fs/xfs/xfs_log_recover.c

   1 /*
   2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_types.h"
  21 #include "xfs_bit.h"
  22 #include "xfs_log.h"
  23 #include "xfs_inum.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_mount.h"
  30 #include "xfs_error.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_alloc_btree.h"
  33 #include "xfs_ialloc_btree.h"
  34 #include "xfs_dir2_sf.h"
  35 #include "xfs_attr_sf.h"
  36 #include "xfs_dinode.h"
  37 #include "xfs_inode.h"
  38 #include "xfs_inode_item.h"
  39 #include "xfs_imap.h"
  40 #include "xfs_alloc.h"
  41 #include "xfs_ialloc.h"
  42 #include "xfs_log_priv.h"
  43 #include "xfs_buf_item.h"
  44 #include "xfs_log_recover.h"
  45 #include "xfs_extfree_item.h"
  46 #include "xfs_trans_priv.h"
  47 #include "xfs_quota.h"
  48 #include "xfs_rw.h"
  49 #include "xfs_utils.h"
  50
  51 STATIC int      xlog_find_zeroed(xlog_t *, xfs_daddr_t *);
  52 STATIC int      xlog_clear_stale_blocks(xlog_t *, xfs_lsn_t);
  53 STATIC void     xlog_recover_insert_item_backq(xlog_recover_item_t **q,
  54                                                xlog_recover_item_t *item);
  55 #if defined(DEBUG)
  56 STATIC void     xlog_recover_check_summary(xlog_t *);
  57 STATIC void     xlog_recover_check_ail(xfs_mount_t *, xfs_log_item_t *, int);
  58 #else
  59 #define xlog_recover_check_summary(log)
  60 #define xlog_recover_check_ail(mp, lip, gen)
  61 #endif
  62
  63
  64 /*
  65  * Sector aligned buffer routines for buffer create/read/write/access
  66  */
  67
  68 #define XLOG_SECTOR_ROUNDUP_BBCOUNT(log, bbs)   \
  69         ( ((log)->l_sectbb_mask && (bbs & (log)->l_sectbb_mask)) ? \
  70         ((bbs + (log)->l_sectbb_mask + 1) & ~(log)->l_sectbb_mask) : (bbs) )
  71 #define XLOG_SECTOR_ROUNDDOWN_BLKNO(log, bno)   ((bno) & ~(log)->l_sectbb_mask)
  72
  73 xfs_buf_t *
  74 xlog_get_bp(
  75         xlog_t          *log,
  76         int             num_bblks)
  77 {
  78         ASSERT(num_bblks > 0);
  79
  80         if (log->l_sectbb_log) {
  81                 if (num_bblks > 1)
  82                         num_bblks += XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
  83                 num_bblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, num_bblks);
  84         }
  85         return xfs_buf_get_noaddr(BBTOB(num_bblks), log->l_mp->m_logdev_targp);
  86 }
  87
  88 void
  89 xlog_put_bp(
  90         xfs_buf_t       *bp)
  91 {
  92         xfs_buf_free(bp);
  93 }
  94
  95
  96 /*
  97  * nbblks should be uint, but oh well.  Just want to catch that 32-bit length.
  98  */
  99 int
 100 xlog_bread(
 101         xlog_t          *log,
 102         xfs_daddr_t     blk_no,
 103         int             nbblks,
 104         xfs_buf_t       *bp)
 105 {
 106         int             error;
 107
 108         if (log->l_sectbb_log) {
 109                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
 110                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
 111         }
 112
 113         ASSERT(nbblks > 0);
 114         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
 115         ASSERT(bp);
 116
 117         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
 118         XFS_BUF_READ(bp);
 119         XFS_BUF_BUSY(bp);
 120         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
 121         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
 122
 123         xfsbdstrat(log->l_mp, bp);
 124         error = xfs_iowait(bp);
 125         if (error)
 126                 xfs_ioerror_alert("xlog_bread", log->l_mp,
 127                                   bp, XFS_BUF_ADDR(bp));
 128         return error;
 129 }
 130
 131 /*
 132  * Write out the buffer at the given block for the given number of blocks.
 133  * The buffer is kept locked across the write and is returned locked.
 134  * This can only be used for synchronous log writes.
 135  */
 136 STATIC int
 137 xlog_bwrite(
 138         xlog_t          *log,
 139         xfs_daddr_t     blk_no,
 140         int             nbblks,
 141         xfs_buf_t       *bp)
 142 {
 143         int             error;
 144
 145         if (log->l_sectbb_log) {
 146                 blk_no = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, blk_no);
 147                 nbblks = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, nbblks);
 148         }
 149
 150         ASSERT(nbblks > 0);
 151         ASSERT(BBTOB(nbblks) <= XFS_BUF_SIZE(bp));
 152
 153         XFS_BUF_SET_ADDR(bp, log->l_logBBstart + blk_no);
 154         XFS_BUF_ZEROFLAGS(bp);
 155         XFS_BUF_BUSY(bp);
 156         XFS_BUF_HOLD(bp);
 157         XFS_BUF_PSEMA(bp, PRIBIO);
 158         XFS_BUF_SET_COUNT(bp, BBTOB(nbblks));
 159         XFS_BUF_SET_TARGET(bp, log->l_mp->m_logdev_targp);
 160
 161         if ((error = xfs_bwrite(log->l_mp, bp)))
 162                 xfs_ioerror_alert("xlog_bwrite", log->l_mp,
 163                                   bp, XFS_BUF_ADDR(bp));
 164         return error;
 165 }
 166
 167 STATIC xfs_caddr_t
 168 xlog_align(
 169         xlog_t          *log,
 170         xfs_daddr_t     blk_no,
 171         int             nbblks,
 172         xfs_buf_t       *bp)
 173 {
 174         xfs_caddr_t     ptr;
 175
 176         if (!log->l_sectbb_log)
 177                 return XFS_BUF_PTR(bp);
 178
 179         ptr = XFS_BUF_PTR(bp) + BBTOB((int)blk_no & log->l_sectbb_mask);
 180         ASSERT(XFS_BUF_SIZE(bp) >=
 181                 BBTOB(nbblks + (blk_no & log->l_sectbb_mask)));
 182         return ptr;
 183 }
 184
 185 #ifdef DEBUG
 186 /*
 187  * dump debug superblock and log record information
 188  */
 189 STATIC void
 190 xlog_header_check_dump(
 191         xfs_mount_t             *mp,
 192         xlog_rec_header_t       *head)
 193 {
 194         int                     b;
 195
 196         cmn_err(CE_DEBUG, "%s:  SB : uuid = ", __func__);
 197         for (b = 0; b < 16; b++)
 198                 cmn_err(CE_DEBUG, "%02x", ((uchar_t *)&mp->m_sb.sb_uuid)[b]);
 199         cmn_err(CE_DEBUG, ", fmt = %d\n", XLOG_FMT);
 200         cmn_err(CE_DEBUG, "    log : uuid = ");
 201         for (b = 0; b < 16; b++)
 202                 cmn_err(CE_DEBUG, "%02x",((uchar_t *)&head->h_fs_uuid)[b]);
 203         cmn_err(CE_DEBUG, ", fmt = %d\n", be32_to_cpu(head->h_fmt));
 204 }
 205 #else
 206 #define xlog_header_check_dump(mp, head)
 207 #endif
 208
 209 /*
 210  * check log record header for recovery
 211  */
 212 STATIC int
 213 xlog_header_check_recover(
 214         xfs_mount_t             *mp,
 215         xlog_rec_header_t       *head)
 216 {
 217         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
 218
 219         /*
 220          * IRIX doesn't write the h_fmt field and leaves it zeroed
 221          * (XLOG_FMT_UNKNOWN). This stops us from trying to recover
 222          * a dirty log created in IRIX.
 223          */
 224         if (unlikely(be32_to_cpu(head->h_fmt) != XLOG_FMT)) {
 225                 xlog_warn(
 226         "XFS: dirty log written in incompatible format - can't recover");
 227                 xlog_header_check_dump(mp, head);
 228                 XFS_ERROR_REPORT("xlog_header_check_recover(1)",
 229                                  XFS_ERRLEVEL_HIGH, mp);
 230                 return XFS_ERROR(EFSCORRUPTED);
 231         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 232                 xlog_warn(
 233         "XFS: dirty log entry has mismatched uuid - can't recover");
 234                 xlog_header_check_dump(mp, head);
 235                 XFS_ERROR_REPORT("xlog_header_check_recover(2)",
 236                                  XFS_ERRLEVEL_HIGH, mp);
 237                 return XFS_ERROR(EFSCORRUPTED);
 238         }
 239         return 0;
 240 }
 241
 242 /*
 243  * read the head block of the log and check the header
 244  */
 245 STATIC int
 246 xlog_header_check_mount(
 247         xfs_mount_t             *mp,
 248         xlog_rec_header_t       *head)
 249 {
 250         ASSERT(be32_to_cpu(head->h_magicno) == XLOG_HEADER_MAGIC_NUM);
 251
 252         if (uuid_is_nil(&head->h_fs_uuid)) {
 253                 /*
 254                  * IRIX doesn't write the h_fs_uuid or h_fmt fields. If
 255                  * h_fs_uuid is nil, we assume this log was last mounted
 256                  * by IRIX and continue.
 257                  */
 258                 xlog_warn("XFS: nil uuid in log - IRIX style log");
 259         } else if (unlikely(!uuid_equal(&mp->m_sb.sb_uuid, &head->h_fs_uuid))) {
 260                 xlog_warn("XFS: log has mismatched uuid - can't recover");
 261                 xlog_header_check_dump(mp, head);
 262                 XFS_ERROR_REPORT("xlog_header_check_mount",
 263                                  XFS_ERRLEVEL_HIGH, mp);
 264                 return XFS_ERROR(EFSCORRUPTED);
 265         }
 266         return 0;
 267 }
 268
 269 STATIC void
 270 xlog_recover_iodone(
 271         struct xfs_buf  *bp)
 272 {
 273         xfs_mount_t     *mp;
 274
 275         ASSERT(XFS_BUF_FSPRIVATE(bp, void *));
 276
 277         if (XFS_BUF_GETERROR(bp)) {
 278                 /*
 279                  * We're not going to bother about retrying
 280                  * this during recovery. One strike!
 281                  */
 282                 mp = XFS_BUF_FSPRIVATE(bp, xfs_mount_t *);
 283                 xfs_ioerror_alert("xlog_recover_iodone",
 284                                   mp, bp, XFS_BUF_ADDR(bp));
 285                 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
 286         }
 287         XFS_BUF_SET_FSPRIVATE(bp, NULL);
 288         XFS_BUF_CLR_IODONE_FUNC(bp);
 289         xfs_biodone(bp);
 290 }
 291
 292 /*
 293  * This routine finds (to an approximation) the first block in the physical
 294  * log which contains the given cycle.  It uses a binary search algorithm.
 295  * Note that the algorithm can not be perfect because the disk will not
 296  * necessarily be perfect.
 297  */
 298 STATIC int
 299 xlog_find_cycle_start(
 300         xlog_t          *log,
 301         xfs_buf_t       *bp,
 302         xfs_daddr_t     first_blk,
 303         xfs_daddr_t     *last_blk,
 304         uint            cycle)
 305 {
 306         xfs_caddr_t     offset;
 307         xfs_daddr_t     mid_blk;
 308         uint            mid_cycle;
 309         int             error;
 310
 311         mid_blk = BLK_AVG(first_blk, *last_blk);
 312         while (mid_blk != first_blk && mid_blk != *last_blk) {
 313                 if ((error = xlog_bread(log, mid_blk, 1, bp)))
 314                         return error;
 315                 offset = xlog_align(log, mid_blk, 1, bp);
 316                 mid_cycle = xlog_get_cycle(offset);
 317                 if (mid_cycle == cycle) {
 318                         *last_blk = mid_blk;
 319                         /* last_half_cycle == mid_cycle */
 320                 } else {
 321                         first_blk = mid_blk;
 322                         /* first_half_cycle == mid_cycle */
 323                 }
 324                 mid_blk = BLK_AVG(first_blk, *last_blk);
 325         }
 326         ASSERT((mid_blk == first_blk && mid_blk+1 == *last_blk) ||
 327                (mid_blk == *last_blk && mid_blk-1 == first_blk));
 328
 329         return 0;
 330 }
 331
 332 /*
 333  * Check that the range of blocks does not contain the cycle number
 334  * given.  The scan needs to occur from front to back and the ptr into the
 335  * region must be updated since a later routine will need to perform another
 336  * test.  If the region is completely good, we end up returning the same
 337  * last block number.
 338  *
 339  * Set blkno to -1 if we encounter no errors.  This is an invalid block number
 340  * since we don't ever expect logs to get this large.
 341  */
 342 STATIC int
 343 xlog_find_verify_cycle(
 344         xlog_t          *log,
 345         xfs_daddr_t     start_blk,
 346         int             nbblks,
 347         uint            stop_on_cycle_no,
 348         xfs_daddr_t     *new_blk)
 349 {
 350         xfs_daddr_t     i, j;
 351         uint            cycle;
 352         xfs_buf_t       *bp;
 353         xfs_daddr_t     bufblks;
 354         xfs_caddr_t     buf = NULL;
 355         int             error = 0;
 356
 357         bufblks = 1 << ffs(nbblks);
 358
 359         while (!(bp = xlog_get_bp(log, bufblks))) {
 360                 /* can't get enough memory to do everything in one big buffer */
 361                 bufblks >>= 1;
 362                 if (bufblks <= log->l_sectbb_log)
 363                         return ENOMEM;
 364         }
 365
 366         for (i = start_blk; i < start_blk + nbblks; i += bufblks) {
 367                 int     bcount;
 368
 369                 bcount = min(bufblks, (start_blk + nbblks - i));
 370
 371                 if ((error = xlog_bread(log, i, bcount, bp)))
 372                         goto out;
 373
 374                 buf = xlog_align(log, i, bcount, bp);
 375                 for (j = 0; j < bcount; j++) {
 376                         cycle = xlog_get_cycle(buf);
 377                         if (cycle == stop_on_cycle_no) {
 378                                 *new_blk = i+j;
 379                                 goto out;
 380                         }
 381
 382                         buf += BBSIZE;
 383                 }
 384         }
 385
 386         *new_blk = -1;
 387
 388 out:
 389         xlog_put_bp(bp);
 390         return error;
 391 }
 392
 393 /*
 394  * Potentially backup over partial log record write.
 395  *
 396  * In the typical case, last_blk is the number of the block directly after
 397  * a good log record.  Therefore, we subtract one to get the block number
 398  * of the last block in the given buffer.  extra_bblks contains the number
 399  * of blocks we would have read on a previous read.  This happens when the
 400  * last log record is split over the end of the physical log.
 401  *
 402  * extra_bblks is the number of blocks potentially verified on a previous
 403  * call to this routine.
 404  */
 405 STATIC int
 406 xlog_find_verify_log_record(
 407         xlog_t                  *log,
 408         xfs_daddr_t             start_blk,
 409         xfs_daddr_t             *last_blk,
 410         int                     extra_bblks)
 411 {
 412         xfs_daddr_t             i;
 413         xfs_buf_t               *bp;
 414         xfs_caddr_t             offset = NULL;
 415         xlog_rec_header_t       *head = NULL;
 416         int                     error = 0;
 417         int                     smallmem = 0;
 418         int                     num_blks = *last_blk - start_blk;
 419         int                     xhdrs;
 420
 421         ASSERT(start_blk != 0 || *last_blk != start_blk);
 422
 423         if (!(bp = xlog_get_bp(log, num_blks))) {
 424                 if (!(bp = xlog_get_bp(log, 1)))
 425                         return ENOMEM;
 426                 smallmem = 1;
 427         } else {
 428                 if ((error = xlog_bread(log, start_blk, num_blks, bp)))
 429                         goto out;
 430                 offset = xlog_align(log, start_blk, num_blks, bp);
 431                 offset += ((num_blks - 1) << BBSHIFT);
 432         }
 433
 434         for (i = (*last_blk) - 1; i >= 0; i--) {
 435                 if (i < start_blk) {
 436                         /* valid log record not found */
 437                         xlog_warn(
 438                 "XFS: Log inconsistent (didn't find previous header)");
 439                         ASSERT(0);
 440                         error = XFS_ERROR(EIO);
 441                         goto out;
 442                 }
 443
 444                 if (smallmem) {
 445                         if ((error = xlog_bread(log, i, 1, bp)))
 446                                 goto out;
 447                         offset = xlog_align(log, i, 1, bp);
 448                 }
 449
 450                 head = (xlog_rec_header_t *)offset;
 451
 452                 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(head->h_magicno))
 453                         break;
 454
 455                 if (!smallmem)
 456                         offset -= BBSIZE;
 457         }
 458
 459         /*
 460          * We hit the beginning of the physical log & still no header.  Return
 461          * to caller.  If caller can handle a return of -1, then this routine
 462          * will be called again for the end of the physical log.
 463          */
 464         if (i == -1) {
 465                 error = -1;
 466                 goto out;
 467         }
 468
 469         /*
 470          * We have the final block of the good log (the first block
 471          * of the log record _before_ the head. So we check the uuid.
 472          */
 473         if ((error = xlog_header_check_mount(log->l_mp, head)))
 474                 goto out;
 475
 476         /*
 477          * We may have found a log record header before we expected one.
 478          * last_blk will be the 1st block # with a given cycle #.  We may end
 479          * up reading an entire log record.  In this case, we don't want to
 480          * reset last_blk.  Only when last_blk points in the middle of a log
 481          * record do we update last_blk.
 482          */
 483         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
 484                 uint    h_size = be32_to_cpu(head->h_size);
 485
 486                 xhdrs = h_size / XLOG_HEADER_CYCLE_SIZE;
 487                 if (h_size % XLOG_HEADER_CYCLE_SIZE)
 488                         xhdrs++;
 489         } else {
 490                 xhdrs = 1;
 491         }
 492
 493         if (*last_blk - i + extra_bblks !=
 494             BTOBB(be32_to_cpu(head->h_len)) + xhdrs)
 495                 *last_blk = i;
 496
 497 out:
 498         xlog_put_bp(bp);
 499         return error;
 500 }
 501
 502 /*
 503  * Head is defined to be the point of the log where the next log write
 504  * write could go.  This means that incomplete LR writes at the end are
 505  * eliminated when calculating the head.  We aren't guaranteed that previous
 506  * LR have complete transactions.  We only know that a cycle number of
 507  * current cycle number -1 won't be present in the log if we start writing
 508  * from our current block number.
 509  *
 510  * last_blk contains the block number of the first block with a given
 511  * cycle number.
 512  *
 513  * Return: zero if normal, non-zero if error.
 514  */
 515 STATIC int
 516 xlog_find_head(
 517         xlog_t          *log,
 518         xfs_daddr_t     *return_head_blk)
 519 {
 520         xfs_buf_t       *bp;
 521         xfs_caddr_t     offset;
 522         xfs_daddr_t     new_blk, first_blk, start_blk, last_blk, head_blk;
 523         int             num_scan_bblks;
 524         uint            first_half_cycle, last_half_cycle;
 525         uint            stop_on_cycle;
 526         int             error, log_bbnum = log->l_logBBsize;
 527
 528         /* Is the end of the log device zeroed? */
 529         if ((error = xlog_find_zeroed(log, &first_blk)) == -1) {
 530                 *return_head_blk = first_blk;
 531
 532                 /* Is the whole lot zeroed? */
 533                 if (!first_blk) {
 534                         /* Linux XFS shouldn't generate totally zeroed logs -
 535                          * mkfs etc write a dummy unmount record to a fresh
 536                          * log so we can store the uuid in there
 537                          */
 538                         xlog_warn("XFS: totally zeroed log");
 539                 }
 540
 541                 return 0;
 542         } else if (error) {
 543                 xlog_warn("XFS: empty log check failed");
 544                 return error;
 545         }
 546
 547         first_blk = 0;                  /* get cycle # of 1st block */
 548         bp = xlog_get_bp(log, 1);
 549         if (!bp)
 550                 return ENOMEM;
 551         if ((error = xlog_bread(log, 0, 1, bp)))
 552                 goto bp_err;
 553         offset = xlog_align(log, 0, 1, bp);
 554         first_half_cycle = xlog_get_cycle(offset);
 555
 556         last_blk = head_blk = log_bbnum - 1;    /* get cycle # of last block */
 557         if ((error = xlog_bread(log, last_blk, 1, bp)))
 558                 goto bp_err;
 559         offset = xlog_align(log, last_blk, 1, bp);
 560         last_half_cycle = xlog_get_cycle(offset);
 561         ASSERT(last_half_cycle != 0);
 562
 563         /*
 564          * If the 1st half cycle number is equal to the last half cycle number,
 565          * then the entire log is stamped with the same cycle number.  In this
 566          * case, head_blk can't be set to zero (which makes sense).  The below
 567          * math doesn't work out properly with head_blk equal to zero.  Instead,
 568          * we set it to log_bbnum which is an invalid block number, but this
 569          * value makes the math correct.  If head_blk doesn't changed through
 570          * all the tests below, *head_blk is set to zero at the very end rather
 571          * than log_bbnum.  In a sense, log_bbnum and zero are the same block
 572          * in a circular file.
