~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/xfs/xfs_log_recover.c

Version: ~ [ linux-5.13-rc7 ] ~ [ linux-5.12.12 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.45 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.127 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.195 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.237 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.273 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.273 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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_shared.h"
 21 #include "xfs_format.h"
 22 #include "xfs_log_format.h"
 23 #include "xfs_trans_resv.h"
 24 #include "xfs_bit.h"
 25 #include "xfs_sb.h"
 26 #include "xfs_mount.h"
 27 #include "xfs_defer.h"
 28 #include "xfs_da_format.h"
 29 #include "xfs_da_btree.h"
 30 #include "xfs_inode.h"
 31 #include "xfs_trans.h"
 32 #include "xfs_log.h"
 33 #include "xfs_log_priv.h"
 34 #include "xfs_log_recover.h"
 35 #include "xfs_inode_item.h"
 36 #include "xfs_extfree_item.h"
 37 #include "xfs_trans_priv.h"
 38 #include "xfs_alloc.h"
 39 #include "xfs_ialloc.h"
 40 #include "xfs_quota.h"
 41 #include "xfs_cksum.h"
 42 #include "xfs_trace.h"
 43 #include "xfs_icache.h"
 44 #include "xfs_bmap_btree.h"
 45 #include "xfs_error.h"
 46 #include "xfs_dir2.h"
 47 #include "xfs_rmap_item.h"
 48 #include "xfs_buf_item.h"
 49 #include "xfs_refcount_item.h"
 50 #include "xfs_bmap_item.h"
 51 
 52 #define BLK_AVG(blk1, blk2)     ((blk1+blk2) >> 1)
 53 
 54 STATIC int
 55 xlog_find_zeroed(
 56         struct xlog     *,
 57         xfs_daddr_t     *);
 58 STATIC int
 59 xlog_clear_stale_blocks(
 60         struct xlog     *,
 61         xfs_lsn_t);
 62 #if defined(DEBUG)
 63 STATIC void
 64 xlog_recover_check_summary(
 65         struct xlog *);
 66 #else
 67 #define xlog_recover_check_summary(log)
 68 #endif
 69 STATIC int
 70 xlog_do_recovery_pass(
 71         struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
 72 
 73 /*
 74  * This structure is used during recovery to record the buf log items which
 75  * have been canceled and should not be replayed.
 76  */
 77 struct xfs_buf_cancel {
 78         xfs_daddr_t             bc_blkno;
 79         uint                    bc_len;
 80         int                     bc_refcount;
 81         struct list_head        bc_list;
 82 };
 83 
 84 /*
 85  * Sector aligned buffer routines for buffer create/read/write/access
 86  */
 87 
 88 /*
 89  * Verify the log-relative block number and length in basic blocks are valid for
 90  * an operation involving the given XFS log buffer. Returns true if the fields
 91  * are valid, false otherwise.
 92  */
 93 static inline bool
 94 xlog_verify_bp(
 95         struct xlog     *log,
 96         xfs_daddr_t     blk_no,
 97         int             bbcount)
 98 {
 99         if (blk_no < 0 || blk_no >= log->l_logBBsize)
100                 return false;
101         if (bbcount <= 0 || (blk_no + bbcount) > log->l_logBBsize)
102                 return false;
103         return true;
104 }
105 
106 /*
107  * Allocate a buffer to hold log data.  The buffer needs to be able
108  * to map to a range of nbblks basic blocks at any valid (basic
109  * block) offset within the log.
110  */
111 STATIC xfs_buf_t *
112 xlog_get_bp(
113         struct xlog     *log,
114         int             nbblks)
115 {
116         struct xfs_buf  *bp;
117 
118         /*
119          * Pass log block 0 since we don't have an addr yet, buffer will be
120          * verified on read.
121          */
122         if (!xlog_verify_bp(log, 0, nbblks)) {
123                 xfs_warn(log->l_mp, "Invalid block length (0x%x) for buffer",
124                         nbblks);
125                 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
126                 return NULL;
127         }
128 
129         /*
130          * We do log I/O in units of log sectors (a power-of-2
131          * multiple of the basic block size), so we round up the
132          * requested size to accommodate the basic blocks required
133          * for complete log sectors.
134          *
135          * In addition, the buffer may be used for a non-sector-
136          * aligned block offset, in which case an I/O of the
137          * requested size could extend beyond the end of the
138          * buffer.  If the requested size is only 1 basic block it
139          * will never straddle a sector boundary, so this won't be
140          * an issue.  Nor will this be a problem if the log I/O is
141          * done in basic blocks (sector size 1).  But otherwise we
142          * extend the buffer by one extra log sector to ensure
143          * there's space to accommodate this possibility.
144          */
145         if (nbblks > 1 && log->l_sectBBsize > 1)
146                 nbblks += log->l_sectBBsize;
147         nbblks = round_up(nbblks, log->l_sectBBsize);
148 
149         bp = xfs_buf_get_uncached(log->l_mp->m_logdev_targp, nbblks, 0);
150         if (bp)
151                 xfs_buf_unlock(bp);
152         return bp;
153 }
154 
155 STATIC void
156 xlog_put_bp(
157         xfs_buf_t       *bp)
158 {
159         xfs_buf_free(bp);
160 }
161 
162 /*
163  * Return the address of the start of the given block number's data
164  * in a log buffer.  The buffer covers a log sector-aligned region.
165  */
166 STATIC char *
167 xlog_align(
168         struct xlog     *log,
169         xfs_daddr_t     blk_no,
170         int             nbblks,
171         struct xfs_buf  *bp)
172 {
173         xfs_daddr_t     offset = blk_no & ((xfs_daddr_t)log->l_sectBBsize - 1);
174 
175         ASSERT(offset + nbblks <= bp->b_length);
176         return bp->b_addr + BBTOB(offset);
177 }
178 
179 
180 /*
181  * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
182  */
183 STATIC int
184 xlog_bread_noalign(
185         struct xlog     *log,
186         xfs_daddr_t     blk_no,
187         int             nbblks,
188         struct xfs_buf  *bp)
189 {
190         int             error;
191 
192         if (!xlog_verify_bp(log, blk_no, nbblks)) {
193                 xfs_warn(log->l_mp,
194                          "Invalid log block/length (0x%llx, 0x%x) for buffer",
195                          blk_no, nbblks);
196                 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
197                 return -EFSCORRUPTED;
198         }
199 
200         blk_no = round_down(blk_no, log->l_sectBBsize);
201         nbblks = round_up(nbblks, log->l_sectBBsize);
202 
203         ASSERT(nbblks > 0);
204         ASSERT(nbblks <= bp->b_length);
205 
206         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
207         bp->b_flags |= XBF_READ;
208         bp->b_io_length = nbblks;
209         bp->b_error = 0;
210 
211         error = xfs_buf_submit_wait(bp);
212         if (error && !XFS_FORCED_SHUTDOWN(log->l_mp))
213                 xfs_buf_ioerror_alert(bp, __func__);
214         return error;
215 }
216 
217 STATIC int
218 xlog_bread(
219         struct xlog     *log,
220         xfs_daddr_t     blk_no,
221         int             nbblks,
222         struct xfs_buf  *bp,
223         char            **offset)
224 {
225         int             error;
226 
227         error = xlog_bread_noalign(log, blk_no, nbblks, bp);
228         if (error)
229                 return error;
230 
231         *offset = xlog_align(log, blk_no, nbblks, bp);
232         return 0;
233 }
234 
235 /*
236  * Read at an offset into the buffer. Returns with the buffer in it's original
237  * state regardless of the result of the read.
238  */
239 STATIC int
240 xlog_bread_offset(
241         struct xlog     *log,
242         xfs_daddr_t     blk_no,         /* block to read from */
243         int             nbblks,         /* blocks to read */
244         struct xfs_buf  *bp,
245         char            *offset)
246 {
247         char            *orig_offset = bp->b_addr;
248         int             orig_len = BBTOB(bp->b_length);
249         int             error, error2;
250 
251         error = xfs_buf_associate_memory(bp, offset, BBTOB(nbblks));
252         if (error)
253                 return error;
254 
255         error = xlog_bread_noalign(log, blk_no, nbblks, bp);
256 
257         /* must reset buffer pointer even on error */
258         error2 = xfs_buf_associate_memory(bp, orig_offset, orig_len);
259         if (error)
260                 return error;
261         return error2;
262 }
263 
264 /*
265  * Write out the buffer at the given block for the given number of blocks.
266  * The buffer is kept locked across the write and is returned locked.
267  * This can only be used for synchronous log writes.
268  */
269 STATIC int
270 xlog_bwrite(
271         struct xlog     *log,
272         xfs_daddr_t     blk_no,
273         int             nbblks,
274         struct xfs_buf  *bp)
275 {
276         int             error;
277 
278         if (!xlog_verify_bp(log, blk_no, nbblks)) {
279                 xfs_warn(log->l_mp,
280                          "Invalid log block/length (0x%llx, 0x%x) for buffer",
281                          blk_no, nbblks);
282                 XFS_ERROR_REPORT(__func__, XFS_ERRLEVEL_HIGH, log->l_mp);
283                 return -EFSCORRUPTED;
284         }
285 
286         blk_no = round_down(blk_no, log->l_sectBBsize);
287         nbblks = round_up(nbblks, log->l_sectBBsize);
288 
289         ASSERT(nbblks > 0);
290         ASSERT(nbblks <= bp->b_length);
291 
292         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
293         xfs_buf_hold(bp);
294         xfs_buf_lock(bp);
295         bp->b_io_length = nbblks;
296         bp->b_error = 0;
297 
298         error = xfs_bwrite(bp);
299         if (error)
300                 xfs_buf_ioerror_alert(bp, __func__);
301         xfs_buf_relse(bp);
302         return error;
303 }
304 
305 #ifdef DEBUG
306 /*
307  * dump debug superblock and log record information
308  */
309 STATIC void
310 xlog_header_check_dump(
311         xfs_mount_t             *mp,
312         xlog_rec_header_t       *head)
313 {
314         xfs_debug(mp, "%s:  SB : uuid = %pU, fmt = %d",
315                 __func__, &mp->m_sb.sb_uuid, XLOG_FMT);
316         xfs_debug(mp, "    log : uuid = %pU, fmt = %d",
317                 &head->h_fs_uuid, be32_to_cpu(head->h_fmt));
318 }
319 #else
320 #define xlog_header_check_dump(mp, head)
321 #endif
322 
323 /*
324  * check log record header for recovery
325  */
326 STATIC int
327 xlog_header_check_recover(
328         xfs_mount_t             *mp,
329         xlog_rec_header_t       *head)
330 {
331         ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
332 
333         /*
334          * IRIX doesn't write the h_fmt field and leaves it zeroed
335          * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
336          * a dirty log created in IRIX.
337          */
338         if (unlikely(head->h_fmt != cpu_to_be32(XLOG_FMT))) {
339                 xfs_warn(mp,
340         "dirty log written in incompatible format - can't recover");
341                 xlog_header_check_dump(mp, head);
342                 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
343                                  XFS_ERRLEVEL_HIGH, mp);
344                 return -EFSCORRUPTED;
345         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
346                 xfs_warn(mp,
347         "dirty log entry has mismatched uuid - can't recover");
348                 xlog_header_check_dump(mp, head);
349                 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
350                                  XFS_ERRLEVEL_HIGH, mp);
351                 return -EFSCORRUPTED;
352         }
353         return 0;
354 }
355 
356 /*
357  * read the head block of the log and check the header
358  */
359 STATIC int
360 xlog_header_check_mount(
361         xfs_mount_t             *mp,
362         xlog_rec_header_t       *head)
363 {
364         ASSERT(head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM));
365 
366         if (uuid_is_null(&head->h_fs_uuid)) {
367                 /*
368                  * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
369                  * h_fs_uuid is null, we assume this log was last mounted
370                  * by IRIX and continue.
371                  */
372                 xfs_warn(mp, "null uuid in log - IRIX style log");
373         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
374                 xfs_warn(mp, "log has mismatched uuid - can't recover");
375                 xlog_header_check_dump(mp, head);
376                 XFS_ERROR_REPORT("xlog_header_check_mount",
377                                  XFS_ERRLEVEL_HIGH, mp);
378                 return -EFSCORRUPTED;
379         }
380         return 0;
381 }
382 
383 STATIC void
384 xlog_recover_iodone(
385         struct xfs_buf  *bp)
386 {
387         if (bp->b_error) {
388                 /*
389                  * We're not going to bother about retrying
390                  * this during recovery. One strike!
391                  */
392                 if (!XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
393                         xfs_buf_ioerror_alert(bp, __func__);
394                         xfs_force_shutdown(bp->b_target->bt_mount,
395                                                 SHUTDOWN_META_IO_ERROR);
396                 }
397         }
398 
399         /*
400          * On v5 supers, a bli could be attached to update the metadata LSN.
401          * Clean it up.
402          */
403         if (bp->b_log_item)
404                 xfs_buf_item_relse(bp);
405         ASSERT(bp->b_log_item == NULL);
406 
407         bp->b_iodone = NULL;
408         xfs_buf_ioend(bp);
409 }
410 
411 /*
412  * This routine finds (to an approximation) the first block in the physical
413  * log which contains the given cycle.  It uses a binary search algorithm.
414  * Note that the algorithm can not be perfect because the disk will not
415  * necessarily be perfect.
416  */
417 STATIC int
418 xlog_find_cycle_start(
419         struct xlog     *log,
420         struct xfs_buf  *bp,
421         xfs_daddr_t     first_blk,
422         xfs_daddr_t     *last_blk,
423         uint            cycle)
424 {
425         char            *offset;
426         xfs_daddr_t     mid_blk;
427         xfs_daddr_t     end_blk;
428         uint            mid_cycle;
429         int             error;
430 
431         end_blk = *last_blk;
432         mid_blk = BLK_AVG(first_blk, end_blk);
433         while (mid_blk != first_blk && mid_blk != end_blk) {
434                 error = xlog_bread(log, mid_blk, 1, bp, &offset);
435                 if (error)
436                         return error;
437                 mid_cycle = xlog_get_cycle(offset);
438                 if (mid_cycle == cycle)
439                         end_blk = mid_blk;   /* last_half_cycle == mid_cycle */
440                 else
441                         first_blk = mid_blk; /* first_half_cycle == mid_cycle */
442                 mid_blk = BLK_AVG(first_blk, end_blk);
443         }
444         ASSERT((mid_blk == first_blk && mid_blk+1 == end_blk) ||
445                (mid_blk == end_blk && mid_blk-1 == first_blk));
446 
447         *last_blk = end_blk;
448 
449         return 0;
450 }
451 
452 /*
453  * Check that a range of blocks does not contain stop_on_cycle_no.
454  * Fill in *new_blk with the block offset where such a block is
455  * found, or with -1 (an invalid block number) if there is no such
456  * block in the range.  The scan needs to occur from front to back
457  * and the pointer into the region must be updated since a later
458  * routine will need to perform another test.
459  */
460 STATIC int
461 xlog_find_verify_cycle(
462         struct xlog     *log,
463         xfs_daddr_t     start_blk,
464         int             nbblks,
465         uint            stop_on_cycle_no,
466         xfs_daddr_t     *new_blk)
467 {
468         xfs_daddr_t     i, j;
469         uint            cycle;
470         xfs_buf_t       *bp;
471         xfs_daddr_t     bufblks;
472         char            *buf = NULL;
473         int             error = 0;
474 
475         /*
476          * Greedily allocate a buffer big enough to handle the full
477          * range of basic blocks we'll be examining.  If that fails,
478          * try a smaller size.  We need to be able to read at least
479          * a log sector, or we're out of luck.
480          */
481         bufblks = 1 << ffs(nbblks);
482         while (bufblks > log->l_logBBsize)
483                 bufblks >>= 1;
484         while (!(bp = xlog_get_bp(log, bufblks))) {
485                 bufblks >>= 1;
486                 if (bufblks < log->l_sectBBsize)
487                         return -ENOMEM;
488         }
489 
490         for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
491                 int     bcount;
492 
493                 bcount = min(bufblks, (start_blk + nbblks - i));
494 
495                 error = xlog_bread(log, i, bcount, bp, &buf);
496                 if (error)
497                         goto out;
498 
499                 for (j = 0; j < bcount; j++) {
500                         cycle = xlog_get_cycle(buf);
501                         if (cycle == stop_on_cycle_no) {
502                                 *new_blk = i+j;
503                                 goto out;
504                         }
505 
506                         buf += BBSIZE;
507                 }
508         }
509 
510         *new_blk = -1;
511 
512 out:
513         xlog_put_bp(bp);
514         return error;
515 }
516 
517 /*
518  * Potentially backup over partial log record write.
519  *
520  * In the typical case, last_blk is the number of the block directly after
521  * a good log record.  Therefore, we subtract one to get the block number
522  * of the last block in the given buffer.  extra_bblks contains the number
523  * of blocks we would have read on a previous read.  This happens when the
524  * last log record is split over the end of the physical log.
525  *
526  * extra_bblks is the number of blocks potentially verified on a previous
527  * call to this routine.
528  */
529 STATIC int
530 xlog_find_verify_log_record(
531         struct xlog             *log,
532         xfs_daddr_t             start_blk,
533         xfs_daddr_t             *last_blk,
534         int                     extra_bblks)
535 {
536         xfs_daddr_t             i;
537         xfs_buf_t               *bp;
538         char                    *offset = NULL;
539         xlog_rec_header_t       *head = NULL;
540         int                     error = 0;
541         int                     smallmem = 0;
542         int                     num_blks = *last_blk - start_blk;
543         int                     xhdrs;
544 
545         ASSERT(start_blk != 0 || *last_blk != start_blk);
546 
547         if (!(bp = xlog_get_bp(log, num_blks))) {
548                 if (!(bp = xlog_get_bp(log, 1)))
549                         return -ENOMEM;
550                 smallmem = 1;
551         } else {
552                 error = xlog_bread(log, start_blk, num_blks, bp, &offset);
553                 if (error)
554                         goto out;
555                 offset += ((num_blks - 1) << BBSHIFT);
556         }
557 
558         for (i = (*last_blk) - 1; i >= 0; i--) {
559                 if (i < start_blk) {
560                         /* valid log record not found */
561                         xfs_warn(log->l_mp,
562                 "Log inconsistent (didn't find previous header)");
563                         ASSERT(0);
564                         error = -EIO;
565                         goto out;
566                 }
567 
568                 if (smallmem) {
569                         error = xlog_bread(log, i, 1, bp, &offset);
570                         if (error)
571                                 goto out;
572                 }
573 
574                 head = (xlog_rec_header_t *)offset;
575 
576                 if (head->h_magicno == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
577                         break;
578 
579                 if (!smallmem)
580                         offset -= BBSIZE;
581         }
582 
583         /*
584          * We hit the beginning of the physical log & still no header.  Return
585          * to caller.  If caller can handle a return of -1, then this routine
586          * will be called again for the end of the physical log.
587          */
588         if (i == -1) {
589                 error = 1;
590                 goto out;
591         }
592 
593         /*
594          * We have the final block of the good log (the first block
595          * of the log record _before_ the head. So we check the uuid.
596          */
597         if ((error = xlog_header_check_mount(log->l_mp, head)))
598                 goto out;
599 
600         /*
601          * We may have found a log record header before we expected one.
602          * last_blk will be the 1st block # with a given cycle #.  We may end
603          * up reading an entire log record.  In this case, we don't want to
604          * reset last_blk.  Only when last_blk points in the middle of a log
605          * record do we update last_blk.
606          */
607         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
608                 uint    h_size = be32_to_cpu(head->h_size);
609 
610                 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
611                 if (h_size % XLOG_HEADER_CYCLE_SIZE)
612                         xhdrs++;
613         } else {
614                 xhdrs = 1;
615         }
616 
617         if (*last_blk - i + extra_bblks !=
618             BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
619                 *last_blk = i;
620 
621 out:
622         xlog_put_bp(bp);
623         return error;
624 }
625 
626 /*
627  * Head is defined to be the point of the log where the next log write
628  * could go.  This means that incomplete LR writes at the end are
629  * eliminated when calculating the head.  We aren't guaranteed that previous
630  * LR have complete transactions.  We only know that a cycle number of
631  * current cycle number -1 won't be present in the log if we start writing
632  * from our current block number.
633  *
634  * last_blk contains the block number of the first block with a given
635  * cycle number.
636  *
637  * Return: zero if normal, non-zero if error.
638  */
639 STATIC int
640 xlog_find_head(
641         struct xlog     *log,
642         xfs_daddr_t     *return_head_blk)
643 {
644         xfs_buf_t       *bp;
645         char            *offset;
646         xfs_daddr_t     new_blk, first_blk, start_blk, last_blk, head_blk;
647         int             num_scan_bblks;
648         uint            first_half_cycle, last_half_cycle;
649         uint            stop_on_cycle;
650         int             error, log_bbnum = log->l_logBBsize;
651 
652         /* Is the end of the log device zeroed? */
653         error = xlog_find_zeroed(log, &first_blk);
654         if (error < 0) {
655                 xfs_warn(log->l_mp, "empty log check failed");
656                 return error;
657         }
658         if (error == 1) {
659                 *return_head_blk = first_blk;
660 
661                 /* Is the whole lot zeroed? */
662                 if (!first_blk) {
663                         /* Linux XFS shouldn't generate totally zeroed logs -
664                          * mkfs etc write a dummy unmount record to a fresh
665                          * log so we can store the uuid in there
666                          */
667                         xfs_warn(log->l_mp, "totally zeroed log");
668                 }
669 
670                 return 0;
671         }
672 
673         first_blk = 0;                  /* get cycle # of 1st block */
674         bp = xlog_get_bp(log, 1);
675         if (!bp)
676                 return -ENOMEM;
677 
678         error = xlog_bread(log, 0, 1, bp, &offset);
679         if (error)
680                 goto bp_err;
681 
682         first_half_cycle = xlog_get_cycle(offset);
683 
684         last_blk = head_blk = log_bbnum - 1;    /* get cycle # of last block */
685         error = xlog_bread(log, last_blk, 1, bp, &offset);
686         if (error)
687                 goto bp_err;
688 
689         last_half_cycle = xlog_get_cycle(offset);
690         ASSERT(last_half_cycle != 0);
691 
692         /*
693          * If the 1st half cycle number is equal to the last half cycle number,
694          * then the entire log is stamped with the same cycle number.  In this
695          * case, head_blk can't be set to zero (which makes sense).  The below
696          * math doesn't work out properly with head_blk equal to zero.  Instead,
697          * we set it to log_bbnum which is an invalid block number, but this
698          * value makes the math correct.  If head_blk doesn't changed through
699          * all the tests below, *head_blk is set to zero at the very end rather
700          * than log_bbnum.  In a sense, log_bbnum and zero are the same block
701          * in a circular file.
702          */
703         if (first_half_cycle == last_half_cycle) {
704                 /*
705                  * In this case we believe that the entire log should have
706                  * cycle number last_half_cycle.  We need to scan backwards
707                  * from the end verifying that there are no holes still
708                  * containing last_half_cycle - 1.  If we find such a hole,
709                  * then the start of that hole will be the new head.  The
710                  * simple case looks like
711                  *        x | x ... | x - 1 | x
712                  * Another case that fits this picture would be
713                  *        x | x + 1 | x ... | x
714                  * In this case the head really is somewhere at the end of the
715                  * log, as one of the latest writes at the beginning was
716                  * incomplete.
717                  * One more case is
718                  *        x | x + 1 | x ... | x - 1 | x
719                  * This is really the combination of the above two cases, and
720                  * the head has to end up at the start of the x-1 hole at the
721                  * end of the log.
722                  *
723                  * In the 256k log case, we will read from the beginning to the
724                  * end of the log and search for cycle numbers equal to x-1.
725                  * We don't worry about the x+1 blocks that we encounter,
726                  * because we know that they cannot be the head since the log
727                  * started with x.
728                  */
729                 head_blk = log_bbnum;
730                 stop_on_cycle = last_half_cycle - 1;
731         } else {
732                 /*
733                  * In this case we want to find the first block with cycle
734                  * number matching last_half_cycle.  We expect the log to be
735                  * some variation on
736                  *        x + 1 ... | x ... | x
737                  * The first block with cycle number x (last_half_cycle) will
738                  * be where the new head belongs.  First we do a binary search
739                  * for the first occurrence of last_half_cycle.  The binary
740                  * search may not be totally accurate, so then we scan back
741                  * from there looking for occurrences of last_half_cycle before
742                  * us.  If that backwards scan wraps around the beginning of
743                  * the log, then we look for occurrences of last_half_cycle - 1
744                  * at the end of the log.  The cases we're looking for look
745                  * like
746                  *                               v binary search stopped here
747                  *        x + 1 ... | x | x + 1 | x ... | x
748                  *                   ^ but we want to locate this spot
749                  * or
750                  *        <---------> less than scan distance
751                  *        x + 1 ... | x ... | x - 1 | x
752                  *                           ^ we want to locate this spot
753                  */
754                 stop_on_cycle = last_half_cycle;
755                 if ((error = xlog_find_cycle_start(log, bp, first_blk,
756                                                 &head_blk, last_half_cycle)))
757                         goto bp_err;
758         }
759 
760         /*
761          * Now validate the answer.  Scan back some number of maximum possible
762          * blocks and make sure each one has the expected cycle number.  The
763          * maximum is determined by the total possible amount of buffering
764          * in the in-core log.  The following number can be made tighter if
765          * we actually look at the block size of the filesystem.
766          */
767         num_scan_bblks = min_t(int, log_bbnum, XLOG_TOTAL_REC_SHIFT(log));
768         if (head_blk >= num_scan_bblks) {
769                 /*
770                  * We are guaranteed that the entire check can be performed
771                  * in one buffer.
772                  */
773                 start_blk = head_blk - num_scan_bblks;
774                 if ((error = xlog_find_verify_cycle(log,
775                                                 start_blk, num_scan_bblks,
776                                                 stop_on_cycle, &new_blk)))
777                         goto bp_err;
778                 if (new_blk != -1)
779                         head_blk = new_blk;
780         } else {                /* need to read 2 parts of log */
781                 /*
782                  * We are going to scan backwards in the log in two parts.
783                  * First we scan the physical end of the log.  In this part
784                  * of the log, we are looking for blocks with cycle number
785                  * last_half_cycle - 1.
786                  * If we find one, then we know that the log starts there, as
787                  * we've found a hole that didn't get written in going around
788                  * the end of the physical log.  The simple case for this is
789                  *        x + 1 ... | x ... | x - 1 | x
790                  *        <---------> less than scan distance
791                  * If all of the blocks at the end of the log have cycle number
792                  * last_half_cycle, then we check the blocks at the start of
793                  * the log looking for occurrences of last_half_cycle.  If we
794                  * find one, then our current estimate for the location of the
795                  * first occurrence of last_half_cycle is wrong and we move
796                  * back to the hole we've found.  This case looks like
797                  *        x + 1 ... | x | x + 1 | x ...
798                  *                               ^ binary search stopped here
799                  * Another case we need to handle that only occurs in 256k
800                  * logs is
801                  *        x + 1 ... | x ... | x+1 | x ...
802                  *                   ^ binary search stops here
803                  * In a 256k log, the scan at the end of the log will see the
804                  * x + 1 blocks.  We need to skip past those since that is
805                  * certainly not the head of the log.  By searching for
806                  * last_half_cycle-1 we accomplish that.
807                  */
808                 ASSERT(head_blk <= INT_MAX &&
809                         (xfs_daddr_t) num_scan_bblks >= head_blk);
810                 start_blk = log_bbnum - (num_scan_bblks - head_blk);
811                 if ((error = xlog_find_verify_cycle(log, start_blk,
812                                         num_scan_bblks - (int)head_blk,
813                                         (stop_on_cycle - 1), &new_blk)))
814                         goto bp_err;
815                 if (new_blk != -1) {
816                         head_blk = new_blk;
817                         goto validate_head;
818                 }
819 
820                 /*
821                  * Scan beginning of log now.  The last part of the physical
822                  * log is good.  This scan needs to verify that it doesn't find
823                  * the last_half_cycle.
