| blob | 9b6b6fab8bc3d15a3e61dcc18d93ac48c9e6a27e |
1 /*
2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 */
35 #include <sys/mountctl.h>
36 #include <sys/namecache.h>
37 #include <sys/buf2.h>
38 #include <vfs/fifofs/fifo.h>
42 /*
43 * USERFS VNOPS
44 */
116 };
130 };
145 };
147 static __inline
148 void
150 {
153 }
161 /*
162 * hammer_vop_fsync { vp, waitfor }
163 *
164 * fsync() an inode to disk and wait for it to be completely committed
165 * such that the information would not be undone if a crash occured after
166 * return.
167 *
168 * NOTE: HAMMER's fsync()'s are going to remain expensive until we implement
169 * a REDO log. A sysctl is provided to relax HAMMER's fsync()
170 * operation.
171 *
172 * Ultimately the combination of a REDO log and use of fast storage
173 * to front-end cluster caches will make fsync fast, but it aint
174 * here yet. And, in anycase, we need real transactional
175 * all-or-nothing features which are not restricted to a single file.
176 */
177 static
178 int
180 {
188 /*
189 * Fsync rule relaxation (default is either full synchronous flush
190 * or REDO semantics with synchronous flush).
191 */
195 mode0:
196 /* no REDO, full synchronous flush */
199 mode1:
200 /* no REDO, full asynchronous flush */
205 /* REDO semantics, synchronous flush */
211 /* REDO semantics, relaxed asynchronous flush */
219 /* ignore the fsync() system call */
223 /* we have to do something */
228 }
230 /*
231 * Fast fsync only needs to flush the UNDO/REDO fifo if
232 * HAMMER_INODE_REDO is non-zero and the only modifications
233 * made to the file are write or write-extends.
234 */
244 }
246 /*
247 * REDO is enabled by fsync(), the idea being we really only
248 * want to lay down REDO records when programs are using
249 * fsync() heavily. The first fsync() on the file starts
250 * the gravy train going and later fsync()s keep it hot by
251 * resetting the redo_count.
252 *
253 * We weren't running REDOs before now so we have to fall
254 * through and do a full fsync of what we have.
255 */
260 }
261 }
262 skip:
264 /*
265 * Do a full flush sequence.
266 *
267 * Attempt to release the vnode while waiting for the inode to
268 * finish flushing. This can really mess up inactive->reclaim
269 * sequences so only do it if the vnode is active.
270 *
271 * WARNING! The VX lock functions must be used. vn_lock() will
272 * fail when this is part of a VOP_RECLAIM sequence.
273 */
285 }
289 }
294 }
296 /*
297 * hammer_vop_read { vp, uio, ioflag, cred }
298 *
299 * MPSAFE (for the cache safe does not require fs_token)
300 */
301 static
302 int
304 {
306 hammer_inode_t ip;
307 hammer_mount_t hmp;
308 off_t offset;
328 /*
329 * Attempt to shortcut directly to the VM object using lwbufs.
330 * This is much faster than instantiating buffer cache buffers.
331 */
340 /*
341 * Allow the UIO's size to override the sequential heuristic.
342 */
349 /*
350 * If reading or writing a huge amount of data we have to break
351 * atomicy and allow the operation to be interrupted by a signal
352 * or it can DOS the machine.
353 */
356 /*
357 * Access the data typically in HAMMER_BUFSIZE blocks via the
358 * buffer cache, but HAMMER may use a variable block size based
359 * on the offset.
360 *
361 * XXX Temporary hack, delay the start transaction while we remain
362 * MPSAFE. NOTE: ino_data.size cannot change while vnode is
363 * locked-shared.
364 */
376 /*
377 * MPSAFE
378 */
384 }
388 }
390 /*
391 * MPUNSAFE
392 */
396 }
398 /*
399 * NOTE: A valid bp has already been acquired, but was not
400 * B_CACHE.
401 */
403 /*
404 * Use file_limit to prevent cluster_read() from
405 * creating buffers of the wrong block size past
406 * the demarc.
407 */
412 }
423 }
427 }
428 skip:
434 }
445 /*
446 * Set B_AGE, data has a lower priority than meta-data.
447 *
448 * Use a hold/unlock/drop sequence to run the uiomove
449 * with the buffer unlocked, avoiding deadlocks against
450 * read()s on mmap()'d spaces.
451 */
459 }
461 finished:
463 /*
464 * Try to update the atime with just the inode lock for maximum
465 * concurrency. If we can't shortcut it we have to get the full
466 * blown transaction.
467 */
471 }
483 }
484 }
486 }
488 /*
489 * hammer_vop_write { vp, uio, ioflag, cred }
490 */
491 static
492 int
494 {
496 hammer_inode_t ip;
497 hammer_mount_t hmp;
498 thread_t td;
501 off_t base_offset;
522 /*
523 * Create a transaction to cover the operations we perform.
524 */
528 /*
529 * Check append mode
530 */
534 /*
535 * Check for illegal write offsets. Valid range is 0...2^63-1.
536 *
537 * NOTE: the base_off assignment is required to work around what
538 * I consider to be a GCC-4 optimization bug.