 573          */
 574         if (first_half_cycle == last_half_cycle) {
 575                 /*
 576                  * In this case we believe that the entire log should have
 577                  * cycle number last_half_cycle.  We need to scan backwards
 578                  * from the end verifying that there are no holes still
 579                  * containing last_half_cycle - 1.  If we find such a hole,
 580                  * then the start of that hole will be the new head.  The
 581                  * simple case looks like
 582                  *        x | x ... | x - 1 | x
 583                  * Another case that fits this picture would be
 584                  *        x | x + 1 | x ... | x
 585                  * In this case the head really is somewhere at the end of the
 586                  * log, as one of the latest writes at the beginning was
 587                  * incomplete.
 588                  * One more case is
 589                  *        x | x + 1 | x ... | x - 1 | x
 590                  * This is really the combination of the above two cases, and
 591                  * the head has to end up at the start of the x-1 hole at the
 592                  * end of the log.
 593                  *
 594                  * In the 256k log case, we will read from the beginning to the
 595                  * end of the log and search for cycle numbers equal to x-1.
 596                  * We don't worry about the x+1 blocks that we encounter,
 597                  * because we know that they cannot be the head since the log
 598                  * started with x.
 599                  */
 600                 head_blk = log_bbnum;
 601                 stop_on_cycle = last_half_cycle - 1;
 602         } else {
 603                 /*
 604                  * In this case we want to find the first block with cycle
 605                  * number matching last_half_cycle.  We expect the log to be
 606                  * some variation on
 607                  *        x + 1 ... | x ...
 608                  * The first block with cycle number x (last_half_cycle) will
 609                  * be where the new head belongs.  First we do a binary search
 610                  * for the first occurrence of last_half_cycle.  The binary
 611                  * search may not be totally accurate, so then we scan back
 612                  * from there looking for occurrences of last_half_cycle before
 613                  * us.  If that backwards scan wraps around the beginning of
 614                  * the log, then we look for occurrences of last_half_cycle - 1
 615                  * at the end of the log.  The cases we're looking for look
 616                  * like
 617                  *        x + 1 ... | x | x + 1 | x ...
 618                  *                               ^ binary search stopped here
 619                  * or
 620                  *        x + 1 ... | x ... | x - 1 | x
 621                  *        <---------> less than scan distance
 622                  */
 623                 stop_on_cycle = last_half_cycle;
 624                 if ((error = xlog_find_cycle_start(log, bp, first_blk,
 625                                                 &head_blk, last_half_cycle)))
 626                         goto bp_err;
 627         }
 628
 629         /*
 630          * Now validate the answer.  Scan back some number of maximum possible
 631          * blocks and make sure each one has the expected cycle number.  The
 632          * maximum is determined by the total possible amount of buffering
 633          * in the in-core log.  The following number can be made tighter if
 634          * we actually look at the block size of the filesystem.
 635          */
 636         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
 637         if (head_blk >= num_scan_bblks) {
 638                 /*
 639                  * We are guaranteed that the entire check can be performed
 640                  * in one buffer.
 641                  */
 642                 start_blk = head_blk - num_scan_bblks;
 643                 if ((error = xlog_find_verify_cycle(log,
 644                                                 start_blk, num_scan_bblks,
 645                                                 stop_on_cycle, &new_blk)))
 646                         goto bp_err;
 647                 if (new_blk != -1)
 648                         head_blk = new_blk;
 649         } else {                /* need to read 2 parts of log */
 650                 /*
 651                  * We are going to scan backwards in the log in two parts.
 652                  * First we scan the physical end of the log.  In this part
 653                  * of the log, we are looking for blocks with cycle number
 654                  * last_half_cycle - 1.
 655                  * If we find one, then we know that the log starts there, as
 656                  * we've found a hole that didn't get written in going around
 657                  * the end of the physical log.  The simple case for this is
 658                  *        x + 1 ... | x ... | x - 1 | x
 659                  *        <---------> less than scan distance
 660                  * If all of the blocks at the end of the log have cycle number
 661                  * last_half_cycle, then we check the blocks at the start of
 662                  * the log looking for occurrences of last_half_cycle.  If we
 663                  * find one, then our current estimate for the location of the
 664                  * first occurrence of last_half_cycle is wrong and we move
 665                  * back to the hole we've found.  This case looks like
 666                  *        x + 1 ... | x | x + 1 | x ...
 667                  *                               ^ binary search stopped here
 668                  * Another case we need to handle that only occurs in 256k
 669                  * logs is
 670                  *        x + 1 ... | x ... | x+1 | x ...
 671                  *                   ^ binary search stops here
 672                  * In a 256k log, the scan at the end of the log will see the
 673                  * x + 1 blocks.  We need to skip past those since that is
 674                  * certainly not the head of the log.  By searching for
 675                  * last_half_cycle-1 we accomplish that.
 676                  */
 677                 start_blk = log_bbnum - num_scan_bblks + head_blk;
 678                 ASSERT(head_blk <= INT_MAX &&
 679                         (xfs_daddr_t) num_scan_bblks - head_blk >= 0);
 680                 if ((error = xlog_find_verify_cycle(log, start_blk,
 681                                         num_scan_bblks - (int)head_blk,
 682                                         (stop_on_cycle - 1), &new_blk)))
 683                         goto bp_err;
 684                 if (new_blk != -1) {
 685                         head_blk = new_blk;
 686                         goto bad_blk;
 687                 }
 688
 689                 /*
 690                  * Scan beginning of log now.  The last part of the physical
 691                  * log is good.  This scan needs to verify that it doesn't find
 692                  * the last_half_cycle.
 693                  */
 694                 start_blk = 0;
 695                 ASSERT(head_blk <= INT_MAX);
 696                 if ((error = xlog_find_verify_cycle(log,
 697                                         start_blk, (int)head_blk,
 698                                         stop_on_cycle, &new_blk)))
 699                         goto bp_err;
 700                 if (new_blk != -1)
 701                         head_blk = new_blk;
 702         }
 703
 704  bad_blk:
 705         /*
 706          * Now we need to make sure head_blk is not pointing to a block in
 707          * the middle of a log record.
 708          */
 709         num_scan_bblks = XLOG_REC_SHIFT(log);
 710         if (head_blk >= num_scan_bblks) {
 711                 start_blk = head_blk - num_scan_bblks; /* don't read head_blk */
 712
 713                 /* start ptr at last block ptr before head_blk */
 714                 if ((error = xlog_find_verify_log_record(log, start_blk,
 715                                                         &head_blk, 0)) == -1) {
 716                         error = XFS_ERROR(EIO);
 717                         goto bp_err;
 718                 } else if (error)
 719                         goto bp_err;
 720         } else {
 721                 start_blk = 0;
 722                 ASSERT(head_blk <= INT_MAX);
 723                 if ((error = xlog_find_verify_log_record(log, start_blk,
 724                                                         &head_blk, 0)) == -1) {
 725                         /* We hit the beginning of the log during our search */
 726                         start_blk = log_bbnum - num_scan_bblks + head_blk;
 727                         new_blk = log_bbnum;
 728                         ASSERT(start_blk <= INT_MAX &&
 729                                 (xfs_daddr_t) log_bbnum-start_blk >= 0);
 730                         ASSERT(head_blk <= INT_MAX);
 731                         if ((error = xlog_find_verify_log_record(log,
 732                                                         start_blk, &new_blk,
 733                                                         (int)head_blk)) == -1) {
 734                                 error = XFS_ERROR(EIO);
 735                                 goto bp_err;
 736                         } else if (error)
 737                                 goto bp_err;
 738                         if (new_blk != log_bbnum)
 739                                 head_blk = new_blk;
 740                 } else if (error)
 741                         goto bp_err;
 742         }
 743
 744         xlog_put_bp(bp);
 745         if (head_blk == log_bbnum)
 746                 *return_head_blk = 0;
 747         else
 748                 *return_head_blk = head_blk;
 749         /*
 750          * When returning here, we have a good block number.  Bad block
 751          * means that during a previous crash, we didn't have a clean break
 752          * from cycle number N to cycle number N-1.  In this case, we need
 753          * to find the first block with cycle number N-1.
 754          */
 755         return 0;
 756
 757  bp_err:
 758         xlog_put_bp(bp);
 759
 760         if (error)
 761             xlog_warn("XFS: failed to find log head");
 762         return error;
 763 }
 764
 765 /*
 766  * Find the sync block number or the tail of the log.
 767  *
 768  * This will be the block number of the last record to have its
 769  * associated buffers synced to disk.  Every log record header has
 770  * a sync lsn embedded in it.  LSNs hold block numbers, so it is easy
 771  * to get a sync block number.  The only concern is to figure out which
 772  * log record header to believe.
 773  *
 774  * The following algorithm uses the log record header with the largest
 775  * lsn.  The entire log record does not need to be valid.  We only care
 776  * that the header is valid.
 777  *
 778  * We could speed up search by using current head_blk buffer, but it is not
 779  * available.
 780  */
 781 int
 782 xlog_find_tail(
 783         xlog_t                  *log,
 784         xfs_daddr_t             *head_blk,
 785         xfs_daddr_t             *tail_blk)
 786 {
 787         xlog_rec_header_t       *rhead;
 788         xlog_op_header_t        *op_head;
 789         xfs_caddr_t             offset = NULL;
 790         xfs_buf_t               *bp;
 791         int                     error, i, found;
 792         xfs_daddr_t             umount_data_blk;
 793         xfs_daddr_t             after_umount_blk;
 794         xfs_lsn_t               tail_lsn;
 795         int                     hblks;
 796
 797         found = 0;
 798
 799         /*
 800          * Find previous log record
 801          */
 802         if ((error = xlog_find_head(log, head_blk)))
 803                 return error;
 804
 805         bp = xlog_get_bp(log, 1);
 806         if (!bp)
 807                 return ENOMEM;
 808         if (*head_blk == 0) {                           /* special case */
 809                 if ((error = xlog_bread(log, 0, 1, bp)))
 810                         goto bread_err;
 811                 offset = xlog_align(log, 0, 1, bp);
 812                 if (xlog_get_cycle(offset) == 0) {
 813                         *tail_blk = 0;
 814                         /* leave all other log inited values alone */
 815                         goto exit;
 816                 }
 817         }
 818
 819         /*
 820          * Search backwards looking for log record header block
 821          */
 822         ASSERT(*head_blk < INT_MAX);
 823         for (i = (int)(*head_blk) - 1; i >= 0; i--) {
 824                 if ((error = xlog_bread(log, i, 1, bp)))
 825                         goto bread_err;
 826                 offset = xlog_align(log, i, 1, bp);
 827                 if (XLOG_HEADER_MAGIC_NUM == be32_to_cpu(*(__be32 *)offset)) {
 828                         found = 1;
 829                         break;
 830                 }
 831         }
 832         /*
 833          * If we haven't found the log record header block, start looking
 834          * again from the end of the physical log.  XXXmiken: There should be
 835          * a check here to make sure we didn't search more than N blocks in
 836          * the previous code.
 837          */
 838         if (!found) {
 839                 for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
 840                         if ((error = xlog_bread(log, i, 1, bp)))
 841                                 goto bread_err;
 842                         offset = xlog_align(log, i, 1, bp);
 843                         if (XLOG_HEADER_MAGIC_NUM ==
 844                             be32_to_cpu(*(__be32 *)offset)) {
 845                                 found = 2;
 846                                 break;
 847                         }
 848                 }
 849         }
 850         if (!found) {
 851                 xlog_warn("XFS: xlog_find_tail: couldn't find sync record");
 852                 ASSERT(0);
 853                 return XFS_ERROR(EIO);
 854         }
 855
 856         /* find blk_no of tail of log */
 857         rhead = (xlog_rec_header_t *)offset;
 858         *tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
 859
 860         /*
 861          * Reset log values according to the state of the log when we
 862          * crashed.  In the case where head_blk == 0, we bump curr_cycle
 863          * one because the next write starts a new cycle rather than
 864          * continuing the cycle of the last good log record.  At this
 865          * point we have guaranteed that all partial log records have been
 866          * accounted for.  Therefore, we know that the last good log record
 867          * written was complete and ended exactly on the end boundary
 868          * of the physical log.
 869          */
 870         log->l_prev_block = i;
 871         log->l_curr_block = (int)*head_blk;
 872         log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
 873         if (found == 2)
 874                 log->l_curr_cycle++;
 875         log->l_tail_lsn = be64_to_cpu(rhead->h_tail_lsn);
 876         log->l_last_sync_lsn = be64_to_cpu(rhead->h_lsn);
 877         log->l_grant_reserve_cycle = log->l_curr_cycle;
 878         log->l_grant_reserve_bytes = BBTOB(log->l_curr_block);
 879         log->l_grant_write_cycle = log->l_curr_cycle;
 880         log->l_grant_write_bytes = BBTOB(log->l_curr_block);
 881
 882         /*
 883          * Look for unmount record.  If we find it, then we know there
 884          * was a clean unmount.  Since 'i' could be the last block in
 885          * the physical log, we convert to a log block before comparing
 886          * to the head_blk.
 887          *
 888          * Save the current tail lsn to use to pass to
 889          * xlog_clear_stale_blocks() below.  We won't want to clear the
 890          * unmount record if there is one, so we pass the lsn of the
 891          * unmount record rather than the block after it.
 892          */
 893         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
 894                 int     h_size = be32_to_cpu(rhead->h_size);
 895                 int     h_version = be32_to_cpu(rhead->h_version);
 896
 897                 if ((h_version & XLOG_VERSION_2) &&
 898                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
 899                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
 900                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
 901                                 hblks++;
 902                 } else {
 903                         hblks = 1;
 904                 }
 905         } else {
 906                 hblks = 1;
 907         }
 908         after_umount_blk = (i + hblks + (int)
 909                 BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
 910         tail_lsn = log->l_tail_lsn;
 911         if (*head_blk == after_umount_blk &&
 912             be32_to_cpu(rhead->h_num_logops) == 1) {
 913                 umount_data_blk = (i + hblks) % log->l_logBBsize;
 914                 if ((error = xlog_bread(log, umount_data_blk, 1, bp))) {
 915                         goto bread_err;
 916                 }
 917                 offset = xlog_align(log, umount_data_blk, 1, bp);
 918                 op_head = (xlog_op_header_t *)offset;
 919                 if (op_head->oh_flags & XLOG_UNMOUNT_TRANS) {
 920                         /*
 921                          * Set tail and last sync so that newly written
 922                          * log records will point recovery to after the
 923                          * current unmount record.
 924                          */
 925                         log->l_tail_lsn =
 926                                 xlog_assign_lsn(log->l_curr_cycle,
 927                                                 after_umount_blk);
 928                         log->l_last_sync_lsn =
 929                                 xlog_assign_lsn(log->l_curr_cycle,
 930                                                 after_umount_blk);
 931                         *tail_blk = after_umount_blk;
 932
 933                         /*
 934                          * Note that the unmount was clean. If the unmount
 935                          * was not clean, we need to know this to rebuild the
 936                          * superblock counters from the perag headers if we
 937                          * have a filesystem using non-persistent counters.
 938                          */
 939                         log->l_mp->m_flags |= XFS_MOUNT_WAS_CLEAN;
 940                 }
 941         }
 942
 943         /*
 944          * Make sure that there are no blocks in front of the head
 945          * with the same cycle number as the head.  This can happen
 946          * because we allow multiple outstanding log writes concurrently,
 947          * and the later writes might make it out before earlier ones.
 948          *
 949          * We use the lsn from before modifying it so that we'll never
 950          * overwrite the unmount record after a clean unmount.
 951          *
 952          * Do this only if we are going to recover the filesystem
 953          *
 954          * NOTE: This used to say "if (!readonly)"
 955          * However on Linux, we can & do recover a read-only filesystem.
 956          * We only skip recovery if NORECOVERY is specified on mount,
 957          * in which case we would not be here.
 958          *
 959          * But... if the -device- itself is readonly, just skip this.
 960          * We can't recover this device anyway, so it won't matter.
 961          */
 962         if (!xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
 963                 error = xlog_clear_stale_blocks(log, tail_lsn);
 964         }
 965
 966 bread_err:
 967 exit:
 968         xlog_put_bp(bp);
 969
 970         if (error)
 971                 xlog_warn("XFS: failed to locate log tail");
 972         return error;
 973 }
 974
 975 /*
 976  * Is the log zeroed at all?
 977  *
 978  * The last binary search should be changed to perform an X block read
 979  * once X becomes small enough.  You can then search linearly through
 980  * the X blocks.  This will cut down on the number of reads we need to do.
 981  *
 982  * If the log is partially zeroed, this routine will pass back the blkno
 983  * of the first block with cycle number 0.  It won't have a complete LR
 984  * preceding it.
 985  *
 986  * Return:
 987  *      0  => the log is completely written to
 988  *      -1 => use *blk_no as the first block of the log
 989  *      >0 => error has occurred
 990  */
 991 STATIC int
 992 xlog_find_zeroed(
 993         xlog_t          *log,
 994         xfs_daddr_t     *blk_no)
 995 {
 996         xfs_buf_t       *bp;
 997         xfs_caddr_t     offset;
 998         uint            first_cycle, last_cycle;
 999         xfs_daddr_t     new_blk, last_blk, start_blk;
1000         xfs_daddr_t     num_scan_bblks;
1001         int             error, log_bbnum = log->l_logBBsize;
1002
1003         *blk_no = 0;
1004
1005         /* check totally zeroed log */
1006         bp = xlog_get_bp(log, 1);
1007         if (!bp)
1008                 return ENOMEM;
1009         if ((error = xlog_bread(log, 0, 1, bp)))
1010                 goto bp_err;
1011         offset = xlog_align(log, 0, 1, bp);
1012         first_cycle = xlog_get_cycle(offset);
1013         if (first_cycle == 0) {         /* completely zeroed log */
1014                 *blk_no = 0;
1015                 xlog_put_bp(bp);
1016                 return -1;
1017         }
1018
1019         /* check partially zeroed log */
1020         if ((error = xlog_bread(log, log_bbnum-1, 1, bp)))
1021                 goto bp_err;
1022         offset = xlog_align(log, log_bbnum-1, 1, bp);
1023         last_cycle = xlog_get_cycle(offset);
1024         if (last_cycle != 0) {          /* log completely written to */
1025                 xlog_put_bp(bp);
1026                 return 0;
1027         } else if (first_cycle != 1) {
1028                 /*
1029                  * If the cycle of the last block is zero, the cycle of
1030                  * the first block must be 1. If it's not, maybe we're
1031                  * not looking at a log... Bail out.
1032                  */
1033                 xlog_warn("XFS: Log inconsistent or not a log (last==0, first!=1)");
1034                 return XFS_ERROR(EINVAL);
1035         }
1036
1037         /* we have a partially zeroed log */
1038         last_blk = log_bbnum-1;
1039         if ((error = xlog_find_cycle_start(log, bp, 0, &last_blk, 0)))
1040                 goto bp_err;
1041
1042         /*
1043          * Validate the answer.  Because there is no way to guarantee that
1044          * the entire log is made up of log records which are the same size,
1045          * we scan over the defined maximum blocks.  At this point, the maximum
1046          * is not chosen to mean anything special.   XXXmiken
1047          */
1048         num_scan_bblks = XLOG_TOTAL_REC_SHIFT(log);
1049         ASSERT(num_scan_bblks <= INT_MAX);
1050
1051         if (last_blk < num_scan_bblks)
1052                 num_scan_bblks = last_blk;
1053         start_blk = last_blk - num_scan_bblks;
1054
1055         /*
1056          * We search for any instances of cycle number 0 that occur before
1057          * our current estimate of the head.  What we're trying to detect is
1058          *        1 ... | 0 | 1 | 0...
