| blob | 3b11f05351517bfef9353f5311e01cf845da9b2b |
1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
26 /*
27 * Copyright (c) 1983,1984,1985,1986,1987,1988,1989 AT&T.
28 * All rights reserved.
29 */
31 /*
32 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
33 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
34 */
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/systm.h>
39 #include <sys/cred.h>
40 #include <sys/time.h>
41 #include <sys/vnode.h>
42 #include <sys/vfs.h>
43 #include <sys/file.h>
44 #include <sys/filio.h>
45 #include <sys/uio.h>
46 #include <sys/buf.h>
47 #include <sys/mman.h>
48 #include <sys/pathname.h>
49 #include <sys/dirent.h>
50 #include <sys/debug.h>
51 #include <sys/vmsystm.h>
52 #include <sys/fcntl.h>
53 #include <sys/flock.h>
54 #include <sys/swap.h>
55 #include <sys/errno.h>
56 #include <sys/strsubr.h>
57 #include <sys/sysmacros.h>
58 #include <sys/kmem.h>
59 #include <sys/cmn_err.h>
60 #include <sys/pathconf.h>
61 #include <sys/utsname.h>
62 #include <sys/dnlc.h>
63 #include <sys/acl.h>
64 #include <sys/systeminfo.h>
65 #include <sys/atomic.h>
66 #include <sys/policy.h>
67 #include <sys/sdt.h>
68 #include <sys/zone.h>
70 #include <rpc/types.h>
71 #include <rpc/auth.h>
72 #include <rpc/clnt.h>
73 #include <rpc/rpc_rdma.h>
75 #include <nfs/nfs.h>
76 #include <nfs/nfs_clnt.h>
77 #include <nfs/rnode.h>
78 #include <nfs/nfs_acl.h>
79 #include <nfs/lm.h>
81 #include <vm/hat.h>
82 #include <vm/as.h>
83 #include <vm/page.h>
84 #include <vm/pvn.h>
85 #include <vm/seg.h>
86 #include <vm/seg_map.h>
87 #include <vm/seg_kpm.h>
88 #include <vm/seg_vn.h>
90 #include <sys/fs_subr.h>
92 #include <sys/ddi.h>
95 cred_t *);
97 stable_how *);
109 caller_context_t *);
118 cred_t *);
130 cred_t *);
132 cred_t *);
135 /*
136 * Error flags used to pass information about certain special errors
137 * which need to be handled specially.
138 */
139 #define NFS_EOF -98
140 #define NFS_VERF_MISMATCH -97
142 /* ALIGN64 aligns the given buffer and adjust buffer size to 64 bit */
143 #define ALIGN64(x, ptr, sz) \
144 x = ((uintptr_t)(ptr)) & (sizeof (uint64_t) - 1); \
145 if (x) { \
146 x = sizeof (uint64_t) - (x); \
147 sz -= (x); \
148 ptr += (x); \
149 }
151 /*
152 * These are the vnode ops routines which implement the vnode interface to
153 * the networked file system. These routines just take their parameters,
154 * make them look networkish by putting the right info into interface structs,
155 * and then calling the appropriate remote routine(s) to do the work.
156 *
157 * Note on directory name lookup cacheing: If we detect a stale fhandle,
158 * we purge the directory cache relative to that vnode. This way, the
159 * user won't get burned by the cache repeatedly. See <nfs/rnode.h> for
160 * more details on rnode locking.
161 */
165 caller_context_t *);
167 caller_context_t *);
169 caller_context_t *);
171 caller_context_t *);
173 caller_context_t *);
175 caller_context_t *);
178 caller_context_t *);
186 vsecattr_t *);
209 caller_context_t *);
222 caller_context_t *);
226 caller_context_t *);
228 caller_context_t *);
230 caller_context_t *);
232 caller_context_t *);
268 /* no separate nfs3_dump */
277 };
279 /*
280 * XXX: This is referenced in modstubs.s
281 */
284 {
286 }
288 /* ARGSUSED */
289 static int
291 {
305 }
308 /*
309 * If there is no cached data or if close-to-open
310 * consistency checking is turned off, we can avoid
311 * the over the wire getattr. Otherwise, if the
312 * file system is mounted readonly, then just verify
313 * the caches are up to date using the normal mechanism.
314 * Else, if the file is not mmap'd, then just mark
315 * the attributes as timed out. They will be refreshed
316 * and the caches validated prior to being used.
317 * Else, the file system is mounted writeable so
318 * force an over the wire GETATTR in order to ensure
319 * that all cached data is valid.
320 */
332 }
337 }
339 /* ARGSUSED */
340 static int
343 {
348 /*
349 * zone_enter(2) prevents processes from changing zones with NFS files
350 * open; if we happen to get here from the wrong zone we can't do
351 * anything over the wire.
352 */
354 /*
355 * We could attempt to clean up locks, except we're sure
356 * that the current process didn't acquire any locks on
357 * the file: any attempt to lock a file belong to another zone
358 * will fail, and one can't lock an NFS file and then change
359 * zones, as that fails too.
360 *
361 * Returning an error here is the sane thing to do. A
362 * subsequent call to VN_RELE() which translates to a
363 * nfs3_inactive() will clean up state: if the zone of the
364 * vnode's origin is still alive and kicking, an async worker
365 * thread will handle the request (from the correct zone), and
366 * everything (minus the commit and final nfs3_getattr_otw()
367 * call) should be OK. If the zone is going away
368 * nfs_async_inactive() will throw away cached pages inline.
369 */
371 }
373 /*
374 * If we are using local locking for this filesystem, then
375 * release all of the SYSV style record locks. Otherwise,
376 * we are doing network locking and we need to release all
377 * of the network locks. All of the locks held by this
378 * process on this file are released no matter what the
379 * incoming reference count is.
380 */
390 /*
391 * If the file has been `unlinked', then purge the
392 * DNLC so that this vnode will get reycled quicker
393 * and the .nfs* file on the server will get removed.
394 */
399 /*
400 * If the file was open for write and there are pages,
401 * then if the file system was mounted using the "no-close-
402 * to-open" semantics, then start an asynchronous flush
403 * of the all of the pages in the file.
404 * else the file system was not mounted using the "no-close-
405 * to-open" semantics, then do a synchronous flush and
406 * commit of all of the dirty and uncommitted pages.
407 *
408 * The asynchronous flush of the pages in the "nocto" path
409 * mostly just associates a cred pointer with the rnode so
410 * writes which happen later will have a better chance of
411 * working. It also starts the data being written to the
412 * server, but without unnecessarily delaying the application.
413 */
427 }
433 }
435 /*
436 * If RWRITEATTR is set, then issue an over the wire GETATTR to
437 * refresh the attribute cache with a set of attributes which
438 * weren't returned from a WRITE. This will enable the close-
439 * to-open processing to work.
440 */
445 }
447 /* ARGSUSED */
448 static int
450 {
452 READ3args args;
453 READ3uiores res;
455 offset_t offset;
456 ssize_t count;
459 failinfo_t fi;
486 }
506 }
518 sv_hostname);
520 }
528 }
533 }
535 /* ARGSUSED */
536 static int
539 {
541 uoff_t off;
542 offset_t diff;
545 caddr_t base;
546 uint_t flags;
567 /*
568 * Bypass VM if caching has been disabled (e.g., locking) or if
569 * using client-side direct I/O and the file is not mmap'd and
570 * there are no cached pages.
571 */
577 }
593 }
594 }
603 /*
604 * Copy data.
605 */
610 S_READ);
613 }
616 /*
617 * If read a whole block or read to eof,
618 * won't need this buffer again soon.
619 */
624 else
631 }
637 }
638 }
642 }
644 /* ARGSUSED */
645 static int
648 {
651 uoff_t off;
652 caddr_t base;
653 uint_t flags;
659 offset_t offset;
661 uint_t bsize;
677 /*
678 * Must serialize if appending.
679 */
685 }
692 }
702 /*
703 * Check to make sure that the process will not exceed
704 * its limit on file size. It is okay to write up to
705 * the limit, but not beyond. Thus, the write which
706 * reaches the limit will be short and the next write
707 * will return an error.
