fs/xfs/xfs_vnodeops.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
  19 #include "xfs.h"
  20 #include "xfs_fs.h"
  21 #include "xfs_types.h"
  22 #include "xfs_bit.h"
  23 #include "xfs_log.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_mount.h"
  29 #include "xfs_da_btree.h"
  30 #include "xfs_bmap_btree.h"
  31 #include "xfs_ialloc_btree.h"
  32 #include "xfs_dinode.h"
  33 #include "xfs_inode.h"
  34 #include "xfs_inode_item.h"
  35 #include "xfs_itable.h"
  36 #include "xfs_ialloc.h"
  37 #include "xfs_alloc.h"
  38 #include "xfs_bmap.h"
  39 #include "xfs_acl.h"
  40 #include "xfs_attr.h"
  41 #include "xfs_error.h"
  42 #include "xfs_quota.h"
  43 #include "xfs_utils.h"
  44 #include "xfs_rtalloc.h"
  45 #include "xfs_trans_space.h"
  46 #include "xfs_log_priv.h"
  47 #include "xfs_filestream.h"
  48 #include "xfs_vnodeops.h"
  49 #include "xfs_trace.h"
  50
  51 /*
  52  * The maximum pathlen is 1024 bytes. Since the minimum file system
  53  * blocksize is 512 bytes, we can get a max of 2 extents back from
  54  * bmapi.
  55  */
  56 #define SYMLINK_MAPS 2
  57
  58 STATIC int
  59 xfs_readlink_bmap(
  60         xfs_inode_t     *ip,
  61         char            *link)
  62 {
  63         xfs_mount_t     *mp = ip->i_mount;
  64         int             pathlen = ip->i_d.di_size;
  65         int             nmaps = SYMLINK_MAPS;
  66         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
  67         xfs_daddr_t     d;
  68         int             byte_cnt;
  69         int             n;
  70         xfs_buf_t       *bp;
  71         int             error = 0;
  72
  73         error = xfs_bmapi_read(ip, 0, XFS_B_TO_FSB(mp, pathlen), mval, &nmaps,
  74                                0);
  75         if (error)
  76                 goto out;
  77
  78         for (n = 0; n < nmaps; n++) {
  79                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
  80                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
  81
  82                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt), 0);
  83                 if (!bp)
  84                         return XFS_ERROR(ENOMEM);
  85                 error = bp->b_error;
  86                 if (error) {
  87                         xfs_buf_ioerror_alert(bp, __func__);
  88                         xfs_buf_relse(bp);
  89                         goto out;
  90                 }
  91                 if (pathlen < byte_cnt)
  92                         byte_cnt = pathlen;
  93                 pathlen -= byte_cnt;
  94
  95                 memcpy(link, bp->b_addr, byte_cnt);
  96                 xfs_buf_relse(bp);
  97         }
  98
  99         link[ip->i_d.di_size] = '\0';
 100         error = 0;
 101
 102  out:
 103         return error;
 104 }
 105
 106 int
 107 xfs_readlink(
 108         xfs_inode_t     *ip,
 109         char            *link)
 110 {
 111         xfs_mount_t     *mp = ip->i_mount;
 112         xfs_fsize_t     pathlen;
 113         int             error = 0;
 114
 115         trace_xfs_readlink(ip);
 116
 117         if (XFS_FORCED_SHUTDOWN(mp))
 118                 return XFS_ERROR(EIO);
 119
 120         xfs_ilock(ip, XFS_ILOCK_SHARED);
 121
 122         pathlen = ip->i_d.di_size;
 123         if (!pathlen)
 124                 goto out;
 125
 126         if (pathlen < 0 || pathlen > MAXPATHLEN) {
 127                 xfs_alert(mp, "%s: inode (%llu) bad symlink length (%lld)",
 128                          __func__, (unsigned long long) ip->i_ino,
 129                          (long long) pathlen);
 130                 ASSERT(0);
 131                 error = XFS_ERROR(EFSCORRUPTED);
 132                 goto out;
 133         }
 134
 135
 136         if (ip->i_df.if_flags & XFS_IFINLINE) {
 137                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
 138                 link[pathlen] = '\0';
 139         } else {
 140                 error = xfs_readlink_bmap(ip, link);
 141         }
 142
 143  out:
 144         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 145         return error;
 146 }
 147
 148 /*
 149  * Flags for xfs_free_eofblocks
 150  */
 151 #define XFS_FREE_EOF_TRYLOCK    (1<<0)
 152
 153 /*
 154  * This is called by xfs_inactive to free any blocks beyond eof
 155  * when the link count isn't zero and by xfs_dm_punch_hole() when
 156  * punching a hole to EOF.
 157  */
 158 STATIC int
 159 xfs_free_eofblocks(
 160         xfs_mount_t     *mp,
 161         xfs_inode_t     *ip,
 162         int             flags)
 163 {
 164         xfs_trans_t     *tp;
 165         int             error;
 166         xfs_fileoff_t   end_fsb;
 167         xfs_fileoff_t   last_fsb;
 168         xfs_filblks_t   map_len;
 169         int             nimaps;
 170         xfs_bmbt_irec_t imap;
 171
 172         /*
 173          * Figure out if there are any blocks beyond the end
 174          * of the file.  If not, then there is nothing to do.
 175          */
 176         end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_ISIZE(ip));
 177         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
 178         if (last_fsb <= end_fsb)
 179                 return 0;
 180         map_len = last_fsb - end_fsb;
 181
 182         nimaps = 1;
 183         xfs_ilock(ip, XFS_ILOCK_SHARED);
 184         error = xfs_bmapi_read(ip, end_fsb, map_len, &imap, &nimaps, 0);
 185         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 186
 187         if (!error && (nimaps != 0) &&
 188             (imap.br_startblock != HOLESTARTBLOCK ||
 189              ip->i_delayed_blks)) {
 190                 /*
 191                  * Attach the dquots to the inode up front.
 192                  */
 193                 error = xfs_qm_dqattach(ip, 0);
 194                 if (error)
 195                         return error;
 196
 197                 /*
 198                  * There are blocks after the end of file.
 199                  * Free them up now by truncating the file to
 200                  * its current size.
 201                  */
 202                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 203
 204                 if (flags & XFS_FREE_EOF_TRYLOCK) {
 205                         if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
 206                                 xfs_trans_cancel(tp, 0);
 207                                 return 0;
 208                         }
 209                 } else {
 210                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 211                 }
 212
 213                 error = xfs_trans_reserve(tp, 0,
 214                                           XFS_ITRUNCATE_LOG_RES(mp),
 215                                           0, XFS_TRANS_PERM_LOG_RES,
 216                                           XFS_ITRUNCATE_LOG_COUNT);
 217                 if (error) {
 218                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 219                         xfs_trans_cancel(tp, 0);
 220                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 221                         return error;
 222                 }
 223
 224                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 225                 xfs_trans_ijoin(tp, ip, 0);
 226
 227                 /*
 228                  * Do not update the on-disk file size.  If we update the
 229                  * on-disk file size and then the system crashes before the
 230                  * contents of the file are flushed to disk then the files
 231                  * may be full of holes (ie NULL files bug).
 232                  */
 233                 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK,
 234                                               XFS_ISIZE(ip));
 235                 if (error) {
 236                         /*
 237                          * If we get an error at this point we simply don't
 238                          * bother truncating the file.
 239                          */
 240                         xfs_trans_cancel(tp,
 241                                          (XFS_TRANS_RELEASE_LOG_RES |
 242                                           XFS_TRANS_ABORT));
 243                 } else {
 244                         error = xfs_trans_commit(tp,
 245                                                 XFS_TRANS_RELEASE_LOG_RES);
 246                 }
 247                 xfs_iunlock(ip, XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL);
 248         }
 249         return error;
 250 }
 251
 252 /*
 253  * Free a symlink that has blocks associated with it.
 254  */
 255 STATIC int
 256 xfs_inactive_symlink_rmt(
 257         xfs_inode_t     *ip,
 258         xfs_trans_t     **tpp)
 259 {
 260         xfs_buf_t       *bp;
 261         int             committed;
 262         int             done;
 263         int             error;
 264         xfs_fsblock_t   first_block;
 265         xfs_bmap_free_t free_list;
 266         int             i;
 267         xfs_mount_t     *mp;
 268         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
 269         int             nmaps;
 270         xfs_trans_t     *ntp;
 271         int             size;
 272         xfs_trans_t     *tp;
 273
 274         tp = *tpp;
 275         mp = ip->i_mount;
 276         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
 277         /*
 278          * We're freeing a symlink that has some
 279          * blocks allocated to it.  Free the
 280          * blocks here.  We know that we've got
 281          * either 1 or 2 extents and that we can
 282          * free them all in one bunmapi call.
 283          */
 284         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
 285         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
 286                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
 287                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 288                 xfs_trans_cancel(tp, 0);
 289                 *tpp = NULL;
 290                 return error;
 291         }
 292         /*
 293          * Lock the inode, fix the size, and join it to the transaction.
 294          * Hold it so in the normal path, we still have it locked for
 295          * the second transaction.  In the error paths we need it
 296          * held so the cancel won't rele it, see below.
 297          */
 298         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
 299         size = (int)ip->i_d.di_size;
 300         ip->i_d.di_size = 0;
 301         xfs_trans_ijoin(tp, ip, 0);
 302         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 303         /*
 304          * Find the block(s) so we can inval and unmap them.
