663                                         ASSERT(split_bblks > 0);
3664                                         error = xlog_bread(log, blk_no,
3665                                                         split_bblks, dbp,
3666                                                         &offset);
3667                                         if (error)
3668                                                 goto bread_err2;
3669                                 }
3670
3671                                 /*
3672                                  * Note: this black magic still works with
3673                                  * large sector sizes (non-512) only because:
3674                                  * - we increased the buffer size originally
3675                                  *   by 1 sector giving us enough extra space
3676                                  *   for the second read;
3677                                  * - the log start is guaranteed to be sector
3678                                  *   aligned;
3679                                  * - we read the log end (LR header start)
3680                                  *   _first_, then the log start (LR header end)
3681                                  *   - order is important.
3682                                  */
3683                                 error = XFS_BUF_SET_PTR(dbp,
3684                                                 offset + BBTOB(split_bblks),
3685                                                 BBTOB(bblks - split_bblks));
3686                                 if (error)
3687                                         goto bread_err2;
3688
3689                                 error = xlog_bread_noalign(log, wrapped_hblks,
3690                                                 bblks - split_bblks,
3691                                                 dbp);
3692                                 if (error)
3693                                         goto bread_err2;
3694
3695                                 error = XFS_BUF_SET_PTR(dbp, offset, h_size);
3696                                 if (error)
3697                                         goto bread_err2;
3698                         }
3699                         xlog_unpack_data(rhead, offset, log);
3700                         if ((error = xlog_recover_process_data(log, rhash,
3701                                                         rhead, offset, pass)))
3702                                 goto bread_err2;
3703                         blk_no += bblks;
3704                 }
3705
3706                 ASSERT(blk_no >= log->l_logBBsize);
3707                 blk_no -= log->l_logBBsize;
3708
3709                 /* read first part of physical log */
3710                 while (blk_no < head_blk) {
3711                         error = xlog_bread(log, blk_no, hblks, hbp, &offset);
3712                         if (error)
3713                                 goto bread_err2;
3714
3715                         rhead = (xlog_rec_header_t *)offset;
3716                         error = xlog_valid_rec_header(log, rhead, blk_no);
3717                         if (error)
3718                                 goto bread_err2;
3719
3720                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3721                         error = xlog_bread(log, blk_no+hblks, bblks, dbp,
3722                                            &offset);
3723                         if (error)
3724                                 goto bread_err2;
3725
3726                         xlog_unpack_data(rhead, offset, log);
3727                         if ((error = xlog_recover_process_data(log, rhash,
3728                                                         rhead, offset, pass)))
3729                                 goto bread_err2;
3730                         blk_no += bblks + hblks;
3731                 }
3732         }
3733
3734  bread_err2:
3735         xlog_put_bp(dbp);
3736  bread_err1:
3737         xlog_put_bp(hbp);
3738         return error;
3739 }
3740
3741 /*
3742  * Do the recovery of the log.  We actually do this in two phases.
3743  * The two passes are necessary in order to implement the function
3744  * of cancelling a record written into the log.  The first pass
3745  * determines those things which have been cancelled, and the
3746  * second pass replays log items normally except for those which
3747  * have been cancelled.  The handling of the replay and cancellations
3748  * takes place in the log item type specific routines.
3749  *
3750  * The table of items which have cancel records in the log is allocated
3751  * and freed at this level, since only here do we know when all of
3752  * the log recovery has been completed.
3753  */
3754 STATIC int
3755 xlog_do_log_recovery(
3756         xlog_t          *log,
3757         xfs_daddr_t     head_blk,
3758         xfs_daddr_t     tail_blk)
3759 {
3760         int             error;
3761
3762         ASSERT(head_blk != tail_blk);
3763
3764         /*
3765          * First do a pass to find all of the cancelled buf log items.
3766          * Store them in the buf_cancel_table for use in the second pass.
3767          */
3768         log->l_buf_cancel_table =
3769                 (xfs_buf_cancel_t **)kmem_zalloc(XLOG_BC_TABLE_SIZE *
3770                                                  sizeof(xfs_buf_cancel_t*),
3771                                                  KM_SLEEP);
3772         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3773                                       XLOG_RECOVER_PASS1);
3774         if (error != 0) {
3775                 kmem_free(log->l_buf_cancel_table);
3776                 log->l_buf_cancel_table = NULL;
3777                 return error;
3778         }
3779         /*
3780          * Then do a second pass to actually recover the items in the log.
3781          * When it is complete free the table of buf cancel items.