708 */
722 }
723 }
728 /*
729 * Bypass VM if caching has been disabled (e.g., locking) or if
730 * using client-side direct I/O and the file is not mmap'd and
731 * there are no cached pages.
732 */
739 uoff_t org_offset;
740 stable_how stab_comm;
742 nfs3_fwrite:
747 /*
748 * A close may have cleared r_error, if so,
749 * propagate ESTALE error return properly
750 */
754 }
760 else
770 }
774 }
789 /*
790 * A close may have cleared r_error, if so,
791 * propagate ESTALE error return properly
792 */
796 }
798 /*
799 * Don't create dirty pages faster than they
800 * can be cleaned so that the system doesn't
801 * get imbalanced. If the async queue is
802 * maxed out, then wait for it to drain before
803 * creating more dirty pages. Also, wait for
804 * any threads doing pagewalks in the vop_getattr
805 * entry points so that they don't block for
806 * long periods.
807 */
823 }
828 }
831 /*
832 * Touch the page and fault it in if it is not in core
833 * before segmap_getmapflt or vpm_data_copy can lock it.
834 * This is to avoid the deadlock if the buffer is mapped
835 * to the same file through mmap which we want to write.
836 */
840 /*
841 * It will use kpm mappings, so no need to
842 * pass an address.
843 */
861 pagecreate);
867 }
868 }
875 /*
876 * Have written a whole block.
877 * Start an asynchronous write
878 * and mark the buffer to
879 * indicate that it won't be
880 * needed again soon.
881 */
889 }
894 }
900 }
901 /*
902 * In the event that we got an access error while
903 * faulting in a page for a write-only file just
904 * force a write.
905 */
908 }
911 bottom:
921 }
923 /*
924 * Flags are composed of {B_ASYNC, B_INVAL, B_FREE, B_DONTNEED}
925 */
926 static int
929 {
933 uchar_t fsdata;
934 stable_how stab_comm;
940 /*
941 * pageio_setup should have set b_addr to 0. This
942 * is correct since we want to do I/O on a page
943 * boundary. bp_mapin will use this addr to calculate
944 * an offset, and then set b_addr to the kernel virtual
945 * address it allocated for us.
946 */
956 /*
957 * Calculate the desired level of stability to write data
958 * on the server and then mark all of the pages to reflect
959 * this.
960 */
968 }
980 /*
981 * If the server wrote pages in a more stable fashion than
982 * was requested, then clear all of the marks in the pages
983 * indicating that COMMIT operations were required.
984 */
989 }
992 }
994 /*
995 * Write to file. Writes to remote server in largest size
996 * chunks that the server can handle. Write is synchronous.
997 */
998 static int
1001 {
1003 WRITE3args args;
1004 WRITE3res res;
1027 else
1038 }
1050 }
1058 "nfs3write: server %s wrote %u, "
1063 }
1072 }
1073 }
1082 tsize;
1084 }
1091 /*
1092 * If the data was written UNSTABLE,
1093 * then might as well stop because
1094 * the whole block will have to get
1095 * rewritten anyway.
1096 */
1100 }
1101 }
1105 }
1106 /*
1107 * Mark the attribute cache as timed out and
1108 * set RWRITEATTR to indicate that the file
1109 * was modified with a WRITE operation and
1110 * that the attributes can not be trusted.
1111 */
1115 }
1119 }
1121 /*
1122 * Read from a file. Reads data in largest chunks our interface can handle.
1123 */
1124 static int
1127 {
1129 READ3args args;
1130 READ3vres res;
1134 failinfo_t fi;
1137 hrtime_t t;
1161 }
1168 else
1188 }
1202 }
1213 }
1226 }
1227 }
1230 }
1232 /* ARGSUSED */
1233 static int
1236 {
1245 }
1246 }
1248 /* ARGSUSED */
1249 static int
1252 {
1258 /*
1259 * If it has been specified that the return value will
1260 * just be used as a hint, and we are only being asked
1261 * for size, fsid or rdevid, then return the client's
1262 * notion of these values without checking to make sure
1263 * that the attribute cache is up to date.
1264 * The whole point is to avoid an over the wire GETATTR
1265 * call.
1266 */
1280 }
1281 }
1283 /*
1284 * Only need to flush pages if asking for the mtime
1285 * and if there any dirty pages or any outstanding
1286 * asynchronous (write) requests for this file.
1287 */
1299 }
1303 }
1304 }
1307 }
1309 /*ARGSUSED4*/
1310 static int
1313 {
1328 vp);
1338 }
1340 static int
1342 {
1344 uint_t mask;
1345 SETATTR3args args;
1346 SETATTR3res res;
1350 mode_t omode;
1352 hrtime_t t;
1359 /*
1360 * Only need to flush pages if there are any pages and
1361 * if the file is marked as dirty in some fashion. The
1362 * file must be flushed so that we can accurately
1363 * determine the size of the file and the cached data
1364 * after the SETATTR returns. A file is considered to
1365 * be dirty if it is either marked with RDIRTY, has
1366 * outstanding i/o's active, or is mmap'd. In this
1367 * last case, we can't tell whether there are dirty
1368 * pages, so we flush just to be sure.
1369 */
1381 }
1382 }
1385 /*
1386 * If the intent is for the server to set the times,
1387 * there is no point in have the mask indicating set mtime or
1388 * atime, because the vap values may be junk, and so result
1389 * in an overflow error. Remove these flags from the vap mask
1390 * before calling in this case, and restore them afterwards.
1391 */
1393 /* Use server times, so don't set the args time fields */
1399 }
1402 }
1404 /* Either do not set times or use the client specified times */
1406 }
1409 /* req time field(s) overflow - return immediately */
1411 }
1419 tryagain:
1436 /*
1437 * Purge the access cache and ACL cache if changing either the
1438 * owner of the file, the group owner, or the mode. These may
1439 * change the access permissions of the file, so purge old
1440 * information and start over again.
1441 */
1451 }
1452 }
1457 }
1461 /*
1462 * If changing the size of the file, invalidate
1463 * any local cached data which is no longer part
1464 * of the file. We also possibly invalidate the
1465 * last page in the file. We could use
1466 * pvn_vpzero(), but this would mark the page as
1467 * modified and require it to be written back to
1468 * the server for no particularly good reason.
1469 * This way, if we access it, then we bring it
1470 * back in. A read should be cheaper than a
1471 * write.
1472 */
1476 }
1478 /*
1479 * Some servers will change the mode to clear the setuid
1480 * and setgid bits when changing the uid or gid. The
1481 * client needs to compensate appropriately.
1482 */
1494 else
1497 }
1498 }
1501 /*
1502 * If we got back a "not synchronized" error, then
1503 * we need to retry with a new guard value. The
1504 * guard value used is the change time. If the
1505 * server returned post_op_attr, then we can just
1506 * retry because we have the latest attributes.
1507 * Otherwise, we issue a GETATTR to get the latest
1508 * attributes and then retry. If we couldn't get
1509 * the attributes this way either, then we give
1510 * up because we can't complete the operation as
1511 * required.
1512 */
1517 }
1519 }
1522 }
1524 static int
1526 {
1529 }
1531 /* ARGSUSED */
1532 static int
1534 {
1536 ACCESS3args args;
1537 ACCESS3res res;
1539 uint32 acc;
1542 failinfo_t fi;
1543 nfs_access_type_t cacc;
1544 hrtime_t t;
1557 }
1561 else
1563 }
1572 ACCESS3_EXECUTE;
1573 }
1581 /*
1582 * ncr and ncrfree both initially
1583 * point to the memory area returned
1584 * by crnetadjust();
1585 * ncrfree not NULL when exiting means
1586 * that we need to release it
1587 */
1590 tryagain:
1597 }
1599 /*
1600 * If the cred can be adjusted, try again
1601 * with the new cred.
1602 */
1607 }
1611 }
1612 }
1627 }
1633 /*
1634 * we just cached results with cred; if cred is the
1635 * adjusted credentials from crnetadjust, we do not want
1636 * to release them before exiting: hence setting ncrfree
1637 * to NULL
1638 */
1642 /*
1643 * If the cred can be adjusted, try again
1644 * with the new cred.