824                  */
825                 start_blk = 0;
826                 ASSERT(head_blk <= INT_MAX);
827                 if ((error = xlog_find_verify_cycle(log,
828                                         start_blk, (int)head_blk,
829                                         stop_on_cycle, &new_blk)))
830                         goto bp_err;
831                 if (new_blk != -1)
832                         head_blk = new_blk;
833         }
834 
835 validate_head:
836         /*
837          * Now we need to make sure head_blk is not pointing to a block in
838          * the middle of a log record.
839          */
840         num_scan_bblks = XLOG_REC_SHIFT(log);
841         if (head_blk >= num_scan_bblks) {
842                 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
843 
844                 /* start ptr at last block ptr before head_blk */
845                 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
846                 if (error == 1)
847                         error = -EIO;
848                 if (error)
849                         goto bp_err;
850         } else {
851                 start_blk = 0;
852                 ASSERT(head_blk <= INT_MAX);
853                 error = xlog_find_verify_log_record(log, start_blk, &head_blk, 0);
854                 if (error < 0)
855                         goto bp_err;
856                 if (error == 1) {
857                         /* We hit the beginning of the log during our search */
858                         start_blk = log_bbnum - (num_scan_bblks - head_blk);
859                         new_blk = log_bbnum;
860                         ASSERT(start_blk <= INT_MAX &&
861                                 (xfs_daddr_t) log_bbnum-start_blk >= 0);
862                         ASSERT(head_blk <= INT_MAX);
863                         error = xlog_find_verify_log_record(log, start_blk,
864                                                         &new_blk, (int)head_blk);
865                         if (error == 1)
866                                 error = -EIO;
867                         if (error)
868                                 goto bp_err;
869                         if (new_blk != log_bbnum)
870                                 head_blk = new_blk;
871                 } else if (error)
872                         goto bp_err;
873         }
874 
875         xlog_put_bp(bp);
876         if (head_blk == log_bbnum)
877                 *return_head_blk = 0;
878         else
879                 *return_head_blk = head_blk;
880         /*
881          * When returning here, we have a good block number.  Bad block
882          * means that during a previous crash, we didn't have a clean break
883          * from cycle number N to cycle number N-1.  In this case, we need
884          * to find the first block with cycle number N-1.
885          */
886         return 0;
887 
888  bp_err:
889         xlog_put_bp(bp);
890 
891         if (error)
892                 xfs_warn(log->l_mp, "failed to find log head");
893         return error;
894 }
895 
896 /*
897  * Seek backwards in the log for log record headers.
898  *
899  * Given a starting log block, walk backwards until we find the provided number
900  * of records or hit the provided tail block. The return value is the number of
901  * records encountered or a negative error code. The log block and buffer
902  * pointer of the last record seen are returned in rblk and rhead respectively.
903  */
904 STATIC int
905 xlog_rseek_logrec_hdr(
906         struct xlog             *log,
907         xfs_daddr_t             head_blk,
908         xfs_daddr_t             tail_blk,
909         int                     count,
910         struct xfs_buf          *bp,
911         xfs_daddr_t             *rblk,
912         struct xlog_rec_header  **rhead,
913         bool                    *wrapped)
914 {
915         int                     i;
916         int                     error;
917         int                     found = 0;
918         char                    *offset = NULL;
919         xfs_daddr_t             end_blk;
920 
921         *wrapped = false;
922 
923         /*
924          * Walk backwards from the head block until we hit the tail or the first
925          * block in the log.
926          */
927         end_blk = head_blk > tail_blk ? tail_blk : 0;
928         for (i = (int) head_blk - 1; i >= end_blk; i--) {
929                 error = xlog_bread(log, i, 1, bp, &offset);
930                 if (error)
931                         goto out_error;
932 
933                 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
934                         *rblk = i;
935                         *rhead = (struct xlog_rec_header *) offset;
936                         if (++found == count)
937                                 break;
938                 }
939         }
940 
941         /*
942          * If we haven't hit the tail block or the log record header count,
943          * start looking again from the end of the physical log. Note that
944          * callers can pass head == tail if the tail is not yet known.
945          */
946         if (tail_blk >= head_blk && found != count) {
947                 for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
948                         error = xlog_bread(log, i, 1, bp, &offset);
949                         if (error)
950                                 goto out_error;
951 
952                         if (*(__be32 *)offset ==
953                             cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
954                                 *wrapped = true;
955                                 *rblk = i;
956                                 *rhead = (struct xlog_rec_header *) offset;
957                                 if (++found == count)
958                                         break;
959                         }
960                 }
961         }
962 
963         return found;
964 
965 out_error:
966         return error;
967 }
968 
969 /*
970  * Seek forward in the log for log record headers.
971  *
972  * Given head and tail blocks, walk forward from the tail block until we find
973  * the provided number of records or hit the head block. The return value is the
974  * number of records encountered or a negative error code. The log block and
975  * buffer pointer of the last record seen are returned in rblk and rhead
976  * respectively.
977  */
978 STATIC int
979 xlog_seek_logrec_hdr(
980         struct xlog             *log,
981         xfs_daddr_t             head_blk,
982         xfs_daddr_t             tail_blk,
983         int                     count,
984         struct xfs_buf          *bp,
985         xfs_daddr_t             *rblk,
986         struct xlog_rec_header  **rhead,
987         bool                    *wrapped)
988 {
989         int                     i;
990         int                     error;
991         int                     found = 0;
992         char                    *offset = NULL;
993         xfs_daddr_t             end_blk;
994 
995         *wrapped = false;
996 
997         /*
998          * Walk forward from the tail block until we hit the head or the last
999          * block in the log.
1000          */
1001         end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
1002         for (i = (int) tail_blk; i <= end_blk; i++) {
1003                 error = xlog_bread(log, i, 1, bp, &offset);
1004                 if (error)
1005                         goto out_error;
1006 
1007                 if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
1008                         *rblk = i;
1009                         *rhead = (struct xlog_rec_header *) offset;
1010                         if (++found == count)
1011                                 break;
1012                 }
1013         }
1014 
1015         /*
1016          * If we haven't hit the head block or the log record header count,
1017          * start looking again from the start of the physical log.
1018          */
1019         if (tail_blk > head_blk && found != count) {
1020                 for (i = 0; i < (int) head_blk; i++) {
1021                         error = xlog_bread(log, i, 1, bp, &offset);
1022                         if (error)
1023                                 goto out_error;
1024 
1025                         if (*(__be32 *)offset ==
1026                             cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
1027                                 *wrapped = true;
1028                                 *rblk = i;
1029                                 *rhead = (struct xlog_rec_header *) offset;
1030                                 if (++found == count)
1031                                         break;
1032                         }
1033                 }
1034         }
1035 
1036         return found;
1037 
1038 out_error:
1039         return error;
1040 }
1041 
1042 /*
1043  * Calculate distance from head to tail (i.e., unused space in the log).
1044  */
1045 static inline int
1046 xlog_tail_distance(
1047         struct xlog     *log,
1048         xfs_daddr_t     head_blk,
1049         xfs_daddr_t     tail_blk)
1050 {
1051         if (head_blk < tail_blk)
1052                 return tail_blk - head_blk;
1053 
1054         return tail_blk + (log->l_logBBsize - head_blk);
1055 }
1056 
1057 /*
1058  * Verify the log tail. This is particularly important when torn or incomplete
1059  * writes have been detected near the front of the log and the head has been
1060  * walked back accordingly.
1061  *
1062  * We also have to handle the case where the tail was pinned and the head
1063  * blocked behind the tail right before a crash. If the tail had been pushed
1064  * immediately prior to the crash and the subsequent checkpoint was only
1065  * partially written, it's possible it overwrote the last referenced tail in the
1066  * log with garbage. This is not a coherency problem because the tail must have
1067  * been pushed before it can be overwritten, but appears as log corruption to
1068  * recovery because we have no way to know the tail was updated if the
1069  * subsequent checkpoint didn't write successfully.
1070  *
1071  * Therefore, CRC check the log from tail to head. If a failure occurs and the
1072  * offending record is within max iclog bufs from the head, walk the tail
1073  * forward and retry until a valid tail is found or corruption is detected out
1074  * of the range of a possible overwrite.
1075  */
1076 STATIC int
1077 xlog_verify_tail(
1078         struct xlog             *log,
1079         xfs_daddr_t             head_blk,
1080         xfs_daddr_t             *tail_blk,
1081         int                     hsize)
1082 {
1083         struct xlog_rec_header  *thead;
1084         struct xfs_buf          *bp;
1085         xfs_daddr_t             first_bad;
1086         int                     error = 0;
1087         bool                    wrapped;
1088         xfs_daddr_t             tmp_tail;
1089         xfs_daddr_t             orig_tail = *tail_blk;
1090 
1091         bp = xlog_get_bp(log, 1);
1092         if (!bp)
1093                 return -ENOMEM;
1094 
1095         /*
1096          * Make sure the tail points to a record (returns positive count on
1097          * success).
1098          */
1099         error = xlog_seek_logrec_hdr(log, head_blk, *tail_blk, 1, bp,
1100                         &tmp_tail, &thead, &wrapped);
1101         if (error < 0)
1102                 goto out;
1103         if (*tail_blk != tmp_tail)
1104                 *tail_blk = tmp_tail;
1105 
1106         /*
1107          * Run a CRC check from the tail to the head. We can't just check
1108          * MAX_ICLOGS records past the tail because the tail may point to stale
1109          * blocks cleared during the search for the head/tail. These blocks are
1110          * overwritten with zero-length records and thus record count is not a
1111          * reliable indicator of the iclog state before a crash.
1112          */
1113         first_bad = 0;
1114         error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1115                                       XLOG_RECOVER_CRCPASS, &first_bad);
1116         while ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1117                 int     tail_distance;
1118 
1119                 /*
1120                  * Is corruption within range of the head? If so, retry from
1121                  * the next record. Otherwise return an error.
1122                  */
1123                 tail_distance = xlog_tail_distance(log, head_blk, first_bad);
1124                 if (tail_distance > BTOBB(XLOG_MAX_ICLOGS * hsize))
1125                         break;
1126 
1127                 /* skip to the next record; returns positive count on success */
1128                 error = xlog_seek_logrec_hdr(log, head_blk, first_bad, 2, bp,
1129                                 &tmp_tail, &thead, &wrapped);
1130                 if (error < 0)
1131                         goto out;
1132 
1133                 *tail_blk = tmp_tail;
1134                 first_bad = 0;
1135                 error = xlog_do_recovery_pass(log, head_blk, *tail_blk,
1136                                               XLOG_RECOVER_CRCPASS, &first_bad);
1137         }
1138 
1139         if (!error && *tail_blk != orig_tail)
1140                 xfs_warn(log->l_mp,
1141                 "Tail block (0x%llx) overwrite detected. Updated to 0x%llx",
1142                          orig_tail, *tail_blk);
1143 out:
1144         xlog_put_bp(bp);
1145         return error;
1146 }
1147 
1148 /*
1149  * Detect and trim torn writes from the head of the log.
1150  *
1151  * Storage without sector atomicity guarantees can result in torn writes in the
1152  * log in the event of a crash. Our only means to detect this scenario is via
1153  * CRC verification. While we can't always be certain that CRC verification
1154  * failure is due to a torn write vs. an unrelated corruption, we do know that
1155  * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
1156  * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
1157  * the log and treat failures in this range as torn writes as a matter of
1158  * policy. In the event of CRC failure, the head is walked back to the last good
1159  * record in the log and the tail is updated from that record and verified.
1160  */
1161 STATIC int
1162 xlog_verify_head(
1163         struct xlog             *log,
1164         xfs_daddr_t             *head_blk,      /* in/out: unverified head */
1165         xfs_daddr_t             *tail_blk,      /* out: tail block */
1166         struct xfs_buf          *bp,
1167         xfs_daddr_t             *rhead_blk,     /* start blk of last record */
1168         struct xlog_rec_header  **rhead,        /* ptr to last record */
1169         bool                    *wrapped)       /* last rec. wraps phys. log */
1170 {
1171         struct xlog_rec_header  *tmp_rhead;
1172         struct xfs_buf          *tmp_bp;
1173         xfs_daddr_t             first_bad;
1174         xfs_daddr_t             tmp_rhead_blk;
1175         int                     found;
1176         int                     error;
1177         bool                    tmp_wrapped;
1178 
1179         /*
1180          * Check the head of the log for torn writes. Search backwards from the
1181          * head until we hit the tail or the maximum number of log record I/Os
1182          * that could have been in flight at one time. Use a temporary buffer so
1183          * we don't trash the rhead/bp pointers from the caller.
1184          */
1185         tmp_bp = xlog_get_bp(log, 1);
1186         if (!tmp_bp)
1187                 return -ENOMEM;
1188         error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
1189                                       XLOG_MAX_ICLOGS, tmp_bp, &tmp_rhead_blk,
1190                                       &tmp_rhead, &tmp_wrapped);
1191         xlog_put_bp(tmp_bp);
1192         if (error < 0)
1193                 return error;
1194 
1195         /*
1196          * Now run a CRC verification pass over the records starting at the
1197          * block found above to the current head. If a CRC failure occurs, the
1198          * log block of the first bad record is saved in first_bad.
1199          */
1200         error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
1201                                       XLOG_RECOVER_CRCPASS, &first_bad);
1202         if ((error == -EFSBADCRC || error == -EFSCORRUPTED) && first_bad) {
1203                 /*
1204                  * We've hit a potential torn write. Reset the error and warn
1205                  * about it.
1206                  */
1207                 error = 0;
1208                 xfs_warn(log->l_mp,
1209 "Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
1210                          first_bad, *head_blk);
1211 
1212                 /*
1213                  * Get the header block and buffer pointer for the last good
1214                  * record before the bad record.
1215                  *
1216                  * Note that xlog_find_tail() clears the blocks at the new head
1217                  * (i.e., the records with invalid CRC) if the cycle number
1218                  * matches the the current cycle.
1219                  */
1220                 found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1, bp,
1221                                               rhead_blk, rhead, wrapped);
1222                 if (found < 0)
1223                         return found;
1224                 if (found == 0)         /* XXX: right thing to do here? */
1225                         return -EIO;
1226 
1227                 /*
1228                  * Reset the head block to the starting block of the first bad
1229                  * log record and set the tail block based on the last good
1230                  * record.
1231                  *
1232                  * Bail out if the updated head/tail match as this indicates
1233                  * possible corruption outside of the acceptable
1234                  * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
1235                  */
1236                 *head_blk = first_bad;
1237                 *tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
1238                 if (*head_blk == *tail_blk) {
1239                         ASSERT(0);
1240                         return 0;
1241                 }
1242         }
1243         if (error)
1244                 return error;
1245 
1246         return xlog_verify_tail(log, *head_blk, tail_blk,
1247                                 be32_to_cpu((*rhead)->h_size));
1248 }
1249 
1250 /*
1251  * Check whether the head of the log points to an unmount record. In other
1252  * words, determine whether the log is clean. If so, update the in-core state
1253  * appropriately.
1254  */
1255 static int
1256 xlog_check_unmount_rec(
1257         struct xlog             *log,
1258         xfs_daddr_t             *head_blk,
1259         xfs_daddr_t             *tail_blk,
1260         struct xlog_rec_header  *rhead,
1261         xfs_daddr_t             rhead_blk,
1262         struct xfs_buf          *bp,
1263         bool                    *clean)
1264 {
1265         struct xlog_op_header   *op_head;
1266         xfs_daddr_t             umount_data_blk;
1267         xfs_daddr_t             after_umount_blk;
1268         int                     hblks;
1269         int                     error;
1270         char                    *offset;
1271 
1272         *clean = false;
1273 
1274         /*
1275          * Look for unmount record. If we find it, then we know there was a
1276          * clean unmount. Since 'i' could be the last block in the physical
1277          * log, we convert to a log block before comparing to the head_blk.
1278          *
1279          * Save the current tail lsn to use to pass to xlog_clear_stale_blocks()
1280          * below. We won't want to clear the unmount record if there is one, so
1281          * we pass the lsn of the unmount record rather than the block after it.
1282          */
1283         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1284                 int     h_size = be32_to_cpu(rhead->h_size);
1285                 int     h_version = be32_to_cpu(rhead->h_version);
1286 
1287                 if ((h_version & XLOG_VERSION_2) &&
1288                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
1289                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
1290                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
1291                                 hblks++;
1292                 } else {
1293                         hblks = 1;
1294                 }
1295         } else {
1296                 hblks = 1;
1297         }
1298         after_umount_blk = rhead_blk + hblks + BTOBB(be32_to_cpu(rhead->h_len));
1299         after_umount_blk = do_mod(after_umount_blk, log->l_logBBsize);
1300         if (*head_blk == after_umount_blk &&
1301             be32_to_cpu(rhead->h_num_logops) == 1) {
1302                 umount_data_blk = rhead_blk + hblks;
1303                 umount_data_blk = do_mod(umount_data_blk, log->l_logBBsize);
1304                 error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
1305                 if (error)
1306                         return error;
1307 
1308                 op_head = (struct xlog_op_header *)offset;
1309                 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
1310                         /*
1311                          * Set tail and last sync so that newly written log
1312                          * records will point recovery to after the current
1313                          * unmount record.
1314                          */
1315                         xlog_assign_atomic_lsn(&log->l_tail_lsn,
1316                                         log->l_curr_cycle, after_umount_blk);
1317                         xlog_assign_atomic_lsn(&log->l_last_sync_lsn,
1318                                         log->l_curr_cycle, after_umount_blk);
1319                         *tail_blk = after_umount_blk;
1320 
1321                         *clean = true;
1322                 }
1323         }
1324 
1325         return 0;
1326 }
1327 
1328 static void
1329 xlog_set_state(
1330         struct xlog             *log,
1331         xfs_daddr_t             head_blk,
1332         struct xlog_rec_header  *rhead,
1333         xfs_daddr_t             rhead_blk,
1334         bool                    bump_cycle)
1335 {
1336         /*
1337          * Reset log values according to the state of the log when we
1338          * crashed.  In the case where head_blk == 0, we bump curr_cycle
1339          * one because the next write starts a new cycle rather than
1340          * continuing the cycle of the last good log record.  At this
1341          * point we have guaranteed that all partial log records have been
1342          * accounted for.  Therefore, we know that the last good log record
1343          * written was complete and ended exactly on the end boundary
1344          * of the physical log.
1345          */
1346         log->l_prev_block = rhead_blk;
1347         log->l_curr_block = (int)head_blk;
1348         log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
1349         if (bump_cycle)
1350                 log->l_curr_cycle++;
1351         atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
1352         atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));
1353         xlog_assign_grant_head(&log->l_reserve_head.grant, log->l_curr_cycle,
1354                                         BBTOB(log->l_curr_block));
1355         xlog_assign_grant_head(&log->l_write_head.grant, log->l_curr_cycle,
1356                                         BBTOB(log->l_curr_block));
1357 }
1358 
1359 /*
1360  * Find the sync block number or the tail of the log.
1361  *
1362  * This will be the block number of the last record to have its
1363  * associated buffers synced to disk.  Every log record header has
1364  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
1365  * to get a sync block number.  The only concern is to figure out which
1366  * log record header to believe.
1367  *
1368  * The following algorithm uses the log record header with the largest
1369  * lsn.  The entire log record does not need to be valid.  We only care
1370  * that the header is valid.
1371  *
1372  * We could speed up search by using current head_blk buffer, but it is not
1373  * available.
1374  */
1375 STATIC int
1376 xlog_find_tail(
1377         struct xlog             *log,
1378         xfs_daddr_t             *head_blk,
1379         xfs_daddr_t             *tail_blk)
1380 {
1381         xlog_rec_header_t       *rhead;
1382         char                    *offset = NULL;
1383         xfs_buf_t               *bp;
1384         int                     error;
1385         xfs_daddr_t             rhead_blk;
1386         xfs_lsn_t               tail_lsn;
1387         bool                    wrapped = false;
1388         bool                    clean = false;
1389 
1390         /*
1391          * Find previous log record
1392          */
1393         if ((error = xlog_find_head(log, head_blk)))
1394                 return error;
1395         ASSERT(*head_blk < INT_MAX);
1396 
1397         bp = xlog_get_bp(log, 1);
1398         if (!bp)
1399                 return -ENOMEM;
1400         if (*head_blk == 0) {                           /* special case */
1401                 error = xlog_bread(log, 0, 1, bp, &offset);
1402                 if (error)
1403                         goto done;
1404 
1405                 if (xlog_get_cycle(offset) == 0) {
1406                         *tail_blk = 0;
1407                         /* leave all other log inited values alone */
1408                         goto done;
1409                 }
1410         }
1411 
1412         /*
1413          * Search backwards through the log looking for the log record header
1414          * block. This wraps all the way back around to the head so something is
1415          * seriously wrong if we can't find it.
1416          */
1417         error = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, bp,
1418                                       &rhead_blk, &rhead, &wrapped);
1419         if (error < 0)
1420                 return error;
1421         if (!error) {
1422                 xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
1423                 return -EIO;
1424         }
1425         *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
1426 
1427         /*
1428          * Set the log state based on the current head record.
1429          */
1430         xlog_set_state(log, *head_blk, rhead, rhead_blk, wrapped);
1431         tail_lsn = atomic64_read(&log->l_tail_lsn);
1432 
1433         /*
1434          * Look for an unmount record at the head of the log. This sets the log
1435          * state to determine whether recovery is necessary.
1436          */
1437         error = xlog_check_unmount_rec(log, head_blk, tail_blk, rhead,
1438                                        rhead_blk, bp, &clean);
1439         if (error)
1440                 goto done;
1441 
1442         /*
1443          * Verify the log head if the log is not clean (e.g., we have anything
1444          * but an unmount record at the head). This uses CRC verification to
1445          * detect and trim torn writes. If discovered, CRC failures are
1446          * considered torn writes and the log head is trimmed accordingly.
1447          *
1448          * Note that we can only run CRC verification when the log is dirty
1449          * because there's no guarantee that the log data behind an unmount
1450          * record is compatible with the current architecture.
1451          */
1452         if (!clean) {
1453                 xfs_daddr_t     orig_head = *head_blk;
1454 
1455                 error = xlog_verify_head(log, head_blk, tail_blk, bp,
1456                                          &rhead_blk, &rhead, &wrapped);
1457                 if (error)
1458                         goto done;
1459 
1460                 /* update in-core state again if the head changed */
1461                 if (*head_blk != orig_head) {
1462                         xlog_set_state(log, *head_blk, rhead, rhead_blk,
1463                                        wrapped);
1464                         tail_lsn = atomic64_read(&log->l_tail_lsn);
1465                         error = xlog_check_unmount_rec(log, head_blk, tail_blk,
1466                                                        rhead, rhead_blk, bp,
1467                                                        &clean);
1468                         if (error)
1469                                 goto done;
1470                 }
1471         }
1472 
1473         /*
1474          * Note that the unmount was clean. If the unmount was not clean, we
1475          * need to know this to rebuild the superblock counters from the perag
1476          * headers if we have a filesystem using non-persistent counters.
1477          */
1478         if (clean)
1479                 log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
1480 
1481         /*
1482          * Make sure that there are no blocks in front of the head
1483          * with the same cycle number as the head.  This can happen
1484          * because we allow multiple outstanding log writes concurrently,
1485          * and the later writes might make it out before earlier ones.
1486          *
1487          * We use the lsn from before modifying it so that we'll never
1488          * overwrite the unmount record after a clean unmount.
1489          *
1490          * Do this only if we are going to recover the filesystem
1491          *
1492          * NOTE: This used to say "if (!readonly)"
1493          * However on Linux, we can & do recover a read-only filesystem.
1494          * We only skip recovery if NORECOVERY is specified on mount,
1495          * in which case we would not be here.
1496          *
1497          * But... if the -device- itself is readonly, just skip this.
1498          * We can't recover this device anyway, so it won't matter.
1499          */
1500         if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp))
1501                 error = xlog_clear_stale_blocks(log, tail_lsn);
1502 
1503 done:
1504         xlog_put_bp(bp);
1505 
1506         if (error)
1507                 xfs_warn(log->l_mp, "failed to locate log tail");
1508         return error;
1509 }
1510 
1511 /*
1512  * Is the log zeroed at all?
1513  *
1514  * The last binary search should be changed to perform an X block read
1515  * once X becomes small enough.  You can then search linearly through
1516  * the X blocks.  This will cut down on the number of reads we need to do.
1517  *
1518  * If the log is partially zeroed, this routine will pass back the blkno
1519  * of the first block with cycle number 0.  It won't have a complete LR
1520  * preceding it.
1521  *
1522  * Return:
1523  *      0  => the log is completely written to
1524  *      1 => use *blk_no as the first block of the log
1525  *      <0 => error has occurred
1526  */
1527 STATIC int
1528 xlog_find_zeroed(
1529         struct xlog     *log,
1530         xfs_daddr_t     *blk_no)
1531 {
1532         xfs_buf_t       *bp;
1533         char            *offset;
1534         uint            first_cycle, last_cycle;
1535         xfs_daddr_t     new_blk, last_blk, start_blk;
1536         xfs_daddr_t     num_scan_bblks;
1537         int             error, log_bbnum = log->l_logBBsize;
1538 
1539         *blk_no = 0;
1540 
1541         /* check totally zeroed log */
1542         bp = xlog_get_bp(log, 1);
1543         if (!bp)
1544                 return -ENOMEM;
1545         error = xlog_bread(log, 0, 1, bp, &offset);
1546         if (error)
1547                 goto bp_err;
1548 
1549         first_cycle = xlog_get_cycle(offset);
1550         if (first_cycle == 0) {         /* completely zeroed log */
1551                 *blk_no = 0;
1552                 xlog_put_bp(bp);
1553                 return 1;
1554         }
1555 
1556         /* check partially zeroed log */
1557         error = xlog_bread(log, log_bbnum-1, 1, bp, &offset);
1558         if (error)
1559                 goto bp_err;
1560 
1561         last_cycle = xlog_get_cycle(offset);
1562         if (last_cycle != 0) {          /* log completely written to */
1563                 xlog_put_bp(bp);
1564                 return 0;
1565         } else if (first_cycle != 1) {
1566                 /*
1567                  * If the cycle of the last block is zero, the cycle of
1568                  * the first block must be 1. If it's not, maybe we're
1569                  * not looking at a log... Bail out.
1570                  */
1571                 xfs_warn(log->l_mp,
1572                         "Log inconsistent or not a log (last==0, first!=1)");
1573                 error = -EINVAL;
1574                 goto bp_err;
1575         }
1576 
1577         /* we have a partially zeroed log */
1578         last_blk = log_bbnum-1;
1579         if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1580                 goto bp_err;
1581 
1582         /*
1583          * Validate the answer.  Because there is no way to guarantee that
1584          * the entire log is made up of log records which are the same size,
1585          * we scan over the defined maximum blocks.  At this point, the maximum
1586          * is not chosen to mean anything special.   XXXmiken
1587          */
1588         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1589         ASSERT(num_scan_bblks <= INT_MAX);
1590 
1591         if (last_blk < num_scan_bblks)
1592                 num_scan_bblks = last_blk;
1593         start_blk = last_blk - num_scan_bblks;
1594 
1595         /*
1596          * We search for any instances of cycle number 0 that occur before
1597          * our current estimate of the head.  What we're trying to detect is
1598          *        1 ... | 0 | 1 | 0...
1599          *                       ^ binary search ends here
1600          */
1601         if ((error = xlog_find_verify_cycle(log, start_blk,
1602                                          (int)num_scan_bblks, 0, &new_blk)))
1603                 goto bp_err;
1604         if (new_blk != -1)
1605                 last_blk = new_blk;
1606 
1607         /*
1608          * Potentially backup over partial log record write.  We don't need
1609          * to search the end of the log because we know it is zero.
