fs/xfs/xfs_vnodeops.c

   1 /*
   2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18
  19 #include "xfs.h"
  20 #include "xfs_fs.h"
  21 #include "xfs_types.h"
  22 #include "xfs_bit.h"
  23 #include "xfs_log.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_mount.h"
  29 #include "xfs_da_btree.h"
  30 #include "xfs_bmap_btree.h"
  31 #include "xfs_ialloc_btree.h"
  32 #include "xfs_dinode.h"
  33 #include "xfs_inode.h"
  34 #include "xfs_inode_item.h"
  35 #include "xfs_itable.h"
  36 #include "xfs_ialloc.h"
  37 #include "xfs_alloc.h"
  38 #include "xfs_bmap.h"
  39 #include "xfs_acl.h"
  40 #include "xfs_attr.h"
  41 #include "xfs_error.h"
  42 #include "xfs_quota.h"
  43 #include "xfs_utils.h"
  44 #include "xfs_rtalloc.h"
  45 #include "xfs_trans_space.h"
  46 #include "xfs_log_priv.h"
  47 #include "xfs_filestream.h"
  48 #include "xfs_vnodeops.h"
  49 #include "xfs_trace.h"
  50 #include "xfs_icache.h"
  51
  52 /*
  53  * The maximum pathlen is 1024 bytes. Since the minimum file system
  54  * blocksize is 512 bytes, we can get a max of 2 extents back from
  55  * bmapi.
  56  */
  57 #define SYMLINK_MAPS 2
  58
  59 STATIC int
  60 xfs_readlink_bmap(
  61         xfs_inode_t     *ip,
  62         char            *link)
  63 {
  64         xfs_mount_t     *mp = ip->i_mount;
  65         int             pathlen = ip->i_d.di_size;
  66         int             nmaps = SYMLINK_MAPS;
  67         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
  68         xfs_daddr_t     d;
  69         int             byte_cnt;
  70         int             n;
  71         xfs_buf_t       *bp;
  72         int             error = 0;
  73
  74         error = xfs_bmapi_read(ip, 0, XFS_B_TO_FSB(mp, pathlen), mval, &nmaps,
  75                                0);
  76         if (error)
  77                 goto out;
  78
  79         for (n = 0; n < nmaps; n++) {
  80                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
  81                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
  82
  83                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt), 0);
  84                 if (!bp)
  85                         return XFS_ERROR(ENOMEM);
  86                 error = bp->b_error;
  87                 if (error) {
  88                         xfs_buf_ioerror_alert(bp, __func__);
  89                         xfs_buf_relse(bp);
  90                         goto out;
  91                 }
  92                 if (pathlen < byte_cnt)
  93                         byte_cnt = pathlen;
  94                 pathlen -= byte_cnt;
  95
  96                 memcpy(link, bp->b_addr, byte_cnt);
  97                 xfs_buf_relse(bp);
  98         }
  99
 100         link[ip->i_d.di_size] = '\0';
 101         error = 0;
 102
 103  out:
 104         return error;
 105 }
 106
 107 int
 108 xfs_readlink(
 109         xfs_inode_t     *ip,
 110         char            *link)
 111 {
 112         xfs_mount_t     *mp = ip->i_mount;
 113         xfs_fsize_t     pathlen;
 114         int             error = 0;
 115
 116         trace_xfs_readlink(ip);
 117
 118         if (XFS_FORCED_SHUTDOWN(mp))
 119                 return XFS_ERROR(EIO);
 120
 121         xfs_ilock(ip, XFS_ILOCK_SHARED);
 122
 123         pathlen = ip->i_d.di_size;
 124         if (!pathlen)
 125                 goto out;
 126
 127         if (pathlen < 0 || pathlen > MAXPATHLEN) {
 128                 xfs_alert(mp, "%s: inode (%llu) bad symlink length (%lld)",
 129                          __func__, (unsigned long long) ip->i_ino,
 130                          (long long) pathlen);
 131                 ASSERT(0);
 132                 error = XFS_ERROR(EFSCORRUPTED);
 133                 goto out;
 134         }
 135
 136
 137         if (ip->i_df.if_flags & XFS_IFINLINE) {
 138                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
 139                 link[pathlen] = '\0';
 140         } else {
 141                 error = xfs_readlink_bmap(ip, link);
 142         }
 143
 144  out:
 145         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 146         return error;
 147 }
 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         bool            need_iolock)
 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)XFS_ISIZE(ip));
 173         last_fsb = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
 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 (need_iolock) {
 201                         if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
 202                                 xfs_trans_cancel(tp, 0);
 203                                 return 0;
 204                         }
 205                 }
 206
 207                 error = xfs_trans_reserve(tp, 0,
 208                                           XFS_ITRUNCATE_LOG_RES(mp),
 209                                           0, XFS_TRANS_PERM_LOG_RES,
 210                                           XFS_ITRUNCATE_LOG_COUNT);
 211                 if (error) {
 212                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 213                         xfs_trans_cancel(tp, 0);
 214                         if (need_iolock)
 215                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 216                         return error;
 217                 }
 218
 219                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 220                 xfs_trans_ijoin(tp, ip, 0);
 221
 222                 /*
 223                  * Do not update the on-disk file size.  If we update the
 224                  * on-disk file size and then the system crashes before the
 225                  * contents of the file are flushed to disk then the files
 226                  * may be full of holes (ie NULL files bug).
 227                  */
 228                 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK,
 229                                               XFS_ISIZE(ip));
 230                 if (error) {
 231                         /*
 232                          * If we get an error at this point we simply don't
 233                          * bother truncating the file.
 234                          */
 235                         xfs_trans_cancel(tp,
 236                                          (XFS_TRANS_RELEASE_LOG_RES |
 237                                           XFS_TRANS_ABORT));
 238                 } else {
 239                         error = xfs_trans_commit(tp,
 240                                                 XFS_TRANS_RELEASE_LOG_RES);
 241                 }
 242
 243                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 244                 if (need_iolock)
 245                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 246         }
 247         return error;
 248 }
 249
 250 /*
 251  * Free a symlink that has blocks associated with it.
 252  */
 253 STATIC int
 254 xfs_inactive_symlink_rmt(
 255         xfs_inode_t     *ip,
 256         xfs_trans_t     **tpp)
 257 {
 258         xfs_buf_t       *bp;
 259         int             committed;
 260         int             done;
 261         int             error;
 262         xfs_fsblock_t   first_block;
 263         xfs_bmap_free_t free_list;
 264         int             i;
 265         xfs_mount_t     *mp;
 266         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
 267         int             nmaps;
 268         xfs_trans_t     *ntp;
 269         int             size;
 270         xfs_trans_t     *tp;
 271
 272         tp = *tpp;
 273         mp = ip->i_mount;
 274         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
 275         /*
 276          * We're freeing a symlink that has some
 277          * blocks allocated to it.  Free the
 278          * blocks here.  We know that we've got
 279          * either 1 or 2 extents and that we can
 280          * free them all in one bunmapi call.
 281          */
 282         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
 283
 284         /*
 285          * Lock the inode, fix the size, and join it to the transaction.
 286          * Hold it so in the normal path, we still have it locked for
 287          * the second transaction.  In the error paths we need it
 288          * held so the cancel won't rele it, see below.
 289          */
 290         size = (int)ip->i_d.di_size;
 291         ip->i_d.di_size = 0;
 292         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 293         /*
 294          * Find the block(s) so we can inval and unmap them.
 295          */
 296         done = 0;
 297         xfs_bmap_init(&free_list, &first_block);
 298         nmaps = ARRAY_SIZE(mval);
 299         error = xfs_bmapi_read(ip, 0, XFS_B_TO_FSB(mp, size),
 300                                 mval, &nmaps, 0);
 301         if (error)
 302                 goto error0;
 303         /*
 304          * Invalidate the block(s).
 305          */
 306         for (i = 0; i < nmaps; i++) {
 307                 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp,
 308                         XFS_FSB_TO_DADDR(mp, mval[i].br_startblock),
 309                         XFS_FSB_TO_BB(mp, mval[i].br_blockcount), 0);
 310                 if (!bp) {
 311                         error = ENOMEM;
 312                         goto error1;
 313                 }
 314                 xfs_trans_binval(tp, bp);
 315         }
 316         /*
 317          * Unmap the dead block(s) to the free_list.
 318          */
 319         if ((error = xfs_bunmapi(tp, ip, 0, size, XFS_BMAPI_METADATA, nmaps,
 320                         &first_block, &free_list, &done)))
 321                 goto error1;
 322         ASSERT(done);
 323         /*
 324          * Commit the first transaction.  This logs the EFI and the inode.
 325          */
 326         if ((error = xfs_bmap_finish(&tp, &free_list, &committed)))
 327                 goto error1;
 328         /*
 329          * The transaction must have been committed, since there were
 330          * actually extents freed by xfs_bunmapi.  See xfs_bmap_finish.