539 */
543 }
548 }
555 }
557 /*
558 * If reading or writing a huge amount of data we have to break
559 * atomicy and allow the operation to be interrupted by a signal
560 * or it can DOS the machine.
561 *
562 * Preset redo_count so we stop generating REDOs earlier if the
563 * limit is exceeded.
564 *
565 * redo_count is heuristical, SMP races are ok
566 */
571 }
573 /*
574 * Access the data typically in HAMMER_BUFSIZE blocks via the
575 * buffer cache, but HAMMER may use a variable block size based
576 * on the offset.
577 */
584 off_t nsize;
593 /*
594 * Control the number of pending records associated with
595 * this inode. If too many have accumulated start a
596 * flush. Try to maintain a pipeline with the flusher.
597 *
598 * NOTE: It is possible for other sources to grow the
599 * records but not necessarily issue another flush,
600 * so use a timeout and ensure that a re-flush occurs.
601 */
609 }
611 }
613 /*
614 * Do not allow HAMMER to blow out the buffer cache. Very
615 * large UIOs can lockout other processes due to bwillwrite()
616 * mechanics.
617 *
618 * The hammer inode is not locked during these operations.
619 * The vnode is locked which can interfere with the pageout
620 * daemon for non-UIO_NOCOPY writes but should not interfere
621 * with the buffer cache. Even so, we cannot afford to
622 * allow the pageout daemon to build up too many dirty buffer
623 * cache buffers.
624 *
625 * Only call this if we aren't being recursively called from
626 * a virtual disk device (vn), else we may deadlock.
627 */
631 /*
632 * Calculate the blocksize at the current offset and figure
633 * out how much we can actually write.
634 */
644 }
649 else
653 nsize,
658 trivial);
661 }
664 /*
665 * Issuing a write with the same data backing the
666 * buffer. Instantiate the buffer to collect the
667 * backing vm pages, then read-in any missing bits.
668 *
669 * This case is used by vop_stdputpages().
670 */
677 }
679 /*
680 * Even though we are entirely overwriting the buffer
681 * we may still have to zero it out to avoid a
682 * mmap/write visibility issue.
683 */
688 /*
689 * If the base offset of the buffer is beyond the
690 * file EOF, we don't have to issue a read.
691 */
696 /*
697 * Partial overwrite, read in any missing bits then
698 * replace the portion being written.
699 */
703 }
709 /*
710 * Generate REDO records if enabled and redo_count will not
711 * exceeded the limit.
712 *
713 * If redo_count exceeds the limit we stop generating records
714 * and clear HAMMER_INODE_REDO. This will cause the next
715 * fsync() to do a full meta-data sync instead of just an
716 * UNDO/REDO fifo update.
717 *
718 * When clearing HAMMER_INODE_REDO any pre-existing REDOs
719 * will still be tracked. The tracks will be terminated
720 * when the related meta-data (including possible data
721 * modifications which are not tracked via REDO) is
722 * flushed.
723 */
729 HAMMER_REDO_WRITE,
734 }
735 }
737 /*
738 * If we screwed up we have to undo any VM size changes we
739 * made.
740 */
748 }
751 }
761 }
766 /*
767 * Once we dirty the buffer any cached zone-X offset
768 * becomes invalid. HAMMER NOTE: no-history mode cannot
769 * allow overwriting over the same data sector unless
770 * we provide UNDOs for the old data, which we don't.
771 */
776 /*
777 * Final buffer disposition.
778 *
779 * Because meta-data updates are deferred, HAMMER is
780 * especially sensitive to excessive bdwrite()s because
781 * the I/O stream is not broken up by disk reads. So the
782 * buffer cache simply cannot keep up.
783 *
784 * WARNING! blksize is variable. cluster_write() is
785 * expected to not blow up if it encounters
786 * buffers that do not match the passed blksize.
787 *
788 * NOTE! Hammer shouldn't need to bawrite()/cluster_write().
789 * The ip->rsv_recs check should burst-flush the data.
790 * If we queue it immediately the buf could be left
791 * locked on the device queue for a very long time.
792 *
793 * However, failing to flush a dirty buffer out when
794 * issued from the pageout daemon can result in a low
795 * memory deadlock against bio_page_alloc(), so we
796 * have to bawrite() on IO_ASYNC as well.
797 *
798 * NOTE! To avoid degenerate stalls due to mismatched block
799 * sizes we only honor IO_DIRECT on the write which
800 * abuts the end of the buffer. However, we must
801 * honor IO_SYNC in case someone is silly enough to
802 * configure a HAMMER file as swap, or when HAMMER
803 * is serving NFS (for commits). Ick ick.
804 */
820 else
825 }
826 }
831 }
833 /*
834 * hammer_vop_access { vp, mode, cred }
835 *
836 * MPSAFE - does not require fs_token
837 */
838 static
839 int
841 {
843 uid_t uid;
844 gid_t gid;
853 }
855 /*
856 * hammer_vop_advlock { vp, id, op, fl, flags }
857 *
858 * MPSAFE - does not require fs_token
859 */
860 static
861 int
863 {
867 }
869 /*
870 * hammer_vop_close { vp, fflag }
871 *
872 * We can only sync-on-close for normal closes. XXX disabled for now.
