sh: Fix PC adjustments for varying opcode length.
[linux-2.6] / fs / xfs / xfs_log_recover.c
1 /*
2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3  * All Rights Reserved.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write the Free Software Foundation,
16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
17  */
18 #include "xfs.h"
19 #include "xfs_fs.h"
20 #include "xfs_types.h"
21 #include "xfs_bit.h"
22 #include "xfs_log.h"
23 #include "xfs_inum.h"
24 #include "xfs_trans.h"
25 #include "xfs_sb.h"
26 #include "xfs_ag.h"
27 #include "xfs_dir2.h"
28 #include "xfs_dmapi.h"
29 #include "xfs_mount.h"
30 #include "xfs_error.h"
31 #include "xfs_bmap_btree.h"
32 #include "xfs_alloc_btree.h"
33 #include "xfs_ialloc_btree.h"
34 #include "xfs_dir2_sf.h"
35 #include "xfs_attr_sf.h"
36 #include "xfs_dinode.h"
37 #include "xfs_inode.h"
38 #include "xfs_inode_item.h"
39 #include "xfs_imap.h"
40 #include "xfs_alloc.h"
41 #include "xfs_ialloc.h"
42 #include "xfs_log_priv.h"
43 #include "xfs_buf_item.h"
44 #include "xfs_log_recover.h"
45 #include "xfs_extfree_item.h"
46 #include "xfs_trans_priv.h"
47 #include "xfs_quota.h"
48 #include "xfs_rw.h"
49
50 STATIC int      xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
51 STATIC int      xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
52 STATIC void     xlog_recover_insert_item_backq(xlog_recover_item_t **q,
53                                                xlog_recover_item_t *item);
54 #if defined(DEBUG)
55 STATIC void     xlog_recover_check_summary(xlog_t *);
56 STATIC void     xlog_recover_check_ail(xfs_mount_t *, xfs_log_item_t *, int);
57 #else
58 #define xlog_recover_check_summary(log)
59 #define xlog_recover_check_ail(mp, lip, gen)
60 #endif
61
62
63 /*
64  * Sector aligned buffer routines for buffer create/read/write/access
65  */
66
67 #define XLOG_SECTOR_ROUNDUP_BBCOUNT(log, bbs)   \
68         ( ((log)->l_sectbb_mask && (bbs & (log)->l_sectbb_mask)) ? \
69         ((bbs + (log)->l_sectbb_mask + 1) & ~(log)->l_sectbb_mask) : (bbs) )
70 #define XLOG_SECTOR_ROUNDDOWN_BLKNO(log, bno)   ((bno) & ~(log)->l_sectbb_mask)
71
72 xfs_buf_t *
73 xlog_get_bp(
74         xlog_t          *log,
75         int             num_bblks)
76 {
77         ASSERT(num_bblks > 0);
78
79         if (log->l_sectbb_log) {
80                 if (num_bblks > 1)
81                         num_bblks += XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
82                 num_bblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, num_bblks);
83         }
84         return xfs_buf_get_noaddr(BBTOB(num_bblks), log->l_mp->m_logdev_targp);
85 }
86
87 void
88 xlog_put_bp(
89         xfs_buf_t       *bp)
90 {
91         xfs_buf_free(bp);
92 }
93
94
95 /*
96  * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
97  */
98 int
99 xlog_bread(
100         xlog_t          *log,
101         xfs_daddr_t     blk_no,
102         int             nbblks,
103         xfs_buf_t       *bp)
104 {
105         int             error;
106
107         if (log->l_sectbb_log) {
108                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
109                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
110         }
111
112         ASSERT(nbblks > 0);
113         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
114         ASSERT(bp);
115
116         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
117         XFS_BUF_READ(bp);
118         XFS_BUF_BUSY(bp);
119         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
120         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
121
122         xfsbdstrat(log->l_mp, bp);
123         if ((error = xfs_iowait(bp)))
124                 xfs_ioerror_alert("xlog_bread", log->l_mp,
125                                   bp, XFS_BUF_ADDR(bp));
126         return error;
127 }
128
129 /*
130  * Write out the buffer at the given block for the given number of blocks.
131  * The buffer is kept locked across the write and is returned locked.
132  * This can only be used for synchronous log writes.
133  */
134 STATIC int
135 xlog_bwrite(
136         xlog_t          *log,
137         xfs_daddr_t     blk_no,
138         int             nbblks,
139         xfs_buf_t       *bp)
140 {
141         int             error;
142
143         if (log->l_sectbb_log) {
144                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
145                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
146         }
147
148         ASSERT(nbblks > 0);
149         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
150
151         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
152         XFS_BUF_ZEROFLAGS(bp);
153         XFS_BUF_BUSY(bp);
154         XFS_BUF_HOLD(bp);
155         XFS_BUF_PSEMA(bp, PRIBIO);
156         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
157         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
158
159         if ((error = xfs_bwrite(log->l_mp, bp)))
160                 xfs_ioerror_alert("xlog_bwrite", log->l_mp,
161                                   bp, XFS_BUF_ADDR(bp));
162         return error;
163 }
164
165 STATIC xfs_caddr_t
166 xlog_align(
167         xlog_t          *log,
168         xfs_daddr_t     blk_no,
169         int             nbblks,
170         xfs_buf_t       *bp)
171 {
172         xfs_caddr_t     ptr;
173
174         if (!log->l_sectbb_log)
175                 return XFS_BUF_PTR(bp);
176
177         ptr = XFS_BUF_PTR(bp) + BBTOB((int)blk_no & log->l_sectbb_mask);
178         ASSERT(XFS_BUF_SIZE(bp) >=
179                 BBTOB(nbblks + (blk_no & log->l_sectbb_mask)));
180         return ptr;
181 }
182
183 #ifdef DEBUG
184 /*
185  * dump debug superblock and log record information
186  */
187 STATIC void
188 xlog_header_check_dump(
189         xfs_mount_t             *mp,
190         xlog_rec_header_t       *head)
191 {
192         int                     b;
193
194         cmn_err(CE_DEBUG, "%s:  SB : uuid = ", __FUNCTION__);
195         for (b = 0; b < 16; b++)
196                 cmn_err(CE_DEBUG, "%02x", ((uchar_t *)&mp->m_sb.sb_uuid)[b]);
197         cmn_err(CE_DEBUG, ", fmt = %d\n", XLOG_FMT);
198         cmn_err(CE_DEBUG, "    log : uuid = ");
199         for (b = 0; b < 16; b++)
200                 cmn_err(CE_DEBUG, "%02x",((uchar_t *)&head->h_fs_uuid)[b]);
201         cmn_err(CE_DEBUG, ", fmt = %d\n", INT_GET(head->h_fmt, ARCH_CONVERT));
202 }
203 #else
204 #define xlog_header_check_dump(mp, head)
205 #endif
206
207 /*
208  * check log record header for recovery
209  */
210 STATIC int
211 xlog_header_check_recover(
212         xfs_mount_t             *mp,
213         xlog_rec_header_t       *head)
214 {
215         ASSERT(INT_GET(head->h_magicno, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM);
216
217         /*
218          * IRIX doesn't write the h_fmt field and leaves it zeroed
219          * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
220          * a dirty log created in IRIX.
221          */
222         if (unlikely(INT_GET(head->h_fmt, ARCH_CONVERT) != XLOG_FMT)) {
223                 xlog_warn(
224         "XFS: dirty log written in incompatible format - can't recover");
225                 xlog_header_check_dump(mp, head);
226                 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
227                                  XFS_ERRLEVEL_HIGH, mp);
228                 return XFS_ERROR(EFSCORRUPTED);
229         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
230                 xlog_warn(
231         "XFS: dirty log entry has mismatched uuid - can't recover");
232                 xlog_header_check_dump(mp, head);
233                 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
234                                  XFS_ERRLEVEL_HIGH, mp);
235                 return XFS_ERROR(EFSCORRUPTED);
236         }
237         return 0;
238 }
239
240 /*
241  * read the head block of the log and check the header
242  */
243 STATIC int
244 xlog_header_check_mount(
245         xfs_mount_t             *mp,
246         xlog_rec_header_t       *head)
247 {
248         ASSERT(INT_GET(head->h_magicno, ARCH_CONVERT) == XLOG_HEADER_MAGIC_NUM);
249
250         if (uuid_is_nil(&head->h_fs_uuid)) {
251                 /*
252                  * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
253                  * h_fs_uuid is nil, we assume this log was last mounted
254                  * by IRIX and continue.
255                  */
256                 xlog_warn("XFS: nil uuid in log - IRIX style log");
257         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
258                 xlog_warn("XFS: log has mismatched uuid - can't recover");
259                 xlog_header_check_dump(mp, head);
260                 XFS_ERROR_REPORT("xlog_header_check_mount",
261                                  XFS_ERRLEVEL_HIGH, mp);
262                 return XFS_ERROR(EFSCORRUPTED);
263         }
264         return 0;
265 }
266
267 STATIC void
268 xlog_recover_iodone(
269         struct xfs_buf  *bp)
270 {
271         xfs_mount_t     *mp;
272
273         ASSERT(XFS_BUF_FSPRIVATE(bp, void *));
274
275         if (XFS_BUF_GETERROR(bp)) {
276                 /*
277                  * We're not going to bother about retrying
278                  * this during recovery. One strike!
279                  */
280                 mp = XFS_BUF_FSPRIVATE(bp, xfs_mount_t *);
281                 xfs_ioerror_alert("xlog_recover_iodone",
282                                   mp, bp, XFS_BUF_ADDR(bp));
283                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
284         }
285         XFS_BUF_SET_FSPRIVATE(bp, NULL);
286         XFS_BUF_CLR_IODONE_FUNC(bp);
287         xfs_biodone(bp);
288 }
289
290 /*
291  * This routine finds (to an approximation) the first block in the physical
292  * log which contains the given cycle.  It uses a binary search algorithm.
293  * Note that the algorithm can not be perfect because the disk will not
294  * necessarily be perfect.
295  */
296 int
297 xlog_find_cycle_start(
298         xlog_t          *log,
299         xfs_buf_t       *bp,
300         xfs_daddr_t     first_blk,
301         xfs_daddr_t     *last_blk,
302         uint            cycle)
303 {
304         xfs_caddr_t     offset;
305         xfs_daddr_t     mid_blk;
306         uint            mid_cycle;
307         int             error;
308
309         mid_blk = BLK_AVG(first_blk, *last_blk);
310         while (mid_blk != first_blk && mid_blk != *last_blk) {
311                 if ((error = xlog_bread(log, mid_blk, 1, bp)))
312                         return error;
313                 offset = xlog_align(log, mid_blk, 1, bp);
314                 mid_cycle = GET_CYCLE(offset, ARCH_CONVERT);
315                 if (mid_cycle == cycle) {
316                         *last_blk = mid_blk;
317                         /* last_half_cycle == mid_cycle */
318                 } else {
319                         first_blk = mid_blk;
320                         /* first_half_cycle == mid_cycle */
321                 }
322                 mid_blk = BLK_AVG(first_blk, *last_blk);
323         }
324         ASSERT((mid_blk == first_blk && mid_blk+1 == *last_blk) ||
325                (mid_blk == *last_blk && mid_blk-1 == first_blk));
326
327         return 0;
328 }
329
330 /*
331  * Check that the range of blocks does not contain the cycle number
332  * given.  The scan needs to occur from front to back and the ptr into the
333  * region must be updated since a later routine will need to perform another
334  * test.  If the region is completely good, we end up returning the same
335  * last block number.
336  *
337  * Set blkno to -1 if we encounter no errors.  This is an invalid block number
338  * since we don't ever expect logs to get this large.
339  */
340 STATIC int
341 xlog_find_verify_cycle(
342         xlog_t          *log,
343         xfs_daddr_t     start_blk,
344         int             nbblks,
345         uint            stop_on_cycle_no,
346         xfs_daddr_t     *new_blk)
347 {
348         xfs_daddr_t     i, j;
349         uint            cycle;
350         xfs_buf_t       *bp;
351         xfs_daddr_t     bufblks;
352         xfs_caddr_t     buf = NULL;
353         int             error = 0;
354
355         bufblks = 1 << ffs(nbblks);
356
357         while (!(bp = xlog_get_bp(log, bufblks))) {
358                 /* can't get enough memory to do everything in one big buffer */
359                 bufblks >>= 1;
360                 if (bufblks <= log->l_sectbb_log)
361                         return ENOMEM;
362         }
363
364         for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
365                 int     bcount;
366
367                 bcount = min(bufblks, (start_blk + nbblks - i));
368
369                 if ((error = xlog_bread(log, i, bcount, bp)))
370                         goto out;
371
372                 buf = xlog_align(log, i, bcount, bp);
373                 for (j = 0; j < bcount; j++) {
374                         cycle = GET_CYCLE(buf, ARCH_CONVERT);
375                         if (cycle == stop_on_cycle_no) {
376                                 *new_blk = i+j;
377                                 goto out;
378                         }
379
380                         buf += BBSIZE;
381                 }
382         }
383
384         *new_blk = -1;
385
386 out:
387         xlog_put_bp(bp);
388         return error;
389 }
390
391 /*
392  * Potentially backup over partial log record write.
393  *
394  * In the typical case, last_blk is the number of the block directly after
395  * a good log record.  Therefore, we subtract one to get the block number
396  * of the last block in the given buffer.  extra_bblks contains the number
397  * of blocks we would have read on a previous read.  This happens when the
398  * last log record is split over the end of the physical log.
399  *
400  * extra_bblks is the number of blocks potentially verified on a previous
401  * call to this routine.
402  */
403 STATIC int
404 xlog_find_verify_log_record(
405         xlog_t                  *log,
406         xfs_daddr_t             start_blk,
407         xfs_daddr_t             *last_blk,
408         int                     extra_bblks)
409 {
410         xfs_daddr_t             i;
411         xfs_buf_t               *bp;
412         xfs_caddr_t             offset = NULL;
413         xlog_rec_header_t       *head = NULL;
414         int                     error = 0;
415         int                     smallmem = 0;
416         int                     num_blks = *last_blk - start_blk;
417         int                     xhdrs;
418
419         ASSERT(start_blk != 0 || *last_blk != start_blk);
420
421         if (!(bp = xlog_get_bp(log, num_blks))) {
422                 if (!(bp = xlog_get_bp(log, 1)))
423                         return ENOMEM;
424                 smallmem = 1;
425         } else {
426                 if ((error = xlog_bread(log, start_blk, num_blks, bp)))
427                         goto out;
428                 offset = xlog_align(log, start_blk, num_blks, bp);
429                 offset += ((num_blks - 1) << BBSHIFT);
430         }
431
432         for (i = (*last_blk) - 1; i >= 0; i--) {
433                 if (i < start_blk) {
434                         /* valid log record not found */
435                         xlog_warn(
436                 "XFS: Log inconsistent (didn't find previous header)");
437                         ASSERT(0);
438                         error = XFS_ERROR(EIO);
439                         goto out;
440                 }
441
442                 if (smallmem) {
443                         if ((error = xlog_bread(log, i, 1, bp)))
444                                 goto out;
445                         offset = xlog_align(log, i, 1, bp);
446                 }
447
448                 head = (xlog_rec_header_t *)offset;
449
450                 if (XLOG_HEADER_MAGIC_NUM ==
451                     INT_GET(head->h_magicno, ARCH_CONVERT))
452                         break;
453
454                 if (!smallmem)
455                         offset -= BBSIZE;
456         }
457
458         /*
459          * We hit the beginning of the physical log & still no header.  Return
460          * to caller.  If caller can handle a return of -1, then this routine
461          * will be called again for the end of the physical log.
462          */
463         if (i == -1) {
464                 error = -1;
465                 goto out;
466         }
467
468         /*
469          * We have the final block of the good log (the first block
470          * of the log record _before_ the head. So we check the uuid.
471          */
472         if ((error = xlog_header_check_mount(log->l_mp, head)))
473                 goto out;
474
475         /*
476          * We may have found a log record header before we expected one.
477          * last_blk will be the 1st block # with a given cycle #.  We may end
478          * up reading an entire log record.  In this case, we don't want to
479          * reset last_blk.  Only when last_blk points in the middle of a log
480          * record do we update last_blk.
481          */
482         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
483                 uint    h_size = INT_GET(head->h_size, ARCH_CONVERT);
484
485                 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
486                 if (h_size % XLOG_HEADER_CYCLE_SIZE)
487                         xhdrs++;
488         } else {
489                 xhdrs = 1;
490         }
491
492         if (*last_blk - i + extra_bblks
493                         != BTOBB(INT_GET(head->h_len, ARCH_CONVERT)) + xhdrs)
494                 *last_blk = i;
495
496 out:
497         xlog_put_bp(bp);
498         return error;
499 }
500
501 /*
502  * Head is defined to be the point of the log where the next log write
503  * write could go.  This means that incomplete LR writes at the end are
504  * eliminated when calculating the head.  We aren't guaranteed that previous
505  * LR have complete transactions.  We only know that a cycle number of
506  * current cycle number -1 won't be present in the log if we start writing
507  * from our current block number.
508  *
509  * last_blk contains the block number of the first block with a given
510  * cycle number.
511  *
512  * Return: zero if normal, non-zero if error.
513  */
514 STATIC int
515 xlog_find_head(
516         xlog_t          *log,
517         xfs_daddr_t     *return_head_blk)
518 {
519         xfs_buf_t       *bp;
520         xfs_caddr_t     offset;
521         xfs_daddr_t     new_blk, first_blk, start_blk, last_blk, head_blk;
522         int             num_scan_bblks;
523         uint            first_half_cycle, last_half_cycle;
524         uint            stop_on_cycle;
525         int             error, log_bbnum = log->l_logBBsize;
526
527         /* Is the end of the log device zeroed? */
528         if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
529                 *return_head_blk = first_blk;
530
531                 /* Is the whole lot zeroed? */
532                 if (!first_blk) {
533                         /* Linux XFS shouldn't generate totally zeroed logs -
534                          * mkfs etc write a dummy unmount record to a fresh
535                          * log so we can store the uuid in there
536                          */
537                         xlog_warn("XFS: totally zeroed log");
538                 }
539
540                 return 0;
541         } else if (error) {
542                 xlog_warn("XFS: empty log check failed");
543                 return error;
544         }
545
546         first_blk = 0;                  /* get cycle # of 1st block */
547         bp = xlog_get_bp(log, 1);
548         if (!bp)
549                 return ENOMEM;
550         if ((error = xlog_bread(log, 0, 1, bp)))
551                 goto bp_err;
552         offset = xlog_align(log, 0, 1, bp);
553         first_half_cycle = GET_CYCLE(offset, ARCH_CONVERT);
554
555         last_blk = head_blk = log_bbnum - 1;    /* get cycle # of last block */
556         if ((error = xlog_bread(log, last_blk, 1, bp)))
557                 goto bp_err;
558         offset = xlog_align(log, last_blk, 1, bp);
559         last_half_cycle = GET_CYCLE(offset, ARCH_CONVERT);
560         ASSERT(last_half_cycle != 0);
561
562         /*
563          * If the 1st half cycle number is equal to the last half cycle number,
564          * then the entire log is stamped with the same cycle number.  In this
565          * case, head_blk can't be set to zero (which makes sense).  The below
566          * math doesn't work out properly with head_blk equal to zero.  Instead,
567          * we set it to log_bbnum which is an invalid block number, but this
568          * value makes the math correct.  If head_blk doesn't changed through
569          * all the tests below, *head_blk is set to zero at the very end rather
570          * than log_bbnum.  In a sense, log_bbnum and zero are the same block
571          * in a circular file.