 331          * The new tp has the extent freeing and EFDs.
 332          */
 333         ASSERT(committed);
 334         /*
 335          * The first xact was committed, so add the inode to the new one.
 336          * Mark it dirty so it will be logged and moved forward in the log as
 337          * part of every commit.
 338          */
 339         xfs_trans_ijoin(tp, ip, 0);
 340         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 341         /*
 342          * Get a new, empty transaction to return to our caller.
 343          */
 344         ntp = xfs_trans_dup(tp);
 345         /*
 346          * Commit the transaction containing extent freeing and EFDs.
 347          * If we get an error on the commit here or on the reserve below,
 348          * we need to unlock the inode since the new transaction doesn't
 349          * have the inode attached.
 350          */
 351         error = xfs_trans_commit(tp, 0);
 352         tp = ntp;
 353         if (error) {
 354                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 355                 goto error0;
 356         }
 357         /*
 358          * transaction commit worked ok so we can drop the extra ticket
 359          * reference that we gained in xfs_trans_dup()
 360          */
 361         xfs_log_ticket_put(tp->t_ticket);
 362
 363         /*
 364          * Remove the memory for extent descriptions (just bookkeeping).
 365          */
 366         if (ip->i_df.if_bytes)
 367                 xfs_idata_realloc(ip, -ip->i_df.if_bytes, XFS_DATA_FORK);
 368         ASSERT(ip->i_df.if_bytes == 0);
 369         /*
 370          * Put an itruncate log reservation in the new transaction
 371          * for our caller.
 372          */
 373         if ((error = xfs_trans_reserve(tp, 0, XFS_ITRUNCATE_LOG_RES(mp), 0,
 374                         XFS_TRANS_PERM_LOG_RES, XFS_ITRUNCATE_LOG_COUNT))) {
 375                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 376                 goto error0;
 377         }
 378
 379         xfs_trans_ijoin(tp, ip, 0);
 380         *tpp = tp;
 381         return 0;
 382
 383  error1:
 384         xfs_bmap_cancel(&free_list);
 385  error0:
 386         return error;
 387 }
 388
 389 int
 390 xfs_release(
 391         xfs_inode_t     *ip)
 392 {
 393         xfs_mount_t     *mp = ip->i_mount;
 394         int             error;
 395
 396         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
 397                 return 0;
 398
 399         /* If this is a read-only mount, don't do this (would generate I/O) */
 400         if (mp->m_flags & XFS_MOUNT_RDONLY)
 401                 return 0;
 402
 403         if (!XFS_FORCED_SHUTDOWN(mp)) {
 404                 int truncated;
 405
 406                 /*
 407                  * If we are using filestreams, and we have an unlinked
 408                  * file that we are processing the last close on, then nothing
 409                  * will be able to reopen and write to this file. Purge this
 410                  * inode from the filestreams cache so that it doesn't delay
 411                  * teardown of the inode.
 412                  */
 413                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
 414                         xfs_filestream_deassociate(ip);
 415
 416                 /*
 417                  * If we previously truncated this file and removed old data
 418                  * in the process, we want to initiate "early" writeout on
 419                  * the last close.  This is an attempt to combat the notorious
 420                  * NULL files problem which is particularly noticeable from a
 421                  * truncate down, buffered (re-)write (delalloc), followed by
 422                  * a crash.  What we are effectively doing here is
 423                  * significantly reducing the time window where we'd otherwise
 424                  * be exposed to that problem.
 425                  */
 426                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
 427                 if (truncated) {
 428                         xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
 429                         if (VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
 430                                 xfs_flush_pages(ip, 0, -1, XBF_ASYNC, FI_NONE);
 431                 }
 432         }
 433
 434         if (ip->i_d.di_nlink == 0)
 435                 return 0;
 436
 437         if ((S_ISREG(ip->i_d.di_mode) &&
 438              (VFS_I(ip)->i_size > 0 ||
 439               (VN_CACHED(VFS_I(ip)) > 0 || ip->i_delayed_blks > 0)) &&
 440              (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
 441             (!(ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
 442
 443                 /*
 444                  * If we can't get the iolock just skip truncating the blocks
 445                  * past EOF because we could deadlock with the mmap_sem
 446                  * otherwise.  We'll get another chance to drop them once the
 447                  * last reference to the inode is dropped, so we'll never leak
 448                  * blocks permanently.
 449                  *
 450                  * Further, check if the inode is being opened, written and
 451                  * closed frequently and we have delayed allocation blocks
 452                  * outstanding (e.g. streaming writes from the NFS server),
 453                  * truncating the blocks past EOF will cause fragmentation to
 454                  * occur.
 455                  *
 456                  * In this case don't do the truncation, either, but we have to
 457                  * be careful how we detect this case. Blocks beyond EOF show
 458                  * up as i_delayed_blks even when the inode is clean, so we
 459                  * need to truncate them away first before checking for a dirty
 460                  * release. Hence on the first dirty close we will still remove
 461                  * the speculative allocation, but after that we will leave it
 462                  * in place.
 463                  */
 464                 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
 465                         return 0;
 466
 467                 error = xfs_free_eofblocks(mp, ip, true);
 468                 if (error)
 469                         return error;
 470
 471                 /* delalloc blocks after truncation means it really is dirty */
 472                 if (ip->i_delayed_blks)
 473                         xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
 474         }
 475         return 0;
 476 }
 477
 478 /*
 479  * xfs_inactive
 480  *
 481  * This is called when the vnode reference count for the vnode
 482  * goes to zero.  If the file has been unlinked, then it must
 483  * now be truncated.  Also, we clear all of the read-ahead state
 484  * kept for the inode here since the file is now closed.
 485  */
 486 int
 487 xfs_inactive(
 488         xfs_inode_t     *ip)
 489 {
 490         xfs_bmap_free_t free_list;
 491         xfs_fsblock_t   first_block;
 492         int             committed;
 493         xfs_trans_t     *tp;
 494         xfs_mount_t     *mp;
 495         int             error;
 496         int             truncate = 0;
 497
 498         /*
 499          * If the inode is already free, then there can be nothing
 500          * to clean up here.
 501          */
 502         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
 503                 ASSERT(ip->i_df.if_real_bytes == 0);
 504                 ASSERT(ip->i_df.if_broot_bytes == 0);
 505                 return VN_INACTIVE_CACHE;
 506         }
 507
 508         mp = ip->i_mount;
 509
 510         error = 0;
 511
 512         /* If this is a read-only mount, don't do this (would generate I/O) */
 513         if (mp->m_flags & XFS_MOUNT_RDONLY)
 514                 goto out;
 515
 516         if (ip->i_d.di_nlink != 0) {
 517                 if ((S_ISREG(ip->i_d.di_mode) &&
 518                     (VFS_I(ip)->i_size > 0 ||
 519                      (VN_CACHED(VFS_I(ip)) > 0 || ip->i_delayed_blks > 0)) &&
 520                     (ip->i_df.if_flags & XFS_IFEXTENTS) &&
 521                     (!(ip->i_d.di_flags &
 522                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
 523                      ip->i_delayed_blks != 0))) {
 524                         error = xfs_free_eofblocks(mp, ip, false);
 525                         if (error)
 526                                 return VN_INACTIVE_CACHE;
 527                 }
 528                 goto out;
 529         }
 530
 531         if (S_ISREG(ip->i_d.di_mode) &&
 532             (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
 533              ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
 534                 truncate = 1;
 535
 536         error = xfs_qm_dqattach(ip, 0);
 537         if (error)
 538                 return VN_INACTIVE_CACHE;
 539
 540         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 541         error = xfs_trans_reserve(tp, 0,
 542                         (truncate || S_ISLNK(ip->i_d.di_mode)) ?
 543                                 XFS_ITRUNCATE_LOG_RES(mp) :
 544                                 XFS_IFREE_LOG_RES(mp),
 545                         0,
 546                         XFS_TRANS_PERM_LOG_RES,
 547                         XFS_ITRUNCATE_LOG_COUNT);
 548         if (error) {
 549                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 550                 xfs_trans_cancel(tp, 0);
 551                 return VN_INACTIVE_CACHE;
 552         }
 553
 554         xfs_ilock(ip, XFS_ILOCK_EXCL);
 555         xfs_trans_ijoin(tp, ip, 0);
 556
 557         if (S_ISLNK(ip->i_d.di_mode)) {
 558                 /*
 559                  * Zero length symlinks _can_ exist.