873 */
874 static
875 int
877 {
878 #if 0
887 else
890 HAMMER_INODE_CLOSEASYNC);
892 }
893 }
894 #endif
896 }
898 /*
899 * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
900 *
901 * The operating system has already ensured that the directory entry
902 * does not exist and done all appropriate namespace locking.
903 */
904 static
905 int
907 {
909 hammer_inode_t dip;
910 hammer_inode_t nip;
912 hammer_mount_t hmp;
924 /*
925 * Create a transaction to cover the operations we perform.
926 */
930 /*
931 * Create a new filesystem object of the requested type. The
932 * returned inode will be referenced and shared-locked to prevent
933 * it from being moved to the flusher.
934 */
944 }
946 /*
947 * Add the new filesystem object to the directory. This will also
948 * bump the inode's link count.
949 */
952 nip);
956 /*
957 * Finish up.
958 */
970 }
972 }
975 }
977 /*
978 * hammer_vop_getattr { vp, vap }
979 *
980 * Retrieve an inode's attribute information. When accessing inodes
981 * historically we fake the atime field to ensure consistent results.
982 * The atime field is stored in the B-Tree element and allowed to be
983 * updated without cycling the element.
984 *
985 * MPSAFE - does not require fs_token
986 */
987 static
988 int
990 {
994 /*
995 * We want the fsid to be different when accessing a filesystem
996 * with different as-of's so programs like diff don't think
997 * the files are the same.
998 *
999 * We also want the fsid to be the same when comparing snapshots,
1000 * or when comparing mirrors (which might be backed by different
1001 * physical devices). HAMMER fsids are based on the PFS's
1002 * shared_uuid field.
1003 *
1004 * XXX there is a chance of collision here. The va_fsid reported
1005 * by stat is different from the more involved fsid used in the
1006 * mount structure.
1007 */
1021 /*
1022 * Special case for @@PFS softlinks. The actual size of the
1023 * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1024 * or for MAX_TID is "@@-1:%05d" == 10 bytes.
1025 *
1026 * Note that userspace hammer command does not allow users to
1027 * create a @@PFS softlink under an existing other PFS (id!=0)
1028 * so the ip localization here for @@PFS softlink is always 0.
1029 */
1037 else
1039 }
1041 /*
1042 * We must provide a consistent atime and mtime for snapshots
1043 * so people can do a 'tar cf - ... | md5' on them and get
1044 * consistent results.
1045 */
1052 }
1063 }
1071 VA_FSID_UUID_VALID;
1081 }
1084 }
1086 /*
1087 * hammer_vop_nresolve { nch, dvp, cred }
1088 *
1089 * Locate the requested directory entry.
1090 */
1091 static
1092 int
1094 {
1097 hammer_mount_t hmp;
1098 hammer_inode_t dip;
1099 hammer_inode_t ip;
1100 hammer_tid_t asof;
1113 /*
1114 * Misc initialization, plus handle as-of name extensions. Look for
1115 * the '@@' extension. Note that as-of files and directories cannot
1116 * be modified.
1117 */
1137 }
1141 }
1142 }
1145 /*
1146 * If this is a PFS we dive into the PFS root inode
1147 */
1157 }
1162 }
1164 }
1166 /*
1167 * If there is no path component the time extension is relative to dip.
1168 * e.g. "fubar/@@<snapshot>"
1169 *
1170 * "." is handled by the kernel, but ".@@<snapshot>" is not.
1171 * e.g. "fubar/.@@<snapshot>"
1172 *
1173 * ".." is handled by the kernel. We do not currently handle
1174 * "..@<snapshot>".
1175 */
1185 }
1190 }
1192 }
1194 /*
1195 * Calculate the namekey and setup the key range for the scan. This
1196 * works kinda like a chained hash table where the lower 32 bits
1197 * of the namekey synthesize the chain.
1198 *
1199 * The key range is inclusive of both key_beg and key_end.
1200 */
1219 /*
1220 * Scan all matching records (the chain), locate the one matching
1221 * the requested path component.
1222 *
1223 * The hammer_ip_*() functions merge in-memory records with on-disk
1224 * records for the purposes of the search.
1225 */
1240 }
1242 }
1243 }
1246 /*
1247 * Lookup the obj_id. This should always succeed. If it does not
1248 * the filesystem may be damaged and we return a dummy inode.
1249 */
1259 localization);
1264 }
1270 }
1275 }
1278 }
1279 done:
1283 }
1285 /*
1286 * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1287 *
1288 * Locate the parent directory of a directory vnode.
1289 *
1290 * dvp is referenced but not locked. *vpp must be returned referenced and
1291 * locked. A parent_obj_id of 0 indicates that we are at the root.
1292 *
1293 * NOTE: as-of sequences are not linked into the directory structure. If
1294 * we are at the root with a different asof then the mount point, reload
1295 * the same directory with the mount point's asof. I'm not sure what this
1296 * will do to NFS. We encode ASOF stamps in NFS file handles so it might not
1297 * get confused, but it hasn't been tested.
