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

   1 /*
   2  * Copyright (c) 2000-2006 Silicon Graphics, Inc.
   3  * Copyright (c) 2012 Red Hat, Inc.
   4  * All Rights Reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License as
   8  * published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it would be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write the Free Software Foundation,
  17  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  18  */
  19
  20 #include "xfs.h"
  21 #include "xfs_fs.h"
  22 #include "xfs_types.h"
  23 #include "xfs_bit.h"
  24 #include "xfs_log.h"
  25 #include "xfs_trans.h"
  26 #include "xfs_sb.h"
  27 #include "xfs_ag.h"
  28 #include "xfs_dir2.h"
  29 #include "xfs_mount.h"
  30 #include "xfs_da_btree.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_ialloc_btree.h"
  33 #include "xfs_dinode.h"
  34 #include "xfs_inode.h"
  35 #include "xfs_inode_item.h"
  36 #include "xfs_itable.h"
  37 #include "xfs_ialloc.h"
  38 #include "xfs_alloc.h"
  39 #include "xfs_bmap.h"
  40 #include "xfs_acl.h"
  41 #include "xfs_attr.h"
  42 #include "xfs_error.h"
  43 #include "xfs_quota.h"
  44 #include "xfs_utils.h"
  45 #include "xfs_rtalloc.h"
  46 #include "xfs_trans_space.h"
  47 #include "xfs_log_priv.h"
  48 #include "xfs_filestream.h"
  49 #include "xfs_vnodeops.h"
  50 #include "xfs_trace.h"
  51 #include "xfs_icache.h"
  52 #include "xfs_symlink.h"
  53
  54
  55 /*
  56  * This is called by xfs_inactive to free any blocks beyond eof
  57  * when the link count isn't zero and by xfs_dm_punch_hole() when
  58  * punching a hole to EOF.
  59  */
  60 int
  61 xfs_free_eofblocks(
  62         xfs_mount_t     *mp,
  63         xfs_inode_t     *ip,
  64         bool            need_iolock)
  65 {
  66         xfs_trans_t     *tp;
  67         int             error;
  68         xfs_fileoff_t   end_fsb;
  69         xfs_fileoff_t   last_fsb;
  70         xfs_filblks_t   map_len;
  71         int             nimaps;
  72         xfs_bmbt_irec_t imap;
  73
  74         /*
  75          * Figure out if there are any blocks beyond the end
  76          * of the file.  If not, then there is nothing to do.
  77          */
  78         end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)XFS_ISIZE(ip));
  79         last_fsb = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
  80         if (last_fsb <= end_fsb)
  81                 return 0;
  82         map_len = last_fsb - end_fsb;
  83
  84         nimaps = 1;
  85         xfs_ilock(ip, XFS_ILOCK_SHARED);
  86         error = xfs_bmapi_read(ip, end_fsb, map_len, &imap, &nimaps, 0);
  87         xfs_iunlock(ip, XFS_ILOCK_SHARED);
  88
  89         if (!error && (nimaps != 0) &&
  90             (imap.br_startblock != HOLESTARTBLOCK ||
  91              ip->i_delayed_blks)) {
  92                 /*
  93                  * Attach the dquots to the inode up front.
  94                  */
  95                 error = xfs_qm_dqattach(ip, 0);
  96                 if (error)
  97                         return error;
  98
  99                 /*
 100                  * There are blocks after the end of file.
 101                  * Free them up now by truncating the file to
 102                  * its current size.
 103                  */
 104                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 105
 106                 if (need_iolock) {
 107                         if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
 108                                 xfs_trans_cancel(tp, 0);
 109                                 return EAGAIN;
 110                         }
 111                 }
 112
 113                 error = xfs_trans_reserve(tp, 0,
 114                                           XFS_ITRUNCATE_LOG_RES(mp),
 115                                           0, XFS_TRANS_PERM_LOG_RES,
 116                                           XFS_ITRUNCATE_LOG_COUNT);
 117                 if (error) {
 118                         ASSERT(XFS_FORCED_SHUTDOWN(mp));
 119                         xfs_trans_cancel(tp, 0);
 120                         if (need_iolock)
 121                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 122                         return error;
 123                 }
 124
 125                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 126                 xfs_trans_ijoin(tp, ip, 0);
 127
 128                 /*
 129                  * Do not update the on-disk file size.  If we update the
 130                  * on-disk file size and then the system crashes before the
 131                  * contents of the file are flushed to disk then the files
 132                  * may be full of holes (ie NULL files bug).
 133                  */
 134                 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK,
 135                                               XFS_ISIZE(ip));
 136                 if (error) {
 137                         /*
 138                          * If we get an error at this point we simply don't
 139                          * bother truncating the file.
 140                          */
 141                         xfs_trans_cancel(tp,
 142                                          (XFS_TRANS_RELEASE_LOG_RES |
 143                                           XFS_TRANS_ABORT));
 144                 } else {
 145                         error = xfs_trans_commit(tp,
 146                                                 XFS_TRANS_RELEASE_LOG_RES);
 147                         if (!error)
 148                                 xfs_inode_clear_eofblocks_tag(ip);
 149                 }
 150
 151                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 152                 if (need_iolock)
 153                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 154         }
 155         return error;
 156 }
 157
 158 int
 159 xfs_release(
 160         xfs_inode_t     *ip)
 161 {
 162         xfs_mount_t     *mp = ip->i_mount;
 163         int             error;
 164
 165         if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
 166                 return 0;
 167
 168         /* If this is a read-only mount, don't do this (would generate I/O) */
 169         if (mp->m_flags & XFS_MOUNT_RDONLY)
 170                 return 0;
 171
 172         if (!XFS_FORCED_SHUTDOWN(mp)) {
 173                 int truncated;
 174
 175                 /*
 176                  * If we are using filestreams, and we have an unlinked
 177                  * file that we are processing the last close on, then nothing
 178                  * will be able to reopen and write to this file. Purge this
 179                  * inode from the filestreams cache so that it doesn't delay
 180                  * teardown of the inode.
 181                  */
 182                 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
 183                         xfs_filestream_deassociate(ip);
 184
 185                 /*
 186                  * If we previously truncated this file and removed old data
 187                  * in the process, we want to initiate "early" writeout on
 188                  * the last close.  This is an attempt to combat the notorious
 189                  * NULL files problem which is particularly noticeable from a
 190                  * truncate down, buffered (re-)write (delalloc), followed by
 191                  * a crash.  What we are effectively doing here is
 192                  * significantly reducing the time window where we'd otherwise
 193                  * be exposed to that problem.
 194                  */
 195                 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
 196                 if (truncated) {
 197                         xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
 198                         if (VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0) {
 199                                 error = -filemap_flush(VFS_I(ip)->i_mapping);
 200                                 if (error)
 201                                         return error;
 202                         }
 203                 }
 204         }
 205
 206         if (ip->i_d.di_nlink == 0)
 207                 return 0;
 208
 209         if (xfs_can_free_eofblocks(ip, false)) {
 210
 211                 /*
 212                  * If we can't get the iolock just skip truncating the blocks
 213                  * past EOF because we could deadlock with the mmap_sem
 214                  * otherwise.  We'll get another chance to drop them once the
 215                  * last reference to the inode is dropped, so we'll never leak
 216                  * blocks permanently.
 217                  *
 218                  * Further, check if the inode is being opened, written and
 219                  * closed frequently and we have delayed allocation blocks
 220                  * outstanding (e.g. streaming writes from the NFS server),
 221                  * truncating the blocks past EOF will cause fragmentation to
 222                  * occur.
 223                  *
 224                  * In this case don't do the truncation, either, but we have to
 225                  * be careful how we detect this case. Blocks beyond EOF show
 226                  * up as i_delayed_blks even when the inode is clean, so we
 227                  * need to truncate them away first before checking for a dirty
 228                  * release. Hence on the first dirty close we will still remove
 229                  * the speculative allocation, but after that we will leave it
 230                  * in place.