572          */
573         if (first_half_cycle == last_half_cycle) {
574                 /*
575                  * In this case we believe that the entire log should have
576                  * cycle number last_half_cycle.  We need to scan backwards
577                  * from the end verifying that there are no holes still
578                  * containing last_half_cycle - 1.  If we find such a hole,
579                  * then the start of that hole will be the new head.  The
580                  * simple case looks like
581                  *        x | x ... | x - 1 | x
582                  * Another case that fits this picture would be
583                  *        x | x + 1 | x ... | x
584                  * In this case the head really is somewhere at the end of the
585                  * log, as one of the latest writes at the beginning was
586                  * incomplete.
587                  * One more case is
588                  *        x | x + 1 | x ... | x - 1 | x
589                  * This is really the combination of the above two cases, and
590                  * the head has to end up at the start of the x-1 hole at the
591                  * end of the log.
592                  *
593                  * In the 256k log case, we will read from the beginning to the
594                  * end of the log and search for cycle numbers equal to x-1.
595                  * We don't worry about the x+1 blocks that we encounter,
596                  * because we know that they cannot be the head since the log
597                  * started with x.
598                  */
599                 head_blk = log_bbnum;
600                 stop_on_cycle = last_half_cycle - 1;
601         } else {
602                 /*
603                  * In this case we want to find the first block with cycle
604                  * number matching last_half_cycle.  We expect the log to be
605                  * some variation on
606                  *        x + 1 ... | x ...
607                  * The first block with cycle number x (last_half_cycle) will
608                  * be where the new head belongs.  First we do a binary search
609                  * for the first occurrence of last_half_cycle.  The binary
610                  * search may not be totally accurate, so then we scan back
611                  * from there looking for occurrences of last_half_cycle before
612                  * us.  If that backwards scan wraps around the beginning of
613                  * the log, then we look for occurrences of last_half_cycle - 1
614                  * at the end of the log.  The cases we're looking for look
615                  * like
616                  *        x + 1 ... | x | x + 1 | x ...
617                  *                               ^ binary search stopped here
618                  * or
619                  *        x + 1 ... | x ... | x - 1 | x
620                  *        <---------> less than scan distance
621                  */
622                 stop_on_cycle = last_half_cycle;
623                 if ((error = xlog_find_cycle_start(log, bp, first_blk,
624                                                 &head_blk, last_half_cycle)))
625                         goto bp_err;
626         }
627
628         /*
629          * Now validate the answer.  Scan back some number of maximum possible
630          * blocks and make sure each one has the expected cycle number.  The
631          * maximum is determined by the total possible amount of buffering
632          * in the in-core log.  The following number can be made tighter if
633          * we actually look at the block size of the filesystem.
634          */
635         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
636         if (head_blk >= num_scan_bblks) {
637                 /*
638                  * We are guaranteed that the entire check can be performed
639                  * in one buffer.
640                  */
641                 start_blk = head_blk - num_scan_bblks;
642                 if ((error = xlog_find_verify_cycle(log,
643                                                 start_blk, num_scan_bblks,
644                                                 stop_on_cycle, &new_blk)))
645                         goto bp_err;
646                 if (new_blk != -1)
647                         head_blk = new_blk;
648         } else {                /* need to read 2 parts of log */
649                 /*
650                  * We are going to scan backwards in the log in two parts.
651                  * First we scan the physical end of the log.  In this part
652                  * of the log, we are looking for blocks with cycle number
653                  * last_half_cycle - 1.
654                  * If we find one, then we know that the log starts there, as
655                  * we've found a hole that didn't get written in going around
656                  * the end of the physical log.  The simple case for this is
657                  *        x + 1 ... | x ... | x - 1 | x
658                  *        <---------> less than scan distance
659                  * If all of the blocks at the end of the log have cycle number
660                  * last_half_cycle, then we check the blocks at the start of
661                  * the log looking for occurrences of last_half_cycle.  If we
662                  * find one, then our current estimate for the location of the
663                  * first occurrence of last_half_cycle is wrong and we move
664                  * back to the hole we've found.  This case looks like
665                  *        x + 1 ... | x | x + 1 | x ...
666                  *                               ^ binary search stopped here
667                  * Another case we need to handle that only occurs in 256k
668                  * logs is
669                  *        x + 1 ... | x ... | x+1 | x ...
670                  *                   ^ binary search stops here
671                  * In a 256k log, the scan at the end of the log will see the
672                  * x + 1 blocks.  We need to skip past those since that is
673                  * certainly not the head of the log.  By searching for
674                  * last_half_cycle-1 we accomplish that.
675                  */
676                 start_blk = log_bbnum - num_scan_bblks + head_blk;
677                 ASSERT(head_blk <= INT_MAX &&
678                         (xfs_daddr_t) num_scan_bblks - head_blk >= 0);
679                 if ((error = xlog_find_verify_cycle(log, start_blk,
680                                         num_scan_bblks - (int)head_blk,
681                                         (stop_on_cycle - 1), &new_blk)))
682                         goto bp_err;
683                 if (new_blk != -1) {
684                         head_blk = new_blk;
685                         goto bad_blk;
686                 }
687
688                 /*
689                  * Scan beginning of log now.  The last part of the physical
690                  * log is good.  This scan needs to verify that it doesn't find
691                  * the last_half_cycle.
692                  */
693                 start_blk = 0;
694                 ASSERT(head_blk <= INT_MAX);
695                 if ((error = xlog_find_verify_cycle(log,
696                                         start_blk, (int)head_blk,
697                                         stop_on_cycle, &new_blk)))
698                         goto bp_err;
699                 if (new_blk != -1)
700                         head_blk = new_blk;
701         }
702
703  bad_blk:
704         /*
705          * Now we need to make sure head_blk is not pointing to a block in
706          * the middle of a log record.
707          */
708         num_scan_bblks = XLOG_REC_SHIFT(log);
709         if (head_blk >= num_scan_bblks) {
710                 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
711
712                 /* start ptr at last block ptr before head_blk */
713                 if ((error = xlog_find_verify_log_record(log, start_blk,
714                                                         &head_blk, 0)) == -1) {
715                         error = XFS_ERROR(EIO);
716                         goto bp_err;
717                 } else if (error)
718                         goto bp_err;
719         } else {
720                 start_blk = 0;
721                 ASSERT(head_blk <= INT_MAX);
722                 if ((error = xlog_find_verify_log_record(log, start_blk,
723                                                         &head_blk, 0)) == -1) {
724                         /* We hit the beginning of the log during our search */
725                         start_blk = log_bbnum - num_scan_bblks + head_blk;
726                         new_blk = log_bbnum;
727                         ASSERT(start_blk <= INT_MAX &&
728                                 (xfs_daddr_t) log_bbnum-start_blk >= 0);
729                         ASSERT(head_blk <= INT_MAX);
730                         if ((error = xlog_find_verify_log_record(log,
731                                                         start_blk, &new_blk,
732                                                         (int)head_blk)) == -1) {
733                                 error = XFS_ERROR(EIO);
734                                 goto bp_err;
735                         } else if (error)
736                                 goto bp_err;
737                         if (new_blk != log_bbnum)
738                                 head_blk = new_blk;
739                 } else if (error)
740                         goto bp_err;
741         }
742
743         xlog_put_bp(bp);
744         if (head_blk == log_bbnum)
745                 *return_head_blk = 0;
746         else
747                 *return_head_blk = head_blk;
748         /*
749          * When returning here, we have a good block number.  Bad block
750          * means that during a previous crash, we didn't have a clean break
751          * from cycle number N to cycle number N-1.  In this case, we need
752          * to find the first block with cycle number N-1.
753          */
754         return 0;
755
756  bp_err:
757         xlog_put_bp(bp);
758
759         if (error)
760             xlog_warn("XFS: failed to find log head");
761         return error;
762 }
763
764 /*
765  * Find the sync block number or the tail of the log.
766  *
767  * This will be the block number of the last record to have its
768  * associated buffers synced to disk.  Every log record header has
769  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
770  * to get a sync block number.  The only concern is to figure out which
771  * log record header to believe.
772  *
773  * The following algorithm uses the log record header with the largest
774  * lsn.  The entire log record does not need to be valid.  We only care
775  * that the header is valid.
776  *
777  * We could speed up search by using current head_blk buffer, but it is not
778  * available.
779  */
780 int
781 xlog_find_tail(
782         xlog_t                  *log,
783         xfs_daddr_t             *head_blk,
784         xfs_daddr_t             *tail_blk)
785 {
786         xlog_rec_header_t       *rhead;
787         xlog_op_header_t        *op_head;
788         xfs_caddr_t             offset = NULL;
789         xfs_buf_t               *bp;
790         int                     error, i, found;
791         xfs_daddr_t             umount_data_blk;
792         xfs_daddr_t             after_umount_blk;
793         xfs_lsn_t               tail_lsn;
794         int                     hblks;
795
796         found = 0;
797
798         /*
799          * Find previous log record
800          */
801         if ((error = xlog_find_head(log, head_blk)))
802                 return error;
803
804         bp = xlog_get_bp(log, 1);
805         if (!bp)
806                 return ENOMEM;
807         if (*head_blk == 0) {                           /* special case */
808                 if ((error = xlog_bread(log, 0, 1, bp)))
809                         goto bread_err;
810                 offset = xlog_align(log, 0, 1, bp);
811                 if (GET_CYCLE(offset, ARCH_CONVERT) == 0) {
812                         *tail_blk = 0;
813                         /* leave all other log inited values alone */
814                         goto exit;
815                 }
816         }
817
818         /*
819          * Search backwards looking for log record header block
820          */
821         ASSERT(*head_blk < INT_MAX);
822         for (i = (int)(*head_blk) - 1; i >= 0; i--) {
823                 if ((error = xlog_bread(log, i, 1, bp)))
824                         goto bread_err;
825                 offset = xlog_align(log, i, 1, bp);
826                 if (XLOG_HEADER_MAGIC_NUM ==
827                     INT_GET(*(uint *)offset, ARCH_CONVERT)) {
828                         found = 1;
829                         break;
830                 }
831         }
832         /*
833          * If we haven't found the log record header block, start looking
834          * again from the end of the physical log.  XXXmiken: There should be
835          * a check here to make sure we didn't search more than N blocks in
836          * the previous code.
837          */
838         if (!found) {
839                 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
840                         if ((error = xlog_bread(log, i, 1, bp)))
841                                 goto bread_err;
842                         offset = xlog_align(log, i, 1, bp);
843                         if (XLOG_HEADER_MAGIC_NUM ==
844                             INT_GET(*(uint*)offset, ARCH_CONVERT)) {
845                                 found = 2;
846                                 break;
847                         }
848                 }
849         }
850         if (!found) {
851                 xlog_warn("XFS: xlog_find_tail: couldn't find sync record");
852                 ASSERT(0);
853                 return XFS_ERROR(EIO);
854         }
855
856         /* find blk_no of tail of log */
857         rhead = (xlog_rec_header_t *)offset;
858         *tail_blk = BLOCK_LSN(INT_GET(rhead->h_tail_lsn, ARCH_CONVERT));
859
860         /*
861          * Reset log values according to the state of the log when we
862          * crashed.  In the case where head_blk == 0, we bump curr_cycle
863          * one because the next write starts a new cycle rather than
864          * continuing the cycle of the last good log record.  At this
865          * point we have guaranteed that all partial log records have been
866          * accounted for.  Therefore, we know that the last good log record
867          * written was complete and ended exactly on the end boundary
868          * of the physical log.
869          */
870         log->l_prev_block = i;
871         log->l_curr_block = (int)*head_blk;
872         log->l_curr_cycle = INT_GET(rhead->h_cycle, ARCH_CONVERT);
873         if (found == 2)
874                 log->l_curr_cycle++;
875         log->l_tail_lsn = INT_GET(rhead->h_tail_lsn, ARCH_CONVERT);
876         log->l_last_sync_lsn = INT_GET(rhead->h_lsn, ARCH_CONVERT);
877         log->l_grant_reserve_cycle = log->l_curr_cycle;
878         log->l_grant_reserve_bytes = BBTOB(log->l_curr_block);
879         log->l_grant_write_cycle = log->l_curr_cycle;
880         log->l_grant_write_bytes = BBTOB(log->l_curr_block);
881
882         /*
883          * Look for unmount record.  If we find it, then we know there
884          * was a clean unmount.  Since 'i' could be the last block in
885          * the physical log, we convert to a log block before comparing
886          * to the head_blk.
887          *
888          * Save the current tail lsn to use to pass to
889          * xlog_clear_stale_blocks() below.  We won't want to clear the
890          * unmount record if there is one, so we pass the lsn of the
891          * unmount record rather than the block after it.
892          */
893         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
894                 int     h_size = INT_GET(rhead->h_size, ARCH_CONVERT);
895                 int     h_version = INT_GET(rhead->h_version, ARCH_CONVERT);
896
897                 if ((h_version & XLOG_VERSION_2) &&
898                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
899                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
900                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
901                                 hblks++;
902                 } else {
903                         hblks = 1;
904                 }
905         } else {
906                 hblks = 1;
907         }
908         after_umount_blk = (i + hblks + (int)
909                 BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT))) % log->l_logBBsize;
910         tail_lsn = log->l_tail_lsn;
911         if (*head_blk == after_umount_blk &&
912             INT_GET(rhead->h_num_logops, ARCH_CONVERT) == 1) {
913                 umount_data_blk = (i + hblks) % log->l_logBBsize;
914                 if ((error = xlog_bread(log, umount_data_blk, 1, bp))) {
915                         goto bread_err;
916                 }
917                 offset = xlog_align(log, umount_data_blk, 1, bp);
918                 op_head = (xlog_op_header_t *)offset;
919                 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
920                         /*
921                          * Set tail and last sync so that newly written
922                          * log records will point recovery to after the
923                          * current unmount record.
924                          */
925                         ASSIGN_ANY_LSN_HOST(log->l_tail_lsn, log->l_curr_cycle,
926                                         after_umount_blk);
927                         ASSIGN_ANY_LSN_HOST(log->l_last_sync_lsn, log->l_curr_cycle,
928                                         after_umount_blk);
929                         *tail_blk = after_umount_blk;
930                 }
931         }
932
933         /*
934          * Make sure that there are no blocks in front of the head
935          * with the same cycle number as the head.  This can happen
936          * because we allow multiple outstanding log writes concurrently,
937          * and the later writes might make it out before earlier ones.
938          *
939          * We use the lsn from before modifying it so that we'll never
940          * overwrite the unmount record after a clean unmount.
941          *
942          * Do this only if we are going to recover the filesystem
943          *
944          * NOTE: This used to say "if (!readonly)"
945          * However on Linux, we can & do recover a read-only filesystem.
946          * We only skip recovery if NORECOVERY is specified on mount,
947          * in which case we would not be here.
948          *
949          * But... if the -device- itself is readonly, just skip this.
950          * We can't recover this device anyway, so it won't matter.
951          */
952         if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
953                 error = xlog_clear_stale_blocks(log, tail_lsn);
954         }
955
956 bread_err:
957 exit:
958         xlog_put_bp(bp);
959
960         if (error)
961                 xlog_warn("XFS: failed to locate log tail");
962         return error;
963 }
964
965 /*
966  * Is the log zeroed at all?
967  *
968  * The last binary search should be changed to perform an X block read
969  * once X becomes small enough.  You can then search linearly through
970  * the X blocks.  This will cut down on the number of reads we need to do.
971  *
972  * If the log is partially zeroed, this routine will pass back the blkno
973  * of the first block with cycle number 0.  It won't have a complete LR
974  * preceding it.
975  *
976  * Return:
977  *      0  => the log is completely written to
978  *      -1 => use *blk_no as the first block of the log
979  *      >0 => error has occurred
980  */
981 int
982 xlog_find_zeroed(
983         xlog_t          *log,
984         xfs_daddr_t     *blk_no)
985 {
986         xfs_buf_t       *bp;
987         xfs_caddr_t     offset;
988         uint            first_cycle, last_cycle;
989         xfs_daddr_t     new_blk, last_blk, start_blk;
990         xfs_daddr_t     num_scan_bblks;
991         int             error, log_bbnum = log->l_logBBsize;
992
993         *blk_no = 0;
994
995         /* check totally zeroed log */
996         bp = xlog_get_bp(log, 1);
997         if (!bp)
998                 return ENOMEM;
999         if ((error = xlog_bread(log, 0, 1, bp)))
1000                 goto bp_err;
1001         offset = xlog_align(log, 0, 1, bp);
1002         first_cycle = GET_CYCLE(offset, ARCH_CONVERT);
1003         if (first_cycle == 0) {         /* completely zeroed log */
1004                 *blk_no = 0;
1005                 xlog_put_bp(bp);
1006                 return -1;
1007         }
1008
1009         /* check partially zeroed log */
1010         if ((error = xlog_bread(log, log_bbnum-1, 1, bp)))
1011                 goto bp_err;
1012         offset = xlog_align(log, log_bbnum-1, 1, bp);
1013         last_cycle = GET_CYCLE(offset, ARCH_CONVERT);
1014         if (last_cycle != 0) {          /* log completely written to */
1015                 xlog_put_bp(bp);
1016                 return 0;
1017         } else if (first_cycle != 1) {
1018                 /*
1019                  * If the cycle of the last block is zero, the cycle of
1020                  * the first block must be 1. If it's not, maybe we're
1021                  * not looking at a log... Bail out.