3782          */
3783         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3784                                       XLOG_RECOVER_PASS2);
3785 #ifdef DEBUG
3786         if (!error) {
3787                 int     i;
3788
3789                 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3790                         ASSERT(log->l_buf_cancel_table[i] == NULL);
3791         }
3792 #endif  /* DEBUG */
3793
3794         kmem_free(log->l_buf_cancel_table);
3795         log->l_buf_cancel_table = NULL;
3796
3797         return error;
3798 }
3799
3800 /*
3801  * Do the actual recovery
3802  */
3803 STATIC int
3804 xlog_do_recover(
3805         xlog_t          *log,
3806         xfs_daddr_t     head_blk,
3807         xfs_daddr_t     tail_blk)
3808 {
3809         int             error;
3810         xfs_buf_t       *bp;
3811         xfs_sb_t        *sbp;
3812
3813         /*
3814          * First replay the images in the log.
3815          */
3816         error = xlog_do_log_recovery(log, head_blk, tail_blk);
3817         if (error) {
3818                 return error;
3819         }
3820
3821         XFS_bflush(log->l_mp->m_ddev_targp);
3822
3823         /*
3824          * If IO errors happened during recovery, bail out.
3825          */
3826         if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3827                 return (EIO);
3828         }
3829
3830         /*
3831          * We now update the tail_lsn since much of the recovery has completed
3832          * and there may be space available to use.  If there were no extent
3833          * or iunlinks, we can free up the entire log and set the tail_lsn to
3834          * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3835          * lsn of the last known good LR on disk.  If there are extent frees
3836          * or iunlinks they will have some entries in the AIL; so we look at
3837          * the AIL to determine how to set the tail_lsn.
3838          */
3839         xlog_assign_tail_lsn(log->l_mp);
3840
3841         /*
3842          * Now that we've finished replaying all buffer and inode
3843          * updates, re-read in the superblock.
3844          */
3845         bp = xfs_getsb(log->l_mp, 0);
3846         XFS_BUF_UNDONE(bp);
3847         ASSERT(!(XFS_BUF_ISWRITE(bp)));
3848         ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
3849         XFS_BUF_READ(bp);
3850         XFS_BUF_UNASYNC(bp);
3851         xfsbdstrat(log->l_mp, bp);
3852         error = xfs_iowait(bp);
3853         if (error) {
3854                 xfs_ioerror_alert("xlog_do_recover",
3855                                   log->l_mp, bp, XFS_BUF_ADDR(bp));
3856                 ASSERT(0);
3857                 xfs_buf_relse(bp);
3858                 return error;
3859         }
3860
3861         /* Convert superblock from on-disk format */
3862         sbp = &log->l_mp->m_sb;
3863         xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
3864         ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3865         ASSERT(xfs_sb_good_version(sbp));
3866         xfs_buf_relse(bp);
3867
3868         /* We've re-read the superblock so re-initialize per-cpu counters */
3869         xfs_icsb_reinit_counters(log->l_mp);
3870
3871         xlog_recover_check_summary(log);
3872
3873         /* Normal transactions can now occur */
3874         log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3875         return 0;
3876 }
3877
3878 /*
3879  * Perform recovery and re-initialize some log variables in xlog_find_tail.
3880  *
3881  * Return error or zero.
3882  */
3883 int
3884 xlog_recover(
3885         xlog_t          *log)
3886 {
3887         xfs_daddr_t     head_blk, tail_blk;
3888         int             error;
3889
3890         /* find the tail of the log */
3891         if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3892                 return error;
3893
3894         if (tail_blk != head_blk) {
3895                 /* There used to be a comment here:
3896                  *
3897                  * disallow recovery on read-only mounts.  note -- mount
3898                  * checks for ENOSPC and turns it into an intelligent
3899                  * error message.
3900                  * ...but this is no longer true.  Now, unless you specify
3901                  * NORECOVERY (in which case this function would never be
3902                  * called), we just go ahead and recover.  We do this all
3903                  * under the vfs layer, so we can get away with it unless
3904                  * the device itself is read-only, in which case we fail.
3905                  */
3906                 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3907                         return error;
3908                 }
3909
3910                 cmn_err(CE_NOTE,
3911                         "Starting XFS recovery on filesystem: %s (logdev: %s)",
3912                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3913                         log->l_mp->m_logname : "internal");
3914
3915                 error = xlog_do_recover(log, head_blk, tail_blk);
3916                 log->l_flags |= XLOG_RECOVERY_NEEDED;
3917         }
3918         return error;
3919 }
3920
3921 /*
3922  * In the first part of recovery we replay inodes and buffers and build
3923  * up the list of extent free items which need to be processed.  Here
3924  * we process the extent free items and clean up the on disk unlinked
3925  * inode lists.  This is separated from the first part of recovery so
3926  * that the root and real-time bitmap inodes can be read in from disk in
3927  * between the two stages.  This is necessary so that we can free space
3928  * in the real-time portion of the file system.
