| blob | 45b8b411d90fe7b17b4d9512812e23d1ffa3681f |
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 }
420 }
424 }
425 skip:
431 }
442 /*
443 * Set B_AGE, data has a lower priority than meta-data.
444 *
445 * Use a hold/unlock/drop sequence to run the uiomove
446 * with the buffer unlocked, avoiding deadlocks against
447 * read()s on mmap()'d spaces.
448 */
456 }
458 finished:
460 /*
461 * Try to update the atime with just the inode lock for maximum
462 * concurrency. If we can't shortcut it we have to get the full
463 * blown transaction.
464 */
468 }
480 }
481 }
483 }
485 /*
486 * hammer_vop_write { vp, uio, ioflag, cred }
487 */
488 static
489 int
491 {
493 hammer_inode_t ip;
494 hammer_mount_t hmp;
495 thread_t td;
498 off_t base_offset;
519 /*
520 * Create a transaction to cover the operations we perform.
521 */
525 /*
526 * Check append mode
527 */
531 /*
532 * Check for illegal write offsets. Valid range is 0...2^63-1.
533 *
534 * NOTE: the base_off assignment is required to work around what
535 * I consider to be a GCC-4 optimization bug.
536 */
540 }
545 }
552 }
554 /*
555 * If reading or writing a huge amount of data we have to break
556 * atomicy and allow the operation to be interrupted by a signal
557 * or it can DOS the machine.
558 *
559 * Preset redo_count so we stop generating REDOs earlier if the
560 * limit is exceeded.
561 *
562 * redo_count is heuristical, SMP races are ok
563 */
568 }
570 /*
571 * Access the data typically in HAMMER_BUFSIZE blocks via the
572 * buffer cache, but HAMMER may use a variable block size based
573 * on the offset.
574 */
581 off_t nsize;
590 /*
591 * Control the number of pending records associated with
592 * this inode. If too many have accumulated start a
593 * flush. Try to maintain a pipeline with the flusher.
594 *
595 * NOTE: It is possible for other sources to grow the
596 * records but not necessarily issue another flush,
597 * so use a timeout and ensure that a re-flush occurs.
598 */
606 }
608 }
610 /*
611 * Do not allow HAMMER to blow out the buffer cache. Very
612 * large UIOs can lockout other processes due to bwillwrite()
613 * mechanics.
614 *
615 * The hammer inode is not locked during these operations.
616 * The vnode is locked which can interfere with the pageout
617 * daemon for non-UIO_NOCOPY writes but should not interfere
618 * with the buffer cache. Even so, we cannot afford to
619 * allow the pageout daemon to build up too many dirty buffer
620 * cache buffers.
621 *
622 * Only call this if we aren't being recursively called from
623 * a virtual disk device (vn), else we may deadlock.
624 */
628 /*
629 * Calculate the blocksize at the current offset and figure
630 * out how much we can actually write.
631 */
641 }
646 else
650 nsize,
655 trivial);
658 }
661 /*
662 * Issuing a write with the same data backing the
663 * buffer. Instantiate the buffer to collect the
664 * backing vm pages, then read-in any missing bits.
665 *
666 * This case is used by vop_stdputpages().
667 */
674 }
676 /*
677 * Even though we are entirely overwriting the buffer
678 * we may still have to zero it out to avoid a
679 * mmap/write visibility issue.
680 */
685 /*
686 * If the base offset of the buffer is beyond the
687 * file EOF, we don't have to issue a read.
688 */
693 /*
694 * Partial overwrite, read in any missing bits then
695 * replace the portion being written.
696 */
700 }
706 /*
707 * Generate REDO records if enabled and redo_count will not
708 * exceeded the limit.
709 *
710 * If redo_count exceeds the limit we stop generating records
711 * and clear HAMMER_INODE_REDO. This will cause the next
712 * fsync() to do a full meta-data sync instead of just an
713 * UNDO/REDO fifo update.
714 *
715 * When clearing HAMMER_INODE_REDO any pre-existing REDOs
716 * will still be tracked. The tracks will be terminated
717 * when the related meta-data (including possible data
718 * modifications which are not tracked via REDO) is
719 * flushed.
720 */
726 HAMMER_REDO_WRITE,
731 }
732 }
734 /*
735 * If we screwed up we have to undo any VM size changes we
736 * made.
737 */
745 }
748 }
758 }
763 /*
764 * Once we dirty the buffer any cached zone-X offset
765 * becomes invalid. HAMMER NOTE: no-history mode cannot
766 * allow overwriting over the same data sector unless
767 * we provide UNDOs for the old data, which we don't.