 231                  */
 232                 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
 233                         return 0;
 234
 235                 error = xfs_free_eofblocks(mp, ip, true);
 236                 if (error && error != EAGAIN)
 237                         return error;
 238
 239                 /* delalloc blocks after truncation means it really is dirty */
 240                 if (ip->i_delayed_blks)
 241                         xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
 242         }
 243         return 0;
 244 }
 245
 246 /*
 247  * xfs_inactive
 248  *
 249  * This is called when the vnode reference count for the vnode
 250  * goes to zero.  If the file has been unlinked, then it must
 251  * now be truncated.  Also, we clear all of the read-ahead state
 252  * kept for the inode here since the file is now closed.
 253  */
 254 int
 255 xfs_inactive(
 256         xfs_inode_t     *ip)
 257 {
 258         xfs_bmap_free_t free_list;
 259         xfs_fsblock_t   first_block;
 260         int             committed;
 261         xfs_trans_t     *tp;
 262         xfs_mount_t     *mp;
 263         int             error;
 264         int             truncate = 0;
 265
 266         /*
 267          * If the inode is already free, then there can be nothing
 268          * to clean up here.
 269          */
 270         if (ip->i_d.di_mode == 0 || is_bad_inode(VFS_I(ip))) {
 271                 ASSERT(ip->i_df.if_real_bytes == 0);
 272                 ASSERT(ip->i_df.if_broot_bytes == 0);
 273                 return VN_INACTIVE_CACHE;
 274         }
 275
 276         mp = ip->i_mount;
 277
 278         error = 0;
 279
 280         /* If this is a read-only mount, don't do this (would generate I/O) */
 281         if (mp->m_flags & XFS_MOUNT_RDONLY)
 282                 goto out;
 283
 284         if (ip->i_d.di_nlink != 0) {
 285                 /*
 286                  * force is true because we are evicting an inode from the
 287                  * cache. Post-eof blocks must be freed, lest we end up with
 288                  * broken free space accounting.
 289                  */
 290                 if (xfs_can_free_eofblocks(ip, true)) {
 291                         error = xfs_free_eofblocks(mp, ip, false);
 292                         if (error)
 293                                 return VN_INACTIVE_CACHE;
 294                 }
 295                 goto out;
 296         }
 297
 298         if (S_ISREG(ip->i_d.di_mode) &&
 299             (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
 300              ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
 301                 truncate = 1;
 302
 303         error = xfs_qm_dqattach(ip, 0);
 304         if (error)
 305                 return VN_INACTIVE_CACHE;
 306
 307         tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 308         error = xfs_trans_reserve(tp, 0,
 309                         (truncate || S_ISLNK(ip->i_d.di_mode)) ?
 310                                 XFS_ITRUNCATE_LOG_RES(mp) :
 311                                 XFS_IFREE_LOG_RES(mp),
 312                         0,
 313                         XFS_TRANS_PERM_LOG_RES,
 314                         XFS_ITRUNCATE_LOG_COUNT);
 315         if (error) {
 316                 ASSERT(XFS_FORCED_SHUTDOWN(mp));
 317                 xfs_trans_cancel(tp, 0);
 318                 return VN_INACTIVE_CACHE;
 319         }
 320
 321         xfs_ilock(ip, XFS_ILOCK_EXCL);
 322         xfs_trans_ijoin(tp, ip, 0);
 323
 324         if (S_ISLNK(ip->i_d.di_mode)) {
 325                 error = xfs_inactive_symlink(ip, &tp);
 326                 if (error)
 327                         goto out_cancel;
 328         } else if (truncate) {
 329                 ip->i_d.di_size = 0;
 330                 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
 331
 332                 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
 333                 if (error)
 334                         goto out_cancel;
 335
 336                 ASSERT(ip->i_d.di_nextents == 0);
 337         }
 338
 339         /*
 340          * If there are attributes associated with the file then blow them away
 341          * now.  The code calls a routine that recursively deconstructs the
 342          * attribute fork.  We need to just commit the current transaction
 343          * because we can't use it for xfs_attr_inactive().
 344          */
 345         if (ip->i_d.di_anextents > 0) {
 346                 ASSERT(ip->i_d.di_forkoff != 0);
 347
 348                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 349                 if (error)
 350                         goto out_unlock;
 351
 352                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 353
 354                 error = xfs_attr_inactive(ip);
 355                 if (error)
 356                         goto out;
 357
 358                 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
 359                 error = xfs_trans_reserve(tp, 0,
 360                                           XFS_IFREE_LOG_RES(mp),
 361                                           0, XFS_TRANS_PERM_LOG_RES,
 362                                           XFS_INACTIVE_LOG_COUNT);
 363                 if (error) {
 364                         xfs_trans_cancel(tp, 0);
 365                         goto out;
 366                 }
 367
 368                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 369                 xfs_trans_ijoin(tp, ip, 0);
 370         }
 371
 372         if (ip->i_afp)
 373                 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
 374
 375         ASSERT(ip->i_d.di_anextents == 0);
 376
 377         /*
 378          * Free the inode.
 379          */
 380         xfs_bmap_init(&free_list, &first_block);
 381         error = xfs_ifree(tp, ip, &free_list);
 382         if (error) {
 383                 /*
 384                  * If we fail to free the inode, shut down.  The cancel
 385                  * might do that, we need to make sure.  Otherwise the
 386                  * inode might be lost for a long time or forever.
 387                  */
 388                 if (!XFS_FORCED_SHUTDOWN(mp)) {
 389                         xfs_notice(mp, "%s: xfs_ifree returned error %d",
 390                                 __func__, error);
 391                         xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
 392                 }
 393                 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
 394         } else {
 395                 /*
 396                  * Credit the quota account(s). The inode is gone.
 397                  */
 398                 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
 399
 400                 /*
 401                  * Just ignore errors at this point.  There is nothing we can
 402                  * do except to try to keep going. Make sure it's not a silent
 403                  * error.
 404                  */
 405                 error = xfs_bmap_finish(&tp,  &free_list, &committed);
 406                 if (error)
 407                         xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
 408                                 __func__, error);
 409                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 410                 if (error)
 411                         xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
 412                                 __func__, error);
 413         }
 414
 415         /*
 416          * Release the dquots held by inode, if any.
 417          */
 418         xfs_qm_dqdetach(ip);
 419 out_unlock:
 420         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 421 out:
 422         return VN_INACTIVE_CACHE;
 423 out_cancel:
 424         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
 425         goto out_unlock;
 426 }
 427
 428 /*
 429  * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
 430  * is allowed, otherwise it has to be an exact match. If a CI match is found,
 431  * ci_name->name will point to a the actual name (caller must free) or
 432  * will be set to NULL if an exact match is found.
 433  */
 434 int
 435 xfs_lookup(
 436         xfs_inode_t             *dp,
 437         struct xfs_name         *name,
 438         xfs_inode_t             **ipp,
 439         struct xfs_name         *ci_name)
 440 {
 441         xfs_ino_t               inum;
 442         int                     error;
 443         uint                    lock_mode;
 444
 445         trace_xfs_lookup(dp, name);
 446
 447         if (XFS_FORCED_SHUTDOWN(dp->i_mount))
 448                 return XFS_ERROR(EIO);
 449
 450         lock_mode = xfs_ilock_map_shared(dp);
 451         error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
 452         xfs_iunlock_map_shared(dp, lock_mode);
 453
 454         if (error)
 455                 goto out;
 456
 457         error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
 458         if (error)
 459                 goto out_free_name;
 460
 461         return 0;
 462
 463 out_free_name:
 464         if (ci_name)
 465                 kmem_free(ci_name->name);
 466 out:
 467         *ipp = NULL;
 468         return error;
 469 }
 470
 471 int
 472 xfs_create(
 473         xfs_inode_t             *dp,
 474         struct xfs_name         *name,
 475         umode_t                 mode,
 476         xfs_dev_t               rdev,
 477         xfs_inode_t             **ipp)
 478 {
 479         int                     is_dir = S_ISDIR(mode);
 480         struct xfs_mount        *mp = dp->i_mount;
 481         struct xfs_inode        *ip = NULL;
 482         struct xfs_trans        *tp = NULL;
 483         int                     error;
 484         xfs_bmap_free_t         free_list;
 485         xfs_fsblock_t           first_block;
 486         bool                    unlock_dp_on_error = false;
 487         uint                    cancel_flags;
 488         int                     committed;
 489         prid_t                  prid;
 490         struct xfs_dquot        *udqp = NULL;
 491         struct xfs_dquot        *gdqp = NULL;
 492         struct xfs_dquot        *pdqp = NULL;
 493         uint                    resblks;
 494         uint                    log_res;
 495         uint                    log_count;
 496
 497         trace_xfs_create(dp, name);
 498
 499         if (XFS_FORCED_SHUTDOWN(mp))
 500                 return XFS_ERROR(EIO);
 501
 502         if (dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
 503                 prid = xfs_get_projid(dp);
 504         else
 505                 prid = XFS_PROJID_DEFAULT;
 506
 507         /*
 508          * Make sure that we have allocated dquot(s) on disk.