1022                  */
1023                 xlog_warn("XFS: Log inconsistent or not a log (last==0, first!=1)");
1024                 return XFS_ERROR(EINVAL);
1025         }
1026
1027         /* we have a partially zeroed log */
1028         last_blk = log_bbnum-1;
1029         if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1030                 goto bp_err;
1031
1032         /*
1033          * Validate the answer.  Because there is no way to guarantee that
1034          * the entire log is made up of log records which are the same size,
1035          * we scan over the defined maximum blocks.  At this point, the maximum
1036          * is not chosen to mean anything special.   XXXmiken
1037          */
1038         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1039         ASSERT(num_scan_bblks <= INT_MAX);
1040
1041         if (last_blk < num_scan_bblks)
1042                 num_scan_bblks = last_blk;
1043         start_blk = last_blk - num_scan_bblks;
1044
1045         /*
1046          * We search for any instances of cycle number 0 that occur before
1047          * our current estimate of the head.  What we're trying to detect is
1048          *        1 ... | 0 | 1 | 0...
1049          *                       ^ binary search ends here
1050          */
1051         if ((error = xlog_find_verify_cycle(log, start_blk,
1052                                          (int)num_scan_bblks, 0, &new_blk)))
1053                 goto bp_err;
1054         if (new_blk != -1)
1055                 last_blk = new_blk;
1056
1057         /*
1058          * Potentially backup over partial log record write.  We don't need
1059          * to search the end of the log because we know it is zero.
1060          */
1061         if ((error = xlog_find_verify_log_record(log, start_blk,
1062                                 &last_blk, 0)) == -1) {
1063             error = XFS_ERROR(EIO);
1064             goto bp_err;
1065         } else if (error)
1066             goto bp_err;
1067
1068         *blk_no = last_blk;
1069 bp_err:
1070         xlog_put_bp(bp);
1071         if (error)
1072                 return error;
1073         return -1;
1074 }
1075
1076 /*
1077  * These are simple subroutines used by xlog_clear_stale_blocks() below
1078  * to initialize a buffer full of empty log record headers and write
1079  * them into the log.
1080  */
1081 STATIC void
1082 xlog_add_record(
1083         xlog_t                  *log,
1084         xfs_caddr_t             buf,
1085         int                     cycle,
1086         int                     block,
1087         int                     tail_cycle,
1088         int                     tail_block)
1089 {
1090         xlog_rec_header_t       *recp = (xlog_rec_header_t *)buf;
1091
1092         memset(buf, 0, BBSIZE);
1093         INT_SET(recp->h_magicno, ARCH_CONVERT, XLOG_HEADER_MAGIC_NUM);
1094         INT_SET(recp->h_cycle, ARCH_CONVERT, cycle);
1095         INT_SET(recp->h_version, ARCH_CONVERT,
1096                         XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb) ? 2 : 1);
1097         ASSIGN_ANY_LSN_DISK(recp->h_lsn, cycle, block);
1098         ASSIGN_ANY_LSN_DISK(recp->h_tail_lsn, tail_cycle, tail_block);
1099         INT_SET(recp->h_fmt, ARCH_CONVERT, XLOG_FMT);
1100         memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1101 }
1102
1103 STATIC int
1104 xlog_write_log_records(
1105         xlog_t          *log,
1106         int             cycle,
1107         int             start_block,
1108         int             blocks,
1109         int             tail_cycle,
1110         int             tail_block)
1111 {
1112         xfs_caddr_t     offset;
1113         xfs_buf_t       *bp;
1114         int             balign, ealign;
1115         int             sectbb = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
1116         int             end_block = start_block + blocks;
1117         int             bufblks;
1118         int             error = 0;
1119         int             i, j = 0;
1120
1121         bufblks = 1 << ffs(blocks);
1122         while (!(bp = xlog_get_bp(log, bufblks))) {
1123                 bufblks >>= 1;
1124                 if (bufblks <= log->l_sectbb_log)
1125                         return ENOMEM;
1126         }
1127
1128         /* We may need to do a read at the start to fill in part of
1129          * the buffer in the starting sector not covered by the first
1130          * write below.
1131          */
1132         balign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, start_block);
1133         if (balign != start_block) {
1134                 if ((error = xlog_bread(log, start_block, 1, bp))) {
1135                         xlog_put_bp(bp);
1136                         return error;
1137                 }
1138                 j = start_block - balign;
1139         }
1140
1141         for (i = start_block; i < end_block; i += bufblks) {
1142                 int             bcount, endcount;
1143
1144                 bcount = min(bufblks, end_block - start_block);
1145                 endcount = bcount - j;
1146
1147                 /* We may need to do a read at the end to fill in part of
1148                  * the buffer in the final sector not covered by the write.
1149                  * If this is the same sector as the above read, skip it.
1150                  */
1151                 ealign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, end_block);
1152                 if (j == 0 && (start_block + endcount > ealign)) {
1153                         offset = XFS_BUF_PTR(bp);
1154                         balign = BBTOB(ealign - start_block);
1155                         XFS_BUF_SET_PTR(bp, offset + balign, BBTOB(sectbb));
1156                         if ((error = xlog_bread(log, ealign, sectbb, bp)))
1157                                 break;
1158                         XFS_BUF_SET_PTR(bp, offset, bufblks);
1159                 }
1160
1161                 offset = xlog_align(log, start_block, endcount, bp);
1162                 for (; j < endcount; j++) {
1163                         xlog_add_record(log, offset, cycle, i+j,
1164                                         tail_cycle, tail_block);
1165                         offset += BBSIZE;
1166                 }
1167                 error = xlog_bwrite(log, start_block, endcount, bp);
1168                 if (error)
1169                         break;
1170                 start_block += endcount;
1171                 j = 0;
1172         }
1173         xlog_put_bp(bp);
1174         return error;
1175 }
1176
1177 /*
1178  * This routine is called to blow away any incomplete log writes out
1179  * in front of the log head.  We do this so that we won't become confused
1180  * if we come up, write only a little bit more, and then crash again.
1181  * If we leave the partial log records out there, this situation could
1182  * cause us to think those partial writes are valid blocks since they
1183  * have the current cycle number.  We get rid of them by overwriting them
1184  * with empty log records with the old cycle number rather than the
1185  * current one.
1186  *
1187  * The tail lsn is passed in rather than taken from
1188  * the log so that we will not write over the unmount record after a
1189  * clean unmount in a 512 block log.  Doing so would leave the log without
1190  * any valid log records in it until a new one was written.  If we crashed
1191  * during that time we would not be able to recover.
1192  */
1193 STATIC int
1194 xlog_clear_stale_blocks(
1195         xlog_t          *log,
1196         xfs_lsn_t       tail_lsn)
1197 {
1198         int             tail_cycle, head_cycle;
1199         int             tail_block, head_block;
1200         int             tail_distance, max_distance;
1201         int             distance;
1202         int             error;
1203
1204         tail_cycle = CYCLE_LSN(tail_lsn);
1205         tail_block = BLOCK_LSN(tail_lsn);
1206         head_cycle = log->l_curr_cycle;
1207         head_block = log->l_curr_block;
1208
1209         /*
1210          * Figure out the distance between the new head of the log
1211          * and the tail.  We want to write over any blocks beyond the
1212          * head that we may have written just before the crash, but
1213          * we don't want to overwrite the tail of the log.
1214          */
1215         if (head_cycle == tail_cycle) {
1216                 /*
1217                  * The tail is behind the head in the physical log,
1218                  * so the distance from the head to the tail is the
1219                  * distance from the head to the end of the log plus
1220                  * the distance from the beginning of the log to the
1221                  * tail.
1222                  */
1223                 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1224                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1225                                          XFS_ERRLEVEL_LOW, log->l_mp);
1226                         return XFS_ERROR(EFSCORRUPTED);
1227                 }
1228                 tail_distance = tail_block + (log->l_logBBsize - head_block);
1229         } else {
1230                 /*
1231                  * The head is behind the tail in the physical log,
1232                  * so the distance from the head to the tail is just
1233                  * the tail block minus the head block.
1234                  */
1235                 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1236                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1237                                          XFS_ERRLEVEL_LOW, log->l_mp);
1238                         return XFS_ERROR(EFSCORRUPTED);
1239                 }
1240                 tail_distance = tail_block - head_block;
1241         }
1242
1243         /*
1244          * If the head is right up against the tail, we can't clear
1245          * anything.
1246          */
1247         if (tail_distance <= 0) {
1248                 ASSERT(tail_distance == 0);
1249                 return 0;
1250         }
1251
1252         max_distance = XLOG_TOTAL_REC_SHIFT(log);
1253         /*
1254          * Take the smaller of the maximum amount of outstanding I/O
1255          * we could have and the distance to the tail to clear out.
1256          * We take the smaller so that we don't overwrite the tail and
1257          * we don't waste all day writing from the head to the tail
1258          * for no reason.
1259          */
1260         max_distance = MIN(max_distance, tail_distance);
1261
1262         if ((head_block + max_distance) <= log->l_logBBsize) {
1263                 /*
1264                  * We can stomp all the blocks we need to without
1265                  * wrapping around the end of the log.  Just do it
1266                  * in a single write.  Use the cycle number of the
1267                  * current cycle minus one so that the log will look like:
1268                  *     n ... | n - 1 ...
1269                  */
1270                 error = xlog_write_log_records(log, (head_cycle - 1),
1271                                 head_block, max_distance, tail_cycle,
1272                                 tail_block);
1273                 if (error)
1274                         return error;
1275         } else {
1276                 /*
1277                  * We need to wrap around the end of the physical log in
1278                  * order to clear all the blocks.  Do it in two separate
1279                  * I/Os.  The first write should be from the head to the
1280                  * end of the physical log, and it should use the current
1281                  * cycle number minus one just like above.
1282                  */
1283                 distance = log->l_logBBsize - head_block;
1284                 error = xlog_write_log_records(log, (head_cycle - 1),
1285                                 head_block, distance, tail_cycle,
1286                                 tail_block);
1287
1288                 if (error)
1289                         return error;
1290
1291                 /*
1292                  * Now write the blocks at the start of the physical log.
1293                  * This writes the remainder of the blocks we want to clear.
1294                  * It uses the current cycle number since we're now on the
1295                  * same cycle as the head so that we get:
1296                  *    n ... n ... | n - 1 ...
1297                  *    ^^^^^ blocks we're writing
1298                  */
1299                 distance = max_distance - (log->l_logBBsize - head_block);
1300                 error = xlog_write_log_records(log, head_cycle, 0, distance,
1301                                 tail_cycle, tail_block);
1302                 if (error)
1303                         return error;
1304         }
1305
1306         return 0;
1307 }
1308
1309 /******************************************************************************
1310  *
1311  *              Log recover routines
1312  *
1313  ******************************************************************************
1314  */
1315
1316 STATIC xlog_recover_t *
1317 xlog_recover_find_tid(
1318         xlog_recover_t          *q,
1319         xlog_tid_t              tid)
1320 {
1321         xlog_recover_t          *p = q;
1322
1323         while (p != NULL) {
1324                 if (p->r_log_tid == tid)
1325                     break;
1326                 p = p->r_next;
1327         }
1328         return p;
1329 }
1330
1331 STATIC void
1332 xlog_recover_put_hashq(
1333         xlog_recover_t          **q,
1334         xlog_recover_t          *trans)
1335 {
1336         trans->r_next = *q;
1337         *q = trans;
1338 }
1339
1340 STATIC void
1341 xlog_recover_add_item(
1342         xlog_recover_item_t     **itemq)
1343 {
1344         xlog_recover_item_t     *item;
1345
1346         item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1347         xlog_recover_insert_item_backq(itemq, item);
1348 }
1349
1350 STATIC int
1351 xlog_recover_add_to_cont_trans(
1352         xlog_recover_t          *trans,
1353         xfs_caddr_t             dp,
1354         int                     len)
1355 {
1356         xlog_recover_item_t     *item;
1357         xfs_caddr_t             ptr, old_ptr;
1358         int                     old_len;
1359
1360         item = trans->r_itemq;
1361         if (item == 0) {
1362                 /* finish copying rest of trans header */
1363                 xlog_recover_add_item(&trans->r_itemq);
1364                 ptr = (xfs_caddr_t) &trans->r_theader +
1365                                 sizeof(xfs_trans_header_t) - len;
1366                 memcpy(ptr, dp, len); /* d, s, l */
1367                 return 0;
1368         }
1369         item = item->ri_prev;
1370
1371         old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1372         old_len = item->ri_buf[item->ri_cnt-1].i_len;
1373
1374         ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1375         memcpy(&ptr[old_len], dp, len); /* d, s, l */
1376         item->ri_buf[item->ri_cnt-1].i_len += len;
1377         item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1378         return 0;
1379 }
1380
1381 /*
1382  * The next region to add is the start of a new region.  It could be
1383  * a whole region or it could be the first part of a new region.  Because
1384  * of this, the assumption here is that the type and size fields of all
1385  * format structures fit into the first 32 bits of the structure.
1386  *
1387  * This works because all regions must be 32 bit aligned.  Therefore, we
1388  * either have both fields or we have neither field.  In the case we have
1389  * neither field, the data part of the region is zero length.  We only have
1390  * a log_op_header and can throw away the header since a new one will appear
1391  * later.  If we have at least 4 bytes, then we can determine how many regions
1392  * will appear in the current log item.
1393  */
1394 STATIC int
1395 xlog_recover_add_to_trans(
1396         xlog_recover_t          *trans,
1397         xfs_caddr_t             dp,
1398         int                     len)
1399 {
1400         xfs_inode_log_format_t  *in_f;                  /* any will do */
1401         xlog_recover_item_t     *item;
1402         xfs_caddr_t             ptr;
1403
1404         if (!len)
1405                 return 0;
1406         item = trans->r_itemq;
1407         if (item == 0) {
1408                 ASSERT(*(uint *)dp == XFS_TRANS_HEADER_MAGIC);
1409                 if (len == sizeof(xfs_trans_header_t))
1410                         xlog_recover_add_item(&trans->r_itemq);
1411                 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1412                 return 0;
1413         }
1414
1415         ptr = kmem_alloc(len, KM_SLEEP);
1416         memcpy(ptr, dp, len);
1417         in_f = (xfs_inode_log_format_t *)ptr;
1418
1419         if (item->ri_prev->ri_total != 0 &&
1420              item->ri_prev->ri_total == item->ri_prev->ri_cnt) {
1421                 xlog_recover_add_item(&trans->r_itemq);
1422         }
1423         item = trans->r_itemq;
1424         item = item->ri_prev;
1425
1426         if (item->ri_total == 0) {              /* first region to be added */
1427                 item->ri_total  = in_f->ilf_size;
1428                 ASSERT(item->ri_total <= XLOG_MAX_REGIONS_IN_ITEM);
1429                 item->ri_buf = kmem_zalloc((item->ri_total *
1430                                             sizeof(xfs_log_iovec_t)), KM_SLEEP);
1431         }
1432         ASSERT(item->ri_total > item->ri_cnt);
1433         /* Description region is ri_buf[0] */
1434         item->ri_buf[item->ri_cnt].i_addr = ptr;
1435         item->ri_buf[item->ri_cnt].i_len  = len;
1436         item->ri_cnt++;
1437         return 0;
1438 }
1439
1440 STATIC void
1441 xlog_recover_new_tid(
1442         xlog_recover_t          **q,
1443         xlog_tid_t              tid,
1444         xfs_lsn_t               lsn)
1445 {
1446         xlog_recover_t          *trans;
1447
1448         trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1449         trans->r_log_tid   = tid;
1450         trans->r_lsn       = lsn;
1451         xlog_recover_put_hashq(q, trans);
1452 }
1453
1454 STATIC int
1455 xlog_recover_unlink_tid(
1456         xlog_recover_t          **q,
1457         xlog_recover_t          *trans)
1458 {
1459         xlog_recover_t          *tp;
1460         int                     found = 0;
1461
1462         ASSERT(trans != 0);
1463         if (trans == *q) {
1464                 *q = (*q)->r_next;
1465         } else {
1466                 tp = *q;
1467                 while (tp != 0) {
1468                         if (tp->r_next == trans) {
1469                                 found = 1;
1470                                 break;
1471                         }
1472                         tp = tp->r_next;
1473                 }
1474                 if (!found) {
1475                         xlog_warn(
1476                              "XFS: xlog_recover_unlink_tid: trans not found");
1477                         ASSERT(0);
1478                         return XFS_ERROR(EIO);
1479                 }
1480                 tp->r_next = tp->r_next->r_next;
1481         }
1482         return 0;
1483 }
1484
1485 STATIC void
1486 xlog_recover_insert_item_backq(
1487         xlog_recover_item_t     **q,
1488         xlog_recover_item_t     *item)
1489 {
1490         if (*q == 0) {
1491                 item->ri_prev = item->ri_next = item;
1492                 *q = item;
1493         } else {
1494                 item->ri_next           = *q;
1495                 item->ri_prev           = (*q)->ri_prev;
1496                 (*q)->ri_prev           = item;
1497                 item->ri_prev->ri_next  = item;
1498         }
1499 }
1500
1501 STATIC void
1502 xlog_recover_insert_item_frontq(
1503         xlog_recover_item_t     **q,
1504         xlog_recover_item_t     *item)
1505 {
1506         xlog_recover_insert_item_backq(q, item);
1507         *q = item;
1508 }
1509
1510 STATIC int
1511 xlog_recover_reorder_trans(
1512         xlog_recover_t          *trans)
1513 {
1514         xlog_recover_item_t     *first_item, *itemq, *itemq_next;
1515         xfs_buf_log_format_t    *buf_f;
1516         ushort                  flags = 0;
1517
1518         first_item = itemq = trans->r_itemq;
1519         trans->r_itemq = NULL;
1520         do {
1521                 itemq_next = itemq->ri_next;
1522                 buf_f = (xfs_buf_log_format_t *)itemq->ri_buf[0].i_addr;
1523
1524                 switch (ITEM_TYPE(itemq)) {
1525                 case XFS_LI_BUF:
1526                         flags = buf_f->blf_flags;
1527                         if (!(flags & XFS_BLI_CANCEL)) {
1528                                 xlog_recover_insert_item_frontq(&trans->r_itemq,
1529                                                                 itemq);
1530                                 break;
1531                         }
1532                 case XFS_LI_INODE:
1533                 case XFS_LI_DQUOT:
1534                 case XFS_LI_QUOTAOFF:
1535                 case XFS_LI_EFD:
1536                 case XFS_LI_EFI:
1537                         xlog_recover_insert_item_backq(&trans->r_itemq, itemq);
1538                         break;
1539                 default:
1540                         xlog_warn(
1541         "XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1542                         ASSERT(0);
1543                         return XFS_ERROR(EIO);
1544                 }
1545                 itemq = itemq_next;
1546         } while (first_item != itemq);
1547         return 0;
1548 }
1549
1550 /*
1551  * Build up the table of buf cancel records so that we don't replay
1552  * cancelled data in the second pass.  For buffer records that are
1553  * not cancel records, there is nothing to do here so we just return.