768 */
773 /*
774 * Final buffer disposition.
775 *
776 * Because meta-data updates are deferred, HAMMER is
777 * especially sensitive to excessive bdwrite()s because
778 * the I/O stream is not broken up by disk reads. So the
779 * buffer cache simply cannot keep up.
780 *
781 * WARNING! blksize is variable. cluster_write() is
782 * expected to not blow up if it encounters
783 * buffers that do not match the passed blksize.
784 *
785 * NOTE! Hammer shouldn't need to bawrite()/cluster_write().
786 * The ip->rsv_recs check should burst-flush the data.
787 * If we queue it immediately the buf could be left
788 * locked on the device queue for a very long time.
789 *
790 * However, failing to flush a dirty buffer out when
791 * issued from the pageout daemon can result in a low
792 * memory deadlock against bio_page_alloc(), so we
793 * have to bawrite() on IO_ASYNC as well.
794 *
795 * NOTE! To avoid degenerate stalls due to mismatched block
796 * sizes we only honor IO_DIRECT on the write which
797 * abuts the end of the buffer. However, we must
798 * honor IO_SYNC in case someone is silly enough to
799 * configure a HAMMER file as swap, or when HAMMER
800 * is serving NFS (for commits). Ick ick.
801 */
817 else
822 }
823 }
828 }
830 /*
831 * hammer_vop_access { vp, mode, cred }
832 *
833 * MPSAFE - does not require fs_token
834 */
835 static
836 int
838 {
840 uid_t uid;
841 gid_t gid;
851 }
853 /*
854 * hammer_vop_advlock { vp, id, op, fl, flags }
855 *
856 * MPSAFE - does not require fs_token
857 */
858 static
859 int
861 {
865 }
867 /*
868 * hammer_vop_close { vp, fflag }
869 *
870 * We can only sync-on-close for normal closes. XXX disabled for now.
871 */
872 static
873 int
875 {
876 #if 0
885 else
888 HAMMER_INODE_CLOSEASYNC);
890 }
891 }
892 #endif
894 }
896 /*
897 * hammer_vop_ncreate { nch, dvp, vpp, cred, vap }
898 *
899 * The operating system has already ensured that the directory entry
900 * does not exist and done all appropriate namespace locking.
901 */
902 static
903 int
905 {
907 hammer_inode_t dip;
908 hammer_inode_t nip;
910 hammer_mount_t hmp;
922 /*
923 * Create a transaction to cover the operations we perform.
924 */
929 /*
930 * Create a new filesystem object of the requested type. The
931 * returned inode will be referenced and shared-locked to prevent
932 * it from being moved to the flusher.
933 */
943 }
945 /*
946 * Add the new filesystem object to the directory. This will also
947 * bump the inode's link count.
948 */
951 nip);
955 /*
956 * Finish up.
957 */
969 }
971 }
974 }
976 /*
977 * hammer_vop_getattr { vp, vap }
978 *
979 * Retrieve an inode's attribute information. When accessing inodes
980 * historically we fake the atime field to ensure consistent results.
981 * The atime field is stored in the B-Tree element and allowed to be
982 * updated without cycling the element.
983 *
984 * MPSAFE - does not require fs_token
985 */
986 static
987 int
989 {
993 /*
994 * We want the fsid to be different when accessing a filesystem
995 * with different as-of's so programs like diff don't think
996 * the files are the same.
997 *
998 * We also want the fsid to be the same when comparing snapshots,
999 * or when comparing mirrors (which might be backed by different
1000 * physical devices). HAMMER fsids are based on the PFS's
1001 * shared_uuid field.
1002 *
1003 * XXX there is a chance of collision here. The va_fsid reported
1004 * by stat is different from the more involved fsid used in the
1005 * mount structure.
1006 */
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 */
1138 }
1142 }
1143 }
1146 /*
1147 * If this is a PFS softlink we dive into the PFS
1148 */
1158 }
1163 }
1165 }
1167 /*
1168 * If there is no path component the time extension is relative to dip.
1169 * e.g. "fubar/@@<snapshot>"
1170 *
1171 * "." is handled by the kernel, but ".@@<snapshot>" is not.
1172 * e.g. "fubar/.@@<snapshot>"
1173 *
1174 * ".." is handled by the kernel. We do not currently handle
1175 * "..@<snapshot>".
1176 */
1186 }
1191 }
1193 }
1195 /*
1196 * Calculate the namekey and setup the key range for the scan. This
1197 * works kinda like a chained hash table where the lower 32 bits
1198 * of the namekey synthesize the chain.
1199 *
1200 * The key range is inclusive of both key_beg and key_end.
