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_inum.h"
  25 #include "xfs_trans.h"
  26 #include "xfs_sb.h"
  27 #include "xfs_ag.h"
  28 #include "xfs_dir2.h"
  29 #include "xfs_mount.h"
  30 #include "xfs_da_btree.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_ialloc_btree.h"
  33 #include "xfs_dinode.h"
  34 #include "xfs_inode.h"
  35 #include "xfs_inode_item.h"
  36 #include "xfs_itable.h"
  37 #include "xfs_ialloc.h"
  38 #include "xfs_alloc.h"
  39 #include "xfs_bmap.h"
  40 #include "xfs_acl.h"
  41 #include "xfs_attr.h"
  42 #include "xfs_rw.h"
  43 #include "xfs_error.h"
  44 #include "xfs_quota.h"
  45 #include "xfs_utils.h"
  46 #include "xfs_rtalloc.h"
  47 #include "xfs_trans_space.h"
  48 #include "xfs_log_priv.h"
  49 #include "xfs_filestream.h"
  50 #include "xfs_vnodeops.h"
  51 #include "xfs_trace.h"
  52
  53 /*
  54  * The maximum pathlen is 1024 bytes. Since the minimum file system
  55  * blocksize is 512 bytes, we can get a max of 2 extents back from
  56  * bmapi.
  57  */
  58 #define SYMLINK_MAPS 2
  59
  60 STATIC int
  61 xfs_readlink_bmap(
  62         xfs_inode_t     *ip,
  63         char            *link)
  64 {
  65         xfs_mount_t     *mp = ip->i_mount;
  66         int             pathlen = ip->i_d.di_size;
  67         int             nmaps = SYMLINK_MAPS;
  68         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
  69         xfs_daddr_t     d;
  70         int             byte_cnt;
  71         int             n;
  72         xfs_buf_t       *bp;
  73         int             error = 0;
  74
  75         error = xfs_bmapi_read(ip, 0, XFS_B_TO_FSB(mp, pathlen), mval, &nmaps,
  76                                0);
  77         if (error)
  78                 goto out;
  79
  80         for (n = 0; n < nmaps; n++) {
  81                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
  82                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
  83
  84                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt),
  85                                   XBF_LOCK | XBF_MAPPED | XBF_DONT_BLOCK);
  86                 if (!bp)
  87                         return XFS_ERROR(ENOMEM);
  88                 error = bp->b_error;
  89                 if (error) {
  90                         xfs_buf_ioerror_alert(bp, __func__);
  91                         xfs_buf_relse(bp);
  92                         goto out;
  93                 }
  94                 if (pathlen < byte_cnt)
  95                         byte_cnt = pathlen;
  96                 pathlen -= byte_cnt;
  97
  98                 memcpy(link, bp->b_addr, byte_cnt);
  99                 xfs_buf_relse(bp);
 100         }
 101
 102         link[ip->i_d.di_size] = '\0';
 103         error = 0;
 104
 105  out:
 106         return error;
 107 }
 108
 109 int
 110 xfs_readlink(
 111         xfs_inode_t     *ip,
 112         char            *link)
 113 {
 114         xfs_mount_t     *mp = ip->i_mount;
 115         int             pathlen;
 116         int             error = 0;
 117
 118         trace_xfs_readlink(ip);
 119
 120         if (XFS_FORCED_SHUTDOWN(mp))
 121                 return XFS_ERROR(EIO);
 122
 123         xfs_ilock(ip, XFS_ILOCK_SHARED);
 124
 125         ASSERT(S_ISLNK(ip->i_d.di_mode));
 126         ASSERT(ip->i_d.di_size <= MAXPATHLEN);
 127
 128         pathlen = ip->i_d.di_size;
 129         if (!pathlen)
 130                 goto out;
 131
 132         if (ip->i_df.if_flags & XFS_IFINLINE) {
 133                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
 134                 link[pathlen] = '\0';
 135         } else {
 136                 error = xfs_readlink_bmap(ip, link);
 137         }
 138
 139  out:
 140         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 141         return error;
 142 }
 143
 144 /*
 145  * Flags for xfs_free_eofblocks
 146  */
 147 #define XFS_FREE_EOF_TRYLOCK    (1<<0)
 148
 149 /*
 150  * This is called by xfs_inactive to free any blocks beyond eof
 151  * when the link count isn't zero and by xfs_dm_punch_hole() when
 152  * punching a hole to EOF.
 153  */
 154 STATIC int
 155 xfs_free_eofblocks(
 156         xfs_mount_t     *mp,
 157         xfs_inode_t     *ip,
 158         int             flags)
 159 {
 160         xfs_trans_t     *tp;
 161         int             error;
 162         xfs_fileoff_t   end_fsb;
 163         xfs_fileoff_t   last_fsb;
 164         xfs_filblks_t   map_len;
 165         int             nimaps;
 166         xfs_bmbt_irec_t imap;
 167
 168         /*
 169          * Figure out if there are any blocks beyond the end
 170          * of the file.  If not, then there is nothing to do.
 171          */
 172         end_fsb = XFS_B_TO_FSB(mp, ((xfs_ufsize_t)ip->i_size));
 173         last_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_MAXIOFFSET(mp));
 174         if (last_fsb <= end_fsb)
 175                 return 0;
 176         map_len = last_fsb - end_fsb;
 177
 178         nimaps = 1;
 179         xfs_ilock(ip, XFS_ILOCK_SHARED);
 180         error = xfs_bmapi_read(ip, end_fsb, map_len, &imap, &nimaps, 0);
 181         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 182
 183         if (!error && (nimaps != 0) &&
 184             (imap.br_startblock != HOLESTARTBLOCK ||
 185              ip->i_delayed_blks)) {
 186                 /*
 187                  * Attach the dquots to the inode up front.
 188                  */
 189                 error = xfs_qm_dqattach(ip, 0);
 190                 if (error)
 191                         return error;
 192
 193                 /*
 194                  * There are blocks after the end of file.
