Pileus Git - ~andy/linux/blob - fs/xfs/xfs_vnodeops.c

   1 /*
   2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18
  19 #include "xfs.h"
  20 #include "xfs_fs.h"
  21 #include "xfs_types.h"
  22 #include "xfs_bit.h"
  23 #include "xfs_log.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_mount.h"
  29 #include "xfs_da_btree.h"
  30 #include "xfs_bmap_btree.h"
  31 #include "xfs_ialloc_btree.h"
  32 #include "xfs_dinode.h"
  33 #include "xfs_inode.h"
  34 #include "xfs_inode_item.h"
  35 #include "xfs_itable.h"
  36 #include "xfs_ialloc.h"
  37 #include "xfs_alloc.h"
  38 #include "xfs_bmap.h"
  39 #include "xfs_acl.h"
  40 #include "xfs_attr.h"
  41 #include "xfs_error.h"
  42 #include "xfs_quota.h"
  43 #include "xfs_utils.h"
  44 #include "xfs_rtalloc.h"
  45 #include "xfs_trans_space.h"
  46 #include "xfs_log_priv.h"
  47 #include "xfs_filestream.h"
  48 #include "xfs_vnodeops.h"
  49 #include "xfs_trace.h"
  50
  51 /*
  52  * The maximum pathlen is 1024 bytes. Since the minimum file system
  53  * blocksize is 512 bytes, we can get a max of 2 extents back from
  54  * bmapi.
  55  */
  56 #define SYMLINK_MAPS 2
  57
  58 STATIC int
  59 xfs_readlink_bmap(
  60         xfs_inode_t     *ip,
  61         char            *link)
  62 {
  63         xfs_mount_t     *mp = ip->i_mount;
  64         int             pathlen = ip->i_d.di_size;
  65         int             nmaps = SYMLINK_MAPS;
  66         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
  67         xfs_daddr_t     d;
  68         int             byte_cnt;
  69         int             n;
  70         xfs_buf_t       *bp;
  71         int             error = 0;
  72
  73         error = xfs_bmapi_read(ip, 0, XFS_B_TO_FSB(mp, pathlen), mval, &nmaps,
  74                                0);
  75         if (error)
  76                 goto out;
  77
  78         for (n = 0; n < nmaps; n++) {
  79                 d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
  80                 byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
  81
  82                 bp = xfs_buf_read(mp->m_ddev_targp, d, BTOBB(byte_cnt), 0);
  83                 if (!bp)
  84                         return XFS_ERROR(ENOMEM);
  85                 error = bp->b_error;
  86                 if (error) {
  87                         xfs_buf_ioerror_alert(bp, __func__);
  88                         xfs_buf_relse(bp);
  89                         goto out;
  90                 }
  91                 if (pathlen < byte_cnt)
  92                         byte_cnt = pathlen;
  93                 pathlen -= byte_cnt;
  94
  95                 memcpy(link, bp->b_addr, byte_cnt);
  96                 xfs_buf_relse(bp);
  97         }
  98
  99         link[ip->i_d.di_size] = '\0';
 100         error = 0;
 101
 102  out:
 103         return error;
 104 }
 105
 106 int
 107 xfs_readlink(
 108         xfs_inode_t     *ip,
 109         char            *link)
 110 {
 111         xfs_mount_t     *mp = ip->i_mount;
 112         xfs_fsize_t     pathlen;
 113         int             error = 0;
 114
 115         trace_xfs_readlink(ip);
 116
 117         if (XFS_FORCED_SHUTDOWN(mp))
 118                 return XFS_ERROR(EIO);
 119
 120         xfs_ilock(ip, XFS_ILOCK_SHARED);
 121
 122         pathlen = ip->i_d.di_size;
 123         if (!pathlen)
 124                 goto out;
 125
 126         if (pathlen < 0 || pathlen > MAXPATHLEN) {
 127                 xfs_alert(mp, "%s: inode (%llu) bad symlink length (%lld)",
 128                          __func__, (unsigned long long) ip->i_ino,
 129                          (long long) pathlen);
 130                 ASSERT(0);
 131                 error = XFS_ERROR(EFSCORRUPTED);
 132                 goto out;
 133         }
 134
 135
 136         if (ip->i_df.if_flags & XFS_IFINLINE) {
 137                 memcpy(link, ip->i_df.if_u1.if_data, pathlen);
 138                 link[pathlen] = '\0';
 139         } else {
 140                 error = xfs_readlink_bmap(ip, link);
 141         }
 142
 143  out:
 144         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 145         return error;
 146 }
 147
 148 /*
 149  * This is called by xfs_inactive to free any blocks beyond eof
 150  * when the link count isn't zero and by xfs_dm_punch_hole() when
 151  * punching a hole to EOF.
 152  */
 153 STATIC int
 154 xfs_free_eofblocks(
 155         xfs_mount_t     *mp,
 156         xfs_inode_t     *ip,
 157         bool            need_iolock)
 158 {
 159         xfs_trans_t     *tp;
 160         int             error;
 161         xfs_fileoff_t   end_fsb;
 162         xfs_fileoff_t   last_fsb;
 163         xfs_filblks_t   map_len;
 164         int             nimaps;
 165         xfs_bmbt_irec_t imap;
 166
 167         /*
 168          * Figure out if there are any blocks beyond the end
 169          * of the file.  If not, then there is nothing to do.
 170          */
 171         end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_ISIZE(ip));
 172         last_fsb = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
 173         if (last_fsb <= end_fsb)
 174                 return 0;
 175         map_len = last_fsb - end_fsb;
 176
 177         nimaps = 1;
 178         xfs_ilock(ip, XFS_ILOCK_SHARED);
 179         error = xfs_bmapi_read(ip, end_fsb, map_len, &imap, &nimaps, 0);
 180         xfs_iunlock(ip, XFS_ILOCK_SHARED);
 181
 182         if (!error && (nimaps != 0) &&
 183             (imap.br_startblock != HOLESTARTBLOCK ||
 184              ip->i_delayed_blks)) {
 185                 /*
 186                  * Attach the dquots to the inode up front.
 187                  */
 188                 error = xfs_qm_dqattach(ip, 0);
 189                 if (error)
 190                         return error;
 191
 192                 /*
 193                  * There are blocks after the end of file.
 194                  * Free them up now by truncating the file to
 195                  * its current size.
 196                  */
 197                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 198
 199                 if (need_iolock) {
 200                         if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
 201                                 xfs_trans_cancel(tp, 0);
 202                                 return 0;
 203                         }
 204                 }
 205
 206                 error = xfs_trans_reserve(tp, 0,
 207                                           XFS_ITRUNCATE_LOG_RES(mp),
 208                                           0, XFS_TRANS_PERM_LOG_RES,
 209                                           XFS_ITRUNCATE_LOG_COUNT);
 210                 if (error) {
 211                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 212                         xfs_trans_cancel(tp, 0);
 213                         if (need_iolock)
 214                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 215                         return error;
 216                 }
 217
 218                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 219                 xfs_trans_ijoin(tp, ip, 0);
 220
 221                 /*
 222                  * Do not update the on-disk file size.  If we update the
 223                  * on-disk file size and then the system crashes before the
 224                  * contents of the file are flushed to disk then the files
 225                  * may be full of holes (ie NULL files bug).
 226                  */
 227                 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK,
 228                                               XFS_ISIZE(ip));
 229                 if (error) {
 230                         /*
 231                          * If we get an error at this point we simply don't
 232                          * bother truncating the file.
 233                          */
 234                         xfs_trans_cancel(tp,
 235                                          (XFS_TRANS_RELEASE_LOG_RES |
 236                                           XFS_TRANS_ABORT));
 237                 } else {
 238                         error = xfs_trans_commit(tp,
 239                                                 XFS_TRANS_RELEASE_LOG_RES);
 240                 }
 241
 242                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 243                 if (need_iolock)
 244                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 245         }
 246         return error;
 247 }
 248
 249 /*
 250  * Free a symlink that has blocks associated with it.
 251  */
 252 STATIC int
 253 xfs_inactive_symlink_rmt(
 254         xfs_inode_t     *ip,
 255         xfs_trans_t     **tpp)
 256 {
 257         xfs_buf_t       *bp;
 258         int             committed;
 259         int             done;
 260         int             error;
 261         xfs_fsblock_t   first_block;
 262         xfs_bmap_free_t free_list;
 263         int             i;
 264         xfs_mount_t     *mp;
 265         xfs_bmbt_irec_t mval[SYMLINK_MAPS];
 266         int             nmaps;
 267         xfs_trans_t     *ntp;
 268         int             size;
 269         xfs_trans_t     *tp;
 270
 271         tp = *tpp;
 272         mp = ip->i_mount;
 273         ASSERT(ip->i_d.di_size > XFS_IFORK_DSIZE(ip));
 274         /*
 275          * We're freeing a symlink that has some
 276          * blocks allocated to it.  Free the
 277          * blocks here.  We know that we've got
 278          * either 1 or 2 extents and that we can
 279          * free them all in one bunmapi call.