1645 */
1650 }
1652 }
1656 }
1662 }
1666 /* ARGSUSED */
1667 static int
1669 {
1671 READLINK3args args;
1672 READLINK3res res;
1673 nfspath3 resdata_backup;
1677 failinfo_t fi;
1678 hrtime_t t;
1680 /*
1681 * Can't readlink anything other than a symbolic link.
1682 */
1699 }
1701 }
1729 }
1734 cr);
1748 }
1751 }
1759 }
1761 /*
1762 * The over the wire error for attempting to readlink something
1763 * other than a symbolic link is ENXIO. However, we need to
1764 * return EINVAL instead of ENXIO, so we map it here.
1765 */
1767 }
1769 /*
1770 * Flush local dirty pages to stable storage on the server.
1771 *
1772 * If FNODSYNC is specified, then there is nothing to do because
1773 * metadata changes are not cached on the client before being
1774 * sent to the server.
1775 */
1776 /* ARGSUSED */
1777 static int
1779 {
1791 }
1793 /*
1794 * Weirdness: if the file was removed or the target of a rename
1795 * operation while it was open, it got renamed instead. Here we
1796 * remove the renamed file.
1797 */
1798 /* ARGSUSED */
1799 static void
1801 {
1806 /*
1807 * If this is coming from the wrong zone, we let someone in the right
1808 * zone take care of it asynchronously. We can get here due to
1809 * VN_RELE() being called from pageout() or fsflush(). This call may
1810 * potentially turn into an expensive no-op if, for instance, v_count
1811 * gets incremented in the meantime, but it's still correct.
1812 */
1816 }
1819 redo:
1821 /*
1822 * Save the vnode pointer for the directory where the
1823 * unlinked-open file got renamed, then set it to NULL
1824 * to prevent another thread from getting here before
1825 * we're done with the remove. While we have the
1826 * statelock, make local copies of the pertinent rnode
1827 * fields. If we weren't to do this in an atomic way, the
1828 * the unl* fields could become inconsistent with respect
1829 * to each other due to a race condition between this
1830 * code and nfs_remove(). See bug report 1034328.
1831 */
1837 REMOVE3args args;
1838 REMOVE3res res;
1841 hrtime_t t;
1851 /*
1852 * If there are any dirty pages left, then flush
1853 * them. This is unfortunate because they just
1854 * may get thrown away during the remove operation,
1855 * but we have to do this for correctness.
1856 */
1867 }
1868 }
1870 /*
1871 * Do the remove operation on the renamed file
1872 */
1897 }
1898 }
1900 /*
1901 * Release stuff held for the remove
1902 */
1907 }
1909 }
1912 }
1914 /*
1915 * Remote file system operations having to do with directory manipulation.
1916 */
1918 /* ARGSUSED */
1919 static int
1923 {
1934 /*
1935 * Are we looking up extended attributes? If so, "dvp" is
1936 * the file or directory for which we want attributes, and
1937 * we need a lookup of the hidden attribute directory
1938 * before we lookup the rest of the path.
1939 */
1954 else
1963 }
1966 }
1971 }
1977 /*
1978 * If vnode is a device, create special vnode.
1979 */
1984 }
1986 out:
1991 }
1995 #ifdef DEBUG
2001 #endif
2003 /* ARGSUSED */
2004 int
2007 {
2012 /*
2013 * If lookup is for "", just return dvp. Don't need
2014 * to send it over the wire, look it up in the dnlc,
2015 * or perform any access checks.
2016 */
2021 }
2023 /*
2024 * Can't do lookups in non-directories.
2025 */
2029 /*
2030 * If we're called with RFSCALL_SOFT, it's important that
2031 * the only rfscall is one we make directly; if we permit
2032 * an access call because we're looking up "." or validating
2033 * a dnlc hit, we'll deadlock because that rfscall will not
2034 * have the RFSCALL_SOFT set.
2035 */
2039 /*
2040 * If lookup is for ".", just return dvp. Don't need
2041 * to send it over the wire or look it up in the dnlc,
2042 * just need to check access.
2043 */
2051 }
2058 }
2060 /*
2061 * Lookup this name in the DNLC. If there was a valid entry,
2062 * then return the results of the lookup.
2063 */
2068 callit:
2072 }
2074 static int
2076 {
2082 /*
2083 * Lookup this name in the DNLC. If successful, then validate
2084 * the caches and then recheck the DNLC. The DNLC is rechecked
2085 * just in case this entry got invalidated during the call
2086 * to nfs3_validate_caches.
2087 *
2088 * An assumption is being made that it is safe to say that a
2089 * file exists which may not on the server. Any operations to
2090 * the server will fail with ESTALE.
2091 */
2092 #ifdef DEBUG
2093 nfs3_lookup_dnlc_lookups++;
2094 #endif
2100 }
2110 }
2113 #ifdef DEBUG
2114 nfs3_lookup_dnlc_neg_hits++;
2115 #endif
2117 }
2119 #ifdef DEBUG
2120 nfs3_lookup_dnlc_hits++;
2121 #endif
2123 }
2124 #ifdef DEBUG
2125 nfs3_lookup_dnlc_disappears++;
2126 #endif
2127 }
2128 #ifdef DEBUG
2129 else
2130 nfs3_lookup_dnlc_misses++;
2131 #endif
2136 }
2138 static int
2141 {
2143 LOOKUP3args args;
2144 LOOKUP3vres res;
2149 failinfo_t fi;
2150 hrtime_t t;
2188 }
2202 }
2204 }
2205 }
2213 }
2215 #ifdef DEBUG
2217 #endif
2219 /* ARGSUSED */
2220 static int
2224 {
2238 top:
2239 /*
2240 * We make a copy of the attributes because the caller does not
2241 * expect us to change what va points to.
2242 */
2245 /*
2246 * If the pathname is "", just use dvp. Don't need
2247 * to send it over the wire, look it up in the dnlc,
2248 * or perform any access checks.
2249 */
2254 /*
2255 * If the pathname is ".", just use dvp. Don't need
2256 * to send it over the wire or look it up in the dnlc,
2257 * just need to check access.
2258 */
2264 }
2267 /*
2268 * We need to go over the wire, just to be sure whether the
2269 * file exists or not. Using the DNLC can be dangerous in
2270 * this case when making a decision regarding existence.
2271 */
2274 }
2281 /*
2282 * If vnode is a device, create special vnode.
2283 */
2288 }
2295 /*
2296 * Existing file was truncated; emit a
2297 * create event.
2298 */
2300 }
2301 }
2302 }
2308 }
2311 }
2315 /*
2316 * Decide what the group-id of the created file should be.
2317 * Set it in attribute list as advisory...
2318 */
2323 }
2332 }
2334 /*
2335 * If this is not an exclusive create, then the CREATE
2336 * request will be made with the GUARDED mode set. This
2337 * means that the server will return EEXIST if the file
2338 * exists. The file could exist because of a retransmitted
2339 * request. In this case, we recover by starting over and
2340 * checking to see whether the file exists. This second
2341 * time through it should and a CREATE request will not be
2342 * sent.
2343 *
2344 * This handles the problem of a dangling CREATE request
2345 * which contains attributes which indicate that the file
2346 * should be truncated. This retransmitted request could
2347 * possibly truncate valid data in the file if not caught
2348 * by the duplicate request mechanism on the server or if
2349 * not caught by other means. The scenario is:
2350 *
2351 * Client transmits CREATE request with size = 0
2352 * Client times out, retransmits request.
2353 * Response to the first request arrives from the server
2354 * and the client proceeds on.
2355 * Client writes data to the file.
2356 * The server now processes retransmitted CREATE request
2357 * and truncates file.
2358 *
2359 * The use of the GUARDED CREATE request prevents this from
2360 * happening because the retransmitted CREATE would fail
2361 * with EEXIST and would not truncate the file.
2362 */
2364 #ifdef DEBUG
2365 nfs3_create_misses++;
2366 #endif
2368 }
2371 }
2375 }
2377 /* ARGSUSED */
2378 static int
2381 {
2383 CREATE3args args;
2384 CREATE3res res;
2390 timestruc_t now;
2391 hrtime_t t;
2397 /*
2398 * Construct the create verifier. This verifier needs
2399 * to be unique between different clients. It also needs
2400 * to vary for each exclusive create request generated
2401 * from the client to the server.