 195                  * Free them up now by truncating the file to
 196                  * its current size.
 197                  */
 198                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 199
 200                 if (flags & XFS_FREE_EOF_TRYLOCK) {
 201                         if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
 202                                 xfs_trans_cancel(tp, 0);
 203                                 return 0;
 204                         }
 205                 } else {
 206                         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 207                 }
 208
 209                 error = xfs_trans_reserve(tp, 0,
 210                                           XFS_ITRUNCATE_LOG_RES(mp),
 211                                           0, XFS_TRANS_PERM_LOG_RES,
 212                                           XFS_ITRUNCATE_LOG_COUNT);
 213                 if (error) {
 214                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 215                         xfs_trans_cancel(tp, 0);
 216                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 217                         return error;
 218                 }
 219
 220                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 221                 xfs_trans_ijoin(tp, ip, 0);
 222
 223                 error = xfs_itruncate_data(&tp, ip, ip->i_size);
 224                 if (error) {
 225                         /*
 226                          * If we get an error at this point we simply don't
 227                          * bother truncating the file.
 228                          */
 229                         xfs_trans_cancel(tp,
 230                                          (XFS_TRANS_RELEASE_LOG_RES |
 231                                           XFS_TRANS_ABORT));
 232                 } else {
 233                         error = xfs_trans_commit(tp,
 234                                                 XFS_TRANS_RELEASE_LOG_RES);
 235                 }
 236                 xfs_iunlock(ip, XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL);
 237         }
 238         return error;
 239 }
 240
 241 /*
 242  * Free a symlink that has blocks associated with it.
 243  */
 244 STATIC int
 245 xfs_inactive_symlink_rmt(
 246         xfs_inode_t     *ip,
 247         xfs_trans_t     **tpp)
 248 {
 249         xfs_buf_t       *bp;
 250         int             committed;
 251         int             done;
 252         int             error;
 253         xfs_fsblock_t   first_block;
 254         xfs_bmap_free_t free_list;
 255         int             i;
 256         xfs_mount_t     *mp;
 257         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
 258         int             nmaps;
 259         xfs_trans_t     *ntp;
 260         int             size;
 261         xfs_trans_t     *tp;
 262
 263         tp = *tpp;
 264         mp = ip->i_mount;
 265         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
 266         /*
 267          * We're freeing a symlink that has some
 268          * blocks allocated to it.  Free the
 269          * blocks here.  We know that we've got
 270          * either 1 or 2 extents and that we can
 271          * free them all in one bunmapi call.
 272          */
 273         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
 274         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
 275                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
 276                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 277                 xfs_trans_cancel(tp, 0);
 278                 *tpp = NULL;
 279                 return error;
 280         }
 281         /*
 282          * Lock the inode, fix the size, and join it to the transaction.
 283          * Hold it so in the normal path, we still have it locked for
 284          * the second transaction.  In the error paths we need it
 285          * held so the cancel won't rele it, see below.
 286          */
 287         xfs_ilock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
 288         size = (int)ip->i_d.di_size;
 289         ip->i_d.di_size = 0;
 290         xfs_trans_ijoin(tp, ip, 0);
 291         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 292         /*
 293          * Find the block(s) so we can inval and unmap them.
 294          */
 295         done = 0;
 296         xfs_bmap_init(&free_list, &first_block);
 297         nmaps = ARRAY_SIZE(mval);
 298         error = xfs_bmapi_read(ip, 0, XFS_B_TO_FSB(mp, size),
 299                                 mval, &nmaps, 0);
 300         if (error)
 301                 goto error0;
 302         /*
 303          * Invalidate the block(s).
 304          */
 305         for (i = 0; i < nmaps; i++) {
 306                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 307                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
 308                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
 309                 if (!bp) {
 310                         error = ENOMEM;
 311                         goto error1;
 312                 }
 313                 xfs_trans_binval(tp, bp);
 314         }
 315         /*
 316          * Unmap the dead block(s) to the free_list.
 317          */
 318         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
 319                         &first_block, &free_list, &done)))
 320                 goto error1;
 321         ASSERT(done);
 322         /*
 323          * Commit the first transaction.  This logs the EFI and the inode.
 324          */
 325         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
 326                 goto error1;
 327         /*
 328          * The transaction must have been committed, since there were
 329          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
 330          * The new tp has the extent freeing and EFDs.
 331          */
 332         ASSERT(committed);
 333         /*
 334          * The first xact was committed, so add the inode to the new one.
 335          * Mark it dirty so it will be logged and moved forward in the log as
 336          * part of every commit.
 337          */
 338         xfs_trans_ijoin(tp, ip, 0);
 339         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 340         /*
 341          * Get a new, empty transaction to return to our caller.
 342          */
 343         ntp = xfs_trans_dup(tp);
 344         /*
 345          * Commit the transaction containing extent freeing and EFDs.
 346          * If we get an error on the commit here or on the reserve below,
 347          * we need to unlock the inode since the new transaction doesn't
 348          * have the inode attached.
 349          */
 350         error = xfs_trans_commit(tp, 0);
 351         tp = ntp;
 352         if (error) {
 353                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 354                 goto error0;
 355         }
 356         /*
 357          * transaction commit worked ok so we can drop the extra ticket
 358          * reference that we gained in xfs_trans_dup()
 359          */
 360         xfs_log_ticket_put(tp->t_ticket);
 361
 362         /*
 363          * Remove the memory for extent descriptions (just bookkeeping).
 364          */
 365         if (ip->i_df.if_bytes)
 366                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
 367         ASSERT(ip->i_df.if_bytes == 0);
 368         /*
 369          * Put an itruncate log reservation in the new transaction
 370          * for our caller.
 371          */
 372         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
 373                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
 374                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 375                 goto error0;
 376         }
 377         /*
 378          * Return with the inode locked but not joined to the transaction.
