| blob | 58bfe629b93316168b5c2ded720b6ea9cf0b6a7a |
1 /*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc. All Rights Reserved.
3 *
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms of version 2 of the GNU General Public License as
6 * published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it would be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
11 *
12 * Further, this software is distributed without any warranty that it is
13 * free of the rightful claim of any third person regarding infringement
14 * or the like. Any license provided herein, whether implied or
15 * otherwise, applies only to this software file. Patent licenses, if
16 * any, provided herein do not apply to combinations of this program with
17 * other software, or any other product whatsoever.
18 *
19 * You should have received a copy of the GNU General Public License along
20 * with this program; if not, write the Free Software Foundation, Inc., 59
21 * Temple Place - Suite 330, Boston MA 02111-1307, USA.
22 *
23 * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
24 * Mountain View, CA 94043, or:
25 *
26 * http://www.sgi.com
27 *
28 * For further information regarding this notice, see:
29 *
30 * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
31 */
74 /*
75 * The maximum pathlen is 1024 bytes. Since the minimum file system
76 * blocksize is 512 bytes, we can get a max of 2 extents back from
77 * bmapi.
78 */
79 #define SYMLINK_MAPS 2
81 /*
82 * For xfs, we check that the file isn't too big to be opened by this kernel.
83 * No other open action is required for regular files. Devices are handled
84 * through the specfs file system, pipes through fifofs. Device and
85 * fifo vnodes are "wrapped" by specfs and fifofs vnodes, respectively,
86 * when a new vnode is first looked up or created.
87 */
88 STATIC int
92 {
103 /*
104 * If it's a directory with any blocks, read-ahead block 0
105 * as we're almost certain to have the next operation be a read there.
106 */
112 }
114 }
117 /*
118 * xfs_getattr
119 */
120 STATIC int
126 {
150 #if XFS_BIG_INUMS
152 #endif
155 /*
156 * Quick exit for non-stat callers
157 */
163 /*
164 * Copy from in-core inode.
165 */
171 /*
172 * Check vnode type block/char vs. everything else.
173 */
185 #if 0
186 /* Large block sizes confuse various
187 * user space programs, so letting the
188 * stripe size through is not a good
189 * idea for now.
190 */
192 /*
193 * If the underlying volume is a stripe, then
194 * return the stripe width in bytes as the
195 * recommended I/O size.
196 */
198 /*
199 * Return the largest of the preferred buffer
200 * sizes since doing small I/Os into larger
201 * buffers causes buffers to be decommissioned.
202 * The value returned is in bytes.
203 */
207 #else
209 /*
210 * Return the largest of the preferred buffer
211 * sizes since doing small I/Os into larger
212 * buffers causes buffers to be decommissioned.
213 * The value returned is in bytes.
214 */
217 #endif
220 /*
221 * If the file blocks are being allocated from a
222 * realtime partition, then return the inode's
223 * realtime extent size or the realtime volume's
224 * extent size.
225 */
229 }
231 }
240 /*
241 * Exit for stat callers. See if any of the rest of the fields
242 * to be filled in are needed.
243 */
249 /*
250 * Convert di_flags to xflags.
251 */
254 /*
255 * Exit for inode revalidate. See if any of the rest of
256 * the fields to be filled in are needed.
257 */
273 else
277 all_done:
281 }
284 /*
285 * xfs_setattr
286 */
287 int
293 {
299 uint lock_flags;
317 /*
318 * Cannot set certain attributes.
319 */
323 }
331 /*
332 * Timestamps do not need to be logged and hence do not
333 * need to be done within a transaction.
334 */
342 }
347 /*
348 * If disk quotas is on, we make sure that the dquots do exist on disk,
349 * before we start any other transactions. Trying to do this later
350 * is messy. We don't care to take a readlock to look at the ids
351 * in inode here, because we can't hold it across the trans_reserve.
352 * If the IDs do change before we take the ilock, we're covered
353 * because the i_*dquot fields will get updated anyway.
354 */
364 }
370 }
376 }
377 /*
378 * We take a reference when we initialize udqp and gdqp,
379 * so it is important that we never blindly double trip on
380 * the same variable. See xfs_create() for an example.
381 */
388 }
390 /*
391 * For the other attributes, we acquire the inode lock and
392 * first do an error checking pass.
393 */
409 }
410 }
420 }
421 }
424 }
428 /* boolean: are we the file owner? */
431 /*
432 * Change various properties of a file.
433 * Only the owner or users with CAP_FOWNER
434 * capability may do these things.
435 */
439 /*
440 * CAP_FOWNER overrides the following restrictions:
441 *
442 * The user ID of the calling process must be equal
443 * to the file owner ID, except in cases where the
444 * CAP_FSETID capability is applicable.
445 */
449 }
451 /*
452 * CAP_FSETID overrides the following restrictions:
453 *
454 * The effective user ID of the calling process shall match
455 * the file owner when setting the set-user-ID and
456 * set-group-ID bits on that file.
457 *
458 * The effective group ID or one of the supplementary group
459 * IDs of the calling process shall match the group owner of
460 * the file when setting the set-group-ID bit on that file
461 */
470 #if 0
471 /* Linux allows this, Irix doesn't. */
474 #endif
477 }
478 }
480 /*
481 * Change file ownership. Must be the owner or privileged.
482 * If the system was configured with the "restricted_chown"
483 * option, the owner is not permitted to give away the file,
484 * and can change the group id only to a group of which he
485 * or she is a member.
486 */
488 /*
489 * These IDs could have changed since we last looked at them.
490 * But, we're assured that if the ownership did change
491 * while we didn't have the inode locked, inode's dquot(s)
492 * would have changed also.
493 */
500 iprojid;
502 /*
503 * CAP_CHOWN overrides the following restrictions:
504 *
505 * If _POSIX_CHOWN_RESTRICTED is defined, this capability
506 * shall override the restriction that a process cannot
507 * change the user ID of a file it owns and the restriction
508 * that the group ID supplied to the chown() function
509 * shall be equal to either the group ID or one of the
510 * supplementary group IDs of the calling process.
511 */
518 }
519 /*
520 * Do a quota reservation only if uid/projid/gid is actually
521 * going to change.
522 */
532 }
533 }
535 /*
536 * Truncate file. Must have write permission and not be a directory.