1554  *
1555  * If we get a cancel record which is already in the table, this indicates
1556  * that the buffer was cancelled multiple times.  In order to ensure
1557  * that during pass 2 we keep the record in the table until we reach its
1558  * last occurrence in the log, we keep a reference count in the cancel
1559  * record in the table to tell us how many times we expect to see this
1560  * record during the second pass.
1561  */
1562 STATIC void
1563 xlog_recover_do_buffer_pass1(
1564         xlog_t                  *log,
1565         xfs_buf_log_format_t    *buf_f)
1566 {
1567         xfs_buf_cancel_t        *bcp;
1568         xfs_buf_cancel_t        *nextp;
1569         xfs_buf_cancel_t        *prevp;
1570         xfs_buf_cancel_t        **bucket;
1571         xfs_daddr_t             blkno = 0;
1572         uint                    len = 0;
1573         ushort                  flags = 0;
1574
1575         switch (buf_f->blf_type) {
1576         case XFS_LI_BUF:
1577                 blkno = buf_f->blf_blkno;
1578                 len = buf_f->blf_len;
1579                 flags = buf_f->blf_flags;
1580                 break;
1581         }
1582
1583         /*
1584          * If this isn't a cancel buffer item, then just return.
1585          */
1586         if (!(flags & XFS_BLI_CANCEL))
1587                 return;
1588
1589         /*
1590          * Insert an xfs_buf_cancel record into the hash table of
1591          * them.  If there is already an identical record, bump
1592          * its reference count.
1593          */
1594         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1595                                           XLOG_BC_TABLE_SIZE];
1596         /*
1597          * If the hash bucket is empty then just insert a new record into
1598          * the bucket.
1599          */
1600         if (*bucket == NULL) {
1601                 bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1602                                                      KM_SLEEP);
1603                 bcp->bc_blkno = blkno;
1604                 bcp->bc_len = len;
1605                 bcp->bc_refcount = 1;
1606                 bcp->bc_next = NULL;
1607                 *bucket = bcp;
1608                 return;
1609         }
1610
1611         /*
1612          * The hash bucket is not empty, so search for duplicates of our
1613          * record.  If we find one them just bump its refcount.  If not
1614          * then add us at the end of the list.
1615          */
1616         prevp = NULL;
1617         nextp = *bucket;
1618         while (nextp != NULL) {
1619                 if (nextp->bc_blkno == blkno && nextp->bc_len == len) {
1620                         nextp->bc_refcount++;
1621                         return;
1622                 }
1623                 prevp = nextp;
1624                 nextp = nextp->bc_next;
1625         }
1626         ASSERT(prevp != NULL);
1627         bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1628                                              KM_SLEEP);
1629         bcp->bc_blkno = blkno;
1630         bcp->bc_len = len;
1631         bcp->bc_refcount = 1;
1632         bcp->bc_next = NULL;
1633         prevp->bc_next = bcp;
1634 }
1635
1636 /*
1637  * Check to see whether the buffer being recovered has a corresponding
1638  * entry in the buffer cancel record table.  If it does then return 1
1639  * so that it will be cancelled, otherwise return 0.  If the buffer is
1640  * actually a buffer cancel item (XFS_BLI_CANCEL is set), then decrement
1641  * the refcount on the entry in the table and remove it from the table
1642  * if this is the last reference.
1643  *
1644  * We remove the cancel record from the table when we encounter its
1645  * last occurrence in the log so that if the same buffer is re-used
1646  * again after its last cancellation we actually replay the changes
1647  * made at that point.
1648  */
1649 STATIC int
1650 xlog_check_buffer_cancelled(
1651         xlog_t                  *log,
1652         xfs_daddr_t             blkno,
1653         uint                    len,
1654         ushort                  flags)
1655 {
1656         xfs_buf_cancel_t        *bcp;
1657         xfs_buf_cancel_t        *prevp;
1658         xfs_buf_cancel_t        **bucket;
1659
1660         if (log->l_buf_cancel_table == NULL) {
1661                 /*
1662                  * There is nothing in the table built in pass one,
1663                  * so this buffer must not be cancelled.
1664                  */
1665                 ASSERT(!(flags & XFS_BLI_CANCEL));
1666                 return 0;
1667         }
1668
1669         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1670                                           XLOG_BC_TABLE_SIZE];
1671         bcp = *bucket;
1672         if (bcp == NULL) {
1673                 /*
1674                  * There is no corresponding entry in the table built
1675                  * in pass one, so this buffer has not been cancelled.
1676                  */
1677                 ASSERT(!(flags & XFS_BLI_CANCEL));
1678                 return 0;
1679         }
1680
1681         /*
1682          * Search for an entry in the buffer cancel table that
1683          * matches our buffer.
1684          */
1685         prevp = NULL;
1686         while (bcp != NULL) {
1687                 if (bcp->bc_blkno == blkno && bcp->bc_len == len) {
1688                         /*
1689                          * We've go a match, so return 1 so that the
1690                          * recovery of this buffer is cancelled.
1691                          * If this buffer is actually a buffer cancel
1692                          * log item, then decrement the refcount on the
1693                          * one in the table and remove it if this is the
1694                          * last reference.
1695                          */
1696                         if (flags & XFS_BLI_CANCEL) {
1697                                 bcp->bc_refcount--;
1698                                 if (bcp->bc_refcount == 0) {
1699                                         if (prevp == NULL) {
1700                                                 *bucket = bcp->bc_next;
1701                                         } else {
1702                                                 prevp->bc_next = bcp->bc_next;
1703                                         }
1704                                         kmem_free(bcp,
1705                                                   sizeof(xfs_buf_cancel_t));
1706                                 }
1707                         }
1708                         return 1;
1709                 }
1710                 prevp = bcp;
1711                 bcp = bcp->bc_next;
1712         }
1713         /*
1714          * We didn't find a corresponding entry in the table, so
1715          * return 0 so that the buffer is NOT cancelled.
1716          */
1717         ASSERT(!(flags & XFS_BLI_CANCEL));
1718         return 0;
1719 }
1720
1721 STATIC int
1722 xlog_recover_do_buffer_pass2(
1723         xlog_t                  *log,
1724         xfs_buf_log_format_t    *buf_f)
1725 {
1726         xfs_daddr_t             blkno = 0;
1727         ushort                  flags = 0;
1728         uint                    len = 0;
1729
1730         switch (buf_f->blf_type) {
1731         case XFS_LI_BUF:
1732                 blkno = buf_f->blf_blkno;
1733                 flags = buf_f->blf_flags;
1734                 len = buf_f->blf_len;
1735                 break;
1736         }
1737
1738         return xlog_check_buffer_cancelled(log, blkno, len, flags);
1739 }
1740
1741 /*
1742  * Perform recovery for a buffer full of inodes.  In these buffers,
1743  * the only data which should be recovered is that which corresponds
1744  * to the di_next_unlinked pointers in the on disk inode structures.
1745  * The rest of the data for the inodes is always logged through the
1746  * inodes themselves rather than the inode buffer and is recovered
1747  * in xlog_recover_do_inode_trans().
1748  *
1749  * The only time when buffers full of inodes are fully recovered is
1750  * when the buffer is full of newly allocated inodes.  In this case
1751  * the buffer will not be marked as an inode buffer and so will be
1752  * sent to xlog_recover_do_reg_buffer() below during recovery.
1753  */
1754 STATIC int
1755 xlog_recover_do_inode_buffer(
1756         xfs_mount_t             *mp,
1757         xlog_recover_item_t     *item,
1758         xfs_buf_t               *bp,
1759         xfs_buf_log_format_t    *buf_f)
1760 {
1761         int                     i;
1762         int                     item_index;
1763         int                     bit;
1764         int                     nbits;
1765         int                     reg_buf_offset;
1766         int                     reg_buf_bytes;
1767         int                     next_unlinked_offset;
1768         int                     inodes_per_buf;
1769         xfs_agino_t             *logged_nextp;
1770         xfs_agino_t             *buffer_nextp;
1771         unsigned int            *data_map = NULL;
1772         unsigned int            map_size = 0;
1773
1774         switch (buf_f->blf_type) {
1775         case XFS_LI_BUF:
1776                 data_map = buf_f->blf_data_map;
1777                 map_size = buf_f->blf_map_size;
1778                 break;
1779         }
1780         /*
1781          * Set the variables corresponding to the current region to
1782          * 0 so that we'll initialize them on the first pass through
1783          * the loop.
1784          */
1785         reg_buf_offset = 0;
1786         reg_buf_bytes = 0;
1787         bit = 0;
1788         nbits = 0;
1789         item_index = 0;
1790         inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1791         for (i = 0; i < inodes_per_buf; i++) {
1792                 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1793                         offsetof(xfs_dinode_t, di_next_unlinked);
1794
1795                 while (next_unlinked_offset >=
1796                        (reg_buf_offset + reg_buf_bytes)) {
1797                         /*
1798                          * The next di_next_unlinked field is beyond
1799                          * the current logged region.  Find the next
1800                          * logged region that contains or is beyond
1801                          * the current di_next_unlinked field.
1802                          */
1803                         bit += nbits;
1804                         bit = xfs_next_bit(data_map, map_size, bit);
1805
1806                         /*
1807                          * If there are no more logged regions in the
1808                          * buffer, then we're done.
1809                          */
1810                         if (bit == -1) {
1811                                 return 0;
1812                         }
1813
1814                         nbits = xfs_contig_bits(data_map, map_size,
1815                                                          bit);
1816                         ASSERT(nbits > 0);
1817                         reg_buf_offset = bit << XFS_BLI_SHIFT;
1818                         reg_buf_bytes = nbits << XFS_BLI_SHIFT;
1819                         item_index++;
1820                 }
1821
1822                 /*
1823                  * If the current logged region starts after the current
1824                  * di_next_unlinked field, then move on to the next
1825                  * di_next_unlinked field.
1826                  */
1827                 if (next_unlinked_offset < reg_buf_offset) {
1828                         continue;
1829                 }
1830
1831                 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1832                 ASSERT((item->ri_buf[item_index].i_len % XFS_BLI_CHUNK) == 0);
1833                 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1834
1835                 /*
1836                  * The current logged region contains a copy of the
1837                  * current di_next_unlinked field.  Extract its value
1838                  * and copy it to the buffer copy.
1839                  */
1840                 logged_nextp = (xfs_agino_t *)
1841                                ((char *)(item->ri_buf[item_index].i_addr) +
1842                                 (next_unlinked_offset - reg_buf_offset));
1843                 if (unlikely(*logged_nextp == 0)) {
1844                         xfs_fs_cmn_err(CE_ALERT, mp,
1845                                 "bad inode buffer log record (ptr = 0x%p, bp = 0x%p).  XFS trying to replay bad (0) inode di_next_unlinked field",
1846                                 item, bp);
1847                         XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1848                                          XFS_ERRLEVEL_LOW, mp);
1849                         return XFS_ERROR(EFSCORRUPTED);
1850                 }
1851
1852                 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1853                                               next_unlinked_offset);
1854                 *buffer_nextp = *logged_nextp;
1855         }
1856
1857         return 0;
1858 }
1859
1860 /*
1861  * Perform a 'normal' buffer recovery.  Each logged region of the
1862  * buffer should be copied over the corresponding region in the
1863  * given buffer.  The bitmap in the buf log format structure indicates
1864  * where to place the logged data.
1865  */
1866 /*ARGSUSED*/
1867 STATIC void
1868 xlog_recover_do_reg_buffer(
1869         xlog_recover_item_t     *item,
1870         xfs_buf_t               *bp,
1871         xfs_buf_log_format_t    *buf_f)
1872 {
1873         int                     i;
1874         int                     bit;
1875         int                     nbits;
1876         unsigned int            *data_map = NULL;
1877         unsigned int            map_size = 0;
1878         int                     error;
1879
1880         switch (buf_f->blf_type) {
1881         case XFS_LI_BUF:
1882                 data_map = buf_f->blf_data_map;
1883                 map_size = buf_f->blf_map_size;
1884                 break;
1885         }
1886         bit = 0;
1887         i = 1;  /* 0 is the buf format structure */
1888         while (1) {
1889                 bit = xfs_next_bit(data_map, map_size, bit);
1890                 if (bit == -1)
1891                         break;
1892                 nbits = xfs_contig_bits(data_map, map_size, bit);
1893                 ASSERT(nbits > 0);
1894                 ASSERT(item->ri_buf[i].i_addr != 0);
1895                 ASSERT(item->ri_buf[i].i_len % XFS_BLI_CHUNK == 0);
1896                 ASSERT(XFS_BUF_COUNT(bp) >=
1897                        ((uint)bit << XFS_BLI_SHIFT)+(nbits<<XFS_BLI_SHIFT));
1898
1899                 /*
1900                  * Do a sanity check if this is a dquot buffer. Just checking
1901                  * the first dquot in the buffer should do. XXXThis is
1902                  * probably a good thing to do for other buf types also.
1903                  */
1904                 error = 0;
1905                 if (buf_f->blf_flags &
1906                    (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
1907                         error = xfs_qm_dqcheck((xfs_disk_dquot_t *)
1908                                                item->ri_buf[i].i_addr,
1909                                                -1, 0, XFS_QMOPT_DOWARN,
1910                                                "dquot_buf_recover");
1911                 }
1912                 if (!error)
1913                         memcpy(xfs_buf_offset(bp,
1914                                 (uint)bit << XFS_BLI_SHIFT),    /* dest */
1915                                 item->ri_buf[i].i_addr,         /* source */
1916                                 nbits<<XFS_BLI_SHIFT);          /* length */
1917                 i++;
1918                 bit += nbits;
1919         }
1920
1921         /* Shouldn't be any more regions */
1922         ASSERT(i == item->ri_total);
1923 }
1924
1925 /*
1926  * Do some primitive error checking on ondisk dquot data structures.
1927  */
1928 int
1929 xfs_qm_dqcheck(
1930         xfs_disk_dquot_t *ddq,
1931         xfs_dqid_t       id,
1932         uint             type,    /* used only when IO_dorepair is true */
1933         uint             flags,
1934         char             *str)
1935 {
1936         xfs_dqblk_t      *d = (xfs_dqblk_t *)ddq;
1937         int             errs = 0;
1938
1939         /*
1940          * We can encounter an uninitialized dquot buffer for 2 reasons:
1941          * 1. If we crash while deleting the quotainode(s), and those blks got
1942          *    used for user data. This is because we take the path of regular
1943          *    file deletion; however, the size field of quotainodes is never
1944          *    updated, so all the tricks that we play in itruncate_finish
1945          *    don't quite matter.
1946          *
1947          * 2. We don't play the quota buffers when there's a quotaoff logitem.
1948          *    But the allocation will be replayed so we'll end up with an
1949          *    uninitialized quota block.
1950          *
1951          * This is all fine; things are still consistent, and we haven't lost
1952          * any quota information. Just don't complain about bad dquot blks.
1953          */
1954         if (be16_to_cpu(ddq->d_magic) != XFS_DQUOT_MAGIC) {
1955                 if (flags & XFS_QMOPT_DOWARN)
1956                         cmn_err(CE_ALERT,
1957                         "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
1958                         str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
1959                 errs++;
1960         }
1961         if (ddq->d_version != XFS_DQUOT_VERSION) {
1962                 if (flags & XFS_QMOPT_DOWARN)
1963                         cmn_err(CE_ALERT,
1964                         "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
1965                         str, id, ddq->d_version, XFS_DQUOT_VERSION);
1966                 errs++;
1967         }
1968
1969         if (ddq->d_flags != XFS_DQ_USER &&
1970             ddq->d_flags != XFS_DQ_PROJ &&
1971             ddq->d_flags != XFS_DQ_GROUP) {
1972                 if (flags & XFS_QMOPT_DOWARN)
1973                         cmn_err(CE_ALERT,
1974                         "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
1975                         str, id, ddq->d_flags);
1976                 errs++;
1977         }
1978
1979         if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
1980                 if (flags & XFS_QMOPT_DOWARN)
1981                         cmn_err(CE_ALERT,
1982                         "%s : ondisk-dquot 0x%p, ID mismatch: "
1983                         "0x%x expected, found id 0x%x",
1984                         str, ddq, id, be32_to_cpu(ddq->d_id));
1985                 errs++;
1986         }
1987
1988         if (!errs && ddq->d_id) {
1989                 if (ddq->d_blk_softlimit &&
1990                     be64_to_cpu(ddq->d_bcount) >=
1991                                 be64_to_cpu(ddq->d_blk_softlimit)) {
1992                         if (!ddq->d_btimer) {
1993                                 if (flags & XFS_QMOPT_DOWARN)
1994                                         cmn_err(CE_ALERT,
1995                                         "%s : Dquot ID 0x%x (0x%p) "
1996                                         "BLK TIMER NOT STARTED",
1997                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
1998                                 errs++;
1999                         }
2000                 }
2001                 if (ddq->d_ino_softlimit &&
2002                     be64_to_cpu(ddq->d_icount) >=
2003                                 be64_to_cpu(ddq->d_ino_softlimit)) {
2004                         if (!ddq->d_itimer) {
2005                                 if (flags & XFS_QMOPT_DOWARN)
2006                                         cmn_err(CE_ALERT,
2007                                         "%s : Dquot ID 0x%x (0x%p) "
2008                                         "INODE TIMER NOT STARTED",
2009                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2010                                 errs++;
2011                         }
2012                 }
2013                 if (ddq->d_rtb_softlimit &&
2014                     be64_to_cpu(ddq->d_rtbcount) >=
2015                                 be64_to_cpu(ddq->d_rtb_softlimit)) {
2016                         if (!ddq->d_rtbtimer) {
2017                                 if (flags & XFS_QMOPT_DOWARN)
2018                                         cmn_err(CE_ALERT,
2019                                         "%s : Dquot ID 0x%x (0x%p) "
2020                                         "RTBLK TIMER NOT STARTED",
2021                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2022                                 errs++;
2023                         }
2024                 }
2025         }
2026
2027         if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2028                 return errs;
2029
2030         if (flags & XFS_QMOPT_DOWARN)
2031                 cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2032
2033         /*
2034          * Typically, a repair is only requested by quotacheck.