 379          */
 380         *tpp = tp;
 381         return 0;
 382
 383  error1:
 384         xfs_bmap_cancel(&free_list);
 385  error0:
 386         /*
 387          * Have to come here with the inode locked and either
 388          * (held and in the transaction) or (not in the transaction).
 389          * If the inode isn't held then cancel would iput it, but
 390          * that's wrong since this is inactive and the vnode ref
 391          * count is 0 already.
 392          * Cancel won't do anything to the inode if held, but it still
 393          * needs to be locked until the cancel is done, if it was
 394          * joined to the transaction.
 395          */
 396         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 397         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
 398         *tpp = NULL;
 399         return error;
 400
 401 }
 402
 403 STATIC int
 404 xfs_inactive_symlink_local(
 405         xfs_inode_t     *ip,
 406         xfs_trans_t     **tpp)
 407 {
 408         int             error;
 409
 410         ASSERT(ip->i_d.di_size <= XFS_IFORK_DSIZE(ip));
 411         /*
 412          * We're freeing a symlink which fit into
 413          * the inode.  Just free the memory used
 414          * to hold the old symlink.
 415          */
 416         error = xfs_trans_reserve(*tpp, 0,
 417                                   XFS_ITRUNCATE_LOG_RES(ip->i_mount),
 418                                   0, XFS_TRANS_PERM_LOG_RES,
 419                                   XFS_ITRUNCATE_LOG_COUNT);
 420
 421         if (error) {
 422                 xfs_trans_cancel(*tpp, 0);
 423                 *tpp = NULL;
 424                 return error;
 425         }
 426         xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
 427
 428         /*
 429          * Zero length symlinks _can_ exist.
 430          */
 431         if (ip->i_df.if_bytes > 0) {
 432                 xfs_idata_realloc(ip,
 433                                   -(ip->i_df.if_bytes),
 434                                   XFS_DATA_FORK);
 435                 ASSERT(ip->i_df.if_bytes == 0);
 436         }
 437         return 0;
 438 }
 439
 440 STATIC int
 441 xfs_inactive_attrs(
 442         xfs_inode_t     *ip,
 443         xfs_trans_t     **tpp)
 444 {
 445         xfs_trans_t     *tp;
 446         int             error;
 447         xfs_mount_t     *mp;
 448
 449         ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
 450         tp = *tpp;
 451         mp = ip->i_mount;
 452         ASSERT(ip->i_d.di_forkoff != 0);
 453         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 454         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 455         if (error)
 456                 goto error_unlock;
 457
 458         error = xfs_attr_inactive(ip);
 459         if (error)
 460                 goto error_unlock;
 461
 462         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 463         error = xfs_trans_reserve(tp, 0,
 464                                   XFS_IFREE_LOG_RES(mp),
 465                                   0, XFS_TRANS_PERM_LOG_RES,
 466                                   XFS_INACTIVE_LOG_COUNT);
 467         if (error)
 468                 goto error_cancel;
 469
 470         xfs_ilock(ip, XFS_ILOCK_EXCL);
 471         xfs_trans_ijoin(tp, ip, 0);
 472         xfs_idestroy_fork(ip, XFS_ATTR_FORK);
 473
 474         ASSERT(ip->i_d.di_anextents == 0);
 475
 476         *tpp = tp;
 477         return 0;
 478
 479 error_cancel:
 480         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 481         xfs_trans_cancel(tp, 0);
 482 error_unlock:
 483         *tpp = NULL;
 484         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 485         return error;
 486 }
 487
 488 int
 489 xfs_release(
 490         xfs_inode_t     *ip)
 491 {
 492         xfs_mount_t     *mp = ip->i_mount;
 493         int             error;
 494
 495         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
 496                 return 0;
 497
 498         /* If this is a read-only mount, don't do this (would generate I/O) */
 499         if (mp->m_flags & XFS_MOUNT_RDONLY)
 500                 return 0;
 501
 502         if (!XFS_FORCED_SHUTDOWN(mp)) {
 503                 int truncated;
 504
 505                 /*
 506                  * If we are using filestreams, and we have an unlinked
 507                  * file that we are processing the last close on, then nothing
 508                  * will be able to reopen and write to this file. Purge this
 509                  * inode from the filestreams cache so that it doesn't delay
 510                  * teardown of the inode.
 511                  */
 512                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
 513                         xfs_filestream_deassociate(ip);
 514
 515                 /*
 516                  * If we previously truncated this file and removed old data
 517                  * in the process, we want to initiate "early" writeout on
 518                  * the last close.  This is an attempt to combat the notorious
 519                  * NULL files problem which is particularly noticeable from a
 520                  * truncate down, buffered (re-)write (delalloc), followed by
 521                  * a crash.  What we are effectively doing here is
 522                  * significantly reducing the time window where we'd otherwise
 523                  * be exposed to that problem.
 524                  */
 525                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
 526                 if (truncated) {
 527                         xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
 528                         if (VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
 529                                 xfs_flush_pages(ip, 0, -1, XBF_ASYNC, FI_NONE);
 530                 }
 531         }
 532
 533         if (ip->i_d.di_nlink == 0)
 534                 return 0;
 535
 536         if ((S_ISREG(ip->i_d.di_mode) &&
 537              ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
 538                ip->i_delayed_blks > 0)) &&
 539              (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
 540             (!(ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
 541
 542                 /*
 543                  * If we can't get the iolock just skip truncating the blocks
 544                  * past EOF because we could deadlock with the mmap_sem
 545                  * otherwise.  We'll get another chance to drop them once the
 546                  * last reference to the inode is dropped, so we'll never leak
 547                  * blocks permanently.
 548                  *
 549                  * Further, check if the inode is being opened, written and
 550                  * closed frequently and we have delayed allocation blocks
 551                  * outstanding (e.g. streaming writes from the NFS server),
 552                  * truncating the blocks past EOF will cause fragmentation to
 553                  * occur.