1610          */
1611         error = xlog_find_verify_log_record(log, start_blk, &last_blk, 0);
1612         if (error == 1)
1613                 error = -EIO;
1614         if (error)
1615                 goto bp_err;
1616 
1617         *blk_no = last_blk;
1618 bp_err:
1619         xlog_put_bp(bp);
1620         if (error)
1621                 return error;
1622         return 1;
1623 }
1624 
1625 /*
1626  * These are simple subroutines used by xlog_clear_stale_blocks() below
1627  * to initialize a buffer full of empty log record headers and write
1628  * them into the log.
1629  */
1630 STATIC void
1631 xlog_add_record(
1632         struct xlog             *log,
1633         char                    *buf,
1634         int                     cycle,
1635         int                     block,
1636         int                     tail_cycle,
1637         int                     tail_block)
1638 {
1639         xlog_rec_header_t       *recp = (xlog_rec_header_t *)buf;
1640 
1641         memset(buf, 0, BBSIZE);
1642         recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1643         recp->h_cycle = cpu_to_be32(cycle);
1644         recp->h_version = cpu_to_be32(
1645                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1646         recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1647         recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1648         recp->h_fmt = cpu_to_be32(XLOG_FMT);
1649         memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1650 }
1651 
1652 STATIC int
1653 xlog_write_log_records(
1654         struct xlog     *log,
1655         int             cycle,
1656         int             start_block,
1657         int             blocks,
1658         int             tail_cycle,
1659         int             tail_block)
1660 {
1661         char            *offset;
1662         xfs_buf_t       *bp;
1663         int             balign, ealign;
1664         int             sectbb = log->l_sectBBsize;
1665         int             end_block = start_block + blocks;
1666         int             bufblks;
1667         int             error = 0;
1668         int             i, j = 0;
1669 
1670         /*
1671          * Greedily allocate a buffer big enough to handle the full
1672          * range of basic blocks to be written.  If that fails, try
1673          * a smaller size.  We need to be able to write at least a
1674          * log sector, or we're out of luck.
1675          */
1676         bufblks = 1 << ffs(blocks);
1677         while (bufblks > log->l_logBBsize)
1678                 bufblks >>= 1;
1679         while (!(bp = xlog_get_bp(log, bufblks))) {
1680                 bufblks >>= 1;
1681                 if (bufblks < sectbb)
1682                         return -ENOMEM;
1683         }
1684 
1685         /* We may need to do a read at the start to fill in part of
1686          * the buffer in the starting sector not covered by the first
1687          * write below.
1688          */
1689         balign = round_down(start_block, sectbb);
1690         if (balign != start_block) {
1691                 error = xlog_bread_noalign(log, start_block, 1, bp);
1692                 if (error)
1693                         goto out_put_bp;
1694 
1695                 j = start_block - balign;
1696         }
1697 
1698         for (i = start_block; i < end_block; i += bufblks) {
1699                 int             bcount, endcount;
1700 
1701                 bcount = min(bufblks, end_block - start_block);
1702                 endcount = bcount - j;
1703 
1704                 /* We may need to do a read at the end to fill in part of
1705                  * the buffer in the final sector not covered by the write.
1706                  * If this is the same sector as the above read, skip it.
1707                  */
1708                 ealign = round_down(end_block, sectbb);
1709                 if (j == 0 && (start_block + endcount > ealign)) {
1710                         offset = bp->b_addr + BBTOB(ealign - start_block);
1711                         error = xlog_bread_offset(log, ealign, sectbb,
1712                                                         bp, offset);
1713                         if (error)
1714                                 break;
1715 
1716                 }
1717 
1718                 offset = xlog_align(log, start_block, endcount, bp);
1719                 for (; j < endcount; j++) {
1720                         xlog_add_record(log, offset, cycle, i+j,
1721                                         tail_cycle, tail_block);
1722                         offset += BBSIZE;
1723                 }
1724                 error = xlog_bwrite(log, start_block, endcount, bp);
1725                 if (error)
1726                         break;
1727                 start_block += endcount;
1728                 j = 0;
1729         }
1730 
1731  out_put_bp:
1732         xlog_put_bp(bp);
1733         return error;
1734 }
1735 
1736 /*
1737  * This routine is called to blow away any incomplete log writes out
1738  * in front of the log head.  We do this so that we won't become confused
1739  * if we come up, write only a little bit more, and then crash again.
1740  * If we leave the partial log records out there, this situation could
1741  * cause us to think those partial writes are valid blocks since they
1742  * have the current cycle number.  We get rid of them by overwriting them
1743  * with empty log records with the old cycle number rather than the
1744  * current one.
1745  *
1746  * The tail lsn is passed in rather than taken from
1747  * the log so that we will not write over the unmount record after a
1748  * clean unmount in a 512 block log.  Doing so would leave the log without
1749  * any valid log records in it until a new one was written.  If we crashed
1750  * during that time we would not be able to recover.
1751  */
1752 STATIC int
1753 xlog_clear_stale_blocks(
1754         struct xlog     *log,
1755         xfs_lsn_t       tail_lsn)
1756 {
1757         int             tail_cycle, head_cycle;
1758         int             tail_block, head_block;
1759         int             tail_distance, max_distance;
1760         int             distance;
1761         int             error;
1762 
1763         tail_cycle = CYCLE_LSN(tail_lsn);
1764         tail_block = BLOCK_LSN(tail_lsn);
1765         head_cycle = log->l_curr_cycle;
1766         head_block = log->l_curr_block;
1767 
1768         /*
1769          * Figure out the distance between the new head of the log
1770          * and the tail.  We want to write over any blocks beyond the
1771          * head that we may have written just before the crash, but
1772          * we don't want to overwrite the tail of the log.
1773          */
1774         if (head_cycle == tail_cycle) {
1775                 /*
1776                  * The tail is behind the head in the physical log,
1777                  * so the distance from the head to the tail is the
1778                  * distance from the head to the end of the log plus
1779                  * the distance from the beginning of the log to the
1780                  * tail.
1781                  */
1782                 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1783                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1784                                          XFS_ERRLEVEL_LOW, log->l_mp);
1785                         return -EFSCORRUPTED;
1786                 }
1787                 tail_distance = tail_block + (log->l_logBBsize - head_block);
1788         } else {
1789                 /*
1790                  * The head is behind the tail in the physical log,
1791                  * so the distance from the head to the tail is just
1792                  * the tail block minus the head block.
1793                  */
1794                 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1795                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1796                                          XFS_ERRLEVEL_LOW, log->l_mp);
1797                         return -EFSCORRUPTED;
1798                 }
1799                 tail_distance = tail_block - head_block;
1800         }
1801 
1802         /*
1803          * If the head is right up against the tail, we can't clear
1804          * anything.
1805          */
1806         if (tail_distance <= 0) {
1807                 ASSERT(tail_distance == 0);
1808                 return 0;
1809         }
1810 
1811         max_distance = XLOG_TOTAL_REC_SHIFT(log);
1812         /*
1813          * Take the smaller of the maximum amount of outstanding I/O
1814          * we could have and the distance to the tail to clear out.
1815          * We take the smaller so that we don't overwrite the tail and
1816          * we don't waste all day writing from the head to the tail
1817          * for no reason.
1818          */
1819         max_distance = MIN(max_distance, tail_distance);
1820 
1821         if ((head_block + max_distance) <= log->l_logBBsize) {
1822                 /*
1823                  * We can stomp all the blocks we need to without
1824                  * wrapping around the end of the log.  Just do it
1825                  * in a single write.  Use the cycle number of the
1826                  * current cycle minus one so that the log will look like:
1827                  *     n ... | n - 1 ...
1828                  */
1829                 error = xlog_write_log_records(log, (head_cycle - 1),
1830                                 head_block, max_distance, tail_cycle,
1831                                 tail_block);
1832                 if (error)
1833                         return error;
1834         } else {
1835                 /*
1836                  * We need to wrap around the end of the physical log in
1837                  * order to clear all the blocks.  Do it in two separate
1838                  * I/Os.  The first write should be from the head to the
1839                  * end of the physical log, and it should use the current
1840                  * cycle number minus one just like above.
1841                  */
1842                 distance = log->l_logBBsize - head_block;
1843                 error = xlog_write_log_records(log, (head_cycle - 1),
1844                                 head_block, distance, tail_cycle,
1845                                 tail_block);
1846 
1847                 if (error)
1848                         return error;
1849 
1850                 /*
1851                  * Now write the blocks at the start of the physical log.
1852                  * This writes the remainder of the blocks we want to clear.
1853                  * It uses the current cycle number since we're now on the
1854                  * same cycle as the head so that we get:
1855                  *    n ... n ... | n - 1 ...
1856                  *    ^^^^^ blocks we're writing
1857                  */
1858                 distance = max_distance - (log->l_logBBsize - head_block);
1859                 error = xlog_write_log_records(log, head_cycle, 0, distance,
1860                                 tail_cycle, tail_block);
1861                 if (error)
1862                         return error;
1863         }
1864 
1865         return 0;
1866 }
1867 
1868 /******************************************************************************
1869  *
1870  *              Log recover routines
1871  *
1872  ******************************************************************************
1873  */
1874 
1875 /*
1876  * Sort the log items in the transaction.
1877  *
1878  * The ordering constraints are defined by the inode allocation and unlink
1879  * behaviour. The rules are:
1880  *
1881  *      1. Every item is only logged once in a given transaction. Hence it
1882  *         represents the last logged state of the item. Hence ordering is
1883  *         dependent on the order in which operations need to be performed so
1884  *         required initial conditions are always met.
1885  *
1886  *      2. Cancelled buffers are recorded in pass 1 in a separate table and
1887  *         there's nothing to replay from them so we can simply cull them
1888  *         from the transaction. However, we can't do that until after we've
1889  *         replayed all the other items because they may be dependent on the
1890  *         cancelled buffer and replaying the cancelled buffer can remove it
1891  *         form the cancelled buffer table. Hence they have tobe done last.
1892  *
1893  *      3. Inode allocation buffers must be replayed before inode items that
1894  *         read the buffer and replay changes into it. For filesystems using the
1895  *         ICREATE transactions, this means XFS_LI_ICREATE objects need to get
1896  *         treated the same as inode allocation buffers as they create and
1897  *         initialise the buffers directly.
1898  *
1899  *      4. Inode unlink buffers must be replayed after inode items are replayed.
1900  *         This ensures that inodes are completely flushed to the inode buffer
1901  *         in a "free" state before we remove the unlinked inode list pointer.
1902  *
1903  * Hence the ordering needs to be inode allocation buffers first, inode items
1904  * second, inode unlink buffers third and cancelled buffers last.
1905  *
1906  * But there's a problem with that - we can't tell an inode allocation buffer
1907  * apart from a regular buffer, so we can't separate them. We can, however,
1908  * tell an inode unlink buffer from the others, and so we can separate them out
1909  * from all the other buffers and move them to last.
1910  *
1911  * Hence, 4 lists, in order from head to tail:
1912  *      - buffer_list for all buffers except cancelled/inode unlink buffers
1913  *      - item_list for all non-buffer items
1914  *      - inode_buffer_list for inode unlink buffers
1915  *      - cancel_list for the cancelled buffers
1916  *
1917  * Note that we add objects to the tail of the lists so that first-to-last
1918  * ordering is preserved within the lists. Adding objects to the head of the
1919  * list means when we traverse from the head we walk them in last-to-first
1920  * order. For cancelled buffers and inode unlink buffers this doesn't matter,
1921  * but for all other items there may be specific ordering that we need to
1922  * preserve.
1923  */
1924 STATIC int
1925 xlog_recover_reorder_trans(
1926         struct xlog             *log,
1927         struct xlog_recover     *trans,
1928         int                     pass)
1929 {
1930         xlog_recover_item_t     *item, *n;
1931         int                     error = 0;
1932         LIST_HEAD(sort_list);
1933         LIST_HEAD(cancel_list);
1934         LIST_HEAD(buffer_list);
1935         LIST_HEAD(inode_buffer_list);
1936         LIST_HEAD(inode_list);
1937 
1938         list_splice_init(&trans->r_itemq, &sort_list);
1939         list_for_each_entry_safe(item, n, &sort_list, ri_list) {
1940                 xfs_buf_log_format_t    *buf_f = item->ri_buf[0].i_addr;
1941 
1942                 switch (ITEM_TYPE(item)) {
1943                 case XFS_LI_ICREATE:
1944                         list_move_tail(&item->ri_list, &buffer_list);
1945                         break;
1946                 case XFS_LI_BUF:
1947                         if (buf_f->blf_flags & XFS_BLF_CANCEL) {
1948                                 trace_xfs_log_recover_item_reorder_head(log,
1949                                                         trans, item, pass);
1950                                 list_move(&item->ri_list, &cancel_list);
1951                                 break;
1952                         }
1953                         if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
1954                                 list_move(&item->ri_list, &inode_buffer_list);
1955                                 break;
1956                         }
1957                         list_move_tail(&item->ri_list, &buffer_list);
1958                         break;
1959                 case XFS_LI_INODE:
1960                 case XFS_LI_DQUOT:
1961                 case XFS_LI_QUOTAOFF:
1962                 case XFS_LI_EFD:
1963                 case XFS_LI_EFI:
1964                 case XFS_LI_RUI:
1965                 case XFS_LI_RUD:
1966                 case XFS_LI_CUI:
1967                 case XFS_LI_CUD:
1968                 case XFS_LI_BUI:
1969                 case XFS_LI_BUD:
1970                         trace_xfs_log_recover_item_reorder_tail(log,
1971                                                         trans, item, pass);
1972                         list_move_tail(&item->ri_list, &inode_list);
1973                         break;
1974                 default:
1975                         xfs_warn(log->l_mp,
1976                                 "%s: unrecognized type of log operation",
1977                                 __func__);
1978                         ASSERT(0);
1979                         /*
1980                          * return the remaining items back to the transaction
1981                          * item list so they can be freed in caller.
1982                          */
1983                         if (!list_empty(&sort_list))
1984                                 list_splice_init(&sort_list, &trans->r_itemq);
1985                         error = -EIO;
1986                         goto out;
1987                 }
1988         }
1989 out:
1990         ASSERT(list_empty(&sort_list));
1991         if (!list_empty(&buffer_list))
1992                 list_splice(&buffer_list, &trans->r_itemq);
1993         if (!list_empty(&inode_list))
1994                 list_splice_tail(&inode_list, &trans->r_itemq);
1995         if (!list_empty(&inode_buffer_list))
1996                 list_splice_tail(&inode_buffer_list, &trans->r_itemq);
1997         if (!list_empty(&cancel_list))
1998                 list_splice_tail(&cancel_list, &trans->r_itemq);
1999         return error;
2000 }
2001 
2002 /*
2003  * Build up the table of buf cancel records so that we don't replay
2004  * cancelled data in the second pass.  For buffer records that are
2005  * not cancel records, there is nothing to do here so we just return.
2006  *
2007  * If we get a cancel record which is already in the table, this indicates
2008  * that the buffer was cancelled multiple times.  In order to ensure
2009  * that during pass 2 we keep the record in the table until we reach its
2010  * last occurrence in the log, we keep a reference count in the cancel
2011  * record in the table to tell us how many times we expect to see this
2012  * record during the second pass.
2013  */
2014 STATIC int
2015 xlog_recover_buffer_pass1(
2016         struct xlog                     *log,
2017         struct xlog_recover_item        *item)
2018 {
2019         xfs_buf_log_format_t    *buf_f = item->ri_buf[0].i_addr;
2020         struct list_head        *bucket;
2021         struct xfs_buf_cancel   *bcp;
2022 
2023         /*
2024          * If this isn't a cancel buffer item, then just return.
2025          */
2026         if (!(buf_f->blf_flags & XFS_BLF_CANCEL)) {
2027                 trace_xfs_log_recover_buf_not_cancel(log, buf_f);
2028                 return 0;
2029         }
2030 
2031         /*
2032          * Insert an xfs_buf_cancel record into the hash table of them.
2033          * If there is already an identical record, bump its reference count.
2034          */
2035         bucket = XLOG_BUF_CANCEL_BUCKET(log, buf_f->blf_blkno);
2036         list_for_each_entry(bcp, bucket, bc_list) {
2037                 if (bcp->bc_blkno == buf_f->blf_blkno &&
2038                     bcp->bc_len == buf_f->blf_len) {
2039                         bcp->bc_refcount++;
2040                         trace_xfs_log_recover_buf_cancel_ref_inc(log, buf_f);
2041                         return 0;
2042                 }
2043         }
2044 
2045         bcp = kmem_alloc(sizeof(struct xfs_buf_cancel), KM_SLEEP);
2046         bcp->bc_blkno = buf_f->blf_blkno;
2047         bcp->bc_len = buf_f->blf_len;
2048         bcp->bc_refcount = 1;
2049         list_add_tail(&bcp->bc_list, bucket);
2050 
2051         trace_xfs_log_recover_buf_cancel_add(log, buf_f);
2052         return 0;
2053 }
2054 
2055 /*
2056  * Check to see whether the buffer being recovered has a corresponding
2057  * entry in the buffer cancel record table. If it is, return the cancel
2058  * buffer structure to the caller.
2059  */
2060 STATIC struct xfs_buf_cancel *
2061 xlog_peek_buffer_cancelled(
2062         struct xlog             *log,
2063         xfs_daddr_t             blkno,
2064         uint                    len,
2065         unsigned short                  flags)
2066 {
2067         struct list_head        *bucket;
2068         struct xfs_buf_cancel   *bcp;
2069 
2070         if (!log->l_buf_cancel_table) {
2071                 /* empty table means no cancelled buffers in the log */
2072                 ASSERT(!(flags & XFS_BLF_CANCEL));
2073                 return NULL;
2074         }
2075 
2076         bucket = XLOG_BUF_CANCEL_BUCKET(log, blkno);
2077         list_for_each_entry(bcp, bucket, bc_list) {
2078                 if (bcp->bc_blkno == blkno && bcp->bc_len == len)
2079                         return bcp;
2080         }
2081 
2082         /*
2083          * We didn't find a corresponding entry in the table, so return 0 so
2084          * that the buffer is NOT cancelled.
2085          */
2086         ASSERT(!(flags & XFS_BLF_CANCEL));
2087         return NULL;
2088 }
2089 
2090 /*
2091  * If the buffer is being cancelled then return 1 so that it will be cancelled,
2092  * otherwise return 0.  If the buffer is actually a buffer cancel item
2093  * (XFS_BLF_CANCEL is set), then decrement the refcount on the entry in the
2094  * table and remove it from the table if this is the last reference.
2095  *
2096  * We remove the cancel record from the table when we encounter its last
2097  * occurrence in the log so that if the same buffer is re-used again after its
2098  * last cancellation we actually replay the changes made at that point.
2099  */
2100 STATIC int
2101 xlog_check_buffer_cancelled(
2102         struct xlog             *log,
2103         xfs_daddr_t             blkno,
2104         uint                    len,
2105         unsigned short                  flags)
2106 {
2107         struct xfs_buf_cancel   *bcp;
2108 
2109         bcp = xlog_peek_buffer_cancelled(log, blkno, len, flags);
2110         if (!bcp)
2111                 return 0;
2112 
2113         /*
2114          * We've go a match, so return 1 so that the recovery of this buffer
2115          * is cancelled.  If this buffer is actually a buffer cancel log
2116          * item, then decrement the refcount on the one in the table and
2117          * remove it if this is the last reference.
2118          */
2119         if (flags & XFS_BLF_CANCEL) {
2120                 if (--bcp->bc_refcount == 0) {
2121                         list_del(&bcp->bc_list);
2122                         kmem_free(bcp);
2123                 }
2124         }
2125         return 1;
2126 }
2127 
2128 /*
2129  * Perform recovery for a buffer full of inodes.  In these buffers, the only
2130  * data which should be recovered is that which corresponds to the
2131  * di_next_unlinked pointers in the on disk inode structures.  The rest of the
2132  * data for the inodes is always logged through the inodes themselves rather
2133  * than the inode buffer and is recovered in xlog_recover_inode_pass2().
2134  *
2135  * The only time when buffers full of inodes are fully recovered is when the
2136  * buffer is full of newly allocated inodes.  In this case the buffer will
2137  * not be marked as an inode buffer and so will be sent to
2138  * xlog_recover_do_reg_buffer() below during recovery.
2139  */
2140 STATIC int
2141 xlog_recover_do_inode_buffer(
2142         struct xfs_mount        *mp,
2143         xlog_recover_item_t     *item,
2144         struct xfs_buf          *bp,
2145         xfs_buf_log_format_t    *buf_f)
2146 {
2147         int                     i;
2148         int                     item_index = 0;
2149         int                     bit = 0;
2150         int                     nbits = 0;
2151         int                     reg_buf_offset = 0;
2152         int                     reg_buf_bytes = 0;
2153         int                     next_unlinked_offset;
2154         int                     inodes_per_buf;
2155         xfs_agino_t             *logged_nextp;
2156         xfs_agino_t             *buffer_nextp;
2157 
2158         trace_xfs_log_recover_buf_inode_buf(mp->m_log, buf_f);
2159 
2160         /*
2161          * Post recovery validation only works properly on CRC enabled
2162          * filesystems.
2163          */
2164         if (xfs_sb_version_hascrc(&mp->m_sb))
2165                 bp->b_ops = &xfs_inode_buf_ops;
2166 
2167         inodes_per_buf = BBTOB(bp->b_io_length) >> mp->m_sb.sb_inodelog;
2168         for (i = 0; i < inodes_per_buf; i++) {
2169                 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
2170                         offsetof(xfs_dinode_t, di_next_unlinked);
2171 
2172                 while (next_unlinked_offset >=
2173                        (reg_buf_offset + reg_buf_bytes)) {
2174                         /*
2175                          * The next di_next_unlinked field is beyond
2176                          * the current logged region.  Find the next
2177                          * logged region that contains or is beyond
2178                          * the current di_next_unlinked field.
2179                          */
2180                         bit += nbits;
2181                         bit = xfs_next_bit(buf_f->blf_data_map,
2182                                            buf_f->blf_map_size, bit);
2183 
2184                         /*
2185                          * If there are no more logged regions in the
2186                          * buffer, then we're done.
2187                          */
2188                         if (bit == -1)
2189                                 return 0;
2190 
2191                         nbits = xfs_contig_bits(buf_f->blf_data_map,
2192                                                 buf_f->blf_map_size, bit);
2193                         ASSERT(nbits > 0);
2194                         reg_buf_offset = bit << XFS_BLF_SHIFT;
2195                         reg_buf_bytes = nbits << XFS_BLF_SHIFT;
2196                         item_index++;
2197                 }
2198 
2199                 /*
2200                  * If the current logged region starts after the current
2201                  * di_next_unlinked field, then move on to the next
2202                  * di_next_unlinked field.
2203                  */
2204                 if (next_unlinked_offset < reg_buf_offset)
2205                         continue;
2206 
2207                 ASSERT(item->ri_buf[item_index].i_addr != NULL);
2208                 ASSERT((item->ri_buf[item_index].i_len % XFS_BLF_CHUNK) == 0);
2209                 ASSERT((reg_buf_offset + reg_buf_bytes) <=
2210                                                         BBTOB(bp->b_io_length));
2211 
2212                 /*
2213                  * The current logged region contains a copy of the
2214                  * current di_next_unlinked field.  Extract its value
2215                  * and copy it to the buffer copy.
2216                  */
2217                 logged_nextp = item->ri_buf[item_index].i_addr +
2218                                 next_unlinked_offset - reg_buf_offset;
2219                 if (unlikely(*logged_nextp == 0)) {
2220                         xfs_alert(mp,
2221                 "Bad inode buffer log record (ptr = "PTR_FMT", bp = "PTR_FMT"). "
2222                 "Trying to replay bad (0) inode di_next_unlinked field.",
2223                                 item, bp);
2224                         XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
2225                                          XFS_ERRLEVEL_LOW, mp);
2226                         return -EFSCORRUPTED;
2227                 }
2228 
2229                 buffer_nextp = xfs_buf_offset(bp, next_unlinked_offset);
2230                 *buffer_nextp = *logged_nextp;
2231 
2232                 /*
2233                  * If necessary, recalculate the CRC in the on-disk inode. We
2234                  * have to leave the inode in a consistent state for whoever
2235                  * reads it next....
2236                  */
2237                 xfs_dinode_calc_crc(mp,
2238                                 xfs_buf_offset(bp, i * mp->m_sb.sb_inodesize));
2239 
2240         }
2241 
2242         return 0;
2243 }
2244 
2245 /*
2246  * V5 filesystems know the age of the buffer on disk being recovered. We can
2247  * have newer objects on disk than we are replaying, and so for these cases we
2248  * don't want to replay the current change as that will make the buffer contents
2249  * temporarily invalid on disk.
2250  *
2251  * The magic number might not match the buffer type we are going to recover
2252  * (e.g. reallocated blocks), so we ignore the xfs_buf_log_format flags.  Hence
2253  * extract the LSN of the existing object in the buffer based on it's current
2254  * magic number.  If we don't recognise the magic number in the buffer, then
2255  * return a LSN of -1 so that the caller knows it was an unrecognised block and
2256  * so can recover the buffer.
2257  *
2258  * Note: we cannot rely solely on magic number matches to determine that the
2259  * buffer has a valid LSN - we also need to verify that it belongs to this
2260  * filesystem, so we need to extract the object's LSN and compare it to that
2261  * which we read from the superblock. If the UUIDs don't match, then we've got a
2262  * stale metadata block from an old filesystem instance that we need to recover
2263  * over the top of.
2264  */
2265 static xfs_lsn_t
2266 xlog_recover_get_buf_lsn(
2267         struct xfs_mount        *mp,
2268         struct xfs_buf          *bp)
2269 {
2270         uint32_t                magic32;
2271         uint16_t                magic16;
2272         uint16_t                magicda;
2273         void                    *blk = bp->b_addr;
2274         uuid_t                  *uuid;
2275         xfs_lsn_t               lsn = -1;
2276 
2277         /* v4 filesystems always recover immediately */
2278         if (!xfs_sb_version_hascrc(&mp->m_sb))
2279                 goto recover_immediately;
2280 
2281         magic32 = be32_to_cpu(*(__be32 *)blk);
2282         switch (magic32) {
2283         case XFS_ABTB_CRC_MAGIC:
2284         case XFS_ABTC_CRC_MAGIC:
2285         case XFS_ABTB_MAGIC:
2286         case XFS_ABTC_MAGIC:
2287         case XFS_RMAP_CRC_MAGIC:
2288         case XFS_REFC_CRC_MAGIC:
2289         case XFS_IBT_CRC_MAGIC:
2290         case XFS_IBT_MAGIC: {
2291                 struct xfs_btree_block *btb = blk;
2292 
2293                 lsn = be64_to_cpu(btb->bb_u.s.bb_lsn);
2294                 uuid = &btb->bb_u.s.bb_uuid;
2295                 break;
2296         }
2297         case XFS_BMAP_CRC_MAGIC:
2298         case XFS_BMAP_MAGIC: {
2299                 struct xfs_btree_block *btb = blk;
2300 
2301                 lsn = be64_to_cpu(btb->bb_u.l.bb_lsn);
2302                 uuid = &btb->bb_u.l.bb_uuid;
2303                 break;
2304         }
2305         case XFS_AGF_MAGIC:
2306                 lsn = be64_to_cpu(((struct xfs_agf *)blk)->agf_lsn);
2307                 uuid = &((struct xfs_agf *)blk)->agf_uuid;
2308                 break;
2309         case XFS_AGFL_MAGIC:
2310                 lsn = be64_to_cpu(((struct xfs_agfl *)blk)->agfl_lsn);
2311                 uuid = &((struct xfs_agfl *)blk)->agfl_uuid;
2312                 break;
2313         case XFS_AGI_MAGIC:
2314                 lsn = be64_to_cpu(((struct xfs_agi *)blk)->agi_lsn);
2315                 uuid = &((struct xfs_agi *)blk)->agi_uuid;
2316                 break;
2317         case XFS_SYMLINK_MAGIC:
2318                 lsn = be64_to_cpu(((struct xfs_dsymlink_hdr *)blk)->sl_lsn);
2319                 uuid = &((struct xfs_dsymlink_hdr *)blk)->sl_uuid;
2320                 break;
2321         case XFS_DIR3_BLOCK_MAGIC:
2322         case XFS_DIR3_DATA_MAGIC:
2323         case XFS_DIR3_FREE_MAGIC:
2324                 lsn = be64_to_cpu(((struct xfs_dir3_blk_hdr *)blk)->lsn);
2325                 uuid = &((struct xfs_dir3_blk_hdr *)blk)->uuid;
2326                 break;
2327         case XFS_ATTR3_RMT_MAGIC:
2328                 /*
2329                  * Remote attr blocks are written synchronously, rather than
2330                  * being logged. That means they do not contain a valid LSN
2331                  * (i.e. transactionally ordered) in them, and hence any time we
2332                  * see a buffer to replay over the top of a remote attribute
2333                  * block we should simply do so.