 560                  */
 561                 if (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) {
 562                         error = xfs_inactive_symlink_rmt(ip, &tp);
 563                         if (error)
 564                                 goto out_cancel;
 565                 } else if (ip->i_df.if_bytes > 0) {
 566                         xfs_idata_realloc(ip, -(ip->i_df.if_bytes),
 567                                           XFS_DATA_FORK);
 568                         ASSERT(ip->i_df.if_bytes == 0);
 569                 }
 570         } else if (truncate) {
 571                 ip->i_d.di_size = 0;
 572                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 573
 574                 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
 575                 if (error)
 576                         goto out_cancel;
 577
 578                 ASSERT(ip->i_d.di_nextents == 0);
 579         }
 580
 581         /*
 582          * If there are attributes associated with the file then blow them away
 583          * now.  The code calls a routine that recursively deconstructs the
 584          * attribute fork.  We need to just commit the current transaction
 585          * because we can't use it for xfs_attr_inactive().
 586          */
 587         if (ip->i_d.di_anextents > 0) {
 588                 ASSERT(ip->i_d.di_forkoff != 0);
 589
 590                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 591                 if (error)
 592                         goto out_unlock;
 593
 594                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 595
 596                 error = xfs_attr_inactive(ip);
 597                 if (error)
 598                         goto out;
 599
 600                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 601                 error = xfs_trans_reserve(tp, 0,
 602                                           XFS_IFREE_LOG_RES(mp),
 603                                           0, XFS_TRANS_PERM_LOG_RES,
 604                                           XFS_INACTIVE_LOG_COUNT);
 605                 if (error) {
 606                         xfs_trans_cancel(tp, 0);
 607                         goto out;
 608                 }
 609
 610                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 611                 xfs_trans_ijoin(tp, ip, 0);
 612         }
 613
 614         if (ip->i_afp)
 615                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
 616
 617         ASSERT(ip->i_d.di_anextents == 0);
 618
 619         /*
 620          * Free the inode.
 621          */
 622         xfs_bmap_init(&free_list, &first_block);
 623         error = xfs_ifree(tp, ip, &free_list);
 624         if (error) {
 625                 /*
 626                  * If we fail to free the inode, shut down.  The cancel
 627                  * might do that, we need to make sure.  Otherwise the
 628                  * inode might be lost for a long time or forever.
 629                  */
 630                 if (!XFS_FORCED_SHUTDOWN(mp)) {
 631                         xfs_notice(mp, "%s: xfs_ifree returned error %d",
 632                                 __func__, error);
 633                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
 634                 }
 635                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
 636         } else {
 637                 /*
 638                  * Credit the quota account(s). The inode is gone.
 639                  */
 640                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
 641
 642                 /*
 643                  * Just ignore errors at this point.  There is nothing we can
 644                  * do except to try to keep going. Make sure it's not a silent
 645                  * error.
 646                  */
 647                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
 648                 if (error)
 649                         xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
 650                                 __func__, error);
 651                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 652                 if (error)
 653                         xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
 654                                 __func__, error);
 655         }
 656
 657         /*
 658          * Release the dquots held by inode, if any.
 659          */
 660         xfs_qm_dqdetach(ip);
 661 out_unlock:
 662         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 663 out:
 664         return VN_INACTIVE_CACHE;
 665 out_cancel:
 666         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 667         goto out_unlock;
 668 }
 669
 670 /*
 671  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
 672  * is allowed, otherwise it has to be an exact match. If a CI match is found,
 673  * ci_name->name will point to a the actual name (caller must free) or
 674  * will be set to NULL if an exact match is found.
 675  */
 676 int
 677 xfs_lookup(
 678         xfs_inode_t             *dp,
 679         struct xfs_name         *name,
 680         xfs_inode_t             **ipp,
 681         struct xfs_name         *ci_name)
 682 {
 683         xfs_ino_t               inum;
 684         int                     error;
 685         uint                    lock_mode;
 686
 687         trace_xfs_lookup(dp, name);
 688
 689         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
 690                 return XFS_ERROR(EIO);
 691
 692         lock_mode = xfs_ilock_map_shared(dp);
 693         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
 694         xfs_iunlock_map_shared(dp, lock_mode);
 695
 696         if (error)
 697                 goto out;
 698
 699         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
 700         if (error)
 701                 goto out_free_name;
 702
 703         return 0;
 704
 705 out_free_name:
 706         if (ci_name)
 707                 kmem_free(ci_name->name);
 708 out:
 709         *ipp = NULL;
 710         return error;
 711 }
 712
 713 int
 714 xfs_create(
 715         xfs_inode_t             *dp,
 716         struct xfs_name         *name,
 717         umode_t                 mode,
 718         xfs_dev_t               rdev,
 719         xfs_inode_t             **ipp)
 720 {
 721         int                     is_dir = S_ISDIR(mode);
 722         struct xfs_mount        *mp = dp->i_mount;
 723         struct xfs_inode        *ip = NULL;
 724         struct xfs_trans        *tp = NULL;
 725         int                     error;
 726         xfs_bmap_free_t         free_list;
 727         xfs_fsblock_t           first_block;
 728         boolean_t               unlock_dp_on_error = B_FALSE;
 729         uint                    cancel_flags;
 730         int                     committed;
 731         prid_t                  prid;
 732         struct xfs_dquot        *udqp = NULL;
 733         struct xfs_dquot        *gdqp = NULL;
 734         uint                    resblks;
 735         uint                    log_res;
 736         uint                    log_count;
 737
 738         trace_xfs_create(dp, name);
 739
 740         if (XFS_FORCED_SHUTDOWN(mp))
 741                 return XFS_ERROR(EIO);
 742
 743         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
 744                 prid = xfs_get_projid(dp);
 745         else
 746                 prid = XFS_PROJID_DEFAULT;
 747
 748         /*
 749          * Make sure that we have allocated dquot(s) on disk.
 750          */
 751         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
 752                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
 753         if (error)
 754                 return error;
 755
 756         if (is_dir) {
 757                 rdev = 0;
 758                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
 759                 log_res = XFS_MKDIR_LOG_RES(mp);
 760                 log_count = XFS_MKDIR_LOG_COUNT;
 761                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
 762         } else {
 763                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
 764                 log_res = XFS_CREATE_LOG_RES(mp);
 765                 log_count = XFS_CREATE_LOG_COUNT;
 766                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
 767         }
 768
 769         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
 770
 771         /*
 772          * Initially assume that the file does not exist and
 773          * reserve the resources for that case.  If that is not
 774          * the case we'll drop the one we have and get a more
 775          * appropriate transaction later.
 776          */
 777         error = xfs_trans_reserve(tp, resblks, log_res, 0,
 778                         XFS_TRANS_PERM_LOG_RES, log_count);
 779         if (error == ENOSPC) {
 780                 /* flush outstanding delalloc blocks and retry */
 781                 xfs_flush_inodes(mp);
 782                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
 783                                 XFS_TRANS_PERM_LOG_RES, log_count);
 784         }
 785         if (error == ENOSPC) {
 786                 /* No space at all so try a "no-allocation" reservation */
 787                 resblks = 0;
 788                 error = xfs_trans_reserve(tp, 0, log_res, 0,
 789                                 XFS_TRANS_PERM_LOG_RES, log_count);
 790         }
 791         if (error) {
 792                 cancel_flags = 0;
 793                 goto out_trans_cancel;
 794         }
 795
 796         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
 797         unlock_dp_on_error = B_TRUE;
 798
 799         xfs_bmap_init(&free_list, &first_block);
 800
 801         /*
 802          * Reserve disk quota and the inode.
 803          */
 804         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
 805         if (error)
 806                 goto out_trans_cancel;
 807
 808         error = xfs_dir_canenter(tp, dp, name, resblks);
 809         if (error)
 810                 goto out_trans_cancel;
 811
 812         /*
 813          * A newly created regular or special file just has one directory
 814          * entry pointing to them, but a directory also the "." entry
 815          * pointing to itself.
 816          */
 817         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
 818                                prid, resblks > 0, &ip, &committed);
 819         if (error) {
 820                 if (error == ENOSPC)
 821                         goto out_trans_cancel;
 822                 goto out_trans_abort;
 823         }
 824
 825         /*
 826          * Now we join the directory inode to the transaction.  We do not do it
 827          * earlier because xfs_dir_ialloc might commit the previous transaction
 828          * (and release all the locks).  An error from here on will result in
 829          * the transaction cancel unlocking dp so don't do it explicitly in the
 830          * error path.
 831          */
 832         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
 833         unlock_dp_on_error = B_FALSE;
 834
 835         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
 836                                         &first_block, &free_list, resblks ?
 837                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
 838         if (error) {
 839                 ASSERT(error != ENOSPC);
 840                 goto out_trans_abort;
 841         }
 842         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 843         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
 844
 845         if (is_dir) {
 846                 error = xfs_dir_init(tp, ip, dp);
 847                 if (error)
 848                         goto out_bmap_cancel;
 849
 850                 error = xfs_bumplink(tp, dp);
 851                 if (error)
 852                         goto out_bmap_cancel;
 853         }
 854
 855         /*
 856          * If this is a synchronous mount, make sure that the
 857          * create transaction goes to disk before returning to
 858          * the user.