 554                  *
 555                  * In this case don't do the truncation, either, but we have to
 556                  * be careful how we detect this case. Blocks beyond EOF show
 557                  * up as i_delayed_blks even when the inode is clean, so we
 558                  * need to truncate them away first before checking for a dirty
 559                  * release. Hence on the first dirty close we will still remove
 560                  * the speculative allocation, but after that we will leave it
 561                  * in place.
 562                  */
 563                 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
 564                         return 0;
 565
 566                 error = xfs_free_eofblocks(mp, ip,
 567                                            XFS_FREE_EOF_TRYLOCK);
 568                 if (error)
 569                         return error;
 570
 571                 /* delalloc blocks after truncation means it really is dirty */
 572                 if (ip->i_delayed_blks)
 573                         xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
 574         }
 575         return 0;
 576 }
 577
 578 /*
 579  * xfs_inactive
 580  *
 581  * This is called when the vnode reference count for the vnode
 582  * goes to zero.  If the file has been unlinked, then it must
 583  * now be truncated.  Also, we clear all of the read-ahead state
 584  * kept for the inode here since the file is now closed.
 585  */
 586 int
 587 xfs_inactive(
 588         xfs_inode_t     *ip)
 589 {
 590         xfs_bmap_free_t free_list;
 591         xfs_fsblock_t   first_block;
 592         int             committed;
 593         xfs_trans_t     *tp;
 594         xfs_mount_t     *mp;
 595         int             error;
 596         int             truncate;
 597
 598         /*
 599          * If the inode is already free, then there can be nothing
 600          * to clean up here.
 601          */
 602         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
 603                 ASSERT(ip->i_df.if_real_bytes == 0);
 604                 ASSERT(ip->i_df.if_broot_bytes == 0);
 605                 return VN_INACTIVE_CACHE;
 606         }
 607
 608         /*
 609          * Only do a truncate if it's a regular file with
 610          * some actual space in it.  It's OK to look at the
 611          * inode's fields without the lock because we're the
 612          * only one with a reference to the inode.
 613          */
 614         truncate = ((ip->i_d.di_nlink == 0) &&
 615             ((ip->i_d.di_size != 0) || (ip->i_size != 0) ||
 616              (ip->i_d.di_nextents > 0) || (ip->i_delayed_blks > 0)) &&
 617             S_ISREG(ip->i_d.di_mode));
 618
 619         mp = ip->i_mount;
 620
 621         error = 0;
 622
 623         /* If this is a read-only mount, don't do this (would generate I/O) */
 624         if (mp->m_flags & XFS_MOUNT_RDONLY)
 625                 goto out;
 626
 627         if (ip->i_d.di_nlink != 0) {
 628                 if ((S_ISREG(ip->i_d.di_mode) &&
 629                      ((ip->i_size > 0) || (VN_CACHED(VFS_I(ip)) > 0 ||
 630                        ip->i_delayed_blks > 0)) &&
 631                       (ip->i_df.if_flags & XFS_IFEXTENTS) &&
 632                      (!(ip->i_d.di_flags &
 633                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
 634                       (ip->i_delayed_blks != 0)))) {
 635                         error = xfs_free_eofblocks(mp, ip, 0);
 636                         if (error)
 637                                 return VN_INACTIVE_CACHE;
 638                 }
 639                 goto out;
 640         }
 641
 642         ASSERT(ip->i_d.di_nlink == 0);
 643
 644         error = xfs_qm_dqattach(ip, 0);
 645         if (error)
 646                 return VN_INACTIVE_CACHE;
 647
 648         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 649         if (truncate) {
 650                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
 651
 652                 error = xfs_trans_reserve(tp, 0,
 653                                           XFS_ITRUNCATE_LOG_RES(mp),
 654                                           0, XFS_TRANS_PERM_LOG_RES,
 655                                           XFS_ITRUNCATE_LOG_COUNT);
 656                 if (error) {
 657                         /* Don't call itruncate_cleanup */
 658                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 659                         xfs_trans_cancel(tp, 0);
 660                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 661                         return VN_INACTIVE_CACHE;
 662                 }
 663
 664                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 665                 xfs_trans_ijoin(tp, ip, 0);
 666
 667                 error = xfs_itruncate_data(&tp, ip, 0);
 668                 if (error) {
 669                         xfs_trans_cancel(tp,
 670                                 XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 671                         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
 672                         return VN_INACTIVE_CACHE;
 673                 }
 674         } else if (S_ISLNK(ip->i_d.di_mode)) {
 675
 676                 /*
 677                  * If we get an error while cleaning up a
 678                  * symlink we bail out.
 679                  */
 680                 error = (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) ?
 681                         xfs_inactive_symlink_rmt(ip, &tp) :
 682                         xfs_inactive_symlink_local(ip, &tp);
 683
 684                 if (error) {
 685                         ASSERT(tp == NULL);
 686                         return VN_INACTIVE_CACHE;
 687                 }
 688
 689                 xfs_trans_ijoin(tp, ip, 0);
 690         } else {
 691                 error = xfs_trans_reserve(tp, 0,
 692                                           XFS_IFREE_LOG_RES(mp),
 693                                           0, XFS_TRANS_PERM_LOG_RES,
 694                                           XFS_INACTIVE_LOG_COUNT);
 695                 if (error) {
 696                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 697                         xfs_trans_cancel(tp, 0);
 698                         return VN_INACTIVE_CACHE;
 699                 }
 700
 701                 xfs_ilock(ip, XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL);
 702                 xfs_trans_ijoin(tp, ip, 0);
 703         }
 704
 705         /*
 706          * If there are attributes associated with the file
 707          * then blow them away now.  The code calls a routine
 708          * that recursively deconstructs the attribute fork.
 709          * We need to just commit the current transaction
 710          * because we can't use it for xfs_attr_inactive().