 305          */
 306         done = 0;
 307         xfs_bmap_init(&free_list, &first_block);
 308         nmaps = ARRAY_SIZE(mval);
 309         error = xfs_bmapi_read(ip, 0, XFS_B_TO_FSB(mp, size),
 310                                 mval, &nmaps, 0);
 311         if (error)
 312                 goto error0;
 313         /*
 314          * Invalidate the block(s).
 315          */
 316         for (i = 0; i < nmaps; i++) {
 317                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 318                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
 319                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
 320                 if (!bp) {
 321                         error = ENOMEM;
 322                         goto error1;
 323                 }
 324                 xfs_trans_binval(tp, bp);
 325         }
 326         /*
 327          * Unmap the dead block(s) to the free_list.
 328          */
 329         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
 330                         &first_block, &free_list, &done)))
 331                 goto error1;
 332         ASSERT(done);
 333         /*
 334          * Commit the first transaction.  This logs the EFI and the inode.
 335          */
 336         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
 337                 goto error1;
 338         /*
 339          * The transaction must have been committed, since there were
 340          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
 341          * The new tp has the extent freeing and EFDs.
 342          */
 343         ASSERT(committed);
 344         /*
 345          * The first xact was committed, so add the inode to the new one.
 346          * Mark it dirty so it will be logged and moved forward in the log as
 347          * part of every commit.
 348          */
 349         xfs_trans_ijoin(tp, ip, 0);
 350         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 351         /*
 352          * Get a new, empty transaction to return to our caller.
 353          */
 354         ntp = xfs_trans_dup(tp);
 355         /*
 356          * Commit the transaction containing extent freeing and EFDs.
 357          * If we get an error on the commit here or on the reserve below,
 358          * we need to unlock the inode since the new transaction doesn't
 359          * have the inode attached.
 360          */
 361         error = xfs_trans_commit(tp, 0);
 362         tp = ntp;
 363         if (error) {
 364                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 365                 goto error0;
 366         }
 367         /*
 368          * transaction commit worked ok so we can drop the extra ticket
 369          * reference that we gained in xfs_trans_dup()
 370          */
 371         xfs_log_ticket_put(tp->t_ticket);
 372
 373         /*
 374          * Remove the memory for extent descriptions (just bookkeeping).
 375          */
 376         if (ip->i_df.if_bytes)
 377                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
 378         ASSERT(ip->i_df.if_bytes == 0);
 379         /*
 380          * Put an itruncate log reservation in the new transaction
 381          * for our caller.
 382          */
 383         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
 384                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
 385                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 386                 goto error0;
 387         }
 388         /*
 389          * Return with the inode locked but not joined to the transaction.
 390          */
 391         *tpp = tp;
 392         return 0;
 393
 394  error1:
 395         xfs_bmap_cancel(&free_list);
 396  error0:
 397         /*
 398          * Have to come here with the inode locked and either
 399          * (held and in the transaction) or (not in the transaction).
 400          * If the inode isn't held then cancel would iput it, but
 401          * that's wrong since this is inactive and the vnode ref
 402          * count is 0 already.
 403          * Cancel won't do anything to the inode if held, but it still
 404          * needs to be locked until the cancel is done, if it was
 405          * joined to the transaction.
 406          */
 407         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 408         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
 409         *tpp = NULL;
 410         return error;
 411
 412 }
 413
 414 STATIC int
 415 xfs_inactive_symlink_local(
 416         xfs_inode_t     *ip,
 417         xfs_trans_t     **tpp)
 418 {
 419         int             error;
 420
 421         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
 422         /*
 423          * We're freeing a symlink which fit into
 424          * the inode.  Just free the memory used
 425          * to hold the old symlink.
 426          */
 427         error = xfs_trans_reserve(*tpp, 0,
 428                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
 429                                   0, XFS_TRANS_PERM_LOG_RES,
 430                                   XFS_ITRUNCATE_LOG_COUNT);
 431
 432         if (error) {
 433                 xfs_trans_cancel(*tpp, 0);
 434                 *tpp = NULL;
 435                 return error;
 436         }
 437         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
 438
 439         /*
 440          * Zero length symlinks _can_ exist.
 441          */
 442         if (ip->i_df.if_bytes > 0) {
 443                 xfs_idata_realloc(ip,
 444                                   -(ip->i_df.if_bytes),
 445                                   XFS_DATA_FORK);
 446                 ASSERT(ip->i_df.if_bytes == 0);
 447         }
 448         return 0;
 449 }
 450
 451 STATIC int
 452 xfs_inactive_attrs(
 453         xfs_inode_t     *ip,
 454         xfs_trans_t     **tpp)
 455 {
 456         xfs_trans_t     *tp;
 457         int             error;
 458         xfs_mount_t     *mp;
 459
 460         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
 461         tp = *tpp;
 462         mp = ip->i_mount;
 463         ASSERT(ip->i_d.di_forkoff != 0);
 464         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 465         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 466         if (error)
 467                 goto error_unlock;
 468
 469         error = xfs_attr_inactive(ip);
 470         if (error)
 471                 goto error_unlock;
 472
 473         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 474         error = xfs_trans_reserve(tp, 0,
 475                                   XFS_IFREE_LOG_RES(mp),
 476                                   0, XFS_TRANS_PERM_LOG_RES,
 477                                   XFS_INACTIVE_LOG_COUNT);
 478         if (error)
 479                 goto error_cancel;
 480
 481         xfs_ilock(ip, XFS_ILOCK_EXCL);
 482         xfs_trans_ijoin(tp, ip, 0);
 483         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
 484
 485         ASSERT(ip->i_d.di_anextents == 0);
 486
 487         *tpp = tp;
 488         return 0;
 489
 490 error_cancel:
 491         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 492         xfs_trans_cancel(tp, 0);
 493 error_unlock:
 494         *tpp = NULL;
 495         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 496         return error;
 497 }
 498
 499 int
 500 xfs_release(
 501         xfs_inode_t     *ip)
 502 {
 503         xfs_mount_t     *mp = ip->i_mount;
 504         int             error;
 505
 506         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
 507                 return 0;
 508
 509         /* If this is a read-only mount, don't do this (would generate I/O) */
 510         if (mp->m_flags & XFS_MOUNT_RDONLY)
 511                 return 0;
 512
 513         if (!XFS_FORCED_SHUTDOWN(mp)) {
 514                 int truncated;
 515
 516                 /*
 517                  * If we are using filestreams, and we have an unlinked
 518                  * file that we are processing the last close on, then nothing
 519                  * will be able to reopen and write to this file. Purge this
 520                  * inode from the filestreams cache so that it doesn't delay
 521                  * teardown of the inode.
 522                  */
 523                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
 524                         xfs_filestream_deassociate(ip);
 525
 526                 /*
 527                  * If we previously truncated this file and removed old data
 528                  * in the process, we want to initiate "early" writeout on
 529                  * the last close.  This is an attempt to combat the notorious
 530                  * NULL files problem which is particularly noticeable from a
 531                  * truncate down, buffered (re-)write (delalloc), followed by
 532                  * a crash.  What we are effectively doing here is
 533                  * significantly reducing the time window where we'd otherwise
 534                  * be exposed to that problem.
 535                  */
 536                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
 537                 if (truncated) {
 538                         xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
 539                         if (VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
 540                                 xfs_flush_pages(ip, 0, -1, XBF_ASYNC, FI_NONE);
 541                 }
 542         }
 543
 544         if (ip->i_d.di_nlink == 0)
 545                 return 0;
 546
 547         if ((S_ISREG(ip->i_d.di_mode) &&
 548              (VFS_I(ip)->i_size > 0 ||
 549               (VN_CACHED(VFS_I(ip)) > 0 || ip->i_delayed_blks > 0)) &&
 550              (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
 551             (!(ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
 552
 553                 /*
 554                  * If we can't get the iolock just skip truncating the blocks
 555                  * past EOF because we could deadlock with the mmap_sem
 556                  * otherwise.  We'll get another chance to drop them once the
 557                  * last reference to the inode is dropped, so we'll never leak
 558                  * blocks permanently.
 559                  *
 560                  * Further, check if the inode is being opened, written and
 561                  * closed frequently and we have delayed allocation blocks
 562                  * outstanding (e.g. streaming writes from the NFS server),
 563                  * truncating the blocks past EOF will cause fragmentation to
 564                  * occur.
 565                  *
 566                  * In this case don't do the truncation, either, but we have to
 567                  * be careful how we detect this case. Blocks beyond EOF show
 568                  * up as i_delayed_blks even when the inode is clean, so we
 569                  * need to truncate them away first before checking for a dirty
 570                  * release. Hence on the first dirty close we will still remove
 571                  * the speculative allocation, but after that we will leave it
 572                  * in place.
 573                  */
 574                 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
 575                         return 0;
 576
 577                 error = xfs_free_eofblocks(mp, ip,
 578                                            XFS_FREE_EOF_TRYLOCK);
 579                 if (error)
 580                         return error;
 581
 582                 /* delalloc blocks after truncation means it really is dirty */
 583                 if (ip->i_delayed_blks)
 584                         xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
 585         }
 586         return 0;
 587 }
 588
 589 /*
 590  * xfs_inactive
 591  *
 592  * This is called when the vnode reference count for the vnode
 593  * goes to zero.  If the file has been unlinked, then it must
 594  * now be truncated.  Also, we clear all of the read-ahead state
 595  * kept for the inode here since the file is now closed.