2035          */
2036         ASSERT(id != -1);
2037         ASSERT(flags & XFS_QMOPT_DQREPAIR);
2038         memset(d, 0, sizeof(xfs_dqblk_t));
2039
2040         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2041         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2042         d->dd_diskdq.d_flags = type;
2043         d->dd_diskdq.d_id = cpu_to_be32(id);
2044
2045         return errs;
2046 }
2047
2048 /*
2049  * Perform a dquot buffer recovery.
2050  * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2051  * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2052  * Else, treat it as a regular buffer and do recovery.
2053  */
2054 STATIC void
2055 xlog_recover_do_dquot_buffer(
2056         xfs_mount_t             *mp,
2057         xlog_t                  *log,
2058         xlog_recover_item_t     *item,
2059         xfs_buf_t               *bp,
2060         xfs_buf_log_format_t    *buf_f)
2061 {
2062         uint                    type;
2063
2064         /*
2065          * Filesystems are required to send in quota flags at mount time.
2066          */
2067         if (mp->m_qflags == 0) {
2068                 return;
2069         }
2070
2071         type = 0;
2072         if (buf_f->blf_flags & XFS_BLI_UDQUOT_BUF)
2073                 type |= XFS_DQ_USER;
2074         if (buf_f->blf_flags & XFS_BLI_PDQUOT_BUF)
2075                 type |= XFS_DQ_PROJ;
2076         if (buf_f->blf_flags & XFS_BLI_GDQUOT_BUF)
2077                 type |= XFS_DQ_GROUP;
2078         /*
2079          * This type of quotas was turned off, so ignore this buffer
2080          */
2081         if (log->l_quotaoffs_flag & type)
2082                 return;
2083
2084         xlog_recover_do_reg_buffer(item, bp, buf_f);
2085 }
2086
2087 /*
2088  * This routine replays a modification made to a buffer at runtime.
2089  * There are actually two types of buffer, regular and inode, which
2090  * are handled differently.  Inode buffers are handled differently
2091  * in that we only recover a specific set of data from them, namely
2092  * the inode di_next_unlinked fields.  This is because all other inode
2093  * data is actually logged via inode records and any data we replay
2094  * here which overlaps that may be stale.
2095  *
2096  * When meta-data buffers are freed at run time we log a buffer item
2097  * with the XFS_BLI_CANCEL bit set to indicate that previous copies
2098  * of the buffer in the log should not be replayed at recovery time.
2099  * This is so that if the blocks covered by the buffer are reused for
2100  * file data before we crash we don't end up replaying old, freed
2101  * meta-data into a user's file.
2102  *
2103  * To handle the cancellation of buffer log items, we make two passes
2104  * over the log during recovery.  During the first we build a table of
2105  * those buffers which have been cancelled, and during the second we
2106  * only replay those buffers which do not have corresponding cancel
2107  * records in the table.  See xlog_recover_do_buffer_pass[1,2] above
2108  * for more details on the implementation of the table of cancel records.
2109  */
2110 STATIC int
2111 xlog_recover_do_buffer_trans(
2112         xlog_t                  *log,
2113         xlog_recover_item_t     *item,
2114         int                     pass)
2115 {
2116         xfs_buf_log_format_t    *buf_f;
2117         xfs_mount_t             *mp;
2118         xfs_buf_t               *bp;
2119         int                     error;
2120         int                     cancel;
2121         xfs_daddr_t             blkno;
2122         int                     len;
2123         ushort                  flags;
2124
2125         buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
2126
2127         if (pass == XLOG_RECOVER_PASS1) {
2128                 /*
2129                  * In this pass we're only looking for buf items
2130                  * with the XFS_BLI_CANCEL bit set.
2131                  */
2132                 xlog_recover_do_buffer_pass1(log, buf_f);
2133                 return 0;
2134         } else {
2135                 /*
2136                  * In this pass we want to recover all the buffers
2137                  * which have not been cancelled and are not
2138                  * cancellation buffers themselves.  The routine
2139                  * we call here will tell us whether or not to
2140                  * continue with the replay of this buffer.
2141                  */
2142                 cancel = xlog_recover_do_buffer_pass2(log, buf_f);
2143                 if (cancel) {
2144                         return 0;
2145                 }
2146         }
2147         switch (buf_f->blf_type) {
2148         case XFS_LI_BUF:
2149                 blkno = buf_f->blf_blkno;
2150                 len = buf_f->blf_len;
2151                 flags = buf_f->blf_flags;
2152                 break;
2153         default:
2154                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
2155                         "xfs_log_recover: unknown buffer type 0x%x, logdev %s",
2156                         buf_f->blf_type, log->l_mp->m_logname ?
2157                         log->l_mp->m_logname : "internal");
2158                 XFS_ERROR_REPORT("xlog_recover_do_buffer_trans",
2159                                  XFS_ERRLEVEL_LOW, log->l_mp);
2160                 return XFS_ERROR(EFSCORRUPTED);
2161         }
2162
2163         mp = log->l_mp;
2164         if (flags & XFS_BLI_INODE_BUF) {
2165                 bp = xfs_buf_read_flags(mp->m_ddev_targp, blkno, len,
2166                                                                 XFS_BUF_LOCK);
2167         } else {
2168                 bp = xfs_buf_read(mp->m_ddev_targp, blkno, len, 0);
2169         }
2170         if (XFS_BUF_ISERROR(bp)) {
2171                 xfs_ioerror_alert("xlog_recover_do..(read#1)", log->l_mp,
2172                                   bp, blkno);
2173                 error = XFS_BUF_GETERROR(bp);
2174                 xfs_buf_relse(bp);
2175                 return error;
2176         }
2177
2178         error = 0;
2179         if (flags & XFS_BLI_INODE_BUF) {
2180                 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2181         } else if (flags &
2182                   (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
2183                 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2184         } else {
2185                 xlog_recover_do_reg_buffer(item, bp, buf_f);
2186         }
2187         if (error)
2188                 return XFS_ERROR(error);
2189
2190         /*
2191          * Perform delayed write on the buffer.  Asynchronous writes will be
2192          * slower when taking into account all the buffers to be flushed.
2193          *
2194          * Also make sure that only inode buffers with good sizes stay in
2195          * the buffer cache.  The kernel moves inodes in buffers of 1 block
2196          * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger.  The inode
2197          * buffers in the log can be a different size if the log was generated
2198          * by an older kernel using unclustered inode buffers or a newer kernel
2199          * running with a different inode cluster size.  Regardless, if the
2200          * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2201          * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2202          * the buffer out of the buffer cache so that the buffer won't
2203          * overlap with future reads of those inodes.
2204          */
2205         if (XFS_DINODE_MAGIC ==
2206             INT_GET(*((__uint16_t *)(xfs_buf_offset(bp, 0))), ARCH_CONVERT) &&
2207             (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2208                         (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2209                 XFS_BUF_STALE(bp);
2210                 error = xfs_bwrite(mp, bp);
2211         } else {
2212                 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2213                        XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2214                 XFS_BUF_SET_FSPRIVATE(bp, mp);
2215                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2216                 xfs_bdwrite(mp, bp);
2217         }
2218
2219         return (error);
2220 }
2221
2222 STATIC int
2223 xlog_recover_do_inode_trans(
2224         xlog_t                  *log,
2225         xlog_recover_item_t     *item,
2226         int                     pass)
2227 {
2228         xfs_inode_log_format_t  *in_f;
2229         xfs_mount_t             *mp;
2230         xfs_buf_t               *bp;
2231         xfs_imap_t              imap;
2232         xfs_dinode_t            *dip;
2233         xfs_ino_t               ino;
2234         int                     len;
2235         xfs_caddr_t             src;
2236         xfs_caddr_t             dest;
2237         int                     error;
2238         int                     attr_index;
2239         uint                    fields;
2240         xfs_dinode_core_t       *dicp;
2241         int                     need_free = 0;
2242
2243         if (pass == XLOG_RECOVER_PASS1) {
2244                 return 0;
2245         }
2246
2247         if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2248                 in_f = (xfs_inode_log_format_t *)item->ri_buf[0].i_addr;
2249         } else {
2250                 in_f = (xfs_inode_log_format_t *)kmem_alloc(
2251                         sizeof(xfs_inode_log_format_t), KM_SLEEP);
2252                 need_free = 1;
2253                 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2254                 if (error)
2255                         goto error;
2256         }
2257         ino = in_f->ilf_ino;
2258         mp = log->l_mp;
2259         if (ITEM_TYPE(item) == XFS_LI_INODE) {
2260                 imap.im_blkno = (xfs_daddr_t)in_f->ilf_blkno;
2261                 imap.im_len = in_f->ilf_len;
2262                 imap.im_boffset = in_f->ilf_boffset;
2263         } else {
2264                 /*
2265                  * It's an old inode format record.  We don't know where
2266                  * its cluster is located on disk, and we can't allow
2267                  * xfs_imap() to figure it out because the inode btrees
2268                  * are not ready to be used.  Therefore do not pass the
2269                  * XFS_IMAP_LOOKUP flag to xfs_imap().  This will give
2270                  * us only the single block in which the inode lives
2271                  * rather than its cluster, so we must make sure to
2272                  * invalidate the buffer when we write it out below.
2273                  */
2274                 imap.im_blkno = 0;
2275                 xfs_imap(log->l_mp, NULL, ino, &imap, 0);
2276         }
2277
2278         /*
2279          * Inode buffers can be freed, look out for it,
2280          * and do not replay the inode.
2281          */
2282         if (xlog_check_buffer_cancelled(log, imap.im_blkno, imap.im_len, 0)) {
2283                 error = 0;
2284                 goto error;
2285         }
2286
2287         bp = xfs_buf_read_flags(mp->m_ddev_targp, imap.im_blkno, imap.im_len,
2288                                                                 XFS_BUF_LOCK);
2289         if (XFS_BUF_ISERROR(bp)) {
2290                 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2291                                   bp, imap.im_blkno);
2292                 error = XFS_BUF_GETERROR(bp);
2293                 xfs_buf_relse(bp);
2294                 goto error;
2295         }
2296         error = 0;
2297         ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2298         dip = (xfs_dinode_t *)xfs_buf_offset(bp, imap.im_boffset);
2299
2300         /*
2301          * Make sure the place we're flushing out to really looks
2302          * like an inode!
2303          */
2304         if (unlikely(INT_GET(dip->di_core.di_magic, ARCH_CONVERT) != XFS_DINODE_MAGIC)) {
2305                 xfs_buf_relse(bp);
2306                 xfs_fs_cmn_err(CE_ALERT, mp,
2307                         "xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2308                         dip, bp, ino);
2309                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(1)",
2310                                  XFS_ERRLEVEL_LOW, mp);
2311                 error = EFSCORRUPTED;
2312                 goto error;
2313         }
2314         dicp = (xfs_dinode_core_t*)(item->ri_buf[1].i_addr);
2315         if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2316                 xfs_buf_relse(bp);
2317                 xfs_fs_cmn_err(CE_ALERT, mp,
2318                         "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2319                         item, ino);
2320                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(2)",
2321                                  XFS_ERRLEVEL_LOW, mp);
2322                 error = EFSCORRUPTED;
2323                 goto error;
2324         }
2325
2326         /* Skip replay when the on disk inode is newer than the log one */
2327         if (dicp->di_flushiter <
2328             INT_GET(dip->di_core.di_flushiter, ARCH_CONVERT)) {
2329                 /*
2330                  * Deal with the wrap case, DI_MAX_FLUSH is less
2331                  * than smaller numbers
2332                  */
2333                 if ((INT_GET(dip->di_core.di_flushiter, ARCH_CONVERT)
2334                                                         == DI_MAX_FLUSH) &&
2335                     (dicp->di_flushiter < (DI_MAX_FLUSH>>1))) {
2336                         /* do nothing */
2337                 } else {
2338                         xfs_buf_relse(bp);
2339                         error = 0;
2340                         goto error;
2341                 }
2342         }
2343         /* Take the opportunity to reset the flush iteration count */
2344         dicp->di_flushiter = 0;
2345
2346         if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2347                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2348                     (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2349                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(3)",
2350                                          XFS_ERRLEVEL_LOW, mp, dicp);
2351                         xfs_buf_relse(bp);
2352                         xfs_fs_cmn_err(CE_ALERT, mp,
2353                                 "xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2354                                 item, dip, bp, ino);
2355                         error = EFSCORRUPTED;
2356                         goto error;
2357                 }
2358         } else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2359                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2360                     (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2361                     (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2362                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(4)",
2363                                              XFS_ERRLEVEL_LOW, mp, dicp);
2364                         xfs_buf_relse(bp);
2365                         xfs_fs_cmn_err(CE_ALERT, mp,
2366                                 "xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2367                                 item, dip, bp, ino);
2368                         error = EFSCORRUPTED;
2369                         goto error;
2370                 }
2371         }
2372         if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2373                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(5)",
2374                                      XFS_ERRLEVEL_LOW, mp, dicp);
2375                 xfs_buf_relse(bp);
2376                 xfs_fs_cmn_err(CE_ALERT, mp,
2377                         "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",
2378                         item, dip, bp, ino,
2379                         dicp->di_nextents + dicp->di_anextents,
2380                         dicp->di_nblocks);
2381                 error = EFSCORRUPTED;
2382                 goto error;
2383         }
2384         if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2385                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(6)",
2386                                      XFS_ERRLEVEL_LOW, mp, dicp);
2387                 xfs_buf_relse(bp);
2388                 xfs_fs_cmn_err(CE_ALERT, mp,
2389                         "xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2390                         item, dip, bp, ino, dicp->di_forkoff);
2391                 error = EFSCORRUPTED;
2392                 goto error;
2393         }
2394         if (unlikely(item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t))) {
2395                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(7)",
2396                                      XFS_ERRLEVEL_LOW, mp, dicp);
2397                 xfs_buf_relse(bp);
2398                 xfs_fs_cmn_err(CE_ALERT, mp,
2399                         "xfs_inode_recover: Bad inode log record length %d, rec ptr 0x%p",
2400                         item->ri_buf[1].i_len, item);
2401                 error = EFSCORRUPTED;
2402                 goto error;
2403         }
2404
2405         /* The core is in in-core format */
2406         xfs_xlate_dinode_core((xfs_caddr_t)&dip->di_core,
2407                               (xfs_dinode_core_t*)item->ri_buf[1].i_addr, -1);
2408
2409         /* the rest is in on-disk format */
2410         if (item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t)) {
2411                 memcpy((xfs_caddr_t) dip + sizeof(xfs_dinode_core_t),
2412                         item->ri_buf[1].i_addr + sizeof(xfs_dinode_core_t),
2413                         item->ri_buf[1].i_len  - sizeof(xfs_dinode_core_t));
2414         }
2415
2416         fields = in_f->ilf_fields;
2417         switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2418         case XFS_ILOG_DEV:
2419                 INT_SET(dip->di_u.di_dev, ARCH_CONVERT, in_f->ilf_u.ilfu_rdev);
2420
2421                 break;
2422         case XFS_ILOG_UUID:
2423                 dip->di_u.di_muuid = in_f->ilf_u.ilfu_uuid;
2424                 break;
2425         }
2426
2427         if (in_f->ilf_size == 2)
2428                 goto write_inode_buffer;
2429         len = item->ri_buf[2].i_len;
2430         src = item->ri_buf[2].i_addr;
2431         ASSERT(in_f->ilf_size <= 4);
2432         ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2433         ASSERT(!(fields & XFS_ILOG_DFORK) ||
2434                (len == in_f->ilf_dsize));
2435
2436         switch (fields & XFS_ILOG_DFORK) {
2437         case XFS_ILOG_DDATA:
2438         case XFS_ILOG_DEXT:
2439                 memcpy(&dip->di_u, src, len);
2440                 break;
2441
2442         case XFS_ILOG_DBROOT:
2443                 xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2444                                  &(dip->di_u.di_bmbt),
2445                                  XFS_DFORK_DSIZE(dip, mp));
2446                 break;
2447
2448         default:
2449                 /*
2450                  * There are no data fork flags set.
2451                  */
2452                 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2453                 break;
2454         }
2455
2456         /*
2457          * If we logged any attribute data, recover it.  There may or
2458          * may not have been any other non-core data logged in this
2459          * transaction.
2460          */
2461         if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2462                 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2463                         attr_index = 3;
2464                 } else {
2465                         attr_index = 2;
2466                 }
2467                 len = item->ri_buf[attr_index].i_len;
2468                 src = item->ri_buf[attr_index].i_addr;
2469                 ASSERT(len == in_f->ilf_asize);
2470
2471                 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2472                 case XFS_ILOG_ADATA:
2473                 case XFS_ILOG_AEXT:
2474                         dest = XFS_DFORK_APTR(dip);
2475                         ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2476                         memcpy(dest, src, len);
2477                         break;
2478
2479                 case XFS_ILOG_ABROOT:
2480                         dest = XFS_DFORK_APTR(dip);
2481                         xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2482                                          (xfs_bmdr_block_t*)dest,
2483                                          XFS_DFORK_ASIZE(dip, mp));
2484                         break;
2485
2486                 default:
2487                         xlog_warn("XFS: xlog_recover_do_inode_trans: Invalid flag");
2488                         ASSERT(0);
2489                         xfs_buf_relse(bp);
2490                         error = EIO;
2491                         goto error;
2492                 }
2493         }
2494
2495 write_inode_buffer:
2496         if (ITEM_TYPE(item) == XFS_LI_INODE) {
2497                 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2498                        XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2499                 XFS_BUF_SET_FSPRIVATE(bp, mp);
2500                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2501                 xfs_bdwrite(mp, bp);
2502         } else {
2503                 XFS_BUF_STALE(bp);
2504                 error = xfs_bwrite(mp, bp);
2505         }
2506
2507 error:
2508         if (need_free)
2509                 kmem_free(in_f, sizeof(*in_f));
2510         return XFS_ERROR(error);
2511 }
2512
2513 /*
2514  * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2515  * structure, so that we know not to do any dquot item or dquot buffer recovery,
2516  * of that type.