537 */
539 /* Short circuit the truncate case for zero length files */
548 }
556 }
557 /*
558 * Make sure that the dquots are attached to the inode.
559 */
562 }
564 /*
565 * Change file access or modified times.
566 */
573 }
574 }
575 }
577 /*
578 * Change extent size or realtime flag.
579 */
581 /*
582 * Can't change extent size if any extents are allocated.
583 */
590 }
592 /*
593 * Can't set extent size unless the file is marked, or
594 * about to be marked as a realtime file.
595 *
596 * This check will be removed when fixed size extents
597 * with buffered data writes is implemented.
598 *
599 */
608 }
610 /*
611 * Can't change realtime flag if any extents are allocated.
612 */
618 }
619 /*
620 * Extent size must be a multiple of the appropriate block
621 * size, if set at all.
622 */
624 xfs_extlen_t size;
633 }
637 }
638 }
639 /*
640 * If realtime flag is set then must have realtime data.
641 */
649 }
650 }
652 /*
653 * Can't modify an immutable/append-only file unless
654 * we have appropriate permission.
655 */
664 }
665 }
667 /*
668 * Now we can make the changes. Before we join the inode
669 * to the transaction, if XFS_AT_SIZE is set then take care of
670 * the part of the truncation that must be done without the
671 * inode lock. This needs to be done before joining the inode
672 * to the transaction, because the inode cannot be unlocked
673 * once it is a part of the transaction.
674 */
687 }
691 XFS_TRANS_PERM_LOG_RES,
692 XFS_ITRUNCATE_LOG_COUNT))) {
697 }
700 }
705 }
707 /* determine whether mandatory locking mode changes */
710 /*
711 * Truncate file. Must have write permission and not be a directory.
712 */
719 /*
720 * signal a sync transaction unless
721 * we're truncating an already unlinked
722 * file on a wsync filesystem
723 */
726 XFS_DATA_FORK,
732 }
733 }
734 /*
735 * Have to do this even if the file's size doesn't change.
736 */
738 }
740 /*
741 * Change file access modes.
742 */
749 }
751 /*
752 * Change file ownership. Must be the owner or privileged.
753 * If the system was configured with the "restricted_chown"
754 * option, the owner is not permitted to give away the file,
755 * and can change the group id only to a group of which he
756 * or she is a member.
757 */
759 /*
760 * CAP_FSETID overrides the following restrictions:
761 *
762 * The set-user-ID and set-group-ID bits of a file will be
763 * cleared upon successful return from chown()
764 */
768 }
770 /*
771 * Change the ownerships and register quota modifications
772 * in the transaction.
773 */
780 }
782 }
790 }
792 }
800 }
802 /*
803 * We may have to rev the inode as well as
804 * the superblock version number since projids didn't
805 * exist before DINODE_VERSION_2 and SB_VERSION_NLINK.
806 */
809 }
813 }
816 /*
817 * Change file access or modified times.
818 */
825 }
831 }
834 }
836 /*
837 * Change XFS-added attributes.
838 */
841 /*
842 * Converting bytes to fs blocks.
843 */
846 }
848 uint di_flags;
850 /* can't set PREALLOC this way, just preserve it */
875 }
876 }
878 }
881 }
883 /*
884 * Change file inode change time only if XFS_AT_CTIME set
885 * AND we have been called by a DMI function.
886 */
893 }
895 /*
896 * Send out timestamp changes that need to be set to the
897 * current time. Not done when called by a DMI function.
898 */
904 /*
905 * If this is a synchronous mount, make sure that the
906 * transaction goes to disk before returning to the user.
907 * This is slightly sub-optimal in that truncates require
908 * two sync transactions instead of one for wsync filesytems.
909 * One for the truncate and one for the timestamps since we
910 * don't want to change the timestamps unless we're sure the
911 * truncate worked. Truncates are less than 1% of the laddis
912 * mix so this probably isn't worth the trouble to optimize.
913 */
920 }
922 /*
923 * If the (regular) file's mandatory locking mode changed, then
924 * notify the vnode. We do this under the inode lock to prevent
925 * racing calls to vop_vnode_change.
926 */
930 mandlock_after);
931 }
935 /*
936 * Release any dquot(s) the inode had kept before chown.
937 */
945 }
952 }
955 abort_return:
957 /* FALLTHROUGH */
958 error_return:
963 }
966 }
968 }
971 /*
972 * xfs_access
973 * Null conversion from vnode mode bits to inode mode bits, as in efs.
974 */
975 STATIC int
980 {
992 }
995 /*
996 * xfs_readlink
997 *
998 */
999 STATIC int
1005 {
1008 xfs_off_t offset;
1015 xfs_daddr_t d;
1039 }
1043 }
1047 }
1049 /*
1050 * See if the symlink is stored inline.
1051 */
1056 }
1058 /*
1059 * Symlink not inline. Call bmap to get it in.
1060 */
1068 }
1081 }
1088 }
1090 }
1093 error_return:
1098 }
1101 /*
1102 * xfs_fsync
1103 *
1104 * This is called to sync the inode and its data out to disk.
1105 * We need to hold the I/O lock while flushing the data, and
1106 * the inode lock while flushing the inode. The inode lock CANNOT
1107 * be held while flushing the data, so acquire after we're done
1108 * with that.
1109 */
1110 STATIC int
1115 xfs_off_t start,
1116 xfs_off_t stop)
1117 {
1132 /*
1133 * We always need to make sure that the required inode state
1134 * is safe on disk. The vnode might be clean but because
1135 * of committed transactions that haven't hit the disk yet.
1136 * Likewise, there could be unflushed non-transactional
1137 * changes to the inode core that have to go to disk.
1138 *
1139 * The following code depends on one assumption: that
1140 * any transaction that changes an inode logs the core
1141 * because it has to change some field in the inode core
1142 * (typically nextents or nblocks). That assumption
1143 * implies that any transactions against an inode will
1144 * catch any non-transactional updates. If inode-altering
1145 * transactions exist that violate this assumption, the
1146 * code breaks. Right now, it figures that if the involved
1147 * update_* field is clear and the inode is unpinned, the
1148 * inode is clean. Either it's been flushed or it's been
1149 * committed and the commit has hit the disk unpinning the inode.
1150 * (Note that xfs_inode_item_format() called at commit clears
1151 * the update_* fields.)