2402 *
2403 * The first attempt is made to use the hostid and a
2404 * unique number on the client. If the hostid has not
2405 * been set, the high resolution time that the exclusive
2406 * create request is being made is used. This will work
2407 * unless two different clients, both with the hostid
2408 * not set, attempt an exclusive create request on the
2409 * same file, at exactly the same clock time. The
2410 * chances of this happening seem small enough to be
2411 * reasonable.
2412 */
2421 }
2422 /*
2423 * Since the server will use this value for the mtime,
2424 * make sure that it can't overflow. Zero out the MSB.
2425 * The actual value does not matter here, only its uniqeness.
2426 */
2429 /*
2430 * Issue the non-exclusive create in guarded mode. This
2431 * may result in some false EEXIST responses for
2432 * retransmitted requests, but these will be handled at
2433 * a higher level. By using GUARDED, duplicate requests
2434 * to do file truncation and possible access problems
2435 * can be avoided.
2436 */
2441 /* req time field(s) overflow - return immediately */
2443 }
2444 }
2458 }
2466 /*
2467 * On exclusive create the times need to be explicitly
2468 * set to clear any potential verifier that may be stored
2469 * in one of these fields (see comment below). This
2470 * is done here to cover the case where no post op attrs
2471 * were returned or a 'invalid' time was returned in
2472 * the attributes.
2473 */
2490 /*
2491 * On an exclusive create, it is possible
2492 * that attributes were returned but those
2493 * postop attributes failed to decode
2494 * properly. If this is the case,
2495 * then most likely the atime or mtime
2496 * were invalid for our client; this
2497 * is caused by the server storing the
2498 * create verifier in one of the time
2499 * fields(most likely mtime).
2500 * So... we are going to setattr just the
2501 * atime/mtime to clear things up.
2502 */
2507 /*
2508 * Setting the times failed.
2509 * Remove the file and return
2510 * the error.
2511 */
2516 }
2517 }
2519 /*
2520 * This handles the non-exclusive case
2521 * and the exclusive case where no post op
2522 * attrs were returned.
2523 */
2530 }
2532 }
2533 }
2535 }
2539 /*
2540 * If doing an exclusive create, then generate
2541 * a SETATTR to set the initial attributes.
2542 * Try to set the mtime and the atime to the
2543 * server's current time. It is somewhat
2544 * expected that these fields will be used to
2545 * store the exclusive create cookie. If not,
2546 * server implementors will need to know that
2547 * a SETATTR will follow an exclusive create
2548 * and the cookie should be destroyed if
2549 * appropriate. This work may have been done
2550 * earlier in this function if post op attrs
2551 * were not available.
2552 *
2553 * The VATTR_GID and VATTR_SIZE bits are turned off
2554 * so that the SETATTR request will not attempt
2555 * to process these. The gid will be set
2556 * separately if appropriate. The size is turned
2557 * off because it is assumed that a new file will
2558 * be created empty and if the file wasn't empty,
2559 * then the exclusive create will have failed
2560 * because the file must have existed already.
2561 * Therefore, no truncate operation is needed.
2562 */
2566 /*
2567 * Couldn't correct the attributes of
2568 * the newly created file and the
2569 * attributes are wrong. Remove the
2570 * file and return an error to the
2571 * application.
2572 */
2576 }
2577 }
2581 /*
2582 * If the gid on the file isn't right, then
2583 * generate a SETATTR to attempt to change
2584 * it. This may or may not work, depending
2585 * upon the server's semantics for allowing
2586 * file ownership changes.
2587 */
2590 }
2592 /*
2593 * If vnode is a device create special vnode
2594 */
2603 }
2606 }
2608 /*
2609 * Special setattr function to take care of rest of atime/mtime
2610 * after successful exclusive create. This function exists to avoid
2611 * handling attributes from the server; exclusive the atime/mtime fields
2612 * may be 'invalid' in client's view and therefore can not be trusted.
2613 */
2614 static int
2616 {
2618 uint_t mask;
2619 SETATTR3args args;
2620 SETATTR3res res;
2623 hrtime_t t;
2626 /* save the caller's mask so that it can be reset later */
2634 /* Use the mask to initialize the arguments */
2638 /* We want to set just atime/mtime on this request */
2654 }
2658 /*
2659 * It is important to pick up the attributes.
2660 * Since this is the exclusive create path, the
2661 * attributes on the initial create were ignored
2662 * and we need these to have the correct info.
2663 */
2665 /*
2666 * No need to do the atime/mtime work again so clear
2667 * the bits.
2668 */
2672 }
2677 }
2679 /* ARGSUSED */
2680 static int
2683 {
2685 MKNOD3args args;
2686 MKNOD3res res;
2690 hrtime_t t;
2701 /* req time field(s) overflow - return immediately */
2703 }
2717 /* req time field(s) overflow - return immediately */
2719 }
2724 }
2738 }
2764 }
2766 }
2768 }
2770 }
2775 }
2777 /*
2778 * If vnode is a device create special vnode
2779 */
2788 }
2790 }
2792 /*
2793 * Weirdness: if the vnode to be removed is open
2794 * we rename it instead of removing it and nfs_inactive
2795 * will remove the new name.
2796 */
2797 /* ARGSUSED */
2798 static int
2800 {
2802 REMOVE3args args;
2803 REMOVE3res res;
2809 hrtime_t t;
2821 }
2827 }
2829 /*
2830 * First just remove the entry from the name cache, as it
2831 * is most likely the only entry for this vp.
2832 */
2835 /*
2836 * If the file has a v_count > 1 then there may be more than one
2837 * entry in the name cache due multiple links or an open file,
2838 * but we don't have the real reference count so flush all
2839 * possible entries.
2840 */
2844 /*
2845 * Now we have the real reference count on the vnode
2846 */
2869 }
2871 }
2874 /*
2875 * We need to flush any dirty pages which happen to
2876 * be hanging around before removing the file. This
2877 * shouldn't happen very often and mostly on file
2878 * systems mounted "nocto".
2879 */
2888 }
2889 }
2902 /*
2903 * The xattr dir may be gone after last attr is removed,
2904 * so flush it from dnlc.
2905 */
2917 cr);
2924 }
2925 }
2926 }
2930 }
2936 }
2938 /* ARGSUSED */
2939 static int
2942 {
2944 LINK3args args;
2945 LINK3res res;
2950 hrtime_t t;
2985 }
2997 cr);
3003 }
3004 }
3009 /*
3010 * Notify the source file of this link operation.
3011 */
3013 }
3015 }
3017 /* ARGSUSED */
3018 static int
3021 {
3030 }
3032 /*
3033 * nfs3rename does the real work of renaming in NFS Version 3.
3034 */
3035 static int
3038 {
3040 RENAME3args args;
3041 RENAME3res res;
3049 hrtime_t t;
3065 }
3072 }
3073 }
3075 /*
3076 * Lookup the target file. If it exists, it needs to be
3077 * checked to see whether it is a mount point and whether
3078 * it is active (open).
3079 */
3082 /*
3083 * If this file has been mounted on, then just
3084 * return busy because renaming to it would remove
3085 * the mounted file system from the name space.
3086 */
3092 }
3094 /*
3095 * Purge the name cache of all references to this vnode
3096 * so that we can check the reference count to infer
3097 * whether it is active or not.
3098 */
3099 /*
3100 * First just remove the entry from the name cache, as it
3101 * is most likely the only entry for this vp.
3102 */
3104 /*
3105 * If the file has a v_count > 1 then there may be more
3106 * than one entry in the name cache due multiple links
3107 * or an open file, but we don't have the real reference
3108 * count so flush all possible entries.
3109 */
3113 /*
3114 * If the vnode is active and is not a directory,
3115 * arrange to rename it to a
3116 * temporary file so that it will continue to be
3117 * accessible. This implements the "unlink-open-file"
3118 * semantics for the target of a rename operation.
3119 * Before doing this though, make sure that the
3120 * source and target files are not already the same.
3121 */
3123 /*
3124 * Lookup the source name.
3125 */
3129 /*
3130 * The source name *should* already exist.
3131 */
3137 }
3139 /*
3140 * Compare the two vnodes. If they are the same,
3141 * just release all held vnodes and return success.