1059          *                       ^ binary search ends here
1060          */
1061         if ((error = xlog_find_verify_cycle(log, start_blk,
1062                                          (int)num_scan_bblks, 0, &new_blk)))
1063                 goto bp_err;
1064         if (new_blk != -1)
1065                 last_blk = new_blk;
1066
1067         /*
1068          * Potentially backup over partial log record write.  We don't need
1069          * to search the end of the log because we know it is zero.
1070          */
1071         if ((error = xlog_find_verify_log_record(log, start_blk,
1072                                 &last_blk, 0)) == -1) {
1073             error = XFS_ERROR(EIO);
1074             goto bp_err;
1075         } else if (error)
1076             goto bp_err;
1077
1078         *blk_no = last_blk;
1079 bp_err:
1080         xlog_put_bp(bp);
1081         if (error)
1082                 return error;
1083         return -1;
1084 }
1085
1086 /*
1087  * These are simple subroutines used by xlog_clear_stale_blocks() below
1088  * to initialize a buffer full of empty log record headers and write
1089  * them into the log.
1090  */
1091 STATIC void
1092 xlog_add_record(
1093         xlog_t                  *log,
1094         xfs_caddr_t             buf,
1095         int                     cycle,
1096         int                     block,
1097         int                     tail_cycle,
1098         int                     tail_block)
1099 {
1100         xlog_rec_header_t       *recp = (xlog_rec_header_t *)buf;
1101
1102         memset(buf, 0, BBSIZE);
1103         recp->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1104         recp->h_cycle = cpu_to_be32(cycle);
1105         recp->h_version = cpu_to_be32(
1106                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1107         recp->h_lsn = cpu_to_be64(xlog_assign_lsn(cycle, block));
1108         recp->h_tail_lsn = cpu_to_be64(xlog_assign_lsn(tail_cycle, tail_block));
1109         recp->h_fmt = cpu_to_be32(XLOG_FMT);
1110         memcpy(&recp->h_fs_uuid, &log->l_mp->m_sb.sb_uuid, sizeof(uuid_t));
1111 }
1112
1113 STATIC int
1114 xlog_write_log_records(
1115         xlog_t          *log,
1116         int             cycle,
1117         int             start_block,
1118         int             blocks,
1119         int             tail_cycle,
1120         int             tail_block)
1121 {
1122         xfs_caddr_t     offset;
1123         xfs_buf_t       *bp;
1124         int             balign, ealign;
1125         int             sectbb = XLOG_SECTOR_ROUNDUP_BBCOUNT(log, 1);
1126         int             end_block = start_block + blocks;
1127         int             bufblks;
1128         int             error = 0;
1129         int             i, j = 0;
1130
1131         bufblks = 1 << ffs(blocks);
1132         while (!(bp = xlog_get_bp(log, bufblks))) {
1133                 bufblks >>= 1;
1134                 if (bufblks <= log->l_sectbb_log)
1135                         return ENOMEM;
1136         }
1137
1138         /* We may need to do a read at the start to fill in part of
1139          * the buffer in the starting sector not covered by the first
1140          * write below.
1141          */
1142         balign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, start_block);
1143         if (balign != start_block) {
1144                 if ((error = xlog_bread(log, start_block, 1, bp))) {
1145                         xlog_put_bp(bp);
1146                         return error;
1147                 }
1148                 j = start_block - balign;
1149         }
1150
1151         for (i = start_block; i < end_block; i += bufblks) {
1152                 int             bcount, endcount;
1153
1154                 bcount = min(bufblks, end_block - start_block);
1155                 endcount = bcount - j;
1156
1157                 /* We may need to do a read at the end to fill in part of
1158                  * the buffer in the final sector not covered by the write.
1159                  * If this is the same sector as the above read, skip it.
1160                  */
1161                 ealign = XLOG_SECTOR_ROUNDDOWN_BLKNO(log, end_block);
1162                 if (j == 0 && (start_block + endcount > ealign)) {
1163                         offset = XFS_BUF_PTR(bp);
1164                         balign = BBTOB(ealign - start_block);
1165                         error = XFS_BUF_SET_PTR(bp, offset + balign,
1166                                                 BBTOB(sectbb));
1167                         if (!error)
1168                                 error = xlog_bread(log, ealign, sectbb, bp);
1169                         if (!error)
1170                                 error = XFS_BUF_SET_PTR(bp, offset, bufblks);
1171                         if (error)
1172                                 break;
1173                 }
1174
1175                 offset = xlog_align(log, start_block, endcount, bp);
1176                 for (; j < endcount; j++) {
1177                         xlog_add_record(log, offset, cycle, i+j,
1178                                         tail_cycle, tail_block);
1179                         offset += BBSIZE;
1180                 }
1181                 error = xlog_bwrite(log, start_block, endcount, bp);
1182                 if (error)
1183                         break;
1184                 start_block += endcount;
1185                 j = 0;
1186         }
1187         xlog_put_bp(bp);
1188         return error;
1189 }
1190
1191 /*
1192  * This routine is called to blow away any incomplete log writes out
1193  * in front of the log head.  We do this so that we won't become confused
1194  * if we come up, write only a little bit more, and then crash again.
1195  * If we leave the partial log records out there, this situation could
1196  * cause us to think those partial writes are valid blocks since they
1197  * have the current cycle number.  We get rid of them by overwriting them
1198  * with empty log records with the old cycle number rather than the
1199  * current one.
1200  *
1201  * The tail lsn is passed in rather than taken from
1202  * the log so that we will not write over the unmount record after a
1203  * clean unmount in a 512 block log.  Doing so would leave the log without
1204  * any valid log records in it until a new one was written.  If we crashed
1205  * during that time we would not be able to recover.
1206  */
1207 STATIC int
1208 xlog_clear_stale_blocks(
1209         xlog_t          *log,
1210         xfs_lsn_t       tail_lsn)
1211 {
1212         int             tail_cycle, head_cycle;
1213         int             tail_block, head_block;
1214         int             tail_distance, max_distance;
1215         int             distance;
1216         int             error;
1217
1218         tail_cycle = CYCLE_LSN(tail_lsn);
1219         tail_block = BLOCK_LSN(tail_lsn);
1220         head_cycle = log->l_curr_cycle;
1221         head_block = log->l_curr_block;
1222
1223         /*
1224          * Figure out the distance between the new head of the log
1225          * and the tail.  We want to write over any blocks beyond the
1226          * head that we may have written just before the crash, but
1227          * we don't want to overwrite the tail of the log.
1228          */
1229         if (head_cycle == tail_cycle) {
1230                 /*
1231                  * The tail is behind the head in the physical log,
1232                  * so the distance from the head to the tail is the
1233                  * distance from the head to the end of the log plus
1234                  * the distance from the beginning of the log to the
1235                  * tail.
1236                  */
1237                 if (unlikely(head_block < tail_block || head_block >= log->l_logBBsize)) {
1238                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(1)",
1239                                          XFS_ERRLEVEL_LOW, log->l_mp);
1240                         return XFS_ERROR(EFSCORRUPTED);
1241                 }
1242                 tail_distance = tail_block + (log->l_logBBsize - head_block);
1243         } else {
1244                 /*
1245                  * The head is behind the tail in the physical log,
1246                  * so the distance from the head to the tail is just
1247                  * the tail block minus the head block.
1248                  */
1249                 if (unlikely(head_block >= tail_block || head_cycle != (tail_cycle + 1))){
1250                         XFS_ERROR_REPORT("xlog_clear_stale_blocks(2)",
1251                                          XFS_ERRLEVEL_LOW, log->l_mp);
1252                         return XFS_ERROR(EFSCORRUPTED);
1253                 }
1254                 tail_distance = tail_block - head_block;
1255         }
1256
1257         /*
1258          * If the head is right up against the tail, we can't clear
1259          * anything.
1260          */
1261         if (tail_distance <= 0) {
1262                 ASSERT(tail_distance == 0);
1263                 return 0;
1264         }
1265
1266         max_distance = XLOG_TOTAL_REC_SHIFT(log);
1267         /*
1268          * Take the smaller of the maximum amount of outstanding I/O
1269          * we could have and the distance to the tail to clear out.
1270          * We take the smaller so that we don't overwrite the tail and
1271          * we don't waste all day writing from the head to the tail
1272          * for no reason.
1273          */
1274         max_distance = MIN(max_distance, tail_distance);
1275
1276         if ((head_block + max_distance) <= log->l_logBBsize) {
1277                 /*
1278                  * We can stomp all the blocks we need to without
1279                  * wrapping around the end of the log.  Just do it
1280                  * in a single write.  Use the cycle number of the
1281                  * current cycle minus one so that the log will look like:
1282                  *     n ... | n - 1 ...
1283                  */
1284                 error = xlog_write_log_records(log, (head_cycle - 1),
1285                                 head_block, max_distance, tail_cycle,
1286                                 tail_block);
1287                 if (error)
1288                         return error;
1289         } else {
1290                 /*
1291                  * We need to wrap around the end of the physical log in
1292                  * order to clear all the blocks.  Do it in two separate
1293                  * I/Os.  The first write should be from the head to the
1294                  * end of the physical log, and it should use the current
1295                  * cycle number minus one just like above.
1296                  */
1297                 distance = log->l_logBBsize - head_block;
1298                 error = xlog_write_log_records(log, (head_cycle - 1),
1299                                 head_block, distance, tail_cycle,
1300                                 tail_block);
1301
1302                 if (error)
1303                         return error;
1304
1305                 /*
1306                  * Now write the blocks at the start of the physical log.
1307                  * This writes the remainder of the blocks we want to clear.
1308                  * It uses the current cycle number since we're now on the
1309                  * same cycle as the head so that we get:
1310                  *    n ... n ... | n - 1 ...
1311                  *    ^^^^^ blocks we're writing
1312                  */
1313                 distance = max_distance - (log->l_logBBsize - head_block);
1314                 error = xlog_write_log_records(log, head_cycle, 0, distance,
1315                                 tail_cycle, tail_block);
1316                 if (error)
1317                         return error;
1318         }
1319
1320         return 0;
1321 }
1322
1323 /******************************************************************************
1324  *
1325  *              Log recover routines
1326  *
1327  ******************************************************************************
1328  */
1329
1330 STATIC xlog_recover_t *
1331 xlog_recover_find_tid(
1332         xlog_recover_t          *q,
1333         xlog_tid_t              tid)
1334 {
1335         xlog_recover_t          *p = q;
1336
1337         while (p != NULL) {
1338                 if (p->r_log_tid == tid)
1339                     break;
1340                 p = p->r_next;
1341         }
1342         return p;
1343 }
1344
1345 STATIC void
1346 xlog_recover_put_hashq(
1347         xlog_recover_t          **q,
1348         xlog_recover_t          *trans)
1349 {
1350         trans->r_next = *q;
1351         *q = trans;
1352 }
1353
1354 STATIC void
1355 xlog_recover_add_item(
1356         xlog_recover_item_t     **itemq)
1357 {
1358         xlog_recover_item_t     *item;
1359
1360         item = kmem_zalloc(sizeof(xlog_recover_item_t), KM_SLEEP);
1361         xlog_recover_insert_item_backq(itemq, item);
1362 }
1363
1364 STATIC int
1365 xlog_recover_add_to_cont_trans(
1366         xlog_recover_t          *trans,
1367         xfs_caddr_t             dp,
1368         int                     len)
1369 {
1370         xlog_recover_item_t     *item;
1371         xfs_caddr_t             ptr, old_ptr;
1372         int                     old_len;
1373
1374         item = trans->r_itemq;
1375         if (item == NULL) {
1376                 /* finish copying rest of trans header */
1377                 xlog_recover_add_item(&trans->r_itemq);
1378                 ptr = (xfs_caddr_t) &trans->r_theader +
1379                                 sizeof(xfs_trans_header_t) - len;
1380                 memcpy(ptr, dp, len); /* d, s, l */
1381                 return 0;
1382         }
1383         item = item->ri_prev;
1384
1385         old_ptr = item->ri_buf[item->ri_cnt-1].i_addr;
1386         old_len = item->ri_buf[item->ri_cnt-1].i_len;
1387
1388         ptr = kmem_realloc(old_ptr, len+old_len, old_len, 0u);
1389         memcpy(&ptr[old_len], dp, len); /* d, s, l */
1390         item->ri_buf[item->ri_cnt-1].i_len += len;
1391         item->ri_buf[item->ri_cnt-1].i_addr = ptr;
1392         return 0;
1393 }
1394
1395 /*
1396  * The next region to add is the start of a new region.  It could be
1397  * a whole region or it could be the first part of a new region.  Because
1398  * of this, the assumption here is that the type and size fields of all
1399  * format structures fit into the first 32 bits of the structure.
1400  *
1401  * This works because all regions must be 32 bit aligned.  Therefore, we
1402  * either have both fields or we have neither field.  In the case we have
1403  * neither field, the data part of the region is zero length.  We only have
1404  * a log_op_header and can throw away the header since a new one will appear
1405  * later.  If we have at least 4 bytes, then we can determine how many regions
1406  * will appear in the current log item.
1407  */
1408 STATIC int
1409 xlog_recover_add_to_trans(
1410         xlog_recover_t          *trans,
1411         xfs_caddr_t             dp,
1412         int                     len)
1413 {
1414         xfs_inode_log_format_t  *in_f;                  /* any will do */
1415         xlog_recover_item_t     *item;
1416         xfs_caddr_t             ptr;
1417
1418         if (!len)
1419                 return 0;
1420         item = trans->r_itemq;
1421         if (item == NULL) {
1422                 /* we need to catch log corruptions here */
1423                 if (*(uint *)dp != XFS_TRANS_HEADER_MAGIC) {
1424                         xlog_warn("XFS: xlog_recover_add_to_trans: "
1425                                   "bad header magic number");
1426                         ASSERT(0);
1427                         return XFS_ERROR(EIO);
1428                 }
1429                 if (len == sizeof(xfs_trans_header_t))
1430                         xlog_recover_add_item(&trans->r_itemq);
1431                 memcpy(&trans->r_theader, dp, len); /* d, s, l */
1432                 return 0;
1433         }
1434
1435         ptr = kmem_alloc(len, KM_SLEEP);
1436         memcpy(ptr, dp, len);
1437         in_f = (xfs_inode_log_format_t *)ptr;
1438
1439         if (item->ri_prev->ri_total != 0 &&
1440              item->ri_prev->ri_total == item->ri_prev->ri_cnt) {
1441                 xlog_recover_add_item(&trans->r_itemq);
1442         }
1443         item = trans->r_itemq;
1444         item = item->ri_prev;
1445
1446         if (item->ri_total == 0) {              /* first region to be added */
1447                 item->ri_total  = in_f->ilf_size;
1448                 ASSERT(item->ri_total <= XLOG_MAX_REGIONS_IN_ITEM);
1449                 item->ri_buf = kmem_zalloc((item->ri_total *
1450                                             sizeof(xfs_log_iovec_t)), KM_SLEEP);
1451         }
1452         ASSERT(item->ri_total > item->ri_cnt);
1453         /* Description region is ri_buf[0] */
1454         item->ri_buf[item->ri_cnt].i_addr = ptr;
1455         item->ri_buf[item->ri_cnt].i_len  = len;
1456         item->ri_cnt++;
1457         return 0;
1458 }
1459
1460 STATIC void
1461 xlog_recover_new_tid(
1462         xlog_recover_t          **q,
1463         xlog_tid_t              tid,
1464         xfs_lsn_t               lsn)
1465 {
1466         xlog_recover_t          *trans;
1467
1468         trans = kmem_zalloc(sizeof(xlog_recover_t), KM_SLEEP);
1469         trans->r_log_tid   = tid;
1470         trans->r_lsn       = lsn;
1471         xlog_recover_put_hashq(q, trans);
1472 }
1473
1474 STATIC int
1475 xlog_recover_unlink_tid(
1476         xlog_recover_t          **q,
1477         xlog_recover_t          *trans)
1478 {
1479         xlog_recover_t          *tp;
1480         int                     found = 0;
1481
1482         ASSERT(trans != NULL);
1483         if (trans == *q) {
1484                 *q = (*q)->r_next;
1485         } else {
1486                 tp = *q;
1487                 while (tp) {
1488                         if (tp->r_next == trans) {
1489                                 found = 1;
1490                                 break;
1491                         }
1492                         tp = tp->r_next;
1493                 }
1494                 if (!found) {
1495                         xlog_warn(
1496                              "XFS: xlog_recover_unlink_tid: trans not found");
1497                         ASSERT(0);
1498                         return XFS_ERROR(EIO);
1499                 }
1500                 tp->r_next = tp->r_next->r_next;
1501         }
1502         return 0;
1503 }
1504
1505 STATIC void
1506 xlog_recover_insert_item_backq(
1507         xlog_recover_item_t     **q,
1508         xlog_recover_item_t     *item)
1509 {
1510         if (*q == NULL) {
1511                 item->ri_prev = item->ri_next = item;
1512                 *q = item;
1513         } else {
1514                 item->ri_next           = *q;
1515                 item->ri_prev           = (*q)->ri_prev;
1516                 (*q)->ri_prev           = item;
1517                 item->ri_prev->ri_next  = item;
1518         }
1519 }
1520
1521 STATIC void
1522 xlog_recover_insert_item_frontq(
1523         xlog_recover_item_t     **q,
1524         xlog_recover_item_t     *item)
1525 {
1526         xlog_recover_insert_item_backq(q, item);
1527         *q = item;
1528 }
1529
1530 STATIC int
1531 xlog_recover_reorder_trans(
1532         xlog_recover_t          *trans)
1533 {
1534         xlog_recover_item_t     *first_item, *itemq, *itemq_next;
1535         xfs_buf_log_format_t    *buf_f;
1536         ushort                  flags = 0;
1537
1538         first_item = itemq = trans->r_itemq;
1539         trans->r_itemq = NULL;
1540         do {
1541                 itemq_next = itemq->ri_next;
1542                 buf_f = (xfs_buf_log_format_t *)itemq->ri_buf[0].i_addr;
1543
1544                 switch (ITEM_TYPE(itemq)) {
1545                 case XFS_LI_BUF:
1546                         flags = buf_f->blf_flags;
1547                         if (!(flags & XFS_BLI_CANCEL)) {
1548                                 xlog_recover_insert_item_frontq(&trans->r_itemq,
1549                                                                 itemq);
1550                                 break;
1551                         }
1552                 case XFS_LI_INODE:
1553                 case XFS_LI_DQUOT:
1554                 case XFS_LI_QUOTAOFF:
1555                 case XFS_LI_EFD:
1556                 case XFS_LI_EFI:
1557                         xlog_recover_insert_item_backq(&trans->r_itemq, itemq);
1558                         break;
1559                 default:
1560                         xlog_warn(
1561         "XFS: xlog_recover_reorder_trans: unrecognized type of log operation");
1562                         ASSERT(0);
1563                         return XFS_ERROR(EIO);
1564                 }
1565                 itemq = itemq_next;
1566         } while (first_item != itemq);
1567         return 0;
1568 }
1569
1570 /*
1571  * Build up the table of buf cancel records so that we don't replay
1572  * cancelled data in the second pass.  For buffer records that are
1573  * not cancel records, there is nothing to do here so we just return.
1574  *
1575  * If we get a cancel record which is already in the table, this indicates
1576  * that the buffer was cancelled multiple times.  In order to ensure
1577  * that during pass 2 we keep the record in the table until we reach its
1578  * last occurrence in the log, we keep a reference count in the cancel
1579  * record in the table to tell us how many times we expect to see this
1580  * record during the second pass.