1152 */
1155 /* If we are flushing data then we care about update_size
1156 * being set, otherwise we care about update_core
1157 */
1161 /*
1162 * Timestamps/size haven't changed since last inode
1163 * flush or inode transaction commit. That means
1164 * either nothing got written or a transaction
1165 * committed which caught the updates. If the
1166 * latter happened and the transaction hasn't
1167 * hit the disk yet, the inode will be still
1168 * be pinned. If it is, force the log.
1169 */
1175 XFS_LOG_FORCE |
1178 }
1181 /*
1182 * Kick off a transaction to log the inode
1183 * core to get the updates. Make it
1184 * sync if FSYNC_WAIT is passed in (which
1185 * is done by everybody but specfs). The
1186 * sync transaction will also force the log.
1187 */
1195 }
1198 /*
1199 * Note - it's possible that we might have pushed
1200 * ourselves out of the way during trans_reserve
1201 * which would flush the inode. But there's no
1202 * guarantee that the inode buffer has actually
1203 * gone out yet (it's delwri). Plus the buffer
1204 * could be pinned anyway if it's part of an
1205 * inode in another recent transaction. So we
1206 * play it safe and fire off the transaction anyway.
1207 */
1216 }
1218 }
1220 /*
1221 * This is called by xfs_inactive to free any blocks beyond eof,
1222 * when the link count isn't zero.
1223 */
1224 STATIC int
1228 {
1231 xfs_fileoff_t end_fsb;
1232 xfs_fileoff_t last_fsb;
1233 xfs_filblks_t map_len;
1235 xfs_bmbt_irec_t imap;
1237 /*
1238 * Figure out if there are any blocks beyond the end
1239 * of the file. If not, then there is nothing to do.
1240 */
1255 /*
1256 * Attach the dquots to the inode up front.
1257 */
1261 /*
1262 * There are blocks after the end of file.
1263 * Free them up now by truncating the file to
1264 * its current size.
1265 */
1268 /*
1269 * Do the xfs_itruncate_start() call before
1270 * reserving any log space because
1271 * itruncate_start will call into the buffer
1272 * cache and we can't
1273 * do that within a transaction.
1274 */
1282 XFS_ITRUNCATE_LOG_COUNT);
1288 }
1292 XFS_IOLOCK_EXCL |
1293 XFS_ILOCK_EXCL);
1298 XFS_DATA_FORK,
1300 /*
1301 * If we get an error at this point we
1302 * simply don't bother truncating the file.
1303 */
1307 XFS_TRANS_ABORT));
1310 XFS_TRANS_RELEASE_LOG_RES,
1311 NULL);
1312 }
1314 }
1316 }
1318 /*
1319 * Free a symlink that has blocks associated with it.
1320 */
1321 STATIC int
1325 {
1330 xfs_fsblock_t first_block;
1331 xfs_bmap_free_t free_list;
1343 /*
1344 * We're freeing a symlink that has some
1345 * blocks allocated to it. Free the
1346 * blocks here. We know that we've got
1347 * either 1 or 2 extents and that we can
1348 * free them all in one bunmapi call.
1349 */
1357 }
1358 /*
1359 * Lock the inode, fix the size, and join it to the transaction.
1360 * Hold it so in the normal path, we still have it locked for
1361 * the second transaction. In the error paths we need it
1362 * held so the cancel won't rele it, see below.
1363 */
1370 /*
1371 * Find the block(s) so we can inval and unmap them.
1372 */
1380 /*
1381 * Invalidate the block(s).
1382 */
1388 }
1389 /*
1390 * Unmap the dead block(s) to the free_list.
1391 */
1396 /*
1397 * Commit the first transaction. This logs the EFI and the inode.
1398 */
1401 /*
1402 * The transaction must have been committed, since there were
1403 * actually extents freed by xfs_bunmapi. See xfs_bmap_finish.
1404 * The new tp has the extent freeing and EFDs.
1405 */
1407 /*
1408 * The first xact was committed, so add the inode to the new one.
1409 * Mark it dirty so it will be logged and moved forward in the log as
1410 * part of every commit.
1411 */
1415 /*
1416 * Get a new, empty transaction to return to our caller.
1417 */
1419 /*
1420 * Commit the transaction containing extent freeing and EFD's.
1421 * If we get an error on the commit here or on the reserve below,
1422 * we need to unlock the inode since the new transaction doesn't
1423 * have the inode attached.
1424 */
1430 }
1431 /*
1432 * Remove the memory for extent descriptions (just bookkeeping).
1433 */
1437 /*
1438 * Put an itruncate log reservation in the new transaction
1439 * for our caller.
1440 */
1445 }
1446 /*
1447 * Return with the inode locked but not joined to the transaction.
1448 */
1452 error1:
1454 error0:
1455 /*
1456 * Have to come here with the inode locked and either
1457 * (held and in the transaction) or (not in the transaction).
1458 * If the inode isn't held then cancel would iput it, but
1459 * that's wrong since this is inactive and the vnode ref
1460 * count is 0 already.
1461 * Cancel won't do anything to the inode if held, but it still
1462 * needs to be locked until the cancel is done, if it was
1463 * joined to the transaction.
1464 */
1470 }
1472 STATIC int
1476 {
1480 /*
1481 * We're freeing a symlink which fit into
1482 * the inode. Just free the memory used
1483 * to hold the old symlink.
1484 */
1488 XFS_ITRUNCATE_LOG_COUNT);
1494 }
1497 /*
1498 * Zero length symlinks _can_ exist.
1499 */
1503 XFS_DATA_FORK);
1505 }
1507 }
1509 /*
1510 *
1511 */
1512 STATIC int
1516 {
1533 }
1539 XFS_INACTIVE_LOG_COUNT);
1546 }
1557 }
1559 STATIC int
1562 {
1573 }
1575 /* If this is a read-only mount, don't do this (would generate I/O) */
1579 #ifdef HAVE_REFCACHE
1580 /* If we are in the NFS reference cache then don't do this now */
1583 #endif
1594 /* Update linux inode block count after free above */
1597 }
1598 }
1601 }
1603 /*
1604 * xfs_inactive
1605 *
1606 * This is called when the vnode reference count for the vnode
1607 * goes to zero. If the file has been unlinked, then it must
1608 * now be truncated. Also, we clear all of the read-ahead state
1609 * kept for the inode here since the file is now closed.