2334                  */
2335                 goto recover_immediately;
2336         case XFS_SB_MAGIC:
2337                 /*
2338                  * superblock uuids are magic. We may or may not have a
2339                  * sb_meta_uuid on disk, but it will be set in the in-core
2340                  * superblock. We set the uuid pointer for verification
2341                  * according to the superblock feature mask to ensure we check
2342                  * the relevant UUID in the superblock.
2343                  */
2344                 lsn = be64_to_cpu(((struct xfs_dsb *)blk)->sb_lsn);
2345                 if (xfs_sb_version_hasmetauuid(&mp->m_sb))
2346                         uuid = &((struct xfs_dsb *)blk)->sb_meta_uuid;
2347                 else
2348                         uuid = &((struct xfs_dsb *)blk)->sb_uuid;
2349                 break;
2350         default:
2351                 break;
2352         }
2353 
2354         if (lsn != (xfs_lsn_t)-1) {
2355                 if (!uuid_equal(&mp->m_sb.sb_meta_uuid, uuid))
2356                         goto recover_immediately;
2357                 return lsn;
2358         }
2359 
2360         magicda = be16_to_cpu(((struct xfs_da_blkinfo *)blk)->magic);
2361         switch (magicda) {
2362         case XFS_DIR3_LEAF1_MAGIC:
2363         case XFS_DIR3_LEAFN_MAGIC:
2364         case XFS_DA3_NODE_MAGIC:
2365                 lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
2366                 uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
2367                 break;
2368         default:
2369                 break;
2370         }
2371 
2372         if (lsn != (xfs_lsn_t)-1) {
2373                 if (!uuid_equal(&mp->m_sb.sb_uuid, uuid))
2374                         goto recover_immediately;
2375                 return lsn;
2376         }
2377 
2378         /*
2379          * We do individual object checks on dquot and inode buffers as they
2380          * have their own individual LSN records. Also, we could have a stale
2381          * buffer here, so we have to at least recognise these buffer types.
2382          *
2383          * A notd complexity here is inode unlinked list processing - it logs
2384          * the inode directly in the buffer, but we don't know which inodes have
2385          * been modified, and there is no global buffer LSN. Hence we need to
2386          * recover all inode buffer types immediately. This problem will be
2387          * fixed by logical logging of the unlinked list modifications.
2388          */
2389         magic16 = be16_to_cpu(*(__be16 *)blk);
2390         switch (magic16) {
2391         case XFS_DQUOT_MAGIC:
2392         case XFS_DINODE_MAGIC:
2393                 goto recover_immediately;
2394         default:
2395                 break;
2396         }
2397 
2398         /* unknown buffer contents, recover immediately */
2399 
2400 recover_immediately:
2401         return (xfs_lsn_t)-1;
2402 
2403 }
2404 
2405 /*
2406  * Validate the recovered buffer is of the correct type and attach the
2407  * appropriate buffer operations to them for writeback. Magic numbers are in a
2408  * few places:
2409  *      the first 16 bits of the buffer (inode buffer, dquot buffer),
2410  *      the first 32 bits of the buffer (most blocks),
2411  *      inside a struct xfs_da_blkinfo at the start of the buffer.
2412  */
2413 static void
2414 xlog_recover_validate_buf_type(
2415         struct xfs_mount        *mp,
2416         struct xfs_buf          *bp,
2417         xfs_buf_log_format_t    *buf_f,
2418         xfs_lsn_t               current_lsn)
2419 {
2420         struct xfs_da_blkinfo   *info = bp->b_addr;
2421         uint32_t                magic32;
2422         uint16_t                magic16;
2423         uint16_t                magicda;
2424         char                    *warnmsg = NULL;
2425 
2426         /*
2427          * We can only do post recovery validation on items on CRC enabled
2428          * fielsystems as we need to know when the buffer was written to be able
2429          * to determine if we should have replayed the item. If we replay old
2430          * metadata over a newer buffer, then it will enter a temporarily
2431          * inconsistent state resulting in verification failures. Hence for now
2432          * just avoid the verification stage for non-crc filesystems
2433          */
2434         if (!xfs_sb_version_hascrc(&mp->m_sb))
2435                 return;
2436 
2437         magic32 = be32_to_cpu(*(__be32 *)bp->b_addr);
2438         magic16 = be16_to_cpu(*(__be16*)bp->b_addr);
2439         magicda = be16_to_cpu(info->magic);
2440         switch (xfs_blft_from_flags(buf_f)) {
2441         case XFS_BLFT_BTREE_BUF:
2442                 switch (magic32) {
2443                 case XFS_ABTB_CRC_MAGIC:
2444                 case XFS_ABTC_CRC_MAGIC:
2445                 case XFS_ABTB_MAGIC:
2446                 case XFS_ABTC_MAGIC:
2447                         bp->b_ops = &xfs_allocbt_buf_ops;
2448                         break;
2449                 case XFS_IBT_CRC_MAGIC:
2450                 case XFS_FIBT_CRC_MAGIC:
2451                 case XFS_IBT_MAGIC:
2452                 case XFS_FIBT_MAGIC:
2453                         bp->b_ops = &xfs_inobt_buf_ops;
2454                         break;
2455                 case XFS_BMAP_CRC_MAGIC:
2456                 case XFS_BMAP_MAGIC:
2457                         bp->b_ops = &xfs_bmbt_buf_ops;
2458                         break;
2459                 case XFS_RMAP_CRC_MAGIC:
2460                         bp->b_ops = &xfs_rmapbt_buf_ops;
2461                         break;
2462                 case XFS_REFC_CRC_MAGIC:
2463                         bp->b_ops = &xfs_refcountbt_buf_ops;
2464                         break;
2465                 default:
2466                         warnmsg = "Bad btree block magic!";
2467                         break;
2468                 }
2469                 break;
2470         case XFS_BLFT_AGF_BUF:
2471                 if (magic32 != XFS_AGF_MAGIC) {
2472                         warnmsg = "Bad AGF block magic!";
2473                         break;
2474                 }
2475                 bp->b_ops = &xfs_agf_buf_ops;
2476                 break;
2477         case XFS_BLFT_AGFL_BUF:
2478                 if (magic32 != XFS_AGFL_MAGIC) {
2479                         warnmsg = "Bad AGFL block magic!";
2480                         break;
2481                 }
2482                 bp->b_ops = &xfs_agfl_buf_ops;
2483                 break;
2484         case XFS_BLFT_AGI_BUF:
2485                 if (magic32 != XFS_AGI_MAGIC) {
2486                         warnmsg = "Bad AGI block magic!";
2487                         break;
2488                 }
2489                 bp->b_ops = &xfs_agi_buf_ops;
2490                 break;
2491         case XFS_BLFT_UDQUOT_BUF:
2492         case XFS_BLFT_PDQUOT_BUF:
2493         case XFS_BLFT_GDQUOT_BUF:
2494 #ifdef CONFIG_XFS_QUOTA
2495                 if (magic16 != XFS_DQUOT_MAGIC) {
2496                         warnmsg = "Bad DQUOT block magic!";
2497                         break;
2498                 }
2499                 bp->b_ops = &xfs_dquot_buf_ops;
2500 #else
2501                 xfs_alert(mp,
2502         "Trying to recover dquots without QUOTA support built in!");
2503                 ASSERT(0);
2504 #endif
2505                 break;
2506         case XFS_BLFT_DINO_BUF:
2507                 if (magic16 != XFS_DINODE_MAGIC) {
2508                         warnmsg = "Bad INODE block magic!";
2509                         break;
2510                 }
2511                 bp->b_ops = &xfs_inode_buf_ops;
2512                 break;
2513         case XFS_BLFT_SYMLINK_BUF:
2514                 if (magic32 != XFS_SYMLINK_MAGIC) {
2515                         warnmsg = "Bad symlink block magic!";
2516                         break;
2517                 }
2518                 bp->b_ops = &xfs_symlink_buf_ops;
2519                 break;
2520         case XFS_BLFT_DIR_BLOCK_BUF:
2521                 if (magic32 != XFS_DIR2_BLOCK_MAGIC &&
2522                     magic32 != XFS_DIR3_BLOCK_MAGIC) {
2523                         warnmsg = "Bad dir block magic!";
2524                         break;
2525                 }
2526                 bp->b_ops = &xfs_dir3_block_buf_ops;
2527                 break;
2528         case XFS_BLFT_DIR_DATA_BUF:
2529                 if (magic32 != XFS_DIR2_DATA_MAGIC &&
2530                     magic32 != XFS_DIR3_DATA_MAGIC) {
2531                         warnmsg = "Bad dir data magic!";
2532                         break;
2533                 }
2534                 bp->b_ops = &xfs_dir3_data_buf_ops;
2535                 break;
2536         case XFS_BLFT_DIR_FREE_BUF:
2537                 if (magic32 != XFS_DIR2_FREE_MAGIC &&
2538                     magic32 != XFS_DIR3_FREE_MAGIC) {
2539                         warnmsg = "Bad dir3 free magic!";
2540                         break;
2541                 }
2542                 bp->b_ops = &xfs_dir3_free_buf_ops;
2543                 break;
2544         case XFS_BLFT_DIR_LEAF1_BUF:
2545                 if (magicda != XFS_DIR2_LEAF1_MAGIC &&
2546                     magicda != XFS_DIR3_LEAF1_MAGIC) {
2547                         warnmsg = "Bad dir leaf1 magic!";
2548                         break;
2549                 }
2550                 bp->b_ops = &xfs_dir3_leaf1_buf_ops;
2551                 break;
2552         case XFS_BLFT_DIR_LEAFN_BUF:
2553                 if (magicda != XFS_DIR2_LEAFN_MAGIC &&
2554                     magicda != XFS_DIR3_LEAFN_MAGIC) {
2555                         warnmsg = "Bad dir leafn magic!";
2556                         break;
2557                 }
2558                 bp->b_ops = &xfs_dir3_leafn_buf_ops;
2559                 break;
2560         case XFS_BLFT_DA_NODE_BUF:
2561                 if (magicda != XFS_DA_NODE_MAGIC &&
2562                     magicda != XFS_DA3_NODE_MAGIC) {
2563                         warnmsg = "Bad da node magic!";
2564                         break;
2565                 }
2566                 bp->b_ops = &xfs_da3_node_buf_ops;
2567                 break;
2568         case XFS_BLFT_ATTR_LEAF_BUF:
2569                 if (magicda != XFS_ATTR_LEAF_MAGIC &&
2570                     magicda != XFS_ATTR3_LEAF_MAGIC) {
2571                         warnmsg = "Bad attr leaf magic!";
2572                         break;
2573                 }
2574                 bp->b_ops = &xfs_attr3_leaf_buf_ops;
2575                 break;
2576         case XFS_BLFT_ATTR_RMT_BUF:
2577                 if (magic32 != XFS_ATTR3_RMT_MAGIC) {
2578                         warnmsg = "Bad attr remote magic!";
2579                         break;
2580                 }
2581                 bp->b_ops = &xfs_attr3_rmt_buf_ops;
2582                 break;
2583         case XFS_BLFT_SB_BUF:
2584                 if (magic32 != XFS_SB_MAGIC) {
2585                         warnmsg = "Bad SB block magic!";
2586                         break;
2587                 }
2588                 bp->b_ops = &xfs_sb_buf_ops;
2589                 break;
2590 #ifdef CONFIG_XFS_RT
2591         case XFS_BLFT_RTBITMAP_BUF:
2592         case XFS_BLFT_RTSUMMARY_BUF:
2593                 /* no magic numbers for verification of RT buffers */
2594                 bp->b_ops = &xfs_rtbuf_ops;
2595                 break;
2596 #endif /* CONFIG_XFS_RT */
2597         default:
2598                 xfs_warn(mp, "Unknown buffer type %d!",
2599                          xfs_blft_from_flags(buf_f));
2600                 break;
2601         }
2602 
2603         /*
2604          * Nothing else to do in the case of a NULL current LSN as this means
2605          * the buffer is more recent than the change in the log and will be
2606          * skipped.
2607          */
2608         if (current_lsn == NULLCOMMITLSN)
2609                 return;
2610 
2611         if (warnmsg) {
2612                 xfs_warn(mp, warnmsg);
2613                 ASSERT(0);
2614         }
2615 
2616         /*
2617          * We must update the metadata LSN of the buffer as it is written out to
2618          * ensure that older transactions never replay over this one and corrupt
2619          * the buffer. This can occur if log recovery is interrupted at some
2620          * point after the current transaction completes, at which point a
2621          * subsequent mount starts recovery from the beginning.
2622          *
2623          * Write verifiers update the metadata LSN from log items attached to
2624          * the buffer. Therefore, initialize a bli purely to carry the LSN to
2625          * the verifier. We'll clean it up in our ->iodone() callback.
2626          */
2627         if (bp->b_ops) {
2628                 struct xfs_buf_log_item *bip;
2629 
2630                 ASSERT(!bp->b_iodone || bp->b_iodone == xlog_recover_iodone);
2631                 bp->b_iodone = xlog_recover_iodone;
2632                 xfs_buf_item_init(bp, mp);
2633                 bip = bp->b_log_item;
2634                 bip->bli_item.li_lsn = current_lsn;
2635         }
2636 }
2637 
2638 /*
2639  * Perform a 'normal' buffer recovery.  Each logged region of the
2640  * buffer should be copied over the corresponding region in the
2641  * given buffer.  The bitmap in the buf log format structure indicates
2642  * where to place the logged data.
2643  */
2644 STATIC void
2645 xlog_recover_do_reg_buffer(
2646         struct xfs_mount        *mp,
2647         xlog_recover_item_t     *item,
2648         struct xfs_buf          *bp,
2649         xfs_buf_log_format_t    *buf_f,
2650         xfs_lsn_t               current_lsn)
2651 {
2652         int                     i;
2653         int                     bit;
2654         int                     nbits;
2655         xfs_failaddr_t          fa;
2656 
2657         trace_xfs_log_recover_buf_reg_buf(mp->m_log, buf_f);
2658 
2659         bit = 0;
2660         i = 1;  /* 0 is the buf format structure */
2661         while (1) {
2662                 bit = xfs_next_bit(buf_f->blf_data_map,
2663                                    buf_f->blf_map_size, bit);
2664                 if (bit == -1)
2665                         break;
2666                 nbits = xfs_contig_bits(buf_f->blf_data_map,
2667                                         buf_f->blf_map_size, bit);
2668                 ASSERT(nbits > 0);
2669                 ASSERT(item->ri_buf[i].i_addr != NULL);
2670                 ASSERT(item->ri_buf[i].i_len % XFS_BLF_CHUNK == 0);
2671                 ASSERT(BBTOB(bp->b_io_length) >=
2672                        ((uint)bit << XFS_BLF_SHIFT) + (nbits << XFS_BLF_SHIFT));
2673 
2674                 /*
2675                  * The dirty regions logged in the buffer, even though
2676                  * contiguous, may span multiple chunks. This is because the
2677                  * dirty region may span a physical page boundary in a buffer
2678                  * and hence be split into two separate vectors for writing into
2679                  * the log. Hence we need to trim nbits back to the length of
2680                  * the current region being copied out of the log.
2681                  */
2682                 if (item->ri_buf[i].i_len < (nbits << XFS_BLF_SHIFT))
2683                         nbits = item->ri_buf[i].i_len >> XFS_BLF_SHIFT;
2684 
2685                 /*
2686                  * Do a sanity check if this is a dquot buffer. Just checking
2687                  * the first dquot in the buffer should do. XXXThis is
2688                  * probably a good thing to do for other buf types also.
2689                  */
2690                 fa = NULL;
2691                 if (buf_f->blf_flags &
2692                    (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2693                         if (item->ri_buf[i].i_addr == NULL) {
2694                                 xfs_alert(mp,
2695                                         "XFS: NULL dquot in %s.", __func__);
2696                                 goto next;
2697                         }
2698                         if (item->ri_buf[i].i_len < sizeof(xfs_disk_dquot_t)) {
2699                                 xfs_alert(mp,
2700                                         "XFS: dquot too small (%d) in %s.",
2701                                         item->ri_buf[i].i_len, __func__);
2702                                 goto next;
2703                         }
2704                         fa = xfs_dquot_verify(mp, item->ri_buf[i].i_addr,
2705                                                -1, 0, 0);
2706                         if (fa) {
2707                                 xfs_alert(mp,
2708         "dquot corrupt at %pS trying to replay into block 0x%llx",
2709                                         fa, bp->b_bn);
2710                                 goto next;
2711                         }
2712                 }
2713 
2714                 memcpy(xfs_buf_offset(bp,
2715                         (uint)bit << XFS_BLF_SHIFT),    /* dest */
2716                         item->ri_buf[i].i_addr,         /* source */
2717                         nbits<<XFS_BLF_SHIFT);          /* length */
2718  next:
2719                 i++;
2720                 bit += nbits;
2721         }
2722 
2723         /* Shouldn't be any more regions */
2724         ASSERT(i == item->ri_total);
2725 
2726         xlog_recover_validate_buf_type(mp, bp, buf_f, current_lsn);
2727 }
2728 
2729 /*
2730  * Perform a dquot buffer recovery.
2731  * Simple algorithm: if we have found a QUOTAOFF log item of the same type
2732  * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2733  * Else, treat it as a regular buffer and do recovery.
2734  *
2735  * Return false if the buffer was tossed and true if we recovered the buffer to
2736  * indicate to the caller if the buffer needs writing.
2737  */
2738 STATIC bool
2739 xlog_recover_do_dquot_buffer(
2740         struct xfs_mount                *mp,
2741         struct xlog                     *log,
2742         struct xlog_recover_item        *item,
2743         struct xfs_buf                  *bp,
2744         struct xfs_buf_log_format       *buf_f)
2745 {
2746         uint                    type;
2747 
2748         trace_xfs_log_recover_buf_dquot_buf(log, buf_f);
2749 
2750         /*
2751          * Filesystems are required to send in quota flags at mount time.
2752          */
2753         if (!mp->m_qflags)
2754                 return false;
2755 
2756         type = 0;
2757         if (buf_f->blf_flags & XFS_BLF_UDQUOT_BUF)
2758                 type |= XFS_DQ_USER;
2759         if (buf_f->blf_flags & XFS_BLF_PDQUOT_BUF)
2760                 type |= XFS_DQ_PROJ;
2761         if (buf_f->blf_flags & XFS_BLF_GDQUOT_BUF)
2762                 type |= XFS_DQ_GROUP;
2763         /*
2764          * This type of quotas was turned off, so ignore this buffer
2765          */
2766         if (log->l_quotaoffs_flag & type)
2767                 return false;
2768 
2769         xlog_recover_do_reg_buffer(mp, item, bp, buf_f, NULLCOMMITLSN);
2770         return true;
2771 }
2772 
2773 /*
2774  * This routine replays a modification made to a buffer at runtime.
2775  * There are actually two types of buffer, regular and inode, which
2776  * are handled differently.  Inode buffers are handled differently
2777  * in that we only recover a specific set of data from them, namely
2778  * the inode di_next_unlinked fields.  This is because all other inode
2779  * data is actually logged via inode records and any data we replay
2780  * here which overlaps that may be stale.
2781  *
2782  * When meta-data buffers are freed at run time we log a buffer item
2783  * with the XFS_BLF_CANCEL bit set to indicate that previous copies
2784  * of the buffer in the log should not be replayed at recovery time.
2785  * This is so that if the blocks covered by the buffer are reused for
2786  * file data before we crash we don't end up replaying old, freed
2787  * meta-data into a user's file.
2788  *
2789  * To handle the cancellation of buffer log items, we make two passes
2790  * over the log during recovery.  During the first we build a table of
2791  * those buffers which have been cancelled, and during the second we
2792  * only replay those buffers which do not have corresponding cancel
2793  * records in the table.  See xlog_recover_buffer_pass[1,2] above
2794  * for more details on the implementation of the table of cancel records.
2795  */
2796 STATIC int
2797 xlog_recover_buffer_pass2(
2798         struct xlog                     *log,
2799         struct list_head                *buffer_list,
2800         struct xlog_recover_item        *item,
2801         xfs_lsn_t                       current_lsn)
2802 {
2803         xfs_buf_log_format_t    *buf_f = item->ri_buf[0].i_addr;
2804         xfs_mount_t             *mp = log->l_mp;
2805         xfs_buf_t               *bp;
2806         int                     error;
2807         uint                    buf_flags;
2808         xfs_lsn_t               lsn;
2809 
2810         /*
2811          * In this pass we only want to recover all the buffers which have
2812          * not been cancelled and are not cancellation buffers themselves.
2813          */
2814         if (xlog_check_buffer_cancelled(log, buf_f->blf_blkno,
2815                         buf_f->blf_len, buf_f->blf_flags)) {
2816                 trace_xfs_log_recover_buf_cancel(log, buf_f);
2817                 return 0;
2818         }
2819 
2820         trace_xfs_log_recover_buf_recover(log, buf_f);
2821 
2822         buf_flags = 0;
2823         if (buf_f->blf_flags & XFS_BLF_INODE_BUF)
2824                 buf_flags |= XBF_UNMAPPED;
2825 
2826         bp = xfs_buf_read(mp->m_ddev_targp, buf_f->blf_blkno, buf_f->blf_len,
2827                           buf_flags, NULL);
2828         if (!bp)
2829                 return -ENOMEM;
2830         error = bp->b_error;
2831         if (error) {
2832                 xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#1)");
2833                 goto out_release;
2834         }
2835 
2836         /*
2837          * Recover the buffer only if we get an LSN from it and it's less than
2838          * the lsn of the transaction we are replaying.
2839          *
2840          * Note that we have to be extremely careful of readahead here.
2841          * Readahead does not attach verfiers to the buffers so if we don't
2842          * actually do any replay after readahead because of the LSN we found
2843          * in the buffer if more recent than that current transaction then we
2844          * need to attach the verifier directly. Failure to do so can lead to
2845          * future recovery actions (e.g. EFI and unlinked list recovery) can
2846          * operate on the buffers and they won't get the verifier attached. This
2847          * can lead to blocks on disk having the correct content but a stale
2848          * CRC.
2849          *
2850          * It is safe to assume these clean buffers are currently up to date.
2851          * If the buffer is dirtied by a later transaction being replayed, then
2852          * the verifier will be reset to match whatever recover turns that
2853          * buffer into.
2854          */
2855         lsn = xlog_recover_get_buf_lsn(mp, bp);
2856         if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
2857                 trace_xfs_log_recover_buf_skip(log, buf_f);
2858                 xlog_recover_validate_buf_type(mp, bp, buf_f, NULLCOMMITLSN);
2859                 goto out_release;
2860         }
2861 
2862         if (buf_f->blf_flags & XFS_BLF_INODE_BUF) {
2863                 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2864                 if (error)
2865                         goto out_release;
2866         } else if (buf_f->blf_flags &
2867                   (XFS_BLF_UDQUOT_BUF|XFS_BLF_PDQUOT_BUF|XFS_BLF_GDQUOT_BUF)) {
2868                 bool    dirty;
2869 
2870                 dirty = xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2871                 if (!dirty)
2872                         goto out_release;
2873         } else {
2874                 xlog_recover_do_reg_buffer(mp, item, bp, buf_f, current_lsn);
2875         }
2876 
2877         /*
2878          * Perform delayed write on the buffer.  Asynchronous writes will be
2879          * slower when taking into account all the buffers to be flushed.
2880          *
2881          * Also make sure that only inode buffers with good sizes stay in
2882          * the buffer cache.  The kernel moves inodes in buffers of 1 block
2883          * or mp->m_inode_cluster_size bytes, whichever is bigger.  The inode
2884          * buffers in the log can be a different size if the log was generated
2885          * by an older kernel using unclustered inode buffers or a newer kernel
2886          * running with a different inode cluster size.  Regardless, if the
2887          * the inode buffer size isn't MAX(blocksize, mp->m_inode_cluster_size)
2888          * for *our* value of mp->m_inode_cluster_size, then we need to keep
2889          * the buffer out of the buffer cache so that the buffer won't
2890          * overlap with future reads of those inodes.
2891          */
2892         if (XFS_DINODE_MAGIC ==
2893             be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2894             (BBTOB(bp->b_io_length) != MAX(log->l_mp->m_sb.sb_blocksize,
2895                         (uint32_t)log->l_mp->m_inode_cluster_size))) {
2896                 xfs_buf_stale(bp);
2897                 error = xfs_bwrite(bp);
2898         } else {
2899                 ASSERT(bp->b_target->bt_mount == mp);
2900                 bp->b_iodone = xlog_recover_iodone;
2901                 xfs_buf_delwri_queue(bp, buffer_list);
2902         }
2903 
2904 out_release:
2905         xfs_buf_relse(bp);
2906         return error;
2907 }
2908 
2909 /*
2910  * Inode fork owner changes
2911  *
2912  * If we have been told that we have to reparent the inode fork, it's because an
2913  * extent swap operation on a CRC enabled filesystem has been done and we are
2914  * replaying it. We need to walk the BMBT of the appropriate fork and change the
2915  * owners of it.
2916  *
2917  * The complexity here is that we don't have an inode context to work with, so
2918  * after we've replayed the inode we need to instantiate one.  This is where the
2919  * fun begins.
2920  *
2921  * We are in the middle of log recovery, so we can't run transactions. That
2922  * means we cannot use cache coherent inode instantiation via xfs_iget(), as
2923  * that will result in the corresponding iput() running the inode through
2924  * xfs_inactive(). If we've just replayed an inode core that changes the link
2925  * count to zero (i.e. it's been unlinked), then xfs_inactive() will run
2926  * transactions (bad!).
2927  *
2928  * So, to avoid this, we instantiate an inode directly from the inode core we've
2929  * just recovered. We have the buffer still locked, and all we really need to
2930  * instantiate is the inode core and the forks being modified. We can do this
2931  * manually, then run the inode btree owner change, and then tear down the
2932  * xfs_inode without having to run any transactions at all.
2933  *
2934  * Also, because we don't have a transaction context available here but need to
2935  * gather all the buffers we modify for writeback so we pass the buffer_list
2936  * instead for the operation to use.