1581  */
1582 STATIC void
1583 xlog_recover_do_buffer_pass1(
1584         xlog_t                  *log,
1585         xfs_buf_log_format_t    *buf_f)
1586 {
1587         xfs_buf_cancel_t        *bcp;
1588         xfs_buf_cancel_t        *nextp;
1589         xfs_buf_cancel_t        *prevp;
1590         xfs_buf_cancel_t        **bucket;
1591         xfs_daddr_t             blkno = 0;
1592         uint                    len = 0;
1593         ushort                  flags = 0;
1594
1595         switch (buf_f->blf_type) {
1596         case XFS_LI_BUF:
1597                 blkno = buf_f->blf_blkno;
1598                 len = buf_f->blf_len;
1599                 flags = buf_f->blf_flags;
1600                 break;
1601         }
1602
1603         /*
1604          * If this isn't a cancel buffer item, then just return.
1605          */
1606         if (!(flags & XFS_BLI_CANCEL))
1607                 return;
1608
1609         /*
1610          * Insert an xfs_buf_cancel record into the hash table of
1611          * them.  If there is already an identical record, bump
1612          * its reference count.
1613          */
1614         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1615                                           XLOG_BC_TABLE_SIZE];
1616         /*
1617          * If the hash bucket is empty then just insert a new record into
1618          * the bucket.
1619          */
1620         if (*bucket == NULL) {
1621                 bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1622                                                      KM_SLEEP);
1623                 bcp->bc_blkno = blkno;
1624                 bcp->bc_len = len;
1625                 bcp->bc_refcount = 1;
1626                 bcp->bc_next = NULL;
1627                 *bucket = bcp;
1628                 return;
1629         }
1630
1631         /*
1632          * The hash bucket is not empty, so search for duplicates of our
1633          * record.  If we find one them just bump its refcount.  If not
1634          * then add us at the end of the list.
1635          */
1636         prevp = NULL;
1637         nextp = *bucket;
1638         while (nextp != NULL) {
1639                 if (nextp->bc_blkno == blkno && nextp->bc_len == len) {
1640                         nextp->bc_refcount++;
1641                         return;
1642                 }
1643                 prevp = nextp;
1644                 nextp = nextp->bc_next;
1645         }
1646         ASSERT(prevp != NULL);
1647         bcp = (xfs_buf_cancel_t *)kmem_alloc(sizeof(xfs_buf_cancel_t),
1648                                              KM_SLEEP);
1649         bcp->bc_blkno = blkno;
1650         bcp->bc_len = len;
1651         bcp->bc_refcount = 1;
1652         bcp->bc_next = NULL;
1653         prevp->bc_next = bcp;
1654 }
1655
1656 /*
1657  * Check to see whether the buffer being recovered has a corresponding
1658  * entry in the buffer cancel record table.  If it does then return 1
1659  * so that it will be cancelled, otherwise return 0.  If the buffer is
1660  * actually a buffer cancel item (XFS_BLI_CANCEL is set), then decrement
1661  * the refcount on the entry in the table and remove it from the table
1662  * if this is the last reference.
1663  *
1664  * We remove the cancel record from the table when we encounter its
1665  * last occurrence in the log so that if the same buffer is re-used
1666  * again after its last cancellation we actually replay the changes
1667  * made at that point.
1668  */
1669 STATIC int
1670 xlog_check_buffer_cancelled(
1671         xlog_t                  *log,
1672         xfs_daddr_t             blkno,
1673         uint                    len,
1674         ushort                  flags)
1675 {
1676         xfs_buf_cancel_t        *bcp;
1677         xfs_buf_cancel_t        *prevp;
1678         xfs_buf_cancel_t        **bucket;
1679
1680         if (log->l_buf_cancel_table == NULL) {
1681                 /*
1682                  * There is nothing in the table built in pass one,
1683                  * so this buffer must not be cancelled.
1684                  */
1685                 ASSERT(!(flags & XFS_BLI_CANCEL));
1686                 return 0;
1687         }
1688
1689         bucket = &log->l_buf_cancel_table[(__uint64_t)blkno %
1690                                           XLOG_BC_TABLE_SIZE];
1691         bcp = *bucket;
1692         if (bcp == NULL) {
1693                 /*
1694                  * There is no corresponding entry in the table built
1695                  * in pass one, so this buffer has not been cancelled.
1696                  */
1697                 ASSERT(!(flags & XFS_BLI_CANCEL));
1698                 return 0;
1699         }
1700
1701         /*
1702          * Search for an entry in the buffer cancel table that
1703          * matches our buffer.
1704          */
1705         prevp = NULL;
1706         while (bcp != NULL) {
1707                 if (bcp->bc_blkno == blkno && bcp->bc_len == len) {
1708                         /*
1709                          * We've go a match, so return 1 so that the
1710                          * recovery of this buffer is cancelled.
1711                          * If this buffer is actually a buffer cancel
1712                          * log item, then decrement the refcount on the
1713                          * one in the table and remove it if this is the
1714                          * last reference.
1715                          */
1716                         if (flags & XFS_BLI_CANCEL) {
1717                                 bcp->bc_refcount--;
1718                                 if (bcp->bc_refcount == 0) {
1719                                         if (prevp == NULL) {
1720                                                 *bucket = bcp->bc_next;
1721                                         } else {
1722                                                 prevp->bc_next = bcp->bc_next;
1723                                         }
1724                                         kmem_free(bcp);
1725                                 }
1726                         }
1727                         return 1;
1728                 }
1729                 prevp = bcp;
1730                 bcp = bcp->bc_next;
1731         }
1732         /*
1733          * We didn't find a corresponding entry in the table, so
1734          * return 0 so that the buffer is NOT cancelled.
1735          */
1736         ASSERT(!(flags & XFS_BLI_CANCEL));
1737         return 0;
1738 }
1739
1740 STATIC int
1741 xlog_recover_do_buffer_pass2(
1742         xlog_t                  *log,
1743         xfs_buf_log_format_t    *buf_f)
1744 {
1745         xfs_daddr_t             blkno = 0;
1746         ushort                  flags = 0;
1747         uint                    len = 0;
1748
1749         switch (buf_f->blf_type) {
1750         case XFS_LI_BUF:
1751                 blkno = buf_f->blf_blkno;
1752                 flags = buf_f->blf_flags;
1753                 len = buf_f->blf_len;
1754                 break;
1755         }
1756
1757         return xlog_check_buffer_cancelled(log, blkno, len, flags);
1758 }
1759
1760 /*
1761  * Perform recovery for a buffer full of inodes.  In these buffers,
1762  * the only data which should be recovered is that which corresponds
1763  * to the di_next_unlinked pointers in the on disk inode structures.
1764  * The rest of the data for the inodes is always logged through the
1765  * inodes themselves rather than the inode buffer and is recovered
1766  * in xlog_recover_do_inode_trans().
1767  *
1768  * The only time when buffers full of inodes are fully recovered is
1769  * when the buffer is full of newly allocated inodes.  In this case
1770  * the buffer will not be marked as an inode buffer and so will be
1771  * sent to xlog_recover_do_reg_buffer() below during recovery.
1772  */
1773 STATIC int
1774 xlog_recover_do_inode_buffer(
1775         xfs_mount_t             *mp,
1776         xlog_recover_item_t     *item,
1777         xfs_buf_t               *bp,
1778         xfs_buf_log_format_t    *buf_f)
1779 {
1780         int                     i;
1781         int                     item_index;
1782         int                     bit;
1783         int                     nbits;
1784         int                     reg_buf_offset;
1785         int                     reg_buf_bytes;
1786         int                     next_unlinked_offset;
1787         int                     inodes_per_buf;
1788         xfs_agino_t             *logged_nextp;
1789         xfs_agino_t             *buffer_nextp;
1790         unsigned int            *data_map = NULL;
1791         unsigned int            map_size = 0;
1792
1793         switch (buf_f->blf_type) {
1794         case XFS_LI_BUF:
1795                 data_map = buf_f->blf_data_map;
1796                 map_size = buf_f->blf_map_size;
1797                 break;
1798         }
1799         /*
1800          * Set the variables corresponding to the current region to
1801          * 0 so that we'll initialize them on the first pass through
1802          * the loop.
1803          */
1804         reg_buf_offset = 0;
1805         reg_buf_bytes = 0;
1806         bit = 0;
1807         nbits = 0;
1808         item_index = 0;
1809         inodes_per_buf = XFS_BUF_COUNT(bp) >> mp->m_sb.sb_inodelog;
1810         for (i = 0; i < inodes_per_buf; i++) {
1811                 next_unlinked_offset = (i * mp->m_sb.sb_inodesize) +
1812                         offsetof(xfs_dinode_t, di_next_unlinked);
1813
1814                 while (next_unlinked_offset >=
1815                        (reg_buf_offset + reg_buf_bytes)) {
1816                         /*
1817                          * The next di_next_unlinked field is beyond
1818                          * the current logged region.  Find the next
1819                          * logged region that contains or is beyond
1820                          * the current di_next_unlinked field.
1821                          */
1822                         bit += nbits;
1823                         bit = xfs_next_bit(data_map, map_size, bit);
1824
1825                         /*
1826                          * If there are no more logged regions in the
1827                          * buffer, then we're done.
1828                          */
1829                         if (bit == -1) {
1830                                 return 0;
1831                         }
1832
1833                         nbits = xfs_contig_bits(data_map, map_size,
1834                                                          bit);
1835                         ASSERT(nbits > 0);
1836                         reg_buf_offset = bit << XFS_BLI_SHIFT;
1837                         reg_buf_bytes = nbits << XFS_BLI_SHIFT;
1838                         item_index++;
1839                 }
1840
1841                 /*
1842                  * If the current logged region starts after the current
1843                  * di_next_unlinked field, then move on to the next
1844                  * di_next_unlinked field.
1845                  */
1846                 if (next_unlinked_offset < reg_buf_offset) {
1847                         continue;
1848                 }
1849
1850                 ASSERT(item->ri_buf[item_index].i_addr != NULL);
1851                 ASSERT((item->ri_buf[item_index].i_len % XFS_BLI_CHUNK) == 0);
1852                 ASSERT((reg_buf_offset + reg_buf_bytes) <= XFS_BUF_COUNT(bp));
1853
1854                 /*
1855                  * The current logged region contains a copy of the
1856                  * current di_next_unlinked field.  Extract its value
1857                  * and copy it to the buffer copy.
1858                  */
1859                 logged_nextp = (xfs_agino_t *)
1860                                ((char *)(item->ri_buf[item_index].i_addr) +
1861                                 (next_unlinked_offset - reg_buf_offset));
1862                 if (unlikely(*logged_nextp == 0)) {
1863                         xfs_fs_cmn_err(CE_ALERT, mp,
1864                                 "bad inode buffer log record (ptr = 0x%p, bp = 0x%p).  XFS trying to replay bad (0) inode di_next_unlinked field",
1865                                 item, bp);
1866                         XFS_ERROR_REPORT("xlog_recover_do_inode_buf",
1867                                          XFS_ERRLEVEL_LOW, mp);
1868                         return XFS_ERROR(EFSCORRUPTED);
1869                 }
1870
1871                 buffer_nextp = (xfs_agino_t *)xfs_buf_offset(bp,
1872                                               next_unlinked_offset);
1873                 *buffer_nextp = *logged_nextp;
1874         }
1875
1876         return 0;
1877 }
1878
1879 /*
1880  * Perform a 'normal' buffer recovery.  Each logged region of the
1881  * buffer should be copied over the corresponding region in the
1882  * given buffer.  The bitmap in the buf log format structure indicates
1883  * where to place the logged data.
1884  */
1885 /*ARGSUSED*/
1886 STATIC void
1887 xlog_recover_do_reg_buffer(
1888         xlog_recover_item_t     *item,
1889         xfs_buf_t               *bp,
1890         xfs_buf_log_format_t    *buf_f)
1891 {
1892         int                     i;
1893         int                     bit;
1894         int                     nbits;
1895         unsigned int            *data_map = NULL;
1896         unsigned int            map_size = 0;
1897         int                     error;
1898
1899         switch (buf_f->blf_type) {
1900         case XFS_LI_BUF:
1901                 data_map = buf_f->blf_data_map;
1902                 map_size = buf_f->blf_map_size;
1903                 break;
1904         }
1905         bit = 0;
1906         i = 1;  /* 0 is the buf format structure */
1907         while (1) {
1908                 bit = xfs_next_bit(data_map, map_size, bit);
1909                 if (bit == -1)
1910                         break;
1911                 nbits = xfs_contig_bits(data_map, map_size, bit);
1912                 ASSERT(nbits > 0);
1913                 ASSERT(item->ri_buf[i].i_addr != NULL);
1914                 ASSERT(item->ri_buf[i].i_len % XFS_BLI_CHUNK == 0);
1915                 ASSERT(XFS_BUF_COUNT(bp) >=
1916                        ((uint)bit << XFS_BLI_SHIFT)+(nbits<<XFS_BLI_SHIFT));
1917
1918                 /*
1919                  * Do a sanity check if this is a dquot buffer. Just checking
1920                  * the first dquot in the buffer should do. XXXThis is
1921                  * probably a good thing to do for other buf types also.
1922                  */
1923                 error = 0;
1924                 if (buf_f->blf_flags &
1925                    (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
1926                         error = xfs_qm_dqcheck((xfs_disk_dquot_t *)
1927                                                item->ri_buf[i].i_addr,
1928                                                -1, 0, XFS_QMOPT_DOWARN,
1929                                                "dquot_buf_recover");
1930                 }
1931                 if (!error)
1932                         memcpy(xfs_buf_offset(bp,
1933                                 (uint)bit << XFS_BLI_SHIFT),    /* dest */
1934                                 item->ri_buf[i].i_addr,         /* source */
1935                                 nbits<<XFS_BLI_SHIFT);          /* length */
1936                 i++;
1937                 bit += nbits;
1938         }
1939
1940         /* Shouldn't be any more regions */
1941         ASSERT(i == item->ri_total);
1942 }
1943
1944 /*
1945  * Do some primitive error checking on ondisk dquot data structures.
1946  */
1947 int
1948 xfs_qm_dqcheck(
1949         xfs_disk_dquot_t *ddq,
1950         xfs_dqid_t       id,
1951         uint             type,    /* used only when IO_dorepair is true */
1952         uint             flags,
1953         char             *str)
1954 {
1955         xfs_dqblk_t      *d = (xfs_dqblk_t *)ddq;
1956         int             errs = 0;
1957
1958         /*
1959          * We can encounter an uninitialized dquot buffer for 2 reasons:
1960          * 1. If we crash while deleting the quotainode(s), and those blks got
1961          *    used for user data. This is because we take the path of regular
1962          *    file deletion; however, the size field of quotainodes is never
1963          *    updated, so all the tricks that we play in itruncate_finish
1964          *    don't quite matter.
1965          *
1966          * 2. We don't play the quota buffers when there's a quotaoff logitem.
1967          *    But the allocation will be replayed so we'll end up with an
1968          *    uninitialized quota block.
1969          *
1970          * This is all fine; things are still consistent, and we haven't lost
1971          * any quota information. Just don't complain about bad dquot blks.
1972          */
1973         if (be16_to_cpu(ddq->d_magic) != XFS_DQUOT_MAGIC) {
1974                 if (flags & XFS_QMOPT_DOWARN)
1975                         cmn_err(CE_ALERT,
1976                         "%s : XFS dquot ID 0x%x, magic 0x%x != 0x%x",
1977                         str, id, be16_to_cpu(ddq->d_magic), XFS_DQUOT_MAGIC);
1978                 errs++;
1979         }
1980         if (ddq->d_version != XFS_DQUOT_VERSION) {
1981                 if (flags & XFS_QMOPT_DOWARN)
1982                         cmn_err(CE_ALERT,
1983                         "%s : XFS dquot ID 0x%x, version 0x%x != 0x%x",
1984                         str, id, ddq->d_version, XFS_DQUOT_VERSION);
1985                 errs++;
1986         }
1987
1988         if (ddq->d_flags != XFS_DQ_USER &&
1989             ddq->d_flags != XFS_DQ_PROJ &&
1990             ddq->d_flags != XFS_DQ_GROUP) {
1991                 if (flags & XFS_QMOPT_DOWARN)
1992                         cmn_err(CE_ALERT,
1993                         "%s : XFS dquot ID 0x%x, unknown flags 0x%x",
1994                         str, id, ddq->d_flags);
1995                 errs++;
1996         }
1997
1998         if (id != -1 && id != be32_to_cpu(ddq->d_id)) {
1999                 if (flags & XFS_QMOPT_DOWARN)
2000                         cmn_err(CE_ALERT,
2001                         "%s : ondisk-dquot 0x%p, ID mismatch: "
2002                         "0x%x expected, found id 0x%x",
2003                         str, ddq, id, be32_to_cpu(ddq->d_id));
2004                 errs++;
2005         }
2006
2007         if (!errs && ddq->d_id) {
2008                 if (ddq->d_blk_softlimit &&
2009                     be64_to_cpu(ddq->d_bcount) >=
2010                                 be64_to_cpu(ddq->d_blk_softlimit)) {
2011                         if (!ddq->d_btimer) {
2012                                 if (flags & XFS_QMOPT_DOWARN)
2013                                         cmn_err(CE_ALERT,
2014                                         "%s : Dquot ID 0x%x (0x%p) "
2015                                         "BLK TIMER NOT STARTED",
2016                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2017                                 errs++;
2018                         }
2019                 }
2020                 if (ddq->d_ino_softlimit &&
2021                     be64_to_cpu(ddq->d_icount) >=
2022                                 be64_to_cpu(ddq->d_ino_softlimit)) {
2023                         if (!ddq->d_itimer) {
2024                                 if (flags & XFS_QMOPT_DOWARN)
2025                                         cmn_err(CE_ALERT,
2026                                         "%s : Dquot ID 0x%x (0x%p) "
2027                                         "INODE TIMER NOT STARTED",
2028                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2029                                 errs++;
2030                         }
2031                 }
2032                 if (ddq->d_rtb_softlimit &&
2033                     be64_to_cpu(ddq->d_rtbcount) >=
2034                                 be64_to_cpu(ddq->d_rtb_softlimit)) {
2035                         if (!ddq->d_rtbtimer) {
2036                                 if (flags & XFS_QMOPT_DOWARN)
2037                                         cmn_err(CE_ALERT,
2038                                         "%s : Dquot ID 0x%x (0x%p) "
2039                                         "RTBLK TIMER NOT STARTED",
2040                                         str, (int)be32_to_cpu(ddq->d_id), ddq);
2041                                 errs++;
2042                         }
2043                 }
2044         }
2045
2046         if (!errs || !(flags & XFS_QMOPT_DQREPAIR))
2047                 return errs;
2048
2049         if (flags & XFS_QMOPT_DOWARN)
2050                 cmn_err(CE_NOTE, "Re-initializing dquot ID 0x%x", id);
2051
2052         /*
2053          * Typically, a repair is only requested by quotacheck.
2054          */
2055         ASSERT(id != -1);
2056         ASSERT(flags & XFS_QMOPT_DQREPAIR);
2057         memset(d, 0, sizeof(xfs_dqblk_t));
2058
2059         d->dd_diskdq.d_magic = cpu_to_be16(XFS_DQUOT_MAGIC);
2060         d->dd_diskdq.d_version = XFS_DQUOT_VERSION;
2061         d->dd_diskdq.d_flags = type;
2062         d->dd_diskdq.d_id = cpu_to_be32(id);
2063
2064         return errs;
2065 }
2066
2067 /*
2068  * Perform a dquot buffer recovery.
2069  * Simple algorithm: if we have found a QUOTAOFF logitem of the same type
2070  * (ie. USR or GRP), then just toss this buffer away; don't recover it.
2071  * Else, treat it as a regular buffer and do recovery.
2072  */
2073 STATIC void
2074 xlog_recover_do_dquot_buffer(
2075         xfs_mount_t             *mp,
2076         xlog_t                  *log,
2077         xlog_recover_item_t     *item,
2078         xfs_buf_t               *bp,
2079         xfs_buf_log_format_t    *buf_f)
2080 {
2081         uint                    type;
2082
2083         /*
2084          * Filesystems are required to send in quota flags at mount time.