 280          */
 281         ASSERT(ip->i_d.di_nextents > 0 && ip->i_d.di_nextents <= 2);
 282
 283         /*
 284          * Lock the inode, fix the size, and join it to the transaction.
 285          * Hold it so in the normal path, we still have it locked for
 286          * the second transaction.  In the error paths we need it
 287          * held so the cancel won't rele it, see below.
 288          */
 289         size = (int)ip->i_d.di_size;
 290         ip->i_d.di_size = 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         xfs_trans_ijoin(tp, ip, 0);
 379         *tpp = tp;
 380         return 0;
 381
 382  error1:
 383         xfs_bmap_cancel(&free_list);
 384  error0:
 385         return error;
 386 }
 387
 388 int
 389 xfs_release(
 390         xfs_inode_t     *ip)
 391 {
 392         xfs_mount_t     *mp = ip->i_mount;
 393         int             error;
 394
 395         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
 396                 return 0;
 397
 398         /* If this is a read-only mount, don't do this (would generate I/O) */
 399         if (mp->m_flags & XFS_MOUNT_RDONLY)
 400                 return 0;
 401
 402         if (!XFS_FORCED_SHUTDOWN(mp)) {
 403                 int truncated;
 404
 405                 /*
 406                  * If we are using filestreams, and we have an unlinked
 407                  * file that we are processing the last close on, then nothing
 408                  * will be able to reopen and write to this file. Purge this
 409                  * inode from the filestreams cache so that it doesn't delay
 410                  * teardown of the inode.
 411                  */
 412                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
 413                         xfs_filestream_deassociate(ip);
 414
 415                 /*
 416                  * If we previously truncated this file and removed old data
 417                  * in the process, we want to initiate "early" writeout on
 418                  * the last close.  This is an attempt to combat the notorious
 419                  * NULL files problem which is particularly noticeable from a
 420                  * truncate down, buffered (re-)write (delalloc), followed by
 421                  * a crash.  What we are effectively doing here is
 422                  * significantly reducing the time window where we'd otherwise
 423                  * be exposed to that problem.
 424                  */
 425                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
 426                 if (truncated) {
 427                         xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
 428                         if (VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0)
 429                                 xfs_flush_pages(ip, 0, -1, XBF_ASYNC, FI_NONE);
 430                 }
 431         }
 432
 433         if (ip->i_d.di_nlink == 0)
 434                 return 0;
 435
 436         if ((S_ISREG(ip->i_d.di_mode) &&
 437              (VFS_I(ip)->i_size > 0 ||
 438               (VN_CACHED(VFS_I(ip)) > 0 || ip->i_delayed_blks > 0)) &&
 439              (ip->i_df.if_flags & XFS_IFEXTENTS))  &&
 440             (!(ip->i_d.di_flags & (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)))) {
 441
 442                 /*
 443                  * If we can't get the iolock just skip truncating the blocks
 444                  * past EOF because we could deadlock with the mmap_sem
 445                  * otherwise.  We'll get another chance to drop them once the
 446                  * last reference to the inode is dropped, so we'll never leak
 447                  * blocks permanently.
 448                  *
 449                  * Further, check if the inode is being opened, written and
 450                  * closed frequently and we have delayed allocation blocks
 451                  * outstanding (e.g. streaming writes from the NFS server),
 452                  * truncating the blocks past EOF will cause fragmentation to
 453                  * occur.
 454                  *
 455                  * In this case don't do the truncation, either, but we have to
 456                  * be careful how we detect this case. Blocks beyond EOF show
 457                  * up as i_delayed_blks even when the inode is clean, so we
 458                  * need to truncate them away first before checking for a dirty
 459                  * release. Hence on the first dirty close we will still remove
 460                  * the speculative allocation, but after that we will leave it
 461                  * in place.
 462                  */
 463                 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
 464                         return 0;
 465
 466                 error = xfs_free_eofblocks(mp, ip, true);
 467                 if (error)
 468                         return error;
 469
 470                 /* delalloc blocks after truncation means it really is dirty */
 471                 if (ip->i_delayed_blks)
 472                         xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
 473         }
 474         return 0;
 475 }
 476
 477 /*
 478  * xfs_inactive
 479  *
 480  * This is called when the vnode reference count for the vnode
 481  * goes to zero.  If the file has been unlinked, then it must
 482  * now be truncated.  Also, we clear all of the read-ahead state
 483  * kept for the inode here since the file is now closed.
 484  */
 485 int
 486 xfs_inactive(
 487         xfs_inode_t     *ip)
 488 {
 489         xfs_bmap_free_t free_list;
 490         xfs_fsblock_t   first_block;
 491         int             committed;
 492         xfs_trans_t     *tp;
 493         xfs_mount_t     *mp;
 494         int             error;
 495         int             truncate = 0;
 496
 497         /*
 498          * If the inode is already free, then there can be nothing
 499          * to clean up here.
 500          */
 501         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
 502                 ASSERT(ip->i_df.if_real_bytes == 0);
 503                 ASSERT(ip->i_df.if_broot_bytes == 0);
 504                 return VN_INACTIVE_CACHE;
 505         }
 506
 507         mp = ip->i_mount;
 508
 509         error = 0;
 510
 511         /* If this is a read-only mount, don't do this (would generate I/O) */
 512         if (mp->m_flags & XFS_MOUNT_RDONLY)
 513                 goto out;
 514
 515         if (ip->i_d.di_nlink != 0) {
 516                 if ((S_ISREG(ip->i_d.di_mode) &&
 517                     (VFS_I(ip)->i_size > 0 ||
 518                      (VN_CACHED(VFS_I(ip)) > 0 || ip->i_delayed_blks > 0)) &&
 519                     (ip->i_df.if_flags & XFS_IFEXTENTS) &&
 520                     (!(ip->i_d.di_flags &
 521                                 (XFS_DIFLAG_PREALLOC | XFS_DIFLAG_APPEND)) ||
 522                      ip->i_delayed_blks != 0))) {
 523                         error = xfs_free_eofblocks(mp, ip, false);
 524                         if (error)
 525                                 return VN_INACTIVE_CACHE;
 526                 }
 527                 goto out;
 528         }
 529
 530         if (S_ISREG(ip->i_d.di_mode) &&
 531             (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
 532              ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
 533                 truncate = 1;
 534
 535         error = xfs_qm_dqattach(ip, 0);
 536         if (error)
 537                 return VN_INACTIVE_CACHE;
 538
 539         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 540         error = xfs_trans_reserve(tp, 0,
 541                         (truncate || S_ISLNK(ip->i_d.di_mode)) ?
 542                                 XFS_ITRUNCATE_LOG_RES(mp) :
 543                                 XFS_IFREE_LOG_RES(mp),
 544                         0,
 545                         XFS_TRANS_PERM_LOG_RES,
 546                         XFS_ITRUNCATE_LOG_COUNT);
 547         if (error) {
 548                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 549                 xfs_trans_cancel(tp, 0);
 550                 return VN_INACTIVE_CACHE;
 551         }
 552
 553         xfs_ilock(ip, XFS_ILOCK_EXCL);
 554         xfs_trans_ijoin(tp, ip, 0);
 555
 556         if (S_ISLNK(ip->i_d.di_mode)) {
 557                 /*
 558                  * Zero length symlinks _can_ exist.
 559                  */
 560                 if (ip->i_d.di_size > XFS_IFORK_DSIZE(ip)) {
 561                         error = xfs_inactive_symlink_rmt(ip, &tp);
 562                         if (error)
 563                                 goto out_cancel;
 564                 } else if (ip->i_df.if_bytes > 0) {
 565                         xfs_idata_realloc(ip, -(ip->i_df.if_bytes),
 566                                           XFS_DATA_FORK);
 567                         ASSERT(ip->i_df.if_bytes == 0);
 568                 }
 569         } else if (truncate) {
 570                 ip->i_d.di_size = 0;
 571                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 572
 573                 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
 574                 if (error)
 575                         goto out_cancel;
 576
 577                 ASSERT(ip->i_d.di_nextents == 0);
 578         }
 579
 580         /*
 581          * If there are attributes associated with the file then blow them away
 582          * now.  The code calls a routine that recursively deconstructs the
 583          * attribute fork.  We need to just commit the current transaction
 584          * because we can't use it for xfs_attr_inactive().
 585          */
 586         if (ip->i_d.di_anextents > 0) {
 587                 ASSERT(ip->i_d.di_forkoff != 0);
 588
 589                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 590                 if (error)
 591                         goto out_unlock;
 592
 593                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 594
 595                 error = xfs_attr_inactive(ip);
 596                 if (error)
 597                         goto out;
 598
 599                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 600                 error = xfs_trans_reserve(tp, 0,
 601                                           XFS_IFREE_LOG_RES(mp),
 602                                           0, XFS_TRANS_PERM_LOG_RES,
 603                                           XFS_INACTIVE_LOG_COUNT);
 604                 if (error) {
 605                         xfs_trans_cancel(tp, 0);
 606                         goto out;
 607                 }
 608
 609                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 610                 xfs_trans_ijoin(tp, ip, 0);
 611         }
 612
 613         if (ip->i_afp)
 614                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
 615
 616         ASSERT(ip->i_d.di_anextents == 0);
 617
 618         /*
 619          * Free the inode.