2937  */
2938 
2939 STATIC int
2940 xfs_recover_inode_owner_change(
2941         struct xfs_mount        *mp,
2942         struct xfs_dinode       *dip,
2943         struct xfs_inode_log_format *in_f,
2944         struct list_head        *buffer_list)
2945 {
2946         struct xfs_inode        *ip;
2947         int                     error;
2948 
2949         ASSERT(in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER));
2950 
2951         ip = xfs_inode_alloc(mp, in_f->ilf_ino);
2952         if (!ip)
2953                 return -ENOMEM;
2954 
2955         /* instantiate the inode */
2956         xfs_inode_from_disk(ip, dip);
2957         ASSERT(ip->i_d.di_version >= 3);
2958 
2959         error = xfs_iformat_fork(ip, dip);
2960         if (error)
2961                 goto out_free_ip;
2962 
2963         if (!xfs_inode_verify_forks(ip)) {
2964                 error = -EFSCORRUPTED;
2965                 goto out_free_ip;
2966         }
2967 
2968         if (in_f->ilf_fields & XFS_ILOG_DOWNER) {
2969                 ASSERT(in_f->ilf_fields & XFS_ILOG_DBROOT);
2970                 error = xfs_bmbt_change_owner(NULL, ip, XFS_DATA_FORK,
2971                                               ip->i_ino, buffer_list);
2972                 if (error)
2973                         goto out_free_ip;
2974         }
2975 
2976         if (in_f->ilf_fields & XFS_ILOG_AOWNER) {
2977                 ASSERT(in_f->ilf_fields & XFS_ILOG_ABROOT);
2978                 error = xfs_bmbt_change_owner(NULL, ip, XFS_ATTR_FORK,
2979                                               ip->i_ino, buffer_list);
2980                 if (error)
2981                         goto out_free_ip;
2982         }
2983 
2984 out_free_ip:
2985         xfs_inode_free(ip);
2986         return error;
2987 }
2988 
2989 STATIC int
2990 xlog_recover_inode_pass2(
2991         struct xlog                     *log,
2992         struct list_head                *buffer_list,
2993         struct xlog_recover_item        *item,
2994         xfs_lsn_t                       current_lsn)
2995 {
2996         struct xfs_inode_log_format     *in_f;
2997         xfs_mount_t             *mp = log->l_mp;
2998         xfs_buf_t               *bp;
2999         xfs_dinode_t            *dip;
3000         int                     len;
3001         char                    *src;
3002         char                    *dest;
3003         int                     error;
3004         int                     attr_index;
3005         uint                    fields;
3006         struct xfs_log_dinode   *ldip;
3007         uint                    isize;
3008         int                     need_free = 0;
3009 
3010         if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
3011                 in_f = item->ri_buf[0].i_addr;
3012         } else {
3013                 in_f = kmem_alloc(sizeof(struct xfs_inode_log_format), KM_SLEEP);
3014                 need_free = 1;
3015                 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
3016                 if (error)
3017                         goto error;
3018         }
3019 
3020         /*
3021          * Inode buffers can be freed, look out for it,
3022          * and do not replay the inode.
3023          */
3024         if (xlog_check_buffer_cancelled(log, in_f->ilf_blkno,
3025                                         in_f->ilf_len, 0)) {
3026                 error = 0;
3027                 trace_xfs_log_recover_inode_cancel(log, in_f);
3028                 goto error;
3029         }
3030         trace_xfs_log_recover_inode_recover(log, in_f);
3031 
3032         bp = xfs_buf_read(mp->m_ddev_targp, in_f->ilf_blkno, in_f->ilf_len, 0,
3033                           &xfs_inode_buf_ops);
3034         if (!bp) {
3035                 error = -ENOMEM;
3036                 goto error;
3037         }
3038         error = bp->b_error;
3039         if (error) {
3040                 xfs_buf_ioerror_alert(bp, "xlog_recover_do..(read#2)");
3041                 goto out_release;
3042         }
3043         ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
3044         dip = xfs_buf_offset(bp, in_f->ilf_boffset);
3045 
3046         /*
3047          * Make sure the place we're flushing out to really looks
3048          * like an inode!
3049          */
3050         if (unlikely(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC))) {
3051                 xfs_alert(mp,
3052         "%s: Bad inode magic number, dip = "PTR_FMT", dino bp = "PTR_FMT", ino = %Ld",
3053                         __func__, dip, bp, in_f->ilf_ino);
3054                 XFS_ERROR_REPORT("xlog_recover_inode_pass2(1)",
3055                                  XFS_ERRLEVEL_LOW, mp);
3056                 error = -EFSCORRUPTED;
3057                 goto out_release;
3058         }
3059         ldip = item->ri_buf[1].i_addr;
3060         if (unlikely(ldip->di_magic != XFS_DINODE_MAGIC)) {
3061                 xfs_alert(mp,
3062                         "%s: Bad inode log record, rec ptr "PTR_FMT", ino %Ld",
3063                         __func__, item, in_f->ilf_ino);
3064                 XFS_ERROR_REPORT("xlog_recover_inode_pass2(2)",
3065                                  XFS_ERRLEVEL_LOW, mp);
3066                 error = -EFSCORRUPTED;
3067                 goto out_release;
3068         }
3069 
3070         /*
3071          * If the inode has an LSN in it, recover the inode only if it's less
3072          * than the lsn of the transaction we are replaying. Note: we still
3073          * need to replay an owner change even though the inode is more recent
3074          * than the transaction as there is no guarantee that all the btree
3075          * blocks are more recent than this transaction, too.
3076          */
3077         if (dip->di_version >= 3) {
3078                 xfs_lsn_t       lsn = be64_to_cpu(dip->di_lsn);
3079 
3080                 if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3081                         trace_xfs_log_recover_inode_skip(log, in_f);
3082                         error = 0;
3083                         goto out_owner_change;
3084                 }
3085         }
3086 
3087         /*
3088          * di_flushiter is only valid for v1/2 inodes. All changes for v3 inodes
3089          * are transactional and if ordering is necessary we can determine that
3090          * more accurately by the LSN field in the V3 inode core. Don't trust
3091          * the inode versions we might be changing them here - use the
3092          * superblock flag to determine whether we need to look at di_flushiter
3093          * to skip replay when the on disk inode is newer than the log one
3094          */
3095         if (!xfs_sb_version_hascrc(&mp->m_sb) &&
3096             ldip->di_flushiter < be16_to_cpu(dip->di_flushiter)) {
3097                 /*
3098                  * Deal with the wrap case, DI_MAX_FLUSH is less
3099                  * than smaller numbers
3100                  */
3101                 if (be16_to_cpu(dip->di_flushiter) == DI_MAX_FLUSH &&
3102                     ldip->di_flushiter < (DI_MAX_FLUSH >> 1)) {
3103                         /* do nothing */
3104                 } else {
3105                         trace_xfs_log_recover_inode_skip(log, in_f);
3106                         error = 0;
3107                         goto out_release;
3108                 }
3109         }
3110 
3111         /* Take the opportunity to reset the flush iteration count */
3112         ldip->di_flushiter = 0;
3113 
3114         if (unlikely(S_ISREG(ldip->di_mode))) {
3115                 if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3116                     (ldip->di_format != XFS_DINODE_FMT_BTREE)) {
3117                         XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(3)",
3118                                          XFS_ERRLEVEL_LOW, mp, ldip);
3119                         xfs_alert(mp,
3120                 "%s: Bad regular inode log record, rec ptr "PTR_FMT", "
3121                 "ino ptr = "PTR_FMT", ino bp = "PTR_FMT", ino %Ld",
3122                                 __func__, item, dip, bp, in_f->ilf_ino);
3123                         error = -EFSCORRUPTED;
3124                         goto out_release;
3125                 }
3126         } else if (unlikely(S_ISDIR(ldip->di_mode))) {
3127                 if ((ldip->di_format != XFS_DINODE_FMT_EXTENTS) &&
3128                     (ldip->di_format != XFS_DINODE_FMT_BTREE) &&
3129                     (ldip->di_format != XFS_DINODE_FMT_LOCAL)) {
3130                         XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(4)",
3131                                              XFS_ERRLEVEL_LOW, mp, ldip);
3132                         xfs_alert(mp,
3133                 "%s: Bad dir inode log record, rec ptr "PTR_FMT", "
3134                 "ino ptr = "PTR_FMT", ino bp = "PTR_FMT", ino %Ld",
3135                                 __func__, item, dip, bp, in_f->ilf_ino);
3136                         error = -EFSCORRUPTED;
3137                         goto out_release;
3138                 }
3139         }
3140         if (unlikely(ldip->di_nextents + ldip->di_anextents > ldip->di_nblocks)){
3141                 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(5)",
3142                                      XFS_ERRLEVEL_LOW, mp, ldip);
3143                 xfs_alert(mp,
3144         "%s: Bad inode log record, rec ptr "PTR_FMT", dino ptr "PTR_FMT", "
3145         "dino bp "PTR_FMT", ino %Ld, total extents = %d, nblocks = %Ld",
3146                         __func__, item, dip, bp, in_f->ilf_ino,
3147                         ldip->di_nextents + ldip->di_anextents,
3148                         ldip->di_nblocks);
3149                 error = -EFSCORRUPTED;
3150                 goto out_release;
3151         }
3152         if (unlikely(ldip->di_forkoff > mp->m_sb.sb_inodesize)) {
3153                 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(6)",
3154                                      XFS_ERRLEVEL_LOW, mp, ldip);
3155                 xfs_alert(mp,
3156         "%s: Bad inode log record, rec ptr "PTR_FMT", dino ptr "PTR_FMT", "
3157         "dino bp "PTR_FMT", ino %Ld, forkoff 0x%x", __func__,
3158                         item, dip, bp, in_f->ilf_ino, ldip->di_forkoff);
3159                 error = -EFSCORRUPTED;
3160                 goto out_release;
3161         }
3162         isize = xfs_log_dinode_size(ldip->di_version);
3163         if (unlikely(item->ri_buf[1].i_len > isize)) {
3164                 XFS_CORRUPTION_ERROR("xlog_recover_inode_pass2(7)",
3165                                      XFS_ERRLEVEL_LOW, mp, ldip);
3166                 xfs_alert(mp,
3167                         "%s: Bad inode log record length %d, rec ptr "PTR_FMT,
3168                         __func__, item->ri_buf[1].i_len, item);
3169                 error = -EFSCORRUPTED;
3170                 goto out_release;
3171         }
3172 
3173         /* recover the log dinode inode into the on disk inode */
3174         xfs_log_dinode_to_disk(ldip, dip);
3175 
3176         /* the rest is in on-disk format */
3177         if (item->ri_buf[1].i_len > isize) {
3178                 memcpy((char *)dip + isize,
3179                         item->ri_buf[1].i_addr + isize,
3180                         item->ri_buf[1].i_len - isize);
3181         }
3182 
3183         fields = in_f->ilf_fields;
3184         if (fields & XFS_ILOG_DEV)
3185                 xfs_dinode_put_rdev(dip, in_f->ilf_u.ilfu_rdev);
3186 
3187         if (in_f->ilf_size == 2)
3188                 goto out_owner_change;
3189         len = item->ri_buf[2].i_len;
3190         src = item->ri_buf[2].i_addr;
3191         ASSERT(in_f->ilf_size <= 4);
3192         ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
3193         ASSERT(!(fields & XFS_ILOG_DFORK) ||
3194                (len == in_f->ilf_dsize));
3195 
3196         switch (fields & XFS_ILOG_DFORK) {
3197         case XFS_ILOG_DDATA:
3198         case XFS_ILOG_DEXT:
3199                 memcpy(XFS_DFORK_DPTR(dip), src, len);
3200                 break;
3201 
3202         case XFS_ILOG_DBROOT:
3203                 xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src, len,
3204                                  (xfs_bmdr_block_t *)XFS_DFORK_DPTR(dip),
3205                                  XFS_DFORK_DSIZE(dip, mp));
3206                 break;
3207 
3208         default:
3209                 /*
3210                  * There are no data fork flags set.
3211                  */
3212                 ASSERT((fields & XFS_ILOG_DFORK) == 0);
3213                 break;
3214         }
3215 
3216         /*
3217          * If we logged any attribute data, recover it.  There may or
3218          * may not have been any other non-core data logged in this
3219          * transaction.
3220          */
3221         if (in_f->ilf_fields & XFS_ILOG_AFORK) {
3222                 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
3223                         attr_index = 3;
3224                 } else {
3225                         attr_index = 2;
3226                 }
3227                 len = item->ri_buf[attr_index].i_len;
3228                 src = item->ri_buf[attr_index].i_addr;
3229                 ASSERT(len == in_f->ilf_asize);
3230 
3231                 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
3232                 case XFS_ILOG_ADATA:
3233                 case XFS_ILOG_AEXT:
3234                         dest = XFS_DFORK_APTR(dip);
3235                         ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
3236                         memcpy(dest, src, len);
3237                         break;
3238 
3239                 case XFS_ILOG_ABROOT:
3240                         dest = XFS_DFORK_APTR(dip);
3241                         xfs_bmbt_to_bmdr(mp, (struct xfs_btree_block *)src,
3242                                          len, (xfs_bmdr_block_t*)dest,
3243                                          XFS_DFORK_ASIZE(dip, mp));
3244                         break;
3245 
3246                 default:
3247                         xfs_warn(log->l_mp, "%s: Invalid flag", __func__);
3248                         ASSERT(0);
3249                         error = -EIO;
3250                         goto out_release;
3251                 }
3252         }
3253 
3254 out_owner_change:
3255         if (in_f->ilf_fields & (XFS_ILOG_DOWNER|XFS_ILOG_AOWNER))
3256                 error = xfs_recover_inode_owner_change(mp, dip, in_f,
3257                                                        buffer_list);
3258         /* re-generate the checksum. */
3259         xfs_dinode_calc_crc(log->l_mp, dip);
3260 
3261         ASSERT(bp->b_target->bt_mount == mp);
3262         bp->b_iodone = xlog_recover_iodone;
3263         xfs_buf_delwri_queue(bp, buffer_list);
3264 
3265 out_release:
3266         xfs_buf_relse(bp);
3267 error:
3268         if (need_free)
3269                 kmem_free(in_f);
3270         return error;
3271 }
3272 
3273 /*
3274  * Recover QUOTAOFF records. We simply make a note of it in the xlog
3275  * structure, so that we know not to do any dquot item or dquot buffer recovery,
3276  * of that type.
3277  */
3278 STATIC int
3279 xlog_recover_quotaoff_pass1(
3280         struct xlog                     *log,
3281         struct xlog_recover_item        *item)
3282 {
3283         xfs_qoff_logformat_t    *qoff_f = item->ri_buf[0].i_addr;
3284         ASSERT(qoff_f);
3285 
3286         /*
3287          * The logitem format's flag tells us if this was user quotaoff,
3288          * group/project quotaoff or both.
3289          */
3290         if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
3291                 log->l_quotaoffs_flag |= XFS_DQ_USER;
3292         if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
3293                 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
3294         if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
3295                 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
3296 
3297         return 0;
3298 }
3299 
3300 /*
3301  * Recover a dquot record
3302  */
3303 STATIC int
3304 xlog_recover_dquot_pass2(
3305         struct xlog                     *log,
3306         struct list_head                *buffer_list,
3307         struct xlog_recover_item        *item,
3308         xfs_lsn_t                       current_lsn)
3309 {
3310         xfs_mount_t             *mp = log->l_mp;
3311         xfs_buf_t               *bp;
3312         struct xfs_disk_dquot   *ddq, *recddq;
3313         xfs_failaddr_t          fa;
3314         int                     error;
3315         xfs_dq_logformat_t      *dq_f;
3316         uint                    type;
3317 
3318 
3319         /*
3320          * Filesystems are required to send in quota flags at mount time.
3321          */
3322         if (mp->m_qflags == 0)
3323                 return 0;
3324 
3325         recddq = item->ri_buf[1].i_addr;
3326         if (recddq == NULL) {
3327                 xfs_alert(log->l_mp, "NULL dquot in %s.", __func__);
3328                 return -EIO;
3329         }
3330         if (item->ri_buf[1].i_len < sizeof(xfs_disk_dquot_t)) {
3331                 xfs_alert(log->l_mp, "dquot too small (%d) in %s.",
3332                         item->ri_buf[1].i_len, __func__);
3333                 return -EIO;
3334         }
3335 
3336         /*
3337          * This type of quotas was turned off, so ignore this record.
3338          */
3339         type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
3340         ASSERT(type);
3341         if (log->l_quotaoffs_flag & type)
3342                 return 0;
3343 
3344         /*
3345          * At this point we know that quota was _not_ turned off.
3346          * Since the mount flags are not indicating to us otherwise, this
3347          * must mean that quota is on, and the dquot needs to be replayed.
3348          * Remember that we may not have fully recovered the superblock yet,
3349          * so we can't do the usual trick of looking at the SB quota bits.
3350          *
3351          * The other possibility, of course, is that the quota subsystem was
3352          * removed since the last mount - ENOSYS.
3353          */
3354         dq_f = item->ri_buf[0].i_addr;
3355         ASSERT(dq_f);
3356         fa = xfs_dquot_verify(mp, recddq, dq_f->qlf_id, 0, 0);
3357         if (fa) {
3358                 xfs_alert(mp, "corrupt dquot ID 0x%x in log at %pS",
3359                                 dq_f->qlf_id, fa);
3360                 return -EIO;
3361         }
3362         ASSERT(dq_f->qlf_len == 1);
3363 
3364         /*
3365          * At this point we are assuming that the dquots have been allocated
3366          * and hence the buffer has valid dquots stamped in it. It should,
3367          * therefore, pass verifier validation. If the dquot is bad, then the
3368          * we'll return an error here, so we don't need to specifically check
3369          * the dquot in the buffer after the verifier has run.
3370          */
3371         error = xfs_trans_read_buf(mp, NULL, mp->m_ddev_targp, dq_f->qlf_blkno,
3372                                    XFS_FSB_TO_BB(mp, dq_f->qlf_len), 0, &bp,
3373                                    &xfs_dquot_buf_ops);
3374         if (error)
3375                 return error;
3376 
3377         ASSERT(bp);
3378         ddq = xfs_buf_offset(bp, dq_f->qlf_boffset);
3379 
3380         /*
3381          * If the dquot has an LSN in it, recover the dquot only if it's less
3382          * than the lsn of the transaction we are replaying.
3383          */
3384         if (xfs_sb_version_hascrc(&mp->m_sb)) {
3385                 struct xfs_dqblk *dqb = (struct xfs_dqblk *)ddq;
3386                 xfs_lsn_t       lsn = be64_to_cpu(dqb->dd_lsn);
3387 
3388                 if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
3389                         goto out_release;
3390                 }
3391         }
3392 
3393         memcpy(ddq, recddq, item->ri_buf[1].i_len);
3394         if (xfs_sb_version_hascrc(&mp->m_sb)) {
3395                 xfs_update_cksum((char *)ddq, sizeof(struct xfs_dqblk),
3396                                  XFS_DQUOT_CRC_OFF);
3397         }
3398 
3399         ASSERT(dq_f->qlf_size == 2);
3400         ASSERT(bp->b_target->bt_mount == mp);
3401         bp->b_iodone = xlog_recover_iodone;
3402         xfs_buf_delwri_queue(bp, buffer_list);
3403 
3404 out_release:
3405         xfs_buf_relse(bp);
3406         return 0;
3407 }
3408 
3409 /*
3410  * This routine is called to create an in-core extent free intent
3411  * item from the efi format structure which was logged on disk.
3412  * It allocates an in-core efi, copies the extents from the format
3413  * structure into it, and adds the efi to the AIL with the given
3414  * LSN.
3415  */
3416 STATIC int
3417 xlog_recover_efi_pass2(
3418         struct xlog                     *log,
3419         struct xlog_recover_item        *item,
3420         xfs_lsn_t                       lsn)
3421 {
3422         int                             error;
3423         struct xfs_mount                *mp = log->l_mp;
3424         struct xfs_efi_log_item         *efip;
3425         struct xfs_efi_log_format       *efi_formatp;
3426 
3427         efi_formatp = item->ri_buf[0].i_addr;
3428 
3429         efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
3430         error = xfs_efi_copy_format(&item->ri_buf[0], &efip->efi_format);
3431         if (error) {
3432                 xfs_efi_item_free(efip);
3433                 return error;
3434         }
3435         atomic_set(&efip->efi_next_extent, efi_formatp->efi_nextents);
3436 
3437         spin_lock(&log->l_ailp->xa_lock);
3438         /*
3439          * The EFI has two references. One for the EFD and one for EFI to ensure
3440          * it makes it into the AIL. Insert the EFI into the AIL directly and
3441          * drop the EFI reference. Note that xfs_trans_ail_update() drops the
3442          * AIL lock.
3443          */
3444         xfs_trans_ail_update(log->l_ailp, &efip->efi_item, lsn);
3445         xfs_efi_release(efip);
3446         return 0;
3447 }
3448 
3449 
3450 /*
3451  * This routine is called when an EFD format structure is found in a committed
3452  * transaction in the log. Its purpose is to cancel the corresponding EFI if it
3453  * was still in the log. To do this it searches the AIL for the EFI with an id
3454  * equal to that in the EFD format structure. If we find it we drop the EFD
3455  * reference, which removes the EFI from the AIL and frees it.
3456  */
3457 STATIC int
3458 xlog_recover_efd_pass2(
3459         struct xlog                     *log,
3460         struct xlog_recover_item        *item)
3461 {
3462         xfs_efd_log_format_t    *efd_formatp;
3463         xfs_efi_log_item_t      *efip = NULL;
3464         xfs_log_item_t          *lip;
3465         uint64_t                efi_id;
3466         struct xfs_ail_cursor   cur;
3467         struct xfs_ail          *ailp = log->l_ailp;
3468 
3469         efd_formatp = item->ri_buf[0].i_addr;
3470         ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
3471                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
3472                (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
3473                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
3474         efi_id = efd_formatp->efd_efi_id;
3475 
3476         /*
3477          * Search for the EFI with the id in the EFD format structure in the
3478          * AIL.
3479          */
3480         spin_lock(&ailp->xa_lock);
3481         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3482         while (lip != NULL) {
3483                 if (lip->li_type == XFS_LI_EFI) {
3484                         efip = (xfs_efi_log_item_t *)lip;
3485                         if (efip->efi_format.efi_id == efi_id) {
3486                                 /*
3487                                  * Drop the EFD reference to the EFI. This
3488                                  * removes the EFI from the AIL and frees it.
3489                                  */
3490                                 spin_unlock(&ailp->xa_lock);
3491                                 xfs_efi_release(efip);
3492                                 spin_lock(&ailp->xa_lock);
3493                                 break;
3494                         }
3495                 }
3496                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3497         }
3498 
3499         xfs_trans_ail_cursor_done(&cur);
3500         spin_unlock(&ailp->xa_lock);
3501 
3502         return 0;
3503 }
3504 
3505 /*
3506  * This routine is called to create an in-core extent rmap update
3507  * item from the rui format structure which was logged on disk.
3508  * It allocates an in-core rui, copies the extents from the format
3509  * structure into it, and adds the rui to the AIL with the given
3510  * LSN.
3511  */
3512 STATIC int
3513 xlog_recover_rui_pass2(
3514         struct xlog                     *log,
3515         struct xlog_recover_item        *item,
3516         xfs_lsn_t                       lsn)
3517 {
3518         int                             error;
3519         struct xfs_mount                *mp = log->l_mp;
3520         struct xfs_rui_log_item         *ruip;
3521         struct xfs_rui_log_format       *rui_formatp;
3522 
3523         rui_formatp = item->ri_buf[0].i_addr;
3524 
3525         ruip = xfs_rui_init(mp, rui_formatp->rui_nextents);
3526         error = xfs_rui_copy_format(&item->ri_buf[0], &ruip->rui_format);
3527         if (error) {
3528                 xfs_rui_item_free(ruip);
3529                 return error;
3530         }
3531         atomic_set(&ruip->rui_next_extent, rui_formatp->rui_nextents);
3532 
3533         spin_lock(&log->l_ailp->xa_lock);
3534         /*
3535          * The RUI has two references. One for the RUD and one for RUI to ensure
3536          * it makes it into the AIL. Insert the RUI into the AIL directly and
3537          * drop the RUI reference. Note that xfs_trans_ail_update() drops the
3538          * AIL lock.
3539          */
3540         xfs_trans_ail_update(log->l_ailp, &ruip->rui_item, lsn);
3541         xfs_rui_release(ruip);
3542         return 0;
3543 }
3544 
3545 
3546 /*
3547  * This routine is called when an RUD format structure is found in a committed
3548  * transaction in the log. Its purpose is to cancel the corresponding RUI if it
3549  * was still in the log. To do this it searches the AIL for the RUI with an id
3550  * equal to that in the RUD format structure. If we find it we drop the RUD
3551  * reference, which removes the RUI from the AIL and frees it.
3552  */
3553 STATIC int
3554 xlog_recover_rud_pass2(
3555         struct xlog                     *log,
3556         struct xlog_recover_item        *item)
3557 {
3558         struct xfs_rud_log_format       *rud_formatp;
3559         struct xfs_rui_log_item         *ruip = NULL;
3560         struct xfs_log_item             *lip;
3561         uint64_t                        rui_id;
3562         struct xfs_ail_cursor           cur;
3563         struct xfs_ail                  *ailp = log->l_ailp;
3564 
3565         rud_formatp = item->ri_buf[0].i_addr;
3566         ASSERT(item->ri_buf[0].i_len == sizeof(struct xfs_rud_log_format));
3567         rui_id = rud_formatp->rud_rui_id;
3568 
3569         /*
3570          * Search for the RUI with the id in the RUD format structure in the
3571          * AIL.
3572          */
3573         spin_lock(&ailp->xa_lock);
3574         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3575         while (lip != NULL) {
3576                 if (lip->li_type == XFS_LI_RUI) {
3577                         ruip = (struct xfs_rui_log_item *)lip;
3578                         if (ruip->rui_format.rui_id == rui_id) {
3579                                 /*
3580                                  * Drop the RUD reference to the RUI. This
3581                                  * removes the RUI from the AIL and frees it.
3582                                  */
3583                                 spin_unlock(&ailp->xa_lock);
3584                                 xfs_rui_release(ruip);
3585                                 spin_lock(&ailp->xa_lock);
3586                                 break;
3587                         }
3588                 }
3589                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3590         }
3591 
3592         xfs_trans_ail_cursor_done(&cur);
3593         spin_unlock(&ailp->xa_lock);
3594 
3595         return 0;
3596 }
3597 
3598 /*
3599  * Copy an CUI format buffer from the given buf, and into the destination
3600  * CUI format structure.  The CUI/CUD items were designed not to need any
3601  * special alignment handling.
3602  */
3603 static int
3604 xfs_cui_copy_format(
3605         struct xfs_log_iovec            *buf,
3606         struct xfs_cui_log_format       *dst_cui_fmt)
3607 {
3608         struct xfs_cui_log_format       *src_cui_fmt;
3609         uint                            len;
3610 
3611         src_cui_fmt = buf->i_addr;
3612         len = xfs_cui_log_format_sizeof(src_cui_fmt->cui_nextents);
3613 
3614         if (buf->i_len == len) {
3615                 memcpy(dst_cui_fmt, src_cui_fmt, len);
3616                 return 0;
3617         }
3618         return -EFSCORRUPTED;
3619 }
3620 
3621 /*
3622  * This routine is called to create an in-core extent refcount update
3623  * item from the cui format structure which was logged on disk.
3624  * It allocates an in-core cui, copies the extents from the format
3625  * structure into it, and adds the cui to the AIL with the given
3626  * LSN.
3627  */
3628 STATIC int
3629 xlog_recover_cui_pass2(
3630         struct xlog                     *log,
3631         struct xlog_recover_item        *item,
3632         xfs_lsn_t                       lsn)
3633 {
3634         int                             error;
3635         struct xfs_mount                *mp = log->l_mp;
3636         struct xfs_cui_log_item         *cuip;
3637         struct xfs_cui_log_format       *cui_formatp;
3638 
3639         cui_formatp = item->ri_buf[0].i_addr;
3640 
3641         cuip = xfs_cui_init(mp, cui_formatp->cui_nextents);
3642         error = xfs_cui_copy_format(&item->ri_buf[0], &cuip->cui_format);
3643         if (error) {
3644                 xfs_cui_item_free(cuip);
3645                 return error;
3646         }
3647         atomic_set(&cuip->cui_next_extent, cui_formatp->cui_nextents);
3648 
3649         spin_lock(&log->l_ailp->xa_lock);
3650         /*
3651          * The CUI has two references. One for the CUD and one for CUI to ensure
3652          * it makes it into the AIL. Insert the CUI into the AIL directly and
3653          * drop the CUI reference. Note that xfs_trans_ail_update() drops the
3654          * AIL lock.
3655          */
3656         xfs_trans_ail_update(log->l_ailp, &cuip->cui_item, lsn);
3657         xfs_cui_release(cuip);
3658         return 0;
3659 }
3660 
3661 
3662 /*
3663  * This routine is called when an CUD format structure is found in a committed
3664  * transaction in the log. Its purpose is to cancel the corresponding CUI if it
3665  * was still in the log. To do this it searches the AIL for the CUI with an id
3666  * equal to that in the CUD format structure. If we find it we drop the CUD
3667  * reference, which removes the CUI from the AIL and frees it.