 509          */
 510         error = xfs_qm_vop_dqalloc(dp, current_fsuid(), current_fsgid(), prid,
 511                                         XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
 512                                         &udqp, &gdqp, &pdqp);
 513         if (error)
 514                 return error;
 515
 516         if (is_dir) {
 517                 rdev = 0;
 518                 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
 519                 log_res = XFS_MKDIR_LOG_RES(mp);
 520                 log_count = XFS_MKDIR_LOG_COUNT;
 521                 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
 522         } else {
 523                 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
 524                 log_res = XFS_CREATE_LOG_RES(mp);
 525                 log_count = XFS_CREATE_LOG_COUNT;
 526                 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
 527         }
 528
 529         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
 530
 531         /*
 532          * Initially assume that the file does not exist and
 533          * reserve the resources for that case.  If that is not
 534          * the case we'll drop the one we have and get a more
 535          * appropriate transaction later.
 536          */
 537         error = xfs_trans_reserve(tp, resblks, log_res, 0,
 538                         XFS_TRANS_PERM_LOG_RES, log_count);
 539         if (error == ENOSPC) {
 540                 /* flush outstanding delalloc blocks and retry */
 541                 xfs_flush_inodes(mp);
 542                 error = xfs_trans_reserve(tp, resblks, log_res, 0,
 543                                 XFS_TRANS_PERM_LOG_RES, log_count);
 544         }
 545         if (error == ENOSPC) {
 546                 /* No space at all so try a "no-allocation" reservation */
 547                 resblks = 0;
 548                 error = xfs_trans_reserve(tp, 0, log_res, 0,
 549                                 XFS_TRANS_PERM_LOG_RES, log_count);
 550         }
 551         if (error) {
 552                 cancel_flags = 0;
 553                 goto out_trans_cancel;
 554         }
 555
 556         xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
 557         unlock_dp_on_error = true;
 558
 559         xfs_bmap_init(&free_list, &first_block);
 560
 561         /*
 562          * Reserve disk quota and the inode.
 563          */
 564         error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
 565                                                 pdqp, resblks, 1, 0);
 566         if (error)
 567                 goto out_trans_cancel;
 568
 569         error = xfs_dir_canenter(tp, dp, name, resblks);
 570         if (error)
 571                 goto out_trans_cancel;
 572
 573         /*
 574          * A newly created regular or special file just has one directory
 575          * entry pointing to them, but a directory also the "." entry
 576          * pointing to itself.
 577          */
 578         error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
 579                                prid, resblks > 0, &ip, &committed);
 580         if (error) {
 581                 if (error == ENOSPC)
 582                         goto out_trans_cancel;
 583                 goto out_trans_abort;
 584         }
 585
 586         /*
 587          * Now we join the directory inode to the transaction.  We do not do it
 588          * earlier because xfs_dir_ialloc might commit the previous transaction
 589          * (and release all the locks).  An error from here on will result in
 590          * the transaction cancel unlocking dp so don't do it explicitly in the
 591          * error path.
 592          */
 593         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
 594         unlock_dp_on_error = false;
 595
 596         error = xfs_dir_createname(tp, dp, name, ip->i_ino,
 597                                         &first_block, &free_list, resblks ?
 598                                         resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
 599         if (error) {
 600                 ASSERT(error != ENOSPC);
 601                 goto out_trans_abort;
 602         }
 603         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 604         xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
 605
 606         if (is_dir) {
 607                 error = xfs_dir_init(tp, ip, dp);
 608                 if (error)
 609                         goto out_bmap_cancel;
 610
 611                 error = xfs_bumplink(tp, dp);
 612                 if (error)
 613                         goto out_bmap_cancel;
 614         }
 615
 616         /*
 617          * If this is a synchronous mount, make sure that the
 618          * create transaction goes to disk before returning to
 619          * the user.
 620          */
 621         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
 622                 xfs_trans_set_sync(tp);
 623
 624         /*
 625          * Attach the dquot(s) to the inodes and modify them incore.
 626          * These ids of the inode couldn't have changed since the new
 627          * inode has been locked ever since it was created.
 628          */
 629         xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
 630
 631         error = xfs_bmap_finish(&tp, &free_list, &committed);
 632         if (error)
 633                 goto out_bmap_cancel;
 634
 635         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 636         if (error)
 637                 goto out_release_inode;
 638
 639         xfs_qm_dqrele(udqp);
 640         xfs_qm_dqrele(gdqp);
 641         xfs_qm_dqrele(pdqp);
 642
 643         *ipp = ip;
 644         return 0;
 645
 646  out_bmap_cancel:
 647         xfs_bmap_cancel(&free_list);
 648  out_trans_abort:
 649         cancel_flags |= XFS_TRANS_ABORT;
 650  out_trans_cancel:
 651         xfs_trans_cancel(tp, cancel_flags);
 652  out_release_inode:
 653         /*
 654          * Wait until after the current transaction is aborted to
 655          * release the inode.  This prevents recursive transactions
 656          * and deadlocks from xfs_inactive.
 657          */
 658         if (ip)
 659                 IRELE(ip);
 660
 661         xfs_qm_dqrele(udqp);
 662         xfs_qm_dqrele(gdqp);
 663         xfs_qm_dqrele(pdqp);
 664
 665         if (unlock_dp_on_error)
 666                 xfs_iunlock(dp, XFS_ILOCK_EXCL);
 667         return error;
 668 }
 669
 670 #ifdef DEBUG
 671 int xfs_locked_n;
 672 int xfs_small_retries;
 673 int xfs_middle_retries;
 674 int xfs_lots_retries;
 675 int xfs_lock_delays;
 676 #endif
 677
 678 /*
 679  * Bump the subclass so xfs_lock_inodes() acquires each lock with
 680  * a different value
 681  */
 682 static inline int
 683 xfs_lock_inumorder(int lock_mode, int subclass)
 684 {
 685         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
 686                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
 687         if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
 688                 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
 689
 690         return lock_mode;
 691 }
 692
 693 /*
 694  * The following routine will lock n inodes in exclusive mode.
 695  * We assume the caller calls us with the inodes in i_ino order.
 696  *
 697  * We need to detect deadlock where an inode that we lock
 698  * is in the AIL and we start waiting for another inode that is locked
 699  * by a thread in a long running transaction (such as truncate). This can
 700  * result in deadlock since the long running trans might need to wait
 701  * for the inode we just locked in order to push the tail and free space
 702  * in the log.
 703  */
 704 void
 705 xfs_lock_inodes(
 706         xfs_inode_t     **ips,
 707         int             inodes,
 708         uint            lock_mode)
 709 {
 710         int             attempts = 0, i, j, try_lock;
 711         xfs_log_item_t  *lp;
 712
 713         ASSERT(ips && (inodes >= 2)); /* we need at least two */
 714
 715         try_lock = 0;
 716         i = 0;
 717
 718 again:
 719         for (; i < inodes; i++) {
 720                 ASSERT(ips[i]);
 721
 722                 if (i && (ips[i] == ips[i-1]))  /* Already locked */
 723                         continue;
 724
 725                 /*
 726                  * If try_lock is not set yet, make sure all locked inodes
 727                  * are not in the AIL.