 859          */
 860         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
 861                 xfs_trans_set_sync(tp);
 862
 863         /*
 864          * Attach the dquot(s) to the inodes and modify them incore.
 865          * These ids of the inode couldn't have changed since the new
 866          * inode has been locked ever since it was created.
 867          */
 868         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
 869
 870         error = xfs_bmap_finish(&tp, &free_list, &committed);
 871         if (error)
 872                 goto out_bmap_cancel;
 873
 874         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 875         if (error)
 876                 goto out_release_inode;
 877
 878         xfs_qm_dqrele(udqp);
 879         xfs_qm_dqrele(gdqp);
 880
 881         *ipp = ip;
 882         return 0;
 883
 884  out_bmap_cancel:
 885         xfs_bmap_cancel(&free_list);
 886  out_trans_abort:
 887         cancel_flags |= XFS_TRANS_ABORT;
 888  out_trans_cancel:
 889         xfs_trans_cancel(tp, cancel_flags);
 890  out_release_inode:
 891         /*
 892          * Wait until after the current transaction is aborted to
 893          * release the inode.  This prevents recursive transactions
 894          * and deadlocks from xfs_inactive.
 895          */
 896         if (ip)
 897                 IRELE(ip);
 898
 899         xfs_qm_dqrele(udqp);
 900         xfs_qm_dqrele(gdqp);
 901
 902         if (unlock_dp_on_error)
 903                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
 904         return error;
 905 }
 906
 907 #ifdef DEBUG
 908 int xfs_locked_n;
 909 int xfs_small_retries;
 910 int xfs_middle_retries;
 911 int xfs_lots_retries;
 912 int xfs_lock_delays;
 913 #endif
 914
 915 /*
 916  * Bump the subclass so xfs_lock_inodes() acquires each lock with
 917  * a different value
 918  */
 919 static inline int
 920 xfs_lock_inumorder(int lock_mode, int subclass)
 921 {
 922         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
 923                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
 924         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
 925                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
 926
 927         return lock_mode;
 928 }
 929
 930 /*
 931  * The following routine will lock n inodes in exclusive mode.
 932  * We assume the caller calls us with the inodes in i_ino order.
 933  *
 934  * We need to detect deadlock where an inode that we lock
 935  * is in the AIL and we start waiting for another inode that is locked
 936  * by a thread in a long running transaction (such as truncate). This can
 937  * result in deadlock since the long running trans might need to wait
 938  * for the inode we just locked in order to push the tail and free space
 939  * in the log.
 940  */
 941 void
 942 xfs_lock_inodes(
 943         xfs_inode_t     **ips,
 944         int             inodes,
 945         uint            lock_mode)
 946 {
 947         int             attempts = 0, i, j, try_lock;
 948         xfs_log_item_t  *lp;
 949
 950         ASSERT(ips && (inodes >= 2)); /* we need at least two */
 951
 952         try_lock = 0;
 953         i = 0;
 954
 955 again:
 956         for (; i < inodes; i++) {
 957                 ASSERT(ips[i]);
 958
 959                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
 960                         continue;
 961
 962                 /*
 963                  * If try_lock is not set yet, make sure all locked inodes
 964                  * are not in the AIL.
 965                  * If any are, set try_lock to be used later.
 966                  */
 967
 968                 if (!try_lock) {
 969                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
 970                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
 971                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
 972                                         try_lock++;
 973                                 }
 974                         }
 975                 }
 976
 977                 /*
 978                  * If any of the previous locks we have locked is in the AIL,
 979                  * we must TRY to get the second and subsequent locks. If
 980                  * we can't get any, we must release all we have
 981                  * and try again.
 982                  */
 983
 984                 if (try_lock) {
 985                         /* try_lock must be 0 if i is 0. */
 986                         /*
 987                          * try_lock means we have an inode locked
 988                          * that is in the AIL.
 989                          */
 990                         ASSERT(i != 0);
 991                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
 992                                 attempts++;
 993
 994                                 /*
 995                                  * Unlock all previous guys and try again.
 996                                  * xfs_iunlock will try to push the tail
 997                                  * if the inode is in the AIL.
 998                                  */
 999
1000                                 for(j = i - 1; j >= 0; j--) {
1001
1002                                         /*
1003                                          * Check to see if we've already
1004                                          * unlocked this one.
1005                                          * Not the first one going back,
1006                                          * and the inode ptr is the same.
1007                                          */
1008                                         if ((j != (i - 1)) && ips[j] ==
1009                                                                 ips[j+1])
1010                                                 continue;
1011
1012                                         xfs_iunlock(ips[j], lock_mode);
1013                                 }
1014
1015                                 if ((attempts % 5) == 0) {
1016                                         delay(1); /* Don't just spin the CPU */
1017 #ifdef DEBUG
1018                                         xfs_lock_delays++;
1019 #endif
1020                                 }
1021                                 i = 0;
1022                                 try_lock = 0;
1023                                 goto again;
1024                         }
1025                 } else {
1026                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1027                 }
1028         }
1029
1030 #ifdef DEBUG
1031         if (attempts) {
1032                 if (attempts < 5) xfs_small_retries++;
1033                 else if (attempts < 100) xfs_middle_retries++;
1034                 else xfs_lots_retries++;
1035         } else {
1036                 xfs_locked_n++;
1037         }
1038 #endif
1039 }
1040
1041 /*
1042  * xfs_lock_two_inodes() can only be used to lock one type of lock
1043  * at a time - the iolock or the ilock, but not both at once. If
1044  * we lock both at once, lockdep will report false positives saying
1045  * we have violated locking orders.
1046  */
1047 void
1048 xfs_lock_two_inodes(
1049         xfs_inode_t             *ip0,
1050         xfs_inode_t             *ip1,
1051         uint                    lock_mode)
1052 {
1053         xfs_inode_t             *temp;
1054         int                     attempts = 0;
1055         xfs_log_item_t          *lp;
1056
1057         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1058                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1059         ASSERT(ip0->i_ino != ip1->i_ino);
1060
1061         if (ip0->i_ino > ip1->i_ino) {
1062                 temp = ip0;
1063                 ip0 = ip1;
1064                 ip1 = temp;
1065         }
1066
1067  again:
1068         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1069
1070         /*
1071          * If the first lock we have locked is in the AIL, we must TRY to get
1072          * the second lock. If we can't get it, we must release the first one
1073          * and try again.
1074          */
1075         lp = (xfs_log_item_t *)ip0->i_itemp;
1076         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1077                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1078                         xfs_iunlock(ip0, lock_mode);
1079                         if ((++attempts % 5) == 0)
1080                                 delay(1); /* Don't just spin the CPU */
1081                         goto again;
1082                 }
1083         } else {
1084                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1085         }
1086 }
1087
1088 int
1089 xfs_remove(
1090         xfs_inode_t             *dp,
1091         struct xfs_name         *name,
1092         xfs_inode_t             *ip)
1093 {
1094         xfs_mount_t             *mp = dp->i_mount;
1095         xfs_trans_t             *tp = NULL;
1096         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1097         int                     error = 0;
1098         xfs_bmap_free_t         free_list;
1099         xfs_fsblock_t           first_block;
1100         int                     cancel_flags;
1101         int                     committed;
1102         int                     link_zero;
1103         uint                    resblks;
1104         uint                    log_count;
1105
1106         trace_xfs_remove(dp, name);
1107
1108         if (XFS_FORCED_SHUTDOWN(mp))
1109                 return XFS_ERROR(EIO);
1110
1111         error = xfs_qm_dqattach(dp, 0);
1112         if (error)
1113                 goto std_return;
1114
1115         error = xfs_qm_dqattach(ip, 0);
1116         if (error)
1117                 goto std_return;
1118
1119         if (is_dir) {
1120                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1121                 log_count = XFS_DEFAULT_LOG_COUNT;
1122         } else {
1123                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1124                 log_count = XFS_REMOVE_LOG_COUNT;
1125         }
1126         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1127
1128         /*
1129          * We try to get the real space reservation first,
1130          * allowing for directory btree deletion(s) implying
1131          * possible bmap insert(s).  If we can't get the space
1132          * reservation then we use 0 instead, and avoid the bmap
1133          * btree insert(s) in the directory code by, if the bmap
1134          * insert tries to happen, instead trimming the LAST
1135          * block from the directory.
1136          */
1137         resblks = XFS_REMOVE_SPACE_RES(mp);
1138         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1139                                   XFS_TRANS_PERM_LOG_RES, log_count);
1140         if (error == ENOSPC) {
1141                 resblks = 0;
1142                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1143                                           XFS_TRANS_PERM_LOG_RES, log_count);
1144         }
1145         if (error) {
1146                 ASSERT(error != ENOSPC);
1147                 cancel_flags = 0;
1148                 goto out_trans_cancel;
1149         }
1150
1151         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1152
1153         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1154         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1155
1156         /*
1157          * If we're removing a directory perform some additional validation.