 620          */
 621         xfs_bmap_init(&free_list, &first_block);
 622         error = xfs_ifree(tp, ip, &free_list);
 623         if (error) {
 624                 /*
 625                  * If we fail to free the inode, shut down.  The cancel
 626                  * might do that, we need to make sure.  Otherwise the
 627                  * inode might be lost for a long time or forever.
 628                  */
 629                 if (!XFS_FORCED_SHUTDOWN(mp)) {
 630                         xfs_notice(mp, "%s: xfs_ifree returned error %d",
 631                                 __func__, error);
 632                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
 633                 }
 634                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
 635         } else {
 636                 /*
 637                  * Credit the quota account(s). The inode is gone.
 638                  */
 639                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
 640
 641                 /*
 642                  * Just ignore errors at this point.  There is nothing we can
 643                  * do except to try to keep going. Make sure it's not a silent
 644                  * error.
 645                  */
 646                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
 647                 if (error)
 648                         xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
 649                                 __func__, error);
 650                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 651                 if (error)
 652                         xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
 653                                 __func__, error);
 654         }
 655
 656         /*
 657          * Release the dquots held by inode, if any.
 658          */
 659         xfs_qm_dqdetach(ip);
 660 out_unlock:
 661         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 662 out:
 663         return VN_INACTIVE_CACHE;
 664 out_cancel:
 665         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 666         goto out_unlock;
 667 }
 668
 669 /*
 670  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
 671  * is allowed, otherwise it has to be an exact match. If a CI match is found,
 672  * ci_name->name will point to a the actual name (caller must free) or
 673  * will be set to NULL if an exact match is found.
 674  */
 675 int
 676 xfs_lookup(
 677         xfs_inode_t             *dp,
 678         struct xfs_name         *name,
 679         xfs_inode_t             **ipp,
 680         struct xfs_name         *ci_name)
 681 {
 682         xfs_ino_t               inum;
 683         int                     error;
 684         uint                    lock_mode;
 685
 686         trace_xfs_lookup(dp, name);
 687
 688         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
 689                 return XFS_ERROR(EIO);
 690
 691         lock_mode = xfs_ilock_map_shared(dp);
 692         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
 693         xfs_iunlock_map_shared(dp, lock_mode);
 694
 695         if (error)
 696                 goto out;
 697
 698         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
 699         if (error)
 700                 goto out_free_name;
 701
 702         return 0;
 703
 704 out_free_name:
 705         if (ci_name)
 706                 kmem_free(ci_name->name);
 707 out:
 708         *ipp = NULL;
 709         return error;
 710 }
 711
 712 int
 713 xfs_create(
 714         xfs_inode_t             *dp,
 715         struct xfs_name         *name,
 716         umode_t                 mode,
 717         xfs_dev_t               rdev,
 718         xfs_inode_t             **ipp)
 719 {
 720         int                     is_dir = S_ISDIR(mode);
 721         struct xfs_mount        *mp = dp->i_mount;
 722         struct xfs_inode        *ip = NULL;
 723         struct xfs_trans        *tp = NULL;
 724         int                     error;
 725         xfs_bmap_free_t         free_list;
 726         xfs_fsblock_t           first_block;
 727         boolean_t               unlock_dp_on_error = B_FALSE;
 728         uint                    cancel_flags;
 729         int                     committed;
 730         prid_t                  prid;
 731         struct xfs_dquot        *udqp = NULL;
 732         struct xfs_dquot        *gdqp = NULL;
 733         uint                    resblks;
 734         uint                    log_res;
 735         uint                    log_count;
 736
 737         trace_xfs_create(dp, name);
 738
 739         if (XFS_FORCED_SHUTDOWN(mp))
 740                 return XFS_ERROR(EIO);
 741
 742         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
 743                 prid = xfs_get_projid(dp);
 744         else
 745                 prid = XFS_PROJID_DEFAULT;
 746
 747         /*
 748          * Make sure that we have allocated dquot(s) on disk.
 749          */
 750         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
 751                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
 752         if (error)
 753                 return error;
 754
 755         if (is_dir) {
 756                 rdev = 0;
 757                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
 758                 log_res = XFS_MKDIR_LOG_RES(mp);
 759                 log_count = XFS_MKDIR_LOG_COUNT;
 760                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
 761         } else {
 762                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
 763                 log_res = XFS_CREATE_LOG_RES(mp);
 764                 log_count = XFS_CREATE_LOG_COUNT;
 765                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
 766         }
 767
 768         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
 769
 770         /*
 771          * Initially assume that the file does not exist and
 772          * reserve the resources for that case.  If that is not
 773          * the case we'll drop the one we have and get a more
 774          * appropriate transaction later.
 775          */
 776         error = xfs_trans_reserve(tp, resblks, log_res, 0,
 777                         XFS_TRANS_PERM_LOG_RES, log_count);
 778         if (error == ENOSPC) {
 779                 /* flush outstanding delalloc blocks and retry */
 780                 xfs_flush_inodes(dp);
 781                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
 782                                 XFS_TRANS_PERM_LOG_RES, log_count);
 783         }
 784         if (error == ENOSPC) {
 785                 /* No space at all so try a "no-allocation" reservation */
 786                 resblks = 0;
 787                 error = xfs_trans_reserve(tp, 0, log_res, 0,
 788                                 XFS_TRANS_PERM_LOG_RES, log_count);
 789         }
 790         if (error) {
 791                 cancel_flags = 0;
 792                 goto out_trans_cancel;
 793         }
 794
 795         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
 796         unlock_dp_on_error = B_TRUE;
 797
 798         xfs_bmap_init(&free_list, &first_block);
 799
 800         /*
 801          * Reserve disk quota and the inode.
 802          */
 803         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
 804         if (error)
 805                 goto out_trans_cancel;
 806
 807         error = xfs_dir_canenter(tp, dp, name, resblks);
 808         if (error)
 809                 goto out_trans_cancel;
 810
 811         /*
 812          * A newly created regular or special file just has one directory
 813          * entry pointing to them, but a directory also the "." entry
 814          * pointing to itself.
 815          */
 816         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
 817                                prid, resblks > 0, &ip, &committed);
 818         if (error) {
 819                 if (error == ENOSPC)
 820                         goto out_trans_cancel;
 821                 goto out_trans_abort;
 822         }
 823
 824         /*
 825          * Now we join the directory inode to the transaction.  We do not do it
 826          * earlier because xfs_dir_ialloc might commit the previous transaction
 827          * (and release all the locks).  An error from here on will result in
 828          * the transaction cancel unlocking dp so don't do it explicitly in the
 829          * error path.
 830          */
 831         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
 832         unlock_dp_on_error = B_FALSE;
 833
 834         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
 835                                         &first_block, &free_list, resblks ?
 836                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
 837         if (error) {
 838                 ASSERT(error != ENOSPC);
 839                 goto out_trans_abort;
 840         }
 841         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 842         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
 843
 844         if (is_dir) {
 845                 error = xfs_dir_init(tp, ip, dp);
 846                 if (error)
 847                         goto out_bmap_cancel;
 848
 849                 error = xfs_bumplink(tp, dp);
 850                 if (error)
 851                         goto out_bmap_cancel;
 852         }
 853
 854         /*
 855          * If this is a synchronous mount, make sure that the
 856          * create transaction goes to disk before returning to
 857          * the user.
 858          */
 859         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
 860                 xfs_trans_set_sync(tp);
 861
 862         /*
 863          * Attach the dquot(s) to the inodes and modify them incore.
 864          * These ids of the inode couldn't have changed since the new
 865          * inode has been locked ever since it was created.
 866          */
 867         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
 868
 869         error = xfs_bmap_finish(&tp, &free_list, &committed);
 870         if (error)
 871                 goto out_bmap_cancel;
 872
 873         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 874         if (error)
 875                 goto out_release_inode;
 876
 877         xfs_qm_dqrele(udqp);
 878         xfs_qm_dqrele(gdqp);
 879
 880         *ipp = ip;
 881         return 0;
 882
 883  out_bmap_cancel:
 884         xfs_bmap_cancel(&free_list);
 885  out_trans_abort:
 886         cancel_flags |= XFS_TRANS_ABORT;
 887  out_trans_cancel:
 888         xfs_trans_cancel(tp, cancel_flags);
 889  out_release_inode:
 890         /*
 891          * Wait until after the current transaction is aborted to
 892          * release the inode.  This prevents recursive transactions
 893          * and deadlocks from xfs_inactive.