1610 */
1611 STATIC int
1615 {
1618 xfs_bmap_free_t free_list;
1619 xfs_fsblock_t first_block;
1631 /*
1632 * If the inode is already free, then there can be nothing
1633 * to clean up here.
1634 */
1639 }
1641 /*
1642 * Only do a truncate if it's a regular file with
1643 * some actual space in it. It's OK to look at the
1644 * inode's fields without the lock because we're the
1645 * only one with a reference to the inode.
1646 */
1656 }
1660 /* If this is a read-only mount, don't do this (would generate I/O) */
1672 /* Update linux inode block count after free above */
1675 }
1677 }
1686 /*
1687 * Do the xfs_itruncate_start() call before
1688 * reserving any log space because itruncate_start
1689 * will call into the buffer cache and we can't
1690 * do that within a transaction.
1691 */
1699 XFS_ITRUNCATE_LOG_COUNT);
1701 /* Don't call itruncate_cleanup */
1706 }
1712 /*
1713 * normally, we have to run xfs_itruncate_finish sync.
1714 * But if filesystem is wsync and we're in the inactive
1715 * path, then we know that nlink == 0, and that the
1716 * xaction that made nlink == 0 is permanently committed
1717 * since xfs_remove runs as a synchronous transaction.
1718 */
1727 }
1730 /*
1731 * If we get an error while cleaning up a
1732 * symlink we bail out.
1733 */
1741 }
1749 XFS_INACTIVE_LOG_COUNT);
1754 }
1759 }
1761 /*
1762 * If there are attributes associated with the file
1763 * then blow them away now. The code calls a routine
1764 * that recursively deconstructs the attribute fork.
1765 * We need to just commit the current transaction
1766 * because we can't use it for xfs_attr_inactive().
1767 */
1770 /*
1771 * If we got an error, the transaction is already
1772 * cancelled, and the inode is unlocked. Just get out.
1773 */
1778 }
1780 /*
1781 * Free the inode.
1782 */
1786 /*
1787 * If we fail to free the inode, shut down. The cancel
1788 * might do that, we need to make sure. Otherwise the
1789 * inode might be lost for a long time or forever.
1790 */
1796 }
1799 /*
1800 * Credit the quota account(s). The inode is gone.
1801 */
1804 /*
1805 * Just ignore errors at this point. There is
1806 * nothing we can do except to try to keep going.
1807 */
1811 }
1812 /*
1813 * Release the dquots held by inode, if any.
1814 */
1819 out:
1821 }
1824 /*
1825 * xfs_lookup
1826 */
1827 STATIC int
1835 {
1837 xfs_ino_t e_inum;
1839 uint lock_mode;
1855 }
1858 }
1861 /*
1862 * xfs_create (create a new file).
1863 */
1864 STATIC int
1871 {
1878 xfs_dev_t rdev;
1880 xfs_bmap_free_t free_list;
1881 xfs_fsblock_t first_block;
1882 boolean_t dp_joined_to_trans;
1884 uint cancel_flags;
1886 xfs_prid_t prid;
1888 uint resblks;
1911 }
1916 /* Return through std_return after this point. */
1923 else
1926 /*
1927 * Make sure that we have allocated dquot(s) on disk.
1928 */
1941 /*
1942 * Initially assume that the file does not exist and
1943 * reserve the resources for that case. If that is not
1944 * the case we'll drop the one we have and get a more
1945 * appropriate transaction later.
1946 */
1953 }
1958 }
1966 /*
1967 * Reserve disk quota and the inode.
1968 */
1984 }
1987 /*
1988 * At this point, we've gotten a newly allocated inode.
1989 * It is locked (and joined to the transaction).
1990 */
1994 /*
1995 * Now we join the directory inode to the transaction.
1996 * We do not do it earlier because xfs_dir_ialloc
1997 * might commit the previous transaction (and release
1998 * all the locks).
1999 */
2011 }
2015 /*
2016 * If this is a synchronous mount, make sure that the
2017 * create transaction goes to disk before returning to
2018 * the user.
2019 */
2022 }
2026 /*
2027 * Attach the dquot(s) to the inodes and modify them incore.
2028 * These ids of the inode couldn't have changed since the new
2029 * inode has been locked ever since it was created.
2030 */
2033 /*
2034 * xfs_trans_commit normally decrements the vnode ref count
2035 * when it unlocks the inode. Since we want to return the
2036 * vnode to the caller, we bump the vnode ref count now.
2037 */
2045 }
2052 }
2057 /*
2058 * Propogate the fact that the vnode changed after the
2059 * xfs_inode locks have been released.
2060 */
2065 /* Fallthrough to std_return with error = 0 */
2067 std_return:
2070 DM_EVENT_POSTCREATE)) {
2076 }
2079 abort_return:
2081 /* FALLTHROUGH */
2082 error_return:
2094 abort_rele:
2095 /*
2096 * Wait until after the current transaction is aborted to
2097 * release the inode. This prevents recursive transactions
2098 * and deadlocks from xfs_inactive.
2099 */
2108 }
2110 #ifdef DEBUG
2111 /*
2112 * Some counters to see if (and how often) we are hitting some deadlock
2113 * prevention code paths.
2114 */
2119 #endif
2121 /*
2122 * The following routine will lock the inodes associated with the
2123 * directory and the named entry in the directory. The locks are
2124 * acquired in increasing inode number.
2125 *
2126 * If the entry is "..", then only the directory is locked. The
2127 * vnode ref count will still include that from the .. entry in
2128 * this case.
2129 *
2130 * There is a deadlock we need to worry about. If the locked directory is
2131 * in the AIL, it might be blocking up the log. The next inode we lock
2132 * could be already locked by another thread waiting for log space (e.g
2133 * a permanent log reservation with a long running transaction (see
2134 * xfs_itruncate_finish)). To solve this, we must check if the directory
2135 * is in the ail and use lock_nowait. If we can't lock, we need to
2136 * drop the inode lock on the directory and try again. xfs_iunlock will
2137 * potentially push the tail if we were holding up the log.
2138 */
2139 STATIC int
2144 {
2146 xfs_ino_t e_inum;
2150 #ifdef DEBUG
2151 xfs_rm_locks++;
2152 #endif
2155 again:
2162 /*
2163 * We want to lock in increasing inum. Since we've already
2164 * acquired the lock on the directory, we may need to release
2165 * if if the inum of the entry turns out to be less.