 711          */
 712         if (ip->i_d.di_anextents > 0) {
 713                 error = xfs_inactive_attrs(ip, &tp);
 714                 /*
 715                  * If we got an error, the transaction is already
 716                  * cancelled, and the inode is unlocked. Just get out.
 717                  */
 718                  if (error)
 719                          return VN_INACTIVE_CACHE;
 720         } else if (ip->i_afp) {
 721                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
 722         }
 723
 724         /*
 725          * Free the inode.
 726          */
 727         xfs_bmap_init(&free_list, &first_block);
 728         error = xfs_ifree(tp, ip, &free_list);
 729         if (error) {
 730                 /*
 731                  * If we fail to free the inode, shut down.  The cancel
 732                  * might do that, we need to make sure.  Otherwise the
 733                  * inode might be lost for a long time or forever.
 734                  */
 735                 if (!XFS_FORCED_SHUTDOWN(mp)) {
 736                         xfs_notice(mp, "%s: xfs_ifree returned error %d",
 737                                 __func__, error);
 738                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
 739                 }
 740                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
 741         } else {
 742                 /*
 743                  * Credit the quota account(s). The inode is gone.
 744                  */
 745                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
 746
 747                 /*
 748                  * Just ignore errors at this point.  There is nothing we can
 749                  * do except to try to keep going. Make sure it's not a silent
 750                  * error.
 751                  */
 752                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
 753                 if (error)
 754                         xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
 755                                 __func__, error);
 756                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 757                 if (error)
 758                         xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
 759                                 __func__, error);
 760         }
 761
 762         /*
 763          * Release the dquots held by inode, if any.
 764          */
 765         xfs_qm_dqdetach(ip);
 766         xfs_iunlock(ip, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
 767
 768  out:
 769         return VN_INACTIVE_CACHE;
 770 }
 771
 772 /*
 773  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
 774  * is allowed, otherwise it has to be an exact match. If a CI match is found,
 775  * ci_name->name will point to a the actual name (caller must free) or
 776  * will be set to NULL if an exact match is found.
 777  */
 778 int
 779 xfs_lookup(
 780         xfs_inode_t             *dp,
 781         struct xfs_name         *name,
 782         xfs_inode_t             **ipp,
 783         struct xfs_name         *ci_name)
 784 {
 785         xfs_ino_t               inum;
 786         int                     error;
 787         uint                    lock_mode;
 788
 789         trace_xfs_lookup(dp, name);
 790
 791         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
 792                 return XFS_ERROR(EIO);
 793
 794         lock_mode = xfs_ilock_map_shared(dp);
 795         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
 796         xfs_iunlock_map_shared(dp, lock_mode);
 797
 798         if (error)
 799                 goto out;
 800
 801         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
 802         if (error)
 803                 goto out_free_name;
 804
 805         return 0;
 806
 807 out_free_name:
 808         if (ci_name)
 809                 kmem_free(ci_name->name);
 810 out:
 811         *ipp = NULL;
 812         return error;
 813 }
 814
 815 int
 816 xfs_create(
 817         xfs_inode_t             *dp,
 818         struct xfs_name         *name,
 819         mode_t                  mode,
 820         xfs_dev_t               rdev,
 821         xfs_inode_t             **ipp)
 822 {
 823         int                     is_dir = S_ISDIR(mode);
 824         struct xfs_mount        *mp = dp->i_mount;
 825         struct xfs_inode        *ip = NULL;
 826         struct xfs_trans        *tp = NULL;
 827         int                     error;
 828         xfs_bmap_free_t         free_list;
 829         xfs_fsblock_t           first_block;
 830         boolean_t               unlock_dp_on_error = B_FALSE;
 831         uint                    cancel_flags;
 832         int                     committed;
 833         prid_t                  prid;
 834         struct xfs_dquot        *udqp = NULL;
 835         struct xfs_dquot        *gdqp = NULL;
 836         uint                    resblks;
 837         uint                    log_res;
 838         uint                    log_count;
 839
 840         trace_xfs_create(dp, name);
 841
 842         if (XFS_FORCED_SHUTDOWN(mp))
 843                 return XFS_ERROR(EIO);
 844
 845         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
 846                 prid = xfs_get_projid(dp);
 847         else
 848                 prid = XFS_PROJID_DEFAULT;
 849
 850         /*
 851          * Make sure that we have allocated dquot(s) on disk.
 852          */
 853         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
 854                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
 855         if (error)
 856                 return error;
 857
 858         if (is_dir) {
 859                 rdev = 0;
 860                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
 861                 log_res = XFS_MKDIR_LOG_RES(mp);
 862                 log_count = XFS_MKDIR_LOG_COUNT;
 863                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
 864         } else {
 865                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
 866                 log_res = XFS_CREATE_LOG_RES(mp);
 867                 log_count = XFS_CREATE_LOG_COUNT;
 868                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
 869         }
 870
 871         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
 872
 873         /*
 874          * Initially assume that the file does not exist and
 875          * reserve the resources for that case.  If that is not
 876          * the case we'll drop the one we have and get a more
 877          * appropriate transaction later.
 878          */
 879         error = xfs_trans_reserve(tp, resblks, log_res, 0,
 880                         XFS_TRANS_PERM_LOG_RES, log_count);
 881         if (error == ENOSPC) {
 882                 /* flush outstanding delalloc blocks and retry */
 883                 xfs_flush_inodes(dp);
 884                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
 885                                 XFS_TRANS_PERM_LOG_RES, log_count);
 886         }
 887         if (error == ENOSPC) {
 888                 /* No space at all so try a "no-allocation" reservation */
 889                 resblks = 0;
 890                 error = xfs_trans_reserve(tp, 0, log_res, 0,
 891                                 XFS_TRANS_PERM_LOG_RES, log_count);
 892         }
 893         if (error) {
 894                 cancel_flags = 0;
 895                 goto out_trans_cancel;
 896         }
 897
 898         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
 899         unlock_dp_on_error = B_TRUE;
 900
 901         /*
 902          * Check for directory link count overflow.