 728                  * If any are, set try_lock to be used later.
 729                  */
 730
 731                 if (!try_lock) {
 732                         for (j = (i - 1); j >= 0 && !try_lock; j--) {
 733                                 lp = (xfs_log_item_t *)ips[j]->i_itemp;
 734                                 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
 735                                         try_lock++;
 736                                 }
 737                         }
 738                 }
 739
 740                 /*
 741                  * If any of the previous locks we have locked is in the AIL,
 742                  * we must TRY to get the second and subsequent locks. If
 743                  * we can't get any, we must release all we have
 744                  * and try again.
 745                  */
 746
 747                 if (try_lock) {
 748                         /* try_lock must be 0 if i is 0. */
 749                         /*
 750                          * try_lock means we have an inode locked
 751                          * that is in the AIL.
 752                          */
 753                         ASSERT(i != 0);
 754                         if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
 755                                 attempts++;
 756
 757                                 /*
 758                                  * Unlock all previous guys and try again.
 759                                  * xfs_iunlock will try to push the tail
 760                                  * if the inode is in the AIL.
 761                                  */
 762
 763                                 for(j = i - 1; j >= 0; j--) {
 764
 765                                         /*
 766                                          * Check to see if we've already
 767                                          * unlocked this one.
 768                                          * Not the first one going back,
 769                                          * and the inode ptr is the same.
 770                                          */
 771                                         if ((j != (i - 1)) && ips[j] ==
 772                                                                 ips[j+1])
 773                                                 continue;
 774
 775                                         xfs_iunlock(ips[j], lock_mode);
 776                                 }
 777
 778                                 if ((attempts % 5) == 0) {
 779                                         delay(1); /* Don't just spin the CPU */
 780 #ifdef DEBUG
 781                                         xfs_lock_delays++;
 782 #endif
 783                                 }
 784                                 i = 0;
 785                                 try_lock = 0;
 786                                 goto again;
 787                         }
 788                 } else {
 789                         xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
 790                 }
 791         }
 792
 793 #ifdef DEBUG
 794         if (attempts) {
 795                 if (attempts < 5) xfs_small_retries++;
 796                 else if (attempts < 100) xfs_middle_retries++;
 797                 else xfs_lots_retries++;
 798         } else {
 799                 xfs_locked_n++;
 800         }
 801 #endif
 802 }
 803
 804 /*
 805  * xfs_lock_two_inodes() can only be used to lock one type of lock
 806  * at a time - the iolock or the ilock, but not both at once. If
 807  * we lock both at once, lockdep will report false positives saying
 808  * we have violated locking orders.
 809  */
 810 void
 811 xfs_lock_two_inodes(
 812         xfs_inode_t             *ip0,
 813         xfs_inode_t             *ip1,
 814         uint                    lock_mode)
 815 {
 816         xfs_inode_t             *temp;
 817         int                     attempts = 0;
 818         xfs_log_item_t          *lp;
 819
 820         if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
 821                 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
 822         ASSERT(ip0->i_ino != ip1->i_ino);
 823
 824         if (ip0->i_ino > ip1->i_ino) {
 825                 temp = ip0;
 826                 ip0 = ip1;
 827                 ip1 = temp;
 828         }
 829
 830  again:
 831         xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
 832
 833         /*
 834          * If the first lock we have locked is in the AIL, we must TRY to get
 835          * the second lock. If we can't get it, we must release the first one
 836          * and try again.
 837          */
 838         lp = (xfs_log_item_t *)ip0->i_itemp;
 839         if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
 840                 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
 841                         xfs_iunlock(ip0, lock_mode);
 842                         if ((++attempts % 5) == 0)
 843                                 delay(1); /* Don't just spin the CPU */
 844                         goto again;
 845                 }
 846         } else {
 847                 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
 848         }
 849 }
 850
 851 int
 852 xfs_remove(
 853         xfs_inode_t             *dp,
 854         struct xfs_name         *name,
 855         xfs_inode_t             *ip)
 856 {
 857         xfs_mount_t             *mp = dp->i_mount;
 858         xfs_trans_t             *tp = NULL;
 859         int                     is_dir = S_ISDIR(ip->i_d.di_mode);
 860         int                     error = 0;
 861         xfs_bmap_free_t         free_list;
 862         xfs_fsblock_t           first_block;
 863         int                     cancel_flags;
 864         int                     committed;
 865         int                     link_zero;
 866         uint                    resblks;
 867         uint                    log_count;
 868
 869         trace_xfs_remove(dp, name);
 870
 871         if (XFS_FORCED_SHUTDOWN(mp))
 872                 return XFS_ERROR(EIO);
 873
 874         error = xfs_qm_dqattach(dp, 0);
 875         if (error)
 876                 goto std_return;
 877
 878         error = xfs_qm_dqattach(ip, 0);
 879         if (error)
 880                 goto std_return;
 881
 882         if (is_dir) {
 883                 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
 884                 log_count = XFS_DEFAULT_LOG_COUNT;
 885         } else {
 886                 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
 887                 log_count = XFS_REMOVE_LOG_COUNT;
 888         }
 889         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
 890
 891         /*
 892          * We try to get the real space reservation first,
 893          * allowing for directory btree deletion(s) implying
 894          * possible bmap insert(s).  If we can't get the space
 895          * reservation then we use 0 instead, and avoid the bmap
 896          * btree insert(s) in the directory code by, if the bmap
 897          * insert tries to happen, instead trimming the LAST
 898          * block from the directory.
 899          */
 900         resblks = XFS_REMOVE_SPACE_RES(mp);
 901         error = xfs_trans_reserve(tp, resblks, XFS_REMOVE_LOG_RES(mp), 0,
 902                                   XFS_TRANS_PERM_LOG_RES, log_count);
 903         if (error == ENOSPC) {
 904                 resblks = 0;
 905                 error = xfs_trans_reserve(tp, 0, XFS_REMOVE_LOG_RES(mp), 0,
 906                                           XFS_TRANS_PERM_LOG_RES, log_count);
 907         }
 908         if (error) {
 909                 ASSERT(error != ENOSPC);
 910                 cancel_flags = 0;
 911                 goto out_trans_cancel;
 912         }
 913
 914         xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
 915
 916         xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
 917         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
 918
 919         /*
 920          * If we're removing a directory perform some additional validation.
 921          */
 922         if (is_dir) {
 923                 ASSERT(ip->i_d.di_nlink >= 2);
 924                 if (ip->i_d.di_nlink != 2) {
 925                         error = XFS_ERROR(ENOTEMPTY);
 926                         goto out_trans_cancel;
 927                 }
 928                 if (!xfs_dir_isempty(ip)) {
 929                         error = XFS_ERROR(ENOTEMPTY);
 930                         goto out_trans_cancel;
 931                 }
 932         }
 933
 934         xfs_bmap_init(&free_list, &first_block);
 935         error = xfs_dir_removename(tp, dp, name, ip->i_ino,
 936                                         &first_block, &free_list, resblks);
 937         if (error) {
 938                 ASSERT(error != ENOENT);
 939                 goto out_bmap_cancel;
 940         }
 941         xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 942
 943         if (is_dir) {
 944                 /*
 945                  * Drop the link from ip's "..".
 946                  */
 947                 error = xfs_droplink(tp, dp);
 948                 if (error)
 949                         goto out_bmap_cancel;
 950
 951                 /*
 952                  * Drop the "." link from ip to self.
 953                  */
 954                 error = xfs_droplink(tp, ip);
 955                 if (error)
 956                         goto out_bmap_cancel;
 957         } else {
 958                 /*
 959                  * When removing a non-directory we need to log the parent
 960                  * inode here.  For a directory this is done implicitly
 961                  * by the xfs_droplink call for the ".." entry.
 962                  */
 963                 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
 964         }
 965
 966         /*
 967          * Drop the link from dp to ip.
 968          */
 969         error = xfs_droplink(tp, ip);
 970         if (error)
 971                 goto out_bmap_cancel;
 972
 973         /*
 974          * Determine if this is the last link while
 975          * we are in the transaction.