1158          */
1159         if (is_dir) {
1160                 ASSERT(ip->i_d.di_nlink >= 2);
1161                 if (ip->i_d.di_nlink != 2) {
1162                         error = XFS_ERROR(ENOTEMPTY);
1163                         goto out_trans_cancel;
1164                 }
1165                 if (!xfs_dir_isempty(ip)) {
1166                         error = XFS_ERROR(ENOTEMPTY);
1167                         goto out_trans_cancel;
1168                 }
1169         }
1170
1171         xfs_bmap_init(&free_list, &first_block);
1172         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1173                                         &first_block, &free_list, resblks);
1174         if (error) {
1175                 ASSERT(error != ENOENT);
1176                 goto out_bmap_cancel;
1177         }
1178         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1179
1180         if (is_dir) {
1181                 /*
1182                  * Drop the link from ip's "..".
1183                  */
1184                 error = xfs_droplink(tp, dp);
1185                 if (error)
1186                         goto out_bmap_cancel;
1187
1188                 /*
1189                  * Drop the "." link from ip to self.
1190                  */
1191                 error = xfs_droplink(tp, ip);
1192                 if (error)
1193                         goto out_bmap_cancel;
1194         } else {
1195                 /*
1196                  * When removing a non-directory we need to log the parent
1197                  * inode here.  For a directory this is done implicitly
1198                  * by the xfs_droplink call for the ".." entry.
1199                  */
1200                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1201         }
1202
1203         /*
1204          * Drop the link from dp to ip.
1205          */
1206         error = xfs_droplink(tp, ip);
1207         if (error)
1208                 goto out_bmap_cancel;
1209
1210         /*
1211          * Determine if this is the last link while
1212          * we are in the transaction.
1213          */
1214         link_zero = (ip->i_d.di_nlink == 0);
1215
1216         /*
1217          * If this is a synchronous mount, make sure that the
1218          * remove transaction goes to disk before returning to
1219          * the user.
1220          */
1221         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1222                 xfs_trans_set_sync(tp);
1223
1224         error = xfs_bmap_finish(&tp, &free_list, &committed);
1225         if (error)
1226                 goto out_bmap_cancel;
1227
1228         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1229         if (error)
1230                 goto std_return;
1231
1232         /*
1233          * If we are using filestreams, kill the stream association.
1234          * If the file is still open it may get a new one but that
1235          * will get killed on last close in xfs_close() so we don't
1236          * have to worry about that.
1237          */
1238         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1239                 xfs_filestream_deassociate(ip);
1240
1241         return 0;
1242
1243  out_bmap_cancel:
1244         xfs_bmap_cancel(&free_list);
1245         cancel_flags |= XFS_TRANS_ABORT;
1246  out_trans_cancel:
1247         xfs_trans_cancel(tp, cancel_flags);
1248  std_return:
1249         return error;
1250 }
1251
1252 int
1253 xfs_link(
1254         xfs_inode_t             *tdp,
1255         xfs_inode_t             *sip,
1256         struct xfs_name         *target_name)
1257 {
1258         xfs_mount_t             *mp = tdp->i_mount;
1259         xfs_trans_t             *tp;
1260         int                     error;
1261         xfs_bmap_free_t         free_list;
1262         xfs_fsblock_t           first_block;
1263         int                     cancel_flags;
1264         int                     committed;
1265         int                     resblks;
1266
1267         trace_xfs_link(tdp, target_name);
1268
1269         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1270
1271         if (XFS_FORCED_SHUTDOWN(mp))
1272                 return XFS_ERROR(EIO);
1273
1274         error = xfs_qm_dqattach(sip, 0);
1275         if (error)
1276                 goto std_return;
1277
1278         error = xfs_qm_dqattach(tdp, 0);
1279         if (error)
1280                 goto std_return;
1281
1282         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1283         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1284         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1285         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1286                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1287         if (error == ENOSPC) {
1288                 resblks = 0;
1289                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1290                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1291         }
1292         if (error) {
1293                 cancel_flags = 0;
1294                 goto error_return;
1295         }
1296
1297         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1298
1299         xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1300         xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1301
1302         /*
1303          * If we are using project inheritance, we only allow hard link
1304          * creation in our tree when the project IDs are the same; else
1305          * the tree quota mechanism could be circumvented.
1306          */
1307         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1308                      (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1309                 error = XFS_ERROR(EXDEV);
1310                 goto error_return;
1311         }
1312
1313         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1314         if (error)
1315                 goto error_return;
1316
1317         xfs_bmap_init(&free_list, &first_block);
1318
1319         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1320                                         &first_block, &free_list, resblks);
1321         if (error)
1322                 goto abort_return;
1323         xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1324         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1325
1326         error = xfs_bumplink(tp, sip);
1327         if (error)
1328                 goto abort_return;
1329
1330         /*
1331          * If this is a synchronous mount, make sure that the
1332          * link transaction goes to disk before returning to
1333          * the user.
1334          */
1335         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1336                 xfs_trans_set_sync(tp);
1337         }
1338
1339         error = xfs_bmap_finish (&tp, &free_list, &committed);
1340         if (error) {
1341                 xfs_bmap_cancel(&free_list);
1342                 goto abort_return;
1343         }
1344
1345         return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1346
1347  abort_return:
1348         cancel_flags |= XFS_TRANS_ABORT;
1349  error_return:
1350         xfs_trans_cancel(tp, cancel_flags);
1351  std_return:
1352         return error;
1353 }
1354
1355 int
1356 xfs_symlink(
1357         xfs_inode_t             *dp,
1358         struct xfs_name         *link_name,
1359         const char              *target_path,
1360         umode_t                 mode,
1361         xfs_inode_t             **ipp)
1362 {
1363         xfs_mount_t             *mp = dp->i_mount;
1364         xfs_trans_t             *tp;
1365         xfs_inode_t             *ip;
1366         int                     error;
1367         int                     pathlen;
1368         xfs_bmap_free_t         free_list;
1369         xfs_fsblock_t           first_block;
1370         boolean_t               unlock_dp_on_error = B_FALSE;
1371         uint                    cancel_flags;
1372         int                     committed;
1373         xfs_fileoff_t           first_fsb;
1374         xfs_filblks_t           fs_blocks;
1375         int                     nmaps;
1376         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
1377         xfs_daddr_t             d;
1378         const char              *cur_chunk;
1379         int                     byte_cnt;
1380         int                     n;
1381         xfs_buf_t               *bp;
1382         prid_t                  prid;
1383         struct xfs_dquot        *udqp, *gdqp;
1384         uint                    resblks;
1385
1386         *ipp = NULL;
1387         error = 0;
1388         ip = NULL;
1389         tp = NULL;
1390
1391         trace_xfs_symlink(dp, link_name);
1392
1393         if (XFS_FORCED_SHUTDOWN(mp))
1394                 return XFS_ERROR(EIO);
1395
1396         /*
1397          * Check component lengths of the target path name.
1398          */
1399         pathlen = strlen(target_path);
1400         if (pathlen >= MAXPATHLEN)      /* total string too long */
1401                 return XFS_ERROR(ENAMETOOLONG);
1402
1403         udqp = gdqp = NULL;
1404         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1405                 prid = xfs_get_projid(dp);
1406         else
1407                 prid = XFS_PROJID_DEFAULT;
1408
1409         /*
1410          * Make sure that we have allocated dquot(s) on disk.
1411          */
1412         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1413                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1414         if (error)
1415                 goto std_return;
1416
1417         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
1418         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1419         /*
1420          * The symlink will fit into the inode data fork?
1421          * There can't be any attributes so we get the whole variable part.
1422          */
1423         if (pathlen <= XFS_LITINO(mp))
1424                 fs_blocks = 0;
1425         else
1426                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
1427         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
1428         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
1429                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1430         if (error == ENOSPC && fs_blocks == 0) {
1431                 resblks = 0;
1432                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
1433                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1434         }
1435         if (error) {
1436                 cancel_flags = 0;
1437                 goto error_return;
1438         }
1439
1440         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1441         unlock_dp_on_error = B_TRUE;
1442
1443         /*
1444          * Check whether the directory allows new symlinks or not.
1445          */
1446         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
1447                 error = XFS_ERROR(EPERM);
1448                 goto error_return;
1449         }
1450
1451         /*
1452          * Reserve disk quota : blocks and inode.
1453          */
1454         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
1455         if (error)
1456                 goto error_return;
1457
1458         /*
1459          * Check for ability to enter directory entry, if no space reserved.