 903          */
 904         if (is_dir && dp->i_d.di_nlink >= XFS_MAXLINK) {
 905                 error = XFS_ERROR(EMLINK);
 906                 goto out_trans_cancel;
 907         }
 908
 909         xfs_bmap_init(&free_list, &first_block);
 910
 911         /*
 912          * Reserve disk quota and the inode.
 913          */
 914         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
 915         if (error)
 916                 goto out_trans_cancel;
 917
 918         error = xfs_dir_canenter(tp, dp, name, resblks);
 919         if (error)
 920                 goto out_trans_cancel;
 921
 922         /*
 923          * A newly created regular or special file just has one directory
 924          * entry pointing to them, but a directory also the "." entry
 925          * pointing to itself.
 926          */
 927         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
 928                                prid, resblks > 0, &ip, &committed);
 929         if (error) {
 930                 if (error == ENOSPC)
 931                         goto out_trans_cancel;
 932                 goto out_trans_abort;
 933         }
 934
 935         /*
 936          * Now we join the directory inode to the transaction.  We do not do it
 937          * earlier because xfs_dir_ialloc might commit the previous transaction
 938          * (and release all the locks).  An error from here on will result in
 939          * the transaction cancel unlocking dp so don't do it explicitly in the
 940          * error path.
 941          */
 942         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
 943         unlock_dp_on_error = B_FALSE;
 944
 945         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
 946                                         &first_block, &free_list, resblks ?
 947                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
 948         if (error) {
 949                 ASSERT(error != ENOSPC);
 950                 goto out_trans_abort;
 951         }
 952         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 953         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
 954
 955         if (is_dir) {
 956                 error = xfs_dir_init(tp, ip, dp);
 957                 if (error)
 958                         goto out_bmap_cancel;
 959
 960                 error = xfs_bumplink(tp, dp);
 961                 if (error)
 962                         goto out_bmap_cancel;
 963         }
 964
 965         /*
 966          * If this is a synchronous mount, make sure that the
 967          * create transaction goes to disk before returning to
 968          * the user.
 969          */
 970         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
 971                 xfs_trans_set_sync(tp);
 972
 973         /*
 974          * Attach the dquot(s) to the inodes and modify them incore.
 975          * These ids of the inode couldn't have changed since the new
 976          * inode has been locked ever since it was created.
 977          */
 978         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
 979
 980         error = xfs_bmap_finish(&tp, &free_list, &committed);
 981         if (error)
 982                 goto out_bmap_cancel;
 983
 984         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 985         if (error)
 986                 goto out_release_inode;
 987
 988         xfs_qm_dqrele(udqp);
 989         xfs_qm_dqrele(gdqp);
 990
 991         *ipp = ip;
 992         return 0;
 993
 994  out_bmap_cancel:
 995         xfs_bmap_cancel(&free_list);
 996  out_trans_abort:
 997         cancel_flags |= XFS_TRANS_ABORT;
 998  out_trans_cancel:
 999         xfs_trans_cancel(tp, cancel_flags);
1000  out_release_inode:
1001         /*
1002          * Wait until after the current transaction is aborted to
1003          * release the inode.  This prevents recursive transactions
1004          * and deadlocks from xfs_inactive.
1005          */
1006         if (ip)
1007                 IRELE(ip);
1008
1009         xfs_qm_dqrele(udqp);
1010         xfs_qm_dqrele(gdqp);
1011
1012         if (unlock_dp_on_error)
1013                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1014         return error;
1015 }
1016
1017 #ifdef DEBUG
1018 int xfs_locked_n;
1019 int xfs_small_retries;
1020 int xfs_middle_retries;
1021 int xfs_lots_retries;
1022 int xfs_lock_delays;
1023 #endif
1024
1025 /*
1026  * Bump the subclass so xfs_lock_inodes() acquires each lock with
1027  * a different value
1028  */
1029 static inline int
1030 xfs_lock_inumorder(int lock_mode, int subclass)
1031 {
1032         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1033                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
1034         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
1035                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
1036
1037         return lock_mode;
1038 }
1039
1040 /*
1041  * The following routine will lock n inodes in exclusive mode.
1042  * We assume the caller calls us with the inodes in i_ino order.
1043  *
1044  * We need to detect deadlock where an inode that we lock
1045  * is in the AIL and we start waiting for another inode that is locked
1046  * by a thread in a long running transaction (such as truncate). This can
1047  * result in deadlock since the long running trans might need to wait
1048  * for the inode we just locked in order to push the tail and free space
1049  * in the log.
1050  */
1051 void
1052 xfs_lock_inodes(
1053         xfs_inode_t     **ips,
1054         int             inodes,
1055         uint            lock_mode)
1056 {
1057         int             attempts = 0, i, j, try_lock;
1058         xfs_log_item_t  *lp;
1059
1060         ASSERT(ips && (inodes >= 2)); /* we need at least two */
1061
1062         try_lock = 0;
1063         i = 0;
1064
1065 again:
1066         for (; i < inodes; i++) {
1067                 ASSERT(ips[i]);
1068
1069                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
1070                         continue;
1071
1072                 /*
1073                  * If try_lock is not set yet, make sure all locked inodes
1074                  * are not in the AIL.
1075                  * If any are, set try_lock to be used later.
1076                  */
1077
1078                 if (!try_lock) {
1079                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
1080                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
1081                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1082                                         try_lock++;
1083                                 }
1084                         }
1085                 }
1086
1087                 /*
1088                  * If any of the previous locks we have locked is in the AIL,
1089                  * we must TRY to get the second and subsequent locks. If
1090                  * we can't get any, we must release all we have
1091                  * and try again.
1092                  */
1093
1094                 if (try_lock) {
1095                         /* try_lock must be 0 if i is 0. */
1096                         /*
1097                          * try_lock means we have an inode locked
1098                          * that is in the AIL.