 894          */
 895         if (ip)
 896                 IRELE(ip);
 897
 898         xfs_qm_dqrele(udqp);
 899         xfs_qm_dqrele(gdqp);
 900
 901         if (unlock_dp_on_error)
 902                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
 903         return error;
 904 }
 905
 906 #ifdef DEBUG
 907 int xfs_locked_n;
 908 int xfs_small_retries;
 909 int xfs_middle_retries;
 910 int xfs_lots_retries;
 911 int xfs_lock_delays;
 912 #endif
 913
 914 /*
 915  * Bump the subclass so xfs_lock_inodes() acquires each lock with
 916  * a different value
 917  */
 918 static inline int
 919 xfs_lock_inumorder(int lock_mode, int subclass)
 920 {
 921         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
 922                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
 923         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
 924                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
 925
 926         return lock_mode;
 927 }
 928
 929 /*
 930  * The following routine will lock n inodes in exclusive mode.
 931  * We assume the caller calls us with the inodes in i_ino order.
 932  *
 933  * We need to detect deadlock where an inode that we lock
 934  * is in the AIL and we start waiting for another inode that is locked
 935  * by a thread in a long running transaction (such as truncate). This can
 936  * result in deadlock since the long running trans might need to wait
 937  * for the inode we just locked in order to push the tail and free space
 938  * in the log.
 939  */
 940 void
 941 xfs_lock_inodes(
 942         xfs_inode_t     **ips,
 943         int             inodes,
 944         uint            lock_mode)
 945 {
 946         int             attempts = 0, i, j, try_lock;
 947         xfs_log_item_t  *lp;
 948
 949         ASSERT(ips && (inodes >= 2)); /* we need at least two */
 950
 951         try_lock = 0;
 952         i = 0;
 953
 954 again:
 955         for (; i < inodes; i++) {
 956                 ASSERT(ips[i]);
 957
 958                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
 959                         continue;
 960
 961                 /*
 962                  * If try_lock is not set yet, make sure all locked inodes
 963                  * are not in the AIL.
 964                  * If any are, set try_lock to be used later.
 965                  */
 966
 967                 if (!try_lock) {
 968                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
 969                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
 970                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
 971                                         try_lock++;
 972                                 }
 973                         }
 974                 }
 975
 976                 /*
 977                  * If any of the previous locks we have locked is in the AIL,
 978                  * we must TRY to get the second and subsequent locks. If
 979                  * we can't get any, we must release all we have
 980                  * and try again.
 981                  */
 982
 983                 if (try_lock) {
 984                         /* try_lock must be 0 if i is 0. */
 985                         /*
 986                          * try_lock means we have an inode locked
 987                          * that is in the AIL.
 988                          */
 989                         ASSERT(i != 0);
 990                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
 991                                 attempts++;
 992
 993                                 /*
 994                                  * Unlock all previous guys and try again.
 995                                  * xfs_iunlock will try to push the tail
 996                                  * if the inode is in the AIL.
 997                                  */
 998
 999                                 for(j = i - 1; j >= 0; j--) {
1000
1001                                         /*
1002                                          * Check to see if we've already
1003                                          * unlocked this one.
1004                                          * Not the first one going back,
1005                                          * and the inode ptr is the same.
1006                                          */
1007                                         if ((j != (i - 1)) && ips[j] ==
1008                                                                 ips[j+1])
1009                                                 continue;
1010
1011                                         xfs_iunlock(ips[j], lock_mode);
1012                                 }
1013
1014                                 if ((attempts % 5) == 0) {
1015                                         delay(1); /* Don't just spin the CPU */
1016 #ifdef DEBUG
1017                                         xfs_lock_delays++;
1018 #endif
1019                                 }
1020                                 i = 0;
1021                                 try_lock = 0;
1022                                 goto again;
1023                         }
1024                 } else {
1025                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
1026                 }
1027         }
1028
1029 #ifdef DEBUG
1030         if (attempts) {
1031                 if (attempts < 5) xfs_small_retries++;
1032                 else if (attempts < 100) xfs_middle_retries++;
1033                 else xfs_lots_retries++;
1034         } else {
1035                 xfs_locked_n++;
1036         }
1037 #endif
1038 }
1039
1040 /*
1041  * xfs_lock_two_inodes() can only be used to lock one type of lock
1042  * at a time - the iolock or the ilock, but not both at once. If
1043  * we lock both at once, lockdep will report false positives saying
1044  * we have violated locking orders.
1045  */
1046 void
1047 xfs_lock_two_inodes(
1048         xfs_inode_t             *ip0,
1049         xfs_inode_t             *ip1,
1050         uint                    lock_mode)
1051 {
1052         xfs_inode_t             *temp;
1053         int                     attempts = 0;
1054         xfs_log_item_t          *lp;
1055
1056         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
1057                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
1058         ASSERT(ip0->i_ino != ip1->i_ino);
1059
1060         if (ip0->i_ino > ip1->i_ino) {
1061                 temp = ip0;
1062                 ip0 = ip1;
1063                 ip1 = temp;
1064         }
1065
1066  again:
1067         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
1068
1069         /*
1070          * If the first lock we have locked is in the AIL, we must TRY to get
1071          * the second lock. If we can't get it, we must release the first one
1072          * and try again.
1073          */
1074         lp = (xfs_log_item_t *)ip0->i_itemp;
1075         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
1076                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
1077                         xfs_iunlock(ip0, lock_mode);
1078                         if ((++attempts % 5) == 0)
1079                                 delay(1); /* Don't just spin the CPU */
1080                         goto again;
1081                 }
1082         } else {
1083                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
1084         }
1085 }
1086
1087 int
1088 xfs_remove(
1089         xfs_inode_t             *dp,
1090         struct xfs_name         *name,
1091         xfs_inode_t             *ip)
1092 {
1093         xfs_mount_t             *mp = dp->i_mount;
1094         xfs_trans_t             *tp = NULL;
1095         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
1096         int                     error = 0;
1097         xfs_bmap_free_t         free_list;
1098         xfs_fsblock_t           first_block;
1099         int                     cancel_flags;
1100         int                     committed;
1101         int                     link_zero;
1102         uint                    resblks;
1103         uint                    log_count;
1104
1105         trace_xfs_remove(dp, name);
1106
1107         if (XFS_FORCED_SHUTDOWN(mp))
1108                 return XFS_ERROR(EIO);
1109
1110         error = xfs_qm_dqattach(dp, 0);
1111         if (error)
1112                 goto std_return;
1113
1114         error = xfs_qm_dqattach(ip, 0);
1115         if (error)
1116                 goto std_return;
1117
1118         if (is_dir) {
1119                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
1120                 log_count = XFS_DEFAULT_LOG_COUNT;
1121         } else {
1122                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
1123                 log_count = XFS_REMOVE_LOG_COUNT;
1124         }
1125         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1126
1127         /*
1128          * We try to get the real space reservation first,
1129          * allowing for directory btree deletion(s) implying
1130          * possible bmap insert(s).  If we can't get the space
1131          * reservation then we use 0 instead, and avoid the bmap
1132          * btree insert(s) in the directory code by, if the bmap
1133          * insert tries to happen, instead trimming the LAST
1134          * block from the directory.
1135          */
1136         resblks = XFS_REMOVE_SPACE_RES(mp);
1137         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
1138                                   XFS_TRANS_PERM_LOG_RES, log_count);
1139         if (error == ENOSPC) {
1140                 resblks = 0;
1141                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
1142                                           XFS_TRANS_PERM_LOG_RES, log_count);
1143         }
1144         if (error) {
1145                 ASSERT(error != ENOSPC);
1146                 cancel_flags = 0;
1147                 goto out_trans_cancel;
1148         }
1149
1150         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
1151
1152         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1153         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1154
1155         /*
1156          * If we're removing a directory perform some additional validation.
1157          */
1158         if (is_dir) {
1159                 ASSERT(ip->i_d.di_nlink >= 2);
1160                 if (ip->i_d.di_nlink != 2) {
1161                         error = XFS_ERROR(ENOTEMPTY);
1162                         goto out_trans_cancel;
1163                 }
1164                 if (!xfs_dir_isempty(ip)) {
1165                         error = XFS_ERROR(ENOTEMPTY);
1166                         goto out_trans_cancel;
1167                 }
1168         }
1169
1170         xfs_bmap_init(&free_list, &first_block);
1171         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
1172                                         &first_block, &free_list, resblks);
1173         if (error) {
1174                 ASSERT(error != ENOENT);
1175                 goto out_bmap_cancel;
1176         }
1177         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1178
1179         if (is_dir) {
1180                 /*
1181                  * Drop the link from ip's "..".
1182                  */
1183                 error = xfs_droplink(tp, dp);
1184                 if (error)
1185                         goto out_bmap_cancel;
1186
1187                 /*
1188                  * Drop the "." link from ip to self.