2085          */
2086         if (mp->m_qflags == 0) {
2087                 return;
2088         }
2089
2090         type = 0;
2091         if (buf_f->blf_flags & XFS_BLI_UDQUOT_BUF)
2092                 type |= XFS_DQ_USER;
2093         if (buf_f->blf_flags & XFS_BLI_PDQUOT_BUF)
2094                 type |= XFS_DQ_PROJ;
2095         if (buf_f->blf_flags & XFS_BLI_GDQUOT_BUF)
2096                 type |= XFS_DQ_GROUP;
2097         /*
2098          * This type of quotas was turned off, so ignore this buffer
2099          */
2100         if (log->l_quotaoffs_flag & type)
2101                 return;
2102
2103         xlog_recover_do_reg_buffer(item, bp, buf_f);
2104 }
2105
2106 /*
2107  * This routine replays a modification made to a buffer at runtime.
2108  * There are actually two types of buffer, regular and inode, which
2109  * are handled differently.  Inode buffers are handled differently
2110  * in that we only recover a specific set of data from them, namely
2111  * the inode di_next_unlinked fields.  This is because all other inode
2112  * data is actually logged via inode records and any data we replay
2113  * here which overlaps that may be stale.
2114  *
2115  * When meta-data buffers are freed at run time we log a buffer item
2116  * with the XFS_BLI_CANCEL bit set to indicate that previous copies
2117  * of the buffer in the log should not be replayed at recovery time.
2118  * This is so that if the blocks covered by the buffer are reused for
2119  * file data before we crash we don't end up replaying old, freed
2120  * meta-data into a user's file.
2121  *
2122  * To handle the cancellation of buffer log items, we make two passes
2123  * over the log during recovery.  During the first we build a table of
2124  * those buffers which have been cancelled, and during the second we
2125  * only replay those buffers which do not have corresponding cancel
2126  * records in the table.  See xlog_recover_do_buffer_pass[1,2] above
2127  * for more details on the implementation of the table of cancel records.
2128  */
2129 STATIC int
2130 xlog_recover_do_buffer_trans(
2131         xlog_t                  *log,
2132         xlog_recover_item_t     *item,
2133         int                     pass)
2134 {
2135         xfs_buf_log_format_t    *buf_f;
2136         xfs_mount_t             *mp;
2137         xfs_buf_t               *bp;
2138         int                     error;
2139         int                     cancel;
2140         xfs_daddr_t             blkno;
2141         int                     len;
2142         ushort                  flags;
2143
2144         buf_f = (xfs_buf_log_format_t *)item->ri_buf[0].i_addr;
2145
2146         if (pass == XLOG_RECOVER_PASS1) {
2147                 /*
2148                  * In this pass we're only looking for buf items
2149                  * with the XFS_BLI_CANCEL bit set.
2150                  */
2151                 xlog_recover_do_buffer_pass1(log, buf_f);
2152                 return 0;
2153         } else {
2154                 /*
2155                  * In this pass we want to recover all the buffers
2156                  * which have not been cancelled and are not
2157                  * cancellation buffers themselves.  The routine
2158                  * we call here will tell us whether or not to
2159                  * continue with the replay of this buffer.
2160                  */
2161                 cancel = xlog_recover_do_buffer_pass2(log, buf_f);
2162                 if (cancel) {
2163                         return 0;
2164                 }
2165         }
2166         switch (buf_f->blf_type) {
2167         case XFS_LI_BUF:
2168                 blkno = buf_f->blf_blkno;
2169                 len = buf_f->blf_len;
2170                 flags = buf_f->blf_flags;
2171                 break;
2172         default:
2173                 xfs_fs_cmn_err(CE_ALERT, log->l_mp,
2174                         "xfs_log_recover: unknown buffer type 0x%x, logdev %s",
2175                         buf_f->blf_type, log->l_mp->m_logname ?
2176                         log->l_mp->m_logname : "internal");
2177                 XFS_ERROR_REPORT("xlog_recover_do_buffer_trans",
2178                                  XFS_ERRLEVEL_LOW, log->l_mp);
2179                 return XFS_ERROR(EFSCORRUPTED);
2180         }
2181
2182         mp = log->l_mp;
2183         if (flags & XFS_BLI_INODE_BUF) {
2184                 bp = xfs_buf_read_flags(mp->m_ddev_targp, blkno, len,
2185                                                                 XFS_BUF_LOCK);
2186         } else {
2187                 bp = xfs_buf_read(mp->m_ddev_targp, blkno, len, 0);
2188         }
2189         if (XFS_BUF_ISERROR(bp)) {
2190                 xfs_ioerror_alert("xlog_recover_do..(read#1)", log->l_mp,
2191                                   bp, blkno);
2192                 error = XFS_BUF_GETERROR(bp);
2193                 xfs_buf_relse(bp);
2194                 return error;
2195         }
2196
2197         error = 0;
2198         if (flags & XFS_BLI_INODE_BUF) {
2199                 error = xlog_recover_do_inode_buffer(mp, item, bp, buf_f);
2200         } else if (flags &
2201                   (XFS_BLI_UDQUOT_BUF|XFS_BLI_PDQUOT_BUF|XFS_BLI_GDQUOT_BUF)) {
2202                 xlog_recover_do_dquot_buffer(mp, log, item, bp, buf_f);
2203         } else {
2204                 xlog_recover_do_reg_buffer(item, bp, buf_f);
2205         }
2206         if (error)
2207                 return XFS_ERROR(error);
2208
2209         /*
2210          * Perform delayed write on the buffer.  Asynchronous writes will be
2211          * slower when taking into account all the buffers to be flushed.
2212          *
2213          * Also make sure that only inode buffers with good sizes stay in
2214          * the buffer cache.  The kernel moves inodes in buffers of 1 block
2215          * or XFS_INODE_CLUSTER_SIZE bytes, whichever is bigger.  The inode
2216          * buffers in the log can be a different size if the log was generated
2217          * by an older kernel using unclustered inode buffers or a newer kernel
2218          * running with a different inode cluster size.  Regardless, if the
2219          * the inode buffer size isn't MAX(blocksize, XFS_INODE_CLUSTER_SIZE)
2220          * for *our* value of XFS_INODE_CLUSTER_SIZE, then we need to keep
2221          * the buffer out of the buffer cache so that the buffer won't
2222          * overlap with future reads of those inodes.
2223          */
2224         if (XFS_DINODE_MAGIC ==
2225             be16_to_cpu(*((__be16 *)xfs_buf_offset(bp, 0))) &&
2226             (XFS_BUF_COUNT(bp) != MAX(log->l_mp->m_sb.sb_blocksize,
2227                         (__uint32_t)XFS_INODE_CLUSTER_SIZE(log->l_mp)))) {
2228                 XFS_BUF_STALE(bp);
2229                 error = xfs_bwrite(mp, bp);
2230         } else {
2231                 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2232                        XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2233                 XFS_BUF_SET_FSPRIVATE(bp, mp);
2234                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2235                 xfs_bdwrite(mp, bp);
2236         }
2237
2238         return (error);
2239 }
2240
2241 STATIC int
2242 xlog_recover_do_inode_trans(
2243         xlog_t                  *log,
2244         xlog_recover_item_t     *item,
2245         int                     pass)
2246 {
2247         xfs_inode_log_format_t  *in_f;
2248         xfs_mount_t             *mp;
2249         xfs_buf_t               *bp;
2250         xfs_imap_t              imap;
2251         xfs_dinode_t            *dip;
2252         xfs_ino_t               ino;
2253         int                     len;
2254         xfs_caddr_t             src;
2255         xfs_caddr_t             dest;
2256         int                     error;
2257         int                     attr_index;
2258         uint                    fields;
2259         xfs_icdinode_t          *dicp;
2260         int                     need_free = 0;
2261
2262         if (pass == XLOG_RECOVER_PASS1) {
2263                 return 0;
2264         }
2265
2266         if (item->ri_buf[0].i_len == sizeof(xfs_inode_log_format_t)) {
2267                 in_f = (xfs_inode_log_format_t *)item->ri_buf[0].i_addr;
2268         } else {
2269                 in_f = (xfs_inode_log_format_t *)kmem_alloc(
2270                         sizeof(xfs_inode_log_format_t), KM_SLEEP);
2271                 need_free = 1;
2272                 error = xfs_inode_item_format_convert(&item->ri_buf[0], in_f);
2273                 if (error)
2274                         goto error;
2275         }
2276         ino = in_f->ilf_ino;
2277         mp = log->l_mp;
2278         if (ITEM_TYPE(item) == XFS_LI_INODE) {
2279                 imap.im_blkno = (xfs_daddr_t)in_f->ilf_blkno;
2280                 imap.im_len = in_f->ilf_len;
2281                 imap.im_boffset = in_f->ilf_boffset;
2282         } else {
2283                 /*
2284                  * It's an old inode format record.  We don't know where
2285                  * its cluster is located on disk, and we can't allow
2286                  * xfs_imap() to figure it out because the inode btrees
2287                  * are not ready to be used.  Therefore do not pass the
2288                  * XFS_IMAP_LOOKUP flag to xfs_imap().  This will give
2289                  * us only the single block in which the inode lives
2290                  * rather than its cluster, so we must make sure to
2291                  * invalidate the buffer when we write it out below.
2292                  */
2293                 imap.im_blkno = 0;
2294                 error = xfs_imap(log->l_mp, NULL, ino, &imap, 0);
2295                 if (error)
2296                         goto error;
2297         }
2298
2299         /*
2300          * Inode buffers can be freed, look out for it,
2301          * and do not replay the inode.
2302          */
2303         if (xlog_check_buffer_cancelled(log, imap.im_blkno, imap.im_len, 0)) {
2304                 error = 0;
2305                 goto error;
2306         }
2307
2308         bp = xfs_buf_read_flags(mp->m_ddev_targp, imap.im_blkno, imap.im_len,
2309                                                                 XFS_BUF_LOCK);
2310         if (XFS_BUF_ISERROR(bp)) {
2311                 xfs_ioerror_alert("xlog_recover_do..(read#2)", mp,
2312                                   bp, imap.im_blkno);
2313                 error = XFS_BUF_GETERROR(bp);
2314                 xfs_buf_relse(bp);
2315                 goto error;
2316         }
2317         error = 0;
2318         ASSERT(in_f->ilf_fields & XFS_ILOG_CORE);
2319         dip = (xfs_dinode_t *)xfs_buf_offset(bp, imap.im_boffset);
2320
2321         /*
2322          * Make sure the place we're flushing out to really looks
2323          * like an inode!
2324          */
2325         if (unlikely(be16_to_cpu(dip->di_core.di_magic) != XFS_DINODE_MAGIC)) {
2326                 xfs_buf_relse(bp);
2327                 xfs_fs_cmn_err(CE_ALERT, mp,
2328                         "xfs_inode_recover: Bad inode magic number, dino ptr = 0x%p, dino bp = 0x%p, ino = %Ld",
2329                         dip, bp, ino);
2330                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(1)",
2331                                  XFS_ERRLEVEL_LOW, mp);
2332                 error = EFSCORRUPTED;
2333                 goto error;
2334         }
2335         dicp = (xfs_icdinode_t *)(item->ri_buf[1].i_addr);
2336         if (unlikely(dicp->di_magic != XFS_DINODE_MAGIC)) {
2337                 xfs_buf_relse(bp);
2338                 xfs_fs_cmn_err(CE_ALERT, mp,
2339                         "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, ino %Ld",
2340                         item, ino);
2341                 XFS_ERROR_REPORT("xlog_recover_do_inode_trans(2)",
2342                                  XFS_ERRLEVEL_LOW, mp);
2343                 error = EFSCORRUPTED;
2344                 goto error;
2345         }
2346
2347         /* Skip replay when the on disk inode is newer than the log one */
2348         if (dicp->di_flushiter < be16_to_cpu(dip->di_core.di_flushiter)) {
2349                 /*
2350                  * Deal with the wrap case, DI_MAX_FLUSH is less
2351                  * than smaller numbers
2352                  */
2353                 if (be16_to_cpu(dip->di_core.di_flushiter) == DI_MAX_FLUSH &&
2354                     dicp->di_flushiter < (DI_MAX_FLUSH >> 1)) {
2355                         /* do nothing */
2356                 } else {
2357                         xfs_buf_relse(bp);
2358                         error = 0;
2359                         goto error;
2360                 }
2361         }
2362         /* Take the opportunity to reset the flush iteration count */
2363         dicp->di_flushiter = 0;
2364
2365         if (unlikely((dicp->di_mode & S_IFMT) == S_IFREG)) {
2366                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2367                     (dicp->di_format != XFS_DINODE_FMT_BTREE)) {
2368                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(3)",
2369                                          XFS_ERRLEVEL_LOW, mp, dicp);
2370                         xfs_buf_relse(bp);
2371                         xfs_fs_cmn_err(CE_ALERT, mp,
2372                                 "xfs_inode_recover: Bad regular inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2373                                 item, dip, bp, ino);
2374                         error = EFSCORRUPTED;
2375                         goto error;
2376                 }
2377         } else if (unlikely((dicp->di_mode & S_IFMT) == S_IFDIR)) {
2378                 if ((dicp->di_format != XFS_DINODE_FMT_EXTENTS) &&
2379                     (dicp->di_format != XFS_DINODE_FMT_BTREE) &&
2380                     (dicp->di_format != XFS_DINODE_FMT_LOCAL)) {
2381                         XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(4)",
2382                                              XFS_ERRLEVEL_LOW, mp, dicp);
2383                         xfs_buf_relse(bp);
2384                         xfs_fs_cmn_err(CE_ALERT, mp,
2385                                 "xfs_inode_recover: Bad dir inode log record, rec ptr 0x%p, ino ptr = 0x%p, ino bp = 0x%p, ino %Ld",
2386                                 item, dip, bp, ino);
2387                         error = EFSCORRUPTED;
2388                         goto error;
2389                 }
2390         }
2391         if (unlikely(dicp->di_nextents + dicp->di_anextents > dicp->di_nblocks)){
2392                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(5)",
2393                                      XFS_ERRLEVEL_LOW, mp, dicp);
2394                 xfs_buf_relse(bp);
2395                 xfs_fs_cmn_err(CE_ALERT, mp,
2396                         "xfs_inode_recover: Bad inode log record, rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, total extents = %d, nblocks = %Ld",
2397                         item, dip, bp, ino,
2398                         dicp->di_nextents + dicp->di_anextents,
2399                         dicp->di_nblocks);
2400                 error = EFSCORRUPTED;
2401                 goto error;
2402         }
2403         if (unlikely(dicp->di_forkoff > mp->m_sb.sb_inodesize)) {
2404                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(6)",
2405                                      XFS_ERRLEVEL_LOW, mp, dicp);
2406                 xfs_buf_relse(bp);
2407                 xfs_fs_cmn_err(CE_ALERT, mp,
2408                         "xfs_inode_recover: Bad inode log rec ptr 0x%p, dino ptr 0x%p, dino bp 0x%p, ino %Ld, forkoff 0x%x",
2409                         item, dip, bp, ino, dicp->di_forkoff);
2410                 error = EFSCORRUPTED;
2411                 goto error;
2412         }
2413         if (unlikely(item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t))) {
2414                 XFS_CORRUPTION_ERROR("xlog_recover_do_inode_trans(7)",
2415                                      XFS_ERRLEVEL_LOW, mp, dicp);
2416                 xfs_buf_relse(bp);
2417                 xfs_fs_cmn_err(CE_ALERT, mp,
2418                         "xfs_inode_recover: Bad inode log record length %d, rec ptr 0x%p",
2419                         item->ri_buf[1].i_len, item);
2420                 error = EFSCORRUPTED;
2421                 goto error;
2422         }
2423
2424         /* The core is in in-core format */
2425         xfs_dinode_to_disk(&dip->di_core,
2426                 (xfs_icdinode_t *)item->ri_buf[1].i_addr);
2427
2428         /* the rest is in on-disk format */
2429         if (item->ri_buf[1].i_len > sizeof(xfs_dinode_core_t)) {
2430                 memcpy((xfs_caddr_t) dip + sizeof(xfs_dinode_core_t),
2431                         item->ri_buf[1].i_addr + sizeof(xfs_dinode_core_t),
2432                         item->ri_buf[1].i_len  - sizeof(xfs_dinode_core_t));
2433         }
2434
2435         fields = in_f->ilf_fields;
2436         switch (fields & (XFS_ILOG_DEV | XFS_ILOG_UUID)) {
2437         case XFS_ILOG_DEV:
2438                 dip->di_u.di_dev = cpu_to_be32(in_f->ilf_u.ilfu_rdev);
2439                 break;
2440         case XFS_ILOG_UUID:
2441                 dip->di_u.di_muuid = in_f->ilf_u.ilfu_uuid;
2442                 break;
2443         }
2444
2445         if (in_f->ilf_size == 2)
2446                 goto write_inode_buffer;
2447         len = item->ri_buf[2].i_len;
2448         src = item->ri_buf[2].i_addr;
2449         ASSERT(in_f->ilf_size <= 4);
2450         ASSERT((in_f->ilf_size == 3) || (fields & XFS_ILOG_AFORK));
2451         ASSERT(!(fields & XFS_ILOG_DFORK) ||
2452                (len == in_f->ilf_dsize));
2453
2454         switch (fields & XFS_ILOG_DFORK) {
2455         case XFS_ILOG_DDATA:
2456         case XFS_ILOG_DEXT:
2457                 memcpy(&dip->di_u, src, len);
2458                 break;
2459
2460         case XFS_ILOG_DBROOT:
2461                 xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2462                                  &(dip->di_u.di_bmbt),
2463                                  XFS_DFORK_DSIZE(dip, mp));
2464                 break;
2465
2466         default:
2467                 /*
2468                  * There are no data fork flags set.
2469                  */
2470                 ASSERT((fields & XFS_ILOG_DFORK) == 0);
2471                 break;
2472         }
2473
2474         /*
2475          * If we logged any attribute data, recover it.  There may or
2476          * may not have been any other non-core data logged in this
2477          * transaction.
2478          */
2479         if (in_f->ilf_fields & XFS_ILOG_AFORK) {
2480                 if (in_f->ilf_fields & XFS_ILOG_DFORK) {
2481                         attr_index = 3;
2482                 } else {
2483                         attr_index = 2;
2484                 }
2485                 len = item->ri_buf[attr_index].i_len;
2486                 src = item->ri_buf[attr_index].i_addr;
2487                 ASSERT(len == in_f->ilf_asize);
2488
2489                 switch (in_f->ilf_fields & XFS_ILOG_AFORK) {
2490                 case XFS_ILOG_ADATA:
2491                 case XFS_ILOG_AEXT:
2492                         dest = XFS_DFORK_APTR(dip);
2493                         ASSERT(len <= XFS_DFORK_ASIZE(dip, mp));
2494                         memcpy(dest, src, len);
2495                         break;
2496
2497                 case XFS_ILOG_ABROOT:
2498                         dest = XFS_DFORK_APTR(dip);
2499                         xfs_bmbt_to_bmdr((xfs_bmbt_block_t *)src, len,
2500                                          (xfs_bmdr_block_t*)dest,
2501                                          XFS_DFORK_ASIZE(dip, mp));
2502                         break;
2503
2504                 default:
2505                         xlog_warn("XFS: xlog_recover_do_inode_trans: Invalid flag");
2506                         ASSERT(0);
2507                         xfs_buf_relse(bp);
2508                         error = EIO;
2509                         goto error;
2510                 }
2511         }
2512
2513 write_inode_buffer:
2514         if (ITEM_TYPE(item) == XFS_LI_INODE) {
2515                 ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2516                        XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2517                 XFS_BUF_SET_FSPRIVATE(bp, mp);
2518                 XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2519                 xfs_bdwrite(mp, bp);
2520         } else {
2521                 XFS_BUF_STALE(bp);
2522                 error = xfs_bwrite(mp, bp);
2523         }
2524
2525 error:
2526         if (need_free)
2527                 kmem_free(in_f);
2528         return XFS_ERROR(error);
2529 }
2530
2531 /*
2532  * Recover QUOTAOFF records. We simply make a note of it in the xlog_t
2533  * structure, so that we know not to do any dquot item or dquot buffer recovery,
2534  * of that type.
2535  */
2536 STATIC int
2537 xlog_recover_do_quotaoff_trans(
2538         xlog_t                  *log,
2539         xlog_recover_item_t     *item,
2540         int                     pass)
2541 {
2542         xfs_qoff_logformat_t    *qoff_f;
2543
2544         if (pass == XLOG_RECOVER_PASS2) {
2545                 return (0);
2546         }
2547
2548         qoff_f = (xfs_qoff_logformat_t *)item->ri_buf[0].i_addr;
2549         ASSERT(qoff_f);
2550
2551         /*
2552          * The logitem format's flag tells us if this was user quotaoff,
2553          * group/project quotaoff or both.