3142 */
3149 }
3151 /*
3152 * Can't mix and match directories and non-
3153 * directories in rename operations. We already
3154 * know that the target is not a directory. If
3155 * the source is a directory, return an error.
3156 */
3163 }
3165 /*
3166 * The target file exists, is not the same as
3167 * the source file, and is active. Link it
3168 * to a temporary filename to avoid having
3169 * the server removing the file completely.
3170 */
3176 }
3184 }
3198 }
3200 }
3201 }
3204 /*
3205 * When renaming directories to be a subdirectory of a
3206 * different parent, the dnlc entry for ".." will no
3207 * longer be valid, so it must be removed.
3208 *
3209 * We do a lookup here to determine whether we are renaming
3210 * a directory and we need to check if we are renaming
3211 * an unlinked file. This might have already been done
3212 * in previous code, so we check ovp == NULL to avoid
3213 * doing it twice.
3214 */
3217 /*
3218 * The source name *should* already exist.
3219 */
3225 }
3227 }
3229 }
3254 }
3256 }
3268 }
3269 /*
3270 * when renaming directories to be a subdirectory of a
3271 * different parent, the dnlc entry for ".." will no
3272 * longer be valid, so it must be removed
3273 */
3280 }
3281 }
3283 /*
3284 * If we are renaming the unlinked file, update the
3285 * r_unldvp and r_unlname as needed.
3286 */
3297 }
3298 }
3299 }
3305 cr);
3306 }
3307 /*
3308 * System V defines rename to return EEXIST, not
3309 * ENOTEMPTY if the target directory is not empty.
3310 * Over the wire, the error is NFSERR_ENOTEMPTY
3311 * which geterrno maps to ENOTEMPTY.
3312 */
3315 }
3325 }
3329 }
3336 }
3338 /* ARGSUSED */
3339 static int
3342 {
3344 MKDIR3args args;
3345 MKDIR3res res;
3350 hrtime_t t;
3356 /*
3357 * Decide what the group-id and set-gid bit of the created directory
3358 * should be. May have to do a setattr to get the gid right.
3359 */
3370 /* req time field(s) overflow - return immediately */
3372 }
3393 }
3406 }
3422 }
3424 }
3425 }
3427 }
3431 }
3436 }
3441 }
3443 /* ARGSUSED */
3444 static int
3447 {
3449 RMDIR3args args;
3450 RMDIR3res res;
3454 hrtime_t t;
3462 /*
3463 * Attempt to prevent a rmdir(".") from succeeding.
3464 */
3469 }
3475 }
3479 /*
3480 * First just remove the entry from the name cache, as it
3481 * is most likely an entry for this vp.
3482 */
3485 /*
3486 * If there vnode reference count is greater than one, then
3487 * there may be additional references in the DNLC which will
3488 * need to be purged. First, trying removing the entry for
3489 * the parent directory and see if that removes the additional
3490 * reference(s). If that doesn't do it, then use dnlc_purge_vp
3491 * to completely remove any references to the directory which
3492 * might still exist in the DNLC.
3493 */
3498 }
3516 }
3528 /*
3529 * System V defines rmdir to return EEXIST, not
3530 * ENOTEMPTY if the directory is not empty. Over
3531 * the wire, the error is NFSERR_ENOTEMPTY which
3532 * geterrno maps to ENOTEMPTY.
3533 */
3536 }
3540 }
3546 }
3548 /* ARGSUSED */
3549 static int
3552 {
3554 SYMLINK3args args;
3555 SYMLINK3res res;
3562 hrtime_t t;
3574 /* req time field(s) overflow - return immediately */
3576 }
3598 }
3616 }
3623 KM_NOSLEEP);
3632 contents;
3638 MAXPATHLEN);
3639 }
3640 }
3641 }
3643 }
3651 }
3652 }
3657 }
3659 #ifdef DEBUG
3665 #endif
3669 /*
3670 * Read directory entries.
3671 * There are some weird things to look out for here. The uio_loffset
3672 * field is either 0 or it is the offset returned from a previous
3673 * readdir. It is an opaque value used by the server to find the
3674 * correct directory block to read. The count field is the number
3675 * of blocks to read on the server. This is advisory only, the server
3676 * may return only one block's worth of entries. Entries may be compressed
3677 * on the server.
3678 */
3679 /* ARGSUSED */
3680 static int
3683 {
3690 #ifdef DEBUG
3692 #endif
3694 rddir_cache srdc;
3695 avl_index_t where;
3703 /*
3704 * Make sure that the directory cache is valid.
3705 */
3708 /*
3709 * Setting nfs_disable_rddir_cache in /etc/system
3710 * allows interoperability with servers that do not
3711 * properly update the attributes of directories.
3712 * Any cached information gets purged before an
3713 * access is made to it.
3714 */
3720 }
3721 }
3723 /*
3724 * It is possible that some servers may not be able to correctly
3725 * handle a large READDIR or READDIRPLUS request due to bugs in
3726 * their implementation. In order to continue to interoperate
3727 * with them, this workaround is provided to limit the maximum
3728 * size of a READDIRPLUS request to 1024. In any case, the request
3729 * size is limited to MAXBSIZE.
3730 */
3735 #ifdef DEBUG
3737 #endif
3738 top:
3739 /*
3740 * Short circuit last readdir which always returns 0 bytes.
3741 * This can be done after the directory has been read through
3742 * completely at least once. This will set r_direof which
3743 * can be used to find the value of the last cookie.
3744 */
3749 #ifdef DEBUG
3750 nfs3_readdir_cache_shorts++;
3751 #endif
3757 }
3758 /*
3759 * Look for a cache entry. Cache entries are identified
3760 * by the NFS cookie value and the byte count requested.
3761 */
3767 /*
3768 * If the cache entry is in the process of being
3769 * filled in, wait until this completes. The
3770 * RDDIRWAIT bit is set to indicate that someone
3771 * is waiting and then the thread currently
3772 * filling the entry is done, it should do a
3773 * cv_broadcast to wakeup all of the threads
3774 * waiting for it to finish.
3775 */
3779 #ifdef DEBUG
3780 nfs3_readdir_cache_waits++;
3781 #endif
3783 /*
3784 * We got interrupted, probably
3785 * the user typed ^C or an alarm
3786 * fired. We free the new entry
3787 * if we allocated one.
3788 */
3796 }
3802 }
3803 /*
3804 * Check to see if a readdir is required to
3805 * fill the entry. If so, mark this entry
3806 * as being filled, remove our reference,
3807 * and branch to the code to fill the entry.
3808 */
3817 }
3818 #ifdef DEBUG
3820 nfs3_readdir_cache_hits++;
3821 #endif
3822 /*
3823 * If an error occurred while attempting
3824 * to fill the cache entry, just return it.
3825 */
3833 }
3835 /*
3836 * The cache entry is complete and good,
3837 * copyout the dirent structs to the calling
3838 * thread.
3839 */
3842 /*
3843 * If no error occurred during the copyout,
3844 * update the offset in the uio struct to
3845 * contain the value of the next cookie
3846 * and set the eof value appropriately.
3847 */
3852 }
3854 /*
3855 * Decide whether to do readahead.
3856 *
3857 * Don't if have already read to the end of
3858 * directory. There is nothing more to read.
3859 *
3860 * Don't if the application is not doing
3861 * lookups in the directory. The readahead
3862 * is only effective if the application can
3863 * be doing work while an async thread is
3864 * handling the over the wire request.
3865 */
3871 else
3880 }
3882 /*
3883 * Check to see whether we found an entry
3884 * for the readahead. If so, we don't need
3885 * to do anything further, so free the new
3886 * entry if one was allocated. Otherwise,
3887 * allocate a new entry, add it to the cache,
3888 * and then initiate an asynchronous readdir
3889 * operation to fill it.
3890 */
3902 }
3910 #ifdef DEBUG
3911 nfs3_readdir_readahead++;
3912 #endif
3915 }
3916 }
3921 }
3923 /*
3924 * Didn't find an entry in the cache. Construct a new empty
3925 * entry and link it into the cache. Other processes attempting
3926 * to access this entry will need to wait until it is filled in.