1189                  */
1190                 error = xfs_droplink(tp, ip);
1191                 if (error)
1192                         goto out_bmap_cancel;
1193         } else {
1194                 /*
1195                  * When removing a non-directory we need to log the parent
1196                  * inode here.  For a directory this is done implicitly
1197                  * by the xfs_droplink call for the ".." entry.
1198                  */
1199                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1200         }
1201
1202         /*
1203          * Drop the link from dp to ip.
1204          */
1205         error = xfs_droplink(tp, ip);
1206         if (error)
1207                 goto out_bmap_cancel;
1208
1209         /*
1210          * Determine if this is the last link while
1211          * we are in the transaction.
1212          */
1213         link_zero = (ip->i_d.di_nlink == 0);
1214
1215         /*
1216          * If this is a synchronous mount, make sure that the
1217          * remove transaction goes to disk before returning to
1218          * the user.
1219          */
1220         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1221                 xfs_trans_set_sync(tp);
1222
1223         error = xfs_bmap_finish(&tp, &free_list, &committed);
1224         if (error)
1225                 goto out_bmap_cancel;
1226
1227         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1228         if (error)
1229                 goto std_return;
1230
1231         /*
1232          * If we are using filestreams, kill the stream association.
1233          * If the file is still open it may get a new one but that
1234          * will get killed on last close in xfs_close() so we don't
1235          * have to worry about that.
1236          */
1237         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1238                 xfs_filestream_deassociate(ip);
1239
1240         return 0;
1241
1242  out_bmap_cancel:
1243         xfs_bmap_cancel(&free_list);
1244         cancel_flags |= XFS_TRANS_ABORT;
1245  out_trans_cancel:
1246         xfs_trans_cancel(tp, cancel_flags);
1247  std_return:
1248         return error;
1249 }
1250
1251 int
1252 xfs_link(
1253         xfs_inode_t             *tdp,
1254         xfs_inode_t             *sip,
1255         struct xfs_name         *target_name)
1256 {
1257         xfs_mount_t             *mp = tdp->i_mount;
1258         xfs_trans_t             *tp;
1259         int                     error;
1260         xfs_bmap_free_t         free_list;
1261         xfs_fsblock_t           first_block;
1262         int                     cancel_flags;
1263         int                     committed;
1264         int                     resblks;
1265
1266         trace_xfs_link(tdp, target_name);
1267
1268         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1269
1270         if (XFS_FORCED_SHUTDOWN(mp))
1271                 return XFS_ERROR(EIO);
1272
1273         error = xfs_qm_dqattach(sip, 0);
1274         if (error)
1275                 goto std_return;
1276
1277         error = xfs_qm_dqattach(tdp, 0);
1278         if (error)
1279                 goto std_return;
1280
1281         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1282         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1283         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1284         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1285                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1286         if (error == ENOSPC) {
1287                 resblks = 0;
1288                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1289                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1290         }
1291         if (error) {
1292                 cancel_flags = 0;
1293                 goto error_return;
1294         }
1295
1296         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1297
1298         xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1299         xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1300
1301         /*
1302          * If we are using project inheritance, we only allow hard link
1303          * creation in our tree when the project IDs are the same; else
1304          * the tree quota mechanism could be circumvented.
1305          */
1306         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1307                      (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1308                 error = XFS_ERROR(EXDEV);
1309                 goto error_return;
1310         }
1311
1312         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1313         if (error)
1314                 goto error_return;
1315
1316         xfs_bmap_init(&free_list, &first_block);
1317
1318         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1319                                         &first_block, &free_list, resblks);
1320         if (error)
1321                 goto abort_return;
1322         xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1323         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1324
1325         error = xfs_bumplink(tp, sip);
1326         if (error)
1327                 goto abort_return;
1328
1329         /*
1330          * If this is a synchronous mount, make sure that the
1331          * link transaction goes to disk before returning to
1332          * the user.
1333          */
1334         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1335                 xfs_trans_set_sync(tp);
1336         }
1337
1338         error = xfs_bmap_finish (&tp, &free_list, &committed);
1339         if (error) {
1340                 xfs_bmap_cancel(&free_list);
1341                 goto abort_return;
1342         }
1343
1344         return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1345
1346  abort_return:
1347         cancel_flags |= XFS_TRANS_ABORT;
1348  error_return:
1349         xfs_trans_cancel(tp, cancel_flags);
1350  std_return:
1351         return error;
1352 }
1353
1354 int
1355 xfs_symlink(
1356         xfs_inode_t             *dp,
1357         struct xfs_name         *link_name,
1358         const char              *target_path,
1359         umode_t                 mode,
1360         xfs_inode_t             **ipp)
1361 {
1362         xfs_mount_t             *mp = dp->i_mount;
1363         xfs_trans_t             *tp;
1364         xfs_inode_t             *ip;
1365         int                     error;
1366         int                     pathlen;
1367         xfs_bmap_free_t         free_list;
1368         xfs_fsblock_t           first_block;
1369         boolean_t               unlock_dp_on_error = B_FALSE;
1370         uint                    cancel_flags;
1371         int                     committed;
1372         xfs_fileoff_t           first_fsb;
1373         xfs_filblks_t           fs_blocks;
1374         int                     nmaps;
1375         xfs_bmbt_irec_t         mval[SYMLINK_MAPS];
1376         xfs_daddr_t             d;
1377         const char              *cur_chunk;
1378         int                     byte_cnt;
1379         int                     n;
1380         xfs_buf_t               *bp;
1381         prid_t                  prid;
1382         struct xfs_dquot        *udqp, *gdqp;
1383         uint                    resblks;
1384
1385         *ipp = NULL;
1386         error = 0;
1387         ip = NULL;
1388         tp = NULL;
1389
1390         trace_xfs_symlink(dp, link_name);
1391
1392         if (XFS_FORCED_SHUTDOWN(mp))
1393                 return XFS_ERROR(EIO);
1394
1395         /*
1396          * Check component lengths of the target path name.
1397          */
1398         pathlen = strlen(target_path);
1399         if (pathlen >= MAXPATHLEN)      /* total string too long */
1400                 return XFS_ERROR(ENAMETOOLONG);
1401
1402         udqp = gdqp = NULL;
1403         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
1404                 prid = xfs_get_projid(dp);
1405         else
1406                 prid = XFS_PROJID_DEFAULT;
1407
1408         /*
1409          * Make sure that we have allocated dquot(s) on disk.
1410          */
1411         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
1412                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT, &udqp, &gdqp);
1413         if (error)
1414                 goto std_return;
1415
1416         tp = xfs_trans_alloc(mp, XFS_TRANS_SYMLINK);
1417         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1418         /*
1419          * The symlink will fit into the inode data fork?
1420          * There can't be any attributes so we get the whole variable part.
1421          */
1422         if (pathlen <= XFS_LITINO(mp))
1423                 fs_blocks = 0;
1424         else
1425                 fs_blocks = XFS_B_TO_FSB(mp, pathlen);
1426         resblks = XFS_SYMLINK_SPACE_RES(mp, link_name->len, fs_blocks);
1427         error = xfs_trans_reserve(tp, resblks, XFS_SYMLINK_LOG_RES(mp), 0,
1428                         XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1429         if (error == ENOSPC && fs_blocks == 0) {
1430                 resblks = 0;
1431                 error = xfs_trans_reserve(tp, 0, XFS_SYMLINK_LOG_RES(mp), 0,
1432                                 XFS_TRANS_PERM_LOG_RES, XFS_SYMLINK_LOG_COUNT);
1433         }
1434         if (error) {
1435                 cancel_flags = 0;
1436                 goto error_return;
1437         }
1438
1439         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1440         unlock_dp_on_error = B_TRUE;
1441
1442         /*
1443          * Check whether the directory allows new symlinks or not.
1444          */
1445         if (dp->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) {
1446                 error = XFS_ERROR(EPERM);
1447                 goto error_return;
1448         }
1449
1450         /*
1451          * Reserve disk quota : blocks and inode.
1452          */
1453         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp, resblks, 1, 0);
1454         if (error)
1455                 goto error_return;
1456
1457         /*
1458          * Check for ability to enter directory entry, if no space reserved.
1459          */
1460         error = xfs_dir_canenter(tp, dp, link_name, resblks);
1461         if (error)
1462                 goto error_return;
1463         /*
1464          * Initialize the bmap freelist prior to calling either
1465          * bmapi or the directory create code.
1466          */
1467         xfs_bmap_init(&free_list, &first_block);
1468
1469         /*
1470          * Allocate an inode for the symlink.
1471          */
1472         error = xfs_dir_ialloc(&tp, dp, S_IFLNK | (mode & ~S_IFMT), 1, 0,
1473                                prid, resblks > 0, &ip, NULL);
1474         if (error) {
1475                 if (error == ENOSPC)
1476                         goto error_return;
1477                 goto error1;
1478         }
1479
1480         /*
1481          * An error after we've joined dp to the transaction will result in the
1482          * transaction cancel unlocking dp so don't do it explicitly in the
1483          * error path.