 596  */
 597 int
 598 xfs_inactive(
 599         xfs_inode_t     *ip)
 600 {
 601         xfs_bmap_free_t free_list;
 602         xfs_fsblock_t   first_block;
 603         int             committed;
 604         xfs_trans_t     *tp;
 605         xfs_mount_t     *mp;
 606         int             error;
 607         int             truncate;
 608
 609         /*
 610          * If the inode is already free, then there can be nothing
 611          * to clean up here.
 612          */
 613         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
 614                 ASSERT(ip->i_df.if_real_bytes == 0);
 615                 ASSERT(ip->i_df.if_broot_bytes == 0);
 616                 return VN_INACTIVE_CACHE;
 617         }
 618
 619         /*
 620          * Only do a truncate if it's a regular file with
 621          * some actual space in it.  It's OK to look at the
 622          * inode's fields without the lock because we're the
 623          * only one with a reference to the inode.
 624          */
 625         truncate = ((ip->i_d.di_nlink == 0) &&
 626             ((ip->i_d.di_size != 0) || XFS_ISIZE(ip) != 0 ||
 627              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
 628             S_ISREG(ip->i_d.di_mode));
 629
 630         mp = ip->i_mount;
 631
 632         error = 0;
 633
 634         /* If this is a read-only mount, don't do this (would generate I/O) */
 635         if (mp->m_flags & XFS_MOUNT_RDONLY)
 636                 goto out;
 637
 638         if (ip->i_d.di_nlink != 0) {
 639                 if ((S_ISREG(ip->i_d.di_mode) &&
 640                     (VFS_I(ip)->i_size > 0 ||
 641                      (VN_CACHED(VFS_I(ip)) > 0 || ip->i_delayed_blks > 0)) &&
 642                     (ip->i_df.if_flags & XFS_IFEXTENTS) &&
 643                     (!(ip->i_d.di_flags &
 644                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
 645                      ip->i_delayed_blks != 0))) {
 646                         error = xfs_free_eofblocks(mp, ip, 0);
 647                         if (error)
 648                                 return VN_INACTIVE_CACHE;
 649                 }
 650                 goto out;
 651         }
 652
 653         ASSERT(ip->i_d.di_nlink == 0);
 654
 655         error = xfs_qm_dqattach(ip, 0);
 656         if (error)
 657                 return VN_INACTIVE_CACHE;
 658
 659         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 660         if (truncate) {
 661                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
 662
 663                 error = xfs_trans_reserve(tp, 0,
 664                                           XFS_ITRUNCATE_LOG_RES(mp),
 665                                           0, XFS_TRANS_PERM_LOG_RES,
 666                                           XFS_ITRUNCATE_LOG_COUNT);
 667                 if (error) {
 668                         /* Don't call itruncate_cleanup */
 669                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 670                         xfs_trans_cancel(tp, 0);
 671                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 672                         return VN_INACTIVE_CACHE;
 673                 }
 674
 675                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 676                 xfs_trans_ijoin(tp, ip, 0);
 677
 678                 ip->i_d.di_size = 0;
 679                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 680
 681                 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
 682                 if (error) {
 683                         xfs_trans_cancel(tp,
 684                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 685                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
 686                         return VN_INACTIVE_CACHE;
 687                 }
 688
 689                 ASSERT(ip->i_d.di_nextents == 0);
 690         } else if (S_ISLNK(ip->i_d.di_mode)) {
 691
 692                 /*
 693                  * If we get an error while cleaning up a
 694                  * symlink we bail out.
 695                  */
 696                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
 697                         xfs_inactive_symlink_rmt(ip, &tp) :
 698                         xfs_inactive_symlink_local(ip, &tp);
 699
 700                 if (error) {
 701                         ASSERT(tp == NULL);
 702                         return VN_INACTIVE_CACHE;
 703                 }
 704
 705                 xfs_trans_ijoin(tp, ip, 0);
 706         } else {
 707                 error = xfs_trans_reserve(tp, 0,
 708                                           XFS_IFREE_LOG_RES(mp),
 709                                           0, XFS_TRANS_PERM_LOG_RES,
 710                                           XFS_INACTIVE_LOG_COUNT);
 711                 if (error) {
 712                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 713                         xfs_trans_cancel(tp, 0);
 714                         return VN_INACTIVE_CACHE;
 715                 }
 716
 717                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
 718                 xfs_trans_ijoin(tp, ip, 0);
 719         }
 720
 721         /*
 722          * If there are attributes associated with the file
 723          * then blow them away now.  The code calls a routine
 724          * that recursively deconstructs the attribute fork.
 725          * We need to just commit the current transaction
 726          * because we can't use it for xfs_attr_inactive().
 727          */
 728         if (ip->i_d.di_anextents > 0) {
 729                 error = xfs_inactive_attrs(ip, &tp);
 730                 /*
 731                  * If we got an error, the transaction is already
 732                  * cancelled, and the inode is unlocked. Just get out.
 733                  */
 734                  if (error)
 735                          return VN_INACTIVE_CACHE;
 736         } else if (ip->i_afp) {
 737                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
 738         }
 739
 740         /*
 741          * Free the inode.
 742          */
 743         xfs_bmap_init(&free_list, &first_block);
 744         error = xfs_ifree(tp, ip, &free_list);
 745         if (error) {
 746                 /*
 747                  * If we fail to free the inode, shut down.  The cancel
 748                  * might do that, we need to make sure.  Otherwise the
 749                  * inode might be lost for a long time or forever.
 750                  */
 751                 if (!XFS_FORCED_SHUTDOWN(mp)) {
 752                         xfs_notice(mp, "%s: xfs_ifree returned error %d",
 753                                 __func__, error);
 754                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
 755                 }
 756                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
 757         } else {
 758                 /*
 759                  * Credit the quota account(s). The inode is gone.
 760                  */
 761                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
 762
 763                 /*
 764                  * Just ignore errors at this point.  There is nothing we can
 765                  * do except to try to keep going. Make sure it's not a silent
 766                  * error.
 767                  */
 768                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
 769                 if (error)
 770                         xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
 771                                 __func__, error);
 772                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 773                 if (error)
 774                         xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
 775                                 __func__, error);
 776         }
 777
 778         /*
 779          * Release the dquots held by inode, if any.
 780          */
 781         xfs_qm_dqdetach(ip);
 782         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
 783
 784  out:
 785         return VN_INACTIVE_CACHE;
 786 }
 787
 788 /*
 789  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
 790  * is allowed, otherwise it has to be an exact match. If a CI match is found,
 791  * ci_name->name will point to a the actual name (caller must free) or
 792  * will be set to NULL if an exact match is found.
 793  */
 794 int
 795 xfs_lookup(
 796         xfs_inode_t             *dp,
 797         struct xfs_name         *name,
 798         xfs_inode_t             **ipp,
 799         struct xfs_name         *ci_name)
 800 {
 801         xfs_ino_t               inum;
 802         int                     error;
 803         uint                    lock_mode;
 804
 805         trace_xfs_lookup(dp, name);
 806
 807         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
 808                 return XFS_ERROR(EIO);
 809
 810         lock_mode = xfs_ilock_map_shared(dp);
 811         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
 812         xfs_iunlock_map_shared(dp, lock_mode);
 813
 814         if (error)
 815                 goto out;
 816
 817         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
 818         if (error)
 819                 goto out_free_name;
 820
 821         return 0;
 822
 823 out_free_name:
 824         if (ci_name)
 825                 kmem_free(ci_name->name);
 826 out:
 827         *ipp = NULL;
 828         return error;
 829 }
 830
 831 int
 832 xfs_create(
 833         xfs_inode_t             *dp,
 834         struct xfs_name         *name,
 835         umode_t                 mode,
 836         xfs_dev_t               rdev,
 837         xfs_inode_t             **ipp)
 838 {
 839         int                     is_dir = S_ISDIR(mode);
 840         struct xfs_mount        *mp = dp->i_mount;
 841         struct xfs_inode        *ip = NULL;
 842         struct xfs_trans        *tp = NULL;
 843         int                     error;
 844         xfs_bmap_free_t         free_list;
 845         xfs_fsblock_t           first_block;
 846         boolean_t               unlock_dp_on_error = B_FALSE;
 847         uint                    cancel_flags;
 848         int                     committed;
 849         prid_t                  prid;
 850         struct xfs_dquot        *udqp = NULL;
 851         struct xfs_dquot        *gdqp = NULL;
 852         uint                    resblks;
 853         uint                    log_res;
 854         uint                    log_count;
 855
 856         trace_xfs_create(dp, name);
 857
 858         if (XFS_FORCED_SHUTDOWN(mp))
 859                 return XFS_ERROR(EIO);
 860
 861         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
 862                 prid = xfs_get_projid(dp);
 863         else
 864                 prid = XFS_PROJID_DEFAULT;
 865
 866         /*
 867          * Make sure that we have allocated dquot(s) on disk.