1460          */
1461         error = xfs_dir_canenter(tp, dp, link_name, resblks);
1462         if (error)
1463                 goto error_return;
1464         /*
1465          * Initialize the bmap freelist prior to calling either
1466          * bmapi or the directory create code.
1467          */
1468         xfs_bmap_init(&free_list, &first_block);
1469
1470         /*
1471          * Allocate an inode for the symlink.
1472          */
1473         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT), 1, 0,
1474                                prid, resblks > 0, &ip, NULL);
1475         if (error) {
1476                 if (error == ENOSPC)
1477                         goto error_return;
1478                 goto error1;
1479         }
1480
1481         /*
1482          * An error after we've joined dp to the transaction will result in the
1483          * transaction cancel unlocking dp so don't do it explicitly in the
1484          * error path.
1485          */
1486         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1487         unlock_dp_on_error = B_FALSE;
1488
1489         /*
1490          * Also attach the dquot(s) to it, if applicable.
1491          */
1492         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
1493
1494         if (resblks)
1495                 resblks -= XFS_IALLOC_SPACE_RES(mp);
1496         /*
1497          * If the symlink will fit into the inode, write it inline.
1498          */
1499         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
1500                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
1501                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
1502                 ip->i_d.di_size = pathlen;
1503
1504                 /*
1505                  * The inode was initially created in extent format.
1506                  */
1507                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
1508                 ip->i_df.if_flags |= XFS_IFINLINE;
1509
1510                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
1511                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
1512
1513         } else {
1514                 first_fsb = 0;
1515                 nmaps = SYMLINK_MAPS;
1516
1517                 error = xfs_bmapi_write(tp, ip, first_fsb, fs_blocks,
1518                                   XFS_BMAPI_METADATA, &first_block, resblks,
1519                                   mval, &nmaps, &free_list);
1520                 if (error)
1521                         goto error2;
1522
1523                 if (resblks)
1524                         resblks -= fs_blocks;
1525                 ip->i_d.di_size = pathlen;
1526                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1527
1528                 cur_chunk = target_path;
1529                 for (n = 0; n < nmaps; n++) {
1530                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
1531                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
1532                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
1533                                                BTOBB(byte_cnt), 0);
1534                         if (!bp) {
1535                                 error = ENOMEM;
1536                                 goto error2;
1537                         }
1538                         if (pathlen < byte_cnt) {
1539                                 byte_cnt = pathlen;
1540                         }
1541                         pathlen -= byte_cnt;
1542
1543                         memcpy(bp->b_addr, cur_chunk, byte_cnt);
1544                         cur_chunk += byte_cnt;
1545
1546                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
1547                 }
1548         }
1549
1550         /*
1551          * Create the directory entry for the symlink.
1552          */
1553         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
1554                                         &first_block, &free_list, resblks);
1555         if (error)
1556                 goto error2;
1557         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1558         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1559
1560         /*
1561          * If this is a synchronous mount, make sure that the
1562          * symlink transaction goes to disk before returning to
1563          * the user.
1564          */
1565         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1566                 xfs_trans_set_sync(tp);
1567         }
1568
1569         error = xfs_bmap_finish(&tp, &free_list, &committed);
1570         if (error) {
1571                 goto error2;
1572         }
1573         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1574         xfs_qm_dqrele(udqp);
1575         xfs_qm_dqrele(gdqp);
1576
1577         *ipp = ip;
1578         return 0;
1579
1580  error2:
1581         IRELE(ip);
1582  error1:
1583         xfs_bmap_cancel(&free_list);
1584         cancel_flags |= XFS_TRANS_ABORT;
1585  error_return:
1586         xfs_trans_cancel(tp, cancel_flags);
1587         xfs_qm_dqrele(udqp);
1588         xfs_qm_dqrele(gdqp);
1589
1590         if (unlock_dp_on_error)
1591                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1592  std_return:
1593         return error;
1594 }
1595
1596 int
1597 xfs_set_dmattrs(
1598         xfs_inode_t     *ip,
1599         u_int           evmask,
1600         u_int16_t       state)
1601 {
1602         xfs_mount_t     *mp = ip->i_mount;
1603         xfs_trans_t     *tp;
1604         int             error;
1605
1606         if (!capable(CAP_SYS_ADMIN))
1607                 return XFS_ERROR(EPERM);
1608
1609         if (XFS_FORCED_SHUTDOWN(mp))
1610                 return XFS_ERROR(EIO);
1611
1612         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
1613         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
1614         if (error) {
1615                 xfs_trans_cancel(tp, 0);
1616                 return error;
1617         }
1618         xfs_ilock(ip, XFS_ILOCK_EXCL);
1619         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1620
1621         ip->i_d.di_dmevmask = evmask;
1622         ip->i_d.di_dmstate  = state;
1623
1624         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1625         error = xfs_trans_commit(tp, 0);
1626
1627         return error;
1628 }
1629
1630 /*
1631  * xfs_alloc_file_space()
1632  *      This routine allocates disk space for the given file.
1633  *
1634  *      If alloc_type == 0, this request is for an ALLOCSP type
1635  *      request which will change the file size.  In this case, no
1636  *      DMAPI event will be generated by the call.  A TRUNCATE event
1637  *      will be generated later by xfs_setattr.
1638  *
1639  *      If alloc_type != 0, this request is for a RESVSP type
1640  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
1641  *      lower block boundary byte address is less than the file's
1642  *      length.
1643  *
1644  * RETURNS:
1645  *       0 on success
1646  *      errno on error
1647  *
1648  */
1649 STATIC int
1650 xfs_alloc_file_space(
1651         xfs_inode_t             *ip,
1652         xfs_off_t               offset,
1653         xfs_off_t               len,
1654         int                     alloc_type,
1655         int                     attr_flags)
1656 {
1657         xfs_mount_t             *mp = ip->i_mount;
1658         xfs_off_t               count;
1659         xfs_filblks_t           allocated_fsb;
1660         xfs_filblks_t           allocatesize_fsb;
1661         xfs_extlen_t            extsz, temp;
1662         xfs_fileoff_t           startoffset_fsb;
1663         xfs_fsblock_t           firstfsb;
1664         int                     nimaps;
1665         int                     quota_flag;
1666         int                     rt;
1667         xfs_trans_t             *tp;
1668         xfs_bmbt_irec_t         imaps[1], *imapp;
1669         xfs_bmap_free_t         free_list;
1670         uint                    qblocks, resblks, resrtextents;
1671         int                     committed;
1672         int                     error;
1673
1674         trace_xfs_alloc_file_space(ip);
1675
1676         if (XFS_FORCED_SHUTDOWN(mp))
1677                 return XFS_ERROR(EIO);
1678
1679         error = xfs_qm_dqattach(ip, 0);
1680         if (error)
1681                 return error;
1682
1683         if (len <= 0)
1684                 return XFS_ERROR(EINVAL);
1685
1686         rt = XFS_IS_REALTIME_INODE(ip);
1687         extsz = xfs_get_extsz_hint(ip);
1688
1689         count = len;
1690         imapp = &imaps[0];
1691         nimaps = 1;
1692         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
1693         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
1694
1695         /*
1696          * Allocate file space until done or until there is an error
1697          */
1698         while (allocatesize_fsb && !error) {
1699                 xfs_fileoff_t   s, e;
1700
1701                 /*
1702                  * Determine space reservations for data/realtime.
1703                  */
1704                 if (unlikely(extsz)) {
1705                         s = startoffset_fsb;
1706                         do_div(s, extsz);
1707                         s *= extsz;
1708                         e = startoffset_fsb + allocatesize_fsb;
1709                         if ((temp = do_mod(startoffset_fsb, extsz)))
1710                                 e += temp;
1711                         if ((temp = do_mod(e, extsz)))
1712                                 e += extsz - temp;
1713                 } else {
1714                         s = 0;
1715                         e = allocatesize_fsb;
1716                 }
1717
1718                 /*
1719                  * The transaction reservation is limited to a 32-bit block
1720                  * count, hence we need to limit the number of blocks we are
1721                  * trying to reserve to avoid an overflow. We can't allocate
1722                  * more than @nimaps extents, and an extent is limited on disk
1723                  * to MAXEXTLEN (21 bits), so use that to enforce the limit.
1724                  */
1725                 resblks = min_t(xfs_fileoff_t, (e - s), (MAXEXTLEN * nimaps));
1726                 if (unlikely(rt)) {
1727                         resrtextents = qblocks = resblks;
1728                         resrtextents /= mp->m_sb.sb_rextsize;
1729                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
1730                         quota_flag = XFS_QMOPT_RES_RTBLKS;
1731                 } else {
1732                         resrtextents = 0;
1733                         resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resblks);
1734                         quota_flag = XFS_QMOPT_RES_REGBLKS;
1735                 }
1736
1737                 /*
1738                  * Allocate and setup the transaction.