1201 */
1220 /*
1221 * Scan all matching records (the chain), locate the one matching
1222 * the requested path component.
1223 *
1224 * The hammer_ip_*() functions merge in-memory records with on-disk
1225 * records for the purposes of the search.
1226 */
1241 }
1243 }
1244 }
1247 /*
1248 * Lookup the obj_id. This should always succeed. If it does not
1249 * the filesystem may be damaged and we return a dummy inode.
1250 */
1260 localization);
1265 }
1271 }
1276 }
1279 }
1280 done:
1284 }
1286 /*
1287 * hammer_vop_nlookupdotdot { dvp, vpp, cred }
1288 *
1289 * Locate the parent directory of a directory vnode.
1290 *
1291 * dvp is referenced but not locked. *vpp must be returned referenced and
1292 * locked. A parent_obj_id of 0 indicates that we are at the root.
1293 *
1294 * NOTE: as-of sequences are not linked into the directory structure. If
1295 * we are at the root with a different asof then the mount point, reload
1296 * the same directory with the mount point's asof. I'm not sure what this
1297 * will do to NFS. We encode ASOF stamps in NFS file handles so it might not
1298 * get confused, but it hasn't been tested.
1299 */
1300 static
1301 int
1303 {
1305 hammer_inode_t dip;
1306 hammer_inode_t ip;
1307 hammer_mount_t hmp;
1310 hammer_tid_t asof;
1317 /*
1318 * Whos are parent? This could be the root of a pseudo-filesystem
1319 * whos parent is in another localization domain.
1320 */
1325 else
1328 /*
1329 * It's probably a PFS root when dip->ino_data.parent_obj_id is 0.
1330 */
1343 }
1344 }
1357 }
1361 }
1363 /*
1364 * hammer_vop_nlink { nch, dvp, vp, cred }
1365 */
1366 static
1367 int
1369 {
1371 hammer_inode_t dip;
1372 hammer_inode_t ip;
1374 hammer_mount_t hmp;
1395 /*
1396 * Create a transaction to cover the operations we perform.
1397 */
1402 /*
1403 * Add the filesystem object to the directory. Note that neither
1404 * dip nor ip are referenced or locked, but their vnodes are
1405 * referenced. This function will bump the inode's link count.
1406 */
1409 ip);
1411 /*
1412 * Finish up.
1413 */
1417 }
1423 }
1425 /*
1426 * hammer_vop_nmkdir { nch, dvp, vpp, cred, vap }
1427 *
1428 * The operating system has already ensured that the directory entry
1429 * does not exist and done all appropriate namespace locking.
1430 */
1431 static
1432 int
1434 {
1436 hammer_inode_t dip;
1437 hammer_inode_t nip;
1439 hammer_mount_t hmp;
1451 /*
1452 * Create a transaction to cover the operations we perform.
1453 */
1458 /*
1459 * Create a new filesystem object of the requested type. The
1460 * returned inode will be referenced but not locked.
1461 */
1470 }
1471 /*
1472 * Add the new filesystem object to the directory. This will also
1473 * bump the inode's link count.
1474 */
1477 nip);
1481 /*
1482 * Finish up.
1483 */
1493 }
1494 }
1500 }
1502 /*
1503 * hammer_vop_nmknod { nch, dvp, vpp, cred, vap }
1504 *
1505 * The operating system has already ensured that the directory entry
1506 * does not exist and done all appropriate namespace locking.
1507 */
1508 static
1509 int
1511 {
1513 hammer_inode_t dip;
1514 hammer_inode_t nip;
1516 hammer_mount_t hmp;
1528 /*
1529 * Create a transaction to cover the operations we perform.
1530 */
1535 /*
1536 * Create a new filesystem object of the requested type. The
1537 * returned inode will be referenced but not locked.
1538 *
1539 * If mknod specifies a directory a pseudo-fs is created.
1540 */
1549 }
1551 /*
1552 * Add the new filesystem object to the directory. This will also
1553 * bump the inode's link count.
1554 */
1557 nip);
1559 /*
1560 * Finish up.
1561 */
1571 }
1572 }
1578 }
1580 /*
1581 * hammer_vop_open { vp, mode, cred, fp }
1582 *
1583 * MPSAFE (does not require fs_token)
1584 */
1585 static
1586 int
1588 {
1589 hammer_inode_t ip;
1597 }
1599 /*
1600 * hammer_vop_print { vp }
1601 */
1602 static
1603 int
1605 {
1607 }
1609 /*
1610 * hammer_vop_readdir { vp, uio, cred, *eofflag, *ncookies, off_t **cookies }
1611 */
1612 static
1613 int
1615 {
1618 hammer_inode_t ip;
1619 hammer_mount_t hmp;
1621 hammer_base_elm_t base;
1626 off_t saveoff;
1646 }
1651 /*
1652 * Handle artificial entries
1653 *
1654 * It should be noted that the minimum value for a directory
1655 * hash key on-media is 0x0000000100000000, so we can use anything
1656 * less then that to represent our 'special' key space.