 868          */
 869         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
 870                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
 871         if (error)
 872                 return error;
 873
 874         if (is_dir) {
 875                 rdev = 0;
 876                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
 877                 log_res = XFS_MKDIR_LOG_RES(mp);
 878                 log_count = XFS_MKDIR_LOG_COUNT;
 879                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
 880         } else {
 881                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
 882                 log_res = XFS_CREATE_LOG_RES(mp);
 883                 log_count = XFS_CREATE_LOG_COUNT;
 884                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
 885         }
 886
 887         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
 888
 889         /*
 890          * Initially assume that the file does not exist and
 891          * reserve the resources for that case.  If that is not
 892          * the case we'll drop the one we have and get a more
 893          * appropriate transaction later.
 894          */
 895         error = xfs_trans_reserve(tp, resblks, log_res, 0,
 896                         XFS_TRANS_PERM_LOG_RES, log_count);
 897         if (error == ENOSPC) {
 898                 /* flush outstanding delalloc blocks and retry */
 899                 xfs_flush_inodes(dp);
 900                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
 901                                 XFS_TRANS_PERM_LOG_RES, log_count);
 902         }
 903         if (error == ENOSPC) {
 904                 /* No space at all so try a "no-allocation" reservation */
 905                 resblks = 0;
 906                 error = xfs_trans_reserve(tp, 0, log_res, 0,
 907                                 XFS_TRANS_PERM_LOG_RES, log_count);
 908         }
 909         if (error) {
 910                 cancel_flags = 0;
 911                 goto out_trans_cancel;
 912         }
 913
 914         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
 915         unlock_dp_on_error = B_TRUE;
 916
 917         xfs_bmap_init(&free_list, &first_block);
 918
 919         /*
 920          * Reserve disk quota and the inode.
 921          */
 922         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
 923         if (error)
 924                 goto out_trans_cancel;
 925
 926         error = xfs_dir_canenter(tp, dp, name, resblks);
 927         if (error)
 928                 goto out_trans_cancel;
 929
 930         /*
 931          * A newly created regular or special file just has one directory
 932          * entry pointing to them, but a directory also the "." entry
 933          * pointing to itself.
 934          */
 935         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
 936                                prid, resblks > 0, &ip, &committed);
 937         if (error) {
 938                 if (error == ENOSPC)
 939                         goto out_trans_cancel;
 940                 goto out_trans_abort;
 941         }
 942
 943         /*
 944          * Now we join the directory inode to the transaction.  We do not do it
 945          * earlier because xfs_dir_ialloc might commit the previous transaction
 946          * (and release all the locks).  An error from here on will result in
 947          * the transaction cancel unlocking dp so don't do it explicitly in the
 948          * error path.
 949          */
 950         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
 951         unlock_dp_on_error = B_FALSE;
 952
 953         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
 954                                         &first_block, &free_list, resblks ?
 955                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
 956         if (error) {
 957                 ASSERT(error != ENOSPC);
 958                 goto out_trans_abort;
 959         }
 960         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 961         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
 962
 963         if (is_dir) {
 964                 error = xfs_dir_init(tp, ip, dp);
 965                 if (error)
 966                         goto out_bmap_cancel;
 967
 968                 error = xfs_bumplink(tp, dp);
 969                 if (error)
 970                         goto out_bmap_cancel;
 971         }
 972
 973         /*
 974          * If this is a synchronous mount, make sure that the
 975          * create transaction goes to disk before returning to
 976          * the user.
 977          */
 978         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
 979                 xfs_trans_set_sync(tp);
 980
 981         /*
 982          * Attach the dquot(s) to the inodes and modify them incore.
 983          * These ids of the inode couldn't have changed since the new
 984          * inode has been locked ever since it was created.
 985          */
 986         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
 987
 988         error = xfs_bmap_finish(&tp, &free_list, &committed);
 989         if (error)
 990                 goto out_bmap_cancel;
 991
 992         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 993         if (error)
 994                 goto out_release_inode;
 995
 996         xfs_qm_dqrele(udqp);
 997         xfs_qm_dqrele(gdqp);
 998
 999         *ipp = ip;
1000         return 0;
1001
1002  out_bmap_cancel:
1003         xfs_bmap_cancel(&free_list);
1004  out_trans_abort:
1005         cancel_flags |= XFS_TRANS_ABORT;
1006  out_trans_cancel:
1007         xfs_trans_cancel(tp, cancel_flags);
1008  out_release_inode:
1009         /*
1010          * Wait until after the current transaction is aborted to
1011          * release the inode.  This prevents recursive transactions
1012          * and deadlocks from xfs_inactive.
1013          */
1014         if (ip)
1015                 IRELE(ip);
1016
1017         xfs_qm_dqrele(udqp);
1018         xfs_qm_dqrele(gdqp);
1019
1020         if (unlock_dp_on_error)
1021                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1022         return error;
1023 }
1024
1025 #ifdef DEBUG
1026 int xfs_locked_n;
1027 int xfs_small_retries;
1028 int xfs_middle_retries;
1029 int xfs_lots_retries;
1030 int xfs_lock_delays;
1031 #endif
1032
1033 /*
1034  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1035  * a different value
1036  */
1037 static inline int
1038 xfs_lock_inumorder(int lock_mode, int subclass)
1039 {
1040         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1041                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1042         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1043                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1044
1045         return lock_mode;
1046 }
1047
1048 /*
1049  * The following routine will lock n inodes in exclusive mode.
1050  * We assume the caller calls us with the inodes in i_ino order.
1051  *
1052  * We need to detect deadlock where an inode that we lock
1053  * is in the AIL and we start waiting for another inode that is locked
1054  * by a thread in a long running transaction (such as truncate). This can
1055  * result in deadlock since the long running trans might need to wait
1056  * for the inode we just locked in order to push the tail and free space
1057  * in the log.
1058  */
1059 void
1060 xfs_lock_inodes(
1061         xfs_inode_t     **ips,
1062         int             inodes,
1063         uint            lock_mode)
1064 {
1065         int             attempts = 0, i, j, try_lock;
1066         xfs_log_item_t  *lp;
1067
1068         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1069
1070         try_lock = 0;
1071         i = 0;
1072
1073 again:
1074         for (; i < inodes; i++) {
1075                 ASSERT(ips[i]);
1076
1077                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1078                         continue;
1079
1080                 /*
1081                  * If try_lock is not set yet, make sure all locked inodes
1082                  * are not in the AIL.
1083                  * If any are, set try_lock to be used later.
1084                  */
1085
1086                 if (!try_lock) {
1087                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1088                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1089                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1090                                         try_lock++;
1091                                 }
1092                         }
1093                 }
1094
1095                 /*
1096                  * If any of the previous locks we have locked is in the AIL,
1097                  * we must TRY to get the second and subsequent locks. If
1098                  * we can't get any, we must release all we have
1099                  * and try again.
1100                  */
1101
1102                 if (try_lock) {
1103                         /* try_lock must be 0 if i is 0. */
1104                         /*
1105                          * try_lock means we have an inode locked
1106                          * that is in the AIL.
1107                          */
1108                         ASSERT(i != 0);
1109                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1110                                 attempts++;
1111
1112                                 /*
1113                                  * Unlock all previous guys and try again.
1114                                  * xfs_iunlock will try to push the tail
1115                                  * if the inode is in the AIL.
1116                                  */
1117
1118                                 for(j = i - 1; j >= 0; j--) {
1119
1120                                         /*
1121                                          * Check to see if we've already
1122                                          * unlocked this one.
1123                                          * Not the first one going back,
1124                                          * and the inode ptr is the same.
1125                                          */
1126                                         if ((j != (i - 1)) && ips[j] ==
1127                                                                 ips[j+1])
1128                                                 continue;
1129
1130                                         xfs_iunlock(ips[j], lock_mode);
1131                                 }
1132
1133                                 if ((attempts % 5) == 0) {
1134                                         delay(1); /* Don't just spin the CPU */
1135 #ifdef DEBUG
1136                                         xfs_lock_delays++;
1137 #endif
1138                                 }
1139                                 i = 0;
1140                                 try_lock = 0;
1141                                 goto again;
1142                         }
1143                 } else {
1144                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1145                 }
1146         }
1147
1148 #ifdef DEBUG
1149         if (attempts) {
1150                 if (attempts < 5) xfs_small_retries++;
1151                 else if (attempts < 100) xfs_middle_retries++;
1152                 else xfs_lots_retries++;
1153         } else {
1154                 xfs_locked_n++;
1155         }
1156 #endif
1157 }
1158
1159 /*
1160  * xfs_lock_two_inodes() can only be used to lock one type of lock
1161  * at a time - the iolock or the ilock, but not both at once. If
1162  * we lock both at once, lockdep will report false positives saying
1163  * we have violated locking orders.
1164  */
1165 void
1166 xfs_lock_two_inodes(
1167         xfs_inode_t             *ip0,
1168         xfs_inode_t             *ip1,
1169         uint                    lock_mode)
1170 {
1171         xfs_inode_t             *temp;
1172         int                     attempts = 0;
1173         xfs_log_item_t          *lp;
1174
1175         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1176                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1177         ASSERT(ip0->i_ino != ip1->i_ino);
1178
1179         if (ip0->i_ino > ip1->i_ino) {
1180                 temp = ip0;
1181                 ip0 = ip1;
1182                 ip1 = temp;
1183         }
1184
1185  again:
1186         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1187
1188         /*
1189          * If the first lock we have locked is in the AIL, we must TRY to get
1190          * the second lock. If we can't get it, we must release the first one
1191          * and try again.