2517  */
2518 STATIC int
2519 xlog_recover_do_quotaoff_trans(
2520         xlog_t                  *log,
2521         xlog_recover_item_t     *item,
2522         int                     pass)
2523 {
2524         xfs_qoff_logformat_t    *qoff_f;
2525
2526         if (pass == XLOG_RECOVER_PASS2) {
2527                 return (0);
2528         }
2529
2530         qoff_f = (xfs_qoff_logformat_t *)item->ri_buf[0].i_addr;
2531         ASSERT(qoff_f);
2532
2533         /*
2534          * The logitem format's flag tells us if this was user quotaoff,
2535          * group/project quotaoff or both.
2536          */
2537         if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2538                 log->l_quotaoffs_flag |= XFS_DQ_USER;
2539         if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2540                 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2541         if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2542                 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2543
2544         return (0);
2545 }
2546
2547 /*
2548  * Recover a dquot record
2549  */
2550 STATIC int
2551 xlog_recover_do_dquot_trans(
2552         xlog_t                  *log,
2553         xlog_recover_item_t     *item,
2554         int                     pass)
2555 {
2556         xfs_mount_t             *mp;
2557         xfs_buf_t               *bp;
2558         struct xfs_disk_dquot   *ddq, *recddq;
2559         int                     error;
2560         xfs_dq_logformat_t      *dq_f;
2561         uint                    type;
2562
2563         if (pass == XLOG_RECOVER_PASS1) {
2564                 return 0;
2565         }
2566         mp = log->l_mp;
2567
2568         /*
2569          * Filesystems are required to send in quota flags at mount time.
2570          */
2571         if (mp->m_qflags == 0)
2572                 return (0);
2573
2574         recddq = (xfs_disk_dquot_t *)item->ri_buf[1].i_addr;
2575         ASSERT(recddq);
2576         /*
2577          * This type of quotas was turned off, so ignore this record.
2578          */
2579         type = INT_GET(recddq->d_flags, ARCH_CONVERT) &
2580                         (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2581         ASSERT(type);
2582         if (log->l_quotaoffs_flag & type)
2583                 return (0);
2584
2585         /*
2586          * At this point we know that quota was _not_ turned off.
2587          * Since the mount flags are not indicating to us otherwise, this
2588          * must mean that quota is on, and the dquot needs to be replayed.
2589          * Remember that we may not have fully recovered the superblock yet,
2590          * so we can't do the usual trick of looking at the SB quota bits.
2591          *
2592          * The other possibility, of course, is that the quota subsystem was
2593          * removed since the last mount - ENOSYS.
2594          */
2595         dq_f = (xfs_dq_logformat_t *)item->ri_buf[0].i_addr;
2596         ASSERT(dq_f);
2597         if ((error = xfs_qm_dqcheck(recddq,
2598                            dq_f->qlf_id,
2599                            0, XFS_QMOPT_DOWARN,
2600                            "xlog_recover_do_dquot_trans (log copy)"))) {
2601                 return XFS_ERROR(EIO);
2602         }
2603         ASSERT(dq_f->qlf_len == 1);
2604
2605         error = xfs_read_buf(mp, mp->m_ddev_targp,
2606                              dq_f->qlf_blkno,
2607                              XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2608                              0, &bp);
2609         if (error) {
2610                 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2611                                   bp, dq_f->qlf_blkno);
2612                 return error;
2613         }
2614         ASSERT(bp);
2615         ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2616
2617         /*
2618          * At least the magic num portion should be on disk because this
2619          * was among a chunk of dquots created earlier, and we did some
2620          * minimal initialization then.
2621          */
2622         if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2623                            "xlog_recover_do_dquot_trans")) {
2624                 xfs_buf_relse(bp);
2625                 return XFS_ERROR(EIO);
2626         }
2627
2628         memcpy(ddq, recddq, item->ri_buf[1].i_len);
2629
2630         ASSERT(dq_f->qlf_size == 2);
2631         ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2632                XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2633         XFS_BUF_SET_FSPRIVATE(bp, mp);
2634         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2635         xfs_bdwrite(mp, bp);
2636
2637         return (0);
2638 }
2639
2640 /*
2641  * This routine is called to create an in-core extent free intent
2642  * item from the efi format structure which was logged on disk.
2643  * It allocates an in-core efi, copies the extents from the format
2644  * structure into it, and adds the efi to the AIL with the given
2645  * LSN.
2646  */
2647 STATIC int
2648 xlog_recover_do_efi_trans(
2649         xlog_t                  *log,
2650         xlog_recover_item_t     *item,
2651         xfs_lsn_t               lsn,
2652         int                     pass)
2653 {
2654         int                     error;
2655         xfs_mount_t             *mp;
2656         xfs_efi_log_item_t      *efip;
2657         xfs_efi_log_format_t    *efi_formatp;
2658         SPLDECL(s);
2659
2660         if (pass == XLOG_RECOVER_PASS1) {
2661                 return 0;
2662         }
2663
2664         efi_formatp = (xfs_efi_log_format_t *)item->ri_buf[0].i_addr;
2665
2666         mp = log->l_mp;
2667         efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2668         if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2669                                          &(efip->efi_format)))) {
2670                 xfs_efi_item_free(efip);
2671                 return error;
2672         }
2673         efip->efi_next_extent = efi_formatp->efi_nextents;
2674         efip->efi_flags |= XFS_EFI_COMMITTED;
2675
2676         AIL_LOCK(mp,s);
2677         /*
2678          * xfs_trans_update_ail() drops the AIL lock.
2679          */
2680         xfs_trans_update_ail(mp, (xfs_log_item_t *)efip, lsn, s);
2681         return 0;
2682 }
2683
2684
2685 /*
2686  * This routine is called when an efd format structure is found in
2687  * a committed transaction in the log.  It's purpose is to cancel
2688  * the corresponding efi if it was still in the log.  To do this
2689  * it searches the AIL for the efi with an id equal to that in the
2690  * efd format structure.  If we find it, we remove the efi from the
2691  * AIL and free it.
2692  */
2693 STATIC void
2694 xlog_recover_do_efd_trans(
2695         xlog_t                  *log,
2696         xlog_recover_item_t     *item,
2697         int                     pass)
2698 {
2699         xfs_mount_t             *mp;
2700         xfs_efd_log_format_t    *efd_formatp;
2701         xfs_efi_log_item_t      *efip = NULL;
2702         xfs_log_item_t          *lip;
2703         int                     gen;
2704         __uint64_t              efi_id;
2705         SPLDECL(s);
2706
2707         if (pass == XLOG_RECOVER_PASS1) {
2708                 return;
2709         }
2710
2711         efd_formatp = (xfs_efd_log_format_t *)item->ri_buf[0].i_addr;
2712         ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2713                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2714                (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2715                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2716         efi_id = efd_formatp->efd_efi_id;
2717
2718         /*
2719          * Search for the efi with the id in the efd format structure
2720          * in the AIL.
2721          */
2722         mp = log->l_mp;
2723         AIL_LOCK(mp,s);
2724         lip = xfs_trans_first_ail(mp, &gen);
2725         while (lip != NULL) {
2726                 if (lip->li_type == XFS_LI_EFI) {
2727                         efip = (xfs_efi_log_item_t *)lip;
2728                         if (efip->efi_format.efi_id == efi_id) {
2729                                 /*
2730                                  * xfs_trans_delete_ail() drops the
2731                                  * AIL lock.
2732                                  */
2733                                 xfs_trans_delete_ail(mp, lip, s);
2734                                 break;
2735                         }
2736                 }
2737                 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
2738         }
2739
2740         /*
2741          * If we found it, then free it up.  If it wasn't there, it
2742          * must have been overwritten in the log.  Oh well.
2743          */
2744         if (lip != NULL) {
2745                 xfs_efi_item_free(efip);
2746         } else {
2747                 AIL_UNLOCK(mp, s);
2748         }
2749 }
2750
2751 /*
2752  * Perform the transaction
2753  *
2754  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2755  * EFIs and EFDs get queued up by adding entries into the AIL for them.
2756  */
2757 STATIC int
2758 xlog_recover_do_trans(
2759         xlog_t                  *log,
2760         xlog_recover_t          *trans,
2761         int                     pass)
2762 {
2763         int                     error = 0;
2764         xlog_recover_item_t     *item, *first_item;
2765
2766         if ((error = xlog_recover_reorder_trans(trans)))
2767                 return error;
2768         first_item = item = trans->r_itemq;
2769         do {
2770                 /*
2771                  * we don't need to worry about the block number being
2772                  * truncated in > 1 TB buffers because in user-land,
2773                  * we're now n32 or 64-bit so xfs_daddr_t is 64-bits so
2774                  * the blknos will get through the user-mode buffer
2775                  * cache properly.  The only bad case is o32 kernels
2776                  * where xfs_daddr_t is 32-bits but mount will warn us
2777                  * off a > 1 TB filesystem before we get here.
2778                  */
2779                 if ((ITEM_TYPE(item) == XFS_LI_BUF)) {
2780                         if  ((error = xlog_recover_do_buffer_trans(log, item,
2781                                                                  pass)))
2782                                 break;
2783                 } else if ((ITEM_TYPE(item) == XFS_LI_INODE)) {
2784                         if ((error = xlog_recover_do_inode_trans(log, item,
2785                                                                 pass)))
2786                                 break;
2787                 } else if (ITEM_TYPE(item) == XFS_LI_EFI) {
2788                         if ((error = xlog_recover_do_efi_trans(log, item, trans->r_lsn,
2789                                                   pass)))
2790                                 break;
2791                 } else if (ITEM_TYPE(item) == XFS_LI_EFD) {
2792                         xlog_recover_do_efd_trans(log, item, pass);
2793                 } else if (ITEM_TYPE(item) == XFS_LI_DQUOT) {
2794                         if ((error = xlog_recover_do_dquot_trans(log, item,
2795                                                                    pass)))
2796                                         break;
2797                 } else if ((ITEM_TYPE(item) == XFS_LI_QUOTAOFF)) {
2798                         if ((error = xlog_recover_do_quotaoff_trans(log, item,
2799                                                                    pass)))
2800                                         break;
2801                 } else {
2802                         xlog_warn("XFS: xlog_recover_do_trans");
2803                         ASSERT(0);
2804                         error = XFS_ERROR(EIO);
2805                         break;
2806                 }
2807                 item = item->ri_next;
2808         } while (first_item != item);
2809
2810         return error;
2811 }
2812
2813 /*
2814  * Free up any resources allocated by the transaction
2815  *
2816  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2817  */
2818 STATIC void
2819 xlog_recover_free_trans(
2820         xlog_recover_t          *trans)
2821 {
2822         xlog_recover_item_t     *first_item, *item, *free_item;
2823         int                     i;
2824
2825         item = first_item = trans->r_itemq;
2826         do {
2827                 free_item = item;
2828                 item = item->ri_next;
2829                  /* Free the regions in the item. */
2830                 for (i = 0; i < free_item->ri_cnt; i++) {
2831                         kmem_free(free_item->ri_buf[i].i_addr,
2832                                   free_item->ri_buf[i].i_len);
2833                 }
2834                 /* Free the item itself */
2835                 kmem_free(free_item->ri_buf,
2836                           (free_item->ri_total * sizeof(xfs_log_iovec_t)));
2837                 kmem_free(free_item, sizeof(xlog_recover_item_t));
2838         } while (first_item != item);
2839         /* Free the transaction recover structure */
2840         kmem_free(trans, sizeof(xlog_recover_t));
2841 }
2842
2843 STATIC int
2844 xlog_recover_commit_trans(
2845         xlog_t                  *log,
2846         xlog_recover_t          **q,
2847         xlog_recover_t          *trans,
2848         int                     pass)
2849 {
2850         int                     error;
2851
2852         if ((error = xlog_recover_unlink_tid(q, trans)))
2853                 return error;
2854         if ((error = xlog_recover_do_trans(log, trans, pass)))
2855                 return error;
2856         xlog_recover_free_trans(trans);                 /* no error */
2857         return 0;
2858 }
2859
2860 STATIC int
2861 xlog_recover_unmount_trans(
2862         xlog_recover_t          *trans)
2863 {
2864         /* Do nothing now */
2865         xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2866         return 0;
2867 }
2868
2869 /*
2870  * There are two valid states of the r_state field.  0 indicates that the
2871  * transaction structure is in a normal state.  We have either seen the
2872  * start of the transaction or the last operation we added was not a partial
2873  * operation.  If the last operation we added to the transaction was a
2874  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2875  *
2876  * NOTE: skip LRs with 0 data length.
2877  */
2878 STATIC int
2879 xlog_recover_process_data(
2880         xlog_t                  *log,
2881         xlog_recover_t          *rhash[],
2882         xlog_rec_header_t       *rhead,
2883         xfs_caddr_t             dp,
2884         int                     pass)
2885 {
2886         xfs_caddr_t             lp;
2887         int                     num_logops;
2888         xlog_op_header_t        *ohead;
2889         xlog_recover_t          *trans;
2890         xlog_tid_t              tid;
2891         int                     error;
2892         unsigned long           hash;
2893         uint                    flags;
2894
2895         lp = dp + INT_GET(rhead->h_len, ARCH_CONVERT);
2896         num_logops = INT_GET(rhead->h_num_logops, ARCH_CONVERT);
2897
2898         /* check the log format matches our own - else we can't recover */
2899         if (xlog_header_check_recover(log->l_mp, rhead))
2900                 return (XFS_ERROR(EIO));
2901
2902         while ((dp < lp) && num_logops) {
2903                 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2904                 ohead = (xlog_op_header_t *)dp;
2905                 dp += sizeof(xlog_op_header_t);
2906                 if (ohead->oh_clientid != XFS_TRANSACTION &&
2907                     ohead->oh_clientid != XFS_LOG) {
2908                         xlog_warn(
2909                 "XFS: xlog_recover_process_data: bad clientid");
2910                         ASSERT(0);
2911                         return (XFS_ERROR(EIO));
2912                 }
2913                 tid = INT_GET(ohead->oh_tid, ARCH_CONVERT);
2914                 hash = XLOG_RHASH(tid);
2915                 trans = xlog_recover_find_tid(rhash[hash], tid);
2916                 if (trans == NULL) {               /* not found; add new tid */
2917                         if (ohead->oh_flags & XLOG_START_TRANS)
2918                                 xlog_recover_new_tid(&rhash[hash], tid,
2919                                         INT_GET(rhead->h_lsn, ARCH_CONVERT));
2920                 } else {
2921                         ASSERT(dp+INT_GET(ohead->oh_len, ARCH_CONVERT) <= lp);
2922                         flags = ohead->oh_flags & ~XLOG_END_TRANS;
2923                         if (flags & XLOG_WAS_CONT_TRANS)
2924                                 flags &= ~XLOG_CONTINUE_TRANS;
2925                         switch (flags) {
2926                         case XLOG_COMMIT_TRANS:
2927                                 error = xlog_recover_commit_trans(log,
2928                                                 &rhash[hash], trans, pass);
2929                                 break;
2930                         case XLOG_UNMOUNT_TRANS:
2931                                 error = xlog_recover_unmount_trans(trans);
2932                                 break;
2933                         case XLOG_WAS_CONT_TRANS:
2934                                 error = xlog_recover_add_to_cont_trans(trans,
2935                                                 dp, INT_GET(ohead->oh_len,
2936                                                         ARCH_CONVERT));
2937                                 break;
2938                         case XLOG_START_TRANS:
2939                                 xlog_warn(
2940                         "XFS: xlog_recover_process_data: bad transaction");
2941                                 ASSERT(0);
2942                                 error = XFS_ERROR(EIO);
2943                                 break;
2944                         case 0:
2945                         case XLOG_CONTINUE_TRANS:
2946                                 error = xlog_recover_add_to_trans(trans,
2947                                                 dp, INT_GET(ohead->oh_len,
2948                                                         ARCH_CONVERT));
2949                                 break;
2950                         default:
2951                                 xlog_warn(
2952                         "XFS: xlog_recover_process_data: bad flag");
2953                                 ASSERT(0);
2954                                 error = XFS_ERROR(EIO);
2955                                 break;
2956                         }
2957                         if (error)
2958                                 return error;
2959                 }
2960                 dp += INT_GET(ohead->oh_len, ARCH_CONVERT);
2961                 num_logops--;
2962         }
2963         return 0;
2964 }
2965
2966 /*
2967  * Process an extent free intent item that was recovered from
2968  * the log.  We need to free the extents that it describes.
2969  */
2970 STATIC void
2971 xlog_recover_process_efi(
2972         xfs_mount_t             *mp,
2973         xfs_efi_log_item_t      *efip)
2974 {
2975         xfs_efd_log_item_t      *efdp;
2976         xfs_trans_t             *tp;
2977         int                     i;
2978         xfs_extent_t            *extp;
2979         xfs_fsblock_t           startblock_fsb;
2980
2981         ASSERT(!(efip->efi_flags & XFS_EFI_RECOVERED));
2982
2983         /*
2984          * First check the validity of the extents described by the
2985          * EFI.  If any are bad, then assume that all are bad and
2986          * just toss the EFI.
2987          */
2988         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2989                 extp = &(efip->efi_format.efi_extents[i]);
2990                 startblock_fsb = XFS_BB_TO_FSB(mp,
2991                                    XFS_FSB_TO_DADDR(mp, extp->ext_start));
2992                 if ((startblock_fsb == 0) ||
2993                     (extp->ext_len == 0) ||
2994                     (startblock_fsb >= mp->m_sb.sb_dblocks) ||
2995                     (extp->ext_len >= mp->m_sb.sb_agblocks)) {
2996                         /*
2997                          * This will pull the EFI from the AIL and
2998                          * free the memory associated with it.
2999                          */
3000                         xfs_efi_release(efip, efip->efi_format.efi_nextents);
3001                         return;
3002                 }
3003         }
3004
3005         tp = xfs_trans_alloc(mp, 0);
3006         xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
3007         efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3008
3009         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3010                 extp = &(efip->efi_format.efi_extents[i]);
3011                 xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3012                 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3013                                          extp->ext_len);
3014         }
3015
3016         efip->efi_flags |= XFS_EFI_RECOVERED;
3017         xfs_trans_commit(tp, 0);
3018 }
3019
3020 /*
3021  * Verify that once we've encountered something other than an EFI
3022  * in the AIL that there are no more EFIs in the AIL.