1484          */
1485         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1486         unlock_dp_on_error = B_FALSE;
1487
1488         /*
1489          * Also attach the dquot(s) to it, if applicable.
1490          */
1491         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp);
1492
1493         if (resblks)
1494                 resblks -= XFS_IALLOC_SPACE_RES(mp);
1495         /*
1496          * If the symlink will fit into the inode, write it inline.
1497          */
1498         if (pathlen <= XFS_IFORK_DSIZE(ip)) {
1499                 xfs_idata_realloc(ip, pathlen, XFS_DATA_FORK);
1500                 memcpy(ip->i_df.if_u1.if_data, target_path, pathlen);
1501                 ip->i_d.di_size = pathlen;
1502
1503                 /*
1504                  * The inode was initially created in extent format.
1505                  */
1506                 ip->i_df.if_flags &= ~(XFS_IFEXTENTS | XFS_IFBROOT);
1507                 ip->i_df.if_flags |= XFS_IFINLINE;
1508
1509                 ip->i_d.di_format = XFS_DINODE_FMT_LOCAL;
1510                 xfs_trans_log_inode(tp, ip, XFS_ILOG_DDATA | XFS_ILOG_CORE);
1511
1512         } else {
1513                 first_fsb = 0;
1514                 nmaps = SYMLINK_MAPS;
1515
1516                 error = xfs_bmapi_write(tp, ip, first_fsb, fs_blocks,
1517                                   XFS_BMAPI_METADATA, &first_block, resblks,
1518                                   mval, &nmaps, &free_list);
1519                 if (error)
1520                         goto error2;
1521
1522                 if (resblks)
1523                         resblks -= fs_blocks;
1524                 ip->i_d.di_size = pathlen;
1525                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1526
1527                 cur_chunk = target_path;
1528                 for (n = 0; n < nmaps; n++) {
1529                         d = XFS_FSB_TO_DADDR(mp, mval[n].br_startblock);
1530                         byte_cnt = XFS_FSB_TO_B(mp, mval[n].br_blockcount);
1531                         bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, d,
1532                                                BTOBB(byte_cnt), 0);
1533                         if (!bp) {
1534                                 error = ENOMEM;
1535                                 goto error2;
1536                         }
1537                         if (pathlen < byte_cnt) {
1538                                 byte_cnt = pathlen;
1539                         }
1540                         pathlen -= byte_cnt;
1541
1542                         memcpy(bp->b_addr, cur_chunk, byte_cnt);
1543                         cur_chunk += byte_cnt;
1544
1545                         xfs_trans_log_buf(tp, bp, 0, byte_cnt - 1);
1546                 }
1547         }
1548
1549         /*
1550          * Create the directory entry for the symlink.
1551          */
1552         error = xfs_dir_createname(tp, dp, link_name, ip->i_ino,
1553                                         &first_block, &free_list, resblks);
1554         if (error)
1555                 goto error2;
1556         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1557         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1558
1559         /*
1560          * If this is a synchronous mount, make sure that the
1561          * symlink transaction goes to disk before returning to
1562          * the user.
1563          */
1564         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1565                 xfs_trans_set_sync(tp);
1566         }
1567
1568         error = xfs_bmap_finish(&tp, &free_list, &committed);
1569         if (error) {
1570                 goto error2;
1571         }
1572         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1573         xfs_qm_dqrele(udqp);
1574         xfs_qm_dqrele(gdqp);
1575
1576         *ipp = ip;
1577         return 0;
1578
1579  error2:
1580         IRELE(ip);
1581  error1:
1582         xfs_bmap_cancel(&free_list);
1583         cancel_flags |= XFS_TRANS_ABORT;
1584  error_return:
1585         xfs_trans_cancel(tp, cancel_flags);
1586         xfs_qm_dqrele(udqp);
1587         xfs_qm_dqrele(gdqp);
1588
1589         if (unlock_dp_on_error)
1590                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1591  std_return:
1592         return error;
1593 }
1594
1595 int
1596 xfs_set_dmattrs(
1597         xfs_inode_t     *ip,
1598         u_int           evmask,
1599         u_int16_t       state)
1600 {
1601         xfs_mount_t     *mp = ip->i_mount;
1602         xfs_trans_t     *tp;
1603         int             error;
1604
1605         if (!capable(CAP_SYS_ADMIN))
1606                 return XFS_ERROR(EPERM);
1607
1608         if (XFS_FORCED_SHUTDOWN(mp))
1609                 return XFS_ERROR(EIO);
1610
1611         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
1612         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
1613         if (error) {
1614                 xfs_trans_cancel(tp, 0);
1615                 return error;
1616         }
1617         xfs_ilock(ip, XFS_ILOCK_EXCL);
1618         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1619
1620         ip->i_d.di_dmevmask = evmask;
1621         ip->i_d.di_dmstate  = state;
1622
1623         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1624         error = xfs_trans_commit(tp, 0);
1625
1626         return error;
1627 }
1628
1629 /*
1630  * xfs_alloc_file_space()
1631  *      This routine allocates disk space for the given file.
1632  *
1633  *      If alloc_type == 0, this request is for an ALLOCSP type
1634  *      request which will change the file size.  In this case, no
1635  *      DMAPI event will be generated by the call.  A TRUNCATE event
1636  *      will be generated later by xfs_setattr.
1637  *
1638  *      If alloc_type != 0, this request is for a RESVSP type
1639  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
1640  *      lower block boundary byte address is less than the file's
1641  *      length.
1642  *
1643  * RETURNS:
1644  *       0 on success
1645  *      errno on error
1646  *
1647  */
1648 STATIC int
1649 xfs_alloc_file_space(
1650         xfs_inode_t             *ip,
1651         xfs_off_t               offset,
1652         xfs_off_t               len,
1653         int                     alloc_type,
1654         int                     attr_flags)
1655 {
1656         xfs_mount_t             *mp = ip->i_mount;
1657         xfs_off_t               count;
1658         xfs_filblks_t           allocated_fsb;
1659         xfs_filblks_t           allocatesize_fsb;
1660         xfs_extlen_t            extsz, temp;
1661         xfs_fileoff_t           startoffset_fsb;
1662         xfs_fsblock_t           firstfsb;
1663         int                     nimaps;
1664         int                     quota_flag;
1665         int                     rt;
1666         xfs_trans_t             *tp;
1667         xfs_bmbt_irec_t         imaps[1], *imapp;
1668         xfs_bmap_free_t         free_list;
1669         uint                    qblocks, resblks, resrtextents;
1670         int                     committed;
1671         int                     error;
1672
1673         trace_xfs_alloc_file_space(ip);
1674
1675         if (XFS_FORCED_SHUTDOWN(mp))
1676                 return XFS_ERROR(EIO);
1677
1678         error = xfs_qm_dqattach(ip, 0);
1679         if (error)
1680                 return error;
1681
1682         if (len <= 0)
1683                 return XFS_ERROR(EINVAL);
1684
1685         rt = XFS_IS_REALTIME_INODE(ip);
1686         extsz = xfs_get_extsz_hint(ip);
1687
1688         count = len;
1689         imapp = &imaps[0];
1690         nimaps = 1;
1691         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
1692         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
1693
1694         /*
1695          * Allocate file space until done or until there is an error
1696          */
1697         while (allocatesize_fsb && !error) {
1698                 xfs_fileoff_t   s, e;
1699
1700                 /*
1701                  * Determine space reservations for data/realtime.
1702                  */
1703                 if (unlikely(extsz)) {
1704                         s = startoffset_fsb;
1705                         do_div(s, extsz);
1706                         s *= extsz;
1707                         e = startoffset_fsb + allocatesize_fsb;
1708                         if ((temp = do_mod(startoffset_fsb, extsz)))
1709                                 e += temp;
1710                         if ((temp = do_mod(e, extsz)))
1711                                 e += extsz - temp;
1712                 } else {
1713                         s = 0;
1714                         e = allocatesize_fsb;
1715                 }
1716
1717                 /*
1718                  * The transaction reservation is limited to a 32-bit block
1719                  * count, hence we need to limit the number of blocks we are
1720                  * trying to reserve to avoid an overflow. We can't allocate
1721                  * more than @nimaps extents, and an extent is limited on disk
1722                  * to MAXEXTLEN (21 bits), so use that to enforce the limit.
1723                  */
1724                 resblks = min_t(xfs_fileoff_t, (e - s), (MAXEXTLEN * nimaps));
1725                 if (unlikely(rt)) {
1726                         resrtextents = qblocks = resblks;
1727                         resrtextents /= mp->m_sb.sb_rextsize;
1728                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
1729                         quota_flag = XFS_QMOPT_RES_RTBLKS;
1730                 } else {
1731                         resrtextents = 0;
1732                         resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resblks);
1733                         quota_flag = XFS_QMOPT_RES_REGBLKS;
1734                 }
1735
1736                 /*
1737                  * Allocate and setup the transaction.