1298 */
1299 static
1300 int
1302 {
1304 hammer_inode_t dip;
1305 hammer_inode_t ip;
1306 hammer_mount_t hmp;
1309 hammer_tid_t asof;
1316 /*
1317 * Whos are parent? This could be the root of a pseudo-filesystem
1318 * whos parent is in another localization domain.
1319 */
1324 else
1327 /*
1328 * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1329 */
1342 }
1343 }
1355 }
1359 }
1361 /*
1362 * hammer_vop_nlink { nch, dvp, vp, cred }
1363 */
1364 static
1365 int
1367 {
1369 hammer_inode_t dip;
1370 hammer_inode_t ip;
1372 hammer_mount_t hmp;
1393 /*
1394 * Create a transaction to cover the operations we perform.
1395 */
1399 /*
1400 * Add the filesystem object to the directory. Note that neither
1401 * dip nor ip are referenced or locked, but their vnodes are
1402 * referenced. This function will bump the inode's link count.
1403 */
1406 ip);
1408 /*
1409 * Finish up.
1410 */
1414 }
1420 }
1422 /*
1423 * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1424 *
1425 * The operating system has already ensured that the directory entry
1426 * does not exist and done all appropriate namespace locking.
1427 */
1428 static
1429 int
1431 {
1433 hammer_inode_t dip;
1434 hammer_inode_t nip;
1436 hammer_mount_t hmp;
1448 /*
1449 * Create a transaction to cover the operations we perform.
1450 */
1454 /*
1455 * Create a new filesystem object of the requested type. The
1456 * returned inode will be referenced but not locked.
1457 */
1466 }
1467 /*
1468 * Add the new filesystem object to the directory. This will also
1469 * bump the inode's link count.
1470 */
1473 nip);
1477 /*
1478 * Finish up.
1479 */
1489 }
1490 }
1496 }
1498 /*
1499 * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1500 *
1501 * The operating system has already ensured that the directory entry
1502 * does not exist and done all appropriate namespace locking.
1503 */
1504 static
1505 int
1507 {
1509 hammer_inode_t dip;
1510 hammer_inode_t nip;
1512 hammer_mount_t hmp;
1524 /*
1525 * Create a transaction to cover the operations we perform.
1526 */
1530 /*
1531 * Create a new filesystem object of the requested type. The
1532 * returned inode will be referenced but not locked.
1533 *
1534 * If mknod specifies a directory a pseudo-fs is created.
1535 */
1544 }
1546 /*
1547 * Add the new filesystem object to the directory. This will also
1548 * bump the inode's link count.
1549 */
1552 nip);
1554 /*
1555 * Finish up.
1556 */
1566 }
1567 }
1573 }
1575 /*
1576 * hammer_vop_open { vp, mode, cred, fp }
1577 *
1578 * MPSAFE (does not require fs_token)
1579 */
1580 static
1581 int
1583 {
1584 hammer_inode_t ip;
1591 }
1593 /*
1594 * hammer_vop_print { vp }
1595 */
1596 static
1597 int
1599 {
1601 }
1603 /*
1604 * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1605 */
1606 static
1607 int
1609 {
1612 hammer_inode_t ip;
1613 hammer_mount_t hmp;
1615 hammer_base_elm_t base;
1620 off_t saveoff;
1639 }
1644 /*
1645 * Handle artificial entries
1646 *
1647 * It should be noted that the minimum value for a directory
1648 * hash key on-media is 0x0000000100000000, so we can use anything
1649 * less then that to represent our 'special' key space.
1650 */
1662 }
1671 }
1680 }
1682 /*
1683 * Key range (begin and end inclusive) to scan. Directory keys
1684 * directly translate to a 64 bit 'seek' position.
1685 */
1717 dtype,
1729 }
1732 done:
1745 }
1752 }
1753 }
1756 }
1758 /*
1759 * hammer_vop_readlink { vp, uio, cred }
1760 */
1761 static
1762 int
1764 {
1767 hammer_inode_t ip;
1768 hammer_mount_t hmp;
1771 hammer_pseudofs_inmem_t pfsm;
1779 /*
1780 * Shortcut if the symlink data was stuffed into ino_data.
1781 *
1782 * Also expand special "@@PFS%05d" softlinks (expansion only
1783 * occurs for non-historical (current) accesses made from the
1784 * primary filesystem).
1785 *
1786 * Note that userspace hammer command does not allow users to
1787 * create a @@PFS softlink under an existing other PFS (id!=0)
1788 * so the ip localization here for @@PFS softlink is always 0.
1789 */
1808 /* vap->va_size == 26 */
1814 /* vap->va_size == 10 */
1818 }
1821 }
1825 }
1829 }
1831 /*
1832 * Long version
1833 */
1837 /*
1838 * Key range (begin and end inclusive) to scan. Directory keys
1839 * directly translate to a 64 bit 'seek' position.