1192          */
1193         lp = (xfs_log_item_t *)ip0->i_itemp;
1194         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1195                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1196                         xfs_iunlock(ip0, lock_mode);
1197                         if ((++attempts % 5) == 0)
1198                                 delay(1); /* Don't just spin the CPU */
1199                         goto again;
1200                 }
1201         } else {
1202                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1203         }
1204 }
1205
1206 int
1207 xfs_remove(
1208         xfs_inode_t             *dp,
1209         struct xfs_name         *name,
1210         xfs_inode_t             *ip)
1211 {
1212         xfs_mount_t             *mp = dp->i_mount;
1213         xfs_trans_t             *tp = NULL;
1214         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1215         int                     error = 0;
1216         xfs_bmap_free_t         free_list;
1217         xfs_fsblock_t           first_block;
1218         int                     cancel_flags;
1219         int                     committed;
1220         int                     link_zero;
1221         uint                    resblks;
1222         uint                    log_count;
1223
1224         trace_xfs_remove(dp, name);
1225
1226         if (XFS_FORCED_SHUTDOWN(mp))
1227                 return XFS_ERROR(EIO);
1228
1229         error = xfs_qm_dqattach(dp, 0);
1230         if (error)
1231                 goto std_return;
1232
1233         error = xfs_qm_dqattach(ip, 0);
1234         if (error)
1235                 goto std_return;
1236
1237         if (is_dir) {
1238                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1239                 log_count = XFS_DEFAULT_LOG_COUNT;
1240         } else {
1241                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1242                 log_count = XFS_REMOVE_LOG_COUNT;
1243         }
1244         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1245
1246         /*
1247          * We try to get the real space reservation first,
1248          * allowing for directory btree deletion(s) implying
1249          * possible bmap insert(s).  If we can't get the space
1250          * reservation then we use 0 instead, and avoid the bmap
1251          * btree insert(s) in the directory code by, if the bmap
1252          * insert tries to happen, instead trimming the LAST
1253          * block from the directory.
1254          */
1255         resblks = XFS_REMOVE_SPACE_RES(mp);
1256         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1257                                   XFS_TRANS_PERM_LOG_RES, log_count);
1258         if (error == ENOSPC) {
1259                 resblks = 0;
1260                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1261                                           XFS_TRANS_PERM_LOG_RES, log_count);
1262         }
1263         if (error) {
1264                 ASSERT(error != ENOSPC);
1265                 cancel_flags = 0;
1266                 goto out_trans_cancel;
1267         }
1268
1269         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1270
1271         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1272         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1273
1274         /*
1275          * If we're removing a directory perform some additional validation.
1276          */
1277         if (is_dir) {
1278                 ASSERT(ip->i_d.di_nlink >= 2);
1279                 if (ip->i_d.di_nlink != 2) {
1280                         error = XFS_ERROR(ENOTEMPTY);
1281                         goto out_trans_cancel;
1282                 }
1283                 if (!xfs_dir_isempty(ip)) {
1284                         error = XFS_ERROR(ENOTEMPTY);
1285                         goto out_trans_cancel;
1286                 }
1287         }
1288
1289         xfs_bmap_init(&free_list, &first_block);
1290         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1291                                         &first_block, &free_list, resblks);
1292         if (error) {
1293                 ASSERT(error != ENOENT);
1294                 goto out_bmap_cancel;
1295         }
1296         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1297
1298         if (is_dir) {
1299                 /*
1300                  * Drop the link from ip's "..".
1301                  */
1302                 error = xfs_droplink(tp, dp);
1303                 if (error)
1304                         goto out_bmap_cancel;
1305
1306                 /*
1307                  * Drop the "." link from ip to self.
1308                  */
1309                 error = xfs_droplink(tp, ip);
1310                 if (error)
1311                         goto out_bmap_cancel;
1312         } else {
1313                 /*
1314                  * When removing a non-directory we need to log the parent
1315                  * inode here.  For a directory this is done implicitly
1316                  * by the xfs_droplink call for the ".." entry.
1317                  */
1318                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1319         }
1320
1321         /*
1322          * Drop the link from dp to ip.
1323          */
1324         error = xfs_droplink(tp, ip);
1325         if (error)
1326                 goto out_bmap_cancel;
1327
1328         /*
1329          * Determine if this is the last link while
1330          * we are in the transaction.
1331          */
1332         link_zero = (ip->i_d.di_nlink == 0);
1333
1334         /*
1335          * If this is a synchronous mount, make sure that the
1336          * remove transaction goes to disk before returning to
1337          * the user.
1338          */
1339         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1340                 xfs_trans_set_sync(tp);
1341
1342         error = xfs_bmap_finish(&tp, &free_list, &committed);
1343         if (error)
1344                 goto out_bmap_cancel;
1345
1346         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1347         if (error)
1348                 goto std_return;
1349
1350         /*
1351          * If we are using filestreams, kill the stream association.
1352          * If the file is still open it may get a new one but that
1353          * will get killed on last close in xfs_close() so we don't
1354          * have to worry about that.
1355          */
1356         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1357                 xfs_filestream_deassociate(ip);
1358
1359         return 0;
1360
1361  out_bmap_cancel:
1362         xfs_bmap_cancel(&free_list);
1363         cancel_flags |= XFS_TRANS_ABORT;
1364  out_trans_cancel:
1365         xfs_trans_cancel(tp, cancel_flags);
1366  std_return:
1367         return error;
1368 }
1369
1370 int
1371 xfs_link(
1372         xfs_inode_t             *tdp,
1373         xfs_inode_t             *sip,
1374         struct xfs_name         *target_name)
1375 {
1376         xfs_mount_t             *mp = tdp->i_mount;
1377         xfs_trans_t             *tp;
1378         int                     error;
1379         xfs_bmap_free_t         free_list;
1380         xfs_fsblock_t           first_block;
1381         int                     cancel_flags;
1382         int                     committed;
1383         int                     resblks;
1384
1385         trace_xfs_link(tdp, target_name);
1386
1387         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1388
1389         if (XFS_FORCED_SHUTDOWN(mp))
1390                 return XFS_ERROR(EIO);
1391
1392         error = xfs_qm_dqattach(sip, 0);
1393         if (error)
1394                 goto std_return;
1395
1396         error = xfs_qm_dqattach(tdp, 0);
1397         if (error)
1398                 goto std_return;
1399
1400         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1401         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1402         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1403         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1404                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1405         if (error == ENOSPC) {
1406                 resblks = 0;
1407                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1408                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1409         }
1410         if (error) {
1411                 cancel_flags = 0;
1412                 goto error_return;
1413         }
1414
1415         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1416
1417         xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1418         xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1419
1420         /*
1421          * If we are using project inheritance, we only allow hard link
1422          * creation in our tree when the project IDs are the same; else
1423          * the tree quota mechanism could be circumvented.
1424          */
1425         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1426                      (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1427                 error = XFS_ERROR(EXDEV);
1428                 goto error_return;
1429         }
1430
1431         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1432         if (error)
1433                 goto error_return;
1434
1435         xfs_bmap_init(&free_list, &first_block);
1436
1437         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1438                                         &first_block, &free_list, resblks);
1439         if (error)
1440                 goto abort_return;
1441         xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1442         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1443
1444         error = xfs_bumplink(tp, sip);
1445         if (error)
1446                 goto abort_return;
1447
1448         /*
1449          * If this is a synchronous mount, make sure that the
1450          * link transaction goes to disk before returning to
1451          * the user.
1452          */
1453         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1454                 xfs_trans_set_sync(tp);
1455         }
1456
1457         error = xfs_bmap_finish (&tp, &free_list, &committed);
1458         if (error) {
1459                 xfs_bmap_cancel(&free_list);
1460                 goto abort_return;
1461         }
1462
1463         return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1464
1465  abort_return:
1466         cancel_flags |= XFS_TRANS_ABORT;
1467  error_return:
1468         xfs_trans_cancel(tp, cancel_flags);
1469  std_return:
1470         return error;
1471 }
1472
1473 int
1474 xfs_symlink(
1475         xfs_inode_t             *dp,
1476         struct xfs_name         *link_name,
1477         const char              *target_path,
1478         umode_t                 mode,
1479         xfs_inode_t             **ipp)
1480 {
1481         xfs_mount_t             *mp = dp->i_mount;
1482         xfs_trans_t             *tp;
1483         xfs_inode_t             *ip;
1484         int                     error;
1485         int                     pathlen;
1486         xfs_bmap_free_t         free_list;
1487         xfs_fsblock_t           first_block;
1488         boolean_t               unlock_dp_on_error = B_FALSE;
1489         uint                    cancel_flags;
1490         int                     committed;
1491         xfs_fileoff_t           first_fsb;
1492         xfs_filblks_t           fs_blocks;
1493         int                     nmaps;
1494         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
1495         xfs_daddr_t             d;
1496         const char              *cur_chunk;
1497         int                     byte_cnt;
1498         int                     n;
1499         xfs_buf_t               *bp;
1500         prid_t                  prid;
1501         struct xfs_dquot        *udqp, *gdqp;
1502         uint                    resblks;
1503
1504         *ipp = NULL;
1505         error = 0;
1506         ip = NULL;
1507         tp = NULL;
1508
1509         trace_xfs_symlink(dp, link_name);
1510
1511         if (XFS_FORCED_SHUTDOWN(mp))
1512                 return XFS_ERROR(EIO);
1513
1514         /*
1515          * Check component lengths of the target path name.