3668  */
3669 STATIC int
3670 xlog_recover_cud_pass2(
3671         struct xlog                     *log,
3672         struct xlog_recover_item        *item)
3673 {
3674         struct xfs_cud_log_format       *cud_formatp;
3675         struct xfs_cui_log_item         *cuip = NULL;
3676         struct xfs_log_item             *lip;
3677         uint64_t                        cui_id;
3678         struct xfs_ail_cursor           cur;
3679         struct xfs_ail                  *ailp = log->l_ailp;
3680 
3681         cud_formatp = item->ri_buf[0].i_addr;
3682         if (item->ri_buf[0].i_len != sizeof(struct xfs_cud_log_format))
3683                 return -EFSCORRUPTED;
3684         cui_id = cud_formatp->cud_cui_id;
3685 
3686         /*
3687          * Search for the CUI with the id in the CUD format structure in the
3688          * AIL.
3689          */
3690         spin_lock(&ailp->xa_lock);
3691         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3692         while (lip != NULL) {
3693                 if (lip->li_type == XFS_LI_CUI) {
3694                         cuip = (struct xfs_cui_log_item *)lip;
3695                         if (cuip->cui_format.cui_id == cui_id) {
3696                                 /*
3697                                  * Drop the CUD reference to the CUI. This
3698                                  * removes the CUI from the AIL and frees it.
3699                                  */
3700                                 spin_unlock(&ailp->xa_lock);
3701                                 xfs_cui_release(cuip);
3702                                 spin_lock(&ailp->xa_lock);
3703                                 break;
3704                         }
3705                 }
3706                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3707         }
3708 
3709         xfs_trans_ail_cursor_done(&cur);
3710         spin_unlock(&ailp->xa_lock);
3711 
3712         return 0;
3713 }
3714 
3715 /*
3716  * Copy an BUI format buffer from the given buf, and into the destination
3717  * BUI format structure.  The BUI/BUD items were designed not to need any
3718  * special alignment handling.
3719  */
3720 static int
3721 xfs_bui_copy_format(
3722         struct xfs_log_iovec            *buf,
3723         struct xfs_bui_log_format       *dst_bui_fmt)
3724 {
3725         struct xfs_bui_log_format       *src_bui_fmt;
3726         uint                            len;
3727 
3728         src_bui_fmt = buf->i_addr;
3729         len = xfs_bui_log_format_sizeof(src_bui_fmt->bui_nextents);
3730 
3731         if (buf->i_len == len) {
3732                 memcpy(dst_bui_fmt, src_bui_fmt, len);
3733                 return 0;
3734         }
3735         return -EFSCORRUPTED;
3736 }
3737 
3738 /*
3739  * This routine is called to create an in-core extent bmap update
3740  * item from the bui format structure which was logged on disk.
3741  * It allocates an in-core bui, copies the extents from the format
3742  * structure into it, and adds the bui to the AIL with the given
3743  * LSN.
3744  */
3745 STATIC int
3746 xlog_recover_bui_pass2(
3747         struct xlog                     *log,
3748         struct xlog_recover_item        *item,
3749         xfs_lsn_t                       lsn)
3750 {
3751         int                             error;
3752         struct xfs_mount                *mp = log->l_mp;
3753         struct xfs_bui_log_item         *buip;
3754         struct xfs_bui_log_format       *bui_formatp;
3755 
3756         bui_formatp = item->ri_buf[0].i_addr;
3757 
3758         if (bui_formatp->bui_nextents != XFS_BUI_MAX_FAST_EXTENTS)
3759                 return -EFSCORRUPTED;
3760         buip = xfs_bui_init(mp);
3761         error = xfs_bui_copy_format(&item->ri_buf[0], &buip->bui_format);
3762         if (error) {
3763                 xfs_bui_item_free(buip);
3764                 return error;
3765         }
3766         atomic_set(&buip->bui_next_extent, bui_formatp->bui_nextents);
3767 
3768         spin_lock(&log->l_ailp->xa_lock);
3769         /*
3770          * The RUI has two references. One for the RUD and one for RUI to ensure
3771          * it makes it into the AIL. Insert the RUI into the AIL directly and
3772          * drop the RUI reference. Note that xfs_trans_ail_update() drops the
3773          * AIL lock.
3774          */
3775         xfs_trans_ail_update(log->l_ailp, &buip->bui_item, lsn);
3776         xfs_bui_release(buip);
3777         return 0;
3778 }
3779 
3780 
3781 /*
3782  * This routine is called when an BUD format structure is found in a committed
3783  * transaction in the log. Its purpose is to cancel the corresponding BUI if it
3784  * was still in the log. To do this it searches the AIL for the BUI with an id
3785  * equal to that in the BUD format structure. If we find it we drop the BUD
3786  * reference, which removes the BUI from the AIL and frees it.
3787  */
3788 STATIC int
3789 xlog_recover_bud_pass2(
3790         struct xlog                     *log,
3791         struct xlog_recover_item        *item)
3792 {
3793         struct xfs_bud_log_format       *bud_formatp;
3794         struct xfs_bui_log_item         *buip = NULL;
3795         struct xfs_log_item             *lip;
3796         uint64_t                        bui_id;
3797         struct xfs_ail_cursor           cur;
3798         struct xfs_ail                  *ailp = log->l_ailp;
3799 
3800         bud_formatp = item->ri_buf[0].i_addr;
3801         if (item->ri_buf[0].i_len != sizeof(struct xfs_bud_log_format))
3802                 return -EFSCORRUPTED;
3803         bui_id = bud_formatp->bud_bui_id;
3804 
3805         /*
3806          * Search for the BUI with the id in the BUD format structure in the
3807          * AIL.
3808          */
3809         spin_lock(&ailp->xa_lock);
3810         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
3811         while (lip != NULL) {
3812                 if (lip->li_type == XFS_LI_BUI) {
3813                         buip = (struct xfs_bui_log_item *)lip;
3814                         if (buip->bui_format.bui_id == bui_id) {
3815                                 /*
3816                                  * Drop the BUD reference to the BUI. This
3817                                  * removes the BUI from the AIL and frees it.
3818                                  */
3819                                 spin_unlock(&ailp->xa_lock);
3820                                 xfs_bui_release(buip);
3821                                 spin_lock(&ailp->xa_lock);
3822                                 break;
3823                         }
3824                 }
3825                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
3826         }
3827 
3828         xfs_trans_ail_cursor_done(&cur);
3829         spin_unlock(&ailp->xa_lock);
3830 
3831         return 0;
3832 }
3833 
3834 /*
3835  * This routine is called when an inode create format structure is found in a
3836  * committed transaction in the log.  It's purpose is to initialise the inodes
3837  * being allocated on disk. This requires us to get inode cluster buffers that
3838  * match the range to be initialised, stamped with inode templates and written
3839  * by delayed write so that subsequent modifications will hit the cached buffer
3840  * and only need writing out at the end of recovery.
3841  */
3842 STATIC int
3843 xlog_recover_do_icreate_pass2(
3844         struct xlog             *log,
3845         struct list_head        *buffer_list,
3846         xlog_recover_item_t     *item)
3847 {
3848         struct xfs_mount        *mp = log->l_mp;
3849         struct xfs_icreate_log  *icl;
3850         xfs_agnumber_t          agno;
3851         xfs_agblock_t           agbno;
3852         unsigned int            count;
3853         unsigned int            isize;
3854         xfs_agblock_t           length;
3855         int                     blks_per_cluster;
3856         int                     bb_per_cluster;
3857         int                     cancel_count;
3858         int                     nbufs;
3859         int                     i;
3860 
3861         icl = (struct xfs_icreate_log *)item->ri_buf[0].i_addr;
3862         if (icl->icl_type != XFS_LI_ICREATE) {
3863                 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad type");
3864                 return -EINVAL;
3865         }
3866 
3867         if (icl->icl_size != 1) {
3868                 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad icl size");
3869                 return -EINVAL;
3870         }
3871 
3872         agno = be32_to_cpu(icl->icl_ag);
3873         if (agno >= mp->m_sb.sb_agcount) {
3874                 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agno");
3875                 return -EINVAL;
3876         }
3877         agbno = be32_to_cpu(icl->icl_agbno);
3878         if (!agbno || agbno == NULLAGBLOCK || agbno >= mp->m_sb.sb_agblocks) {
3879                 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad agbno");
3880                 return -EINVAL;
3881         }
3882         isize = be32_to_cpu(icl->icl_isize);
3883         if (isize != mp->m_sb.sb_inodesize) {
3884                 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad isize");
3885                 return -EINVAL;
3886         }
3887         count = be32_to_cpu(icl->icl_count);
3888         if (!count) {
3889                 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad count");
3890                 return -EINVAL;
3891         }
3892         length = be32_to_cpu(icl->icl_length);
3893         if (!length || length >= mp->m_sb.sb_agblocks) {
3894                 xfs_warn(log->l_mp, "xlog_recover_do_icreate_trans: bad length");
3895                 return -EINVAL;
3896         }
3897 
3898         /*
3899          * The inode chunk is either full or sparse and we only support
3900          * m_ialloc_min_blks sized sparse allocations at this time.
3901          */
3902         if (length != mp->m_ialloc_blks &&
3903             length != mp->m_ialloc_min_blks) {
3904                 xfs_warn(log->l_mp,
3905                          "%s: unsupported chunk length", __FUNCTION__);
3906                 return -EINVAL;
3907         }
3908 
3909         /* verify inode count is consistent with extent length */
3910         if ((count >> mp->m_sb.sb_inopblog) != length) {
3911                 xfs_warn(log->l_mp,
3912                          "%s: inconsistent inode count and chunk length",
3913                          __FUNCTION__);
3914                 return -EINVAL;
3915         }
3916 
3917         /*
3918          * The icreate transaction can cover multiple cluster buffers and these
3919          * buffers could have been freed and reused. Check the individual
3920          * buffers for cancellation so we don't overwrite anything written after
3921          * a cancellation.
3922          */
3923         blks_per_cluster = xfs_icluster_size_fsb(mp);
3924         bb_per_cluster = XFS_FSB_TO_BB(mp, blks_per_cluster);
3925         nbufs = length / blks_per_cluster;
3926         for (i = 0, cancel_count = 0; i < nbufs; i++) {
3927                 xfs_daddr_t     daddr;
3928 
3929                 daddr = XFS_AGB_TO_DADDR(mp, agno,
3930                                          agbno + i * blks_per_cluster);
3931                 if (xlog_check_buffer_cancelled(log, daddr, bb_per_cluster, 0))
3932                         cancel_count++;
3933         }
3934 
3935         /*
3936          * We currently only use icreate for a single allocation at a time. This
3937          * means we should expect either all or none of the buffers to be
3938          * cancelled. Be conservative and skip replay if at least one buffer is
3939          * cancelled, but warn the user that something is awry if the buffers
3940          * are not consistent.
3941          *
3942          * XXX: This must be refined to only skip cancelled clusters once we use
3943          * icreate for multiple chunk allocations.
3944          */
3945         ASSERT(!cancel_count || cancel_count == nbufs);
3946         if (cancel_count) {
3947                 if (cancel_count != nbufs)
3948                         xfs_warn(mp,
3949         "WARNING: partial inode chunk cancellation, skipped icreate.");
3950                 trace_xfs_log_recover_icreate_cancel(log, icl);
3951                 return 0;
3952         }
3953 
3954         trace_xfs_log_recover_icreate_recover(log, icl);
3955         return xfs_ialloc_inode_init(mp, NULL, buffer_list, count, agno, agbno,
3956                                      length, be32_to_cpu(icl->icl_gen));
3957 }
3958 
3959 STATIC void
3960 xlog_recover_buffer_ra_pass2(
3961         struct xlog                     *log,
3962         struct xlog_recover_item        *item)
3963 {
3964         struct xfs_buf_log_format       *buf_f = item->ri_buf[0].i_addr;
3965         struct xfs_mount                *mp = log->l_mp;
3966 
3967         if (xlog_peek_buffer_cancelled(log, buf_f->blf_blkno,
3968                         buf_f->blf_len, buf_f->blf_flags)) {
3969                 return;
3970         }
3971 
3972         xfs_buf_readahead(mp->m_ddev_targp, buf_f->blf_blkno,
3973                                 buf_f->blf_len, NULL);
3974 }
3975 
3976 STATIC void
3977 xlog_recover_inode_ra_pass2(
3978         struct xlog                     *log,
3979         struct xlog_recover_item        *item)
3980 {
3981         struct xfs_inode_log_format     ilf_buf;
3982         struct xfs_inode_log_format     *ilfp;
3983         struct xfs_mount                *mp = log->l_mp;
3984         int                     error;
3985 
3986         if (item->ri_buf[0].i_len == sizeof(struct xfs_inode_log_format)) {
3987                 ilfp = item->ri_buf[0].i_addr;
3988         } else {
3989                 ilfp = &ilf_buf;
3990                 memset(ilfp, 0, sizeof(*ilfp));
3991                 error = xfs_inode_item_format_convert(&item->ri_buf[0], ilfp);
3992                 if (error)
3993                         return;
3994         }
3995 
3996         if (xlog_peek_buffer_cancelled(log, ilfp->ilf_blkno, ilfp->ilf_len, 0))
3997                 return;
3998 
3999         xfs_buf_readahead(mp->m_ddev_targp, ilfp->ilf_blkno,
4000                                 ilfp->ilf_len, &xfs_inode_buf_ra_ops);
4001 }
4002 
4003 STATIC void
4004 xlog_recover_dquot_ra_pass2(
4005         struct xlog                     *log,
4006         struct xlog_recover_item        *item)
4007 {
4008         struct xfs_mount        *mp = log->l_mp;
4009         struct xfs_disk_dquot   *recddq;
4010         struct xfs_dq_logformat *dq_f;
4011         uint                    type;
4012         int                     len;
4013 
4014 
4015         if (mp->m_qflags == 0)
4016                 return;
4017 
4018         recddq = item->ri_buf[1].i_addr;
4019         if (recddq == NULL)
4020                 return;
4021         if (item->ri_buf[1].i_len < sizeof(struct xfs_disk_dquot))
4022                 return;
4023 
4024         type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
4025         ASSERT(type);
4026         if (log->l_quotaoffs_flag & type)
4027                 return;
4028 
4029         dq_f = item->ri_buf[0].i_addr;
4030         ASSERT(dq_f);
4031         ASSERT(dq_f->qlf_len == 1);
4032 
4033         len = XFS_FSB_TO_BB(mp, dq_f->qlf_len);
4034         if (xlog_peek_buffer_cancelled(log, dq_f->qlf_blkno, len, 0))
4035                 return;
4036 
4037         xfs_buf_readahead(mp->m_ddev_targp, dq_f->qlf_blkno, len,
4038                           &xfs_dquot_buf_ra_ops);
4039 }
4040 
4041 STATIC void
4042 xlog_recover_ra_pass2(
4043         struct xlog                     *log,
4044         struct xlog_recover_item        *item)
4045 {
4046         switch (ITEM_TYPE(item)) {
4047         case XFS_LI_BUF:
4048                 xlog_recover_buffer_ra_pass2(log, item);
4049                 break;
4050         case XFS_LI_INODE:
4051                 xlog_recover_inode_ra_pass2(log, item);
4052                 break;
4053         case XFS_LI_DQUOT:
4054                 xlog_recover_dquot_ra_pass2(log, item);
4055                 break;
4056         case XFS_LI_EFI:
4057         case XFS_LI_EFD:
4058         case XFS_LI_QUOTAOFF:
4059         case XFS_LI_RUI:
4060         case XFS_LI_RUD:
4061         case XFS_LI_CUI:
4062         case XFS_LI_CUD:
4063         case XFS_LI_BUI:
4064         case XFS_LI_BUD:
4065         default:
4066                 break;
4067         }
4068 }
4069 
4070 STATIC int
4071 xlog_recover_commit_pass1(
4072         struct xlog                     *log,
4073         struct xlog_recover             *trans,
4074         struct xlog_recover_item        *item)
4075 {
4076         trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS1);
4077 
4078         switch (ITEM_TYPE(item)) {
4079         case XFS_LI_BUF:
4080                 return xlog_recover_buffer_pass1(log, item);
4081         case XFS_LI_QUOTAOFF:
4082                 return xlog_recover_quotaoff_pass1(log, item);
4083         case XFS_LI_INODE:
4084         case XFS_LI_EFI:
4085         case XFS_LI_EFD:
4086         case XFS_LI_DQUOT:
4087         case XFS_LI_ICREATE:
4088         case XFS_LI_RUI:
4089         case XFS_LI_RUD:
4090         case XFS_LI_CUI:
4091         case XFS_LI_CUD:
4092         case XFS_LI_BUI:
4093         case XFS_LI_BUD:
4094                 /* nothing to do in pass 1 */
4095                 return 0;
4096         default:
4097                 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
4098                         __func__, ITEM_TYPE(item));
4099                 ASSERT(0);
4100                 return -EIO;
4101         }
4102 }
4103 
4104 STATIC int
4105 xlog_recover_commit_pass2(
4106         struct xlog                     *log,
4107         struct xlog_recover             *trans,
4108         struct list_head                *buffer_list,
4109         struct xlog_recover_item        *item)
4110 {
4111         trace_xfs_log_recover_item_recover(log, trans, item, XLOG_RECOVER_PASS2);
4112 
4113         switch (ITEM_TYPE(item)) {
4114         case XFS_LI_BUF:
4115                 return xlog_recover_buffer_pass2(log, buffer_list, item,
4116                                                  trans->r_lsn);
4117         case XFS_LI_INODE:
4118                 return xlog_recover_inode_pass2(log, buffer_list, item,
4119                                                  trans->r_lsn);
4120         case XFS_LI_EFI:
4121                 return xlog_recover_efi_pass2(log, item, trans->r_lsn);
4122         case XFS_LI_EFD:
4123                 return xlog_recover_efd_pass2(log, item);
4124         case XFS_LI_RUI:
4125                 return xlog_recover_rui_pass2(log, item, trans->r_lsn);
4126         case XFS_LI_RUD:
4127                 return xlog_recover_rud_pass2(log, item);
4128         case XFS_LI_CUI:
4129                 return xlog_recover_cui_pass2(log, item, trans->r_lsn);
4130         case XFS_LI_CUD:
4131                 return xlog_recover_cud_pass2(log, item);
4132         case XFS_LI_BUI:
4133                 return xlog_recover_bui_pass2(log, item, trans->r_lsn);
4134         case XFS_LI_BUD:
4135                 return xlog_recover_bud_pass2(log, item);
4136         case XFS_LI_DQUOT:
4137                 return xlog_recover_dquot_pass2(log, buffer_list, item,
4138                                                 trans->r_lsn);
4139         case XFS_LI_ICREATE:
4140                 return xlog_recover_do_icreate_pass2(log, buffer_list, item);
4141         case XFS_LI_QUOTAOFF:
4142                 /* nothing to do in pass2 */
4143                 return 0;
4144         default:
4145                 xfs_warn(log->l_mp, "%s: invalid item type (%d)",
4146                         __func__, ITEM_TYPE(item));
4147                 ASSERT(0);
4148                 return -EIO;
4149         }
4150 }
4151 
4152 STATIC int
4153 xlog_recover_items_pass2(
4154         struct xlog                     *log,
4155         struct xlog_recover             *trans,
4156         struct list_head                *buffer_list,
4157         struct list_head                *item_list)
4158 {
4159         struct xlog_recover_item        *item;
4160         int                             error = 0;
4161 
4162         list_for_each_entry(item, item_list, ri_list) {
4163                 error = xlog_recover_commit_pass2(log, trans,
4164                                           buffer_list, item);
4165                 if (error)
4166                         return error;
4167         }
4168 
4169         return error;
4170 }
4171 
4172 /*
4173  * Perform the transaction.
4174  *
4175  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
4176  * EFIs and EFDs get queued up by adding entries into the AIL for them.
4177  */
4178 STATIC int
4179 xlog_recover_commit_trans(
4180         struct xlog             *log,
4181         struct xlog_recover     *trans,
4182         int                     pass,
4183         struct list_head        *buffer_list)
4184 {
4185         int                             error = 0;
4186         int                             items_queued = 0;
4187         struct xlog_recover_item        *item;
4188         struct xlog_recover_item        *next;
4189         LIST_HEAD                       (ra_list);
4190         LIST_HEAD                       (done_list);
4191 
4192         #define XLOG_RECOVER_COMMIT_QUEUE_MAX 100
4193 
4194         hlist_del_init(&trans->r_list);
4195 
4196         error = xlog_recover_reorder_trans(log, trans, pass);
4197         if (error)
4198                 return error;
4199 
4200         list_for_each_entry_safe(item, next, &trans->r_itemq, ri_list) {
4201                 switch (pass) {
4202                 case XLOG_RECOVER_PASS1:
4203                         error = xlog_recover_commit_pass1(log, trans, item);
4204                         break;
4205                 case XLOG_RECOVER_PASS2:
4206                         xlog_recover_ra_pass2(log, item);
4207                         list_move_tail(&item->ri_list, &ra_list);
4208                         items_queued++;
4209                         if (items_queued >= XLOG_RECOVER_COMMIT_QUEUE_MAX) {
4210                                 error = xlog_recover_items_pass2(log, trans,
4211                                                 buffer_list, &ra_list);
4212                                 list_splice_tail_init(&ra_list, &done_list);
4213                                 items_queued = 0;
4214                         }
4215 
4216                         break;
4217                 default:
4218                         ASSERT(0);
4219                 }
4220 
4221                 if (error)
4222                         goto out;
4223         }
4224 
4225 out:
4226         if (!list_empty(&ra_list)) {
4227                 if (!error)
4228                         error = xlog_recover_items_pass2(log, trans,
4229                                         buffer_list, &ra_list);
4230                 list_splice_tail_init(&ra_list, &done_list);
4231         }
4232 
4233         if (!list_empty(&done_list))
4234                 list_splice_init(&done_list, &trans->r_itemq);
4235 
4236         return error;
4237 }
4238 
4239 STATIC void
4240 xlog_recover_add_item(
4241         struct list_head        *head)
4242 {
4243         xlog_recover_item_t     *item;
4244 
4245         item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
4246         INIT_LIST_HEAD(&item->ri_list);
4247         list_add_tail(&item->ri_list, head);
4248 }
4249 
4250 STATIC int
4251 xlog_recover_add_to_cont_trans(
4252         struct xlog             *log,
4253         struct xlog_recover     *trans,
4254         char                    *dp,
4255         int                     len)
4256 {
4257         xlog_recover_item_t     *item;
4258         char                    *ptr, *old_ptr;
4259         int                     old_len;
4260 
4261         /*
4262          * If the transaction is empty, the header was split across this and the
4263          * previous record. Copy the rest of the header.
4264          */
4265         if (list_empty(&trans->r_itemq)) {
4266                 ASSERT(len <= sizeof(struct xfs_trans_header));
4267                 if (len > sizeof(struct xfs_trans_header)) {
4268                         xfs_warn(log->l_mp, "%s: bad header length", __func__);
4269                         return -EIO;
4270                 }
4271 
4272                 xlog_recover_add_item(&trans->r_itemq);
4273                 ptr = (char *)&trans->r_theader +
4274                                 sizeof(struct xfs_trans_header) - len;
4275                 memcpy(ptr, dp, len);
4276                 return 0;
4277         }
4278 
4279         /* take the tail entry */
4280         item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
4281 
4282         old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
4283         old_len = item->ri_buf[item->ri_cnt-1].i_len;
4284 
4285         ptr = kmem_realloc(old_ptr, len + old_len, KM_SLEEP);
4286         memcpy(&ptr[old_len], dp, len);
4287         item->ri_buf[item->ri_cnt-1].i_len += len;
4288         item->ri_buf[item->ri_cnt-1].i_addr = ptr;
4289         trace_xfs_log_recover_item_add_cont(log, trans, item, 0);
4290         return 0;
4291 }
4292 
4293 /*
4294  * The next region to add is the start of a new region.  It could be
4295  * a whole region or it could be the first part of a new region.  Because
4296  * of this, the assumption here is that the type and size fields of all
4297  * format structures fit into the first 32 bits of the structure.
4298  *
4299  * This works because all regions must be 32 bit aligned.  Therefore, we
4300  * either have both fields or we have neither field.  In the case we have
4301  * neither field, the data part of the region is zero length.  We only have
4302  * a log_op_header and can throw away the header since a new one will appear
4303  * later.  If we have at least 4 bytes, then we can determine how many regions
4304  * will appear in the current log item.
4305  */
4306 STATIC int
4307 xlog_recover_add_to_trans(
4308         struct xlog             *log,
4309         struct xlog_recover     *trans,
4310         char                    *dp,
4311         int                     len)
4312 {
4313         struct xfs_inode_log_format     *in_f;                  /* any will do */
4314         xlog_recover_item_t     *item;
4315         char                    *ptr;
4316 
4317         if (!len)
4318                 return 0;
4319         if (list_empty(&trans->r_itemq)) {
4320                 /* we need to catch log corruptions here */
4321                 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
4322                         xfs_warn(log->l_mp, "%s: bad header magic number",
4323                                 __func__);
4324                         ASSERT(0);
4325                         return -EIO;
4326                 }
4327 
4328                 if (len > sizeof(struct xfs_trans_header)) {
4329                         xfs_warn(log->l_mp, "%s: bad header length", __func__);
4330                         ASSERT(0);
4331                         return -EIO;
4332                 }
4333 
4334                 /*
4335                  * The transaction header can be arbitrarily split across op
4336                  * records. If we don't have the whole thing here, copy what we
4337                  * do have and handle the rest in the next record.
4338                  */
4339                 if (len == sizeof(struct xfs_trans_header))
4340                         xlog_recover_add_item(&trans->r_itemq);
4341                 memcpy(&trans->r_theader, dp, len);
4342                 return 0;
4343         }
4344 
4345         ptr = kmem_alloc(len, KM_SLEEP);
4346         memcpy(ptr, dp, len);
4347         in_f = (struct xfs_inode_log_format *)ptr;
4348 
4349         /* take the tail entry */
4350         item = list_entry(trans->r_itemq.prev, xlog_recover_item_t, ri_list);
4351         if (item->ri_total != 0 &&
4352              item->ri_total == item->ri_cnt) {
4353                 /* tail item is in use, get a new one */
4354                 xlog_recover_add_item(&trans->r_itemq);
4355                 item = list_entry(trans->r_itemq.prev,
4356                                         xlog_recover_item_t, ri_list);
4357         }
4358 
4359         if (item->ri_total == 0) {              /* first region to be added */
4360                 if (in_f->ilf_size == 0 ||
4361                     in_f->ilf_size > XLOG_MAX_REGIONS_IN_ITEM) {
4362                         xfs_warn(log->l_mp,
4363                 "bad number of regions (%d) in inode log format",
4364                                   in_f->ilf_size);
4365                         ASSERT(0);
4366                         kmem_free(ptr);
4367                         return -EIO;
4368                 }
4369 
4370                 item->ri_total = in_f->ilf_size;
4371                 item->ri_buf =
4372                         kmem_zalloc(item->ri_total * sizeof(xfs_log_iovec_t),
4373                                     KM_SLEEP);
4374         }
4375         ASSERT(item->ri_total > item->ri_cnt);
4376         /* Description region is ri_buf[0] */
4377         item->ri_buf[item->ri_cnt].i_addr = ptr;
4378         item->ri_buf[item->ri_cnt].i_len  = len;
4379         item->ri_cnt++;
4380         trace_xfs_log_recover_item_add(log, trans, item, 0);
4381         return 0;
4382 }
4383 
4384 /*
4385  * Free up any resources allocated by the transaction
4386  *
4387  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
4388  */
4389 STATIC void
4390 xlog_recover_free_trans(
4391         struct xlog_recover     *trans)
4392 {
4393         xlog_recover_item_t     *item, *n;
4394         int                     i;
4395 
4396         hlist_del_init(&trans->r_list);
4397 
4398         list_for_each_entry_safe(item, n, &trans->r_itemq, ri_list) {
4399                 /* Free the regions in the item. */
4400                 list_del(&item->ri_list);
4401                 for (i = 0; i < item->ri_cnt; i++)
4402                         kmem_free(item->ri_buf[i].i_addr);
4403                 /* Free the item itself */
4404                 kmem_free(item->ri_buf);
4405                 kmem_free(item);
4406         }
4407         /* Free the transaction recover structure */
4408         kmem_free(trans);
4409 }
4410 
4411 /*
4412  * On error or completion, trans is freed.