1840 */
1842 HAMMER_LOCALIZE_MISC;
1857 HAMMER_SYMLINK_NAME_OFF);
1860 HAMMER_SYMLINK_NAME_OFF,
1862 }
1863 }
1868 }
1870 /*
1871 * hammer_vop_nremove { nch, dvp, cred }
1872 */
1873 static
1874 int
1876 {
1878 hammer_inode_t dip;
1879 hammer_mount_t hmp;
1888 }
1898 }
1900 /*
1901 * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1902 */
1903 static
1904 int
1906 {
1910 hammer_inode_t fdip;
1911 hammer_inode_t tdip;
1912 hammer_inode_t ip;
1913 hammer_mount_t hmp;
1950 /*
1951 * Remove tncp from the target directory and then link ip as
1952 * tncp. XXX pass trans to dounlink
1953 *
1954 * Force the inode sync-time to match the transaction so it is
1955 * in-sync with the creation of the target directory entry.
1956 */
1962 ip);
1967 }
1968 }
1972 /*
1973 * Locate the record in the originating directory and remove it.
1974 *
1975 * Calculate the namekey and setup the key range for the scan. This
1976 * works kinda like a chained hash table where the lower 32 bits
1977 * of the namekey synthesize the chain.
1978 *
1979 * The key range is inclusive of both key_beg and key_end.
1980 */
1983 retry:
1999 /*
2000 * Scan all matching records (the chain), locate the one matching
2001 * the requested path component.
2002 *
2003 * The hammer_ip_*() functions merge in-memory records with on-disk
2004 * records for the purposes of the search.
2005 */
2015 }
2017 }
2019 /*
2020 * If all is ok we have to get the inode so we can adjust nlinks.
2021 *
2022 * WARNING: hammer_ip_del_direntry() may have to terminate the
2023 * cursor to avoid a recursion. It's ok to call hammer_done_cursor()
2024 * twice.
2025 */
2029 /*
2030 * XXX A deadlock here will break rename's atomicy for the purposes
2031 * of crash recovery.
2032 */
2036 }
2038 /*
2039 * Cleanup and tell the kernel that the rename succeeded.
2040 *
2041 * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2042 * without formally acquiring the vp since the vp might
2043 * have zero refs on it, or in the middle of a reclaim,
2044 * etc.
2045 */
2060 }
2062 }
2063 }
2065 failed:
2069 }
2071 /*
2072 * hammer_vop_nrmdir { nch, dvp, cred }
2073 */
2074 static
2075 int
2077 {
2079 hammer_inode_t dip;
2080 hammer_mount_t hmp;
2089 }
2099 }
2101 /*
2102 * hammer_vop_markatime { vp, cred }
2103 */
2104 static
2105 int
2107 {
2109 hammer_inode_t ip;
2110 hammer_mount_t hmp;
2129 }
2131 /*
2132 * hammer_vop_setattr { vp, vap, cred }
2133 */
2134 static
2135 int
2137 {
2139 hammer_inode_t ip;
2141 hammer_mount_t hmp;
2147 #if 0
2149 #endif
2165 }
2182 }
2186 }
2187 }
2189 }
2193 }
2198 uuid_t uuid_uid;
2199 uuid_t uuid_gid;
2217 }
2219 }
2220 }
2227 /*
2228 * Log the operation if in fast-fsync mode or if
2229 * there are unterminated redo write records present.
2230 *
2231 * The second check is needed so the recovery code
2232 * properly truncates write redos even if nominal
2233 * REDO operations is turned off due to excessive
2234 * writes, because the related records might be
2235 * destroyed and never lay down a TERM_WRITE.
2236 */
2241 HAMMER_REDO_TRUNC,
2243 }
2246 /*
2247 * XXX break atomicy, we can deadlock the backend
2248 * if we do not release the lock. Probably not a
2249 * big deal here.
2250 */
2253 blksize,
2269 }
2272 /* XXX safe to use SDIRTY instead of DDIRTY here? */
2275 /*
2276 * On-media truncation is cached in the inode until
2277 * the inode is synchronized. We must immediately
2278 * handle any frontend records.
2279 */
2288 }
2289 }
2291 #if 0
2292 /*
2293 * When truncating, nvtruncbuf() may have cleaned out
2294 * a portion of the last block on-disk in the buffer
2295 * cache. We must clean out any frontend records
2296 * for blocks beyond the new last block.
2297 */
2303 }
2304 #endif
2313 }
2323 }
2325 }
2330 }
2335 }
2348 }
2349 }
2350 done:
2357 }
2359 /*
2360 * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2361 */
2362 static
2363 int
2365 {
2367 hammer_inode_t dip;
2368 hammer_inode_t nip;
2369 hammer_record_t record;
2371 hammer_mount_t hmp;
2386 /*
2387 * Create a transaction to cover the operations we perform.
2388 */
2392 /*
2393 * Create a new filesystem object of the requested type. The
2394 * returned inode will be referenced but not locked.
2395 */
2405 }
2407 /*
2408 * Add a record representing the symlink. symlink stores the link
2409 * as pure data, not a string, and is no \0 terminated.