2554          */
2555         if (qoff_f->qf_flags & XFS_UQUOTA_ACCT)
2556                 log->l_quotaoffs_flag |= XFS_DQ_USER;
2557         if (qoff_f->qf_flags & XFS_PQUOTA_ACCT)
2558                 log->l_quotaoffs_flag |= XFS_DQ_PROJ;
2559         if (qoff_f->qf_flags & XFS_GQUOTA_ACCT)
2560                 log->l_quotaoffs_flag |= XFS_DQ_GROUP;
2561
2562         return (0);
2563 }
2564
2565 /*
2566  * Recover a dquot record
2567  */
2568 STATIC int
2569 xlog_recover_do_dquot_trans(
2570         xlog_t                  *log,
2571         xlog_recover_item_t     *item,
2572         int                     pass)
2573 {
2574         xfs_mount_t             *mp;
2575         xfs_buf_t               *bp;
2576         struct xfs_disk_dquot   *ddq, *recddq;
2577         int                     error;
2578         xfs_dq_logformat_t      *dq_f;
2579         uint                    type;
2580
2581         if (pass == XLOG_RECOVER_PASS1) {
2582                 return 0;
2583         }
2584         mp = log->l_mp;
2585
2586         /*
2587          * Filesystems are required to send in quota flags at mount time.
2588          */
2589         if (mp->m_qflags == 0)
2590                 return (0);
2591
2592         recddq = (xfs_disk_dquot_t *)item->ri_buf[1].i_addr;
2593         ASSERT(recddq);
2594         /*
2595          * This type of quotas was turned off, so ignore this record.
2596          */
2597         type = recddq->d_flags & (XFS_DQ_USER | XFS_DQ_PROJ | XFS_DQ_GROUP);
2598         ASSERT(type);
2599         if (log->l_quotaoffs_flag & type)
2600                 return (0);
2601
2602         /*
2603          * At this point we know that quota was _not_ turned off.
2604          * Since the mount flags are not indicating to us otherwise, this
2605          * must mean that quota is on, and the dquot needs to be replayed.
2606          * Remember that we may not have fully recovered the superblock yet,
2607          * so we can't do the usual trick of looking at the SB quota bits.
2608          *
2609          * The other possibility, of course, is that the quota subsystem was
2610          * removed since the last mount - ENOSYS.
2611          */
2612         dq_f = (xfs_dq_logformat_t *)item->ri_buf[0].i_addr;
2613         ASSERT(dq_f);
2614         if ((error = xfs_qm_dqcheck(recddq,
2615                            dq_f->qlf_id,
2616                            0, XFS_QMOPT_DOWARN,
2617                            "xlog_recover_do_dquot_trans (log copy)"))) {
2618                 return XFS_ERROR(EIO);
2619         }
2620         ASSERT(dq_f->qlf_len == 1);
2621
2622         error = xfs_read_buf(mp, mp->m_ddev_targp,
2623                              dq_f->qlf_blkno,
2624                              XFS_FSB_TO_BB(mp, dq_f->qlf_len),
2625                              0, &bp);
2626         if (error) {
2627                 xfs_ioerror_alert("xlog_recover_do..(read#3)", mp,
2628                                   bp, dq_f->qlf_blkno);
2629                 return error;
2630         }
2631         ASSERT(bp);
2632         ddq = (xfs_disk_dquot_t *)xfs_buf_offset(bp, dq_f->qlf_boffset);
2633
2634         /*
2635          * At least the magic num portion should be on disk because this
2636          * was among a chunk of dquots created earlier, and we did some
2637          * minimal initialization then.
2638          */
2639         if (xfs_qm_dqcheck(ddq, dq_f->qlf_id, 0, XFS_QMOPT_DOWARN,
2640                            "xlog_recover_do_dquot_trans")) {
2641                 xfs_buf_relse(bp);
2642                 return XFS_ERROR(EIO);
2643         }
2644
2645         memcpy(ddq, recddq, item->ri_buf[1].i_len);
2646
2647         ASSERT(dq_f->qlf_size == 2);
2648         ASSERT(XFS_BUF_FSPRIVATE(bp, void *) == NULL ||
2649                XFS_BUF_FSPRIVATE(bp, xfs_mount_t *) == mp);
2650         XFS_BUF_SET_FSPRIVATE(bp, mp);
2651         XFS_BUF_SET_IODONE_FUNC(bp, xlog_recover_iodone);
2652         xfs_bdwrite(mp, bp);
2653
2654         return (0);
2655 }
2656
2657 /*
2658  * This routine is called to create an in-core extent free intent
2659  * item from the efi format structure which was logged on disk.
2660  * It allocates an in-core efi, copies the extents from the format
2661  * structure into it, and adds the efi to the AIL with the given
2662  * LSN.
2663  */
2664 STATIC int
2665 xlog_recover_do_efi_trans(
2666         xlog_t                  *log,
2667         xlog_recover_item_t     *item,
2668         xfs_lsn_t               lsn,
2669         int                     pass)
2670 {
2671         int                     error;
2672         xfs_mount_t             *mp;
2673         xfs_efi_log_item_t      *efip;
2674         xfs_efi_log_format_t    *efi_formatp;
2675
2676         if (pass == XLOG_RECOVER_PASS1) {
2677                 return 0;
2678         }
2679
2680         efi_formatp = (xfs_efi_log_format_t *)item->ri_buf[0].i_addr;
2681
2682         mp = log->l_mp;
2683         efip = xfs_efi_init(mp, efi_formatp->efi_nextents);
2684         if ((error = xfs_efi_copy_format(&(item->ri_buf[0]),
2685                                          &(efip->efi_format)))) {
2686                 xfs_efi_item_free(efip);
2687                 return error;
2688         }
2689         efip->efi_next_extent = efi_formatp->efi_nextents;
2690         efip->efi_flags |= XFS_EFI_COMMITTED;
2691
2692         spin_lock(&mp->m_ail_lock);
2693         /*
2694          * xfs_trans_update_ail() drops the AIL lock.
2695          */
2696         xfs_trans_update_ail(mp, (xfs_log_item_t *)efip, lsn);
2697         return 0;
2698 }
2699
2700
2701 /*
2702  * This routine is called when an efd format structure is found in
2703  * a committed transaction in the log.  It's purpose is to cancel
2704  * the corresponding efi if it was still in the log.  To do this
2705  * it searches the AIL for the efi with an id equal to that in the
2706  * efd format structure.  If we find it, we remove the efi from the
2707  * AIL and free it.
2708  */
2709 STATIC void
2710 xlog_recover_do_efd_trans(
2711         xlog_t                  *log,
2712         xlog_recover_item_t     *item,
2713         int                     pass)
2714 {
2715         xfs_mount_t             *mp;
2716         xfs_efd_log_format_t    *efd_formatp;
2717         xfs_efi_log_item_t      *efip = NULL;
2718         xfs_log_item_t          *lip;
2719         int                     gen;
2720         __uint64_t              efi_id;
2721
2722         if (pass == XLOG_RECOVER_PASS1) {
2723                 return;
2724         }
2725
2726         efd_formatp = (xfs_efd_log_format_t *)item->ri_buf[0].i_addr;
2727         ASSERT((item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_32_t) +
2728                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_32_t)))) ||
2729                (item->ri_buf[0].i_len == (sizeof(xfs_efd_log_format_64_t) +
2730                 ((efd_formatp->efd_nextents - 1) * sizeof(xfs_extent_64_t)))));
2731         efi_id = efd_formatp->efd_efi_id;
2732
2733         /*
2734          * Search for the efi with the id in the efd format structure
2735          * in the AIL.
2736          */
2737         mp = log->l_mp;
2738         spin_lock(&mp->m_ail_lock);
2739         lip = xfs_trans_first_ail(mp, &gen);
2740         while (lip != NULL) {
2741                 if (lip->li_type == XFS_LI_EFI) {
2742                         efip = (xfs_efi_log_item_t *)lip;
2743                         if (efip->efi_format.efi_id == efi_id) {
2744                                 /*
2745                                  * xfs_trans_delete_ail() drops the
2746                                  * AIL lock.
2747                                  */
2748                                 xfs_trans_delete_ail(mp, lip);
2749                                 xfs_efi_item_free(efip);
2750                                 return;
2751                         }
2752                 }
2753                 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
2754         }
2755         spin_unlock(&mp->m_ail_lock);
2756 }
2757
2758 /*
2759  * Perform the transaction
2760  *
2761  * If the transaction modifies a buffer or inode, do it now.  Otherwise,
2762  * EFIs and EFDs get queued up by adding entries into the AIL for them.
2763  */
2764 STATIC int
2765 xlog_recover_do_trans(
2766         xlog_t                  *log,
2767         xlog_recover_t          *trans,
2768         int                     pass)
2769 {
2770         int                     error = 0;
2771         xlog_recover_item_t     *item, *first_item;
2772
2773         if ((error = xlog_recover_reorder_trans(trans)))
2774                 return error;
2775         first_item = item = trans->r_itemq;
2776         do {
2777                 /*
2778                  * we don't need to worry about the block number being
2779                  * truncated in > 1 TB buffers because in user-land,
2780                  * we're now n32 or 64-bit so xfs_daddr_t is 64-bits so
2781                  * the blknos will get through the user-mode buffer
2782                  * cache properly.  The only bad case is o32 kernels
2783                  * where xfs_daddr_t is 32-bits but mount will warn us
2784                  * off a > 1 TB filesystem before we get here.
2785                  */
2786                 if ((ITEM_TYPE(item) == XFS_LI_BUF)) {
2787                         if  ((error = xlog_recover_do_buffer_trans(log, item,
2788                                                                  pass)))
2789                                 break;
2790                 } else if ((ITEM_TYPE(item) == XFS_LI_INODE)) {
2791                         if ((error = xlog_recover_do_inode_trans(log, item,
2792                                                                 pass)))
2793                                 break;
2794                 } else if (ITEM_TYPE(item) == XFS_LI_EFI) {
2795                         if ((error = xlog_recover_do_efi_trans(log, item, trans->r_lsn,
2796                                                   pass)))
2797                                 break;
2798                 } else if (ITEM_TYPE(item) == XFS_LI_EFD) {
2799                         xlog_recover_do_efd_trans(log, item, pass);
2800                 } else if (ITEM_TYPE(item) == XFS_LI_DQUOT) {
2801                         if ((error = xlog_recover_do_dquot_trans(log, item,
2802                                                                    pass)))
2803                                         break;
2804                 } else if ((ITEM_TYPE(item) == XFS_LI_QUOTAOFF)) {
2805                         if ((error = xlog_recover_do_quotaoff_trans(log, item,
2806                                                                    pass)))
2807                                         break;
2808                 } else {
2809                         xlog_warn("XFS: xlog_recover_do_trans");
2810                         ASSERT(0);
2811                         error = XFS_ERROR(EIO);
2812                         break;
2813                 }
2814                 item = item->ri_next;
2815         } while (first_item != item);
2816
2817         return error;
2818 }
2819
2820 /*
2821  * Free up any resources allocated by the transaction
2822  *
2823  * Remember that EFIs, EFDs, and IUNLINKs are handled later.
2824  */
2825 STATIC void
2826 xlog_recover_free_trans(
2827         xlog_recover_t          *trans)
2828 {
2829         xlog_recover_item_t     *first_item, *item, *free_item;
2830         int                     i;
2831
2832         item = first_item = trans->r_itemq;
2833         do {
2834                 free_item = item;
2835                 item = item->ri_next;
2836                  /* Free the regions in the item. */
2837                 for (i = 0; i < free_item->ri_cnt; i++) {
2838                         kmem_free(free_item->ri_buf[i].i_addr);
2839                 }
2840                 /* Free the item itself */
2841                 kmem_free(free_item->ri_buf);
2842                 kmem_free(free_item);
2843         } while (first_item != item);
2844         /* Free the transaction recover structure */
2845         kmem_free(trans);
2846 }
2847
2848 STATIC int
2849 xlog_recover_commit_trans(
2850         xlog_t                  *log,
2851         xlog_recover_t          **q,
2852         xlog_recover_t          *trans,
2853         int                     pass)
2854 {
2855         int                     error;
2856
2857         if ((error = xlog_recover_unlink_tid(q, trans)))
2858                 return error;
2859         if ((error = xlog_recover_do_trans(log, trans, pass)))
2860                 return error;
2861         xlog_recover_free_trans(trans);                 /* no error */
2862         return 0;
2863 }
2864
2865 STATIC int
2866 xlog_recover_unmount_trans(
2867         xlog_recover_t          *trans)
2868 {
2869         /* Do nothing now */
2870         xlog_warn("XFS: xlog_recover_unmount_trans: Unmount LR");
2871         return 0;
2872 }
2873
2874 /*
2875  * There are two valid states of the r_state field.  0 indicates that the
2876  * transaction structure is in a normal state.  We have either seen the
2877  * start of the transaction or the last operation we added was not a partial
2878  * operation.  If the last operation we added to the transaction was a
2879  * partial operation, we need to mark r_state with XLOG_WAS_CONT_TRANS.
2880  *
2881  * NOTE: skip LRs with 0 data length.
2882  */
2883 STATIC int
2884 xlog_recover_process_data(
2885         xlog_t                  *log,
2886         xlog_recover_t          *rhash[],
2887         xlog_rec_header_t       *rhead,
2888         xfs_caddr_t             dp,
2889         int                     pass)
2890 {
2891         xfs_caddr_t             lp;
2892         int                     num_logops;
2893         xlog_op_header_t        *ohead;
2894         xlog_recover_t          *trans;
2895         xlog_tid_t              tid;
2896         int                     error;
2897         unsigned long           hash;
2898         uint                    flags;
2899
2900         lp = dp + be32_to_cpu(rhead->h_len);
2901         num_logops = be32_to_cpu(rhead->h_num_logops);
2902
2903         /* check the log format matches our own - else we can't recover */
2904         if (xlog_header_check_recover(log->l_mp, rhead))
2905                 return (XFS_ERROR(EIO));
2906
2907         while ((dp < lp) && num_logops) {
2908                 ASSERT(dp + sizeof(xlog_op_header_t) <= lp);
2909                 ohead = (xlog_op_header_t *)dp;
2910                 dp += sizeof(xlog_op_header_t);
2911                 if (ohead->oh_clientid != XFS_TRANSACTION &&
2912                     ohead->oh_clientid != XFS_LOG) {
2913                         xlog_warn(
2914                 "XFS: xlog_recover_process_data: bad clientid");
2915                         ASSERT(0);
2916                         return (XFS_ERROR(EIO));
2917                 }
2918                 tid = be32_to_cpu(ohead->oh_tid);
2919                 hash = XLOG_RHASH(tid);
2920                 trans = xlog_recover_find_tid(rhash[hash], tid);
2921                 if (trans == NULL) {               /* not found; add new tid */
2922                         if (ohead->oh_flags & XLOG_START_TRANS)
2923                                 xlog_recover_new_tid(&rhash[hash], tid,
2924                                         be64_to_cpu(rhead->h_lsn));
2925                 } else {
2926                         if (dp + be32_to_cpu(ohead->oh_len) > lp) {
2927                                 xlog_warn(
2928                         "XFS: xlog_recover_process_data: bad length");
2929                                 WARN_ON(1);
2930                                 return (XFS_ERROR(EIO));
2931                         }
2932                         flags = ohead->oh_flags & ~XLOG_END_TRANS;
2933                         if (flags & XLOG_WAS_CONT_TRANS)
2934                                 flags &= ~XLOG_CONTINUE_TRANS;
2935                         switch (flags) {
2936                         case XLOG_COMMIT_TRANS:
2937                                 error = xlog_recover_commit_trans(log,
2938                                                 &rhash[hash], trans, pass);
2939                                 break;
2940                         case XLOG_UNMOUNT_TRANS:
2941                                 error = xlog_recover_unmount_trans(trans);
2942                                 break;
2943                         case XLOG_WAS_CONT_TRANS:
2944                                 error = xlog_recover_add_to_cont_trans(trans,
2945                                                 dp, be32_to_cpu(ohead->oh_len));
2946                                 break;
2947                         case XLOG_START_TRANS:
2948                                 xlog_warn(
2949                         "XFS: xlog_recover_process_data: bad transaction");
2950                                 ASSERT(0);
2951                                 error = XFS_ERROR(EIO);
2952                                 break;
2953                         case 0:
2954                         case XLOG_CONTINUE_TRANS:
2955                                 error = xlog_recover_add_to_trans(trans,
2956                                                 dp, be32_to_cpu(ohead->oh_len));
2957                                 break;
2958                         default:
2959                                 xlog_warn(
2960                         "XFS: xlog_recover_process_data: bad flag");
2961                                 ASSERT(0);
2962                                 error = XFS_ERROR(EIO);
2963                                 break;
2964                         }
2965                         if (error)
2966                                 return error;
2967                 }
2968                 dp += be32_to_cpu(ohead->oh_len);
2969                 num_logops--;
2970         }
2971         return 0;
2972 }
2973
2974 /*
2975  * Process an extent free intent item that was recovered from
2976  * the log.  We need to free the extents that it describes.
2977  */
2978 STATIC int
2979 xlog_recover_process_efi(
2980         xfs_mount_t             *mp,
2981         xfs_efi_log_item_t      *efip)
2982 {
2983         xfs_efd_log_item_t      *efdp;
2984         xfs_trans_t             *tp;
2985         int                     i;
2986         int                     error = 0;
2987         xfs_extent_t            *extp;
2988         xfs_fsblock_t           startblock_fsb;
2989
2990         ASSERT(!(efip->efi_flags & XFS_EFI_RECOVERED));
2991
2992         /*
2993          * First check the validity of the extents described by the
2994          * EFI.  If any are bad, then assume that all are bad and
2995          * just toss the EFI.
2996          */
2997         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
2998                 extp = &(efip->efi_format.efi_extents[i]);
2999                 startblock_fsb = XFS_BB_TO_FSB(mp,
3000                                    XFS_FSB_TO_DADDR(mp, extp->ext_start));
3001                 if ((startblock_fsb == 0) ||
3002                     (extp->ext_len == 0) ||
3003                     (startblock_fsb >= mp->m_sb.sb_dblocks) ||
3004                     (extp->ext_len >= mp->m_sb.sb_agblocks)) {
3005                         /*
3006                          * This will pull the EFI from the AIL and
3007                          * free the memory associated with it.
3008                          */
3009                         xfs_efi_release(efip, efip->efi_format.efi_nextents);
3010                         return XFS_ERROR(EIO);
3011                 }
3012         }
3013
3014         tp = xfs_trans_alloc(mp, 0);
3015         error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0, 0, 0);
3016         if (error)
3017                 goto abort_error;
3018         efdp = xfs_trans_get_efd(tp, efip, efip->efi_format.efi_nextents);
3019
3020         for (i = 0; i < efip->efi_format.efi_nextents; i++) {
3021                 extp = &(efip->efi_format.efi_extents[i]);
3022                 error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
3023                 if (error)
3024                         goto abort_error;
3025                 xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
3026                                          extp->ext_len);
3027         }
3028
3029         efip->efi_flags |= XFS_EFI_RECOVERED;
3030         error = xfs_trans_commit(tp, 0);
3031         return error;
3032
3033 abort_error:
3034         xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3035         return error;
3036 }
3037
3038 /*
3039  * Verify that once we've encountered something other than an EFI
3040  * in the AIL that there are no more EFIs in the AIL.
3041  */
3042 #if defined(DEBUG)
3043 STATIC void
3044 xlog_recover_check_ail(
3045         xfs_mount_t             *mp,
3046         xfs_log_item_t          *lip,
3047         int                     gen)
3048 {
3049         int                     orig_gen = gen;
3050
3051         do {
3052                 ASSERT(lip->li_type != XFS_LI_EFI);
3053                 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3054                 /*
3055                  * The check will be bogus if we restart from the
3056                  * beginning of the AIL, so ASSERT that we don't.
3057                  * We never should since we're holding the AIL lock
3058                  * the entire time.