1739                  */
1740                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
1741                 error = xfs_trans_reserve(tp, resblks,
1742                                           XFS_WRITE_LOG_RES(mp), resrtextents,
1743                                           XFS_TRANS_PERM_LOG_RES,
1744                                           XFS_WRITE_LOG_COUNT);
1745                 /*
1746                  * Check for running out of space
1747                  */
1748                 if (error) {
1749                         /*
1750                          * Free the transaction structure.
1751                          */
1752                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
1753                         xfs_trans_cancel(tp, 0);
1754                         break;
1755                 }
1756                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1757                 error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks,
1758                                                       0, quota_flag);
1759                 if (error)
1760                         goto error1;
1761
1762                 xfs_trans_ijoin(tp, ip, 0);
1763
1764                 xfs_bmap_init(&free_list, &firstfsb);
1765                 error = xfs_bmapi_write(tp, ip, startoffset_fsb,
1766                                         allocatesize_fsb, alloc_type, &firstfsb,
1767                                         0, imapp, &nimaps, &free_list);
1768                 if (error) {
1769                         goto error0;
1770                 }
1771
1772                 /*
1773                  * Complete the transaction
1774                  */
1775                 error = xfs_bmap_finish(&tp, &free_list, &committed);
1776                 if (error) {
1777                         goto error0;
1778                 }
1779
1780                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1781                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1782                 if (error) {
1783                         break;
1784                 }
1785
1786                 allocated_fsb = imapp->br_blockcount;
1787
1788                 if (nimaps == 0) {
1789                         error = XFS_ERROR(ENOSPC);
1790                         break;
1791                 }
1792
1793                 startoffset_fsb += allocated_fsb;
1794                 allocatesize_fsb -= allocated_fsb;
1795         }
1796
1797         return error;
1798
1799 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
1800         xfs_bmap_cancel(&free_list);
1801         xfs_trans_unreserve_quota_nblks(tp, ip, (long)qblocks, 0, quota_flag);
1802
1803 error1: /* Just cancel transaction */
1804         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1805         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1806         return error;
1807 }
1808
1809 /*
1810  * Zero file bytes between startoff and endoff inclusive.
1811  * The iolock is held exclusive and no blocks are buffered.
1812  *
1813  * This function is used by xfs_free_file_space() to zero
1814  * partial blocks when the range to free is not block aligned.
1815  * When unreserving space with boundaries that are not block
1816  * aligned we round up the start and round down the end
1817  * boundaries and then use this function to zero the parts of
1818  * the blocks that got dropped during the rounding.
1819  */
1820 STATIC int
1821 xfs_zero_remaining_bytes(
1822         xfs_inode_t             *ip,
1823         xfs_off_t               startoff,
1824         xfs_off_t               endoff)
1825 {
1826         xfs_bmbt_irec_t         imap;
1827         xfs_fileoff_t           offset_fsb;
1828         xfs_off_t               lastoffset;
1829         xfs_off_t               offset;
1830         xfs_buf_t               *bp;
1831         xfs_mount_t             *mp = ip->i_mount;
1832         int                     nimap;
1833         int                     error = 0;
1834
1835         /*
1836          * Avoid doing I/O beyond eof - it's not necessary
1837          * since nothing can read beyond eof.  The space will
1838          * be zeroed when the file is extended anyway.
1839          */
1840         if (startoff >= XFS_ISIZE(ip))
1841                 return 0;
1842
1843         if (endoff > XFS_ISIZE(ip))
1844                 endoff = XFS_ISIZE(ip);
1845
1846         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
1847                                         mp->m_rtdev_targp : mp->m_ddev_targp,
1848                                   BTOBB(mp->m_sb.sb_blocksize), 0);
1849         if (!bp)
1850                 return XFS_ERROR(ENOMEM);
1851
1852         xfs_buf_unlock(bp);
1853
1854         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
1855                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1856                 nimap = 1;
1857                 error = xfs_bmapi_read(ip, offset_fsb, 1, &imap, &nimap, 0);
1858                 if (error || nimap < 1)
1859                         break;
1860                 ASSERT(imap.br_blockcount >= 1);
1861                 ASSERT(imap.br_startoff == offset_fsb);
1862                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
1863                 if (lastoffset > endoff)
1864                         lastoffset = endoff;
1865                 if (imap.br_startblock == HOLESTARTBLOCK)
1866                         continue;
1867                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1868                 if (imap.br_state == XFS_EXT_UNWRITTEN)
1869                         continue;
1870                 XFS_BUF_UNDONE(bp);
1871                 XFS_BUF_UNWRITE(bp);
1872                 XFS_BUF_READ(bp);
1873                 XFS_BUF_SET_ADDR(bp, xfs_fsb_to_db(ip, imap.br_startblock));
1874                 xfsbdstrat(mp, bp);
1875                 error = xfs_buf_iowait(bp);
1876                 if (error) {
1877                         xfs_buf_ioerror_alert(bp,
1878                                         "xfs_zero_remaining_bytes(read)");
1879                         break;
1880                 }
1881                 memset(bp->b_addr +
1882                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
1883                       0, lastoffset - offset + 1);
1884                 XFS_BUF_UNDONE(bp);
1885                 XFS_BUF_UNREAD(bp);
1886                 XFS_BUF_WRITE(bp);
1887                 xfsbdstrat(mp, bp);
1888                 error = xfs_buf_iowait(bp);
1889                 if (error) {
1890                         xfs_buf_ioerror_alert(bp,
1891                                         "xfs_zero_remaining_bytes(write)");
1892                         break;
1893                 }
1894         }
1895         xfs_buf_free(bp);
1896         return error;
1897 }
1898
1899 /*
1900  * xfs_free_file_space()
1901  *      This routine frees disk space for the given file.
1902  *
1903  *      This routine is only called by xfs_change_file_space
1904  *      for an UNRESVSP type call.
1905  *
1906  * RETURNS:
1907  *       0 on success
1908  *      errno on error
1909  *
1910  */
1911 STATIC int
1912 xfs_free_file_space(
1913         xfs_inode_t             *ip,
1914         xfs_off_t               offset,
1915         xfs_off_t               len,
1916         int                     attr_flags)
1917 {
1918         int                     committed;
1919         int                     done;
1920         xfs_fileoff_t           endoffset_fsb;
1921         int                     error;
1922         xfs_fsblock_t           firstfsb;
1923         xfs_bmap_free_t         free_list;
1924         xfs_bmbt_irec_t         imap;
1925         xfs_off_t               ioffset;
1926         xfs_extlen_t            mod=0;
1927         xfs_mount_t             *mp;
1928         int                     nimap;
1929         uint                    resblks;
1930         uint                    rounding;
1931         int                     rt;
1932         xfs_fileoff_t           startoffset_fsb;
1933         xfs_trans_t             *tp;
1934         int                     need_iolock = 1;
1935
1936         mp = ip->i_mount;
1937
1938         trace_xfs_free_file_space(ip);
1939
1940         error = xfs_qm_dqattach(ip, 0);
1941         if (error)
1942                 return error;
1943
1944         error = 0;
1945         if (len <= 0)   /* if nothing being freed */
1946                 return error;
1947         rt = XFS_IS_REALTIME_INODE(ip);
1948         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
1949         endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
1950
1951         if (attr_flags & XFS_ATTR_NOLOCK)
1952                 need_iolock = 0;
1953         if (need_iolock) {
1954                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1955                 /* wait for the completion of any pending DIOs */
1956                 inode_dio_wait(VFS_I(ip));
1957         }
1958
1959         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
1960         ioffset = offset & ~(rounding - 1);
1961
1962         if (VN_CACHED(VFS_I(ip)) != 0) {
1963                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
1964                 if (error)
1965                         goto out_unlock_iolock;
1966         }
1967
1968         /*
1969          * Need to zero the stuff we're not freeing, on disk.
1970          * If it's a realtime file & can't use unwritten extents then we
1971          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
1972          * will take care of it for us.