1657 */
1669 }
1678 }
1687 }
1689 /*
1690 * Key range (begin and end inclusive) to scan. Directory keys
1691 * directly translate to a 64 bit 'seek' position.
1692 */
1724 dtype,
1736 }
1739 done:
1752 }
1759 }
1760 }
1763 }
1765 /*
1766 * hammer_vop_readlink { vp, uio, cred }
1767 */
1768 static
1769 int
1771 {
1774 hammer_inode_t ip;
1775 hammer_mount_t hmp;
1778 hammer_pseudofs_inmem_t pfsm;
1786 /*
1787 * Shortcut if the symlink data was stuffed into ino_data.
1788 *
1789 * Also expand special "@@PFS%05d" softlinks (expansion only
1790 * occurs for non-historical (current) accesses made from the
1791 * primary filesystem).
1792 *
1793 * Note that userspace hammer command does not allow users to
1794 * create a @@PFS softlink under an existing other PFS (id!=0)
1795 * so the ip localization here for @@PFS softlink is always 0.
1796 */
1815 /* vap->va_size == 26 */
1821 /* vap->va_size == 10 */
1825 }
1828 }
1832 }
1836 }
1838 /*
1839 * Long version
1840 */
1845 /*
1846 * Key range (begin and end inclusive) to scan. Directory keys
1847 * directly translate to a 64 bit 'seek' position.
1848 */
1850 HAMMER_LOCALIZE_MISC;
1865 HAMMER_SYMLINK_NAME_OFF);
1868 HAMMER_SYMLINK_NAME_OFF,
1870 }
1871 }
1876 }
1878 /*
1879 * hammer_vop_nremove { nch, dvp, cred }
1880 */
1881 static
1882 int
1884 {
1886 hammer_inode_t dip;
1887 hammer_mount_t hmp;
1896 }
1907 }
1909 /*
1910 * hammer_vop_nrename { fnch, tnch, fdvp, tdvp, cred }
1911 */
1912 static
1913 int
1915 {
1919 hammer_inode_t fdip;
1920 hammer_inode_t tdip;
1921 hammer_inode_t ip;
1922 hammer_mount_t hmp;
1960 /*
1961 * Remove tncp from the target directory and then link ip as
1962 * tncp. XXX pass trans to dounlink
1963 *
1964 * Force the inode sync-time to match the transaction so it is
1965 * in-sync with the creation of the target directory entry.
1966 */
1972 ip);
1977 }
1978 }
1982 /*
1983 * Locate the record in the originating directory and remove it.
1984 *
1985 * Calculate the namekey and setup the key range for the scan. This
1986 * works kinda like a chained hash table where the lower 32 bits
1987 * of the namekey synthesize the chain.
1988 *
1989 * The key range is inclusive of both key_beg and key_end.
1990 */
1993 retry:
2009 /*
2010 * Scan all matching records (the chain), locate the one matching
2011 * the requested path component.
2012 *
2013 * The hammer_ip_*() functions merge in-memory records with on-disk
2014 * records for the purposes of the search.
2015 */
2025 }
2027 }
2029 /*
2030 * If all is ok we have to get the inode so we can adjust nlinks.
2031 *
2032 * WARNING: hammer_ip_del_direntry() may have to terminate the
2033 * cursor to avoid a recursion. It's ok to call hammer_done_cursor()
2034 * twice.
2035 */
2039 /*
2040 * XXX A deadlock here will break rename's atomicy for the purposes
2041 * of crash recovery.
2042 */
2046 }
2048 /*
2049 * Cleanup and tell the kernel that the rename succeeded.
2050 *
2051 * NOTE: ip->vp, if non-NULL, cannot be directly referenced
2052 * without formally acquiring the vp since the vp might
2053 * have zero refs on it, or in the middle of a reclaim,
2054 * etc.