3059                  */
3060                 ASSERT(gen == orig_gen);
3061         } while (lip != NULL);
3062 }
3063 #endif  /* DEBUG */
3064
3065 /*
3066  * When this is called, all of the EFIs which did not have
3067  * corresponding EFDs should be in the AIL.  What we do now
3068  * is free the extents associated with each one.
3069  *
3070  * Since we process the EFIs in normal transactions, they
3071  * will be removed at some point after the commit.  This prevents
3072  * us from just walking down the list processing each one.
3073  * We'll use a flag in the EFI to skip those that we've already
3074  * processed and use the AIL iteration mechanism's generation
3075  * count to try to speed this up at least a bit.
3076  *
3077  * When we start, we know that the EFIs are the only things in
3078  * the AIL.  As we process them, however, other items are added
3079  * to the AIL.  Since everything added to the AIL must come after
3080  * everything already in the AIL, we stop processing as soon as
3081  * we see something other than an EFI in the AIL.
3082  */
3083 STATIC int
3084 xlog_recover_process_efis(
3085         xlog_t                  *log)
3086 {
3087         xfs_log_item_t          *lip;
3088         xfs_efi_log_item_t      *efip;
3089         int                     gen;
3090         xfs_mount_t             *mp;
3091         int                     error = 0;
3092
3093         mp = log->l_mp;
3094         spin_lock(&mp->m_ail_lock);
3095
3096         lip = xfs_trans_first_ail(mp, &gen);
3097         while (lip != NULL) {
3098                 /*
3099                  * We're done when we see something other than an EFI.
3100                  */
3101                 if (lip->li_type != XFS_LI_EFI) {
3102                         xlog_recover_check_ail(mp, lip, gen);
3103                         break;
3104                 }
3105
3106                 /*
3107                  * Skip EFIs that we've already processed.
3108                  */
3109                 efip = (xfs_efi_log_item_t *)lip;
3110                 if (efip->efi_flags & XFS_EFI_RECOVERED) {
3111                         lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3112                         continue;
3113                 }
3114
3115                 spin_unlock(&mp->m_ail_lock);
3116                 error = xlog_recover_process_efi(mp, efip);
3117                 if (error)
3118                         return error;
3119                 spin_lock(&mp->m_ail_lock);
3120                 lip = xfs_trans_next_ail(mp, lip, &gen, NULL);
3121         }
3122         spin_unlock(&mp->m_ail_lock);
3123         return error;
3124 }
3125
3126 /*
3127  * This routine performs a transaction to null out a bad inode pointer
3128  * in an agi unlinked inode hash bucket.
3129  */
3130 STATIC void
3131 xlog_recover_clear_agi_bucket(
3132         xfs_mount_t     *mp,
3133         xfs_agnumber_t  agno,
3134         int             bucket)
3135 {
3136         xfs_trans_t     *tp;
3137         xfs_agi_t       *agi;
3138         xfs_buf_t       *agibp;
3139         int             offset;
3140         int             error;
3141
3142         tp = xfs_trans_alloc(mp, XFS_TRANS_CLEAR_AGI_BUCKET);
3143         error = xfs_trans_reserve(tp, 0, XFS_CLEAR_AGI_BUCKET_LOG_RES(mp), 0, 0, 0);
3144         if (!error)
3145                 error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
3146                                    XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3147                                    XFS_FSS_TO_BB(mp, 1), 0, &agibp);
3148         if (error)
3149                 goto out_abort;
3150
3151         error = EINVAL;
3152         agi = XFS_BUF_TO_AGI(agibp);
3153         if (be32_to_cpu(agi->agi_magicnum) != XFS_AGI_MAGIC)
3154                 goto out_abort;
3155
3156         agi->agi_unlinked[bucket] = cpu_to_be32(NULLAGINO);
3157         offset = offsetof(xfs_agi_t, agi_unlinked) +
3158                  (sizeof(xfs_agino_t) * bucket);
3159         xfs_trans_log_buf(tp, agibp, offset,
3160                           (offset + sizeof(xfs_agino_t) - 1));
3161
3162         error = xfs_trans_commit(tp, 0);
3163         if (error)
3164                 goto out_error;
3165         return;
3166
3167 out_abort:
3168         xfs_trans_cancel(tp, XFS_TRANS_ABORT);
3169 out_error:
3170         xfs_fs_cmn_err(CE_WARN, mp, "xlog_recover_clear_agi_bucket: "
3171                         "failed to clear agi %d. Continuing.", agno);
3172         return;
3173 }
3174
3175 /*
3176  * xlog_iunlink_recover
3177  *
3178  * This is called during recovery to process any inodes which
3179  * we unlinked but not freed when the system crashed.  These
3180  * inodes will be on the lists in the AGI blocks.  What we do
3181  * here is scan all the AGIs and fully truncate and free any
3182  * inodes found on the lists.  Each inode is removed from the
3183  * lists when it has been fully truncated and is freed.  The
3184  * freeing of the inode and its removal from the list must be
3185  * atomic.
3186  */
3187 void
3188 xlog_recover_process_iunlinks(
3189         xlog_t          *log)
3190 {
3191         xfs_mount_t     *mp;
3192         xfs_agnumber_t  agno;
3193         xfs_agi_t       *agi;
3194         xfs_buf_t       *agibp;
3195         xfs_buf_t       *ibp;
3196         xfs_dinode_t    *dip;
3197         xfs_inode_t     *ip;
3198         xfs_agino_t     agino;
3199         xfs_ino_t       ino;
3200         int             bucket;
3201         int             error;
3202         uint            mp_dmevmask;
3203
3204         mp = log->l_mp;
3205
3206         /*
3207          * Prevent any DMAPI event from being sent while in this function.
3208          */
3209         mp_dmevmask = mp->m_dmevmask;
3210         mp->m_dmevmask = 0;
3211
3212         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
3213                 /*
3214                  * Find the agi for this ag.
3215                  */
3216                 agibp = xfs_buf_read(mp->m_ddev_targp,
3217                                 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)),
3218                                 XFS_FSS_TO_BB(mp, 1), 0);
3219                 if (XFS_BUF_ISERROR(agibp)) {
3220                         xfs_ioerror_alert("xlog_recover_process_iunlinks(#1)",
3221                                 log->l_mp, agibp,
3222                                 XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp)));
3223                 }
3224                 agi = XFS_BUF_TO_AGI(agibp);
3225                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agi->agi_magicnum));
3226
3227                 for (bucket = 0; bucket < XFS_AGI_UNLINKED_BUCKETS; bucket++) {
3228
3229                         agino = be32_to_cpu(agi->agi_unlinked[bucket]);
3230                         while (agino != NULLAGINO) {
3231
3232                                 /*
3233                                  * Release the agi buffer so that it can
3234                                  * be acquired in the normal course of the
3235                                  * transaction to truncate and free the inode.
3236                                  */
3237                                 xfs_buf_relse(agibp);
3238
3239                                 ino = XFS_AGINO_TO_INO(mp, agno, agino);
3240                                 error = xfs_iget(mp, NULL, ino, 0, 0, &ip, 0);
3241                                 ASSERT(error || (ip != NULL));
3242
3243                                 if (!error) {
3244                                         /*
3245                                          * Get the on disk inode to find the
3246                                          * next inode in the bucket.
3247                                          */
3248                                         error = xfs_itobp(mp, NULL, ip, &dip,
3249                                                         &ibp, 0, 0,
3250                                                         XFS_BUF_LOCK);
3251                                         ASSERT(error || (dip != NULL));
3252                                 }
3253
3254                                 if (!error) {
3255                                         ASSERT(ip->i_d.di_nlink == 0);
3256
3257                                         /* setup for the next pass */
3258                                         agino = be32_to_cpu(
3259                                                         dip->di_next_unlinked);
3260                                         xfs_buf_relse(ibp);
3261                                         /*
3262                                          * Prevent any DMAPI event from
3263                                          * being sent when the
3264                                          * reference on the inode is
3265                                          * dropped.
3266                                          */
3267                                         ip->i_d.di_dmevmask = 0;
3268
3269                                         /*
3270                                          * If this is a new inode, handle
3271                                          * it specially.  Otherwise,
3272                                          * just drop our reference to the
3273                                          * inode.  If there are no
3274                                          * other references, this will
3275                                          * send the inode to
3276                                          * xfs_inactive() which will
3277                                          * truncate the file and free
3278                                          * the inode.
3279                                          */
3280                                         if (ip->i_d.di_mode == 0)
3281                                                 xfs_iput_new(ip, 0);
3282                                         else
3283                                                 IRELE(ip);
3284                                 } else {
3285                                         /*
3286                                          * We can't read in the inode
3287                                          * this bucket points to, or
3288                                          * this inode is messed up.  Just
3289                                          * ditch this bucket of inodes.  We
3290                                          * will lose some inodes and space,
3291                                          * but at least we won't hang.  Call
3292                                          * xlog_recover_clear_agi_bucket()
3293                                          * to perform a transaction to clear
3294                                          * the inode pointer in the bucket.
3295                                          */
3296                                         xlog_recover_clear_agi_bucket(mp, agno,
3297                                                         bucket);
3298
3299                                         agino = NULLAGINO;
3300                                 }
3301
3302                                 /*
3303                                  * Reacquire the agibuffer and continue around
3304                                  * the loop.
3305                                  */
3306                                 agibp = xfs_buf_read(mp->m_ddev_targp,
3307                                                 XFS_AG_DADDR(mp, agno,
3308                                                         XFS_AGI_DADDR(mp)),
3309                                                 XFS_FSS_TO_BB(mp, 1), 0);
3310                                 if (XFS_BUF_ISERROR(agibp)) {
3311                                         xfs_ioerror_alert(
3312                                 "xlog_recover_process_iunlinks(#2)",
3313                                                 log->l_mp, agibp,
3314                                                 XFS_AG_DADDR(mp, agno,
3315                                                         XFS_AGI_DADDR(mp)));
3316                                 }
3317                                 agi = XFS_BUF_TO_AGI(agibp);
3318                                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(
3319                                         agi->agi_magicnum));
3320                         }
3321                 }
3322
3323                 /*
3324                  * Release the buffer for the current agi so we can
3325                  * go on to the next one.
3326                  */
3327                 xfs_buf_relse(agibp);
3328         }
3329
3330         mp->m_dmevmask = mp_dmevmask;
3331 }
3332
3333
3334 #ifdef DEBUG
3335 STATIC void
3336 xlog_pack_data_checksum(
3337         xlog_t          *log,
3338         xlog_in_core_t  *iclog,
3339         int             size)
3340 {
3341         int             i;
3342         __be32          *up;
3343         uint            chksum = 0;
3344
3345         up = (__be32 *)iclog->ic_datap;
3346         /* divide length by 4 to get # words */
3347         for (i = 0; i < (size >> 2); i++) {
3348                 chksum ^= be32_to_cpu(*up);
3349                 up++;
3350         }
3351         iclog->ic_header.h_chksum = cpu_to_be32(chksum);
3352 }
3353 #else
3354 #define xlog_pack_data_checksum(log, iclog, size)
3355 #endif
3356
3357 /*
3358  * Stamp cycle number in every block
3359  */
3360 void
3361 xlog_pack_data(
3362         xlog_t                  *log,
3363         xlog_in_core_t          *iclog,
3364         int                     roundoff)
3365 {
3366         int                     i, j, k;
3367         int                     size = iclog->ic_offset + roundoff;
3368         __be32                  cycle_lsn;
3369         xfs_caddr_t             dp;
3370         xlog_in_core_2_t        *xhdr;
3371
3372         xlog_pack_data_checksum(log, iclog, size);
3373
3374         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
3375
3376         dp = iclog->ic_datap;
3377         for (i = 0; i < BTOBB(size) &&
3378                 i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3379                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
3380                 *(__be32 *)dp = cycle_lsn;
3381                 dp += BBSIZE;
3382         }
3383
3384         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3385                 xhdr = (xlog_in_core_2_t *)&iclog->ic_header;
3386                 for ( ; i < BTOBB(size); i++) {
3387                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3388                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3389                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
3390                         *(__be32 *)dp = cycle_lsn;
3391                         dp += BBSIZE;
3392                 }
3393
3394                 for (i = 1; i < log->l_iclog_heads; i++) {
3395                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
3396                 }
3397         }
3398 }
3399
3400 #if defined(DEBUG) && defined(XFS_LOUD_RECOVERY)
3401 STATIC void
3402 xlog_unpack_data_checksum(
3403         xlog_rec_header_t       *rhead,
3404         xfs_caddr_t             dp,
3405         xlog_t                  *log)
3406 {
3407         __be32                  *up = (__be32 *)dp;
3408         uint                    chksum = 0;
3409         int                     i;
3410
3411         /* divide length by 4 to get # words */
3412         for (i=0; i < be32_to_cpu(rhead->h_len) >> 2; i++) {
3413                 chksum ^= be32_to_cpu(*up);
3414                 up++;
3415         }
3416         if (chksum != be32_to_cpu(rhead->h_chksum)) {
3417             if (rhead->h_chksum ||
3418                 ((log->l_flags & XLOG_CHKSUM_MISMATCH) == 0)) {
3419                     cmn_err(CE_DEBUG,
3420                         "XFS: LogR chksum mismatch: was (0x%x) is (0x%x)\n",
3421                             be32_to_cpu(rhead->h_chksum), chksum);
3422                     cmn_err(CE_DEBUG,
3423 "XFS: Disregard message if filesystem was created with non-DEBUG kernel");
3424                     if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3425                             cmn_err(CE_DEBUG,
3426                                 "XFS: LogR this is a LogV2 filesystem\n");
3427                     }
3428                     log->l_flags |= XLOG_CHKSUM_MISMATCH;
3429             }
3430         }
3431 }
3432 #else
3433 #define xlog_unpack_data_checksum(rhead, dp, log)
3434 #endif
3435
3436 STATIC void
3437 xlog_unpack_data(
3438         xlog_rec_header_t       *rhead,
3439         xfs_caddr_t             dp,
3440         xlog_t                  *log)
3441 {
3442         int                     i, j, k;
3443         xlog_in_core_2_t        *xhdr;
3444
3445         for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
3446                   i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {
3447                 *(__be32 *)dp = *(__be32 *)&rhead->h_cycle_data[i];
3448                 dp += BBSIZE;
3449         }
3450
3451         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3452                 xhdr = (xlog_in_core_2_t *)rhead;
3453                 for ( ; i < BTOBB(be32_to_cpu(rhead->h_len)); i++) {
3454                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3455                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3456                         *(__be32 *)dp = xhdr[j].hic_xheader.xh_cycle_data[k];
3457                         dp += BBSIZE;
3458                 }
3459         }
3460
3461         xlog_unpack_data_checksum(rhead, dp, log);
3462 }
3463
3464 STATIC int
3465 xlog_valid_rec_header(
3466         xlog_t                  *log,
3467         xlog_rec_header_t       *rhead,
3468         xfs_daddr_t             blkno)
3469 {
3470         int                     hlen;
3471
3472         if (unlikely(be32_to_cpu(rhead->h_magicno) != XLOG_HEADER_MAGIC_NUM)) {
3473                 XFS_ERROR_REPORT("xlog_valid_rec_header(1)",
3474                                 XFS_ERRLEVEL_LOW, log->l_mp);
3475                 return XFS_ERROR(EFSCORRUPTED);
3476         }
3477         if (unlikely(
3478             (!rhead->h_version ||
3479             (be32_to_cpu(rhead->h_version) & (~XLOG_VERSION_OKBITS))))) {
3480                 xlog_warn("XFS: %s: unrecognised log version (%d).",
3481                         __func__, be32_to_cpu(rhead->h_version));
3482                 return XFS_ERROR(EIO);
3483         }
3484
3485         /* LR body must have data or it wouldn't have been written */
3486         hlen = be32_to_cpu(rhead->h_len);
3487         if (unlikely( hlen <= 0 || hlen > INT_MAX )) {
3488                 XFS_ERROR_REPORT("xlog_valid_rec_header(2)",
3489                                 XFS_ERRLEVEL_LOW, log->l_mp);
3490                 return XFS_ERROR(EFSCORRUPTED);
3491         }
3492         if (unlikely( blkno > log->l_logBBsize || blkno > INT_MAX )) {
3493                 XFS_ERROR_REPORT("xlog_valid_rec_header(3)",
3494                                 XFS_ERRLEVEL_LOW, log->l_mp);
3495                 return XFS_ERROR(EFSCORRUPTED);
3496         }
3497         return 0;
3498 }
3499
3500 /*
3501  * Read the log from tail to head and process the log records found.
3502  * Handle the two cases where the tail and head are in the same cycle
3503  * and where the active portion of the log wraps around the end of
3504  * the physical log separately.  The pass parameter is passed through
3505  * to the routines called to process the data and is not looked at
3506  * here.
3507  */
3508 STATIC int
3509 xlog_do_recovery_pass(
3510         xlog_t                  *log,
3511         xfs_daddr_t             head_blk,
3512         xfs_daddr_t             tail_blk,
3513         int                     pass)
3514 {
3515         xlog_rec_header_t       *rhead;
3516         xfs_daddr_t             blk_no;
3517         xfs_caddr_t             bufaddr, offset;
3518         xfs_buf_t               *hbp, *dbp;
3519         int                     error = 0, h_size;
3520         int                     bblks, split_bblks;
3521         int                     hblks, split_hblks, wrapped_hblks;
3522         xlog_recover_t          *rhash[XLOG_RHASH_SIZE];
3523
3524         ASSERT(head_blk != tail_blk);
3525
3526         /*
3527          * Read the header of the tail block and get the iclog buffer size from
3528          * h_size.  Use this to tell how many sectors make up the log header.
3529          */
3530         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
3531                 /*
3532                  * When using variable length iclogs, read first sector of
3533                  * iclog header and extract the header size from it.  Get a
3534                  * new hbp that is the correct size.
3535                  */
3536                 hbp = xlog_get_bp(log, 1);
3537                 if (!hbp)
3538                         return ENOMEM;
3539                 if ((error = xlog_bread(log, tail_blk, 1, hbp)))
3540                         goto bread_err1;
3541                 offset = xlog_align(log, tail_blk, 1, hbp);
3542                 rhead = (xlog_rec_header_t *)offset;
3543                 error = xlog_valid_rec_header(log, rhead, tail_blk);
3544                 if (error)
3545                         goto bread_err1;
3546                 h_size = be32_to_cpu(rhead->h_size);
3547                 if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
3548                     (h_size > XLOG_HEADER_CYCLE_SIZE)) {
3549                         hblks = h_size / XLOG_HEADER_CYCLE_SIZE;
3550                         if (h_size % XLOG_HEADER_CYCLE_SIZE)
3551                                 hblks++;
3552                         xlog_put_bp(hbp);
3553                         hbp = xlog_get_bp(log, hblks);
3554                 } else {
3555                         hblks = 1;
3556                 }
3557         } else {
3558                 ASSERT(log->l_sectbb_log == 0);
3559                 hblks = 1;
3560                 hbp = xlog_get_bp(log, 1);
3561                 h_size = XLOG_BIG_RECORD_BSIZE;
3562         }
3563
3564         if (!hbp)
3565                 return ENOMEM;
3566         dbp = xlog_get_bp(log, BTOBB(h_size));
3567         if (!dbp) {
3568                 xlog_put_bp(hbp);
3569                 return ENOMEM;
3570         }
3571
3572         memset(rhash, 0, sizeof(rhash));
3573         if (tail_blk <= head_blk) {
3574                 for (blk_no = tail_blk; blk_no < head_blk; ) {
3575                         if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3576                                 goto bread_err2;
3577                         offset = xlog_align(log, blk_no, hblks, hbp);
3578                         rhead = (xlog_rec_header_t *)offset;
3579                         error = xlog_valid_rec_header(log, rhead, blk_no);
3580                         if (error)
3581                                 goto bread_err2;
3582
3583                         /* blocks in data section */
3584                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3585                         error = xlog_bread(log, blk_no + hblks, bblks, dbp);
3586                         if (error)
3587                                 goto bread_err2;
3588                         offset = xlog_align(log, blk_no + hblks, bblks, dbp);
3589                         xlog_unpack_data(rhead, offset, log);
3590                         if ((error = xlog_recover_process_data(log,
3591                                                 rhash, rhead, offset, pass)))
3592                                 goto bread_err2;
3593                         blk_no += bblks + hblks;
3594                 }
3595         } else {
3596                 /*
3597                  * Perform recovery around the end of the physical log.