 976          */
 977         link_zero = (ip->i_d.di_nlink == 0);
 978
 979         /*
 980          * If this is a synchronous mount, make sure that the
 981          * remove transaction goes to disk before returning to
 982          * the user.
 983          */
 984         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
 985                 xfs_trans_set_sync(tp);
 986
 987         error = xfs_bmap_finish(&tp, &free_list, &committed);
 988         if (error)
 989                 goto out_bmap_cancel;
 990
 991         error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
 992         if (error)
 993                 goto std_return;
 994
 995         /*
 996          * If we are using filestreams, kill the stream association.
 997          * If the file is still open it may get a new one but that
 998          * will get killed on last close in xfs_close() so we don't
 999          * have to worry about that.
1000          */
1001         if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
1002                 xfs_filestream_deassociate(ip);
1003
1004         return 0;
1005
1006  out_bmap_cancel:
1007         xfs_bmap_cancel(&free_list);
1008         cancel_flags |= XFS_TRANS_ABORT;
1009  out_trans_cancel:
1010         xfs_trans_cancel(tp, cancel_flags);
1011  std_return:
1012         return error;
1013 }
1014
1015 int
1016 xfs_link(
1017         xfs_inode_t             *tdp,
1018         xfs_inode_t             *sip,
1019         struct xfs_name         *target_name)
1020 {
1021         xfs_mount_t             *mp = tdp->i_mount;
1022         xfs_trans_t             *tp;
1023         int                     error;
1024         xfs_bmap_free_t         free_list;
1025         xfs_fsblock_t           first_block;
1026         int                     cancel_flags;
1027         int                     committed;
1028         int                     resblks;
1029
1030         trace_xfs_link(tdp, target_name);
1031
1032         ASSERT(!S_ISDIR(sip->i_d.di_mode));
1033
1034         if (XFS_FORCED_SHUTDOWN(mp))
1035                 return XFS_ERROR(EIO);
1036
1037         error = xfs_qm_dqattach(sip, 0);
1038         if (error)
1039                 goto std_return;
1040
1041         error = xfs_qm_dqattach(tdp, 0);
1042         if (error)
1043                 goto std_return;
1044
1045         tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1046         cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1047         resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1048         error = xfs_trans_reserve(tp, resblks, XFS_LINK_LOG_RES(mp), 0,
1049                         XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1050         if (error == ENOSPC) {
1051                 resblks = 0;
1052                 error = xfs_trans_reserve(tp, 0, XFS_LINK_LOG_RES(mp), 0,
1053                                 XFS_TRANS_PERM_LOG_RES, XFS_LINK_LOG_COUNT);
1054         }
1055         if (error) {
1056                 cancel_flags = 0;
1057                 goto error_return;
1058         }
1059
1060         xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1061
1062         xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1063         xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1064
1065         /*
1066          * If we are using project inheritance, we only allow hard link
1067          * creation in our tree when the project IDs are the same; else
1068          * the tree quota mechanism could be circumvented.
1069          */
1070         if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1071                      (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1072                 error = XFS_ERROR(EXDEV);
1073                 goto error_return;
1074         }
1075
1076         error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1077         if (error)
1078                 goto error_return;
1079
1080         xfs_bmap_init(&free_list, &first_block);
1081
1082         error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1083                                         &first_block, &free_list, resblks);
1084         if (error)
1085                 goto abort_return;
1086         xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1087         xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1088
1089         error = xfs_bumplink(tp, sip);
1090         if (error)
1091                 goto abort_return;
1092
1093         /*
1094          * If this is a synchronous mount, make sure that the
1095          * link transaction goes to disk before returning to
1096          * the user.
1097          */
1098         if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1099                 xfs_trans_set_sync(tp);
1100         }
1101
1102         error = xfs_bmap_finish (&tp, &free_list, &committed);
1103         if (error) {
1104                 xfs_bmap_cancel(&free_list);
1105                 goto abort_return;
1106         }
1107
1108         return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1109
1110  abort_return:
1111         cancel_flags |= XFS_TRANS_ABORT;
1112  error_return:
1113         xfs_trans_cancel(tp, cancel_flags);
1114  std_return:
1115         return error;
1116 }
1117
1118 int
1119 xfs_set_dmattrs(
1120         xfs_inode_t     *ip,
1121         u_int           evmask,
1122         u_int16_t       state)
1123 {
1124         xfs_mount_t     *mp = ip->i_mount;
1125         xfs_trans_t     *tp;
1126         int             error;
1127
1128         if (!capable(CAP_SYS_ADMIN))
1129                 return XFS_ERROR(EPERM);
1130
1131         if (XFS_FORCED_SHUTDOWN(mp))
1132                 return XFS_ERROR(EIO);
1133
1134         tp = xfs_trans_alloc(mp, XFS_TRANS_SET_DMATTRS);
1135         error = xfs_trans_reserve(tp, 0, XFS_ICHANGE_LOG_RES (mp), 0, 0, 0);
1136         if (error) {
1137                 xfs_trans_cancel(tp, 0);
1138                 return error;
1139         }
1140         xfs_ilock(ip, XFS_ILOCK_EXCL);
1141         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1142
1143         ip->i_d.di_dmevmask = evmask;
1144         ip->i_d.di_dmstate  = state;
1145
1146         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1147         error = xfs_trans_commit(tp, 0);
1148
1149         return error;
1150 }
1151
1152 /*
1153  * xfs_alloc_file_space()
1154  *      This routine allocates disk space for the given file.
1155  *
1156  *      If alloc_type == 0, this request is for an ALLOCSP type
1157  *      request which will change the file size.  In this case, no
1158  *      DMAPI event will be generated by the call.  A TRUNCATE event
1159  *      will be generated later by xfs_setattr.
1160  *
1161  *      If alloc_type != 0, this request is for a RESVSP type
1162  *      request, and a DMAPI DM_EVENT_WRITE will be generated if the
1163  *      lower block boundary byte address is less than the file's
1164  *      length.
1165  *
1166  * RETURNS:
1167  *       0 on success
1168  *      errno on error
1169  *
1170  */
1171 STATIC int
1172 xfs_alloc_file_space(
1173         xfs_inode_t             *ip,
1174         xfs_off_t               offset,
1175         xfs_off_t               len,
1176         int                     alloc_type,
1177         int                     attr_flags)
1178 {
1179         xfs_mount_t             *mp = ip->i_mount;
1180         xfs_off_t               count;
1181         xfs_filblks_t           allocated_fsb;
1182         xfs_filblks_t           allocatesize_fsb;
1183         xfs_extlen_t            extsz, temp;
1184         xfs_fileoff_t           startoffset_fsb;
1185         xfs_fsblock_t           firstfsb;
1186         int                     nimaps;
1187         int                     quota_flag;
1188         int                     rt;
1189         xfs_trans_t             *tp;
1190         xfs_bmbt_irec_t         imaps[1], *imapp;
1191         xfs_bmap_free_t         free_list;
1192         uint                    qblocks, resblks, resrtextents;
1193         int                     committed;
1194         int                     error;
1195
1196         trace_xfs_alloc_file_space(ip);
1197
1198         if (XFS_FORCED_SHUTDOWN(mp))
1199                 return XFS_ERROR(EIO);
1200
1201         error = xfs_qm_dqattach(ip, 0);
1202         if (error)
1203                 return error;
1204
1205         if (len <= 0)
1206                 return XFS_ERROR(EINVAL);
1207
1208         rt = XFS_IS_REALTIME_INODE(ip);
1209         extsz = xfs_get_extsz_hint(ip);
1210
1211         count = len;
1212         imapp = &imaps[0];
1213         nimaps = 1;
1214         startoffset_fsb = XFS_B_TO_FSBT(mp, offset);
1215         allocatesize_fsb = XFS_B_TO_FSB(mp, count);
1216
1217         /*
1218          * Allocate file space until done or until there is an error
1219          */
1220         while (allocatesize_fsb && !error) {
1221                 xfs_fileoff_t   s, e;
1222
1223                 /*
1224                  * Determine space reservations for data/realtime.