3927 *
3928 * Since kmem_alloc may block, another pass through the cache
3929 * will need to be taken to make sure that another process
3930 * hasn't already added an entry to the cache for this request.
3931 */
3938 }
3940 /*
3941 * Add this entry to the cache.
3942 */
3947 bottom:
3948 #ifdef DEBUG
3950 nfs3_readdir_cache_misses++;
3951 #endif
3952 /*
3953 * Do the readdir. This routine decides whether to use
3954 * READDIR or READDIRPLUS.
3955 */
3958 /*
3959 * If this operation failed, just return the error which occurred.
3960 */
3964 /*
3965 * Since the RPC operation will have taken sometime and blocked
3966 * this process, another pass through the cache will need to be
3967 * taken to find the correct cache entry. It is possible that
3968 * the correct cache entry will not be there (although one was
3969 * added) because the directory changed during the RPC operation
3970 * and the readdir cache was flushed. In this case, just start
3971 * over. It is hoped that this will not happen too often... :-)
3972 */
3975 /* NOTREACHED */
3976 }
3978 static int
3980 {
3988 /*
3989 * Issue the proper request.
3990 *
3991 * If the server does not support READDIRPLUS, then use READDIR.
3992 *
3993 * Otherwise --
3994 * Issue a READDIRPLUS if reading to fill an empty cache or if
3995 * an application has performed a lookup in the directory which
3996 * required an over the wire lookup. The use of READDIRPLUS
3997 * will help to (re)populate the DNLC.
3998 */
4005 }
4017 }
4026 }
4028 static void
4030 {
4032 READDIR3args args;
4033 READDIR3vres res;
4034 vattr_t dva;
4039 hrtime_t t;
4050 /*
4051 * NFS client failover support
4052 * suppress failover unless we have a zero cookie
4053 */
4061 }
4063 #ifdef DEBUG
4065 #else
4067 #endif
4081 }
4094 }
4105 }
4112 }
4122 err:
4126 }
4128 /*
4129 * Read directory entries.
4130 * There are some weird things to look out for here. The uio_loffset
4131 * field is either 0 or it is the offset returned from a previous
4132 * readdir. It is an opaque value used by the server to find the
4133 * correct directory block to read. The count field is the number
4134 * of blocks to read on the server. This is advisory only, the server
4135 * may return only one block's worth of entries. Entries may be compressed
4136 * on the server.
4137 */
4138 static void
4140 {
4142 READDIRPLUS3args args;
4143 READDIRPLUS3vres res;
4144 vattr_t dva;
4160 /*
4161 * NFS client failover support
4162 * suppress failover unless we have a zero cookie
4163 */
4171 }
4173 #ifdef DEBUG
4175 #else
4177 #endif
4192 }
4205 }
4209 }
4220 }
4222 }
4229 }
4240 err:
4244 }
4246 #ifdef DEBUG
4248 #endif
4250 static int
4252 {
4257 offset_t offset;
4274 }
4276 read_again:
4282 /*
4283 * Didn't get it all because we hit EOF,
4284 * zero all the memory beyond the EOF.
4285 */
4286 /* bzero(rdaddr + */
4289 }
4293 /*
4294 * We didn't read anything at all as we are
4295 * past EOF. Return an error indicator back
4296 * but don't destroy the pages (yet).
4297 */
4299 }
4312 }
4314 }
4326 }
4328 write_again:
4334 #ifdef DEBUG
4338 offset);
4342 }
4343 #endif
4358 }
4360 }
4363 /*
4364 * Don't print EDQUOT errors on the console.
4365 * Don't print asynchronous EACCES errors.
4366 * Don't print EFBIG errors.
4367 * Print all other write errors.
4368 */
4373 /*
4374 * Update r_error and r_flags as appropriate.
4375 * If the error was ESTALE, then mark the
4376 * rnode as not being writeable and save
4377 * the error status. Otherwise, save any
4378 * errors which occur from asynchronous
4379 * page invalidations. Any errors occurring
4380 * from other operations should be saved
4381 * by the caller.
4382 */
4393 }
4395 }
4399 /*
4400 * A close may have cleared r_error, if so,
4401 * propagate ESTALE error return properly
4402 */
4405 }
4406 }
4414 }
4416 /* ARGSUSED */
4417 static int
4419 {
4429 }
4433 }
4435 /* ARGSUSED2 */
4436 static int
4438 {
4444 }
4451 }
4455 }
4457 /* ARGSUSED */
4458 static void
4460 {
4464 }
4466 /* ARGSUSED */
4467 static int
4469 {
4471 /*
4472 * Because we stuff the readdir cookie into the offset field
4473 * someone may attempt to do an lseek with the cookie which
4474 * we want to succeed.
4475 */
4481 }
4483 /*
4484 * number of nfs3_bsize blocks to read ahead.
4485 */
4488 #ifdef DEBUG
4490 #endif
4492 /*
4493 * Return all the pages from [off..off+len) in file
4494 */
4495 /* ARGSUSED */
4496 static int
4500 {
4513 /*
4514 * Now valididate that the caches are up to date.
4515 */
4522 retry:
4525 /*
4526 * Don't create dirty pages faster than they
4527 * can be cleaned so that the system doesn't
4528 * get imbalanced. If the async queue is
4529 * maxed out, then wait for it to drain before
4530 * creating more dirty pages. Also, wait for
4531 * any threads doing pagewalks in the vop_getattr
4532 * entry points so that they don't block for
4533 * long periods.
4534 */
4540 }
4542 /*
4543 * If we are getting called as a side effect of an nfs_write()
4544 * operation the local file size might not be extended yet.
4545 * In this case we want to be able to return pages of zeroes.
4546 */
4550 }
4563 }
4566 }
4568 /*
4569 * Called from pvn_getpages to get a particular page.
4570 */
4571 /* ARGSUSED */
4572 static int
4576 {
4578 uint_t bsize;
4581 uoff_t lbn;
4582 uoff_t io_off;
4583 uoff_t blkoff;
4584 uoff_t rablkoff;
4586 uint_t blksize;
4599 reread:
4611 /*
4612 * Queueing up the readahead before doing the synchronous read
4613 * results in a significant increase in read throughput because
4614 * of the increased parallelism between the async threads and
4615 * the process context.
4616 */
4622 /*
4623 * Calculate the number of readaheads to do.
4624 * a) No readaheads at offset = 0.
4625 * b) Do maximum(nfs3_nra) readaheads when the readahead
4626 * window is closed.
4627 * c) Do readaheads between 1 to (nfs3_nra - 1) depending
4628 * upon how far the readahead window is open or close.
4629 * d) No readaheads if rp->r_nextr is not within the scope
4630 * of the readahead window (random i/o).
4631 */
4641 else
4652 }
4653 readahead--;
4655 /*
4656 * Indicate that we did a readahead so
4657 * readahead offset is not updated
4658 * by the synchronous read below.
4659 */
4662 /*
4663 * set readahead offset to
4664 * offset of last async readahead
4665 * request.
4666 */
4668 }
4670 }
4672 again:
4676 nfs3_readahead);
4678 /*
4679 * Block for this page is not allocated, or the offset
4680 * is beyond the current allocation size, or we're
4681 * allocating a swap slot and the page was not found,
4682 * so allocate it and return a zero page.
4683 */
4692 /*
4693 * Need to go to server to get a BLOCK, exception to
4694 * that being while reading at offset = 0 or doing
4695 * random i/o, in that case read only a PAGE.
4696 */
4700 /*
4701 * If only a block or less is left in
4702 * the file, read all that is remaining.
4703 */
4705 /*
4706 * Trying to access beyond EOF,
4707 * set up to get at least one page.
4708 */
4723 /*
4724 * Some other thread has entered the page,
4725 * so just use it.
4726 */
4730 /*
4731 * Now round the request size up to page boundaries.
4732 * This ensures that the entire page will be
4733 * initialized to zeroes if EOF is encountered.
4734 */
4740 /*
4741 * pageio_setup should have set b_addr to 0. This
4742 * is correct since we want to do I/O on a page
4743 * boundary. bp_mapin will use this addr to calculate
4744 * an offset, and then set b_addr to the kernel virtual
4745 * address it allocated for us.