2166 */
2168 /*
2169 * We are already in the right order, so just
2170 * lock on the inode of the entry.
2171 * We need to use nowait if dp is in the AIL.
2172 */
2177 attempts++;
2178 #ifdef DEBUG
2179 xfs_rm_attempts++;
2180 #endif
2182 /*
2183 * Unlock dp and try again.
2184 * xfs_iunlock will try to push the tail
2185 * if the inode is in the AIL.
2186 */
2192 #ifdef DEBUG
2193 xfs_rm_lock_delays++;
2194 #endif
2195 }
2197 }
2200 }
2207 }
2208 /* else e_inum == dp->i_ino */
2209 /* This can happen if we're asked to lock /x/..
2210 * the entry is "..", which is also the parent directory.
2211 */
2214 }
2216 #ifdef DEBUG
2222 #endif
2224 /*
2225 * The following routine will lock n inodes in exclusive mode.
2226 * We assume the caller calls us with the inodes in i_ino order.
2227 *
2228 * We need to detect deadlock where an inode that we lock
2229 * is in the AIL and we start waiting for another inode that is locked
2230 * by a thread in a long running transaction (such as truncate). This can
2231 * result in deadlock since the long running trans might need to wait
2232 * for the inode we just locked in order to push the tail and free space
2233 * in the log.
2234 */
2235 void
2240 uint lock_mode)
2241 {
2253 }
2255 again:
2262 /*
2263 * If try_lock is not set yet, make sure all locked inodes
2264 * are not in the AIL.
2265 * If any are, set try_lock to be used later.
2266 */
2272 try_lock++;
2273 }
2274 }
2275 }
2277 /*
2278 * If any of the previous locks we have locked is in the AIL,
2279 * we must TRY to get the second and subsequent locks. If
2280 * we can't get any, we must release all we have
2281 * and try again.
2282 */
2285 /* try_lock must be 0 if i is 0. */
2286 /*
2287 * try_lock means we have an inode locked
2288 * that is in the AIL.
2289 */
2292 attempts++;
2294 /*
2295 * Unlock all previous guys and try again.
2296 * xfs_iunlock will try to push the tail
2297 * if the inode is in the AIL.
2298 */
2302 /*
2303 * Check to see if we've already
2304 * unlocked this one.
2305 * Not the first one going back,
2306 * and the inode ptr is the same.
2307 */
2313 }
2317 #ifdef DEBUG
2318 xfs_lock_delays++;
2319 #endif
2320 }
2324 }
2327 }
2328 }
2330 #ifdef DEBUG
2336 xfs_locked_n++;
2337 }
2338 #endif
2339 }
2341 #ifdef DEBUG
2342 #define REMOVE_DEBUG_TRACE(x) {remove_which_error_return = (x);}
2345 #define REMOVE_DEBUG_TRACE(x)
2349 /*
2350 * xfs_remove
2351 *
2352 */
2353 STATIC int
2358 {
2365 xfs_bmap_free_t free_list;
2366 xfs_fsblock_t first_block;
2371 uint resblks;
2391 }
2393 /* From this point on, return through std_return */
2396 /*
2397 * We need to get a reference to ip before we get our log
2398 * reservation. The reason for this is that we cannot call
2399 * xfs_iget for an inode for which we do not have a reference
2400 * once we've acquired a log reservation. This is because the
2401 * inode we are trying to get might be in xfs_inactive going
2402 * for a log reservation. Since we'll have to wait for the
2403 * inactive code to complete before returning from xfs_iget,
2404 * we need to make sure that we don't have log space reserved
2405 * when we call xfs_iget. Instead we get an unlocked referece
2406 * to the inode before getting our log reservation.
2407 */
2412 }
2427 }
2431 /*
2432 * We try to get the real space reservation first,
2433 * allowing for directory btree deletion(s) implying
2434 * possible bmap insert(s). If we can't get the space
2435 * reservation then we use 0 instead, and avoid the bmap
2436 * btree insert(s) in the directory code by, if the bmap
2437 * insert tries to happen, instead trimming the LAST
2438 * block from the directory.
2439 */
2447 }
2454 }
2462 }
2464 /*
2465 * At this point, we've gotten both the directory and the entry
2466 * inodes locked.
2467 */
2470 /*
2471 * Increment vnode ref count only in this case since
2472 * there's an extra vnode reference in the case where
2473 * dp == ip.
2474 */
2477 }
2479 /*
2480 * Entry must exist since we did a lookup in xfs_lock_dir_and_entry.
2481 */
2489 }
2499 }
2501 /* Determine if this is the last link while
2502 * we are in the transaction.
2503 */
2506 /*
2507 * Take an extra ref on the inode so that it doesn't
2508 * go to xfs_inactive() from within the commit.
2509 */
2512 /*
2513 * If this is a synchronous mount, make sure that the
2514 * remove transaction goes to disk before returning to
2515 * the user.
2516 */
2519 }
2525 }
2531 }
2533 /*
2534 * Before we drop our extra reference to the inode, purge it
2535 * from the refcache if it is there. By waiting until afterwards
2536 * to do the IRELE, we ensure that we won't go inactive in the
2537 * xfs_refcache_purge_ip routine (although that would be OK).
2538 */
2543 /*
2544 * Let interposed file systems know about removed links.
2545 */
2550 /* Fall through to std_return with error = 0 */
2551 std_return:
2553 DM_EVENT_POSTREMOVE)) {
2558 }
2561 error1:
2567 error_rele:
2568 /*
2569 * In this case make sure to not release the inode until after
2570 * the current transaction is aborted. Releasing it beforehand
2571 * can cause us to go to xfs_inactive and start a recursive
2572 * transaction which can easily deadlock with the current one.
2573 */
2578 /*
2579 * Before we drop our extra reference to the inode, purge it
2580 * from the refcache if it is there. By waiting until afterwards
2581 * to do the IRELE, we ensure that we won't go inactive in the
2582 * xfs_refcache_purge_ip routine (although that would be OK).
2583 */
2589 }
2592 /*
2593 * xfs_link
2594 *
2595 */
2596 STATIC int
2602 {
2608 xfs_bmap_free_t free_list;
2609 xfs_fsblock_t first_block;
2640 }
2642 /* Return through std_return after this point. */
2659 }
2663 }
2671 }
2675 /*
2676 * Increment vnode ref counts since xfs_trans_commit &
2677 * xfs_trans_cancel will both unlock the inodes and
2678 * decrement the associated ref counts.