1225                  */
1226                 if (unlikely(extsz)) {
1227                         s = startoffset_fsb;
1228                         do_div(s, extsz);
1229                         s *= extsz;
1230                         e = startoffset_fsb + allocatesize_fsb;
1231                         if ((temp = do_mod(startoffset_fsb, extsz)))
1232                                 e += temp;
1233                         if ((temp = do_mod(e, extsz)))
1234                                 e += extsz - temp;
1235                 } else {
1236                         s = 0;
1237                         e = allocatesize_fsb;
1238                 }
1239
1240                 /*
1241                  * The transaction reservation is limited to a 32-bit block
1242                  * count, hence we need to limit the number of blocks we are
1243                  * trying to reserve to avoid an overflow. We can't allocate
1244                  * more than @nimaps extents, and an extent is limited on disk
1245                  * to MAXEXTLEN (21 bits), so use that to enforce the limit.
1246                  */
1247                 resblks = min_t(xfs_fileoff_t, (e - s), (MAXEXTLEN * nimaps));
1248                 if (unlikely(rt)) {
1249                         resrtextents = qblocks = resblks;
1250                         resrtextents /= mp->m_sb.sb_rextsize;
1251                         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
1252                         quota_flag = XFS_QMOPT_RES_RTBLKS;
1253                 } else {
1254                         resrtextents = 0;
1255                         resblks = qblocks = XFS_DIOSTRAT_SPACE_RES(mp, resblks);
1256                         quota_flag = XFS_QMOPT_RES_REGBLKS;
1257                 }
1258
1259                 /*
1260                  * Allocate and setup the transaction.
1261                  */
1262                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
1263                 error = xfs_trans_reserve(tp, resblks,
1264                                           XFS_WRITE_LOG_RES(mp), resrtextents,
1265                                           XFS_TRANS_PERM_LOG_RES,
1266                                           XFS_WRITE_LOG_COUNT);
1267                 /*
1268                  * Check for running out of space
1269                  */
1270                 if (error) {
1271                         /*
1272                          * Free the transaction structure.
1273                          */
1274                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
1275                         xfs_trans_cancel(tp, 0);
1276                         break;
1277                 }
1278                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1279                 error = xfs_trans_reserve_quota_nblks(tp, ip, qblocks,
1280                                                       0, quota_flag);
1281                 if (error)
1282                         goto error1;
1283
1284                 xfs_trans_ijoin(tp, ip, 0);
1285
1286                 xfs_bmap_init(&free_list, &firstfsb);
1287                 error = xfs_bmapi_write(tp, ip, startoffset_fsb,
1288                                         allocatesize_fsb, alloc_type, &firstfsb,
1289                                         0, imapp, &nimaps, &free_list);
1290                 if (error) {
1291                         goto error0;
1292                 }
1293
1294                 /*
1295                  * Complete the transaction
1296                  */
1297                 error = xfs_bmap_finish(&tp, &free_list, &committed);
1298                 if (error) {
1299                         goto error0;
1300                 }
1301
1302                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1303                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1304                 if (error) {
1305                         break;
1306                 }
1307
1308                 allocated_fsb = imapp->br_blockcount;
1309
1310                 if (nimaps == 0) {
1311                         error = XFS_ERROR(ENOSPC);
1312                         break;
1313                 }
1314
1315                 startoffset_fsb += allocated_fsb;
1316                 allocatesize_fsb -= allocated_fsb;
1317         }
1318
1319         return error;
1320
1321 error0: /* Cancel bmap, unlock inode, unreserve quota blocks, cancel trans */
1322         xfs_bmap_cancel(&free_list);
1323         xfs_trans_unreserve_quota_nblks(tp, ip, (long)qblocks, 0, quota_flag);
1324
1325 error1: /* Just cancel transaction */
1326         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1327         xfs_iunlock(ip, XFS_ILOCK_EXCL);
1328         return error;
1329 }
1330
1331 /*
1332  * Zero file bytes between startoff and endoff inclusive.
1333  * The iolock is held exclusive and no blocks are buffered.
1334  *
1335  * This function is used by xfs_free_file_space() to zero
1336  * partial blocks when the range to free is not block aligned.
1337  * When unreserving space with boundaries that are not block
1338  * aligned we round up the start and round down the end
1339  * boundaries and then use this function to zero the parts of
1340  * the blocks that got dropped during the rounding.
1341  */
1342 STATIC int
1343 xfs_zero_remaining_bytes(
1344         xfs_inode_t             *ip,
1345         xfs_off_t               startoff,
1346         xfs_off_t               endoff)
1347 {
1348         xfs_bmbt_irec_t         imap;
1349         xfs_fileoff_t           offset_fsb;
1350         xfs_off_t               lastoffset;
1351         xfs_off_t               offset;
1352         xfs_buf_t               *bp;
1353         xfs_mount_t             *mp = ip->i_mount;
1354         int                     nimap;
1355         int                     error = 0;
1356
1357         /*
1358          * Avoid doing I/O beyond eof - it's not necessary
1359          * since nothing can read beyond eof.  The space will
1360          * be zeroed when the file is extended anyway.
1361          */
1362         if (startoff >= XFS_ISIZE(ip))
1363                 return 0;
1364
1365         if (endoff > XFS_ISIZE(ip))
1366                 endoff = XFS_ISIZE(ip);
1367
1368         bp = xfs_buf_get_uncached(XFS_IS_REALTIME_INODE(ip) ?
1369                                         mp->m_rtdev_targp : mp->m_ddev_targp,
1370                                   BTOBB(mp->m_sb.sb_blocksize), 0);
1371         if (!bp)
1372                 return XFS_ERROR(ENOMEM);
1373
1374         xfs_buf_unlock(bp);
1375
1376         for (offset = startoff; offset <= endoff; offset = lastoffset + 1) {
1377                 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1378                 nimap = 1;
1379                 error = xfs_bmapi_read(ip, offset_fsb, 1, &imap, &nimap, 0);
1380                 if (error || nimap < 1)
1381                         break;
1382                 ASSERT(imap.br_blockcount >= 1);
1383                 ASSERT(imap.br_startoff == offset_fsb);
1384                 lastoffset = XFS_FSB_TO_B(mp, imap.br_startoff + 1) - 1;
1385                 if (lastoffset > endoff)
1386                         lastoffset = endoff;
1387                 if (imap.br_startblock == HOLESTARTBLOCK)
1388                         continue;
1389                 ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1390                 if (imap.br_state == XFS_EXT_UNWRITTEN)
1391                         continue;
1392                 XFS_BUF_UNDONE(bp);
1393                 XFS_BUF_UNWRITE(bp);
1394                 XFS_BUF_READ(bp);
1395                 XFS_BUF_SET_ADDR(bp, xfs_fsb_to_db(ip, imap.br_startblock));
1396                 xfsbdstrat(mp, bp);
1397                 error = xfs_buf_iowait(bp);
1398                 if (error) {
1399                         xfs_buf_ioerror_alert(bp,
1400                                         "xfs_zero_remaining_bytes(read)");
1401                         break;
1402                 }
1403                 memset(bp->b_addr +
1404                         (offset - XFS_FSB_TO_B(mp, imap.br_startoff)),
1405                       0, lastoffset - offset + 1);
1406                 XFS_BUF_UNDONE(bp);
1407                 XFS_BUF_UNREAD(bp);
1408                 XFS_BUF_WRITE(bp);
1409                 xfsbdstrat(mp, bp);
1410                 error = xfs_buf_iowait(bp);
1411                 if (error) {
1412                         xfs_buf_ioerror_alert(bp,
1413                                         "xfs_zero_remaining_bytes(write)");
1414                         break;
1415                 }
1416         }
1417         xfs_buf_free(bp);
1418         return error;
1419 }
1420
1421 /*
1422  * xfs_free_file_space()
1423  *      This routine frees disk space for the given file.
1424  *
1425  *      This routine is only called by xfs_change_file_space
1426  *      for an UNRESVSP type call.