2055 */
2070 }
2072 }
2073 }
2075 failed:
2079 }
2081 /*
2082 * hammer_vop_nrmdir { nch, dvp, cred }
2083 */
2084 static
2085 int
2087 {
2089 hammer_inode_t dip;
2090 hammer_mount_t hmp;
2099 }
2110 }
2112 /*
2113 * hammer_vop_markatime { vp, cred }
2114 */
2115 static
2116 int
2118 {
2120 hammer_inode_t ip;
2121 hammer_mount_t hmp;
2141 }
2143 /*
2144 * hammer_vop_setattr { vp, vap, cred }
2145 */
2146 static
2147 int
2149 {
2151 hammer_inode_t ip;
2153 hammer_mount_t hmp;
2159 #if 0
2161 #endif
2177 }
2195 }
2199 }
2200 }
2202 }
2206 }
2211 uuid_t uuid_uid;
2212 uuid_t uuid_gid;
2230 }
2232 }
2233 }
2240 /*
2241 * Log the operation if in fast-fsync mode or if
2242 * there are unterminated redo write records present.
2243 *
2244 * The second check is needed so the recovery code
2245 * properly truncates write redos even if nominal
2246 * REDO operations is turned off due to excessive
2247 * writes, because the related records might be
2248 * destroyed and never lay down a TERM_WRITE.
2249 */
2254 HAMMER_REDO_TRUNC,
2256 }
2259 /*
2260 * XXX break atomicy, we can deadlock the backend
2261 * if we do not release the lock. Probably not a
2262 * big deal here.
2263 */
2266 blksize,
2282 }
2285 /* XXX safe to use SDIRTY instead of DDIRTY here? */
2288 /*
2289 * On-media truncation is cached in the inode until
2290 * the inode is synchronized. We must immediately
2291 * handle any frontend records.
2292 */
2301 }
2302 }
2304 #if 0
2305 /*
2306 * When truncating, nvtruncbuf() may have cleaned out
2307 * a portion of the last block on-disk in the buffer
2308 * cache. We must clean out any frontend records
2309 * for blocks beyond the new last block.
2310 */
2316 }
2317 #endif
2326 }
2336 }
2338 }
2343 }
2348 }
2361 }
2362 }
2363 done:
2370 }
2372 /*
2373 * hammer_vop_nsymlink { nch, dvp, vpp, cred, vap, target }
2374 */
2375 static
2376 int
2378 {
2380 hammer_inode_t dip;
2381 hammer_inode_t nip;
2382 hammer_record_t record;
2384 hammer_mount_t hmp;
2399 /*
2400 * Create a transaction to cover the operations we perform.
2401 */
2406 /*
2407 * Create a new filesystem object of the requested type. The
2408 * returned inode will be referenced but not locked.
2409 */
2419 }
2421 /*
2422 * Add a record representing the symlink. symlink stores the link
2423 * as pure data, not a string, and is no \0 terminated.
2424 */
2435 HAMMER_LOCALIZE_MISC;
2442 }
2444 /*
2445 * Set the file size to the length of the link.
2446 */
2450 }
2451 }
2456 /*
2457 * Finish up.
2458 */
2469 }
2470 }
2474 }
2476 /*
2477 * hammer_vop_nwhiteout { nch, dvp, cred, flags }
2478 */
2479 static
2480 int
2482 {
2484 hammer_inode_t dip;
2485 hammer_mount_t hmp;
2494 }
2505 }
2507 /*
2508 * hammer_vop_ioctl { vp, command, data, fflag, cred }
2509 */
2510 static
2511 int
2513 {
2524 }
2526 static
2527 int
2529 {
2536 };
2537 hammer_mount_t hmp;
2554 else
2559 /*
2560 * Call standard mountctl VOP function
2561 * so we get user mount flags.
2562 */
2574 }
2581 }
2584 }
2586 /*
2587 * hammer_vop_strategy { vp, bio }
2588 *
2589 * Strategy call, used for regular file read & write only. Note that the
2590 * bp may represent a cluster.
2591 *
2592 * To simplify operation and allow better optimizations in the future,
2593 * this code does not make any assumptions with regards to buffer alignment
2594 * or size.
2595 */
2596 static
2597 int
2599 {
2617 }
2619 /* hammer_dump_dedup_cache(((hammer_inode_t)ap->a_vp->v_data)->hmp); */
2622 }
2624 /*
2625 * Read from a regular file. Iterate the related records and fill in the
2626 * BIO/BUF. Gaps are zero-filled.
2627 *
2628 * The support code in hammer_object.c should be used to deal with mixed
2629 * in-memory and on-disk records.
2630 *
2631 * NOTE: Can be called from the cluster code with an oversized buf.
2632 *
2633 * XXX atime update
2634 */
2635 static
2636 int
2638 {
2640 hammer_inode_t ip;
2641 hammer_inode_t dip;
2642 hammer_mount_t hmp;
2644 hammer_base_elm_t base;
2645 hammer_off_t disk_offset;
2663 /*
2664 * The zone-2 disk offset may have been set by the cluster code via
2665 * a BMAP operation, or else should be NOOFFSET.