1516          */
1517         pathlen = strlen(target_path);
1518         if (pathlen >= MAXPATHLEN)      /* total string too long */
1519                 return XFS_ERROR(ENAMETOOLONG);
1520
1521         udqp = gdqp = NULL;
1522         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1523                 prid = xfs_get_projid(dp);
1524         else
1525                 prid = XFS_PROJID_DEFAULT;
1526
1527         /*
1528          * Make sure that we have allocated dquot(s) on disk.
1529          */
1530         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1531                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1532         if (error)
1533                 goto std_return;
1534
1535         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
1536         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1537         /*
1538          * The symlink will fit into the inode data fork?
1539          * There can't be any attributes so we get the whole variable part.
1540          */
1541         if (pathlen <= XFS_LITINO(mp))
1542                 fs_blocks = 0;
1543         else
1544                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
1545         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
1546         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
1547                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1548         if (error == ENOSPC && fs_blocks == 0) {
1549                 resblks = 0;
1550                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
1551                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1552         }
1553         if (error) {
1554                 cancel_flags = 0;
1555                 goto error_return;
1556         }
1557
1558         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1559         unlock_dp_on_error = B_TRUE;
1560
1561         /*
1562          * Check whether the directory allows new symlinks or not.
1563          */
1564         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
1565                 error = XFS_ERROR(EPERM);
1566                 goto error_return;
1567         }
1568
1569         /*
1570          * Reserve disk quota : blocks and inode.
1571          */
1572         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
1573         if (error)
1574                 goto error_return;
1575
1576         /*
1577          * Check for ability to enter directory entry, if no space reserved.
1578          */
1579         error = xfs_dir_canenter(tp, dp, link_name, resblks);
1580         if (error)
1581                 goto error_return;
1582         /*
1583          * Initialize the bmap freelist prior to calling either
1584          * bmapi or the directory create code.
1585          */
1586         xfs_bmap_init(&free_list, &first_block);
1587
1588         /*
1589          * Allocate an inode for the symlink.
1590          */
1591         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT), 1, 0,
1592                                prid, resblks > 0, &ip, NULL);
1593         if (error) {
1594                 if (error == ENOSPC)
1595                         goto error_return;
1596                 goto error1;
1597         }
1598
1599         /*
1600          * An error after we've joined dp to the transaction will result in the
1601          * transaction cancel unlocking dp so don't do it explicitly in the
1602          * error path.
1603          */
1604         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1605         unlock_dp_on_error = B_FALSE;
1606
1607         /*
1608          * Also attach the dquot(s) to it, if applicable.
1609          */
1610         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
1611
1612         if (resblks)
1613                 resblks -= XFS_IALLOC_SPACE_RES(mp);
1614         /*
1615          * If the symlink will fit into the inode, write it inline.
1616          */
1617         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
1618                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
1619                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
1620                 ip->i_d.di_size = pathlen;
1621
1622                 /*
1623                  * The inode was initially created in extent format.
1624                  */
1625                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
1626                 ip->i_df.if_flags |= XFS_IFINLINE;
1627
1628                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
1629                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
1630
1631         } else {
1632                 first_fsb = 0;
1633                 nmaps = SYMLINK_MAPS;
1634
1635                 error = xfs_bmapi_write(tp, ip, first_fsb, fs_blocks,
1636                                   XFS_BMAPI_METADATA, &first_block, resblks,
1637                                   mval, &nmaps, &free_list);
1638                 if (error)
1639                         goto error2;
1640
1641                 if (resblks)
1642                         resblks -= fs_blocks;
1643                 ip->i_d.di_size = pathlen;
1644                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1645
1646                 cur_chunk = target_path;
1647                 for (n = 0; n < nmaps; n++) {
1648                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
1649                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
1650                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
1651                                                BTOBB(byte_cnt), 0);
1652                         if (!bp) {
1653                                 error = ENOMEM;
1654                                 goto error2;
1655                         }
1656                         if (pathlen < byte_cnt) {
1657                                 byte_cnt = pathlen;
1658                         }
1659                         pathlen -= byte_cnt;
1660
1661                         memcpy(bp->b_addr, cur_chunk, byte_cnt);
1662                         cur_chunk += byte_cnt;
1663
1664                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
1665                 }
1666         }
1667
1668         /*
1669          * Create the directory entry for the symlink.
1670          */
1671         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
1672                                         &first_block, &free_list, resblks);
1673         if (error)
1674                 goto error2;
1675         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1676         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1677
1678         /*
1679          * If this is a synchronous mount, make sure that the
1680          * symlink transaction goes to disk before returning to
1681          * the user.
1682          */
1683         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1684                 xfs_trans_set_sync(tp);
1685         }
1686
1687         error = xfs_bmap_finish(&tp, &free_list, &committed);
1688         if (error) {
1689                 goto error2;
1690         }
1691         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1692         xfs_qm_dqrele(udqp);
1693         xfs_qm_dqrele(gdqp);
1694
1695         *ipp = ip;
1696         return 0;
1697
1698  error2:
1699         IRELE(ip);
1700  error1:
1701         xfs_bmap_cancel(&free_list);
1702         cancel_flags |= XFS_TRANS_ABORT;
1703  error_return:
1704         xfs_trans_cancel(tp, cancel_flags);
1705         xfs_qm_dqrele(udqp);
1706         xfs_qm_dqrele(gdqp);
1707
1708         if (unlock_dp_on_error)
1709                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1710  std_return:
1711         return error;
1712 }
1713
1714 int
1715 xfs_set_dmattrs(
1716         xfs_inode_t     *ip,
1717         u_int           evmask,
1718         u_int16_t       state)
1719 {
1720         xfs_mount_t     *mp = ip->i_mount;
1721         xfs_trans_t     *tp;
1722         int             error;
1723
1724         if (!capable(CAP_SYS_ADMIN))
1725                 return XFS_ERROR(EPERM);
1726
1727         if (XFS_FORCED_SHUTDOWN(mp))
1728                 return XFS_ERROR(EIO);
1729
1730         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
1731         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
1732         if (error) {
1733                 xfs_trans_cancel(tp, 0);
1734                 return error;
1735         }
1736         xfs_ilock(ip, XFS_ILOCK_EXCL);
1737         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1738
1739         ip->i_d.di_dmevmask = evmask;
1740         ip->i_d.di_dmstate  = state;
1741
1742         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1743         error = xfs_trans_commit(tp, 0);
1744
1745         return error;
1746 }
1747
1748 /*
1749  * xfs_alloc_file_space()
1750  *      This routine allocates disk space for the given file.
1751  *
1752  *      If alloc_type == 0, this request is for an ALLOCSP type
1753  *      request which will change the file size.  In this case, no
1754  *      DMAPI event will be generated by the call.  A TRUNCATE event
1755  *      will be generated later by xfs_setattr.
1756  *
1757  *      If alloc_type != 0, this request is for a RESVSP type
1758  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
1759  *      lower block boundary byte address is less than the file's
1760  *      length.
1761  *
1762  * RETURNS:
1763  *       0 on success
1764  *      errno on error
1765  *
1766  */
1767 STATIC int
1768 xfs_alloc_file_space(
1769         xfs_inode_t             *ip,
1770         xfs_off_t               offset,
1771         xfs_off_t               len,
1772         int                     alloc_type,
1773         int                     attr_flags)
1774 {
1775         xfs_mount_t             *mp = ip->i_mount;
1776         xfs_off_t               count;
1777         xfs_filblks_t           allocated_fsb;
1778         xfs_filblks_t           allocatesize_fsb;
1779         xfs_extlen_t            extsz, temp;
1780         xfs_fileoff_t           startoffset_fsb;
1781         xfs_fsblock_t           firstfsb;
1782         int                     nimaps;
1783         int                     quota_flag;
1784         int                     rt;
1785         xfs_trans_t             *tp;
1786         xfs_bmbt_irec_t         imaps[1], *imapp;
1787         xfs_bmap_free_t         free_list;
1788         uint                    qblocks, resblks, resrtextents;
1789         int                     committed;
1790         int                     error;
1791
1792         trace_xfs_alloc_file_space(ip);
1793
1794         if (XFS_FORCED_SHUTDOWN(mp))
1795                 return XFS_ERROR(EIO);
1796
1797         error = xfs_qm_dqattach(ip, 0);
1798         if (error)
1799                 return error;
1800
1801         if (len <= 0)
1802                 return XFS_ERROR(EINVAL);
1803
1804         rt = XFS_IS_REALTIME_INODE(ip);
1805         extsz = xfs_get_extsz_hint(ip);
1806
1807         count = len;
1808         imapp = &imaps[0];
1809         nimaps = 1;
1810         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
1811         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
1812
1813         /*
1814          * Allocate file space until done or until there is an error
1815          */
1816         while (allocatesize_fsb && !error) {
1817                 xfs_fileoff_t   s, e;
1818
1819                 /*
1820                  * Determine space reservations for data/realtime.