3023  */
3024 #if defined(DEBUG)
3025 STATIC void
3026 xlog_recover_check_ail(
3027         xfs_mount_t             *mp,
3028         xfs_log_item_t          *lip,
3029         int                     gen)
3030 {
3031         int                     orig_gen = gen;
3032
3033         do {
3034                 ASSERT(lip->li_type != XFS_LI_EFI);
3035                 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3036                 /*
3037                  * The check will be bogus if we restart from the
3038                  * beginning of the AIL, so ASSERT that we don't.
3039                  * We never should since we're holding the AIL lock
3040                  * the entire time.
3041                  */
3042                 ASSERT(gen == orig_gen);
3043         } while (lip != NULL);
3044 }
3045 #endif  /* DEBUG */
3046
3047 /*
3048  * When this is called, all of the EFIs which did not have
3049  * corresponding EFDs should be in the AIL.  What we do now
3050  * is free the extents associated with each one.
3051  *
3052  * Since we process the EFIs in normal transactions, they
3053  * will be removed at some point after the commit.  This prevents
3054  * us from just walking down the list processing each one.
3055  * We'll use a flag in the EFI to skip those that we've already
3056  * processed and use the AIL iteration mechanism's generation
3057  * count to try to speed this up at least a bit.
3058  *
3059  * When we start, we know that the EFIs are the only things in
3060  * the AIL.  As we process them, however, other items are added
3061  * to the AIL.  Since everything added to the AIL must come after
3062  * everything already in the AIL, we stop processing as soon as
3063  * we see something other than an EFI in the AIL.
3064  */
3065 STATIC void
3066 xlog_recover_process_efis(
3067         xlog_t                  *log)
3068 {
3069         xfs_log_item_t          *lip;
3070         xfs_efi_log_item_t      *efip;
3071         int                     gen;
3072         xfs_mount_t             *mp;
3073         SPLDECL(s);
3074
3075         mp = log->l_mp;
3076         AIL_LOCK(mp,s);
3077
3078         lip = xfs_trans_first_ail(mp, &gen);
3079         while (lip != NULL) {
3080                 /*
3081                  * We're done when we see something other than an EFI.
3082                  */
3083                 if (lip->li_type != XFS_LI_EFI) {
3084                         xlog_recover_check_ail(mp, lip, gen);
3085                         break;
3086                 }
3087
3088                 /*
3089                  * Skip EFIs that we've already processed.
3090                  */
3091                 efip = (xfs_efi_log_item_t *)lip;
3092                 if (efip->efi_flags & XFS_EFI_RECOVERED) {
3093                         lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3094                         continue;
3095                 }
3096
3097                 AIL_UNLOCK(mp, s);
3098                 xlog_recover_process_efi(mp, efip);
3099                 AIL_LOCK(mp,s);
3100                 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3101         }
3102         AIL_UNLOCK(mp, s);
3103 }
3104
3105 /*
3106  * This routine performs a transaction to null out a bad inode pointer
3107  * in an agi unlinked inode hash bucket.
3108  */
3109 STATIC void
3110 xlog_recover_clear_agi_bucket(
3111         xfs_mount_t     *mp,
3112         xfs_agnumber_t  agno,
3113         int             bucket)
3114 {
3115         xfs_trans_t     *tp;
3116         xfs_agi_t       *agi;
3117         xfs_buf_t       *agibp;
3118         int             offset;
3119         int             error;
3120
3121         tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3122         xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp), 0, 0, 0);
3123
3124         error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
3125                                    XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3126                                    XFS_FSS_TO_BB(mp, 1), 0, &agibp);
3127         if (error) {
3128                 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3129                 return;
3130         }
3131
3132         agi = XFS_BUF_TO_AGI(agibp);
3133         if (be32_to_cpu(agi->agi_magicnum) != XFS_AGI_MAGIC) {
3134                 xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3135                 return;
3136         }
3137
3138         agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3139         offset = offsetof(xfs_agi_t, agi_unlinked) +
3140                  (sizeof(xfs_agino_t) * bucket);
3141         xfs_trans_log_buf(tp, agibp, offset,
3142                           (offset + sizeof(xfs_agino_t) - 1));
3143
3144         (void) xfs_trans_commit(tp, 0);
3145 }
3146
3147 /*
3148  * xlog_iunlink_recover
3149  *
3150  * This is called during recovery to process any inodes which
3151  * we unlinked but not freed when the system crashed.  These
3152  * inodes will be on the lists in the AGI blocks.  What we do
3153  * here is scan all the AGIs and fully truncate and free any
3154  * inodes found on the lists.  Each inode is removed from the
3155  * lists when it has been fully truncated and is freed.  The
3156  * freeing of the inode and its removal from the list must be
3157  * atomic.
3158  */
3159 void
3160 xlog_recover_process_iunlinks(
3161         xlog_t          *log)
3162 {
3163         xfs_mount_t     *mp;
3164         xfs_agnumber_t  agno;
3165         xfs_agi_t       *agi;
3166         xfs_buf_t       *agibp;
3167         xfs_buf_t       *ibp;
3168         xfs_dinode_t    *dip;
3169         xfs_inode_t     *ip;
3170         xfs_agino_t     agino;
3171         xfs_ino_t       ino;
3172         int             bucket;
3173         int             error;
3174         uint            mp_dmevmask;
3175
3176         mp = log->l_mp;
3177
3178         /*
3179          * Prevent any DMAPI event from being sent while in this function.
3180          */
3181         mp_dmevmask = mp->m_dmevmask;
3182         mp->m_dmevmask = 0;
3183
3184         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3185                 /*
3186                  * Find the agi for this ag.
3187                  */
3188                 agibp = xfs_buf_read(mp->m_ddev_targp,
3189                                 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3190                                 XFS_FSS_TO_BB(mp, 1), 0);
3191                 if (XFS_BUF_ISERROR(agibp)) {
3192                         xfs_ioerror_alert("xlog_recover_process_iunlinks(#1)",
3193                                 log->l_mp, agibp,
3194                                 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)));
3195                 }
3196                 agi = XFS_BUF_TO_AGI(agibp);
3197                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agi->agi_magicnum));
3198
3199                 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3200
3201                         agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3202                         while (agino != NULLAGINO) {
3203
3204                                 /*
3205                                  * Release the agi buffer so that it can
3206                                  * be acquired in the normal course of the
3207                                  * transaction to truncate and free the inode.
3208                                  */
3209                                 xfs_buf_relse(agibp);
3210
3211                                 ino = XFS_AGINO_TO_INO(mp, agno, agino);
3212                                 error = xfs_iget(mp, NULL, ino, 0, 0, &ip, 0);
3213                                 ASSERT(error || (ip != NULL));
3214
3215                                 if (!error) {
3216                                         /*
3217                                          * Get the on disk inode to find the
3218                                          * next inode in the bucket.
3219                                          */
3220                                         error = xfs_itobp(mp, NULL, ip, &dip,
3221                                                         &ibp, 0, 0);
3222                                         ASSERT(error || (dip != NULL));
3223                                 }
3224
3225                                 if (!error) {
3226                                         ASSERT(ip->i_d.di_nlink == 0);
3227
3228                                         /* setup for the next pass */
3229                                         agino = INT_GET(dip->di_next_unlinked,
3230                                                         ARCH_CONVERT);
3231                                         xfs_buf_relse(ibp);
3232                                         /*
3233                                          * Prevent any DMAPI event from
3234                                          * being sent when the
3235                                          * reference on the inode is
3236                                          * dropped.
3237                                          */
3238                                         ip->i_d.di_dmevmask = 0;
3239
3240                                         /*
3241                                          * If this is a new inode, handle
3242                                          * it specially.  Otherwise,
3243                                          * just drop our reference to the
3244                                          * inode.  If there are no
3245                                          * other references, this will
3246                                          * send the inode to
3247                                          * xfs_inactive() which will
3248                                          * truncate the file and free
3249                                          * the inode.
3250                                          */
3251                                         if (ip->i_d.di_mode == 0)
3252                                                 xfs_iput_new(ip, 0);
3253                                         else
3254                                                 VN_RELE(XFS_ITOV(ip));
3255                                 } else {
3256                                         /*
3257                                          * We can't read in the inode
3258                                          * this bucket points to, or
3259                                          * this inode is messed up.  Just
3260                                          * ditch this bucket of inodes.  We
3261                                          * will lose some inodes and space,
3262                                          * but at least we won't hang.  Call
3263                                          * xlog_recover_clear_agi_bucket()
3264                                          * to perform a transaction to clear
3265                                          * the inode pointer in the bucket.
3266                                          */
3267                                         xlog_recover_clear_agi_bucket(mp, agno,
3268                                                         bucket);
3269
3270                                         agino = NULLAGINO;
3271                                 }
3272
3273                                 /*
3274                                  * Reacquire the agibuffer and continue around
3275                                  * the loop.
3276                                  */
3277                                 agibp = xfs_buf_read(mp->m_ddev_targp,
3278                                                 XFS_AG_DADDR(mp, agno,
3279                                                         XFS_AGI_DADDR(mp)),
3280                                                 XFS_FSS_TO_BB(mp, 1), 0);
3281                                 if (XFS_BUF_ISERROR(agibp)) {
3282                                         xfs_ioerror_alert(
3283                                 "xlog_recover_process_iunlinks(#2)",
3284                                                 log->l_mp, agibp,
3285                                                 XFS_AG_DADDR(mp, agno,
3286                                                         XFS_AGI_DADDR(mp)));
3287                                 }
3288                                 agi = XFS_BUF_TO_AGI(agibp);
3289                                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(
3290                                         agi->agi_magicnum));
3291                         }
3292                 }
3293
3294                 /*
3295                  * Release the buffer for the current agi so we can
3296                  * go on to the next one.
3297                  */
3298                 xfs_buf_relse(agibp);
3299         }
3300
3301         mp->m_dmevmask = mp_dmevmask;
3302 }
3303
3304
3305 #ifdef DEBUG
3306 STATIC void
3307 xlog_pack_data_checksum(
3308         xlog_t          *log,
3309         xlog_in_core_t  *iclog,
3310         int             size)
3311 {
3312         int             i;
3313         uint            *up;
3314         uint            chksum = 0;
3315
3316         up = (uint *)iclog->ic_datap;
3317         /* divide length by 4 to get # words */
3318         for (i = 0; i < (size >> 2); i++) {
3319                 chksum ^= INT_GET(*up, ARCH_CONVERT);
3320                 up++;
3321         }
3322         INT_SET(iclog->ic_header.h_chksum, ARCH_CONVERT, chksum);
3323 }
3324 #else
3325 #define xlog_pack_data_checksum(log, iclog, size)
3326 #endif
3327
3328 /*
3329  * Stamp cycle number in every block
3330  */
3331 void
3332 xlog_pack_data(
3333         xlog_t                  *log,
3334         xlog_in_core_t          *iclog,
3335         int                     roundoff)
3336 {
3337         int                     i, j, k;
3338         int                     size = iclog->ic_offset + roundoff;
3339         uint                    cycle_lsn;
3340         xfs_caddr_t             dp;
3341         xlog_in_core_2_t        *xhdr;
3342
3343         xlog_pack_data_checksum(log, iclog, size);
3344
3345         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3346
3347         dp = iclog->ic_datap;
3348         for (i = 0; i < BTOBB(size) &&
3349                 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3350                 iclog->ic_header.h_cycle_data[i] = *(uint *)dp;
3351                 *(uint *)dp = cycle_lsn;
3352                 dp += BBSIZE;
3353         }
3354
3355         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3356                 xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3357                 for ( ; i < BTOBB(size); i++) {
3358                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3359                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3360                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(uint *)dp;
3361                         *(uint *)dp = cycle_lsn;
3362                         dp += BBSIZE;
3363                 }
3364
3365                 for (i = 1; i < log->l_iclog_heads; i++) {
3366                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3367                 }
3368         }
3369 }
3370
3371 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
3372 STATIC void
3373 xlog_unpack_data_checksum(
3374         xlog_rec_header_t       *rhead,
3375         xfs_caddr_t             dp,
3376         xlog_t                  *log)
3377 {
3378         uint                    *up = (uint *)dp;
3379         uint                    chksum = 0;
3380         int                     i;
3381
3382         /* divide length by 4 to get # words */
3383         for (i=0; i < INT_GET(rhead->h_len, ARCH_CONVERT) >> 2; i++) {
3384                 chksum ^= INT_GET(*up, ARCH_CONVERT);
3385                 up++;
3386         }
3387         if (chksum != INT_GET(rhead->h_chksum, ARCH_CONVERT)) {
3388             if (rhead->h_chksum ||
3389                 ((log->l_flags & XLOG_CHKSUM_MISMATCH) == 0)) {
3390                     cmn_err(CE_DEBUG,
3391                         "XFS: LogR chksum mismatch: was (0x%x) is (0x%x)\n",
3392                             INT_GET(rhead->h_chksum, ARCH_CONVERT), chksum);
3393                     cmn_err(CE_DEBUG,
3394 "XFS: Disregard message if filesystem was created with non-DEBUG kernel");
3395                     if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3396                             cmn_err(CE_DEBUG,
3397                                 "XFS: LogR this is a LogV2 filesystem\n");
3398                     }
3399                     log->l_flags |= XLOG_CHKSUM_MISMATCH;
3400             }
3401         }
3402 }
3403 #else
3404 #define xlog_unpack_data_checksum(rhead, dp, log)
3405 #endif
3406
3407 STATIC void
3408 xlog_unpack_data(
3409         xlog_rec_header_t       *rhead,
3410         xfs_caddr_t             dp,
3411         xlog_t                  *log)
3412 {
3413         int                     i, j, k;
3414         xlog_in_core_2_t        *xhdr;
3415
3416         for (i = 0; i < BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT)) &&
3417                   i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3418                 *(uint *)dp = *(uint *)&rhead->h_cycle_data[i];
3419                 dp += BBSIZE;
3420         }
3421
3422         if (XFS_SB_VERSION_HASLOGV2(&log->l_mp->m_sb)) {
3423                 xhdr = (xlog_in_core_2_t *)rhead;
3424                 for ( ; i < BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT)); i++) {
3425                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3426                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3427                         *(uint *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3428                         dp += BBSIZE;
3429                 }
3430         }
3431
3432         xlog_unpack_data_checksum(rhead, dp, log);
3433 }
3434
3435 STATIC int
3436 xlog_valid_rec_header(
3437         xlog_t                  *log,
3438         xlog_rec_header_t       *rhead,
3439         xfs_daddr_t             blkno)
3440 {
3441         int                     hlen;
3442
3443         if (unlikely(
3444             (INT_GET(rhead->h_magicno, ARCH_CONVERT) !=
3445                         XLOG_HEADER_MAGIC_NUM))) {
3446                 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3447                                 XFS_ERRLEVEL_LOW, log->l_mp);
3448                 return XFS_ERROR(EFSCORRUPTED);
3449         }
3450         if (unlikely(
3451             (!rhead->h_version ||
3452             (INT_GET(rhead->h_version, ARCH_CONVERT) &
3453                         (~XLOG_VERSION_OKBITS)) != 0))) {
3454                 xlog_warn("XFS: %s: unrecognised log version (%d).",
3455                         __FUNCTION__, INT_GET(rhead->h_version, ARCH_CONVERT));
3456                 return XFS_ERROR(EIO);
3457         }
3458
3459         /* LR body must have data or it wouldn't have been written */
3460         hlen = INT_GET(rhead->h_len, ARCH_CONVERT);
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             bufaddr, 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         xlog_recover_t          *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                 if ((error = xlog_bread(log, tail_blk, 1, hbp)))
3514                         goto bread_err1;
3515                 offset = xlog_align(log, tail_blk, 1, hbp);
3516                 rhead = (xlog_rec_header_t *)offset;
3517                 error = xlog_valid_rec_header(log, rhead, tail_blk);
3518                 if (error)
3519                         goto bread_err1;
3520                 h_size = INT_GET(rhead->h_size, ARCH_CONVERT);
3521                 if ((INT_GET(rhead->h_version, ARCH_CONVERT)
3522                                 & XLOG_VERSION_2) &&
3523                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3524                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3525                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
3526                                 hblks++;
3527                         xlog_put_bp(hbp);
3528                         hbp = xlog_get_bp(log, hblks);
3529                 } else {
3530                         hblks = 1;
3531                 }
3532         } else {
3533                 ASSERT(log->l_sectbb_log == 0);
3534                 hblks = 1;
3535                 hbp = xlog_get_bp(log, 1);
3536                 h_size = XLOG_BIG_RECORD_BSIZE;
3537         }
3538
3539         if (!hbp)
3540                 return ENOMEM;
3541         dbp = xlog_get_bp(log, BTOBB(h_size));
3542         if (!dbp) {
3543                 xlog_put_bp(hbp);
3544                 return ENOMEM;
3545         }
3546
3547         memset(rhash, 0, sizeof(rhash));
3548         if (tail_blk <= head_blk) {
3549                 for (blk_no = tail_blk; blk_no < head_blk; ) {
3550                         if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3551                                 goto bread_err2;
3552                         offset = xlog_align(log, blk_no, hblks, hbp);
3553                         rhead = (xlog_rec_header_t *)offset;
3554                         error = xlog_valid_rec_header(log, rhead, blk_no);
3555                         if (error)
3556                                 goto bread_err2;
3557
3558                         /* blocks in data section */
3559                         bblks = (int)BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT));
3560                         error = xlog_bread(log, blk_no + hblks, bblks, dbp);
3561                         if (error)
3562                                 goto bread_err2;
3563                         offset = xlog_align(log, blk_no + hblks, bblks, dbp);
3564                         xlog_unpack_data(rhead, offset, log);
3565                         if ((error = xlog_recover_process_data(log,
3566                                                 rhash, rhead, offset, pass)))
3567                                 goto bread_err2;
3568                         blk_no += bblks + hblks;
3569                 }
3570         } else {
3571                 /*
3572                  * Perform recovery around the end of the physical log.
3573                  * When the head is not on the same cycle number as the tail,
3574                  * we can't do a sequential recovery as above.