2666 *
2667 * Checking the high bits for a match against zone-2 should suffice.
2668 *
2669 * In cases where a lot of data duplication is present it may be
2670 * more beneficial to drop through and doubule-buffer through the
2671 * device.
2672 */
2680 }
2682 /*
2683 * Try to shortcut requests for double_buffer mode too.
2684 * Since this mode runs through the device buffer cache
2685 * only compatible buffer sizes (meaning those generated
2686 * by normal filesystem buffers) are legal.
2687 */
2693 }
2694 }
2696 /*
2697 * Well, that sucked. Do it the hard way. If all the stars are
2698 * aligned we may still be able to issue a direct-read.
2699 */
2704 /*
2705 * Key range (begin and end inclusive) to scan. Note that the key's
2706 * stored in the actual records represent BASE+LEN, not BASE. The
2707 * first record containing bio_offset will have a key > bio_offset.
2708 */
2710 HAMMER_LOCALIZE_MISC;
2721 #if 0
2727 #endif
2728 {
2735 else
2737 }
2740 /*
2741 * Set NOSWAPCACHE for cursor data extraction if double buffering
2742 * is disabled or (if the file is not marked cacheable via chflags
2743 * and vm.swapcache_use_chflags is enabled).
2744 */
2747 vm_swapcache_use_chflags)) {
2749 }
2755 /*
2756 * Get the base file offset of the record. The key for
2757 * data records is (base + bytes) rather then (base).
2758 */
2762 /*
2763 * Calculate the gap, if any, and zero-fill it.
2764 *
2765 * n is the offset of the start of the record verses our
2766 * current seek offset in the bio.
2767 */
2775 }
2777 /*
2778 * Calculate the data offset in the record and the number
2779 * of bytes we can copy.
2780 *
2781 * There are two degenerate cases. First, boff may already
2782 * be at bp->b_bufsize. Secondly, the data offset within
2783 * the record may exceed the record's size.
2784 */
2793 }
2795 /*
2796 * Deal with cached truncations. This cool bit of code
2797 * allows truncate()/ftruncate() to avoid having to sync
2798 * the file.
2799 *
2800 * If the frontend is truncated then all backend records are
2801 * subject to the frontend's truncation.
2802 *
2803 * If the backend is truncated then backend records on-disk
2804 * (but not in-memory) are subject to the backend's
2805 * truncation. In-memory records owned by the backend
2806 * represent data written after the truncation point on the
2807 * backend and must not be truncated.
2808 *
2809 * Truncate operations deal with frontend buffer cache
2810 * buffers and frontend-owned in-memory records synchronously.
2811 */
2819 }
2820 }
2827 }
2828 }
2830 /*
2831 * Try to issue a direct read into our bio if possible,
2832 * otherwise resolve the element data into a hammer_buffer
2833 * and copy.
2834 *
2835 * The buffer on-disk should be zerod past any real
2836 * truncation point, but may not be for any synthesized
2837 * truncation point from above.
2838 *
2839 * NOTE: disk_offset is only valid if the cursor data is
2840 * on-disk.
2841 */
2848 /*
2849 * Direct read case
2850 */
2858 /*
2859 * Async I/O case for reading from backing store
2860 * and copying the data to the filesystem buffer.
2861 * live-dedup has to verify the data anyway if it
2862 * gets a hit later so we can just add the entry
2863 * now.
2864 */
2878 }
2879 }
2883 /*
2884 * We have to be sure that the only elements added to the
2885 * dedup cache are those which are already on-media.
2886 */
2890 /*
2891 * Iterate until we have filled the request.
2892 */
2897 }
2899 /*
2900 * There may have been a gap after the last record
2901 */
2907 /* boff = bp->b_bufsize; */
2908 }
2910 /*
2911 * Disallow swapcache operation on the vnode buffer if double
2912 * buffering is enabled, the swapcache will get the data via
2913 * the block device buffer.
2914 */
2918 /*
2919 * Cleanup
2920 */
2927 done:
2928 /*
2929 * Cache the b-tree node for the last data read in cache[1].
2930 *
2931 * If we hit the file EOF then also cache the node in the
2932 * governing directory's cache[3], it will be used to initialize
2933 * the new inode's cache[1] for any inodes looked up via the directory.
2934 *
2935 * This doesn't reduce disk accesses since the B-Tree chain is
2936 * likely cached, but it does reduce cpu overhead when looking
2937 * up file offsets for cpdup/tar/cpio style iterations.