4746 */
4756 /*
4757 * If doing a write beyond what we believe is EOF,
4758 * don't bother trying to read the pages from the
4759 * server, we'll just zero the pages here. We
4760 * don't check that the rw flag is S_WRITE here
4761 * because some implementations may attempt a
4762 * read access to the buffer before copying data.
4763 */
4771 }
4773 /*
4774 * Unmap the buffer before freeing it.
4775 */
4785 /*
4786 * If doing a write system call just return
4787 * zeroed pages, else user tried to get pages
4788 * beyond EOF, return error. We don't check
4789 * that the rw flag is S_WRITE here because
4790 * some implementations may attempt a read
4791 * access to the buffer before copying data.
4792 */
4795 else
4797 }
4803 }
4804 }
4805 }
4807 out:
4815 }
4820 /*
4821 * Page exists in the cache, acquire the appropriate lock.
4822 * If this fails, start all over again.
4823 */
4825 #ifdef DEBUG
4826 nfs3_lostpage++;
4827 #endif
4829 }
4833 }
4839 }
4841 static void
4844 {
4847 uoff_t io_off;
4859 /*
4860 * If less than a block left in file read less
4861 * than a block.
4862 */
4870 /*
4871 * The isra flag passed to the kluster function is 1, we may have
4872 * gotten a return value of NULL for a variety of reasons (# of free
4873 * pages < minfree, someone entered the page on the vnode etc). In all
4874 * cases, we want to punt on the readahead.
4875 */
4879 /*
4880 * Now round the request size up to page boundaries.
4881 * This ensures that the entire page will be
4882 * initialized to zeroes if EOF is encountered.
4883 */
4889 /*
4890 * pageio_setup should have set b_addr to 0. This is correct since
4891 * we want to do I/O on a page boundary. bp_mapin() will use this addr
4892 * to calculate an offset, and then set b_addr to the kernel virtual
4893 * address it allocated for us.
4894 */
4904 /*
4905 * If doing a write beyond what we believe is EOF, don't bother trying
4906 * to read the pages from the server, we'll just zero the pages here.
4907 * We don't check that the rw flag is S_WRITE here because some
4908 * implementations may attempt a read access to the buffer before
4909 * copying data.
4910 */
4921 }
4923 /*
4924 * Unmap the buffer before freeing it.
4925 */
4936 /*
4937 * In case of error set readahead offset
4938 * to the lowest offset.
4939 * pvn_read_done() calls VN_DISPOSE to destroy the pages
4940 */
4946 }
4947 }
4949 /*
4950 * Flags are composed of {B_INVAL, B_FREE, B_DONTNEED, B_FORCE}
4951 * If len == 0, do from off to EOF.
4952 *
4953 * The normal cases should be len == 0 && off == 0 (entire vp list),
4954 * len == MAXBSIZE (from segmap_release actions), and len == PAGESIZE
4955 * (from pageout).
4956 */
4957 /* ARGSUSED */
4958 static int
4961 {
4967 /*
4968 * XXX - Why should this check be made here?
4969 */
4988 }
4990 /*
4991 * Write out a single page, possibly klustering adjacent dirty pages.
4992 */
4993 int
4996 {
4997 uoff_t io_off;
4998 uoff_t lbn_off;
4999 uoff_t lbn;
5001 uint_t bsize;
5017 /*
5018 * Find a kluster that fits in one block, or in
5019 * one page if pages are bigger than blocks. If
5020 * there is less file space allocated than a whole
5021 * page, we'll shorten the i/o request below.
5022 */
5026 /*
5027 * pvn_write_kluster shouldn't have returned a page with offset
5028 * behind the original page we were given. Verify that.
5029 */
5032 /*
5033 * Now pp will have the list of kept dirty pages marked for
5034 * write back. It will also handle invalidation and freeing
5035 * of pages that are not dirty. Check for page length rounding
5036 * problems.
5037 */
5041 }
5042 /*
5043 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
5044 * consistent value of r_size. RMODINPROGRESS is set in writerp().
5045 * When RMODINPROGRESS is set it indicates that a uiomove() is in
5046 * progress and the r_size has not been made consistent with the
5047 * new size of the file. When the uiomove() completes the r_size is
5048 * updated and the RMODINPROGRESS flag is cleared.
5049 *
5050 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
5051 * consistent value of r_size. Without this handshaking, it is
5052 * possible that nfs(3)_bio() picks up the old value of r_size
5053 * before the uiomove() in writerp() completes. This will result
5054 * in the write through nfs(3)_bio() being dropped.
5055 *
5056 * More precisely, there is a window between the time the uiomove()
5057 * completes and the time the r_size is updated. If a fop_putpage()
5058 * operation intervenes in this window, the page will be picked up,
5059 * because it is dirty (it will be unlocked, unless it was
5060 * pagecreate'd). When the page is picked up as dirty, the dirty
5061 * bit is reset (pvn_getdirty()). In nfs(3)write(), r_size is
5062 * checked. This will still be the old size. Therefore the page will
5063 * not be written out. When segmap_release() calls fop_putpage(),
5064 * the page will be found to be clean and the write will be dropped.
5065 */
5072 /*
5073 * A write is in progress for this region of the file.
5074 * If we did not detect RMODINPROGRESS here then this
5075 * path through nfs_putapage() would eventually go to
5076 * nfs(3)_bio() and may not write out all of the data
5077 * in the pages. We end up losing data. So we decide
5078 * to set the modified bit on each page in the page
5079 * list and mark the rnode with RDIRTY. This write
5080 * will be restarted at some later time.
5081 */
5089 }
5097 }
5099 }
5103 nfs3_sync_putapage);
5112 }
5114 static int
5117 {
5136 }
5139 /*
5140 * If this was not an async thread, then try again to
5141 * write out the pages, but this time, also destroy
5142 * them whether or not the write is successful. This
5143 * will prevent memory from filling up with these
5144 * pages and destroying them is the only alternative
5145 * if they can't be written out.
5146 *
5147 * Don't do this if this is an async thread because
5148 * when the pages are unlocked in pvn_write_done,
5149 * some other thread could have come along, locked
5150 * them, and queued for an async thread. It would be
5151 * possible for all of the async threads to be tied
5152 * up waiting to lock the pages again and they would
5153 * all already be locked and waiting for an async
5154 * thread to handle them. Deadlock.
5155 */
5159 }
5167 }
5171 }
5174 }
5176 /* ARGSUSED */
5177 static int
5181 {
5199 /*
5200 * If there is cached data and if close-to-open consistency
5201 * checking is not turned off and if the file system is not
5202 * mounted readonly, then force an over the wire getattr.
5203 * Otherwise, just invoke nfs3getattr to get a copy of the
5204 * attributes. The attribute cache will be used unless it
5205 * is timed out and if it is, then an over the wire getattr
5206 * will be issued.
5207 */
5212 else
5217 /*
5218 * Check to see if the vnode is currently marked as not cachable.
5219 * This means portions of the file are locked (through fop_frlock).
5220 * In this case the map request must be refused. We use
5221 * rp->r_lkserlock to avoid a race with concurrent lock requests.
5222 */
5225 /*
5226 * Atomically increment r_inmap after acquiring r_rwlock. The
5227 * idea here is to acquire r_rwlock to block read/write and
5228 * not to protect r_inmap. r_inmap will inform nfs3_read/write()
5229 * that we are in nfs3_map(). Now, r_rwlock is acquired in order
5230 * and we can prevent the deadlock that would have occurred
5231 * when nfs3_addmap() would have acquired it out of order.
5232 *
5233 * Since we are not protecting r_inmap by any lock, we do not
5234 * hold any lock when we decrement it. We atomically decrement
5235 * r_inmap after we release r_lkserlock.
5236 */
5246 }
5251 }
5253 /*
5254 * Don't allow concurrent locks and mapping if mandatory locking is
5255 * enabled.
5256 */
5261 }
5268 }
5284 done:
5288 }
5290 /* ARGSUSED */
5291 static int
5295 {
5307 }
5309 /* ARGSUSED */
5310 static int
5314 {
5315 netobj lm_fh3;
5323 /* check for valid cmd parameter */
5327 /* Verify l_type. */
5343 }
5345 /* check the validity of the lock range */
5351 /*
5352 * If the filesystem is mounted using local locking, pass the
5353 * request off to the local locking code.