1738                  */
1739                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
1740                 error = xfs_trans_reserve(tp, resblks,
1741                                           XFS_WRITE_LOG_RES(mp), resrtextents,
1742                                           XFS_TRANS_PERM_LOG_RES,
1743                                           XFS_WRITE_LOG_COUNT);
1744                 /*
1745                  * Check for running out of space
1746                  */
1747                 if (error) {
1748                         /*
1749                          * Free the transaction structure.
1750                          */
1751                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
1752                         xfs_trans_cancel(tp, 0);
1753                         break;
1754                 }
1755                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1756                 error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks,
1757                                                       0, quota_flag);
1758                 if (error)
1759                         goto error1;
1760
1761                 xfs_trans_ijoin(tp, ip, 0);
1762
1763                 xfs_bmap_init(&free_list, &firstfsb);
1764                 error = xfs_bmapi_write(tp, ip, startoffset_fsb,
1765                                         allocatesize_fsb, alloc_type, &firstfsb,
1766                                         0, imapp, &nimaps, &free_list);
1767                 if (error) {
1768                         goto error0;
1769                 }
1770
1771                 /*
1772                  * Complete the transaction
1773                  */
1774                 error = xfs_bmap_finish(&tp, &free_list, &committed);
1775                 if (error) {
1776                         goto error0;
1777                 }
1778
1779                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1780                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1781                 if (error) {
1782                         break;
1783                 }
1784
1785                 allocated_fsb = imapp->br_blockcount;
1786
1787                 if (nimaps == 0) {
1788                         error = XFS_ERROR(ENOSPC);
1789                         break;
1790                 }
1791
1792                 startoffset_fsb += allocated_fsb;
1793                 allocatesize_fsb -= allocated_fsb;
1794         }
1795
1796         return error;
1797
1798 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
1799         xfs_bmap_cancel(&free_list);
1800         xfs_trans_unreserve_quota_nblks(tp, ip, (long)qblocks, 0, quota_flag);
1801
1802 error1: /* Just cancel transaction */
1803         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1804         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1805         return error;
1806 }
1807
1808 /*
1809  * Zero file bytes between startoff and endoff inclusive.
1810  * The iolock is held exclusive and no blocks are buffered.
1811  *
1812  * This function is used by xfs_free_file_space() to zero
1813  * partial blocks when the range to free is not block aligned.
1814  * When unreserving space with boundaries that are not block
1815  * aligned we round up the start and round down the end
1816  * boundaries and then use this function to zero the parts of
1817  * the blocks that got dropped during the rounding.
1818  */
1819 STATIC int
1820 xfs_zero_remaining_bytes(
1821         xfs_inode_t             *ip,
1822         xfs_off_t               startoff,
1823         xfs_off_t               endoff)
1824 {
1825         xfs_bmbt_irec_t         imap;
1826         xfs_fileoff_t           offset_fsb;
1827         xfs_off_t               lastoffset;
1828         xfs_off_t               offset;
1829         xfs_buf_t               *bp;
1830         xfs_mount_t             *mp = ip->i_mount;
1831         int                     nimap;
1832         int                     error = 0;
1833
1834         /*
1835          * Avoid doing I/O beyond eof - it's not necessary
1836          * since nothing can read beyond eof.  The space will
1837          * be zeroed when the file is extended anyway.
1838          */
1839         if (startoff >= XFS_ISIZE(ip))
1840                 return 0;
1841
1842         if (endoff > XFS_ISIZE(ip))
1843                 endoff = XFS_ISIZE(ip);
1844
1845         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
1846                                         mp->m_rtdev_targp : mp->m_ddev_targp,
1847                                   BTOBB(mp->m_sb.sb_blocksize), 0);
1848         if (!bp)
1849                 return XFS_ERROR(ENOMEM);
1850
1851         xfs_buf_unlock(bp);
1852
1853         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
1854                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1855                 nimap = 1;
1856                 error = xfs_bmapi_read(ip, offset_fsb, 1, &imap, &nimap, 0);
1857                 if (error || nimap < 1)
1858                         break;
1859                 ASSERT(imap.br_blockcount >= 1);
1860                 ASSERT(imap.br_startoff == offset_fsb);
1861                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
1862                 if (lastoffset > endoff)
1863                         lastoffset = endoff;
1864                 if (imap.br_startblock == HOLESTARTBLOCK)
1865                         continue;
1866                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1867                 if (imap.br_state == XFS_EXT_UNWRITTEN)
1868                         continue;
1869                 XFS_BUF_UNDONE(bp);
1870                 XFS_BUF_UNWRITE(bp);
1871                 XFS_BUF_READ(bp);
1872                 XFS_BUF_SET_ADDR(bp, xfs_fsb_to_db(ip, imap.br_startblock));
1873                 xfsbdstrat(mp, bp);
1874                 error = xfs_buf_iowait(bp);
1875                 if (error) {
1876                         xfs_buf_ioerror_alert(bp,
1877                                         "xfs_zero_remaining_bytes(read)");
1878                         break;
1879                 }
1880                 memset(bp->b_addr +
1881                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
1882                       0, lastoffset - offset + 1);
1883                 XFS_BUF_UNDONE(bp);
1884                 XFS_BUF_UNREAD(bp);
1885                 XFS_BUF_WRITE(bp);
1886                 xfsbdstrat(mp, bp);
1887                 error = xfs_buf_iowait(bp);
1888                 if (error) {
1889                         xfs_buf_ioerror_alert(bp,
1890                                         "xfs_zero_remaining_bytes(write)");
1891                         break;
1892                 }
1893         }
1894         xfs_buf_free(bp);
1895         return error;
1896 }
1897
1898 /*
1899  * xfs_free_file_space()
1900  *      This routine frees disk space for the given file.
1901  *
1902  *      This routine is only called by xfs_change_file_space
1903  *      for an UNRESVSP type call.
1904  *
1905  * RETURNS:
1906  *       0 on success
1907  *      errno on error
1908  *
1909  */
1910 STATIC int
1911 xfs_free_file_space(
1912         xfs_inode_t             *ip,
1913         xfs_off_t               offset,
1914         xfs_off_t               len,
1915         int                     attr_flags)
1916 {
1917         int                     committed;
1918         int                     done;
1919         xfs_fileoff_t           endoffset_fsb;
1920         int                     error;
1921         xfs_fsblock_t           firstfsb;
1922         xfs_bmap_free_t         free_list;
1923         xfs_bmbt_irec_t         imap;
1924         xfs_off_t               ioffset;
1925         xfs_extlen_t            mod=0;
1926         xfs_mount_t             *mp;
1927         int                     nimap;
1928         uint                    resblks;
1929         uint                    rounding;
1930         int                     rt;
1931         xfs_fileoff_t           startoffset_fsb;
1932         xfs_trans_t             *tp;
1933         int                     need_iolock = 1;
1934
1935         mp = ip->i_mount;
1936
1937         trace_xfs_free_file_space(ip);
1938
1939         error = xfs_qm_dqattach(ip, 0);
1940         if (error)
1941                 return error;
1942
1943         error = 0;
1944         if (len <= 0)   /* if nothing being freed */
1945                 return error;
1946         rt = XFS_IS_REALTIME_INODE(ip);
1947         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
1948         endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
1949
1950         if (attr_flags & XFS_ATTR_NOLOCK)
1951                 need_iolock = 0;
1952         if (need_iolock) {
1953                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1954                 /* wait for the completion of any pending DIOs */
1955                 inode_dio_wait(VFS_I(ip));
1956         }
1957
1958         rounding = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
1959         ioffset = offset & ~(rounding - 1);
1960
1961         if (VN_CACHED(VFS_I(ip)) != 0) {
1962                 error = xfs_flushinval_pages(ip, ioffset, -1, FI_REMAPF_LOCKED);
1963                 if (error)
1964                         goto out_unlock_iolock;
1965         }
1966
1967         /*
1968          * Need to zero the stuff we're not freeing, on disk.
1969          * If it's a realtime file & can't use unwritten extents then we
1970          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
1971          * will take care of it for us.