2938 */
2947 }
2948 }
2953 }
2955 /*
2956 * BMAP operation - used to support cluster_read() only.
2957 *
2958 * (struct vnode *vp, off_t loffset, off_t *doffsetp, int *runp, int *runb)
2959 *
2960 * This routine may return EOPNOTSUPP if the opration is not supported for
2961 * the specified offset. The contents of the pointer arguments do not
2962 * need to be initialized in that case.
2963 *
2964 * If a disk address is available and properly aligned return 0 with
2965 * *doffsetp set to the zone-2 address, and *runp / *runb set appropriately
2966 * to the run-length relative to that offset. Callers may assume that
2967 * *doffsetp is valid if 0 is returned, even if *runp is not sufficiently
2968 * large, so return EOPNOTSUPP if it is not sufficiently large.
2969 */
2970 static
2971 int
2973 {
2975 hammer_inode_t ip;
2976 hammer_mount_t hmp;
2978 hammer_base_elm_t base;
2985 hammer_off_t last_disk_offset;
2986 hammer_off_t disk_offset;
2995 /*
2996 * We can only BMAP regular files. We can't BMAP database files,
2997 * directories, etc.
2998 */
3002 /*
3003 * bmap is typically called with runp/runb both NULL when used
3004 * for writing. We do not support BMAP for writing atm.
3005 */
3009 /*
3010 * Scan the B-Tree to acquire blockmap addresses, then translate
3011 * to raw addresses.
3012 */
3018 /*
3019 * Key range (begin and end inclusive) to scan. Note that the key's
3020 * stored in the actual records represent BASE+LEN, not BASE. The
3021 * first record containing bio_offset will have a key > bio_offset.
3022 */
3024 HAMMER_LOCALIZE_MISC;
3031 else
3047 else
3057 /*
3058 * Get the base file offset of the record. The key for
3059 * data records is (base + bytes) rather then (base).
3060 *
3061 * NOTE: rec_offset + rec_len may exceed the end-of-file.
3062 * The extra bytes should be zero on-disk and the BMAP op
3063 * should still be ok.
3064 */
3069 /*
3070 * Incorporate any cached truncation.
3071 *
3072 * NOTE: Modifications to rec_len based on synthesized
3073 * truncation points remove the guarantee that any extended
3074 * data on disk is zero (since the truncations may not have
3075 * taken place on-media yet).
3076 */
3084 }
3085 }
3092 }
3093 }
3095 /*
3096 * Accumulate information. If we have hit a discontiguous
3097 * block reset base_offset unless we are already beyond the
3098 * requested offset. If we are, that's it, we stop.
3099 */
3110 }
3116 }
3119 }
3128 /*
3129 * If we couldn't find any records or the records we did find were
3130 * all behind the requested offset, return failure. A forward
3131 * truncation can leave a hole w/ no on-disk records.
3132 */
3136 /*
3137 * Figure out the block size at the requested offset and adjust
3138 * our limits so the cluster_read() does not create inappropriately
3139 * sized buffer cache buffers.
3140 */
3144 }
3149 }
3151 /*
3152 * Returning EOPNOTSUPP simply prevents the direct-IO optimization
3153 * from occuring.
3154 */
3158 /*
3159 * Only large-data zones can be direct-IOd
3160 */
3164 /*
3165 * doffsetp is not aligned or the forward run size does
3166 * not cover a whole buffer, disallow the direct I/O.
3167 */
3170 /*
3171 * We're good.
3172 */
3177 }
3181 }
3183 }
3185 }
3187 /*
3188 * Write to a regular file. Because this is a strategy call the OS is
3189 * trying to actually get data onto the media.
3190 */
3191 static
3192 int
3194 {
3195 hammer_record_t record;
3196 hammer_mount_t hmp;
3197 hammer_inode_t ip;
3217 }
3221 /*
3222 * Disallow swapcache operation on the vnode buffer if double
3223 * buffering is enabled, the swapcache will get the data via
3224 * the block device buffer.
3225 */
3229 /*
3230 * Interlock with inode destruction (no in-kernel or directory
3231 * topology visibility). If we queue new IO while trying to
3232 * destroy the inode we can deadlock the vtrunc call in
3233 * hammer_inode_unloadable_check().
3234 *
3235 * Besides, there's no point flushing a bp associated with an
3236 * inode that is being destroyed on-media and has no kernel
3237 * references.
3238 */
3245 }
3247 /*
3248 * Reserve space and issue a direct-write from the front-end.
3249 * NOTE: The direct_io code will hammer_bread/bcopy smaller
3250 * allocations.