1427  *
1428  * RETURNS:
1429  *       0 on success
1430  *      errno on error
1431  *
1432  */
1433 STATIC int
1434 xfs_free_file_space(
1435         xfs_inode_t             *ip,
1436         xfs_off_t               offset,
1437         xfs_off_t               len,
1438         int                     attr_flags)
1439 {
1440         int                     committed;
1441         int                     done;
1442         xfs_fileoff_t           endoffset_fsb;
1443         int                     error;
1444         xfs_fsblock_t           firstfsb;
1445         xfs_bmap_free_t         free_list;
1446         xfs_bmbt_irec_t         imap;
1447         xfs_off_t               ioffset;
1448         xfs_extlen_t            mod=0;
1449         xfs_mount_t             *mp;
1450         int                     nimap;
1451         uint                    resblks;
1452         xfs_off_t               rounding;
1453         int                     rt;
1454         xfs_fileoff_t           startoffset_fsb;
1455         xfs_trans_t             *tp;
1456         int                     need_iolock = 1;
1457
1458         mp = ip->i_mount;
1459
1460         trace_xfs_free_file_space(ip);
1461
1462         error = xfs_qm_dqattach(ip, 0);
1463         if (error)
1464                 return error;
1465
1466         error = 0;
1467         if (len <= 0)   /* if nothing being freed */
1468                 return error;
1469         rt = XFS_IS_REALTIME_INODE(ip);
1470         startoffset_fsb = XFS_B_TO_FSB(mp, offset);
1471         endoffset_fsb = XFS_B_TO_FSBT(mp, offset + len);
1472
1473         if (attr_flags & XFS_ATTR_NOLOCK)
1474                 need_iolock = 0;
1475         if (need_iolock) {
1476                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1477                 /* wait for the completion of any pending DIOs */
1478                 inode_dio_wait(VFS_I(ip));
1479         }
1480
1481         rounding = max_t(xfs_off_t, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
1482         ioffset = offset & ~(rounding - 1);
1483         error = -filemap_write_and_wait_range(VFS_I(ip)->i_mapping,
1484                                               ioffset, -1);
1485         if (error)
1486                 goto out_unlock_iolock;
1487         truncate_pagecache_range(VFS_I(ip), ioffset, -1);
1488
1489         /*
1490          * Need to zero the stuff we're not freeing, on disk.
1491          * If it's a realtime file & can't use unwritten extents then we
1492          * actually need to zero the extent edges.  Otherwise xfs_bunmapi
1493          * will take care of it for us.
1494          */
1495         if (rt && !xfs_sb_version_hasextflgbit(&mp->m_sb)) {
1496                 nimap = 1;
1497                 error = xfs_bmapi_read(ip, startoffset_fsb, 1,
1498                                         &imap, &nimap, 0);
1499                 if (error)
1500                         goto out_unlock_iolock;
1501                 ASSERT(nimap == 0 || nimap == 1);
1502                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
1503                         xfs_daddr_t     block;
1504
1505                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1506                         block = imap.br_startblock;
1507                         mod = do_div(block, mp->m_sb.sb_rextsize);
1508                         if (mod)
1509                                 startoffset_fsb += mp->m_sb.sb_rextsize - mod;
1510                 }
1511                 nimap = 1;
1512                 error = xfs_bmapi_read(ip, endoffset_fsb - 1, 1,
1513                                         &imap, &nimap, 0);
1514                 if (error)
1515                         goto out_unlock_iolock;
1516                 ASSERT(nimap == 0 || nimap == 1);
1517                 if (nimap && imap.br_startblock != HOLESTARTBLOCK) {
1518                         ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1519                         mod++;
1520                         if (mod && (mod != mp->m_sb.sb_rextsize))
1521                                 endoffset_fsb -= mod;
1522                 }
1523         }
1524         if ((done = (endoffset_fsb <= startoffset_fsb)))
1525                 /*
1526                  * One contiguous piece to clear
1527                  */
1528                 error = xfs_zero_remaining_bytes(ip, offset, offset + len - 1);
1529         else {
1530                 /*
1531                  * Some full blocks, possibly two pieces to clear
1532                  */
1533                 if (offset < XFS_FSB_TO_B(mp, startoffset_fsb))
1534                         error = xfs_zero_remaining_bytes(ip, offset,
1535                                 XFS_FSB_TO_B(mp, startoffset_fsb) - 1);
1536                 if (!error &&
1537                     XFS_FSB_TO_B(mp, endoffset_fsb) < offset + len)
1538                         error = xfs_zero_remaining_bytes(ip,
1539                                 XFS_FSB_TO_B(mp, endoffset_fsb),
1540                                 offset + len - 1);
1541         }
1542
1543         /*
1544          * free file space until done or until there is an error
1545          */
1546         resblks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
1547         while (!error && !done) {
1548
1549                 /*
1550                  * allocate and setup the transaction. Allow this
1551                  * transaction to dip into the reserve blocks to ensure
1552                  * the freeing of the space succeeds at ENOSPC.
1553                  */
1554                 tp = xfs_trans_alloc(mp, XFS_TRANS_DIOSTRAT);
1555                 tp->t_flags |= XFS_TRANS_RESERVE;
1556                 error = xfs_trans_reserve(tp,
1557                                           resblks,
1558                                           XFS_WRITE_LOG_RES(mp),
1559                                           0,
1560                                           XFS_TRANS_PERM_LOG_RES,
1561                                           XFS_WRITE_LOG_COUNT);
1562
1563                 /*
1564                  * check for running out of space
1565                  */
1566                 if (error) {
1567                         /*
1568                          * Free the transaction structure.
1569                          */
1570                         ASSERT(error == ENOSPC || XFS_FORCED_SHUTDOWN(mp));
1571                         xfs_trans_cancel(tp, 0);
1572                         break;
1573                 }
1574                 xfs_ilock(ip, XFS_ILOCK_EXCL);
1575                 error = xfs_trans_reserve_quota(tp, mp,
1576                                 ip->i_udquot, ip->i_gdquot, ip->i_pdquot,
1577                                 resblks, 0, XFS_QMOPT_RES_REGBLKS);
1578                 if (error)
1579                         goto error1;
1580
1581                 xfs_trans_ijoin(tp, ip, 0);
1582
1583                 /*
1584                  * issue the bunmapi() call to free the blocks
1585                  */
1586                 xfs_bmap_init(&free_list, &firstfsb);
1587                 error = xfs_bunmapi(tp, ip, startoffset_fsb,
1588                                   endoffset_fsb - startoffset_fsb,
1589                                   0, 2, &firstfsb, &free_list, &done);
1590                 if (error) {
1591                         goto error0;
1592                 }
1593
1594                 /*
1595                  * complete the transaction
1596                  */
1597                 error = xfs_bmap_finish(&tp, &free_list, &committed);
1598                 if (error) {
1599                         goto error0;
1600                 }
1601
1602                 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1603                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1604         }
1605
1606  out_unlock_iolock:
1607         if (need_iolock)
1608                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1609         return error;
1610
1611  error0:
1612         xfs_bmap_cancel(&free_list);
1613  error1:
1614         xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1615         xfs_iunlock(ip, need_iolock ? (XFS_ILOCK_EXCL | XFS_IOLOCK_EXCL) :
1616                     XFS_ILOCK_EXCL);
1617         return error;
1618 }
1619
1620
1621 STATIC int
1622 xfs_zero_file_space(
1623         struct xfs_inode        *ip,
1624         xfs_off_t               offset,
1625         xfs_off_t               len,
1626         int                     attr_flags)
1627 {
1628         struct xfs_mount        *mp = ip->i_mount;
1629         uint                    granularity;
1630         xfs_off_t               start_boundary;
1631         xfs_off_t               end_boundary;
1632         int                     error;
1633
1634         granularity = max_t(uint, 1 << mp->m_sb.sb_blocklog, PAGE_CACHE_SIZE);
1635
1636         /*
1637          * Round the range of extents we are going to convert inwards.  If the
1638          * offset is aligned, then it doesn't get changed so we zero from the
1639          * start of the block offset points to.