1972          */
1973         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
1974                 nimap = 1;
1975                 error = xfs_bmapi_read(ip, startoffset_fsb, 1,
1976                                         &imap, &nimap, 0);
1977                 if (error)
1978                         goto out_unlock_iolock;
1979                 ASSERT(nimap == 0 || nimap == 1);
1980                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
1981                         xfs_daddr_t     block;
1982
1983                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1984                         block = imap.br_startblock;
1985                         mod = do_div(block, mp->m_sb.sb_rextsize);
1986                         if (mod)
1987                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
1988                 }
1989                 nimap = 1;
1990                 error = xfs_bmapi_read(ip, endoffset_fsb - 1, 1,
1991                                         &imap, &nimap, 0);
1992                 if (error)
1993                         goto out_unlock_iolock;
1994                 ASSERT(nimap == 0 || nimap == 1);
1995                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
1996                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1997                         mod++;
1998                         if (mod && (mod != mp->m_sb.sb_rextsize))
1999                                 endoffset_fsb -= mod;
2000                 }
2001         }
2002         if ((done = (endoffset_fsb <= startoffset_fsb)))
2003                 /*
2004                  * One contiguous piece to clear
2005                  */
2006                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
2007         else {
2008                 /*
2009                  * Some full blocks, possibly two pieces to clear
2010                  */
2011                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
2012                         error = xfs_zero_remaining_bytes(ip, offset,
2013                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
2014                 if (!error &&
2015                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
2016                         error = xfs_zero_remaining_bytes(ip,
2017                                 XFS_FSB_TO_B(mp, endoffset_fsb),
2018                                 offset + len - 1);
2019         }
2020
2021         /*
2022          * free file space until done or until there is an error
2023          */
2024         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
2025         while (!error && !done) {
2026
2027                 /*
2028                  * allocate and setup the transaction. Allow this
2029                  * transaction to dip into the reserve blocks to ensure
2030                  * the freeing of the space succeeds at ENOSPC.
2031                  */
2032                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
2033                 tp->t_flags |= XFS_TRANS_RESERVE;
2034                 error = xfs_trans_reserve(tp,
2035                                           resblks,
2036                                           XFS_WRITE_LOG_RES(mp),
2037                                           0,
2038                                           XFS_TRANS_PERM_LOG_RES,
2039                                           XFS_WRITE_LOG_COUNT);
2040
2041                 /*
2042                  * check for running out of space
2043                  */
2044                 if (error) {
2045                         /*
2046                          * Free the transaction structure.
2047                          */
2048                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
2049                         xfs_trans_cancel(tp, 0);
2050                         break;
2051                 }
2052                 xfs_ilock(ip, XFS_ILOCK_EXCL);
2053                 error = xfs_trans_reserve_quota(tp, mp,
2054                                 ip->i_udquot, ip->i_gdquot,
2055                                 resblks, 0, XFS_QMOPT_RES_REGBLKS);
2056                 if (error)
2057                         goto error1;
2058
2059                 xfs_trans_ijoin(tp, ip, 0);
2060
2061                 /*
2062                  * issue the bunmapi() call to free the blocks
2063                  */
2064                 xfs_bmap_init(&free_list, &firstfsb);
2065                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
2066                                   endoffset_fsb - startoffset_fsb,
2067                                   0, 2, &firstfsb, &free_list, &done);
2068                 if (error) {
2069                         goto error0;
2070                 }
2071
2072                 /*
2073                  * complete the transaction
2074                  */
2075                 error = xfs_bmap_finish(&tp, &free_list, &committed);
2076                 if (error) {
2077                         goto error0;
2078                 }
2079
2080                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2081                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2082         }
2083
2084  out_unlock_iolock:
2085         if (need_iolock)
2086                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2087         return error;
2088
2089  error0:
2090         xfs_bmap_cancel(&free_list);
2091  error1:
2092         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
2093         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
2094                     XFS_ILOCK_EXCL);
2095         return error;
2096 }
2097
2098 /*
2099  * xfs_change_file_space()
2100  *      This routine allocates or frees disk space for the given file.
2101  *      The user specified parameters are checked for alignment and size
2102  *      limitations.
2103  *
2104  * RETURNS:
2105  *       0 on success
2106  *      errno on error
2107  *
2108  */
2109 int
2110 xfs_change_file_space(
2111         xfs_inode_t     *ip,
2112         int             cmd,
2113         xfs_flock64_t   *bf,
2114         xfs_off_t       offset,
2115         int             attr_flags)
2116 {
2117         xfs_mount_t     *mp = ip->i_mount;
2118         int             clrprealloc;
2119         int             error;
2120         xfs_fsize_t     fsize;
2121         int             setprealloc;
2122         xfs_off_t       startoffset;
2123         xfs_off_t       llen;
2124         xfs_trans_t     *tp;
2125         struct iattr    iattr;
2126         int             prealloc_type;
2127
2128         if (!S_ISREG(ip->i_d.di_mode))
2129                 return XFS_ERROR(EINVAL);
2130
2131         switch (bf->l_whence) {
2132         case 0: /*SEEK_SET*/
2133                 break;
2134         case 1: /*SEEK_CUR*/
2135                 bf->l_start += offset;
2136                 break;
2137         case 2: /*SEEK_END*/
2138                 bf->l_start += XFS_ISIZE(ip);
2139                 break;
2140         default:
2141                 return XFS_ERROR(EINVAL);
2142         }
2143
2144         llen = bf->l_len > 0 ? bf->l_len - 1 : bf->l_len;
2145
2146         if (bf->l_start < 0 ||
2147             bf->l_start > mp->m_super->s_maxbytes ||
2148             bf->l_start + llen < 0 ||
2149             bf->l_start + llen > mp->m_super->s_maxbytes)
2150                 return XFS_ERROR(EINVAL);
2151
2152         bf->l_whence = 0;
2153
2154         startoffset = bf->l_start;
2155         fsize = XFS_ISIZE(ip);
2156
2157         /*
2158          * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
2159          * file space.
2160          * These calls do NOT zero the data space allocated to the file,
2161          * nor do they change the file size.
2162          *
2163          * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
2164          * space.
2165          * These calls cause the new file data to be zeroed and the file
2166          * size to be changed.
2167          */
2168         setprealloc = clrprealloc = 0;
2169         prealloc_type = XFS_BMAPI_PREALLOC;
2170
2171         switch (cmd) {
2172         case XFS_IOC_ZERO_RANGE:
2173                 prealloc_type |= XFS_BMAPI_CONVERT;
2174                 xfs_tosspages(ip, startoffset, startoffset + bf->l_len, 0);
2175                 /* FALLTHRU */
2176         case XFS_IOC_RESVSP:
2177         case XFS_IOC_RESVSP64:
2178                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
2179                                                 prealloc_type, attr_flags);
2180                 if (error)
2181                         return error;
2182                 setprealloc = 1;
2183                 break;
2184
2185         case XFS_IOC_UNRESVSP:
2186         case XFS_IOC_UNRESVSP64:
2187                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
2188                                                                 attr_flags)))
2189                         return error;
2190                 break;
2191
2192         case XFS_IOC_ALLOCSP:
2193         case XFS_IOC_ALLOCSP64:
2194         case XFS_IOC_FREESP:
2195         case XFS_IOC_FREESP64:
2196                 /*
2197                  * These operations actually do IO when extending the file, but
2198                  * the allocation is done seperately to the zeroing that is
2199                  * done. This set of operations need to be serialised against
2200                  * other IO operations, such as truncate and buffered IO. We
2201                  * need to take the IOLOCK here to serialise the allocation and
2202                  * zeroing IO to prevent other IOLOCK holders (e.g. getbmap,
2203                  * truncate, direct IO) from racing against the transient
2204                  * allocated but not written state we can have here.
2205                  */
2206                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
2207                 if (startoffset > fsize) {
2208                         error = xfs_alloc_file_space(ip, fsize,
2209                                         startoffset - fsize, 0,
2210                                         attr_flags | XFS_ATTR_NOLOCK);
2211                         if (error) {
2212                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2213                                 break;
2214                         }
2215                 }
2216
2217                 iattr.ia_valid = ATTR_SIZE;
2218                 iattr.ia_size = startoffset;
2219
2220                 error = xfs_setattr_size(ip, &iattr,
2221                                          attr_flags | XFS_ATTR_NOLOCK);
2222                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
2223
2224                 if (error)
2225                         return error;
2226
2227                 clrprealloc = 1;
2228                 break;
2229
2230         default:
2231                 ASSERT(0);
2232                 return XFS_ERROR(EINVAL);
2233         }
2234
2235         /*
2236          * update the inode timestamp, mode, and prealloc flag bits
2237          */
2238         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
2239
2240         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
2241                                       0, 0, 0))) {
2242                 /* ASSERT(0); */
2243                 xfs_trans_cancel(tp, 0);
2244                 return error;
2245         }
2246
2247         xfs_ilock(ip, XFS_ILOCK_EXCL);
2248         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2249
2250         if ((attr_flags & XFS_ATTR_DMI) == 0) {
2251                 ip->i_d.di_mode &= ~S_ISUID;
2252
2253                 /*
2254                  * Note that we don't have to worry about mandatory
2255                  * file locking being disabled here because we only
2256                  * clear the S_ISGID bit if the Group execute bit is
2257                  * on, but if it was on then mandatory locking wouldn't
2258                  * have been enabled.
2259                  */
2260                 if (ip->i_d.di_mode & S_IXGRP)
2261                         ip->i_d.di_mode &= ~S_ISGID;
2262
2263                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2264         }
2265         if (setprealloc)
2266                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
2267         else if (clrprealloc)
2268                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
2269
2270         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2271         if (attr_flags & XFS_ATTR_SYNC)
2272                 xfs_trans_set_sync(tp);
2273         return xfs_trans_commit(tp, 0);
2274 }