3575                  */
3576                 blk_no = tail_blk;
3577                 while (blk_no < log->l_logBBsize) {
3578                         /*
3579                          * Check for header wrapping around physical end-of-log
3580                          */
3581                         offset = NULL;
3582                         split_hblks = 0;
3583                         wrapped_hblks = 0;
3584                         if (blk_no + hblks <= log->l_logBBsize) {
3585                                 /* Read header in one read */
3586                                 error = xlog_bread(log, blk_no, hblks, hbp);
3587                                 if (error)
3588                                         goto bread_err2;
3589                                 offset = xlog_align(log, blk_no, hblks, hbp);
3590                         } else {
3591                                 /* This LR is split across physical log end */
3592                                 if (blk_no != log->l_logBBsize) {
3593                                         /* some data before physical log end */
3594                                         ASSERT(blk_no <= INT_MAX);
3595                                         split_hblks = log->l_logBBsize - (int)blk_no;
3596                                         ASSERT(split_hblks > 0);
3597                                         if ((error = xlog_bread(log, blk_no,
3598                                                         split_hblks, hbp)))
3599                                                 goto bread_err2;
3600                                         offset = xlog_align(log, blk_no,
3601                                                         split_hblks, hbp);
3602                                 }
3603                                 /*
3604                                  * Note: this black magic still works with
3605                                  * large sector sizes (non-512) only because:
3606                                  * - we increased the buffer size originally
3607                                  *   by 1 sector giving us enough extra space
3608                                  *   for the second read;
3609                                  * - the log start is guaranteed to be sector
3610                                  *   aligned;
3611                                  * - we read the log end (LR header start)
3612                                  *   _first_, then the log start (LR header end)
3613                                  *   - order is important.
3614                                  */
3615                                 bufaddr = XFS_BUF_PTR(hbp);
3616                                 XFS_BUF_SET_PTR(hbp,
3617                                                 bufaddr + BBTOB(split_hblks),
3618                                                 BBTOB(hblks - split_hblks));
3619                                 wrapped_hblks = hblks - split_hblks;
3620                                 error = xlog_bread(log, 0, wrapped_hblks, hbp);
3621                                 if (error)
3622                                         goto bread_err2;
3623                                 XFS_BUF_SET_PTR(hbp, bufaddr, BBTOB(hblks));
3624                                 if (!offset)
3625                                         offset = xlog_align(log, 0,
3626                                                         wrapped_hblks, hbp);
3627                         }
3628                         rhead = (xlog_rec_header_t *)offset;
3629                         error = xlog_valid_rec_header(log, rhead,
3630                                                 split_hblks ? blk_no : 0);
3631                         if (error)
3632                                 goto bread_err2;
3633
3634                         bblks = (int)BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT));
3635                         blk_no += hblks;
3636
3637                         /* Read in data for log record */
3638                         if (blk_no + bblks <= log->l_logBBsize) {
3639                                 error = xlog_bread(log, blk_no, bblks, dbp);
3640                                 if (error)
3641                                         goto bread_err2;
3642                                 offset = xlog_align(log, blk_no, bblks, dbp);
3643                         } else {
3644                                 /* This log record is split across the
3645                                  * physical end of log */
3646                                 offset = NULL;
3647                                 split_bblks = 0;
3648                                 if (blk_no != log->l_logBBsize) {
3649                                         /* some data is before the physical
3650                                          * end of log */
3651                                         ASSERT(!wrapped_hblks);
3652                                         ASSERT(blk_no <= INT_MAX);
3653                                         split_bblks =
3654                                                 log->l_logBBsize - (int)blk_no;
3655                                         ASSERT(split_bblks > 0);
3656                                         if ((error = xlog_bread(log, blk_no,
3657                                                         split_bblks, dbp)))
3658                                                 goto bread_err2;
3659                                         offset = xlog_align(log, blk_no,
3660                                                         split_bblks, dbp);
3661                                 }
3662                                 /*
3663                                  * Note: this black magic still works with
3664                                  * large sector sizes (non-512) only because:
3665                                  * - we increased the buffer size originally
3666                                  *   by 1 sector giving us enough extra space
3667                                  *   for the second read;
3668                                  * - the log start is guaranteed to be sector
3669                                  *   aligned;
3670                                  * - we read the log end (LR header start)
3671                                  *   _first_, then the log start (LR header end)
3672                                  *   - order is important.
3673                                  */
3674                                 bufaddr = XFS_BUF_PTR(dbp);
3675                                 XFS_BUF_SET_PTR(dbp,
3676                                                 bufaddr + BBTOB(split_bblks),
3677                                                 BBTOB(bblks - split_bblks));
3678                                 if ((error = xlog_bread(log, wrapped_hblks,
3679                                                 bblks - split_bblks, dbp)))
3680                                         goto bread_err2;
3681                                 XFS_BUF_SET_PTR(dbp, bufaddr, h_size);
3682                                 if (!offset)
3683                                         offset = xlog_align(log, wrapped_hblks,
3684                                                 bblks - split_bblks, dbp);
3685                         }
3686                         xlog_unpack_data(rhead, offset, log);
3687                         if ((error = xlog_recover_process_data(log, rhash,
3688                                                         rhead, offset, pass)))
3689                                 goto bread_err2;
3690                         blk_no += bblks;
3691                 }
3692
3693                 ASSERT(blk_no >= log->l_logBBsize);
3694                 blk_no -= log->l_logBBsize;
3695
3696                 /* read first part of physical log */
3697                 while (blk_no < head_blk) {
3698                         if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3699                                 goto bread_err2;
3700                         offset = xlog_align(log, blk_no, hblks, hbp);
3701                         rhead = (xlog_rec_header_t *)offset;
3702                         error = xlog_valid_rec_header(log, rhead, blk_no);
3703                         if (error)
3704                                 goto bread_err2;
3705                         bblks = (int)BTOBB(INT_GET(rhead->h_len, ARCH_CONVERT));
3706                         if ((error = xlog_bread(log, blk_no+hblks, bblks, dbp)))
3707                                 goto bread_err2;
3708                         offset = xlog_align(log, blk_no+hblks, bblks, dbp);
3709                         xlog_unpack_data(rhead, offset, log);
3710                         if ((error = xlog_recover_process_data(log, rhash,
3711                                                         rhead, offset, pass)))
3712                                 goto bread_err2;
3713                         blk_no += bblks + hblks;
3714                 }
3715         }
3716
3717  bread_err2:
3718         xlog_put_bp(dbp);
3719  bread_err1:
3720         xlog_put_bp(hbp);
3721         return error;
3722 }
3723
3724 /*
3725  * Do the recovery of the log.  We actually do this in two phases.
3726  * The two passes are necessary in order to implement the function
3727  * of cancelling a record written into the log.  The first pass
3728  * determines those things which have been cancelled, and the
3729  * second pass replays log items normally except for those which
3730  * have been cancelled.  The handling of the replay and cancellations
3731  * takes place in the log item type specific routines.
3732  *
3733  * The table of items which have cancel records in the log is allocated
3734  * and freed at this level, since only here do we know when all of
3735  * the log recovery has been completed.
3736  */
3737 STATIC int
3738 xlog_do_log_recovery(
3739         xlog_t          *log,
3740         xfs_daddr_t     head_blk,
3741         xfs_daddr_t     tail_blk)
3742 {
3743         int             error;
3744
3745         ASSERT(head_blk != tail_blk);
3746
3747         /*
3748          * First do a pass to find all of the cancelled buf log items.
3749          * Store them in the buf_cancel_table for use in the second pass.
3750          */
3751         log->l_buf_cancel_table =
3752                 (xfs_buf_cancel_t **)kmem_zalloc(XLOG_BC_TABLE_SIZE *
3753                                                  sizeof(xfs_buf_cancel_t*),
3754                                                  KM_SLEEP);
3755         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3756                                       XLOG_RECOVER_PASS1);
3757         if (error != 0) {
3758                 kmem_free(log->l_buf_cancel_table,
3759                           XLOG_BC_TABLE_SIZE * sizeof(xfs_buf_cancel_t*));
3760                 log->l_buf_cancel_table = NULL;
3761                 return error;
3762         }
3763         /*
3764          * Then do a second pass to actually recover the items in the log.
3765          * When it is complete free the table of buf cancel items.
3766          */
3767         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3768                                       XLOG_RECOVER_PASS2);
3769 #ifdef DEBUG
3770         if (!error) {
3771                 int     i;
3772
3773                 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3774                         ASSERT(log->l_buf_cancel_table[i] == NULL);
3775         }
3776 #endif  /* DEBUG */
3777
3778         kmem_free(log->l_buf_cancel_table,
3779                   XLOG_BC_TABLE_SIZE * sizeof(xfs_buf_cancel_t*));
3780         log->l_buf_cancel_table = NULL;
3781
3782         return error;
3783 }
3784
3785 /*
3786  * Do the actual recovery
3787  */
3788 STATIC int
3789 xlog_do_recover(
3790         xlog_t          *log,
3791         xfs_daddr_t     head_blk,
3792         xfs_daddr_t     tail_blk)
3793 {
3794         int             error;
3795         xfs_buf_t       *bp;
3796         xfs_sb_t        *sbp;
3797
3798         /*
3799          * First replay the images in the log.
3800          */
3801         error = xlog_do_log_recovery(log, head_blk, tail_blk);
3802         if (error) {
3803                 return error;
3804         }
3805
3806         XFS_bflush(log->l_mp->m_ddev_targp);
3807
3808         /*
3809          * If IO errors happened during recovery, bail out.
3810          */
3811         if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3812                 return (EIO);
3813         }
3814
3815         /*
3816          * We now update the tail_lsn since much of the recovery has completed
3817          * and there may be space available to use.  If there were no extent
3818          * or iunlinks, we can free up the entire log and set the tail_lsn to
3819          * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3820          * lsn of the last known good LR on disk.  If there are extent frees
3821          * or iunlinks they will have some entries in the AIL; so we look at
3822          * the AIL to determine how to set the tail_lsn.
3823          */
3824         xlog_assign_tail_lsn(log->l_mp);
3825
3826         /*
3827          * Now that we've finished replaying all buffer and inode
3828          * updates, re-read in the superblock.
3829          */
3830         bp = xfs_getsb(log->l_mp, 0);
3831         XFS_BUF_UNDONE(bp);
3832         XFS_BUF_READ(bp);
3833         xfsbdstrat(log->l_mp, bp);
3834         if ((error = xfs_iowait(bp))) {
3835                 xfs_ioerror_alert("xlog_do_recover",
3836                                   log->l_mp, bp, XFS_BUF_ADDR(bp));
3837                 ASSERT(0);
3838                 xfs_buf_relse(bp);
3839                 return error;
3840         }
3841
3842         /* Convert superblock from on-disk format */
3843         sbp = &log->l_mp->m_sb;
3844         xfs_xlatesb(XFS_BUF_TO_SBP(bp), sbp, 1, XFS_SB_ALL_BITS);
3845         ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3846         ASSERT(XFS_SB_GOOD_VERSION(sbp));
3847         xfs_buf_relse(bp);
3848
3849         /* We've re-read the superblock so re-initialize per-cpu counters */
3850         xfs_icsb_reinit_counters(log->l_mp);
3851
3852         xlog_recover_check_summary(log);
3853
3854         /* Normal transactions can now occur */
3855         log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3856         return 0;
3857 }
3858
3859 /*
3860  * Perform recovery and re-initialize some log variables in xlog_find_tail.
3861  *
3862  * Return error or zero.
3863  */
3864 int
3865 xlog_recover(
3866         xlog_t          *log)
3867 {
3868         xfs_daddr_t     head_blk, tail_blk;
3869         int             error;
3870
3871         /* find the tail of the log */
3872         if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3873                 return error;
3874
3875         if (tail_blk != head_blk) {
3876                 /* There used to be a comment here:
3877                  *
3878                  * disallow recovery on read-only mounts.  note -- mount
3879                  * checks for ENOSPC and turns it into an intelligent
3880                  * error message.
3881                  * ...but this is no longer true.  Now, unless you specify
3882                  * NORECOVERY (in which case this function would never be
3883                  * called), we just go ahead and recover.  We do this all
3884                  * under the vfs layer, so we can get away with it unless
3885                  * the device itself is read-only, in which case we fail.
3886                  */
3887                 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3888                         return error;
3889                 }
3890
3891                 cmn_err(CE_NOTE,
3892                         "Starting XFS recovery on filesystem: %s (logdev: %s)",
3893                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3894                         log->l_mp->m_logname : "internal");
3895
3896                 error = xlog_do_recover(log, head_blk, tail_blk);
3897                 log->l_flags |= XLOG_RECOVERY_NEEDED;
3898         }
3899         return error;
3900 }
3901
3902 /*
3903  * In the first part of recovery we replay inodes and buffers and build
3904  * up the list of extent free items which need to be processed.  Here
3905  * we process the extent free items and clean up the on disk unlinked
3906  * inode lists.  This is separated from the first part of recovery so
3907  * that the root and real-time bitmap inodes can be read in from disk in
3908  * between the two stages.  This is necessary so that we can free space
3909  * in the real-time portion of the file system.
3910  */
3911 int
3912 xlog_recover_finish(
3913         xlog_t          *log,
3914         int             mfsi_flags)
3915 {
3916         /*
3917          * Now we're ready to do the transactions needed for the
3918          * rest of recovery.  Start with completing all the extent
3919          * free intent records and then process the unlinked inode
3920          * lists.  At this point, we essentially run in normal mode
3921          * except that we're still performing recovery actions
3922          * rather than accepting new requests.
3923          */
3924         if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3925                 xlog_recover_process_efis(log);
3926                 /*
3927                  * Sync the log to get all the EFIs out of the AIL.
3928                  * This isn't absolutely necessary, but it helps in
3929                  * case the unlink transactions would have problems
3930                  * pushing the EFIs out of the way.
3931                  */
3932                 xfs_log_force(log->l_mp, (xfs_lsn_t)0,
3933                               (XFS_LOG_FORCE | XFS_LOG_SYNC));
3934
3935                 if ( (mfsi_flags & XFS_MFSI_NOUNLINK) == 0 ) {
3936                         xlog_recover_process_iunlinks(log);
3937                 }
3938
3939                 xlog_recover_check_summary(log);
3940
3941                 cmn_err(CE_NOTE,
3942                         "Ending XFS recovery on filesystem: %s (logdev: %s)",
3943                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3944                         log->l_mp->m_logname : "internal");
3945                 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3946         } else {
3947                 cmn_err(CE_DEBUG,
3948                         "!Ending clean XFS mount for filesystem: %s\n",
3949                         log->l_mp->m_fsname);
3950         }
3951         return 0;
3952 }
3953
3954
3955 #if defined(DEBUG)
3956 /*
3957  * Read all of the agf and agi counters and check that they
3958  * are consistent with the superblock counters.
3959  */
3960 void
3961 xlog_recover_check_summary(
3962         xlog_t          *log)
3963 {
3964         xfs_mount_t     *mp;
3965         xfs_agf_t       *agfp;
3966         xfs_agi_t       *agip;
3967         xfs_buf_t       *agfbp;
3968         xfs_buf_t       *agibp;
3969         xfs_daddr_t     agfdaddr;
3970         xfs_daddr_t     agidaddr;
3971         xfs_buf_t       *sbbp;
3972 #ifdef XFS_LOUD_RECOVERY
3973         xfs_sb_t        *sbp;
3974 #endif
3975         xfs_agnumber_t  agno;
3976         __uint64_t      freeblks;
3977         __uint64_t      itotal;
3978         __uint64_t      ifree;
3979
3980         mp = log->l_mp;
3981
3982         freeblks = 0LL;
3983         itotal = 0LL;
3984         ifree = 0LL;
3985         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3986                 agfdaddr = XFS_AG_DADDR(mp, agno, XFS_AGF_DADDR(mp));
3987                 agfbp = xfs_buf_read(mp->m_ddev_targp, agfdaddr,
3988                                 XFS_FSS_TO_BB(mp, 1), 0);
3989                 if (XFS_BUF_ISERROR(agfbp)) {
3990                         xfs_ioerror_alert("xlog_recover_check_summary(agf)",
3991                                                 mp, agfbp, agfdaddr);
3992                 }
3993                 agfp = XFS_BUF_TO_AGF(agfbp);
3994                 ASSERT(XFS_AGF_MAGIC == be32_to_cpu(agfp->agf_magicnum));
3995                 ASSERT(XFS_AGF_GOOD_VERSION(be32_to_cpu(agfp->agf_versionnum)));
3996                 ASSERT(be32_to_cpu(agfp->agf_seqno) == agno);
3997
3998                 freeblks += be32_to_cpu(agfp->agf_freeblks) +
3999                             be32_to_cpu(agfp->agf_flcount);
4000                 xfs_buf_relse(agfbp);
4001
4002                 agidaddr = XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp));
4003                 agibp = xfs_buf_read(mp->m_ddev_targp, agidaddr,
4004                                 XFS_FSS_TO_BB(mp, 1), 0);
4005                 if (XFS_BUF_ISERROR(agibp)) {
4006                         xfs_ioerror_alert("xlog_recover_check_summary(agi)",
4007                                           mp, agibp, agidaddr);
4008                 }
4009                 agip = XFS_BUF_TO_AGI(agibp);
4010                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agip->agi_magicnum));
4011                 ASSERT(XFS_AGI_GOOD_VERSION(be32_to_cpu(agip->agi_versionnum)));
4012                 ASSERT(be32_to_cpu(agip->agi_seqno) == agno);
4013
4014                 itotal += be32_to_cpu(agip->agi_count);
4015                 ifree += be32_to_cpu(agip->agi_freecount);
4016                 xfs_buf_relse(agibp);
4017         }
4018
4019         sbbp = xfs_getsb(mp, 0);
4020 #ifdef XFS_LOUD_RECOVERY
4021         sbp = &mp->m_sb;
4022         xfs_xlatesb(XFS_BUF_TO_SBP(sbbp), sbp, 1, XFS_SB_ALL_BITS);
4023         cmn_err(CE_NOTE,
4024                 "xlog_recover_check_summary: sb_icount %Lu itotal %Lu",
4025                 sbp->sb_icount, itotal);
4026         cmn_err(CE_NOTE,
4027                 "xlog_recover_check_summary: sb_ifree %Lu itotal %Lu",
4028                 sbp->sb_ifree, ifree);
4029         cmn_err(CE_NOTE,
4030                 "xlog_recover_check_summary: sb_fdblocks %Lu freeblks %Lu",
4031                 sbp->sb_fdblocks, freeblks);
4032 #if 0
4033         /*
4034          * This is turned off until I account for the allocation
4035          * btree blocks which live in free space.
4036          */
4037         ASSERT(sbp->sb_icount == itotal);
4038         ASSERT(sbp->sb_ifree == ifree);
4039         ASSERT(sbp->sb_fdblocks == freeblks);
4040 #endif
4041 #endif
4042         xfs_buf_relse(sbbp);
4043 }
4044 #endif /* DEBUG */