2679 */
2685 /*
2686 * If the source has too many links, we can't make any more to it.
2687 */
2691 }
2693 /*
2694 * If we are using project inheritance, we only allow hard link
2695 * creation in our tree when the project IDs are the same; else
2696 * the tree quota mechanism could be circumvented.
2697 */
2702 }
2706 target_namelen)))
2713 resblks);
2723 }
2725 /*
2726 * If this is a synchronous mount, make sure that the
2727 * link transaction goes to disk before returning to
2728 * the user.
2729 */
2732 }
2738 }
2743 }
2745 /* Fall through to std_return with error = 0. */
2746 std_return:
2748 DM_EVENT_POSTLINK)) {
2753 }
2756 abort_return:
2758 /* FALLTHROUGH */
2759 error_return:
2763 }
2764 /*
2765 * xfs_mkdir
2766 *
2767 */
2768 STATIC int
2775 {
2785 xfs_bmap_free_t free_list;
2786 xfs_fsblock_t first_block;
2788 boolean_t dp_joined_to_trans;
2791 xfs_prid_t prid;
2793 uint resblks;
2818 }
2820 /* Return through std_return after this point. */
2830 else
2833 /*
2834 * Make sure that we have allocated dquot(s) on disk.
2835 */
2850 XFS_TRANS_PERM_LOG_RES,
2851 XFS_MKDIR_LOG_COUNT);
2852 }
2857 }
2861 /*
2862 * Check for directory link count overflow.
2863 */
2867 }
2869 /*
2870 * Reserve disk quota and the inode.
2871 */
2879 /*
2880 * create the directory inode.
2881 */
2889 }
2892 /*
2893 * Now we add the directory inode to the transaction.
2894 * We waited until now since xfs_dir_ialloc might start
2895 * a new transaction. Had we joined the transaction
2896 * earlier, the locks might have gotten released.
2897 */
2910 }
2913 /*
2914 * Bump the in memory version number of the parent directory
2915 * so that other processes accessing it will recognize that
2916 * the directory has changed.
2917 */
2923 }
2929 }
2938 /*
2939 * Attach the dquots to the new inode and modify the icount incore.
2940 */
2943 /*
2944 * If this is a synchronous mount, make sure that the
2945 * mkdir transaction goes to disk before returning to
2946 * the user.
2947 */
2950 }
2956 }
2963 }
2965 /* Fall through to std_return with error = 0 or errno from
2966 * xfs_trans_commit. */
2968 std_return:
2971 DM_EVENT_POSTCREATE)) {
2975 DM_RIGHT_NULL,
2978 }
2981 error2:
2982 error1:
2984 abort_return:
2986 error_return:
2993 }
2996 }
2999 /*
3000 * xfs_rmdir
3001 *
3002 */
3003 STATIC int
3008 {
3015 xfs_bmap_free_t free_list;
3016 xfs_fsblock_t first_block;
3023 uint resblks;
3042 }
3044 /* Return through std_return after this point. */
3048 /*
3049 * We need to get a reference to cdp before we get our log
3050 * reservation. The reason for this is that we cannot call
3051 * xfs_iget for an inode for which we do not have a reference
3052 * once we've acquired a log reservation. This is because the
3053 * inode we are trying to get might be in xfs_inactive going
3054 * for a log reservation. Since we'll have to wait for the
3055 * inactive code to complete before returning from xfs_iget,
3056 * we need to make sure that we don't have log space reserved
3057 * when we call xfs_iget. Instead we get an unlocked referece
3058 * to the inode before getting our log reservation.
3059 */
3064 }
3068 /*
3069 * Get the dquots for the inodes.
3070 */
3078 }
3082 /*
3083 * We try to get the real space reservation first,
3084 * allowing for directory btree deletion(s) implying
3085 * possible bmap insert(s). If we can't get the space
3086 * reservation then we use 0 instead, and avoid the bmap
3087 * btree insert(s) in the directory code by, if the bmap
3088 * insert tries to happen, instead trimming the LAST
3089 * block from the directory.
3090 */
3098 }
3104 }
3107 /*
3108 * Now lock the child directory inode and the parent directory
3109 * inode in the proper order. This will take care of validating
3110 * that the directory entry for the child directory inode has
3111 * not changed while we were obtaining a log reservation.
3112 */
3118 }
3122 /*
3123 * Only increment the parent directory vnode count if
3124 * we didn't bump it in looking up cdp. The only time
3125 * we don't bump it is when we're looking up ".".
3126 */
3128 }
3137 }
3141 }
3147 }
3151 /*
3152 * Bump the in memory generation count on the parent
3153 * directory so that other can know that it has changed.
3154 */
3157 /*
3158 * Drop the link from cdp's "..".
3159 */
3163 }
3165 /*
3166 * Drop the link from dp to cdp.
3167 */
3171 }
3173 /*
3174 * Drop the "." link from cdp to self.
3175 */
3179 }
3181 /* Determine these before committing transaction */
3184 /*
3185 * Take an extra ref on the child vnode so that it
3186 * does not go to xfs_inactive() from within the commit.
3187 */
3190 /*
3191 * If this is a synchronous mount, make sure that the
3192 * rmdir transaction goes to disk before returning to
3193 * the user.
3194 */
3197 }
3203 XFS_TRANS_ABORT));
3206 }
3212 }
3215 /*
3216 * Let interposed file systems know about removed links.
3217 */
3222 /* Fall through to std_return with error = 0 or the errno
3223 * from xfs_trans_commit. */
3224 std_return:
3231 }
3234 error1:
3237 error_return:
3240 }
3243 /*
3244 * xfs_readdir
3245 *
3246 * Read dp's entries starting at uiop->uio_offset and translate them into
3247 * bufsize bytes worth of struct dirents starting at bufbase.
3248 */
3249 STATIC int
3255 {
3259 uint lock_mode;
3260 xfs_off_t start_offset;
3268 }
3275 }
3278 }
3281 /*
3282 * xfs_symlink
3283 *
3284 */
3285 STATIC int
3293 {
3300 xfs_bmap_free_t free_list;
3301 xfs_fsblock_t first_block;
3302 boolean_t dp_joined_to_trans;
3304 uint cancel_flags;
3306 xfs_fileoff_t first_fsb;
3307 xfs_filblks_t fs_blocks;
3310 xfs_daddr_t d;
3315 xfs_prid_t prid;
3317 uint resblks;
3338 /*
3339 * Check component lengths of the target path name.