1821                  */
1822                 if (unlikely(extsz)) {
1823                         s = startoffset_fsb;
1824                         do_div(s, extsz);
1825                         s *= extsz;
1826                         e = startoffset_fsb + allocatesize_fsb;
1827                         if ((temp = do_mod(startoffset_fsb, extsz)))
1828                                 e += temp;
1829                         if ((temp = do_mod(e, extsz)))
1830                                 e += extsz - temp;
1831                 } else {
1832                         s = 0;
1833                         e = allocatesize_fsb;
1834                 }
1835
1836                 /*
1837                  * The transaction reservation is limited to a 32-bit block
1838                  * count, hence we need to limit the number of blocks we are
1839                  * trying to reserve to avoid an overflow. We can't allocate
1840                  * more than @nimaps extents, and an extent is limited on disk
1841                  * to MAXEXTLEN (21 bits), so use that to enforce the limit.
1842                  */
1843                 resblks = min_t(xfs_fileoff_t, (e - s), (MAXEXTLEN * nimaps));
1844                 if (unlikely(rt)) {
1845                         resrtextents = qblocks = resblks;
1846                         resrtextents /= mp->m_sb.sb_rextsize;
1847                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
1848                         quota_flag = XFS_QMOPT_RES_RTBLKS;
1849                 } else {
1850                         resrtextents = 0;
1851                         resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resblks);
1852                         quota_flag = XFS_QMOPT_RES_REGBLKS;
1853                 }
1854
1855                 /*
1856                  * Allocate and setup the transaction.
1857                  */
1858                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
1859                 error = xfs_trans_reserve(tp, resblks,
1860                                           XFS_WRITE_LOG_RES(mp), resrtextents,
1861                                           XFS_TRANS_PERM_LOG_RES,
1862                                           XFS_WRITE_LOG_COUNT);
1863                 /*
1864                  * Check for running out of space
1865                  */
1866                 if (error) {
1867                         /*
1868                          * Free the transaction structure.
1869                          */
1870                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
1871                         xfs_trans_cancel(tp, 0);
1872                         break;
1873                 }
1874                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1875                 error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks,
1876                                                       0, quota_flag);
1877                 if (error)
1878                         goto error1;
1879
1880                 xfs_trans_ijoin(tp, ip, 0);
1881
1882                 xfs_bmap_init(&free_list, &firstfsb);
1883                 error = xfs_bmapi_write(tp, ip, startoffset_fsb,
1884                                         allocatesize_fsb, alloc_type, &firstfsb,
1885                                         0, imapp, &nimaps, &free_list);
1886                 if (error) {
1887                         goto error0;
1888                 }
1889
1890                 /*
1891                  * Complete the transaction
1892                  */
1893                 error = xfs_bmap_finish(&tp, &free_list, &committed);
1894                 if (error) {
1895                         goto error0;
1896                 }
1897
1898                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1899                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1900                 if (error) {
1901                         break;
1902                 }
1903
1904                 allocated_fsb = imapp->br_blockcount;
1905
1906                 if (nimaps == 0) {
1907                         error = XFS_ERROR(ENOSPC);
1908                         break;
1909                 }
1910
1911                 startoffset_fsb += allocated_fsb;
1912                 allocatesize_fsb -= allocated_fsb;
1913         }
1914
1915         return error;
1916
1917 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
1918         xfs_bmap_cancel(&free_list);
1919         xfs_trans_unreserve_quota_nblks(tp, ip, (long)qblocks, 0, quota_flag);
1920
1921 error1: /* Just cancel transaction */
1922         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1923         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1924         return error;
1925 }
1926
1927 /*
1928  * Zero file bytes between startoff and endoff inclusive.
1929  * The iolock is held exclusive and no blocks are buffered.
1930  *
1931  * This function is used by xfs_free_file_space() to zero
1932  * partial blocks when the range to free is not block aligned.
1933  * When unreserving space with boundaries that are not block
1934  * aligned we round up the start and round down the end
1935  * boundaries and then use this function to zero the parts of
1936  * the blocks that got dropped during the rounding.
1937  */
1938 STATIC int
1939 xfs_zero_remaining_bytes(
1940         xfs_inode_t             *ip,
1941         xfs_off_t               startoff,
1942         xfs_off_t               endoff)
1943 {
1944         xfs_bmbt_irec_t         imap;
1945         xfs_fileoff_t           offset_fsb;
1946         xfs_off_t               lastoffset;
1947         xfs_off_t               offset;
1948         xfs_buf_t               *bp;
1949         xfs_mount_t             *mp = ip->i_mount;
1950         int                     nimap;
1951         int                     error = 0;
1952
1953         /*
1954          * Avoid doing I/O beyond eof - it's not necessary
1955          * since nothing can read beyond eof.  The space will
1956          * be zeroed when the file is extended anyway.
1957          */
1958         if (startoff >= XFS_ISIZE(ip))
1959                 return 0;
1960
1961         if (endoff > XFS_ISIZE(ip))
1962                 endoff = XFS_ISIZE(ip);
1963
1964         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
1965                                         mp->m_rtdev_targp : mp->m_ddev_targp,
1966                                   BTOBB(mp->m_sb.sb_blocksize), 0);
1967         if (!bp)
1968                 return XFS_ERROR(ENOMEM);
1969
1970         xfs_buf_unlock(bp);
1971
1972         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
1973                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1974                 nimap = 1;
1975                 error = xfs_bmapi_read(ip, offset_fsb, 1, &imap, &nimap, 0);
1976                 if (error || nimap < 1)
1977                         break;
1978                 ASSERT(imap.br_blockcount >= 1);
1979                 ASSERT(imap.br_startoff == offset_fsb);
1980                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
1981                 if (lastoffset > endoff)
1982                         lastoffset = endoff;
1983                 if (imap.br_startblock == HOLESTARTBLOCK)
1984                         continue;
1985                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1986                 if (imap.br_state == XFS_EXT_UNWRITTEN)
1987                         continue;
1988                 XFS_BUF_UNDONE(bp);
1989                 XFS_BUF_UNWRITE(bp);
1990                 XFS_BUF_READ(bp);
1991                 XFS_BUF_SET_ADDR(bp, xfs_fsb_to_db(ip, imap.br_startblock));
1992                 xfsbdstrat(mp, bp);
1993                 error = xfs_buf_iowait(bp);
1994                 if (error) {
1995                         xfs_buf_ioerror_alert(bp,
1996                                         "xfs_zero_remaining_bytes(read)");
1997                         break;
1998                 }
1999                 memset(bp->b_addr +
2000                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
2001                       0, lastoffset - offset + 1);
2002                 XFS_BUF_UNDONE(bp);
2003                 XFS_BUF_UNREAD(bp);
2004                 XFS_BUF_WRITE(bp);
2005                 xfsbdstrat(mp, bp);
2006                 error = xfs_buf_iowait(bp);
2007                 if (error) {
2008                         xfs_buf_ioerror_alert(bp,
2009                                         "xfs_zero_remaining_bytes(write)");
2010                         break;
2011                 }
2012         }
2013         xfs_buf_free(bp);
2014         return error;
2015 }
2016
2017 /*
2018  * xfs_free_file_space()
2019  *      This routine frees disk space for the given file.
2020  *
2021  *      This routine is only called by xfs_change_file_space
2022  *      for an UNRESVSP type call.
2023  *
2024  * RETURNS:
2025  *       0 on success
2026  *      errno on error
2027  *
2028  */
2029 STATIC int
2030 xfs_free_file_space(
2031         xfs_inode_t             *ip,
2032         xfs_off_t               offset,
2033         xfs_off_t               len,
2034         int                     attr_flags)
2035 {
2036         int                     committed;
2037         int                     done;
2038         xfs_fileoff_t           endoffset_fsb;
2039         int                     error;
2040         xfs_fsblock_t           firstfsb;
2041         xfs_bmap_free_t         free_list;
2042         xfs_bmbt_irec_t         imap;
2043         xfs_off_t               ioffset;
2044         xfs_extlen_t            mod=0;
2045         xfs_mount_t             *mp;
2046         int                     nimap;
2047         uint                    resblks;
2048         uint                    rounding;
2049         int                     rt;
2050         xfs_fileoff_t           startoffset_fsb;
2051         xfs_trans_t             *tp;
2052         int                     need_iolock = 1;
2053
2054         mp = ip->i_mount;
2055
2056         trace_xfs_free_file_space(ip);
2057
2058         error = xfs_qm_dqattach(ip, 0);
2059         if (error)
2060                 return error;
2061
2062         error = 0;
2063         if (len <= 0)   /* if nothing being freed */
2064                 return error;
2065         rt = XFS_IS_REALTIME_INODE(ip);
2066         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
2067         endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
2068
2069         if (attr_flags & XFS_ATTR_NOLOCK)
2070                 need_iolock = 0;
2071         if (need_iolock) {
2072                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
2073                 /* wait for the completion of any pending DIOs */
2074                 inode_dio_wait(VFS_I(ip));
2075         }
2076
2077         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
2078         ioffset = offset & ~(rounding - 1);
2079
2080         if (VN_CACHED(VFS_I(ip)) != 0) {
2081                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
2082                 if (error)
2083                         goto out_unlock_iolock;
2084         }
2085
2086         /*
2087          * Need to zero the stuff we're not freeing, on disk.
2088          * If it's a realtime file & can't use unwritten extents then we
2089          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
2090          * will take care of it for us.