1640          */
1641         start_boundary = round_up(offset, granularity);
1642         end_boundary = round_down(offset + len, granularity);
1643
1644         ASSERT(start_boundary >= offset);
1645         ASSERT(end_boundary <= offset + len);
1646
1647         if (!(attr_flags & XFS_ATTR_NOLOCK))
1648                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1649
1650         if (start_boundary < end_boundary - 1) {
1651                 /* punch out the page cache over the conversion range */
1652                 truncate_pagecache_range(VFS_I(ip), start_boundary,
1653                                          end_boundary - 1);
1654                 /* convert the blocks */
1655                 error = xfs_alloc_file_space(ip, start_boundary,
1656                                         end_boundary - start_boundary - 1,
1657                                         XFS_BMAPI_PREALLOC | XFS_BMAPI_CONVERT,
1658                                         attr_flags);
1659                 if (error)
1660                         goto out_unlock;
1661
1662                 /* We've handled the interior of the range, now for the edges */
1663                 if (start_boundary != offset)
1664                         error = xfs_iozero(ip, offset, start_boundary - offset);
1665                 if (error)
1666                         goto out_unlock;
1667
1668                 if (end_boundary != offset + len)
1669                         error = xfs_iozero(ip, end_boundary,
1670                                            offset + len - end_boundary);
1671
1672         } else {
1673                 /*
1674                  * It's either a sub-granularity range or the range spanned lies
1675                  * partially across two adjacent blocks.
1676                  */
1677                 error = xfs_iozero(ip, offset, len);
1678         }
1679
1680 out_unlock:
1681         if (!(attr_flags & XFS_ATTR_NOLOCK))
1682                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1683         return error;
1684
1685 }
1686
1687 /*
1688  * xfs_change_file_space()
1689  *      This routine allocates or frees disk space for the given file.
1690  *      The user specified parameters are checked for alignment and size
1691  *      limitations.
1692  *
1693  * RETURNS:
1694  *       0 on success
1695  *      errno on error
1696  *
1697  */
1698 int
1699 xfs_change_file_space(
1700         xfs_inode_t     *ip,
1701         int             cmd,
1702         xfs_flock64_t   *bf,
1703         xfs_off_t       offset,
1704         int             attr_flags)
1705 {
1706         xfs_mount_t     *mp = ip->i_mount;
1707         int             clrprealloc;
1708         int             error;
1709         xfs_fsize_t     fsize;
1710         int             setprealloc;
1711         xfs_off_t       startoffset;
1712         xfs_trans_t     *tp;
1713         struct iattr    iattr;
1714
1715         if (!S_ISREG(ip->i_d.di_mode))
1716                 return XFS_ERROR(EINVAL);
1717
1718         switch (bf->l_whence) {
1719         case 0: /*SEEK_SET*/
1720                 break;
1721         case 1: /*SEEK_CUR*/
1722                 bf->l_start += offset;
1723                 break;
1724         case 2: /*SEEK_END*/
1725                 bf->l_start += XFS_ISIZE(ip);
1726                 break;
1727         default:
1728                 return XFS_ERROR(EINVAL);
1729         }
1730
1731         /*
1732          * length of <= 0 for resv/unresv/zero is invalid.  length for
1733          * alloc/free is ignored completely and we have no idea what userspace
1734          * might have set it to, so set it to zero to allow range
1735          * checks to pass.
1736          */
1737         switch (cmd) {
1738         case XFS_IOC_ZERO_RANGE:
1739         case XFS_IOC_RESVSP:
1740         case XFS_IOC_RESVSP64:
1741         case XFS_IOC_UNRESVSP:
1742         case XFS_IOC_UNRESVSP64:
1743                 if (bf->l_len <= 0)
1744                         return XFS_ERROR(EINVAL);
1745                 break;
1746         default:
1747                 bf->l_len = 0;
1748                 break;
1749         }
1750
1751         if (bf->l_start < 0 ||
1752             bf->l_start > mp->m_super->s_maxbytes ||
1753             bf->l_start + bf->l_len < 0 ||
1754             bf->l_start + bf->l_len >= mp->m_super->s_maxbytes)
1755                 return XFS_ERROR(EINVAL);
1756
1757         bf->l_whence = 0;
1758
1759         startoffset = bf->l_start;
1760         fsize = XFS_ISIZE(ip);
1761
1762         setprealloc = clrprealloc = 0;
1763         switch (cmd) {
1764         case XFS_IOC_ZERO_RANGE:
1765                 error = xfs_zero_file_space(ip, startoffset, bf->l_len,
1766                                                 attr_flags);
1767                 if (error)
1768                         return error;
1769                 setprealloc = 1;
1770                 break;
1771
1772         case XFS_IOC_RESVSP:
1773         case XFS_IOC_RESVSP64:
1774                 error = xfs_alloc_file_space(ip, startoffset, bf->l_len,
1775                                                 XFS_BMAPI_PREALLOC, attr_flags);
1776                 if (error)
1777                         return error;
1778                 setprealloc = 1;
1779                 break;
1780
1781         case XFS_IOC_UNRESVSP:
1782         case XFS_IOC_UNRESVSP64:
1783                 if ((error = xfs_free_file_space(ip, startoffset, bf->l_len,
1784                                                                 attr_flags)))
1785                         return error;
1786                 break;
1787
1788         case XFS_IOC_ALLOCSP:
1789         case XFS_IOC_ALLOCSP64:
1790         case XFS_IOC_FREESP:
1791         case XFS_IOC_FREESP64:
1792                 /*
1793                  * These operations actually do IO when extending the file, but
1794                  * the allocation is done seperately to the zeroing that is
1795                  * done. This set of operations need to be serialised against
1796                  * other IO operations, such as truncate and buffered IO. We
1797                  * need to take the IOLOCK here to serialise the allocation and
1798                  * zeroing IO to prevent other IOLOCK holders (e.g. getbmap,
1799                  * truncate, direct IO) from racing against the transient
1800                  * allocated but not written state we can have here.
1801                  */
1802                 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1803                 if (startoffset > fsize) {
1804                         error = xfs_alloc_file_space(ip, fsize,
1805                                         startoffset - fsize, 0,
1806                                         attr_flags | XFS_ATTR_NOLOCK);
1807                         if (error) {
1808                                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1809                                 break;
1810                         }
1811                 }
1812
1813                 iattr.ia_valid = ATTR_SIZE;
1814                 iattr.ia_size = startoffset;
1815
1816                 error = xfs_setattr_size(ip, &iattr,
1817                                          attr_flags | XFS_ATTR_NOLOCK);
1818                 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1819
1820                 if (error)
1821                         return error;
1822
1823                 clrprealloc = 1;
1824                 break;
1825
1826         default:
1827                 ASSERT(0);
1828                 return XFS_ERROR(EINVAL);
1829         }
1830
1831         /*
1832          * update the inode timestamp, mode, and prealloc flag bits
1833          */
1834         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITEID);
1835
1836         if ((error = xfs_trans_reserve(tp, 0, XFS_WRITEID_LOG_RES(mp),
1837                                       0, 0, 0))) {
1838                 /* ASSERT(0); */
1839                 xfs_trans_cancel(tp, 0);
1840                 return error;
1841         }
1842
1843         xfs_ilock(ip, XFS_ILOCK_EXCL);
1844         xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1845
1846         if ((attr_flags & XFS_ATTR_DMI) == 0) {
1847                 ip->i_d.di_mode &= ~S_ISUID;
1848
1849                 /*
1850                  * Note that we don't have to worry about mandatory
1851                  * file locking being disabled here because we only
1852                  * clear the S_ISGID bit if the Group execute bit is
1853                  * on, but if it was on then mandatory locking wouldn't
1854                  * have been enabled.
1855                  */
1856                 if (ip->i_d.di_mode & S_IXGRP)
1857                         ip->i_d.di_mode &= ~S_ISGID;
1858
1859                 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1860         }
1861         if (setprealloc)
1862                 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
1863         else if (clrprealloc)
1864                 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
1865
1866         xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1867         if (attr_flags & XFS_ATTR_SYNC)
1868                 xfs_trans_set_sync(tp);
1869         return xfs_trans_commit(tp, 0);
1870 }