3340 */
3349 /*
3350 * Skip any slashes.
3351 */
3353 total++;
3354 path++;
3355 }
3357 /*
3358 * Count up to the next slash or end of path.
3359 * Error out if the component is bigger than MAXNAMELEN.
3360 */
3365 }
3366 }
3367 }
3368 }
3376 }
3378 /* Return through std_return after this point. */
3385 else
3388 /*
3389 * Make sure that we have allocated dquot(s) on disk.
3390 */
3399 /*
3400 * The symlink will fit into the inode data fork?
3401 * There can't be any attributes so we get the whole variable part.
3402 */
3405 else
3414 }
3419 }
3423 /*
3424 * Check whether the directory allows new symlinks or not.
3425 */
3429 }
3431 /*
3432 * Reserve disk quota : blocks and inode.
3433 */
3438 /*
3439 * Check for ability to enter directory entry, if no space reserved.
3440 */
3444 /*
3445 * Initialize the bmap freelist prior to calling either
3446 * bmapi or the directory create code.
3447 */
3450 /*
3451 * Allocate an inode for the symlink.
3452 */
3459 }
3466 /*
3467 * Also attach the dquot(s) to it, if applicable.
3468 */
3473 /*
3474 * If the symlink will fit into the inode, write it inline.
3475 */
3481 /*
3482 * The inode was initially created in extent format.
3483 */
3500 }
3516 }
3523 }
3524 }
3526 /*
3527 * Create the directory entry for the symlink.
3528 */
3533 }
3537 /*
3538 * Bump the in memory version number of the parent directory
3539 * so that other processes accessing it will recognize that
3540 * the directory has changed.
3541 */
3544 /*
3545 * If this is a synchronous mount, make sure that the
3546 * symlink transaction goes to disk before returning to
3547 * the user.
3548 */
3551 }
3553 /*
3554 * xfs_trans_commit normally decrements the vnode ref count
3555 * when it unlocks the inode. Since we want to return the
3556 * vnode to the caller, we bump the vnode ref count now.
3557 */
3563 }
3568 /* Fall through to std_return with error = 0 or errno from
3569 * xfs_trans_commit */
3570 std_return:
3572 DM_EVENT_POSTSYMLINK)) {
3578 }
3586 }
3589 error2:
3591 error1:
3594 error_return:
3601 }
3604 }
3607 /*
3608 * xfs_fid2
3609 *
3610 * A fid routine that takes a pointer to a previously allocated
3611 * fid structure (like xfs_fast_fid) but uses a 64 bit inode number.
3612 */
3613 STATIC int
3617 {
3629 /*
3630 * use memcpy because the inode is a long long and there's no
3631 * assurance that xfid->fid_ino is properly aligned.
3632 */
3637 }
3640 /*
3641 * xfs_rwlock
3642 */
3643 int
3646 vrwlock_t locktype)
3647 {
3665 }
3668 }
3671 /*
3672 * xfs_rwunlock
3673 */
3674 void
3677 vrwlock_t locktype)
3678 {
3687 /*
3688 * In the write case, we may have added a new entry to
3689 * the reference cache. This might store a pointer to
3690 * an inode to be released in this inode. If it is there,
3691 * clear the pointer and release the inode after unlocking
3692 * this one.
3693 */
3699 }
3701 }
3703 STATIC int
3707 {
3720 /*
3721 * Bypass inodes which have already been cleaned by
3722 * the inode flush clustering code inside xfs_iflush
3723 */
3731 xfs_lsn_t sync_lsn;
3744 }
3745 }
3747 /*
3748 * We make this non-blocking if the inode is contended,
3749 * return EAGAIN to indicate to the caller that they
3750 * did not succeed. This prevents the flush path from
3751 * blocking on inodes inside another operation right
3752 * now, they get caught later by xfs_sync.
3753 */
3767 }
3770 }
3774 else
3779 }
3782 }
3785 int
3788 u_int evmask,
3789 u_int16_t state,
3791 {
3811 }
3823 }
3826 /*
3827 * xfs_reclaim
3828 */
3829 STATIC int
3832 {
3843 /* bad inode, get out here ASAP */
3847 }
3854 /* If we have nothing to flush with this inode then complete the
3855 * teardown now, otherwise break the link between the xfs inode
3856 * and the linux inode and clean up the xfs inode later. This
3857 * avoids flushing the inode to disk during the delete operation
3858 * itself.
3859 */
3867 /* Protect sync from us */
3873 }
3875 }
3877 int
3882 {
3890 /* The hash lock here protects a thread in xfs_iget_core from
3891 * racing with us on linking the inode back with a vnode.
3892 * Once we have the XFS_IRECLAIM flag set it will not touch
3893 * us.
3894 */
3902 }
3904 }
3908 /*
3909 * If the inode is still dirty, then flush it out. If the inode
3910 * is not in the AIL, then it will be OK to flush it delwri as
3911 * long as xfs_iflush() does not keep any references to the inode.
3912 * We leave that decision up to xfs_iflush() since it has the
3913 * knowledge of whether it's OK to simply do a delwri flush of
3914 * the inode or whether we need to wait until the inode is
3915 * pulled from the AIL.
3916 * We get the flush lock regardless, though, just to make sure
3917 * we don't free it while it is being flushed.
3918 */
3923 }
3929 /*
3930 * If we hit an error, typically because of filesystem
3931 * shutdown, we don't need to let vn_reclaim to know
3932 * because we're gonna reclaim the inode anyway.
3933 */
3937 }
3939 }
3946 /*
3947 * We are not interested in doing an iflush if we're
3948 * in the process of shutting down the filesystem forcibly.
3949 * So, just reclaim the inode.
3950 */
3953 }
3955 reclaim:
3958 }
3960 int
3962 {
3978 }
3979 }
3982 XFS_IFLUSH_DELWRI_ELSE_ASYNC);
3985 }
3988 }
3992 }
3994 /*
3995 * xfs_alloc_file_space()
3996 * This routine allocates disk space for the given file.