1099                          */
1100                         ASSERT(i != 0);
1101                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
1102                                 attempts++;
1103
1104                                 /*
1105                                  * Unlock all previous guys and try again.
1106                                  * xfs_iunlock will try to push the tail
1107                                  * if the inode is in the AIL.
1108                                  */
1109
1110                                 for(j = i - 1; j >= 0; j--) {
1111
1112                                         /*
1113                                          * Check to see if we've already
1114                                          * unlocked this one.
1115                                          * Not the first one going back,
1116                                          * and the inode ptr is the same.
1117                                          */
1118                                         if ((j != (i - 1)) && ips[j] ==
1119                                                                 ips[j+1])
1120                                                 continue;
1121
1122                                         xfs_iunlock(ips[j], lock_mode);
1123                                 }
1124
1125                                 if ((attempts % 5) == 0) {
1126                                         delay(1); /* Don't just spin the CPU */
1127 #ifdef DEBUG
1128                                         xfs_lock_delays++;
1129 #endif
1130                                 }
1131                                 i = 0;
1132                                 try_lock = 0;
1133                                 goto again;
1134                         }
1135                 } else {
1136                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1137                 }
1138         }
1139
1140 #ifdef DEBUG
1141         if (attempts) {
1142                 if (attempts < 5) xfs_small_retries++;
1143                 else if (attempts < 100) xfs_middle_retries++;
1144                 else xfs_lots_retries++;
1145         } else {
1146                 xfs_locked_n++;
1147         }
1148 #endif
1149 }
1150
1151 /*
1152  * xfs_lock_two_inodes() can only be used to lock one type of lock
1153  * at a time - the iolock or the ilock, but not both at once. If
1154  * we lock both at once, lockdep will report false positives saying
1155  * we have violated locking orders.
1156  */
1157 void
1158 xfs_lock_two_inodes(
1159         xfs_inode_t             *ip0,
1160         xfs_inode_t             *ip1,
1161         uint                    lock_mode)
1162 {
1163         xfs_inode_t             *temp;
1164         int                     attempts = 0;
1165         xfs_log_item_t          *lp;
1166
1167         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1168                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1169         ASSERT(ip0->i_ino != ip1->i_ino);
1170
1171         if (ip0->i_ino > ip1->i_ino) {
1172                 temp = ip0;
1173                 ip0 = ip1;
1174                 ip1 = temp;
1175         }
1176
1177  again:
1178         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1179
1180         /*
1181          * If the first lock we have locked is in the AIL, we must TRY to get
1182          * the second lock. If we can't get it, we must release the first one
1183          * and try again.
1184          */
1185         lp = (xfs_log_item_t *)ip0->i_itemp;
1186         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1187                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1188                         xfs_iunlock(ip0, lock_mode);
1189                         if ((++attempts % 5) == 0)
1190                                 delay(1); /* Don't just spin the CPU */
1191                         goto again;
1192                 }
1193         } else {
1194                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1195         }
1196 }
1197
1198 int
1199 xfs_remove(
1200         xfs_inode_t             *dp,
1201         struct xfs_name         *name,
1202         xfs_inode_t             *ip)
1203 {
1204         xfs_mount_t             *mp = dp->i_mount;
1205         xfs_trans_t             *tp = NULL;
1206         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1207         int                     error = 0;
1208         xfs_bmap_free_t         free_list;
1209         xfs_fsblock_t           first_block;
1210         int                     cancel_flags;
1211         int                     committed;
1212         int                     link_zero;
1213         uint                    resblks;
1214         uint                    log_count;
1215
1216         trace_xfs_remove(dp, name);
1217
1218         if (XFS_FORCED_SHUTDOWN(mp))
1219                 return XFS_ERROR(EIO);
1220
1221         error = xfs_qm_dqattach(dp, 0);
1222         if (error)
1223                 goto std_return;
1224
1225         error = xfs_qm_dqattach(ip, 0);
1226         if (error)
1227                 goto std_return;
1228
1229         if (is_dir) {
1230                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1231                 log_count = XFS_DEFAULT_LOG_COUNT;
1232         } else {
1233                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1234                 log_count = XFS_REMOVE_LOG_COUNT;
1235         }
1236         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1237
1238         /*
1239          * We try to get the real space reservation first,
1240          * allowing for directory btree deletion(s) implying
1241          * possible bmap insert(s).  If we can't get the space
1242          * reservation then we use 0 instead, and avoid the bmap
1243          * btree insert(s) in the directory code by, if the bmap
1244          * insert tries to happen, instead trimming the LAST
1245          * block from the directory.
1246          */
1247         resblks = XFS_REMOVE_SPACE_RES(mp);
1248         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1249                                   XFS_TRANS_PERM_LOG_RES, log_count);
1250         if (error == ENOSPC) {
1251                 resblks = 0;
1252                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1253                                           XFS_TRANS_PERM_LOG_RES, log_count);
1254         }
1255         if (error) {
1256                 ASSERT(error != ENOSPC);
1257                 cancel_flags = 0;
1258                 goto out_trans_cancel;
1259         }
1260
1261         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1262
1263         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1264         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1265
1266         /*
1267          * If we're removing a directory perform some additional validation.
1268          */
1269         if (is_dir) {
1270                 ASSERT(ip->i_d.di_nlink >= 2);
1271                 if (ip->i_d.di_nlink != 2) {
1272                         error = XFS_ERROR(ENOTEMPTY);
1273                         goto out_trans_cancel;
1274                 }
1275                 if (!xfs_dir_isempty(ip)) {
1276                         error = XFS_ERROR(ENOTEMPTY);
1277                         goto out_trans_cancel;
1278                 }
1279         }
1280
1281         xfs_bmap_init(&free_list, &first_block);
1282         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1283                                         &first_block, &free_list, resblks);
1284         if (error) {
1285                 ASSERT(error != ENOENT);
1286                 goto out_bmap_cancel;
1287         }
1288         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1289
1290         if (is_dir) {
1291                 /*
1292                  * Drop the link from ip's "..".