4413  */
4414 STATIC int
4415 xlog_recovery_process_trans(
4416         struct xlog             *log,
4417         struct xlog_recover     *trans,
4418         char                    *dp,
4419         unsigned int            len,
4420         unsigned int            flags,
4421         int                     pass,
4422         struct list_head        *buffer_list)
4423 {
4424         int                     error = 0;
4425         bool                    freeit = false;
4426 
4427         /* mask off ophdr transaction container flags */
4428         flags &= ~XLOG_END_TRANS;
4429         if (flags & XLOG_WAS_CONT_TRANS)
4430                 flags &= ~XLOG_CONTINUE_TRANS;
4431 
4432         /*
4433          * Callees must not free the trans structure. We'll decide if we need to
4434          * free it or not based on the operation being done and it's result.
4435          */
4436         switch (flags) {
4437         /* expected flag values */
4438         case 0:
4439         case XLOG_CONTINUE_TRANS:
4440                 error = xlog_recover_add_to_trans(log, trans, dp, len);
4441                 break;
4442         case XLOG_WAS_CONT_TRANS:
4443                 error = xlog_recover_add_to_cont_trans(log, trans, dp, len);
4444                 break;
4445         case XLOG_COMMIT_TRANS:
4446                 error = xlog_recover_commit_trans(log, trans, pass,
4447                                                   buffer_list);
4448                 /* success or fail, we are now done with this transaction. */
4449                 freeit = true;
4450                 break;
4451 
4452         /* unexpected flag values */
4453         case XLOG_UNMOUNT_TRANS:
4454                 /* just skip trans */
4455                 xfs_warn(log->l_mp, "%s: Unmount LR", __func__);
4456                 freeit = true;
4457                 break;
4458         case XLOG_START_TRANS:
4459         default:
4460                 xfs_warn(log->l_mp, "%s: bad flag 0x%x", __func__, flags);
4461                 ASSERT(0);
4462                 error = -EIO;
4463                 break;
4464         }
4465         if (error || freeit)
4466                 xlog_recover_free_trans(trans);
4467         return error;
4468 }
4469 
4470 /*
4471  * Lookup the transaction recovery structure associated with the ID in the
4472  * current ophdr. If the transaction doesn't exist and the start flag is set in
4473  * the ophdr, then allocate a new transaction for future ID matches to find.
4474  * Either way, return what we found during the lookup - an existing transaction
4475  * or nothing.
4476  */
4477 STATIC struct xlog_recover *
4478 xlog_recover_ophdr_to_trans(
4479         struct hlist_head       rhash[],
4480         struct xlog_rec_header  *rhead,
4481         struct xlog_op_header   *ohead)
4482 {
4483         struct xlog_recover     *trans;
4484         xlog_tid_t              tid;
4485         struct hlist_head       *rhp;
4486 
4487         tid = be32_to_cpu(ohead->oh_tid);
4488         rhp = &rhash[XLOG_RHASH(tid)];
4489         hlist_for_each_entry(trans, rhp, r_list) {
4490                 if (trans->r_log_tid == tid)
4491                         return trans;
4492         }
4493 
4494         /*
4495          * skip over non-start transaction headers - we could be
4496          * processing slack space before the next transaction starts
4497          */
4498         if (!(ohead->oh_flags & XLOG_START_TRANS))
4499                 return NULL;
4500 
4501         ASSERT(be32_to_cpu(ohead->oh_len) == 0);
4502 
4503         /*
4504          * This is a new transaction so allocate a new recovery container to
4505          * hold the recovery ops that will follow.
4506          */
4507         trans = kmem_zalloc(sizeof(struct xlog_recover), KM_SLEEP);
4508         trans->r_log_tid = tid;
4509         trans->r_lsn = be64_to_cpu(rhead->h_lsn);
4510         INIT_LIST_HEAD(&trans->r_itemq);
4511         INIT_HLIST_NODE(&trans->r_list);
4512         hlist_add_head(&trans->r_list, rhp);
4513 
4514         /*
4515          * Nothing more to do for this ophdr. Items to be added to this new
4516          * transaction will be in subsequent ophdr containers.
4517          */
4518         return NULL;
4519 }
4520 
4521 STATIC int
4522 xlog_recover_process_ophdr(
4523         struct xlog             *log,
4524         struct hlist_head       rhash[],
4525         struct xlog_rec_header  *rhead,
4526         struct xlog_op_header   *ohead,
4527         char                    *dp,
4528         char                    *end,
4529         int                     pass,
4530         struct list_head        *buffer_list)
4531 {
4532         struct xlog_recover     *trans;
4533         unsigned int            len;
4534         int                     error;
4535 
4536         /* Do we understand who wrote this op? */
4537         if (ohead->oh_clientid != XFS_TRANSACTION &&
4538             ohead->oh_clientid != XFS_LOG) {
4539                 xfs_warn(log->l_mp, "%s: bad clientid 0x%x",
4540                         __func__, ohead->oh_clientid);
4541                 ASSERT(0);
4542                 return -EIO;
4543         }
4544 
4545         /*
4546          * Check the ophdr contains all the data it is supposed to contain.
4547          */
4548         len = be32_to_cpu(ohead->oh_len);
4549         if (dp + len > end) {
4550                 xfs_warn(log->l_mp, "%s: bad length 0x%x", __func__, len);
4551                 WARN_ON(1);
4552                 return -EIO;
4553         }
4554 
4555         trans = xlog_recover_ophdr_to_trans(rhash, rhead, ohead);
4556         if (!trans) {
4557                 /* nothing to do, so skip over this ophdr */
4558                 return 0;
4559         }
4560 
4561         /*
4562          * The recovered buffer queue is drained only once we know that all
4563          * recovery items for the current LSN have been processed. This is
4564          * required because:
4565          *
4566          * - Buffer write submission updates the metadata LSN of the buffer.
4567          * - Log recovery skips items with a metadata LSN >= the current LSN of
4568          *   the recovery item.
4569          * - Separate recovery items against the same metadata buffer can share
4570          *   a current LSN. I.e., consider that the LSN of a recovery item is
4571          *   defined as the starting LSN of the first record in which its
4572          *   transaction appears, that a record can hold multiple transactions,
4573          *   and/or that a transaction can span multiple records.
4574          *
4575          * In other words, we are allowed to submit a buffer from log recovery
4576          * once per current LSN. Otherwise, we may incorrectly skip recovery
4577          * items and cause corruption.
4578          *
4579          * We don't know up front whether buffers are updated multiple times per
4580          * LSN. Therefore, track the current LSN of each commit log record as it
4581          * is processed and drain the queue when it changes. Use commit records
4582          * because they are ordered correctly by the logging code.
4583          */
4584         if (log->l_recovery_lsn != trans->r_lsn &&
4585             ohead->oh_flags & XLOG_COMMIT_TRANS) {
4586                 error = xfs_buf_delwri_submit(buffer_list);
4587                 if (error)
4588                         return error;
4589                 log->l_recovery_lsn = trans->r_lsn;
4590         }
4591 
4592         return xlog_recovery_process_trans(log, trans, dp, len,
4593                                            ohead->oh_flags, pass, buffer_list);
4594 }
4595 
4596 /*
4597  * There are two valid states of the r_state field.  0 indicates that the
4598  * transaction structure is in a normal state.  We have either seen the
4599  * start of the transaction or the last operation we added was not a partial
4600  * operation.  If the last operation we added to the transaction was a
4601  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
4602  *
4603  * NOTE: skip LRs with 0 data length.
4604  */
4605 STATIC int
4606 xlog_recover_process_data(
4607         struct xlog             *log,
4608         struct hlist_head       rhash[],
4609         struct xlog_rec_header  *rhead,
4610         char                    *dp,
4611         int                     pass,
4612         struct list_head        *buffer_list)
4613 {
4614         struct xlog_op_header   *ohead;
4615         char                    *end;
4616         int                     num_logops;
4617         int                     error;
4618 
4619         end = dp + be32_to_cpu(rhead->h_len);
4620         num_logops = be32_to_cpu(rhead->h_num_logops);
4621 
4622         /* check the log format matches our own - else we can't recover */
4623         if (xlog_header_check_recover(log->l_mp, rhead))
4624                 return -EIO;
4625 
4626         trace_xfs_log_recover_record(log, rhead, pass);
4627         while ((dp < end) && num_logops) {
4628 
4629                 ohead = (struct xlog_op_header *)dp;
4630                 dp += sizeof(*ohead);
4631                 ASSERT(dp <= end);
4632 
4633                 /* errors will abort recovery */
4634                 error = xlog_recover_process_ophdr(log, rhash, rhead, ohead,
4635                                                    dp, end, pass, buffer_list);
4636                 if (error)
4637                         return error;
4638 
4639                 dp += be32_to_cpu(ohead->oh_len);
4640                 num_logops--;
4641         }
4642         return 0;
4643 }
4644 
4645 /* Recover the EFI if necessary. */
4646 STATIC int
4647 xlog_recover_process_efi(
4648         struct xfs_mount                *mp,
4649         struct xfs_ail                  *ailp,
4650         struct xfs_log_item             *lip)
4651 {
4652         struct xfs_efi_log_item         *efip;
4653         int                             error;
4654 
4655         /*
4656          * Skip EFIs that we've already processed.
4657          */
4658         efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4659         if (test_bit(XFS_EFI_RECOVERED, &efip->efi_flags))
4660                 return 0;
4661 
4662         spin_unlock(&ailp->xa_lock);
4663         error = xfs_efi_recover(mp, efip);
4664         spin_lock(&ailp->xa_lock);
4665 
4666         return error;
4667 }
4668 
4669 /* Release the EFI since we're cancelling everything. */
4670 STATIC void
4671 xlog_recover_cancel_efi(
4672         struct xfs_mount                *mp,
4673         struct xfs_ail                  *ailp,
4674         struct xfs_log_item             *lip)
4675 {
4676         struct xfs_efi_log_item         *efip;
4677 
4678         efip = container_of(lip, struct xfs_efi_log_item, efi_item);
4679 
4680         spin_unlock(&ailp->xa_lock);
4681         xfs_efi_release(efip);
4682         spin_lock(&ailp->xa_lock);
4683 }
4684 
4685 /* Recover the RUI if necessary. */
4686 STATIC int
4687 xlog_recover_process_rui(
4688         struct xfs_mount                *mp,
4689         struct xfs_ail                  *ailp,
4690         struct xfs_log_item             *lip)
4691 {
4692         struct xfs_rui_log_item         *ruip;
4693         int                             error;
4694 
4695         /*
4696          * Skip RUIs that we've already processed.
4697          */
4698         ruip = container_of(lip, struct xfs_rui_log_item, rui_item);
4699         if (test_bit(XFS_RUI_RECOVERED, &ruip->rui_flags))
4700                 return 0;
4701 
4702         spin_unlock(&ailp->xa_lock);
4703         error = xfs_rui_recover(mp, ruip);
4704         spin_lock(&ailp->xa_lock);
4705 
4706         return error;
4707 }
4708 
4709 /* Release the RUI since we're cancelling everything. */
4710 STATIC void
4711 xlog_recover_cancel_rui(
4712         struct xfs_mount                *mp,
4713         struct xfs_ail                  *ailp,
4714         struct xfs_log_item             *lip)
4715 {
4716         struct xfs_rui_log_item         *ruip;
4717 
4718         ruip = container_of(lip, struct xfs_rui_log_item, rui_item);
4719 
4720         spin_unlock(&ailp->xa_lock);
4721         xfs_rui_release(ruip);
4722         spin_lock(&ailp->xa_lock);
4723 }
4724 
4725 /* Recover the CUI if necessary. */
4726 STATIC int
4727 xlog_recover_process_cui(
4728         struct xfs_mount                *mp,
4729         struct xfs_ail                  *ailp,
4730         struct xfs_log_item             *lip,
4731         struct xfs_defer_ops            *dfops)
4732 {
4733         struct xfs_cui_log_item         *cuip;
4734         int                             error;
4735 
4736         /*
4737          * Skip CUIs that we've already processed.
4738          */
4739         cuip = container_of(lip, struct xfs_cui_log_item, cui_item);
4740         if (test_bit(XFS_CUI_RECOVERED, &cuip->cui_flags))
4741                 return 0;
4742 
4743         spin_unlock(&ailp->xa_lock);
4744         error = xfs_cui_recover(mp, cuip, dfops);
4745         spin_lock(&ailp->xa_lock);
4746 
4747         return error;
4748 }
4749 
4750 /* Release the CUI since we're cancelling everything. */
4751 STATIC void
4752 xlog_recover_cancel_cui(
4753         struct xfs_mount                *mp,
4754         struct xfs_ail                  *ailp,
4755         struct xfs_log_item             *lip)
4756 {
4757         struct xfs_cui_log_item         *cuip;
4758 
4759         cuip = container_of(lip, struct xfs_cui_log_item, cui_item);
4760 
4761         spin_unlock(&ailp->xa_lock);
4762         xfs_cui_release(cuip);
4763         spin_lock(&ailp->xa_lock);
4764 }
4765 
4766 /* Recover the BUI if necessary. */
4767 STATIC int
4768 xlog_recover_process_bui(
4769         struct xfs_mount                *mp,
4770         struct xfs_ail                  *ailp,
4771         struct xfs_log_item             *lip,
4772         struct xfs_defer_ops            *dfops)
4773 {
4774         struct xfs_bui_log_item         *buip;
4775         int                             error;
4776 
4777         /*
4778          * Skip BUIs that we've already processed.
4779          */
4780         buip = container_of(lip, struct xfs_bui_log_item, bui_item);
4781         if (test_bit(XFS_BUI_RECOVERED, &buip->bui_flags))
4782                 return 0;
4783 
4784         spin_unlock(&ailp->xa_lock);
4785         error = xfs_bui_recover(mp, buip, dfops);
4786         spin_lock(&ailp->xa_lock);
4787 
4788         return error;
4789 }
4790 
4791 /* Release the BUI since we're cancelling everything. */
4792 STATIC void
4793 xlog_recover_cancel_bui(
4794         struct xfs_mount                *mp,
4795         struct xfs_ail                  *ailp,
4796         struct xfs_log_item             *lip)
4797 {
4798         struct xfs_bui_log_item         *buip;
4799 
4800         buip = container_of(lip, struct xfs_bui_log_item, bui_item);
4801 
4802         spin_unlock(&ailp->xa_lock);
4803         xfs_bui_release(buip);
4804         spin_lock(&ailp->xa_lock);
4805 }
4806 
4807 /* Is this log item a deferred action intent? */
4808 static inline bool xlog_item_is_intent(struct xfs_log_item *lip)
4809 {
4810         switch (lip->li_type) {
4811         case XFS_LI_EFI:
4812         case XFS_LI_RUI:
4813         case XFS_LI_CUI:
4814         case XFS_LI_BUI:
4815                 return true;
4816         default:
4817                 return false;
4818         }
4819 }
4820 
4821 /* Take all the collected deferred ops and finish them in order. */
4822 static int
4823 xlog_finish_defer_ops(
4824         struct xfs_mount        *mp,
4825         struct xfs_defer_ops    *dfops)
4826 {
4827         struct xfs_trans        *tp;
4828         int64_t                 freeblks;
4829         uint                    resblks;
4830         int                     error;
4831 
4832         /*
4833          * We're finishing the defer_ops that accumulated as a result of
4834          * recovering unfinished intent items during log recovery.  We
4835          * reserve an itruncate transaction because it is the largest
4836          * permanent transaction type.  Since we're the only user of the fs
4837          * right now, take 93% (15/16) of the available free blocks.  Use
4838          * weird math to avoid a 64-bit division.
4839          */
4840         freeblks = percpu_counter_sum(&mp->m_fdblocks);
4841         if (freeblks <= 0)
4842                 return -ENOSPC;
4843         resblks = min_t(int64_t, UINT_MAX, freeblks);
4844         resblks = (resblks * 15) >> 4;
4845         error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, resblks,
4846                         0, XFS_TRANS_RESERVE, &tp);
4847         if (error)
4848                 return error;
4849 
4850         error = xfs_defer_finish(&tp, dfops);
4851         if (error)
4852                 goto out_cancel;
4853 
4854         return xfs_trans_commit(tp);
4855 
4856 out_cancel:
4857         xfs_trans_cancel(tp);
4858         return error;
4859 }
4860 
4861 /*
4862  * When this is called, all of the log intent items which did not have
4863  * corresponding log done items should be in the AIL.  What we do now
4864  * is update the data structures associated with each one.
4865  *
4866  * Since we process the log intent items in normal transactions, they
4867  * will be removed at some point after the commit.  This prevents us
4868  * from just walking down the list processing each one.  We'll use a
4869  * flag in the intent item to skip those that we've already processed
4870  * and use the AIL iteration mechanism's generation count to try to
4871  * speed this up at least a bit.
4872  *
4873  * When we start, we know that the intents are the only things in the
4874  * AIL.  As we process them, however, other items are added to the
4875  * AIL.
4876  */
4877 STATIC int
4878 xlog_recover_process_intents(
4879         struct xlog             *log)
4880 {
4881         struct xfs_defer_ops    dfops;
4882         struct xfs_ail_cursor   cur;
4883         struct xfs_log_item     *lip;
4884         struct xfs_ail          *ailp;
4885         xfs_fsblock_t           firstfsb;
4886         int                     error = 0;
4887 #if defined(DEBUG) || defined(XFS_WARN)
4888         xfs_lsn_t               last_lsn;
4889 #endif
4890 
4891         ailp = log->l_ailp;
4892         spin_lock(&ailp->xa_lock);
4893         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4894 #if defined(DEBUG) || defined(XFS_WARN)
4895         last_lsn = xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block);
4896 #endif
4897         xfs_defer_init(&dfops, &firstfsb);
4898         while (lip != NULL) {
4899                 /*
4900                  * We're done when we see something other than an intent.
4901                  * There should be no intents left in the AIL now.
4902                  */
4903                 if (!xlog_item_is_intent(lip)) {
4904 #ifdef DEBUG
4905                         for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4906                                 ASSERT(!xlog_item_is_intent(lip));
4907 #endif
4908                         break;
4909                 }
4910 
4911                 /*
4912                  * We should never see a redo item with a LSN higher than
4913                  * the last transaction we found in the log at the start
4914                  * of recovery.
4915                  */
4916                 ASSERT(XFS_LSN_CMP(last_lsn, lip->li_lsn) >= 0);
4917 
4918                 /*
4919                  * NOTE: If your intent processing routine can create more
4920                  * deferred ops, you /must/ attach them to the dfops in this
4921                  * routine or else those subsequent intents will get
4922                  * replayed in the wrong order!
4923                  */
4924                 switch (lip->li_type) {
4925                 case XFS_LI_EFI:
4926                         error = xlog_recover_process_efi(log->l_mp, ailp, lip);
4927                         break;
4928                 case XFS_LI_RUI:
4929                         error = xlog_recover_process_rui(log->l_mp, ailp, lip);
4930                         break;
4931                 case XFS_LI_CUI:
4932                         error = xlog_recover_process_cui(log->l_mp, ailp, lip,
4933                                         &dfops);
4934                         break;
4935                 case XFS_LI_BUI:
4936                         error = xlog_recover_process_bui(log->l_mp, ailp, lip,
4937                                         &dfops);
4938                         break;
4939                 }
4940                 if (error)
4941                         goto out;
4942                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
4943         }
4944 out:
4945         xfs_trans_ail_cursor_done(&cur);
4946         spin_unlock(&ailp->xa_lock);
4947         if (error)
4948                 xfs_defer_cancel(&dfops);
4949         else
4950                 error = xlog_finish_defer_ops(log->l_mp, &dfops);
4951 
4952         return error;
4953 }
4954 
4955 /*
4956  * A cancel occurs when the mount has failed and we're bailing out.
4957  * Release all pending log intent items so they don't pin the AIL.
4958  */
4959 STATIC int
4960 xlog_recover_cancel_intents(
4961         struct xlog             *log)
4962 {
4963         struct xfs_log_item     *lip;
4964         int                     error = 0;
4965         struct xfs_ail_cursor   cur;
4966         struct xfs_ail          *ailp;
4967 
4968         ailp = log->l_ailp;
4969         spin_lock(&ailp->xa_lock);
4970         lip = xfs_trans_ail_cursor_first(ailp, &cur, 0);
4971         while (lip != NULL) {
4972                 /*
4973                  * We're done when we see something other than an intent.
4974                  * There should be no intents left in the AIL now.
4975                  */
4976                 if (!xlog_item_is_intent(lip)) {
4977 #ifdef DEBUG
4978                         for (; lip; lip = xfs_trans_ail_cursor_next(ailp, &cur))
4979                                 ASSERT(!xlog_item_is_intent(lip));
4980 #endif
4981                         break;
4982                 }
4983 
4984                 switch (lip->li_type) {
4985                 case XFS_LI_EFI:
4986                         xlog_recover_cancel_efi(log->l_mp, ailp, lip);
4987                         break;
4988                 case XFS_LI_RUI:
4989                         xlog_recover_cancel_rui(log->l_mp, ailp, lip);
4990                         break;
4991                 case XFS_LI_CUI:
4992                         xlog_recover_cancel_cui(log->l_mp, ailp, lip);
4993                         break;
4994                 case XFS_LI_BUI:
4995                         xlog_recover_cancel_bui(log->l_mp, ailp, lip);
4996                         break;
4997                 }
4998 
4999                 lip = xfs_trans_ail_cursor_next(ailp, &cur);
5000         }
5001 
5002         xfs_trans_ail_cursor_done(&cur);
5003         spin_unlock(&ailp->xa_lock);
5004         return error;
5005 }
5006 
5007 /*
5008  * This routine performs a transaction to null out a bad inode pointer
5009  * in an agi unlinked inode hash bucket.
5010  */
5011 STATIC void
5012 xlog_recover_clear_agi_bucket(
5013         xfs_mount_t     *mp,
5014         xfs_agnumber_t  agno,
5015         int             bucket)
5016 {
5017         xfs_trans_t     *tp;
5018         xfs_agi_t       *agi;
5019         xfs_buf_t       *agibp;
5020         int             offset;
5021         int             error;
5022 
5023         error = xfs_trans_alloc(mp, &M_RES(mp)->tr_clearagi, 0, 0, 0, &tp);
5024         if (error)
5025                 goto out_error;
5026 
5027         error = xfs_read_agi(mp, tp, agno, &agibp);
5028         if (error)
5029                 goto out_abort;
5030 
5031         agi = XFS_BUF_TO_AGI(agibp);
5032         agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
5033         offset = offsetof(xfs_agi_t, agi_unlinked) +
5034                  (sizeof(xfs_agino_t) * bucket);
5035         xfs_trans_log_buf(tp, agibp, offset,
5036                           (offset + sizeof(xfs_agino_t) - 1));
5037 
5038         error = xfs_trans_commit(tp);
5039         if (error)
5040                 goto out_error;
5041         return;
5042 
5043 out_abort:
5044         xfs_trans_cancel(tp);
5045 out_error:
5046         xfs_warn(mp, "%s: failed to clear agi %d. Continuing.", __func__, agno);
5047         return;
5048 }
5049 
5050 STATIC xfs_agino_t
5051 xlog_recover_process_one_iunlink(
5052         struct xfs_mount                *mp,
5053         xfs_agnumber_t                  agno,
5054         xfs_agino_t                     agino,
5055         int                             bucket)
5056 {
5057         struct xfs_buf                  *ibp;
5058         struct xfs_dinode               *dip;
5059         struct xfs_inode                *ip;
5060         xfs_ino_t                       ino;
5061         int                             error;
5062 
5063         ino = XFS_AGINO_TO_INO(mp, agno, agino);
5064         error = xfs_iget(mp, NULL, ino, 0, 0, &ip);
5065         if (error)
5066                 goto fail;
5067 
5068         /*
5069          * Get the on disk inode to find the next inode in the bucket.
5070          */
5071         error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &ibp, 0, 0);
5072         if (error)
5073                 goto fail_iput;
5074 
5075         xfs_iflags_clear(ip, XFS_IRECOVERY);
5076         ASSERT(VFS_I(ip)->i_nlink == 0);
5077         ASSERT(VFS_I(ip)->i_mode != 0);
5078 
5079         /* setup for the next pass */
5080         agino = be32_to_cpu(dip->di_next_unlinked);
5081         xfs_buf_relse(ibp);
5082 
5083         /*
5084          * Prevent any DMAPI event from being sent when the reference on
5085          * the inode is dropped.
5086          */
5087         ip->i_d.di_dmevmask = 0;
5088 
5089         IRELE(ip);
5090         return agino;
5091 
5092  fail_iput:
5093         IRELE(ip);
5094  fail:
5095         /*
5096          * We can't read in the inode this bucket points to, or this inode
5097          * is messed up.  Just ditch this bucket of inodes.  We will lose
5098          * some inodes and space, but at least we won't hang.
5099          *
5100          * Call xlog_recover_clear_agi_bucket() to perform a transaction to
5101          * clear the inode pointer in the bucket.
5102          */
5103         xlog_recover_clear_agi_bucket(mp, agno, bucket);
5104         return NULLAGINO;
5105 }
5106 
5107 /*
5108  * xlog_iunlink_recover
5109  *
5110  * This is called during recovery to process any inodes which
5111  * we unlinked but not freed when the system crashed.  These
5112  * inodes will be on the lists in the AGI blocks.  What we do
5113  * here is scan all the AGIs and fully truncate and free any
5114  * inodes found on the lists.  Each inode is removed from the
5115  * lists when it has been fully truncated and is freed.  The
5116  * freeing of the inode and its removal from the list must be
5117  * atomic.
5118  */
5119 STATIC void
5120 xlog_recover_process_iunlinks(
5121         struct xlog     *log)
5122 {
5123         xfs_mount_t     *mp;
5124         xfs_agnumber_t  agno;
5125         xfs_agi_t       *agi;
5126         xfs_buf_t       *agibp;
5127         xfs_agino_t     agino;
5128         int             bucket;
5129         int             error;
5130         uint            mp_dmevmask;
5131 
5132         mp = log->l_mp;
5133 
5134         /*
5135          * Prevent any DMAPI event from being sent while in this function.
5136          */
5137         mp_dmevmask = mp->m_dmevmask;
5138         mp->m_dmevmask = 0;
5139 
5140         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
5141                 /*
5142                  * Find the agi for this ag.
5143                  */
5144                 error = xfs_read_agi(mp, NULL, agno, &agibp);
5145                 if (error) {
5146                         /*
5147                          * AGI is b0rked. Don't process it.
5148                          *
5149                          * We should probably mark the filesystem as corrupt
5150                          * after we've recovered all the ag's we can....
5151                          */
5152                         continue;
5153                 }
5154                 /*
5155                  * Unlock the buffer so that it can be acquired in the normal
5156                  * course of the transaction to truncate and free each inode.
5157                  * Because we are not racing with anyone else here for the AGI
5158                  * buffer, we don't even need to hold it locked to read the
5159                  * initial unlinked bucket entries out of the buffer. We keep
5160                  * buffer reference though, so that it stays pinned in memory
5161                  * while we need the buffer.