3997 *
3998 * If alloc_type == 0, this request is for an ALLOCSP type
3999 * request which will change the file size. In this case, no
4000 * DMAPI event will be generated by the call. A TRUNCATE event
4001 * will be generated later by xfs_setattr.
4002 *
4003 * If alloc_type != 0, this request is for a RESVSP type
4004 * request, and a DMAPI DM_EVENT_WRITE will be generated if the
4005 * lower block boundary byte address is less than the file's
4006 * length.
4007 *
4008 * RETURNS:
4009 * 0 on success
4010 * errno on error
4011 *
4012 */
4013 STATIC int
4016 xfs_off_t offset,
4017 xfs_off_t len,
4020 {
4021 xfs_filblks_t allocated_fsb;
4022 xfs_filblks_t allocatesize_fsb;
4024 xfs_off_t count;
4025 xfs_filblks_t datablocks;
4027 xfs_fsblock_t firstfsb;
4028 xfs_bmap_free_t free_list;
4034 uint resblks;
4037 xfs_fileoff_t startoffset_fsb;
4047 /*
4048 * determine if this is a realtime file
4049 */
4053 else
4072 /* Generate a DMAPI event if needed. */
4076 xfs_off_t end_dmi_offset;
4086 }
4088 /*
4089 * allocate file space until done or until there is an error
4090 */
4091 retry:
4093 /*
4094 * determine if reserving space on
4095 * the data or realtime partition.
4096 */
4104 rtextsize);
4110 }
4112 /*
4113 * allocate and setup the transaction
4114 */
4118 resblks,
4120 numrtextents,
4121 XFS_TRANS_PERM_LOG_RES,
4122 XFS_WRITE_LOG_COUNT);
4124 /*
4125 * check for running out of space
4126 */
4128 /*
4129 * Free the transaction structure.
4130 */
4134 }
4144 /*
4145 * issue the bmapi() call to allocate the blocks
4146 */
4154 }
4156 /*
4157 * complete the transaction
4158 */
4162 }
4168 }
4175 }
4179 }
4180 dmapi_enospc_check:
4190 /* else fall through with error from XFS_SEND_DATA */
4191 }
4195 error0:
4197 error1:
4201 }
4203 /*
4204 * Zero file bytes between startoff and endoff inclusive.
4205 * The iolock is held exclusive and no blocks are buffered.
4206 */
4207 STATIC int
4210 xfs_off_t startoff,
4211 xfs_off_t endoff)
4212 {
4213 xfs_bmbt_irec_t imap;
4214 xfs_fileoff_t offset_fsb;
4215 xfs_off_t lastoffset;
4216 xfs_off_t offset;
4252 }
4264 }
4265 }
4268 }
4270 /*
4271 * xfs_free_file_space()
4272 * This routine frees disk space for the given file.
4273 *
4274 * This routine is only called by xfs_change_file_space
4275 * for an UNRESVSP type call.
4276 *
4277 * RETURNS:
4278 * 0 on success
4279 * errno on error
4280 *
4281 */
4282 STATIC int
4285 xfs_off_t offset,
4286 xfs_off_t len,
4288 {
4292 xfs_off_t end_dmi_offset;
4293 xfs_fileoff_t endoffset_fsb;
4295 xfs_fsblock_t firstfsb;
4296 xfs_bmap_free_t free_list;
4297 xfs_off_t ilen;
4298 xfs_bmbt_irec_t imap;
4299 xfs_off_t ioffset;
4303 uint resblks;
4306 xfs_fileoff_t startoffset_fsb;
4336 }
4356 }
4358 /*
4359 * Need to zero the stuff we're not freeing, on disk.
4360 * If its a realtime file & can't use unwritten extents then we
4361 * actually need to zero the extent edges. Otherwise xfs_bunmapi
4362 * will take care of it for us.
4363 */
4372 xfs_daddr_t block;
4379 }
4388 mod++;
4391 }
4392 }
4394 /*
4395 * One contiguous piece to clear
4396 */
4399 /*
4400 * Some full blocks, possibly two pieces to clear
4401 */
4410 }
4412 /*
4413 * free file space until done or until there is an error
4414 */
4418 /*
4419 * allocate and setup the transaction
4420 */
4423 resblks,
4426 XFS_TRANS_PERM_LOG_RES,
4427 XFS_WRITE_LOG_COUNT);
4429 /*
4430 * check for running out of space
4431 */
4433 /*
4434 * Free the transaction structure.
4435 */
4439 }
4450 /*
4451 * issue the bunmapi() call to free the blocks
4452 */
4459 }
4461 /*
4462 * complete the transaction
4463 */
4467 }
4471 }
4473 out_unlock_iolock:
4478 error0:
4480 error1:
4483 XFS_ILOCK_EXCL);
4485 }
4487 /*
4488 * xfs_change_file_space()
4489 * This routine allocates or frees disk space for the given file.
4490 * The user specified parameters are checked for alignment and size
4491 * limitations.
4492 *
4493 * RETURNS:
4494 * 0 on success
4495 * errno on error
4496 *
4497 */
4498 int
4503 xfs_off_t offset,
4506 {
4509 xfs_fsize_t fsize;
4513 xfs_off_t startoffset;
4514 xfs_off_t llen;
4516 vattr_t va;
4525 /*
4526 * must be a regular file and have write permission
4527 */
4536 }
4551 }
4566 /*
4567 * XFS_IOC_RESVSP and XFS_IOC_UNRESVSP will reserve or unreserve
4568 * file space.
4569 * These calls do NOT zero the data space allocated to the file,
4570 * nor do they change the file size.
4571 *
4572 * XFS_IOC_ALLOCSP and XFS_IOC_FREESP will allocate and free file
4573 * space.
4574 * These calls cause the new file data to be zeroed and the file
4575 * size to be changed.
4576 */
4592 attr_flags)))
4605 }
4621 }
4623 /*
4624 * update the inode timestamp, mode, and prealloc flag bits
4625 */
4630 /* ASSERT(0); */
4633 }
4643 /*
4644 * Note that we don't have to worry about mandatory
4645 * file locking being disabled here because we only
4646 * clear the S_ISGID bit if the Group execute bit is
4647 * on, but if it was on then mandatory locking wouldn't
4648 * have been enabled.
4649 */
4654 }
4668 }
4670 vnodeops_t xfs_vnodeops = {
4674 #ifdef HAVE_SENDFILE
4676 #endif
4710 };