3929  */
3930 int
3931 xlog_recover_finish(
3932         xlog_t          *log)
3933 {
3934         /*
3935          * Now we're ready to do the transactions needed for the
3936          * rest of recovery.  Start with completing all the extent
3937          * free intent records and then process the unlinked inode
3938          * lists.  At this point, we essentially run in normal mode
3939          * except that we're still performing recovery actions
3940          * rather than accepting new requests.
3941          */
3942         if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3943                 int     error;
3944                 error = xlog_recover_process_efis(log);
3945                 if (error) {
3946                         cmn_err(CE_ALERT,
3947                                 "Failed to recover EFIs on filesystem: %s",
3948                                 log->l_mp->m_fsname);
3949                         return error;
3950                 }
3951                 /*
3952                  * Sync the log to get all the EFIs out of the AIL.
3953                  * This isn't absolutely necessary, but it helps in
3954                  * case the unlink transactions would have problems
3955                  * pushing the EFIs out of the way.
3956                  */
3957                 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
3958
3959                 xlog_recover_process_iunlinks(log);
3960
3961                 xlog_recover_check_summary(log);
3962
3963                 cmn_err(CE_NOTE,
3964                         "Ending XFS recovery on filesystem: %s (logdev: %s)",
3965                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3966                         log->l_mp->m_logname : "internal");
3967                 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3968         } else {
3969                 cmn_err(CE_DEBUG,
3970                         "!Ending clean XFS mount for filesystem: %s\n",
3971                         log->l_mp->m_fsname);
3972         }
3973         return 0;
3974 }
3975
3976
3977 #if defined(DEBUG)
3978 /*
3979  * Read all of the agf and agi counters and check that they
3980  * are consistent with the superblock counters.
3981  */
3982 void
3983 xlog_recover_check_summary(
3984         xlog_t          *log)
3985 {
3986         xfs_mount_t     *mp;
3987         xfs_agf_t       *agfp;
3988         xfs_buf_t       *agfbp;
3989         xfs_buf_t       *agibp;
3990         xfs_buf_t       *sbbp;
3991 #ifdef XFS_LOUD_RECOVERY
3992         xfs_sb_t        *sbp;
3993 #endif
3994         xfs_agnumber_t  agno;
3995         __uint64_t      freeblks;
3996         __uint64_t      itotal;
3997         __uint64_t      ifree;
3998         int             error;
3999
4000         mp = log->l_mp;
4001
4002         freeblks = 0LL;
4003         itotal = 0LL;
4004         ifree = 0LL;
4005         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4006                 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
4007                 if (error) {
4008                         xfs_fs_cmn_err(CE_ALERT, mp,
4009                                         "xlog_recover_check_summary(agf)"
4010                                         "agf read failed agno %d error %d",
4011                                                         agno, error);
4012                 } else {
4013                         agfp = XFS_BUF_TO_AGF(agfbp);
4014                         freeblks += be32_to_cpu(agfp->agf_freeblks) +
4015                                     be32_to_cpu(agfp->agf_flcount);
4016                         xfs_buf_relse(agfbp);
4017                 }
4018
4019                 error = xfs_read_agi(mp, NULL, agno, &agibp);
4020                 if (!error) {
4021                         struct xfs_agi  *agi = XFS_BUF_TO_AGI(agibp);
4022
4023                         itotal += be32_to_cpu(agi->agi_count);
4024                         ifree += be32_to_cpu(agi->agi_freecount);
4025                         xfs_buf_relse(agibp);
4026                 }
4027         }
4028
4029         sbbp = xfs_getsb(mp, 0);
4030 #ifdef XFS_LOUD_RECOVERY
4031         sbp = &mp->m_sb;
4032         xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(sbbp));
4033         cmn_err(CE_NOTE,
4034                 "xlog_recover_check_summary: sb_icount %Lu itotal %Lu",
4035                 sbp->sb_icount, itotal);
4036         cmn_err(CE_NOTE,
4037                 "xlog_recover_check_summary: sb_ifree %Lu itotal %Lu",
4038                 sbp->sb_ifree, ifree);
4039         cmn_err(CE_NOTE,
4040                 "xlog_recover_check_summary: sb_fdblocks %Lu freeblks %Lu",
4041                 sbp->sb_fdblocks, freeblks);
4042 #if 0
4043         /*
4044          * This is turned off until I account for the allocation
4045          * btree blocks which live in free space.
4046          */
4047         ASSERT(sbp->sb_icount == itotal);
4048         ASSERT(sbp->sb_ifree == ifree);
4049         ASSERT(sbp->sb_fdblocks == freeblks);
4050 #endif
4051 #endif
4052         xfs_buf_relse(sbbp);
4053 }
4054 #endif /* DEBUG */