2091          */
2092         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
2093                 nimap = 1;
2094                 error = xfs_bmapi_read(ip, startoffset_fsb, 1,
2095                                         &imap, &nimap, 0);
2096                 if (error)
2097                         goto out_unlock_iolock;
2098                 ASSERT(nimap == 0 || nimap == 1);
2099                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2100                         xfs_daddr_t     block;
2101
2102                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2103                         block = imap.br_startblock;
2104                         mod = do_div(block, mp->m_sb.sb_rextsize);
2105                         if (mod)
2106                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
2107                 }
2108                 nimap = 1;
2109                 error = xfs_bmapi_read(ip, endoffset_fsb - 1, 1,
2110                                         &imap, &nimap, 0);
2111                 if (error)
2112                         goto out_unlock_iolock;
2113                 ASSERT(nimap == 0 || nimap == 1);
2114                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
2115                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
2116                         mod++;
2117                         if (mod && (mod != mp->m_sb.sb_rextsize))
2118                                 endoffset_fsb -= mod;
2119                 }
2120         }
2121         if ((done = (endoffset_fsb <= startoffset_fsb)))
2122                 /*
2123                  * One contiguous piece to clear
2124                  */
2125                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
2126         else {
2127                 /*
2128                  * Some full blocks, possibly two pieces to clear
2129                  */
2130                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
2131                         error = xfs_zero_remaining_bytes(ip, offset,
2132                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
2133                 if (!error &&
2134                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
2135                         error = xfs_zero_remaining_bytes(ip,
2136                                 XFS_FSB_TO_B(mp, endoffset_fsb),
2137                                 offset + len - 1);
2138         }
2139
2140         /*
2141          * free file space until done or until there is an error
2142          */
2143         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2144         while (!error && !done) {
2145
2146                 /*
2147                  * allocate and setup the transaction. Allow this
2148                  * transaction to dip into the reserve blocks to ensure
2149                  * the freeing of the space succeeds at ENOSPC.
2150                  */
2151                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2152                 tp->t_flags |= XFS_TRANS_RESERVE;
2153                 error = xfs_trans_reserve(tp,
2154                                           resblks,
2155                                           XFS_WRITE_LOG_RES(mp),
2156                                           0,
2157                                           XFS_TRANS_PERM_LOG_RES,
2158                                           XFS_WRITE_LOG_COUNT);
2159
2160                 /*
2161                  * check for running out of space
2162                  */
2163                 if (error) {
2164                         /*
2165                          * Free the transaction structure.
2166                          */
2167                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2168                         xfs_trans_cancel(tp, 0);
2169                         break;
2170                 }
2171                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2172                 error = xfs_trans_reserve_quota(tp, mp,
2173                                 ip->i_udquot, ip->i_gdquot,
2174                                 resblks, 0, XFS_QMOPT_RES_REGBLKS);
2175                 if (error)
2176                         goto error1;
2177
2178                 xfs_trans_ijoin(tp, ip, 0);
2179
2180                 /*
2181                  * issue the bunmapi() call to free the blocks
2182                  */
2183                 xfs_bmap_init(&free_list, &firstfsb);
2184                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
2185                                   endoffset_fsb - startoffset_fsb,
2186                                   0, 2, &firstfsb, &free_list, &done);
2187                 if (error) {
2188                         goto error0;
2189                 }
2190
2191                 /*
2192                  * complete the transaction
2193                  */
2194                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2195                 if (error) {
2196                         goto error0;
2197                 }
2198
2199                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2200                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2201         }
2202
2203  out_unlock_iolock:
2204         if (need_iolock)
2205                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2206         return error;
2207
2208  error0:
2209         xfs_bmap_cancel(&free_list);
2210  error1:
2211         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2212         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
2213                     XFS_ILOCK_EXCL);
2214         return error;
2215 }
2216
2217 /*
2218  * xfs_change_file_space()
2219  *      This routine allocates or frees disk space for the given file.
2220  *      The user specified parameters are checked for alignment and size
2221  *      limitations.
2222  *
2223  * RETURNS:
2224  *       0 on success
2225  *      errno on error
2226  *
2227  */
2228 int
2229 xfs_change_file_space(
2230         xfs_inode_t     *ip,
2231         int             cmd,
2232         xfs_flock64_t   *bf,
2233         xfs_off_t       offset,
2234         int             attr_flags)
2235 {
2236         xfs_mount_t     *mp = ip->i_mount;
2237         int             clrprealloc;
2238         int             error;
2239         xfs_fsize_t     fsize;
2240         int             setprealloc;
2241         xfs_off_t       startoffset;
2242         xfs_off_t       llen;
2243         xfs_trans_t     *tp;
2244         struct iattr    iattr;
2245         int             prealloc_type;
2246
2247         if (!S_ISREG(ip->i_d.di_mode))
2248                 return XFS_ERROR(EINVAL);
2249
2250         switch (bf->l_whence) {
2251         case 0: /*SEEK_SET*/
2252                 break;
2253         case 1: /*SEEK_CUR*/
2254                 bf->l_start += offset;
2255                 break;
2256         case 2: /*SEEK_END*/
2257                 bf->l_start += XFS_ISIZE(ip);
2258                 break;
2259         default:
2260                 return XFS_ERROR(EINVAL);
2261         }
2262
2263         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
2264
2265         if (   (bf->l_start < 0)
2266             || (bf->l_start > XFS_MAXIOFFSET(mp))
2267             || (bf->l_start + llen < 0)
2268             || (bf->l_start + llen > XFS_MAXIOFFSET(mp)))
2269                 return XFS_ERROR(EINVAL);
2270
2271         bf->l_whence = 0;
2272
2273         startoffset = bf->l_start;
2274         fsize = XFS_ISIZE(ip);
2275
2276         /*
2277          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
2278          * file space.
2279          * These calls do NOT zero the data space allocated to the file,
2280          * nor do they change the file size.
2281          *
2282          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
2283          * space.
2284          * These calls cause the new file data to be zeroed and the file
2285          * size to be changed.
2286          */
2287         setprealloc = clrprealloc = 0;
2288         prealloc_type = XFS_BMAPI_PREALLOC;
2289
2290         switch (cmd) {
2291         case XFS_IOC_ZERO_RANGE:
2292                 prealloc_type |= XFS_BMAPI_CONVERT;
2293                 xfs_tosspages(ip, startoffset, startoffset + bf->l_len, 0);
2294                 /* FALLTHRU */
2295         case XFS_IOC_RESVSP:
2296         case XFS_IOC_RESVSP64:
2297                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
2298                                                 prealloc_type, attr_flags);
2299                 if (error)
2300                         return error;
2301                 setprealloc = 1;
2302                 break;
2303
2304         case XFS_IOC_UNRESVSP:
2305         case XFS_IOC_UNRESVSP64:
2306                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
2307                                                                 attr_flags)))
2308                         return error;
2309                 break;
2310
2311         case XFS_IOC_ALLOCSP:
2312         case XFS_IOC_ALLOCSP64:
2313         case XFS_IOC_FREESP:
2314         case XFS_IOC_FREESP64:
2315                 /*
2316                  * These operations actually do IO when extending the file, but
2317                  * the allocation is done seperately to the zeroing that is
2318                  * done. This set of operations need to be serialised against
2319                  * other IO operations, such as truncate and buffered IO. We
2320                  * need to take the IOLOCK here to serialise the allocation and
2321                  * zeroing IO to prevent other IOLOCK holders (e.g. getbmap,
2322                  * truncate, direct IO) from racing against the transient
2323                  * allocated but not written state we can have here.
2324                  */
2325                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
2326                 if (startoffset > fsize) {
2327                         error = xfs_alloc_file_space(ip, fsize,
2328                                         startoffset - fsize, 0,
2329                                         attr_flags | XFS_ATTR_NOLOCK);
2330                         if (error) {
2331                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2332                                 break;
2333                         }
2334                 }
2335
2336                 iattr.ia_valid = ATTR_SIZE;
2337                 iattr.ia_size = startoffset;
2338
2339                 error = xfs_setattr_size(ip, &iattr,
2340                                          attr_flags | XFS_ATTR_NOLOCK);
2341                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2342
2343                 if (error)
2344                         return error;
2345
2346                 clrprealloc = 1;
2347                 break;
2348
2349         default:
2350                 ASSERT(0);
2351                 return XFS_ERROR(EINVAL);
2352         }
2353
2354         /*
2355          * update the inode timestamp, mode, and prealloc flag bits
2356          */
2357         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
2358
2359         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
2360                                       0, 0, 0))) {
2361                 /* ASSERT(0); */
2362                 xfs_trans_cancel(tp, 0);
2363                 return error;
2364         }
2365
2366         xfs_ilock(ip, XFS_ILOCK_EXCL);
2367         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2368
2369         if ((attr_flags & XFS_ATTR_DMI) == 0) {
2370                 ip->i_d.di_mode &= ~S_ISUID;
2371
2372                 /*
2373                  * Note that we don't have to worry about mandatory
2374                  * file locking being disabled here because we only
2375                  * clear the S_ISGID bit if the Group execute bit is
2376                  * on, but if it was on then mandatory locking wouldn't
2377                  * have been enabled.
2378                  */
2379                 if (ip->i_d.di_mode & S_IXGRP)
2380                         ip->i_d.di_mode &= ~S_ISGID;
2381
2382                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2383         }
2384         if (setprealloc)
2385                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
2386         else if (clrprealloc)
2387                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
2388
2389         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2390         if (attr_flags & XFS_ATTR_SYNC)
2391                 xfs_trans_set_sync(tp);
2392         return xfs_trans_commit(tp, 0);
2393 }