2410 */
2421 HAMMER_LOCALIZE_MISC;
2428 }
2430 /*
2431 * Set the file size to the length of the link.
2432 */
2436 }
2437 }
2442 /*
2443 * Finish up.
2444 */
2455 }
2456 }
2460 }
2462 /*
2463 * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2464 */
2465 static
2466 int
2468 {
2470 hammer_inode_t dip;
2471 hammer_mount_t hmp;
2480 }
2490 }
2492 /*
2493 * hammer_vop_ioctl { vp, command, data, fflag, cred }
2494 */
2495 static
2496 int
2498 {
2508 }
2510 static
2511 int
2513 {
2520 };
2521 hammer_mount_t hmp;
2538 else
2543 /*
2544 * Call standard mountctl VOP function
2545 * so we get user mount flags.
2546 */
2558 }
2565 }
2568 }
2570 /*
2571 * hammer_vop_strategy { vp, bio }
2572 *
2573 * Strategy call, used for regular file read & write only. Note that the
2574 * bp may represent a cluster.
2575 *
2576 * To simplify operation and allow better optimizations in the future,
2577 * this code does not make any assumptions with regards to buffer alignment
2578 * or size.
2579 */
2580 static
2581 int
2583 {
2601 }
2603 /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2606 }
2608 /*
2609 * Read from a regular file. Iterate the related records and fill in the
2610 * BIO/BUF. Gaps are zero-filled.
2611 *
2612 * The support code in hammer_object.c should be used to deal with mixed
2613 * in-memory and on-disk records.
2614 *
2615 * NOTE: Can be called from the cluster code with an oversized buf.
2616 *
2617 * XXX atime update
2618 */
2619 static
2620 int
2622 {
2624 hammer_inode_t ip;
2625 hammer_inode_t dip;
2626 hammer_mount_t hmp;
2628 hammer_base_elm_t base;
2629 hammer_off_t disk_offset;
2647 /*
2648 * The zone-2 disk offset may have been set by the cluster code via
2649 * a BMAP operation, or else should be NOOFFSET.
2650 *
2651 * Checking the high bits for a match against zone-2 should suffice.
2652 *
2653 * In cases where a lot of data duplication is present it may be
2654 * more beneficial to drop through and doubule-buffer through the
2655 * device.
2656 */
2664 }
2666 /*
2667 * Try to shortcut requests for double_buffer mode too.
2668 * Since this mode runs through the device buffer cache
2669 * only compatible buffer sizes (meaning those generated
2670 * by normal filesystem buffers) are legal.
2671 */
2677 }
2678 }
2680 /*
2681 * Well, that sucked. Do it the hard way. If all the stars are
2682 * aligned we may still be able to issue a direct-read.
2683 */
2688 /*
2689 * Key range (begin and end inclusive) to scan. Note that the key's
2690 * stored in the actual records represent BASE+LEN, not BASE. The
2691 * first record containing bio_offset will have a key > bio_offset.
2692 */
2694 HAMMER_LOCALIZE_MISC;
2705 #if 0
2711 #endif
2712 {
2719 else
2721 }
2724 /*
2725 * Set NOSWAPCACHE for cursor data extraction if double buffering
2726 * is disabled or (if the file is not marked cacheable via chflags
2727 * and vm.swapcache_use_chflags is enabled).
2728 */
2731 vm_swapcache_use_chflags)) {
2733 }
2739 /*
2740 * Get the base file offset of the record. The key for
2741 * data records is (base + bytes) rather then (base).
2742 */
2746 /*
2747 * Calculate the gap, if any, and zero-fill it.
2748 *
2749 * n is the offset of the start of the record verses our
2750 * current seek offset in the bio.
2751 */
2759 }
2761 /*
2762 * Calculate the data offset in the record and the number
2763 * of bytes we can copy.
2764 *
2765 * There are two degenerate cases. First, boff may already
2766 * be at bp->b_bufsize. Secondly, the data offset within
2767 * the record may exceed the record's size.
2768 */
2777 }
2779 /*
2780 * Deal with cached truncations. This cool bit of code
2781 * allows truncate()/ftruncate() to avoid having to sync
2782 * the file.
2783 *
2784 * If the frontend is truncated then all backend records are
2785 * subject to the frontend's truncation.
2786 *
2787 * If the backend is truncated then backend records on-disk
2788 * (but not in-memory) are subject to the backend's
2789 * truncation. In-memory records owned by the backend
2790 * represent data written after the truncation point on the
2791 * backend and must not be truncated.
2792 *
2793 * Truncate operations deal with frontend buffer cache
2794 * buffers and frontend-owned in-memory records synchronously.
2795 */
2803 }
2804 }
2811 }
2812 }
2814 /*
2815 * Try to issue a direct read into our bio if possible,
2816 * otherwise resolve the element data into a hammer_buffer
2817 * and copy.
2818 *
2819 * The buffer on-disk should be zerod past any real
2820 * truncation point, but may not be for any synthesized
2821 * truncation point from above.
2822 *
2823 * NOTE: disk_offset is only valid if the cursor data is
2824 * on-disk.
2825 */
2832 /*
2833 * Direct read case
2834 */
2842 /*
2843 * Async I/O case for reading from backing store
2844 * and copying the data to the filesystem buffer.
2845 * live-dedup has to verify the data anyway if it
2846 * gets a hit later so we can just add the entry
2847 * now.
2848 */
2862 }
2863 }
2867 /*
2868 * We have to be sure that the only elements added to the
2869 * dedup cache are those which are already on-media.
2870 */
2874 /*
2875 * Iterate until we have filled the request.
2876 */
2881 }
2883 /*
2884 * There may have been a gap after the last record
2885 */
2891 /* boff = bp->b_bufsize; */
2892 }
2894 /*
2895 * Disallow swapcache operation on the vnode buffer if double
2896 * buffering is enabled, the swapcache will get the data via
2897 * the block device buffer.
2898 */
2902 /*
2903 * Cleanup
2904 */
2911 done:
2912 /*
2913 * Cache the b-tree node for the last data read in cache[1].
2914 *
2915 * If we hit the file EOF then also cache the node in the
2916 * governing directory's cache[3], it will be used to initialize
2917 * the new inode's cache[1] for any inodes looked up via the directory.
2918 *
2919 * This doesn't reduce disk accesses since the B-Tree chain is
2920 * likely cached, but it does reduce cpu overhead when looking
2921 * up file offsets for cpdup/tar/cpio style iterations.
2922 */
2931 }
2932 }
2937 }
2939 /*
2940 * BMAP operation - used to support cluster_read() only.
2941 *
2942 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2943 *
2944 * This routine may return EOPNOTSUPP if the opration is not supported for
2945 * the specified offset. The contents of the pointer arguments do not
2946 * need to be initialized in that case.
2947 *
2948 * If a disk address is available and properly aligned return 0 with
2949 * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2950 * to the run-length relative to that offset. Callers may assume that
2951 * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2952 * large, so return EOPNOTSUPP if it is not sufficiently large.
2953 */
2954 static
2955 int
2957 {
2959 hammer_inode_t ip;
2960 hammer_mount_t hmp;
2962 hammer_base_elm_t base;
2969 hammer_off_t last_disk_offset;
2970 hammer_off_t disk_offset;
2978 /*
2979 * We can only BMAP regular files. We can't BMAP database files,
2980 * directories, etc.
2981 */
2985 /*
2986 * bmap is typically called with runp/runb both NULL when used
2987 * for writing. We do not support BMAP for writing atm.
2988 */
2992 /*
2993 * Scan the B-Tree to acquire blockmap addresses, then translate
2994 * to raw addresses.
2995 */
3001 /*
3002 * Key range (begin and end inclusive) to scan. Note that the key's
3003 * stored in the actual records represent BASE+LEN, not BASE. The
3004 * first record containing bio_offset will have a key > bio_offset.
3005 */
3007 HAMMER_LOCALIZE_MISC;
3014 else
3030 else
3040 /*
3041 * Get the base file offset of the record. The key for
3042 * data records is (base + bytes) rather then (base).
3043 *
3044 * NOTE: rec_offset + rec_len may exceed the end-of-file.
3045 * The extra bytes should be zero on-disk and the BMAP op
3046 * should still be ok.
3047 */
3052 /*
3053 * Incorporate any cached truncation.
3054 *
3055 * NOTE: Modifications to rec_len based on synthesized
3056 * truncation points remove the guarantee that any extended
3057 * data on disk is zero (since the truncations may not have
3058 * taken place on-media yet).
3059 */
3067 }
3068 }
3075 }
3076 }
3078 /*
3079 * Accumulate information. If we have hit a discontiguous
3080 * block reset base_offset unless we are already beyond the
3081 * requested offset. If we are, that's it, we stop.
3082 */
3093 }
3099 }
3102 }
3111 /*
3112 * If we couldn't find any records or the records we did find were
3113 * all behind the requested offset, return failure. A forward
3114 * truncation can leave a hole w/ no on-disk records.
3115 */
3119 /*
3120 * Figure out the block size at the requested offset and adjust
3121 * our limits so the cluster_read() does not create inappropriately
3122 * sized buffer cache buffers.
3123 */
3127 }
3132 }
3134 /*
3135 * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3136 * from occuring.
3137 */
3141 /*
3142 * Only large-data zones can be direct-IOd
3143 */
3147 /*
3148 * doffsetp is not aligned or the forward run size does
3149 * not cover a whole buffer, disallow the direct I/O.
3150 */
3153 /*
3154 * We're good.
3155 */
3160 }
3164 }
3166 }
3168 }
3170 /*
3171 * Write to a regular file. Because this is a strategy call the OS is
3172 * trying to actually get data onto the media.
3173 */
3174 static
3175 int
3177 {
3178 hammer_record_t record;
3179 hammer_mount_t hmp;
3180 hammer_inode_t ip;
3200 }
3204 /*
3205 * Disallow swapcache operation on the vnode buffer if double
3206 * buffering is enabled, the swapcache will get the data via
3207 * the block device buffer.
3208 */
3212 /*
3213 * Interlock with inode destruction (no in-kernel or directory
3214 * topology visibility). If we queue new IO while trying to
3215 * destroy the inode we can deadlock the vtrunc call in
3216 * hammer_inode_unloadable_check().
3217 *
3218 * Besides, there's no point flushing a bp associated with an
3219 * inode that is being destroyed on-media and has no kernel
3220 * references.
3221 */
3228 }
3230 /*
3231 * Reserve space and issue a direct-write from the front-end.
3232 * NOTE: The direct_io code will hammer_bread/bcopy smaller
3233 * allocations.
3234 *
3235 * An in-memory record will be installed to reference the storage
3236 * until the flusher can get to it.
3237 *
3238 * Since we own the high level bio the front-end will not try to
3239 * do a direct-read until the write completes.
3240 *
3241 * NOTE: The only time we do not reserve a full-sized buffers
3242 * worth of data is if the file is small. We do not try to
3243 * allocate a fragment (from the small-data zone) at the end of
3244 * an otherwise large file as this can lead to wildly separated
3245 * data.
3246 */
3251 else
3257 /*
3258 * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3259 * in hammer_vop_write(). We must flag the record so the proper
3260 * REDO_TERM_WRITE entry is generated during the flush.
3261 */
3266 }
3273 }
3281 }
3284 }
3286 /*
3287 * dounlink - disconnect a directory entry
3288 *
3289 * XXX whiteout support not really in yet
3290 */
3291 static int
3295 {
3297 hammer_inode_t dip;
3298 hammer_inode_t ip;
3299 hammer_mount_t hmp;
3305 /*
3306 * Calculate the namekey and setup the key range for the scan. This
3307 * works kinda like a chained hash table where the lower 32 bits
3308 * of the namekey synthesize the chain.
3309 *
3310 * The key range is inclusive of both key_beg and key_end.
3311 */
3321 retry:
3337 /*
3338 * Scan all matching records (the chain), locate the one matching
3339 * the requested path component. info->last_error contains the
3340 * error code on search termination and could be 0, ENOENT, or
3341 * something else.
3342 *
3343 * The hammer_ip_*() functions merge in-memory records with on-disk
3344 * records for the purposes of the search.
3345 */
3357 }
3359 }
3361 /*
3362 * If all is ok we have to get the inode so we can adjust nlinks.
3363 * To avoid a deadlock with the flusher we must release the inode
3364 * lock on the directory when acquiring the inode for the entry.
3365 *
3366 * If the target is a directory, it must be empty.
3367 */
3382 }
3384 /*
3385 * If isdir >= 0 we validate that the entry is or is not a
3386 * directory. If isdir < 0 we don't care.
3387 */
3395 }
3396 }
3398 /*
3399 * If we are trying to remove a directory the directory must
3400 * be empty.
3401 *
3402 * The check directory code can loop and deadlock/retry. Our
3403 * own cursor's node locks must be released to avoid a 3-way
3404 * deadlock with the flusher if the check directory code
3405 * blocks.
3406 *
3407 * If any changes whatsoever have been made to the cursor
3408 * set EDEADLK and retry.
3409 *
3410 * WARNING: See warnings in hammer_unlock_cursor()
3411 * function.
3412 */
3414 HAMMER_OBJTYPE_DIRECTORY) {
3423 }
3424 }
3426 /*
3427 * Delete the directory entry.
3428 *
3429 * WARNING: hammer_ip_del_direntry() may have to terminate
3430 * the cursor to avoid a deadlock. It is ok to call
3431 * hammer_done_cursor() twice.
3432 */
3436 }
3439 /*
3440 * Tell the namecache that we are now unlinked.
3441 */
3444 /*
3445 * NOTE: ip->vp, if non-NULL, cannot be directly
3446 * referenced without formally acquiring the
3447 * vp since the vp might have zero refs on it,
3448 * or in the middle of a reclaim, etc.
3449 *
3450 * NOTE: The cache_setunresolved() can rip the vp
3451 * out from under us since the vp may not have
3452 * any refs, in which case ip->vp will be NULL
3453 * from the outset.
3454 */
3462 #if 0
3463 /*
3464 * Don't do this, it can deadlock
3465 * on concurrent rm's of hardlinks.
3466 * Shouldn't be needed any more.
3467 */
3469 #endif
3472 }
3474 }
3475 }
3480 }
3485 }
3487 /************************************************************************
3488 * FIFO AND SPECFS OPS *
3489 ************************************************************************
3490 *
3491 */
3492 static int
3494 {
3495 /* XXX update itimes */
3497 }
3499 static int
3501 {
3505 /* XXX update access time */
3507 }
3509 static int
3511 {
3515 /* XXX update access time */
3517 }
3519 static
3520 int
3522 {
3529 }
3531 /************************************************************************
3532 * KQFILTER OPS *
3533 ************************************************************************
3534 *
3535 */
3551 static
3552 int
3554 {
3570 }
3577 }
3579 static void
3581 {
3585 }
3587 static int
3589 {
3593 off_t off;
3598 }
3606 }
3608 static int
3610 {
3615 }
3617 static int
3619 {
3625 }
3627 }