3251 *
3252 * An in-memory record will be installed to reference the storage
3253 * until the flusher can get to it.
3254 *
3255 * Since we own the high level bio the front-end will not try to
3256 * do a direct-read until the write completes.
3257 *
3258 * NOTE: The only time we do not reserve a full-sized buffers
3259 * worth of data is if the file is small. We do not try to
3260 * allocate a fragment (from the small-data zone) at the end of
3261 * an otherwise large file as this can lead to wildly separated
3262 * data.
3263 */
3268 else
3274 /*
3275 * B_VFSFLAG1 indicates that a REDO_WRITE entry was generated
3276 * in hammer_vop_write(). We must flag the record so the proper
3277 * REDO_TERM_WRITE entry is generated during the flush.
3278 */
3283 }
3290 }
3298 }
3301 }
3303 /*
3304 * dounlink - disconnect a directory entry
3305 *
3306 * XXX whiteout support not really in yet
3307 */
3308 static int
3312 {
3314 hammer_inode_t dip;
3315 hammer_inode_t ip;
3316 hammer_mount_t hmp;
3322 /*
3323 * Calculate the namekey and setup the key range for the scan. This
3324 * works kinda like a chained hash table where the lower 32 bits
3325 * of the namekey synthesize the chain.
3326 *
3327 * The key range is inclusive of both key_beg and key_end.
3328 */
3338 retry:
3354 /*
3355 * Scan all matching records (the chain), locate the one matching
3356 * the requested path component. info->last_error contains the
3357 * error code on search termination and could be 0, ENOENT, or
3358 * something else.
3359 *
3360 * The hammer_ip_*() functions merge in-memory records with on-disk
3361 * records for the purposes of the search.
3362 */
3374 }
3376 }
3378 /*
3379 * If all is ok we have to get the inode so we can adjust nlinks.
3380 * To avoid a deadlock with the flusher we must release the inode
3381 * lock on the directory when acquiring the inode for the entry.
3382 *
3383 * If the target is a directory, it must be empty.
3384 */
3399 }
3401 /*
3402 * If isdir >= 0 we validate that the entry is or is not a
3403 * directory. If isdir < 0 we don't care.
3404 */
3412 }
3413 }
3415 /*
3416 * If we are trying to remove a directory the directory must
3417 * be empty.
3418 *
3419 * The check directory code can loop and deadlock/retry. Our
3420 * own cursor's node locks must be released to avoid a 3-way
3421 * deadlock with the flusher if the check directory code
3422 * blocks.
3423 *
3424 * If any changes whatsoever have been made to the cursor
3425 * set EDEADLK and retry.
3426 *
3427 * WARNING: See warnings in hammer_unlock_cursor()
3428 * function.
3429 */
3431 HAMMER_OBJTYPE_DIRECTORY) {
3440 }
3441 }
3443 /*
3444 * Delete the directory entry.
3445 *
3446 * WARNING: hammer_ip_del_direntry() may have to terminate
3447 * the cursor to avoid a deadlock. It is ok to call
3448 * hammer_done_cursor() twice.
3449 */
3453 }
3456 /*
3457 * Tell the namecache that we are now unlinked.
3458 */
3461 /*
3462 * NOTE: ip->vp, if non-NULL, cannot be directly
3463 * referenced without formally acquiring the
3464 * vp since the vp might have zero refs on it,
3465 * or in the middle of a reclaim, etc.
3466 *
3467 * NOTE: The cache_setunresolved() can rip the vp
3468 * out from under us since the vp may not have
3469 * any refs, in which case ip->vp will be NULL
3470 * from the outset.
3471 */
3479 #if 0
3480 /*
3481 * Don't do this, it can deadlock
3482 * on concurrent rm's of hardlinks.
3483 * Shouldn't be needed any more.
3484 */
3486 #endif
3489 }
3491 }
3492 }
3497 }
3502 }
3504 /************************************************************************
3505 * FIFO AND SPECFS OPS *
3506 ************************************************************************
3507 *
3508 */
3509 static int
3511 {
3512 /* XXX update itimes */
3514 }
3516 static int
3518 {
3522 /* XXX update access time */
3524 }
3526 static int
3528 {
3532 /* XXX update access time */
3534 }
3536 static
3537 int
3539 {
3546 }
3548 /************************************************************************
3549 * KQFILTER OPS *
3550 ************************************************************************
3551 *
3552 */
3568 static
3569 int
3571 {
3587 }
3594 }
3596 static void
3598 {
3602 }
3604 static int
3606 {
3610 off_t off;
3615 }
3623 }
3625 static int
3627 {
3632 }
3634 static int
3636 {
3642 }
3644 }