1973          */
1974         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
1975                 nimap = 1;
1976                 error = xfs_bmapi_read(ip, startoffset_fsb, 1,
1977                                         &imap, &nimap, 0);
1978                 if (error)
1979                         goto out_unlock_iolock;
1980                 ASSERT(nimap == 0 || nimap == 1);
1981                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
1982                         xfs_daddr_t     block;
1983
1984                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1985                         block = imap.br_startblock;
1986                         mod = do_div(block, mp->m_sb.sb_rextsize);
1987                         if (mod)
1988                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
1989                 }
1990                 nimap = 1;
1991                 error = xfs_bmapi_read(ip, endoffset_fsb - 1, 1,
1992                                         &imap, &nimap, 0);
1993                 if (error)
1994                         goto out_unlock_iolock;
1995                 ASSERT(nimap == 0 || nimap == 1);
1996                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
1997                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1998                         mod++;
1999                         if (mod && (mod != mp->m_sb.sb_rextsize))
2000                                 endoffset_fsb -= mod;
2001                 }
2002         }
2003         if ((done = (endoffset_fsb <= startoffset_fsb)))
2004                 /*
2005                  * One contiguous piece to clear
2006                  */
2007                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
2008         else {
2009                 /*
2010                  * Some full blocks, possibly two pieces to clear
2011                  */
2012                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
2013                         error = xfs_zero_remaining_bytes(ip, offset,
2014                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
2015                 if (!error &&
2016                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
2017                         error = xfs_zero_remaining_bytes(ip,
2018                                 XFS_FSB_TO_B(mp, endoffset_fsb),
2019                                 offset + len - 1);
2020         }
2021
2022         /*
2023          * free file space until done or until there is an error
2024          */
2025         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2026         while (!error && !done) {
2027
2028                 /*
2029                  * allocate and setup the transaction. Allow this
2030                  * transaction to dip into the reserve blocks to ensure
2031                  * the freeing of the space succeeds at ENOSPC.
2032                  */
2033                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2034                 tp->t_flags |= XFS_TRANS_RESERVE;
2035                 error = xfs_trans_reserve(tp,
2036                                           resblks,
2037                                           XFS_WRITE_LOG_RES(mp),
2038                                           0,
2039                                           XFS_TRANS_PERM_LOG_RES,
2040                                           XFS_WRITE_LOG_COUNT);
2041
2042                 /*
2043                  * check for running out of space
2044                  */
2045                 if (error) {
2046                         /*
2047                          * Free the transaction structure.
2048                          */
2049                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2050                         xfs_trans_cancel(tp, 0);
2051                         break;
2052                 }
2053                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2054                 error = xfs_trans_reserve_quota(tp, mp,
2055                                 ip->i_udquot, ip->i_gdquot,
2056                                 resblks, 0, XFS_QMOPT_RES_REGBLKS);
2057                 if (error)
2058                         goto error1;
2059
2060                 xfs_trans_ijoin(tp, ip, 0);
2061
2062                 /*
2063                  * issue the bunmapi() call to free the blocks
2064                  */
2065                 xfs_bmap_init(&free_list, &firstfsb);
2066                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
2067                                   endoffset_fsb - startoffset_fsb,
2068                                   0, 2, &firstfsb, &free_list, &done);
2069                 if (error) {
2070                         goto error0;
2071                 }
2072
2073                 /*
2074                  * complete the transaction
2075                  */
2076                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2077                 if (error) {
2078                         goto error0;
2079                 }
2080
2081                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2082                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2083         }
2084
2085  out_unlock_iolock:
2086         if (need_iolock)
2087                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2088         return error;
2089
2090  error0:
2091         xfs_bmap_cancel(&free_list);
2092  error1:
2093         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2094         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
2095                     XFS_ILOCK_EXCL);
2096         return error;
2097 }
2098
2099 /*
2100  * xfs_change_file_space()
2101  *      This routine allocates or frees disk space for the given file.
2102  *      The user specified parameters are checked for alignment and size
2103  *      limitations.
2104  *
2105  * RETURNS:
2106  *       0 on success
2107  *      errno on error
2108  *
2109  */
2110 int
2111 xfs_change_file_space(
2112         xfs_inode_t     *ip,
2113         int             cmd,
2114         xfs_flock64_t   *bf,
2115         xfs_off_t       offset,
2116         int             attr_flags)
2117 {
2118         xfs_mount_t     *mp = ip->i_mount;
2119         int             clrprealloc;
2120         int             error;
2121         xfs_fsize_t     fsize;
2122         int             setprealloc;
2123         xfs_off_t       startoffset;
2124         xfs_off_t       llen;
2125         xfs_trans_t     *tp;
2126         struct iattr    iattr;
2127         int             prealloc_type;
2128
2129         if (!S_ISREG(ip->i_d.di_mode))
2130                 return XFS_ERROR(EINVAL);
2131
2132         switch (bf->l_whence) {
2133         case 0: /*SEEK_SET*/
2134                 break;
2135         case 1: /*SEEK_CUR*/
2136                 bf->l_start += offset;
2137                 break;
2138         case 2: /*SEEK_END*/
2139                 bf->l_start += XFS_ISIZE(ip);
2140                 break;
2141         default:
2142                 return XFS_ERROR(EINVAL);
2143         }
2144
2145         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
2146
2147         if (bf->l_start < 0 ||
2148             bf->l_start > mp->m_super->s_maxbytes ||
2149             bf->l_start + llen < 0 ||
2150             bf->l_start + llen > mp->m_super->s_maxbytes)
2151                 return XFS_ERROR(EINVAL);
2152
2153         bf->l_whence = 0;
2154
2155         startoffset = bf->l_start;
2156         fsize = XFS_ISIZE(ip);
2157
2158         /*
2159          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
2160          * file space.
2161          * These calls do NOT zero the data space allocated to the file,
2162          * nor do they change the file size.
2163          *
2164          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
2165          * space.
2166          * These calls cause the new file data to be zeroed and the file
2167          * size to be changed.
2168          */
2169         setprealloc = clrprealloc = 0;
2170         prealloc_type = XFS_BMAPI_PREALLOC;
2171
2172         switch (cmd) {
2173         case XFS_IOC_ZERO_RANGE:
2174                 prealloc_type |= XFS_BMAPI_CONVERT;
2175                 xfs_tosspages(ip, startoffset, startoffset + bf->l_len, 0);
2176                 /* FALLTHRU */
2177         case XFS_IOC_RESVSP:
2178         case XFS_IOC_RESVSP64:
2179                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
2180                                                 prealloc_type, attr_flags);
2181                 if (error)
2182                         return error;
2183                 setprealloc = 1;
2184                 break;
2185
2186         case XFS_IOC_UNRESVSP:
2187         case XFS_IOC_UNRESVSP64:
2188                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
2189                                                                 attr_flags)))
2190                         return error;
2191                 break;
2192
2193         case XFS_IOC_ALLOCSP:
2194         case XFS_IOC_ALLOCSP64:
2195         case XFS_IOC_FREESP:
2196         case XFS_IOC_FREESP64:
2197                 /*
2198                  * These operations actually do IO when extending the file, but
2199                  * the allocation is done seperately to the zeroing that is
2200                  * done. This set of operations need to be serialised against
2201                  * other IO operations, such as truncate and buffered IO. We
2202                  * need to take the IOLOCK here to serialise the allocation and
2203                  * zeroing IO to prevent other IOLOCK holders (e.g. getbmap,
2204                  * truncate, direct IO) from racing against the transient
2205                  * allocated but not written state we can have here.
2206                  */
2207                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
2208                 if (startoffset > fsize) {
2209                         error = xfs_alloc_file_space(ip, fsize,
2210                                         startoffset - fsize, 0,
2211                                         attr_flags | XFS_ATTR_NOLOCK);
2212                         if (error) {
2213                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2214                                 break;
2215                         }
2216                 }
2217
2218                 iattr.ia_valid = ATTR_SIZE;
2219                 iattr.ia_size = startoffset;
2220
2221                 error = xfs_setattr_size(ip, &iattr,
2222                                          attr_flags | XFS_ATTR_NOLOCK);
2223                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2224
2225                 if (error)
2226                         return error;
2227
2228                 clrprealloc = 1;
2229                 break;
2230
2231         default:
2232                 ASSERT(0);
2233                 return XFS_ERROR(EINVAL);
2234         }
2235
2236         /*
2237          * update the inode timestamp, mode, and prealloc flag bits
2238          */
2239         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
2240
2241         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
2242                                       0, 0, 0))) {
2243                 /* ASSERT(0); */
2244                 xfs_trans_cancel(tp, 0);
2245                 return error;
2246         }
2247
2248         xfs_ilock(ip, XFS_ILOCK_EXCL);
2249         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2250
2251         if ((attr_flags & XFS_ATTR_DMI) == 0) {
2252                 ip->i_d.di_mode &= ~S_ISUID;
2253
2254                 /*
2255                  * Note that we don't have to worry about mandatory
2256                  * file locking being disabled here because we only
2257                  * clear the S_ISGID bit if the Group execute bit is
2258                  * on, but if it was on then mandatory locking wouldn't
2259                  * have been enabled.
2260                  */
2261                 if (ip->i_d.di_mode & S_IXGRP)
2262                         ip->i_d.di_mode &= ~S_ISGID;
2263
2264                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2265         }
2266         if (setprealloc)
2267                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
2268         else if (clrprealloc)
2269                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
2270
2271         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2272         if (attr_flags & XFS_ATTR_SYNC)
2273                 xfs_trans_set_sync(tp);
2274         return xfs_trans_commit(tp, 0);
2275 }