1293                  */
1294                 error = xfs_droplink(tp, dp);
1295                 if (error)
1296                         goto out_bmap_cancel;
1297
1298                 /*
1299                  * Drop the "." link from ip to self.
1300                  */
1301                 error = xfs_droplink(tp, ip);
1302                 if (error)
1303                         goto out_bmap_cancel;
1304         } else {
1305                 /*
1306                  * When removing a non-directory we need to log the parent
1307                  * inode here.  For a directory this is done implicitly
1308                  * by the xfs_droplink call for the ".." entry.
1309                  */
1310                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1311         }
1312
1313         /*
1314          * Drop the link from dp to ip.
1315          */
1316         error = xfs_droplink(tp, ip);
1317         if (error)
1318                 goto out_bmap_cancel;
1319
1320         /*
1321          * Determine if this is the last link while
1322          * we are in the transaction.
1323          */
1324         link_zero = (ip->i_d.di_nlink == 0);
1325
1326         /*
1327          * If this is a synchronous mount, make sure that the
1328          * remove transaction goes to disk before returning to
1329          * the user.
1330          */
1331         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1332                 xfs_trans_set_sync(tp);
1333
1334         error = xfs_bmap_finish(&tp, &free_list, &committed);
1335         if (error)
1336                 goto out_bmap_cancel;
1337
1338         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1339         if (error)
1340                 goto std_return;
1341
1342         /*
1343          * If we are using filestreams, kill the stream association.
1344          * If the file is still open it may get a new one but that
1345          * will get killed on last close in xfs_close() so we don't
1346          * have to worry about that.
1347          */
1348         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1349                 xfs_filestream_deassociate(ip);
1350
1351         return 0;
1352
1353  out_bmap_cancel:
1354         xfs_bmap_cancel(&free_list);
1355         cancel_flags |= XFS_TRANS_ABORT;
1356  out_trans_cancel:
1357         xfs_trans_cancel(tp, cancel_flags);
1358  std_return:
1359         return error;
1360 }
1361
1362 int
1363 xfs_link(
1364         xfs_inode_t             *tdp,
1365         xfs_inode_t             *sip,
1366         struct xfs_name         *target_name)
1367 {
1368         xfs_mount_t             *mp = tdp->i_mount;
1369         xfs_trans_t             *tp;
1370         int                     error;
1371         xfs_bmap_free_t         free_list;
1372         xfs_fsblock_t           first_block;
1373         int                     cancel_flags;
1374         int                     committed;
1375         int                     resblks;
1376
1377         trace_xfs_link(tdp, target_name);
1378
1379         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1380
1381         if (XFS_FORCED_SHUTDOWN(mp))
1382                 return XFS_ERROR(EIO);
1383
1384         error = xfs_qm_dqattach(sip, 0);
1385         if (error)
1386                 goto std_return;
1387
1388         error = xfs_qm_dqattach(tdp, 0);
1389         if (error)
1390                 goto std_return;
1391
1392         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1393         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1394         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1395         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1396                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1397         if (error == ENOSPC) {
1398                 resblks = 0;
1399                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1400                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1401         }
1402         if (error) {
1403                 cancel_flags = 0;
1404                 goto error_return;
1405         }
1406
1407         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1408
1409         xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1410         xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1411
1412         /*
1413          * If the source has too many links, we can't make any more to it.
1414          */
1415         if (sip->i_d.di_nlink >= XFS_MAXLINK) {
1416                 error = XFS_ERROR(EMLINK);
1417                 goto error_return;
1418         }
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         mode_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, 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 >= ip->i_size)
1959                 return 0;
1960
1961         if (endoff > ip->i_size)
1962                 endoff = ip->i_size;
1963
1964         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
1965                                         mp->m_rtdev_targp : mp->m_ddev_targp,
1966                                 mp->m_sb.sb_blocksize, XBF_DONT_BLOCK);
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 += ip->i_size;
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 = ip->i_size;
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                 if (startoffset > fsize) {
2316                         error = xfs_alloc_file_space(ip, fsize,
2317                                         startoffset - fsize, 0, attr_flags);
2318                         if (error)
2319                                 break;
2320                 }
2321
2322                 iattr.ia_valid = ATTR_SIZE;
2323                 iattr.ia_size = startoffset;
2324
2325                 error = xfs_setattr_size(ip, &iattr, attr_flags);
2326
2327                 if (error)
2328                         return error;
2329
2330                 clrprealloc = 1;
2331                 break;
2332
2333         default:
2334                 ASSERT(0);
2335                 return XFS_ERROR(EINVAL);
2336         }
2337
2338         /*
2339          * update the inode timestamp, mode, and prealloc flag bits
2340          */
2341         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
2342
2343         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
2344                                       0, 0, 0))) {
2345                 /* ASSERT(0); */
2346                 xfs_trans_cancel(tp, 0);
2347                 return error;
2348         }
2349
2350         xfs_ilock(ip, XFS_ILOCK_EXCL);
2351         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2352
2353         if ((attr_flags & XFS_ATTR_DMI) == 0) {
2354                 ip->i_d.di_mode &= ~S_ISUID;
2355
2356                 /*
2357                  * Note that we don't have to worry about mandatory
2358                  * file locking being disabled here because we only
2359                  * clear the S_ISGID bit if the Group execute bit is
2360                  * on, but if it was on then mandatory locking wouldn't
2361                  * have been enabled.
2362                  */
2363                 if (ip->i_d.di_mode & S_IXGRP)
2364                         ip->i_d.di_mode &= ~S_ISGID;
2365
2366                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2367         }
2368         if (setprealloc)
2369                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
2370         else if (clrprealloc)
2371                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
2372
2373         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2374         if (attr_flags & XFS_ATTR_SYNC)
2375                 xfs_trans_set_sync(tp);
2376         return xfs_trans_commit(tp, 0);
2377 }