5162                  */
5163                 agi = XFS_BUF_TO_AGI(agibp);
5164                 xfs_buf_unlock(agibp);
5165 
5166                 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
5167                         agino = be32_to_cpu(agi->agi_unlinked[bucket]);
5168                         while (agino != NULLAGINO) {
5169                                 agino = xlog_recover_process_one_iunlink(mp,
5170                                                         agno, agino, bucket);
5171                         }
5172                 }
5173                 xfs_buf_rele(agibp);
5174         }
5175 
5176         mp->m_dmevmask = mp_dmevmask;
5177 }
5178 
5179 STATIC int
5180 xlog_unpack_data(
5181         struct xlog_rec_header  *rhead,
5182         char                    *dp,
5183         struct xlog             *log)
5184 {
5185         int                     i, j, k;
5186 
5187         for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
5188                   i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
5189                 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
5190                 dp += BBSIZE;
5191         }
5192 
5193         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
5194                 xlog_in_core_2_t *xhdr = (xlog_in_core_2_t *)rhead;
5195                 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
5196                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
5197                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
5198                         *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
5199                         dp += BBSIZE;
5200                 }
5201         }
5202 
5203         return 0;
5204 }
5205 
5206 /*
5207  * CRC check, unpack and process a log record.
5208  */
5209 STATIC int
5210 xlog_recover_process(
5211         struct xlog             *log,
5212         struct hlist_head       rhash[],
5213         struct xlog_rec_header  *rhead,
5214         char                    *dp,
5215         int                     pass,
5216         struct list_head        *buffer_list)
5217 {
5218         int                     error;
5219         __le32                  old_crc = rhead->h_crc;
5220         __le32                  crc;
5221 
5222 
5223         crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
5224 
5225         /*
5226          * Nothing else to do if this is a CRC verification pass. Just return
5227          * if this a record with a non-zero crc. Unfortunately, mkfs always
5228          * sets old_crc to 0 so we must consider this valid even on v5 supers.
5229          * Otherwise, return EFSBADCRC on failure so the callers up the stack
5230          * know precisely what failed.
5231          */
5232         if (pass == XLOG_RECOVER_CRCPASS) {
5233                 if (old_crc && crc != old_crc)
5234                         return -EFSBADCRC;
5235                 return 0;
5236         }
5237 
5238         /*
5239          * We're in the normal recovery path. Issue a warning if and only if the
5240          * CRC in the header is non-zero. This is an advisory warning and the
5241          * zero CRC check prevents warnings from being emitted when upgrading
5242          * the kernel from one that does not add CRCs by default.
5243          */
5244         if (crc != old_crc) {
5245                 if (old_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
5246                         xfs_alert(log->l_mp,
5247                 "log record CRC mismatch: found 0x%x, expected 0x%x.",
5248                                         le32_to_cpu(old_crc),
5249                                         le32_to_cpu(crc));
5250                         xfs_hex_dump(dp, 32);
5251                 }
5252 
5253                 /*
5254                  * If the filesystem is CRC enabled, this mismatch becomes a
5255                  * fatal log corruption failure.
5256                  */
5257                 if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
5258                         return -EFSCORRUPTED;
5259         }
5260 
5261         error = xlog_unpack_data(rhead, dp, log);
5262         if (error)
5263                 return error;
5264 
5265         return xlog_recover_process_data(log, rhash, rhead, dp, pass,
5266                                          buffer_list);
5267 }
5268 
5269 STATIC int
5270 xlog_valid_rec_header(
5271         struct xlog             *log,
5272         struct xlog_rec_header  *rhead,
5273         xfs_daddr_t             blkno)
5274 {
5275         int                     hlen;
5276 
5277         if (unlikely(rhead->h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))) {
5278                 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
5279                                 XFS_ERRLEVEL_LOW, log->l_mp);
5280                 return -EFSCORRUPTED;
5281         }
5282         if (unlikely(
5283             (!rhead->h_version ||
5284             (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
5285                 xfs_warn(log->l_mp, "%s: unrecognised log version (%d).",
5286                         __func__, be32_to_cpu(rhead->h_version));
5287                 return -EIO;
5288         }
5289 
5290         /* LR body must have data or it wouldn't have been written */
5291         hlen = be32_to_cpu(rhead->h_len);
5292         if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
5293                 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
5294                                 XFS_ERRLEVEL_LOW, log->l_mp);
5295                 return -EFSCORRUPTED;
5296         }
5297         if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
5298                 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
5299                                 XFS_ERRLEVEL_LOW, log->l_mp);
5300                 return -EFSCORRUPTED;
5301         }
5302         return 0;
5303 }
5304 
5305 /*
5306  * Read the log from tail to head and process the log records found.
5307  * Handle the two cases where the tail and head are in the same cycle
5308  * and where the active portion of the log wraps around the end of
5309  * the physical log separately.  The pass parameter is passed through
5310  * to the routines called to process the data and is not looked at
5311  * here.
5312  */
5313 STATIC int
5314 xlog_do_recovery_pass(
5315         struct xlog             *log,
5316         xfs_daddr_t             head_blk,
5317         xfs_daddr_t             tail_blk,
5318         int                     pass,
5319         xfs_daddr_t             *first_bad)     /* out: first bad log rec */
5320 {
5321         xlog_rec_header_t       *rhead;
5322         xfs_daddr_t             blk_no, rblk_no;
5323         xfs_daddr_t             rhead_blk;
5324         char                    *offset;
5325         xfs_buf_t               *hbp, *dbp;
5326         int                     error = 0, h_size, h_len;
5327         int                     error2 = 0;
5328         int                     bblks, split_bblks;
5329         int                     hblks, split_hblks, wrapped_hblks;
5330         int                     i;
5331         struct hlist_head       rhash[XLOG_RHASH_SIZE];
5332         LIST_HEAD               (buffer_list);
5333 
5334         ASSERT(head_blk != tail_blk);
5335         blk_no = rhead_blk = tail_blk;
5336 
5337         for (i = 0; i < XLOG_RHASH_SIZE; i++)
5338                 INIT_HLIST_HEAD(&rhash[i]);
5339 
5340         /*
5341          * Read the header of the tail block and get the iclog buffer size from
5342          * h_size.  Use this to tell how many sectors make up the log header.
5343          */
5344         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
5345                 /*
5346                  * When using variable length iclogs, read first sector of
5347                  * iclog header and extract the header size from it.  Get a
5348                  * new hbp that is the correct size.
5349                  */
5350                 hbp = xlog_get_bp(log, 1);
5351                 if (!hbp)
5352                         return -ENOMEM;
5353 
5354                 error = xlog_bread(log, tail_blk, 1, hbp, &offset);
5355                 if (error)
5356                         goto bread_err1;
5357 
5358                 rhead = (xlog_rec_header_t *)offset;
5359                 error = xlog_valid_rec_header(log, rhead, tail_blk);
5360                 if (error)
5361                         goto bread_err1;
5362 
5363                 /*
5364                  * xfsprogs has a bug where record length is based on lsunit but
5365                  * h_size (iclog size) is hardcoded to 32k. Now that we
5366                  * unconditionally CRC verify the unmount record, this means the
5367                  * log buffer can be too small for the record and cause an
5368                  * overrun.
5369                  *
5370                  * Detect this condition here. Use lsunit for the buffer size as
5371                  * long as this looks like the mkfs case. Otherwise, return an
5372                  * error to avoid a buffer overrun.
5373                  */
5374                 h_size = be32_to_cpu(rhead->h_size);
5375                 h_len = be32_to_cpu(rhead->h_len);
5376                 if (h_len > h_size) {
5377                         if (h_len <= log->l_mp->m_logbsize &&
5378                             be32_to_cpu(rhead->h_num_logops) == 1) {
5379                                 xfs_warn(log->l_mp,
5380                 "invalid iclog size (%d bytes), using lsunit (%d bytes)",
5381                                          h_size, log->l_mp->m_logbsize);
5382                                 h_size = log->l_mp->m_logbsize;
5383                         } else
5384                                 return -EFSCORRUPTED;
5385                 }
5386 
5387                 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
5388                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
5389                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
5390                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
5391                                 hblks++;
5392                         xlog_put_bp(hbp);
5393                         hbp = xlog_get_bp(log, hblks);
5394                 } else {
5395                         hblks = 1;
5396                 }
5397         } else {
5398                 ASSERT(log->l_sectBBsize == 1);
5399                 hblks = 1;
5400                 hbp = xlog_get_bp(log, 1);
5401                 h_size = XLOG_BIG_RECORD_BSIZE;
5402         }
5403 
5404         if (!hbp)
5405                 return -ENOMEM;
5406         dbp = xlog_get_bp(log, BTOBB(h_size));
5407         if (!dbp) {
5408                 xlog_put_bp(hbp);
5409                 return -ENOMEM;
5410         }
5411 
5412         memset(rhash, 0, sizeof(rhash));
5413         if (tail_blk > head_blk) {
5414                 /*
5415                  * Perform recovery around the end of the physical log.
5416                  * When the head is not on the same cycle number as the tail,
5417                  * we can't do a sequential recovery.
5418                  */
5419                 while (blk_no < log->l_logBBsize) {
5420                         /*
5421                          * Check for header wrapping around physical end-of-log
5422                          */
5423                         offset = hbp->b_addr;
5424                         split_hblks = 0;
5425                         wrapped_hblks = 0;
5426                         if (blk_no + hblks <= log->l_logBBsize) {
5427                                 /* Read header in one read */
5428                                 error = xlog_bread(log, blk_no, hblks, hbp,
5429                                                    &offset);
5430                                 if (error)
5431                                         goto bread_err2;
5432                         } else {
5433                                 /* This LR is split across physical log end */
5434                                 if (blk_no != log->l_logBBsize) {
5435                                         /* some data before physical log end */
5436                                         ASSERT(blk_no <= INT_MAX);
5437                                         split_hblks = log->l_logBBsize - (int)blk_no;
5438                                         ASSERT(split_hblks > 0);
5439                                         error = xlog_bread(log, blk_no,
5440                                                            split_hblks, hbp,
5441                                                            &offset);
5442                                         if (error)
5443                                                 goto bread_err2;
5444                                 }
5445 
5446                                 /*
5447                                  * Note: this black magic still works with
5448                                  * large sector sizes (non-512) only because:
5449                                  * - we increased the buffer size originally
5450                                  *   by 1 sector giving us enough extra space
5451                                  *   for the second read;
5452                                  * - the log start is guaranteed to be sector
5453                                  *   aligned;
5454                                  * - we read the log end (LR header start)
5455                                  *   _first_, then the log start (LR header end)
5456                                  *   - order is important.
5457                                  */
5458                                 wrapped_hblks = hblks - split_hblks;
5459                                 error = xlog_bread_offset(log, 0,
5460                                                 wrapped_hblks, hbp,
5461                                                 offset + BBTOB(split_hblks));
5462                                 if (error)
5463                                         goto bread_err2;
5464                         }
5465                         rhead = (xlog_rec_header_t *)offset;
5466                         error = xlog_valid_rec_header(log, rhead,
5467                                                 split_hblks ? blk_no : 0);
5468                         if (error)
5469                                 goto bread_err2;
5470 
5471                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
5472                         blk_no += hblks;
5473 
5474                         /*
5475                          * Read the log record data in multiple reads if it
5476                          * wraps around the end of the log. Note that if the
5477                          * header already wrapped, blk_no could point past the
5478                          * end of the log. The record data is contiguous in
5479                          * that case.
5480                          */
5481                         if (blk_no + bblks <= log->l_logBBsize ||
5482                             blk_no >= log->l_logBBsize) {
5483                                 /* mod blk_no in case the header wrapped and
5484                                  * pushed it beyond the end of the log */
5485                                 rblk_no = do_mod(blk_no, log->l_logBBsize);
5486                                 error = xlog_bread(log, rblk_no, bblks, dbp,
5487                                                    &offset);
5488                                 if (error)
5489                                         goto bread_err2;
5490                         } else {
5491                                 /* This log record is split across the
5492                                  * physical end of log */
5493                                 offset = dbp->b_addr;
5494                                 split_bblks = 0;
5495                                 if (blk_no != log->l_logBBsize) {
5496                                         /* some data is before the physical
5497                                          * end of log */
5498                                         ASSERT(!wrapped_hblks);
5499                                         ASSERT(blk_no <= INT_MAX);
5500                                         split_bblks =
5501                                                 log->l_logBBsize - (int)blk_no;
5502                                         ASSERT(split_bblks > 0);
5503                                         error = xlog_bread(log, blk_no,
5504                                                         split_bblks, dbp,
5505                                                         &offset);
5506                                         if (error)
5507                                                 goto bread_err2;
5508                                 }
5509 
5510                                 /*
5511                                  * Note: this black magic still works with
5512                                  * large sector sizes (non-512) only because:
5513                                  * - we increased the buffer size originally
5514                                  *   by 1 sector giving us enough extra space
5515                                  *   for the second read;
5516                                  * - the log start is guaranteed to be sector
5517                                  *   aligned;
5518                                  * - we read the log end (LR header start)
5519                                  *   _first_, then the log start (LR header end)
5520                                  *   - order is important.
5521                                  */
5522                                 error = xlog_bread_offset(log, 0,
5523                                                 bblks - split_bblks, dbp,
5524                                                 offset + BBTOB(split_bblks));
5525                                 if (error)
5526                                         goto bread_err2;
5527                         }
5528 
5529                         error = xlog_recover_process(log, rhash, rhead, offset,
5530                                                      pass, &buffer_list);
5531                         if (error)
5532                                 goto bread_err2;
5533 
5534                         blk_no += bblks;
5535                         rhead_blk = blk_no;
5536                 }
5537 
5538                 ASSERT(blk_no >= log->l_logBBsize);
5539                 blk_no -= log->l_logBBsize;
5540                 rhead_blk = blk_no;
5541         }
5542 
5543         /* read first part of physical log */
5544         while (blk_no < head_blk) {
5545                 error = xlog_bread(log, blk_no, hblks, hbp, &offset);
5546                 if (error)
5547                         goto bread_err2;
5548 
5549                 rhead = (xlog_rec_header_t *)offset;
5550                 error = xlog_valid_rec_header(log, rhead, blk_no);
5551                 if (error)
5552                         goto bread_err2;
5553 
5554                 /* blocks in data section */
5555                 bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
5556                 error = xlog_bread(log, blk_no+hblks, bblks, dbp,
5557                                    &offset);
5558                 if (error)
5559                         goto bread_err2;
5560 
5561                 error = xlog_recover_process(log, rhash, rhead, offset, pass,
5562                                              &buffer_list);
5563                 if (error)
5564                         goto bread_err2;
5565 
5566                 blk_no += bblks + hblks;
5567                 rhead_blk = blk_no;
5568         }
5569 
5570  bread_err2:
5571         xlog_put_bp(dbp);
5572  bread_err1:
5573         xlog_put_bp(hbp);
5574 
5575         /*
5576          * Submit buffers that have been added from the last record processed,
5577          * regardless of error status.
5578          */
5579         if (!list_empty(&buffer_list))
5580                 error2 = xfs_buf_delwri_submit(&buffer_list);
5581 
5582         if (error && first_bad)
5583                 *first_bad = rhead_blk;
5584 
5585         /*
5586          * Transactions are freed at commit time but transactions without commit
5587          * records on disk are never committed. Free any that may be left in the
5588          * hash table.
5589          */
5590         for (i = 0; i < XLOG_RHASH_SIZE; i++) {
5591                 struct hlist_node       *tmp;
5592                 struct xlog_recover     *trans;
5593 
5594                 hlist_for_each_entry_safe(trans, tmp, &rhash[i], r_list)
5595                         xlog_recover_free_trans(trans);
5596         }
5597 
5598         return error ? error : error2;
5599 }
5600 
5601 /*
5602  * Do the recovery of the log.  We actually do this in two phases.
5603  * The two passes are necessary in order to implement the function
5604  * of cancelling a record written into the log.  The first pass
5605  * determines those things which have been cancelled, and the
5606  * second pass replays log items normally except for those which
5607  * have been cancelled.  The handling of the replay and cancellations
5608  * takes place in the log item type specific routines.
5609  *
5610  * The table of items which have cancel records in the log is allocated
5611  * and freed at this level, since only here do we know when all of
5612  * the log recovery has been completed.
5613  */
5614 STATIC int
5615 xlog_do_log_recovery(
5616         struct xlog     *log,
5617         xfs_daddr_t     head_blk,
5618         xfs_daddr_t     tail_blk)
5619 {
5620         int             error, i;
5621 
5622         ASSERT(head_blk != tail_blk);
5623 
5624         /*
5625          * First do a pass to find all of the cancelled buf log items.
5626          * Store them in the buf_cancel_table for use in the second pass.
5627          */
5628         log->l_buf_cancel_table = kmem_zalloc(XLOG_BC_TABLE_SIZE *
5629                                                  sizeof(struct list_head),
5630                                                  KM_SLEEP);
5631         for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
5632                 INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
5633 
5634         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
5635                                       XLOG_RECOVER_PASS1, NULL);
5636         if (error != 0) {
5637                 kmem_free(log->l_buf_cancel_table);
5638                 log->l_buf_cancel_table = NULL;
5639                 return error;
5640         }
5641         /*
5642          * Then do a second pass to actually recover the items in the log.
5643          * When it is complete free the table of buf cancel items.
5644          */
5645         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
5646                                       XLOG_RECOVER_PASS2, NULL);
5647 #ifdef DEBUG
5648         if (!error) {
5649                 int     i;
5650 
5651                 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
5652                         ASSERT(list_empty(&log->l_buf_cancel_table[i]));
5653         }
5654 #endif  /* DEBUG */
5655 
5656         kmem_free(log->l_buf_cancel_table);
5657         log->l_buf_cancel_table = NULL;
5658 
5659         return error;
5660 }
5661 
5662 /*
5663  * Do the actual recovery
5664  */
5665 STATIC int
5666 xlog_do_recover(
5667         struct xlog     *log,
5668         xfs_daddr_t     head_blk,
5669         xfs_daddr_t     tail_blk)
5670 {
5671         struct xfs_mount *mp = log->l_mp;
5672         int             error;
5673         xfs_buf_t       *bp;
5674         xfs_sb_t        *sbp;
5675 
5676         trace_xfs_log_recover(log, head_blk, tail_blk);
5677 
5678         /*
5679          * First replay the images in the log.
5680          */
5681         error = xlog_do_log_recovery(log, head_blk, tail_blk);
5682         if (error)
5683                 return error;
5684 
5685         /*
5686          * If IO errors happened during recovery, bail out.
5687          */
5688         if (XFS_FORCED_SHUTDOWN(mp)) {
5689                 return -EIO;
5690         }
5691 
5692         /*
5693          * We now update the tail_lsn since much of the recovery has completed
5694          * and there may be space available to use.  If there were no extent
5695          * or iunlinks, we can free up the entire log and set the tail_lsn to
5696          * be the last_sync_lsn.  This was set in xlog_find_tail to be the
5697          * lsn of the last known good LR on disk.  If there are extent frees
5698          * or iunlinks they will have some entries in the AIL; so we look at
5699          * the AIL to determine how to set the tail_lsn.
5700          */
5701         xlog_assign_tail_lsn(mp);
5702 
5703         /*
5704          * Now that we've finished replaying all buffer and inode
5705          * updates, re-read in the superblock and reverify it.
5706          */
5707         bp = xfs_getsb(mp, 0);
5708         bp->b_flags &= ~(XBF_DONE | XBF_ASYNC);
5709         ASSERT(!(bp->b_flags & XBF_WRITE));
5710         bp->b_flags |= XBF_READ;
5711         bp->b_ops = &xfs_sb_buf_ops;
5712 
5713         error = xfs_buf_submit_wait(bp);
5714         if (error) {
5715                 if (!XFS_FORCED_SHUTDOWN(mp)) {
5716                         xfs_buf_ioerror_alert(bp, __func__);
5717                         ASSERT(0);
5718                 }
5719                 xfs_buf_relse(bp);
5720                 return error;
5721         }
5722 
5723         /* Convert superblock from on-disk format */
5724         sbp = &mp->m_sb;
5725         xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
5726         xfs_buf_relse(bp);
5727 
5728         /* re-initialise in-core superblock and geometry structures */
5729         xfs_reinit_percpu_counters(mp);
5730         error = xfs_initialize_perag(mp, sbp->sb_agcount, &mp->m_maxagi);
5731         if (error) {
5732                 xfs_warn(mp, "Failed post-recovery per-ag init: %d", error);
5733                 return error;
5734         }
5735         mp->m_alloc_set_aside = xfs_alloc_set_aside(mp);
5736 
5737         xlog_recover_check_summary(log);
5738 
5739         /* Normal transactions can now occur */
5740         log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
5741         return 0;
5742 }
5743 
5744 /*
5745  * Perform recovery and re-initialize some log variables in xlog_find_tail.
5746  *
5747  * Return error or zero.
5748  */
5749 int
5750 xlog_recover(
5751         struct xlog     *log)
5752 {
5753         xfs_daddr_t     head_blk, tail_blk;
5754         int             error;
5755 
5756         /* find the tail of the log */
5757         error = xlog_find_tail(log, &head_blk, &tail_blk);
5758         if (error)
5759                 return error;
5760 
5761         /*
5762          * The superblock was read before the log was available and thus the LSN
5763          * could not be verified. Check the superblock LSN against the current
5764          * LSN now that it's known.
5765          */
5766         if (xfs_sb_version_hascrc(&log->l_mp->m_sb) &&
5767             !xfs_log_check_lsn(log->l_mp, log->l_mp->m_sb.sb_lsn))
5768                 return -EINVAL;
5769 
5770         if (tail_blk != head_blk) {
5771                 /* There used to be a comment here:
5772                  *
5773                  * disallow recovery on read-only mounts.  note -- mount
5774                  * checks for ENOSPC and turns it into an intelligent
5775                  * error message.
5776                  * ...but this is no longer true.  Now, unless you specify
5777                  * NORECOVERY (in which case this function would never be
5778                  * called), we just go ahead and recover.  We do this all
5779                  * under the vfs layer, so we can get away with it unless
5780                  * the device itself is read-only, in which case we fail.
5781                  */
5782                 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
5783                         return error;
5784                 }
5785 
5786                 /*
5787                  * Version 5 superblock log feature mask validation. We know the
5788                  * log is dirty so check if there are any unknown log features
5789                  * in what we need to recover. If there are unknown features
5790                  * (e.g. unsupported transactions, then simply reject the
5791                  * attempt at recovery before touching anything.
5792                  */
5793                 if (XFS_SB_VERSION_NUM(&log->l_mp->m_sb) == XFS_SB_VERSION_5 &&
5794                     xfs_sb_has_incompat_log_feature(&log->l_mp->m_sb,
5795                                         XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN)) {
5796                         xfs_warn(log->l_mp,
5797 "Superblock has unknown incompatible log features (0x%x) enabled.",
5798                                 (log->l_mp->m_sb.sb_features_log_incompat &
5799                                         XFS_SB_FEAT_INCOMPAT_LOG_UNKNOWN));
5800                         xfs_warn(log->l_mp,
5801 "The log can not be fully and/or safely recovered by this kernel.");
5802                         xfs_warn(log->l_mp,
5803 "Please recover the log on a kernel that supports the unknown features.");
5804                         return -EINVAL;
5805                 }
5806 
5807                 /*
5808                  * Delay log recovery if the debug hook is set. This is debug
5809                  * instrumention to coordinate simulation of I/O failures with
5810                  * log recovery.
5811                  */
5812                 if (xfs_globals.log_recovery_delay) {
5813                         xfs_notice(log->l_mp,
5814                                 "Delaying log recovery for %d seconds.",
5815                                 xfs_globals.log_recovery_delay);
5816                         msleep(xfs_globals.log_recovery_delay * 1000);
5817                 }
5818 
5819                 xfs_notice(log->l_mp, "Starting recovery (logdev: %s)",
5820                                 log->l_mp->m_logname ? log->l_mp->m_logname
5821                                                      : "internal");
5822 
5823                 error = xlog_do_recover(log, head_blk, tail_blk);
5824                 log->l_flags |= XLOG_RECOVERY_NEEDED;
5825         }
5826         return error;
5827 }
5828 
5829 /*
5830  * In the first part of recovery we replay inodes and buffers and build
5831  * up the list of extent free items which need to be processed.  Here
5832  * we process the extent free items and clean up the on disk unlinked
5833  * inode lists.  This is separated from the first part of recovery so
5834  * that the root and real-time bitmap inodes can be read in from disk in
5835  * between the two stages.  This is necessary so that we can free space
5836  * in the real-time portion of the file system.
5837  */
5838 int
5839 xlog_recover_finish(
5840         struct xlog     *log)
5841 {
5842         /*
5843          * Now we're ready to do the transactions needed for the
5844          * rest of recovery.  Start with completing all the extent
5845          * free intent records and then process the unlinked inode
5846          * lists.  At this point, we essentially run in normal mode
5847          * except that we're still performing recovery actions
5848          * rather than accepting new requests.
5849          */
5850         if (log->l_flags & XLOG_RECOVERY_NEEDED) {
5851                 int     error;
5852                 error = xlog_recover_process_intents(log);
5853                 if (error) {
5854                         xfs_alert(log->l_mp, "Failed to recover intents");
5855                         return error;
5856                 }
5857 
5858                 /*
5859                  * Sync the log to get all the intents out of the AIL.
5860                  * This isn't absolutely necessary, but it helps in
5861                  * case the unlink transactions would have problems
5862                  * pushing the intents out of the way.
5863                  */
5864                 xfs_log_force(log->l_mp, XFS_LOG_SYNC);
5865 
5866                 xlog_recover_process_iunlinks(log);
5867 
5868                 xlog_recover_check_summary(log);
5869 
5870                 xfs_notice(log->l_mp, "Ending recovery (logdev: %s)",
5871                                 log->l_mp->m_logname ? log->l_mp->m_logname
5872                                                      : "internal");
5873                 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
5874         } else {
5875                 xfs_info(log->l_mp, "Ending clean mount");
5876         }
5877         return 0;
5878 }
5879 
5880 int
5881 xlog_recover_cancel(
5882         struct xlog     *log)
5883 {
5884         int             error = 0;
5885 
5886         if (log->l_flags & XLOG_RECOVERY_NEEDED)
5887                 error = xlog_recover_cancel_intents(log);
5888 
5889         return error;
5890 }
5891 
5892 #if defined(DEBUG)
5893 /*
5894  * Read all of the agf and agi counters and check that they
5895  * are consistent with the superblock counters.
5896  */
5897 STATIC void
5898 xlog_recover_check_summary(
5899         struct xlog     *log)
5900 {
5901         xfs_mount_t     *mp;
5902         xfs_agf_t       *agfp;
5903         xfs_buf_t       *agfbp;
5904         xfs_buf_t       *agibp;
5905         xfs_agnumber_t  agno;
5906         uint64_t        freeblks;
5907         uint64_t        itotal;
5908         uint64_t        ifree;
5909         int             error;
5910 
5911         mp = log->l_mp;
5912 
5913         freeblks = 0LL;
5914         itotal = 0LL;
5915         ifree = 0LL;
5916         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
5917                 error = xfs_read_agf(mp, NULL, agno, 0, &agfbp);
5918                 if (error) {
5919                         xfs_alert(mp, "%s agf read failed agno %d error %d",
5920                                                 __func__, agno, error);
5921                 } else {
5922                         agfp = XFS_BUF_TO_AGF(agfbp);
5923                         freeblks += be32_to_cpu(agfp->agf_freeblks) +
5924                                     be32_to_cpu(agfp->agf_flcount);
5925                         xfs_buf_relse(agfbp);
5926                 }
5927 
5928                 error = xfs_read_agi(mp, NULL, agno, &agibp);
5929                 if (error) {
5930                         xfs_alert(mp, "%s agi read failed agno %d error %d",
5931                                                 __func__, agno, error);
5932                 } else {
5933                         struct xfs_agi  *agi = XFS_BUF_TO_AGI(agibp);
5934 
5935                         itotal += be32_to_cpu(agi->agi_count);
5936                         ifree += be32_to_cpu(agi->agi_freecount);
5937                         xfs_buf_relse(agibp);
5938                 }
5939         }
5940 }
5941 #endif /* DEBUG */
5942 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp