| blob | c8150ed998975f8b0d3455b293b247b3a07227c3 |
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;
330 /*
331 * Attempt to shortcut directly to the VM object using lwbufs.
332 * This is much faster than instantiating buffer cache buffers.
333 */
342 /*
343 * Allow the UIO's size to override the sequential heuristic.
344 */
351 /*
352 * If reading or writing a huge amount of data we have to break
353 * atomicy and allow the operation to be interrupted by a signal
354 * or it can DOS the machine.
355 */
358 /*
359 * Access the data typically in HAMMER_BUFSIZE blocks via the
360 * buffer cache, but HAMMER may use a variable block size based
361 * on the offset.
362 *
363 * XXX Temporary hack, delay the start transaction while we remain
364 * MPSAFE. NOTE: ino_data.size cannot change while vnode is
365 * locked-shared.
366 */
378 /*
379 * MPSAFE
380 */
386 }
390 }
392 /*
393 * MPUNSAFE
394 */
398 }
400 /*
401 * NOTE: A valid bp has already been acquired, but was not
402 * B_CACHE.
403 */
405 /*
406 * Use file_limit to prevent cluster_read() from
407 * creating buffers of the wrong block size past
408 * the demarc.
409 */
414 }
425 }
429 }
430 skip:
436 }
447 /*
448 * Set B_AGE, data has a lower priority than meta-data.
449 *
450 * Use a hold/unlock/drop sequence to run the uiomove
451 * with the buffer unlocked, avoiding deadlocks against
452 * read()s on mmap()'d spaces.
453 */
461 }
463 finished:
465 /*
466 * Try to update the atime with just the inode lock for maximum
467 * concurrency. If we can't shortcut it we have to get the full
468 * blown transaction.
469 */
473 }
485 }
486 }
488 }
490 /*
491 * hammer_vop_write { vp, uio, ioflag, cred }
492 */
493 static
494 int
496 {
498 hammer_inode_t ip;
499 hammer_mount_t hmp;
500 thread_t td;
504 off_t base_offset;
526 /*
527 * Create a transaction to cover the operations we perform.
528 */
532 /*
533 * Use v_lastwrite_ts if file not open for writing
534 * (i.e. a late msync)
535 */
542 }
545 }
547 /*
548 * Check append mode
549 */
553 /*
554 * Check for illegal write offsets. Valid range is 0...2^63-1.
555 *
556 * NOTE: the base_off assignment is required to work around what
557 * I consider to be a GCC-4 optimization bug.
558 */
562 }
567 }
574 }
576 /*
577 * If reading or writing a huge amount of data we have to break
578 * atomicy and allow the operation to be interrupted by a signal
579 * or it can DOS the machine.
580 *
581 * Preset redo_count so we stop generating REDOs earlier if the
582 * limit is exceeded.
583 *
584 * redo_count is heuristical, SMP races are ok
585 */
590 }
592 /*
593 * Access the data typically in HAMMER_BUFSIZE blocks via the
594 * buffer cache, but HAMMER may use a variable block size based
595 * on the offset.
596 */
603 off_t nsize;
612 /*
613 * Control the number of pending records associated with
614 * this inode. If too many have accumulated start a
615 * flush. Try to maintain a pipeline with the flusher.
616 *
617 * NOTE: It is possible for other sources to grow the
618 * records but not necessarily issue another flush,
619 * so use a timeout and ensure that a re-flush occurs.
620 */
628 }
630 }
632 /*
633 * Do not allow HAMMER to blow out the buffer cache. Very
634 * large UIOs can lockout other processes due to bwillwrite()
635 * mechanics.
636 *
637 * The hammer inode is not locked during these operations.
638 * The vnode is locked which can interfere with the pageout
639 * daemon for non-UIO_NOCOPY writes but should not interfere
640 * with the buffer cache. Even so, we cannot afford to
641 * allow the pageout daemon to build up too many dirty buffer
642 * cache buffers.
643 *
644 * Only call this if we aren't being recursively called from
645 * a virtual disk device (vn), else we may deadlock.
646 */
650 /*
651 * Calculate the blocksize at the current offset and figure
652 * out how much we can actually write.
653 */
663 }
668 else
672 nsize,
677 trivial);
680 }
683 /*
684 * Issuing a write with the same data backing the
685 * buffer. Instantiate the buffer to collect the
686 * backing vm pages, then read-in any missing bits.
687 *
688 * This case is used by vop_stdputpages().
689 */
696 }
698 /*
699 * Even though we are entirely overwriting the buffer
700 * we may still have to zero it out to avoid a
701 * mmap/write visibility issue.
702 */
707 /*
708 * If the base offset of the buffer is beyond the
709 * file EOF, we don't have to issue a read.
710 */
715 /*
716 * Partial overwrite, read in any missing bits then
717 * replace the portion being written.
718 */
722 }
728 /*
729 * Generate REDO records if enabled and redo_count will not
730 * exceeded the limit.
731 *
732 * If redo_count exceeds the limit we stop generating records
733 * and clear HAMMER_INODE_REDO. This will cause the next
734 * fsync() to do a full meta-data sync instead of just an
735 * UNDO/REDO fifo update.
736 *
737 * When clearing HAMMER_INODE_REDO any pre-existing REDOs
738 * will still be tracked. The tracks will be terminated
739 * when the related meta-data (including possible data
740 * modifications which are not tracked via REDO) is
741 * flushed.
742 */
748 HAMMER_REDO_WRITE,
753 }
754 }
756 /*
757 * If we screwed up we have to undo any VM size changes we
758 * made.
759 */
767 }
770 }
780 }
785 /*
786 * Once we dirty the buffer any cached zone-X offset
787 * becomes invalid. HAMMER NOTE: no-history mode cannot
788 * allow overwriting over the same data sector unless
789 * we provide UNDOs for the old data, which we don't.
790 */
795 /*
796 * Final buffer disposition.
797 *
798 * Because meta-data updates are deferred, HAMMER is
799 * especially sensitive to excessive bdwrite()s because
800 * the I/O stream is not broken up by disk reads. So the
801 * buffer cache simply cannot keep up.
802 *
803 * WARNING! blksize is variable. cluster_write() is
804 * expected to not blow up if it encounters
805 * buffers that do not match the passed blksize.
806 *
807 * NOTE! Hammer shouldn't need to bawrite()/cluster_write().
808 * The ip->rsv_recs check should burst-flush the data.
809 * If we queue it immediately the buf could be left
810 * locked on the device queue for a very long time.
811 *
812 * However, failing to flush a dirty buffer out when
813 * issued from the pageout daemon can result in a low
814 * memory deadlock against bio_page_alloc(), so we
815 * have to bawrite() on IO_ASYNC as well.
816 *
817 * NOTE! To avoid degenerate stalls due to mismatched block
818 * sizes we only honor IO_DIRECT on the write which
819 * abuts the end of the buffer. However, we must
820 * honor IO_SYNC in case someone is silly enough to
821 * configure a HAMMER file as swap, or when HAMMER
822 * is serving NFS (for commits). Ick ick.
823 */
839 else
844 }
845 }
850 }
852 /*
853 * hammer_vop_access { vp, mode, cred }
854 *
855 * MPSAFE - does not require fs_token
856 */
857 static
858 int
860 {
862 uid_t uid;
863 gid_t gid;
872 }
874 /*
875 * hammer_vop_advlock { vp, id, op, fl, flags }
876 *
877 * MPSAFE - does not require fs_token
878 */
879 static
880 int
882 {
886 }
888 /*
889 * hammer_vop_close { vp, fflag }
890 *
891 * We can only sync-on-close for normal closes. XXX disabled for now.
892 */
893 static
894 int
896 {
897 #if 0
906 else
909 HAMMER_INODE_CLOSEASYNC);
911 }
912 }
913 #endif
915 }
917 /*
918 * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
919 *
920 * The operating system has already ensured that the directory entry
921 * does not exist and done all appropriate namespace locking.
922 */
923 static
924 int
926 {
928 hammer_inode_t dip;
929 hammer_inode_t nip;
931 hammer_mount_t hmp;
943 /*
944 * Create a transaction to cover the operations we perform.
945 */
949 /*
950 * Create a new filesystem object of the requested type. The
951 * returned inode will be referenced and shared-locked to prevent
952 * it from being moved to the flusher.
953 */
963 }
965 /*
966 * Add the new filesystem object to the directory. This will also
967 * bump the inode's link count.
968 */
971 nip);
975 /*
976 * Finish up.
977 */
989 }
991 }
994 }
996 /*
997 * hammer_vop_getattr { vp, vap }
998 *
999 * Retrieve an inode's attribute information. When accessing inodes
1000 * historically we fake the atime field to ensure consistent results.
1001 * The atime field is stored in the B-Tree element and allowed to be
1002 * updated without cycling the element.
1003 *
1004 * MPSAFE - does not require fs_token
1005 */
1006 static
1007 int
1009 {
1013 /*
1014 * We want the fsid to be different when accessing a filesystem
1015 * with different as-of's so programs like diff don't think
1016 * the files are the same.
1017 *
1018 * We also want the fsid to be the same when comparing snapshots,
1019 * or when comparing mirrors (which might be backed by different
1020 * physical devices). HAMMER fsids are based on the PFS's
1021 * shared_uuid field.
1022 *
1023 * XXX there is a chance of collision here. The va_fsid reported
1024 * by stat is different from the more involved fsid used in the
1025 * mount structure.
1026 */
1040 /*
1041 * Special case for @@PFS softlinks. The actual size of the
1042 * expanded softlink is "@@0x%016llx:%05d" == 26 bytes.
1043 * or for MAX_TID is "@@-1:%05d" == 10 bytes.
1044 *
1045 * Note that userspace hammer command does not allow users to
1046 * create a @@PFS softlink under an existing other PFS (id!=0)
1047 * so the ip localization here for @@PFS softlink is always 0.
1048 */
1056 else
1058 }
1060 /*
1061 * We must provide a consistent atime and mtime for snapshots
1062 * so people can do a 'tar cf - ... | md5' on them and get
1063 * consistent results.
1064 */
1071 }
1082 }
1090 VA_FSID_UUID_VALID;
1100 }
1103 }
1105 /*
1106 * hammer_vop_nresolve { nch, dvp, cred }
1107 *
1108 * Locate the requested directory entry.
1109 */
1110 static
1111 int
1113 {
1116 hammer_mount_t hmp;
1117 hammer_inode_t dip;
1118 hammer_inode_t ip;
1119 hammer_tid_t asof;
1132 /*
1133 * Misc initialization, plus handle as-of name extensions. Look for
1134 * the '@@' extension. Note that as-of files and directories cannot
1135 * be modified.
1136 */
1156 }
1160 }
1161 }
1164 /*
1165 * If this is a PFS we dive into the PFS root inode
1166 */
1176 }
1181 }
1183 }
1185 /*
1186 * If there is no path component the time extension is relative to dip.
1187 * e.g. "fubar/@@<snapshot>"
1188 *
1189 * "." is handled by the kernel, but ".@@<snapshot>" is not.
1190 * e.g. "fubar/.@@<snapshot>"
1191 *
1192 * ".." is handled by the kernel. We do not currently handle
1193 * "..@<snapshot>".
1194 */
1204 }
1209 }
1211 }
1213 /*
1214 * Calculate the namekey and setup the key range for the scan. This
1215 * works kinda like a chained hash table where the lower 32 bits
1216 * of the namekey synthesize the chain.
1217 *
1218 * The key range is inclusive of both key_beg and key_end.
1219 */
1238 /*
1239 * Scan all matching records (the chain), locate the one matching
1240 * the requested path component.
1241 *
1242 * The hammer_ip_*() functions merge in-memory records with on-disk
1243 * records for the purposes of the search.
1244 */
1259 }
1261 }
1262 }
1265 /*
1266 * Lookup the obj_id. This should always succeed. If it does not
1267 * the filesystem may be damaged and we return a dummy inode.
1268 */
1278 localization);
1283 }
1289 }
1294 }
1297 }
1298 done:
1302 }
1304 /*
1305 * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1306 *
1307 * Locate the parent directory of a directory vnode.
1308 *
1309 * dvp is referenced but not locked. *vpp must be returned referenced and
1310 * locked. A parent_obj_id of 0 indicates that we are at the root.
1311 *
1312 * NOTE: as-of sequences are not linked into the directory structure. If
1313 * we are at the root with a different asof then the mount point, reload
1314 * the same directory with the mount point's asof. I'm not sure what this
1315 * will do to NFS. We encode ASOF stamps in NFS file handles so it might not
1316 * get confused, but it hasn't been tested.
1317 */
1318 static
1319 int
1321 {
1323 hammer_inode_t dip;
1324 hammer_inode_t ip;
1325 hammer_mount_t hmp;
1328 hammer_tid_t asof;
1335 /*
1336 * Whos are parent? This could be the root of a pseudo-filesystem
1337 * whos parent is in another localization domain.
1338 */
1343 else
1346 /*
1347 * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1348 */
1361 }
1362 }
1374 }
1378 }
1380 /*
1381 * hammer_vop_nlink { nch, dvp, vp, cred }
1382 */
1383 static
1384 int
1386 {
1388 hammer_inode_t dip;
1389 hammer_inode_t ip;
1391 hammer_mount_t hmp;
1412 /*
1413 * Create a transaction to cover the operations we perform.
1414 */
1418 /*
1419 * Add the filesystem object to the directory. Note that neither
1420 * dip nor ip are referenced or locked, but their vnodes are
1421 * referenced. This function will bump the inode's link count.
1422 */
1425 ip);
1427 /*
1428 * Finish up.
1429 */
1433 }
1439 }
1441 /*
1442 * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1443 *
1444 * The operating system has already ensured that the directory entry
1445 * does not exist and done all appropriate namespace locking.
1446 */
1447 static
1448 int
1450 {
1452 hammer_inode_t dip;
1453 hammer_inode_t nip;
1455 hammer_mount_t hmp;
1467 /*
1468 * Create a transaction to cover the operations we perform.
1469 */
1473 /*
1474 * Create a new filesystem object of the requested type. The
1475 * returned inode will be referenced but not locked.
1476 */
1485 }
1486 /*
1487 * Add the new filesystem object to the directory. This will also
1488 * bump the inode's link count.
1489 */
1492 nip);
1496 /*
1497 * Finish up.
1498 */
1508 }
1509 }
1515 }
1517 /*
1518 * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1519 *
1520 * The operating system has already ensured that the directory entry
1521 * does not exist and done all appropriate namespace locking.
1522 */
1523 static
1524 int
1526 {
1528 hammer_inode_t dip;
1529 hammer_inode_t nip;
1531 hammer_mount_t hmp;
1543 /*
1544 * Create a transaction to cover the operations we perform.
1545 */
1549 /*
1550 * Create a new filesystem object of the requested type. The
1551 * returned inode will be referenced but not locked.
1552 *
1553 * If mknod specifies a directory a pseudo-fs is created.
1554 */
1563 }
1565 /*
1566 * Add the new filesystem object to the directory. This will also
1567 * bump the inode's link count.
1568 */
1571 nip);
1573 /*
1574 * Finish up.
1575 */
1585 }
1586 }
1592 }
1594 /*
1595 * hammer_vop_open { vp, mode, cred, fp }
1596 *
1597 * MPSAFE (does not require fs_token)
1598 */
1599 static
1600 int
1602 {
1603 hammer_inode_t ip;
1610 }
1612 /*
1613 * hammer_vop_print { vp }
1614 */
1615 static
1616 int
1618 {
1620 }
1622 /*
1623 * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1624 */
1625 static
1626 int
1628 {
1631 hammer_inode_t ip;
1632 hammer_mount_t hmp;
1634 hammer_base_elm_t base;
1639 off_t saveoff;
1658 }
1663 /*
1664 * Handle artificial entries
1665 *
1666 * It should be noted that the minimum value for a directory
1667 * hash key on-media is 0x0000000100000000, so we can use anything
1668 * less then that to represent our 'special' key space.
1669 */
1681 }
1690 }
1699 }
1701 /*
1702 * Key range (begin and end inclusive) to scan. Directory keys
1703 * directly translate to a 64 bit 'seek' position.
1704 */
1736 dtype,
1748 }
1751 done:
1764 }
1771 }
1772 }
1775 }
1777 /*
1778 * hammer_vop_readlink { vp, uio, cred }
1779 */
1780 static
1781 int
1783 {
1786 hammer_inode_t ip;
1787 hammer_mount_t hmp;
1790 hammer_pseudofs_inmem_t pfsm;
1798 /*
1799 * Shortcut if the symlink data was stuffed into ino_data.
1800 *
1801 * Also expand special "@@PFS%05d" softlinks (expansion only
1802 * occurs for non-historical (current) accesses made from the
1803 * primary filesystem).
1804 *
1805 * Note that userspace hammer command does not allow users to
1806 * create a @@PFS softlink under an existing other PFS (id!=0)
1807 * so the ip localization here for @@PFS softlink is always 0.
1808 */
1827 /* vap->va_size == 26 */
1833 /* vap->va_size == 10 */
1837 }
1840 }
1844 }
1848 }
1850 /*
1851 * Long version
1852 */
1856 /*
1857 * Key range (begin and end inclusive) to scan. Directory keys
1858 * directly translate to a 64 bit 'seek' position.
1859 */
1861 HAMMER_LOCALIZE_MISC;
1876 HAMMER_SYMLINK_NAME_OFF);
1879 HAMMER_SYMLINK_NAME_OFF,
1881 }
1882 }
1887 }
1889 /*
1890 * hammer_vop_nremove { nch, dvp, cred }
1891 */
1892 static
1893 int
1895 {
1897 hammer_inode_t dip;
1898 hammer_mount_t hmp;
1907 }
1917 }
1919 /*
1920 * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1921 */
1922 static
1923 int
1925 {
1929 hammer_inode_t fdip;
1930 hammer_inode_t tdip;
1931 hammer_inode_t ip;
1932 hammer_mount_t hmp;
1969 /*
1970 * Remove tncp from the target directory and then link ip as
1971 * tncp. XXX pass trans to dounlink
1972 *
1973 * Force the inode sync-time to match the transaction so it is
1974 * in-sync with the creation of the target directory entry.
1975 */
1981 ip);
1986 }
1987 }
1991 /*
1992 * Locate the record in the originating directory and remove it.
1993 *
1994 * Calculate the namekey and setup the key range for the scan. This
1995 * works kinda like a chained hash table where the lower 32 bits
1996 * of the namekey synthesize the chain.
1997 *
1998 * The key range is inclusive of both key_beg and key_end.
1999 */
2002 retry:
2018 /*
2019 * Scan all matching records (the chain), locate the one matching
2020 * the requested path component.
2021 *
2022 * The hammer_ip_*() functions merge in-memory records with on-disk
2023 * records for the purposes of the search.
2024 */
2034 }
2036 }
2038 /*
2039 * If all is ok we have to get the inode so we can adjust nlinks.
2040 *
2041 * WARNING: hammer_ip_del_direntry() may have to terminate the
2042 * cursor to avoid a recursion. It's ok to call hammer_done_cursor()
2043 * twice.
2044 */
2048 /*
2049 * XXX A deadlock here will break rename's atomicy for the purposes
2050 * of crash recovery.
2051 */
2055 }
2057 /*
2058 * Cleanup and tell the kernel that the rename succeeded.
2059 *
2060 * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2061 * without formally acquiring the vp since the vp might
2062 * have zero refs on it, or in the middle of a reclaim,
2063 * etc.
2064 */
2079 }
2081 }
2082 }
2084 failed:
2088 }
2090 /*
2091 * hammer_vop_nrmdir { nch, dvp, cred }
2092 */
2093 static
2094 int
2096 {
2098 hammer_inode_t dip;
2099 hammer_mount_t hmp;
2108 }
2118 }
2120 /*
2121 * hammer_vop_markatime { vp, cred }
2122 */
2123 static
2124 int
2126 {
2128 hammer_inode_t ip;
2129 hammer_mount_t hmp;
2148 }
2150 /*
2151 * hammer_vop_setattr { vp, vap, cred }
2152 */
2153 static
2154 int
2156 {
2158 hammer_inode_t ip;
2160 hammer_mount_t hmp;
2166 #if 0
2168 #endif
2184 }
2201 }
2205 }
2206 }
2208 }
2212 }
2217 hammer_uuid_t uuid_uid;
2218 hammer_uuid_t uuid_gid;
2234 }
2236 }
2237 }
2244 /*
2245 * Log the operation if in fast-fsync mode or if
2246 * there are unterminated redo write records present.
2247 *
2248 * The second check is needed so the recovery code
2249 * properly truncates write redos even if nominal
2250 * REDO operations is turned off due to excessive
2251 * writes, because the related records might be
2252 * destroyed and never lay down a TERM_WRITE.
2253 */
2258 HAMMER_REDO_TRUNC,
2260 }
2263 /*
2264 * XXX break atomicy, we can deadlock the backend
2265 * if we do not release the lock. Probably not a
2266 * big deal here.
2267 */
2270 blksize,
2286 }
2289 /* XXX safe to use SDIRTY instead of DDIRTY here? */
2293 /*
2294 * On-media truncation is cached in the inode until
2295 * the inode is synchronized. We must immediately
2296 * handle any frontend records.
2297 */
2306 }
2307 }
2309 #if 0
2310 /*
2311 * When truncating, nvtruncbuf() may have cleaned out
2312 * a portion of the last block on-disk in the buffer
2313 * cache. We must clean out any frontend records
2314 * for blocks beyond the new last block.
2315 */
2321 }
2322 #endif
2331 }
2342 }
2344 }
2349 }
2355 }
2368 }
2369 }
2370 done:
2377 }
2379 /*
2380 * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2381 */
2382 static
2383 int
2385 {
2387 hammer_inode_t dip;
2388 hammer_inode_t nip;
2389 hammer_record_t record;
2391 hammer_mount_t hmp;
2406 /*
2407 * Create a transaction to cover the operations we perform.
2408 */
2412 /*
2413 * Create a new filesystem object of the requested type. The
2414 * returned inode will be referenced but not locked.
2415 */
2425 }
2427 /*
2428 * Add a record representing the symlink. symlink stores the link
2429 * as pure data, not a string, and is no \0 terminated.
2430 */
2441 HAMMER_LOCALIZE_MISC;
2448 }
2450 /*
2451 * Set the file size to the length of the link.
2452 */
2456 }
2457 }
2462 /*
2463 * Finish up.
2464 */
2475 }
2476 }
2480 }
2482 /*
2483 * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2484 */
2485 static
2486 int
2488 {
2490 hammer_inode_t dip;
2491 hammer_mount_t hmp;
2500 }
2510 }
2512 /*
2513 * hammer_vop_ioctl { vp, command, data, fflag, cred }
2514 */
2515 static
2516 int
2518 {
2528 }
2530 static
2531 int
2533 {
2540 };
2541 hammer_mount_t hmp;
2558 else
2563 /*
2564 * Call standard mountctl VOP function
2565 * so we get user mount flags.
2566 */
2578 }
2585 }
2588 }
2590 /*
2591 * hammer_vop_strategy { vp, bio }
2592 *
2593 * Strategy call, used for regular file read & write only. Note that the
2594 * bp may represent a cluster.
2595 *
2596 * To simplify operation and allow better optimizations in the future,
2597 * this code does not make any assumptions with regards to buffer alignment
2598 * or size.
2599 */
2600 static
2601 int
2603 {
2621 }
2623 }
2625 /*
2626 * Read from a regular file. Iterate the related records and fill in the
2627 * BIO/BUF. Gaps are zero-filled.
2628 *
2629 * The support code in hammer_object.c should be used to deal with mixed
2630 * in-memory and on-disk records.
2631 *
2632 * NOTE: Can be called from the cluster code with an oversized buf.
2633 *
2634 * XXX atime update
2635 */
2636 static
2637 int
2639 {
2641 hammer_inode_t ip;
2642 hammer_inode_t dip;
2643 hammer_mount_t hmp;
2645 hammer_base_elm_t base;
2646 hammer_off_t disk_offset;
2664 /*
2665 * The zone-2 disk offset may have been set by the cluster code via
2666 * a BMAP operation, or else should be NOOFFSET.
2667 *
2668 * Checking the high bits for a match against zone-2 should suffice.
2669 *
2670 * In cases where a lot of data duplication is present it may be
2671 * more beneficial to drop through and doubule-buffer through the
2672 * device.
2673 */
2681 }
2683 /*
2684 * Try to shortcut requests for double_buffer mode too.
2685 * Since this mode runs through the device buffer cache
2686 * only compatible buffer sizes (meaning those generated
2687 * by normal filesystem buffers) are legal.
2688 */
2694 }
2695 }
2697 /*
2698 * Well, that sucked. Do it the hard way. If all the stars are
2699 * aligned we may still be able to issue a direct-read.
2700 */
2705 /*
2706 * Key range (begin and end inclusive) to scan. Note that the key's
2707 * stored in the actual records represent BASE+LEN, not BASE. The
2708 * first record containing bio_offset will have a key > bio_offset.
2709 */
2711 HAMMER_LOCALIZE_MISC;
2722 #if 0
2728 #endif
2729 {
2736 else
2738 }
2741 /*
2742 * Set NOSWAPCACHE for cursor data extraction if double buffering
2743 * is disabled or (if the file is not marked cacheable via chflags
2744 * and vm.swapcache_use_chflags is enabled).
2745 */
2748 vm_swapcache_use_chflags)) {
2750 }
2756 /*
2757 * Get the base file offset of the record. The key for
2758 * data records is (base + bytes) rather then (base).
2759 */
2763 /*
2764 * Calculate the gap, if any, and zero-fill it.
2765 *
2766 * n is the offset of the start of the record verses our
2767 * current seek offset in the bio.
2768 */
2776 }
2778 /*
2779 * Calculate the data offset in the record and the number
2780 * of bytes we can copy.
2781 *
2782 * There are two degenerate cases. First, boff may already
2783 * be at bp->b_bufsize. Secondly, the data offset within
2784 * the record may exceed the record's size.
2785 */
2794 }
2796 /*
2797 * Deal with cached truncations. This cool bit of code
2798 * allows truncate()/ftruncate() to avoid having to sync
2799 * the file.
2800 *
2801 * If the frontend is truncated then all backend records are
2802 * subject to the frontend's truncation.
2803 *
2804 * If the backend is truncated then backend records on-disk
2805 * (but not in-memory) are subject to the backend's
2806 * truncation. In-memory records owned by the backend
2807 * represent data written after the truncation point on the
2808 * backend and must not be truncated.
2809 *
2810 * Truncate operations deal with frontend buffer cache
2811 * buffers and frontend-owned in-memory records synchronously.
2812 */
2820 }
2821 }
2828 }
2829 }
2831 /*
2832 * Try to issue a direct read into our bio if possible,
2833 * otherwise resolve the element data into a hammer_buffer
2834 * and copy.
2835 *
2836 * The buffer on-disk should be zerod past any real
2837 * truncation point, but may not be for any synthesized
2838 * truncation point from above.
2839 *
2840 * NOTE: disk_offset is only valid if the cursor data is
2841 * on-disk.
2842 */
2849 /*
2850 * Direct read case
2851 */
2857 /*
2858 * Async I/O case for reading from backing store
2859 * and copying the data to the filesystem buffer.
2860 */
2870 }
2871 }
2875 /*
2876 * Iterate until we have filled the request.
2877 */
2882 }
2884 /*
2885 * There may have been a gap after the last record
2886 */
2892 /* boff = bp->b_bufsize; */
2893 }
2895 /*
2896 * Disallow swapcache operation on the vnode buffer if double
2897 * buffering is enabled, the swapcache will get the data via
2898 * the block device buffer.
2899 */
2903 /*
2904 * Cleanup
2905 */
2912 done:
2913 /*
2914 * Cache the b-tree node for the last data read in cache[1].
2915 *
2916 * If we hit the file EOF then also cache the node in the
2917 * governing directory's cache[3], it will be used to initialize
2918 * the new inode's cache[1] for any inodes looked up via the directory.
2919 *
2920 * This doesn't reduce disk accesses since the B-Tree chain is
2921 * likely cached, but it does reduce cpu overhead when looking
2922 * up file offsets for cpdup/tar/cpio style iterations.
2923 */
2932 }
2933 }
2938 }
2940 /*
2941 * BMAP operation - used to support cluster_read() only.
2942 *
2943 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2944 *
2945 * This routine may return EOPNOTSUPP if the opration is not supported for
2946 * the specified offset. The contents of the pointer arguments do not
2947 * need to be initialized in that case.
2948 *
2949 * If a disk address is available and properly aligned return 0 with
2950 * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2951 * to the run-length relative to that offset. Callers may assume that
2952 * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2953 * large, so return EOPNOTSUPP if it is not sufficiently large.
2954 */
2955 static
2956 int
2958 {
2960 hammer_inode_t ip;
2961 hammer_mount_t hmp;
2963 hammer_base_elm_t base;
2970 hammer_off_t last_disk_offset;
2971 hammer_off_t disk_offset;
2979 /*
2980 * We can only BMAP regular files. We can't BMAP database files,
2981 * directories, etc.
2982 */
2986 /*
2987 * bmap is typically called with runp/runb both NULL when used
2988 * for writing. We do not support BMAP for writing atm.
2989 */
2993 /*
2994 * Scan the B-Tree to acquire blockmap addresses, then translate
2995 * to raw addresses.
2996 */
3002 /*
3003 * Key range (begin and end inclusive) to scan. Note that the key's
3004 * stored in the actual records represent BASE+LEN, not BASE. The
3005 * first record containing bio_offset will have a key > bio_offset.
3006 */
3008 HAMMER_LOCALIZE_MISC;
3015 else
3031 else
3041 /*
3042 * Get the base file offset of the record. The key for
3043 * data records is (base + bytes) rather then (base).
3044 *
3045 * NOTE: rec_offset + rec_len may exceed the end-of-file.
3046 * The extra bytes should be zero on-disk and the BMAP op
3047 * should still be ok.
3048 */
3053 /*
3054 * Incorporate any cached truncation.
3055 *
3056 * NOTE: Modifications to rec_len based on synthesized
3057 * truncation points remove the guarantee that any extended
3058 * data on disk is zero (since the truncations may not have
3059 * taken place on-media yet).
3060 */
3068 }
3069 }
3076 }
3077 }
3079 /*
3080 * Accumulate information. If we have hit a discontiguous
3081 * block reset base_offset unless we are already beyond the
3082 * requested offset. If we are, that's it, we stop.
3083 */
3094 }
3097 }
3099 }
3108 /*
3109 * If we couldn't find any records or the records we did find were
3110 * all behind the requested offset, return failure. A forward
3111 * truncation can leave a hole w/ no on-disk records.
3112 */
3116 /*
3117 * Figure out the block size at the requested offset and adjust
3118 * our limits so the cluster_read() does not create inappropriately
3119 * sized buffer cache buffers.
3120 */
3124 }
3129 }
3131 /*
3132 * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3133 * from occuring.
3134 */
3138 /*
3139 * Only large-data zones can be direct-IOd
3140 */
3144 /*
3145 * doffsetp is not aligned or the forward run size does
3146 * not cover a whole buffer, disallow the direct I/O.
3147 */
3150 /*
3151 * We're good.
3152 */
3157 }
3161 }
3163 }
3165 }
3167 /*
3168 * Write to a regular file. Because this is a strategy call the OS is
3169 * trying to actually get data onto the media.
3170 */
3171 static
3172 int
3174 {
3175 hammer_record_t record;
3176 hammer_mount_t hmp;
3177 hammer_inode_t ip;
3197 }
3201 /*
3202 * Disallow swapcache operation on the vnode buffer if double
3203 * buffering is enabled, the swapcache will get the data via
3204 * the block device buffer.
3205 */
3209 /*
3210 * Interlock with inode destruction (no in-kernel or directory
3211 * topology visibility). If we queue new IO while trying to
3212 * destroy the inode we can deadlock the vtrunc call in
3213 * hammer_inode_unloadable_check().
3214 *
3215 * Besides, there's no point flushing a bp associated with an
3216 * inode that is being destroyed on-media and has no kernel
3217 * references.
3218 */
3225 }
3227 /*
3228 * Reserve space and issue a direct-write from the front-end.
3229 * NOTE: The direct_io code will hammer_bread/bcopy smaller
3230 * allocations.
3231 *
3232 * An in-memory record will be installed to reference the storage
3233 * until the flusher can get to it.
3234 *
3235 * Since we own the high level bio the front-end will not try to
3236 * do a direct-read until the write completes.
3237 *
3238 * NOTE: The only time we do not reserve a full-sized buffers
3239 * worth of data is if the file is small. We do not try to
3240 * allocate a fragment (from the small-data zone) at the end of
3241 * an otherwise large file as this can lead to wildly separated
3242 * data.
3243 */
3248 else
3254 /*
3255 * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3256 * in hammer_vop_write(). We must flag the record so the proper
3257 * REDO_TERM_WRITE entry is generated during the flush.
3258 */
3263 }
3272 }
3275 }
3277 /*
3278 * dounlink - disconnect a directory entry
3279 *
3280 * XXX whiteout support not really in yet
3281 */
3282 static int
3286 {
3288 hammer_inode_t dip;
3289 hammer_inode_t ip;
3290 hammer_mount_t hmp;
3296 /*
3297 * Calculate the namekey and setup the key range for the scan. This
3298 * works kinda like a chained hash table where the lower 32 bits
3299 * of the namekey synthesize the chain.
3300 *
3301 * The key range is inclusive of both key_beg and key_end.
3302 */
3312 retry:
3328 /*
3329 * Scan all matching records (the chain), locate the one matching
3330 * the requested path component. info->last_error contains the
3331 * error code on search termination and could be 0, ENOENT, or
3332 * something else.
3333 *
3334 * The hammer_ip_*() functions merge in-memory records with on-disk
3335 * records for the purposes of the search.
3336 */
3348 }
3350 }
3352 /*
3353 * If all is ok we have to get the inode so we can adjust nlinks.
3354 * To avoid a deadlock with the flusher we must release the inode
3355 * lock on the directory when acquiring the inode for the entry.
3356 *
3357 * If the target is a directory, it must be empty.
3358 */
3373 }
3375 /*
3376 * If isdir >= 0 we validate that the entry is or is not a
3377 * directory. If isdir < 0 we don't care.
3378 */
3386 }
3387 }
3389 /*
3390 * If we are trying to remove a directory the directory must
3391 * be empty.
3392 *
3393 * The check directory code can loop and deadlock/retry. Our
3394 * own cursor's node locks must be released to avoid a 3-way
3395 * deadlock with the flusher if the check directory code
3396 * blocks.
3397 *
3398 * If any changes whatsoever have been made to the cursor
3399 * set EDEADLK and retry.
3400 *
3401 * WARNING: See warnings in hammer_unlock_cursor()
3402 * function.
3403 */
3405 HAMMER_OBJTYPE_DIRECTORY) {
3414 }
3415 }
3417 /*
3418 * Delete the directory entry.
3419 *
3420 * WARNING: hammer_ip_del_direntry() may have to terminate
3421 * the cursor to avoid a deadlock. It is ok to call
3422 * hammer_done_cursor() twice.
3423 */
3427 }
3430 /*
3431 * Tell the namecache that we are now unlinked.
3432 */
3435 /*
3436 * NOTE: ip->vp, if non-NULL, cannot be directly
3437 * referenced without formally acquiring the
3438 * vp since the vp might have zero refs on it,
3439 * or in the middle of a reclaim, etc.
3440 *
3441 * NOTE: The cache_setunresolved() can rip the vp
3442 * out from under us since the vp may not have
3443 * any refs, in which case ip->vp will be NULL
3444 * from the outset.
3445 */
3453 #if 0
3454 /*
3455 * Don't do this, it can deadlock
3456 * on concurrent rm's of hardlinks.
3457 * Shouldn't be needed any more.
3458 */
3460 #endif
3463 }
3465 }
3466 }
3471 }
3476 }
3478 /************************************************************************
3479 * FIFO AND SPECFS OPS *
3480 ************************************************************************
3481 *
3482 */
3483 static int
3485 {
3486 /* XXX update itimes */
3488 }
3490 static int
3492 {
3496 /* XXX update access time */
3498 }
3500 static int
3502 {
3506 /* XXX update access time */
3508 }
3510 static
3511 int
3513 {
3520 }
3522 /************************************************************************
3523 * KQFILTER OPS *
3524 ************************************************************************
3525 *
3526 */
3542 static
3543 int
3545 {
3561 }
3568 }
3570 static void
3572 {
3576 }
3578 static int
3580 {
3584 off_t off;
3589 }
3597 }
3599 static int
3601 {
3606 }
3608 static int
3610 {
3616 }
3618 }