3598                  * When the head is not on the same cycle number as the tail,
3599                  * we can't do a sequential recovery as above.
3600                  */
3601                 blk_no = tail_blk;
3602                 while (blk_no < log->l_logBBsize) {
3603                         /*
3604                          * Check for header wrapping around physical end-of-log
3605                          */
3606                         offset = NULL;
3607                         split_hblks = 0;
3608                         wrapped_hblks = 0;
3609                         if (blk_no + hblks <= log->l_logBBsize) {
3610                                 /* Read header in one read */
3611                                 error = xlog_bread(log, blk_no, hblks, hbp);
3612                                 if (error)
3613                                         goto bread_err2;
3614                                 offset = xlog_align(log, blk_no, hblks, hbp);
3615                         } else {
3616                                 /* This LR is split across physical log end */
3617                                 if (blk_no != log->l_logBBsize) {
3618                                         /* some data before physical log end */
3619                                         ASSERT(blk_no <= INT_MAX);
3620                                         split_hblks = log->l_logBBsize - (int)blk_no;
3621                                         ASSERT(split_hblks > 0);
3622                                         if ((error = xlog_bread(log, blk_no,
3623                                                         split_hblks, hbp)))
3624                                                 goto bread_err2;
3625                                         offset = xlog_align(log, blk_no,
3626                                                         split_hblks, hbp);
3627                                 }
3628                                 /*
3629                                  * Note: this black magic still works with
3630                                  * large sector sizes (non-512) only because:
3631                                  * - we increased the buffer size originally
3632                                  *   by 1 sector giving us enough extra space
3633                                  *   for the second read;
3634                                  * - the log start is guaranteed to be sector
3635                                  *   aligned;
3636                                  * - we read the log end (LR header start)
3637                                  *   _first_, then the log start (LR header end)
3638                                  *   - order is important.
3639                                  */
3640                                 wrapped_hblks = hblks - split_hblks;
3641                                 bufaddr = XFS_BUF_PTR(hbp);
3642                                 error = XFS_BUF_SET_PTR(hbp,
3643                                                 bufaddr + BBTOB(split_hblks),
3644                                                 BBTOB(hblks - split_hblks));
3645                                 if (!error)
3646                                         error = xlog_bread(log, 0,
3647                                                         wrapped_hblks, hbp);
3648                                 if (!error)
3649                                         error = XFS_BUF_SET_PTR(hbp, bufaddr,
3650                                                         BBTOB(hblks));
3651                                 if (error)
3652                                         goto bread_err2;
3653                                 if (!offset)
3654                                         offset = xlog_align(log, 0,
3655                                                         wrapped_hblks, hbp);
3656                         }
3657                         rhead = (xlog_rec_header_t *)offset;
3658                         error = xlog_valid_rec_header(log, rhead,
3659                                                 split_hblks ? blk_no : 0);
3660                         if (error)
3661                                 goto bread_err2;
3662
3663                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3664                         blk_no += hblks;
3665
3666                         /* Read in data for log record */
3667                         if (blk_no + bblks <= log->l_logBBsize) {
3668                                 error = xlog_bread(log, blk_no, bblks, dbp);
3669                                 if (error)
3670                                         goto bread_err2;
3671                                 offset = xlog_align(log, blk_no, bblks, dbp);
3672                         } else {
3673                                 /* This log record is split across the
3674                                  * physical end of log */
3675                                 offset = NULL;
3676                                 split_bblks = 0;
3677                                 if (blk_no != log->l_logBBsize) {
3678                                         /* some data is before the physical
3679                                          * end of log */
3680                                         ASSERT(!wrapped_hblks);
3681                                         ASSERT(blk_no <= INT_MAX);
3682                                         split_bblks =
3683                                                 log->l_logBBsize - (int)blk_no;
3684                                         ASSERT(split_bblks > 0);
3685                                         if ((error = xlog_bread(log, blk_no,
3686                                                         split_bblks, dbp)))
3687                                                 goto bread_err2;
3688                                         offset = xlog_align(log, blk_no,
3689                                                         split_bblks, dbp);
3690                                 }
3691                                 /*
3692                                  * Note: this black magic still works with
3693                                  * large sector sizes (non-512) only because:
3694                                  * - we increased the buffer size originally
3695                                  *   by 1 sector giving us enough extra space
3696                                  *   for the second read;
3697                                  * - the log start is guaranteed to be sector
3698                                  *   aligned;
3699                                  * - we read the log end (LR header start)
3700                                  *   _first_, then the log start (LR header end)
3701                                  *   - order is important.
3702                                  */
3703                                 bufaddr = XFS_BUF_PTR(dbp);
3704                                 error = XFS_BUF_SET_PTR(dbp,
3705                                                 bufaddr + BBTOB(split_bblks),
3706                                                 BBTOB(bblks - split_bblks));
3707                                 if (!error)
3708                                         error = xlog_bread(log, wrapped_hblks,
3709                                                         bblks - split_bblks,
3710                                                         dbp);
3711                                 if (!error)
3712                                         error = XFS_BUF_SET_PTR(dbp, bufaddr,
3713                                                         h_size);
3714                                 if (error)
3715                                         goto bread_err2;
3716                                 if (!offset)
3717                                         offset = xlog_align(log, wrapped_hblks,
3718                                                 bblks - split_bblks, dbp);
3719                         }
3720                         xlog_unpack_data(rhead, offset, log);
3721                         if ((error = xlog_recover_process_data(log, rhash,
3722                                                         rhead, offset, pass)))
3723                                 goto bread_err2;
3724                         blk_no += bblks;
3725                 }
3726
3727                 ASSERT(blk_no >= log->l_logBBsize);
3728                 blk_no -= log->l_logBBsize;
3729
3730                 /* read first part of physical log */
3731                 while (blk_no < head_blk) {
3732                         if ((error = xlog_bread(log, blk_no, hblks, hbp)))
3733                                 goto bread_err2;
3734                         offset = xlog_align(log, blk_no, hblks, hbp);
3735                         rhead = (xlog_rec_header_t *)offset;
3736                         error = xlog_valid_rec_header(log, rhead, blk_no);
3737                         if (error)
3738                                 goto bread_err2;
3739                         bblks = (int)BTOBB(be32_to_cpu(rhead->h_len));
3740                         if ((error = xlog_bread(log, blk_no+hblks, bblks, dbp)))
3741                                 goto bread_err2;
3742                         offset = xlog_align(log, blk_no+hblks, bblks, dbp);
3743                         xlog_unpack_data(rhead, offset, log);
3744                         if ((error = xlog_recover_process_data(log, rhash,
3745                                                         rhead, offset, pass)))
3746                                 goto bread_err2;
3747                         blk_no += bblks + hblks;
3748                 }
3749         }
3750
3751  bread_err2:
3752         xlog_put_bp(dbp);
3753  bread_err1:
3754         xlog_put_bp(hbp);
3755         return error;
3756 }
3757
3758 /*
3759  * Do the recovery of the log.  We actually do this in two phases.
3760  * The two passes are necessary in order to implement the function
3761  * of cancelling a record written into the log.  The first pass
3762  * determines those things which have been cancelled, and the
3763  * second pass replays log items normally except for those which
3764  * have been cancelled.  The handling of the replay and cancellations
3765  * takes place in the log item type specific routines.
3766  *
3767  * The table of items which have cancel records in the log is allocated
3768  * and freed at this level, since only here do we know when all of
3769  * the log recovery has been completed.
3770  */
3771 STATIC int
3772 xlog_do_log_recovery(
3773         xlog_t          *log,
3774         xfs_daddr_t     head_blk,
3775         xfs_daddr_t     tail_blk)
3776 {
3777         int             error;
3778
3779         ASSERT(head_blk != tail_blk);
3780
3781         /*
3782          * First do a pass to find all of the cancelled buf log items.
3783          * Store them in the buf_cancel_table for use in the second pass.
3784          */
3785         log->l_buf_cancel_table =
3786                 (xfs_buf_cancel_t **)kmem_zalloc(XLOG_BC_TABLE_SIZE *
3787                                                  sizeof(xfs_buf_cancel_t*),
3788                                                  KM_SLEEP);
3789         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3790                                       XLOG_RECOVER_PASS1);
3791         if (error != 0) {
3792                 kmem_free(log->l_buf_cancel_table);
3793                 log->l_buf_cancel_table = NULL;
3794                 return error;
3795         }
3796         /*
3797          * Then do a second pass to actually recover the items in the log.
3798          * When it is complete free the table of buf cancel items.
3799          */
3800         error = xlog_do_recovery_pass(log, head_blk, tail_blk,
3801                                       XLOG_RECOVER_PASS2);
3802 #ifdef DEBUG
3803         if (!error) {
3804                 int     i;
3805
3806                 for (i = 0; i < XLOG_BC_TABLE_SIZE; i++)
3807                         ASSERT(log->l_buf_cancel_table[i] == NULL);
3808         }
3809 #endif  /* DEBUG */
3810
3811         kmem_free(log->l_buf_cancel_table);
3812         log->l_buf_cancel_table = NULL;
3813
3814         return error;
3815 }
3816
3817 /*
3818  * Do the actual recovery
3819  */
3820 STATIC int
3821 xlog_do_recover(
3822         xlog_t          *log,
3823         xfs_daddr_t     head_blk,
3824         xfs_daddr_t     tail_blk)
3825 {
3826         int             error;
3827         xfs_buf_t       *bp;
3828         xfs_sb_t        *sbp;
3829
3830         /*
3831          * First replay the images in the log.
3832          */
3833         error = xlog_do_log_recovery(log, head_blk, tail_blk);
3834         if (error) {
3835                 return error;
3836         }
3837
3838         XFS_bflush(log->l_mp->m_ddev_targp);
3839
3840         /*
3841          * If IO errors happened during recovery, bail out.
3842          */
3843         if (XFS_FORCED_SHUTDOWN(log->l_mp)) {
3844                 return (EIO);
3845         }
3846
3847         /*
3848          * We now update the tail_lsn since much of the recovery has completed
3849          * and there may be space available to use.  If there were no extent
3850          * or iunlinks, we can free up the entire log and set the tail_lsn to
3851          * be the last_sync_lsn.  This was set in xlog_find_tail to be the
3852          * lsn of the last known good LR on disk.  If there are extent frees
3853          * or iunlinks they will have some entries in the AIL; so we look at
3854          * the AIL to determine how to set the tail_lsn.
3855          */
3856         xlog_assign_tail_lsn(log->l_mp);
3857
3858         /*
3859          * Now that we've finished replaying all buffer and inode
3860          * updates, re-read in the superblock.
3861          */
3862         bp = xfs_getsb(log->l_mp, 0);
3863         XFS_BUF_UNDONE(bp);
3864         ASSERT(!(XFS_BUF_ISWRITE(bp)));
3865         ASSERT(!(XFS_BUF_ISDELAYWRITE(bp)));
3866         XFS_BUF_READ(bp);
3867         XFS_BUF_UNASYNC(bp);
3868         xfsbdstrat(log->l_mp, bp);
3869         error = xfs_iowait(bp);
3870         if (error) {
3871                 xfs_ioerror_alert("xlog_do_recover",
3872                                   log->l_mp, bp, XFS_BUF_ADDR(bp));
3873                 ASSERT(0);
3874                 xfs_buf_relse(bp);
3875                 return error;
3876         }
3877
3878         /* Convert superblock from on-disk format */
3879         sbp = &log->l_mp->m_sb;
3880         xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(bp));
3881         ASSERT(sbp->sb_magicnum == XFS_SB_MAGIC);
3882         ASSERT(xfs_sb_good_version(sbp));
3883         xfs_buf_relse(bp);
3884
3885         /* We've re-read the superblock so re-initialize per-cpu counters */
3886         xfs_icsb_reinit_counters(log->l_mp);
3887
3888         xlog_recover_check_summary(log);
3889
3890         /* Normal transactions can now occur */
3891         log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
3892         return 0;
3893 }
3894
3895 /*
3896  * Perform recovery and re-initialize some log variables in xlog_find_tail.
3897  *
3898  * Return error or zero.
3899  */
3900 int
3901 xlog_recover(
3902         xlog_t          *log)
3903 {
3904         xfs_daddr_t     head_blk, tail_blk;
3905         int             error;
3906
3907         /* find the tail of the log */
3908         if ((error = xlog_find_tail(log, &head_blk, &tail_blk)))
3909                 return error;
3910
3911         if (tail_blk != head_blk) {
3912                 /* There used to be a comment here:
3913                  *
3914                  * disallow recovery on read-only mounts.  note -- mount
3915                  * checks for ENOSPC and turns it into an intelligent
3916                  * error message.
3917                  * ...but this is no longer true.  Now, unless you specify
3918                  * NORECOVERY (in which case this function would never be
3919                  * called), we just go ahead and recover.  We do this all
3920                  * under the vfs layer, so we can get away with it unless
3921                  * the device itself is read-only, in which case we fail.
3922                  */
3923                 if ((error = xfs_dev_is_read_only(log->l_mp, "recovery"))) {
3924                         return error;
3925                 }
3926
3927                 cmn_err(CE_NOTE,
3928                         "Starting XFS recovery on filesystem: %s (logdev: %s)",
3929                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3930                         log->l_mp->m_logname : "internal");
3931
3932                 error = xlog_do_recover(log, head_blk, tail_blk);
3933                 log->l_flags |= XLOG_RECOVERY_NEEDED;
3934         }
3935         return error;
3936 }
3937
3938 /*
3939  * In the first part of recovery we replay inodes and buffers and build
3940  * up the list of extent free items which need to be processed.  Here
3941  * we process the extent free items and clean up the on disk unlinked
3942  * inode lists.  This is separated from the first part of recovery so
3943  * that the root and real-time bitmap inodes can be read in from disk in
3944  * between the two stages.  This is necessary so that we can free space
3945  * in the real-time portion of the file system.
3946  */
3947 int
3948 xlog_recover_finish(
3949         xlog_t          *log)
3950 {
3951         /*
3952          * Now we're ready to do the transactions needed for the
3953          * rest of recovery.  Start with completing all the extent
3954          * free intent records and then process the unlinked inode
3955          * lists.  At this point, we essentially run in normal mode
3956          * except that we're still performing recovery actions
3957          * rather than accepting new requests.
3958          */
3959         if (log->l_flags & XLOG_RECOVERY_NEEDED) {
3960                 int     error;
3961                 error = xlog_recover_process_efis(log);
3962                 if (error) {
3963                         cmn_err(CE_ALERT,
3964                                 "Failed to recover EFIs on filesystem: %s",
3965                                 log->l_mp->m_fsname);
3966                         return error;
3967                 }
3968                 /*
3969                  * Sync the log to get all the EFIs out of the AIL.
3970                  * This isn't absolutely necessary, but it helps in
3971                  * case the unlink transactions would have problems
3972                  * pushing the EFIs out of the way.
3973                  */
3974                 xfs_log_force(log->l_mp, (xfs_lsn_t)0,
3975                               (XFS_LOG_FORCE | XFS_LOG_SYNC));
3976
3977                 xlog_recover_process_iunlinks(log);
3978
3979                 xlog_recover_check_summary(log);
3980
3981                 cmn_err(CE_NOTE,
3982                         "Ending XFS recovery on filesystem: %s (logdev: %s)",
3983                         log->l_mp->m_fsname, log->l_mp->m_logname ?
3984                         log->l_mp->m_logname : "internal");
3985                 log->l_flags &= ~XLOG_RECOVERY_NEEDED;
3986         } else {
3987                 cmn_err(CE_DEBUG,
3988                         "!Ending clean XFS mount for filesystem: %s\n",
3989                         log->l_mp->m_fsname);
3990         }
3991         return 0;
3992 }
3993
3994
3995 #if defined(DEBUG)
3996 /*
3997  * Read all of the agf and agi counters and check that they
3998  * are consistent with the superblock counters.
3999  */
4000 void
4001 xlog_recover_check_summary(
4002         xlog_t          *log)
4003 {
4004         xfs_mount_t     *mp;
4005         xfs_agf_t       *agfp;
4006         xfs_agi_t       *agip;
4007         xfs_buf_t       *agfbp;
4008         xfs_buf_t       *agibp;
4009         xfs_daddr_t     agfdaddr;
4010         xfs_daddr_t     agidaddr;
4011         xfs_buf_t       *sbbp;
4012 #ifdef XFS_LOUD_RECOVERY
4013         xfs_sb_t        *sbp;
4014 #endif
4015         xfs_agnumber_t  agno;
4016         __uint64_t      freeblks;
4017         __uint64_t      itotal;
4018         __uint64_t      ifree;
4019
4020         mp = log->l_mp;
4021
4022         freeblks = 0LL;
4023         itotal = 0LL;
4024         ifree = 0LL;
4025         for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
4026                 agfdaddr = XFS_AG_DADDR(mp, agno, XFS_AGF_DADDR(mp));
4027                 agfbp = xfs_buf_read(mp->m_ddev_targp, agfdaddr,
4028                                 XFS_FSS_TO_BB(mp, 1), 0);
4029                 if (XFS_BUF_ISERROR(agfbp)) {
4030                         xfs_ioerror_alert("xlog_recover_check_summary(agf)",
4031                                                 mp, agfbp, agfdaddr);
4032                 }
4033                 agfp = XFS_BUF_TO_AGF(agfbp);
4034                 ASSERT(XFS_AGF_MAGIC == be32_to_cpu(agfp->agf_magicnum));
4035                 ASSERT(XFS_AGF_GOOD_VERSION(be32_to_cpu(agfp->agf_versionnum)));
4036                 ASSERT(be32_to_cpu(agfp->agf_seqno) == agno);
4037
4038                 freeblks += be32_to_cpu(agfp->agf_freeblks) +
4039                             be32_to_cpu(agfp->agf_flcount);
4040                 xfs_buf_relse(agfbp);
4041
4042                 agidaddr = XFS_AG_DADDR(mp, agno, XFS_AGI_DADDR(mp));
4043                 agibp = xfs_buf_read(mp->m_ddev_targp, agidaddr,
4044                                 XFS_FSS_TO_BB(mp, 1), 0);
4045                 if (XFS_BUF_ISERROR(agibp)) {
4046                         xfs_ioerror_alert("xlog_recover_check_summary(agi)",
4047                                           mp, agibp, agidaddr);
4048                 }
4049                 agip = XFS_BUF_TO_AGI(agibp);
4050                 ASSERT(XFS_AGI_MAGIC == be32_to_cpu(agip->agi_magicnum));
4051                 ASSERT(XFS_AGI_GOOD_VERSION(be32_to_cpu(agip->agi_versionnum)));
4052                 ASSERT(be32_to_cpu(agip->agi_seqno) == agno);
4053
4054                 itotal += be32_to_cpu(agip->agi_count);
4055                 ifree += be32_to_cpu(agip->agi_freecount);
4056                 xfs_buf_relse(agibp);
4057         }
4058
4059         sbbp = xfs_getsb(mp, 0);
4060 #ifdef XFS_LOUD_RECOVERY
4061         sbp = &mp->m_sb;
4062         xfs_sb_from_disk(sbp, XFS_BUF_TO_SBP(sbbp));
4063         cmn_err(CE_NOTE,
4064                 "xlog_recover_check_summary: sb_icount %Lu itotal %Lu",
4065                 sbp->sb_icount, itotal);
4066         cmn_err(CE_NOTE,
4067                 "xlog_recover_check_summary: sb_ifree %Lu itotal %Lu",
4068                 sbp->sb_ifree, ifree);
4069         cmn_err(CE_NOTE,
4070                 "xlog_recover_check_summary: sb_fdblocks %Lu freeblks %Lu",
4071                 sbp->sb_fdblocks, freeblks);
4072 #if 0
4073         /*
4074          * This is turned off until I account for the allocation
4075          * btree blocks which live in free space.
4076          */
4077         ASSERT(sbp->sb_icount == itotal);
4078         ASSERT(sbp->sb_ifree == ifree);
4079         ASSERT(sbp->sb_fdblocks == freeblks);
4080 #endif
4081 #endif
4082         xfs_buf_relse(sbbp);
4083 }
4084 #endif /* DEBUG */