5354 */
5357 /*
5358 * For complete safety, we should be holding
5359 * r_lkserlock. However, we can't call
5360 * lm_safelock and then fs_frlock while
5361 * holding r_lkserlock, so just invoke
5362 * lm_safelock and expect that this will
5363 * catch enough of the cases.
5364 */
5367 }
5369 }
5373 /*
5374 * Check whether the given lock request can proceed, given the
5375 * current file mappings.
5376 */
5383 }
5384 }
5386 /*
5387 * Flush the cache after waiting for async I/O to finish. For new
5388 * locks, this is so that the process gets the latest bits from the
5389 * server. For unlocks, this is so that other clients see the
5390 * latest bits once the file has been unlocked. If currently dirty
5391 * pages can't be flushed, then don't allow a lock to be set. But
5392 * allow unlocks to succeed, to avoid having orphan locks on the
5393 * server.
5394 */
5409 }
5414 }
5425 }
5429 }
5430 }
5431 }
5436 /*
5437 * Call the lock manager to do the real work of contacting
5438 * the server and obtaining the lock.
5439 */
5445 done:
5448 }
5450 /*
5451 * Free storage space associated with the specified vnode. The portion
5452 * to be freed is specified by bfp->l_start and bfp->l_len (already
5453 * normalized to a "whence" of 0).
5454 *
5455 * This is an experimental facility whose continued existence is not
5456 * guaranteed. Currently, we only support the special case
5457 * of l_len == 0, meaning free to end of file.
5458 */
5459 /* ARGSUSED */
5460 static int
5463 {
5478 /*
5479 * ftruncate should not change the ctime and
5480 * mtime if we truncate the file to its
5481 * previous size.
5482 */
5495 }
5498 }
5500 /* ARGSUSED */
5501 static int
5503 {
5506 }
5508 /*
5509 * Setup and add an address space callback to do the work of the delmap call.
5510 * The callback will (and must be) deleted in the actual callback function.
5511 *
5512 * This is done in order to take care of the problem that we have with holding
5513 * the address space's a_lock for a long period of time (e.g. if the NFS server
5514 * is down). Callbacks will be executed in the address space code while the
5515 * a_lock is not held. Holding the address space's a_lock causes things such
5516 * as ps and fork to hang because they are trying to acquire this lock as well.
5517 */
5518 /* ARGSUSED */
5519 static int
5523 {
5532 /*
5533 * A process may not change zones if it has NFS pages mmap'ed
5534 * in, so we can't legitimately get here from the wrong zone.
5535 */
5540 /*
5541 * The way that the address space of this process deletes its mapping
5542 * of this file is via the following call chains:
5543 * - as_free()->segop_unmap()/segvn_unmap()->fop_delmap()/nfs3_delmap()
5544 * - as_unmap()->segop_unmap()/segvn_unmap()->fop_delmap()/nfs3_delmap()
5545 *
5546 * With the use of address space callbacks we are allowed to drop the
5547 * address space lock, a_lock, while executing the NFS operations that
5548 * need to go over the wire. Returning EAGAIN to the caller of this
5549 * function is what drives the execution of the callback that we add
5550 * below. The callback will be executed by the address space code
5551 * after dropping the a_lock. When the callback is finished, since
5552 * we dropped the a_lock, it must be re-acquired and segvn_unmap()
5553 * is called again on the same segment to finish the rest of the work
5554 * that needs to happen during unmapping.
5555 *
5556 * This action of calling back into the segment driver causes
5557 * nfs3_delmap() to get called again, but since the callback was
5558 * already executed at this point, it already did the work and there
5559 * is nothing left for us to do.
5560 *
5561 * To Summarize:
5562 * - The first time nfs3_delmap is called by the current thread is when
5563 * we add the caller associated with this delmap to the delmap caller
5564 * list, add the callback, and return EAGAIN.
5565 * - The second time in this call chain when nfs3_delmap is called we
5566 * will find this caller in the delmap caller list and realize there
5567 * is no more work to do thus removing this caller from the list and
5568 * returning the error that was set in the callback execution.
5569 */
5572 /*
5573 * 'error' is from the actual delmap operations. To avoid
5574 * hangs, we need to handle the return of EAGAIN differently
5575 * since this is what drives the callback execution.
5576 * In this case, we don't want to return EAGAIN and do the
5577 * callback execution because there are none to execute.
5578 */
5581 else
5583 }
5585 /* current caller was not in the list */
5608 }
5610 /*
5611 * Remove some pages from an mmap'd vnode. Just update the
5612 * count of pages. If doing close-to-open, then flush and
5613 * commit all of the pages associated with this file.
5614 * Otherwise, start an asynchronous page flush to write out
5615 * any dirty pages. This will also associate a credential
5616 * with the rnode which can be used to write the pages.
5617 */
5618 /* ARGSUSED */
5619 static void
5621 {
5633 /*
5634 * Initiate a page flush and potential commit if there are
5635 * pages, the file system was not mounted readonly, the segment
5636 * was mapped shared, and the pages themselves were writeable.
5637 */
5643 /*
5644 * If this is a cross-zone access a sync putpage won't work, so
5645 * the best we can do is try an async putpage. That seems
5646 * better than something more draconian such as discarding the
5647 * dirty pages.
5648 */
5653 else
5661 }
5672 }
5676 #ifdef DEBUG
5679 #endif
5681 /* ARGSUSED */
5682 static int
5685 {
5687 PATHCONF3args args;
5688 PATHCONF3res res;
5690 failinfo_t fi;
5692 hrtime_t t;
5696 /*
5697 * Large file spec - need to base answer on info stored
5698 * on original FSINFO response.
5699 */
5709 }
5713 }
5716 }
5719 }
5722 }
5725 }
5732 }
5737 }
5768 }
5769 }
5771 }
5779 }
5802 }
5804 #ifdef DEBUG
5805 nfs3_pathconf_cache_hits++;
5806 #endif
5808 }
5810 }
5811 #ifdef DEBUG
5812 nfs3_pathconf_cache_misses++;
5813 #endif
5845 }
5847 }
5867 }
5871 }
5874 }
5876 /*
5877 * Called by async thread to do synchronous pageio. Do the i/o, wait
5878 * for it to complete, and cleanup the page list when done.
5879 */
5880 static int
5883 {
5890 else
5893 }
5895 /* ARGSUSED */
5896 static int
5899 {
5915 nfs3_sync_pageio);
5923 }
5925 /* ARGSUSED */
5926 static void
5929 {
5934 offset3 offset;
5935 count3 len;
5936 k_sigset_t smask;
5938 /*
5939 * We should get called with fl equal to either B_FREE or
5940 * B_INVAL. Any other value is illegal.
5941 *
5942 * The page that we are either supposed to free or destroy
5943 * should be exclusive locked and its io lock should not
5944 * be held.
5945 */
5950 /*
5951 * If the page doesn't need to be committed or we shouldn't
5952 * even bother attempting to commit it, then just make sure
5953 * that the p_fsdata byte is clear and then either free or
5954 * destroy the page as appropriate.
5955 */
5960 else
5963 }
5965 /*
5966 * If there is a page invalidation operation going on, then
5967 * if this is one of the pages being destroyed, then just
5968 * clear the p_fsdata byte and then either free or destroy
5969 * the page as appropriate.
5970 */
5977 else
5980 }
5982 /*
5983 * If we are freeing this page and someone else is already
5984 * waiting to do a commit, then just unlock the page and
5985 * return. That other thread will take care of commiting
5986 * this page. The page can be freed sometime after the
5987 * commit has finished. Otherwise, if the page is marked
5988 * as delay commit, then we may be getting called from
5989 * pvn_write_done, one page at a time. This could result
5990 * in one commit per page, so we end up doing lots of small
5991 * commits instead of fewer larger commits. This is bad,
5992 * we want do as few commits as possible.
5993 */
5999 }
6005 }
6006 }
6008 /*
6009 * Check to see if there is a signal which would prevent an
6010 * attempt to commit the pages from being successful. If so,
6011 * then don't bother with all of the work to gather pages and
6012 * generate the unsuccessful RPC. Just return from here and
6013 * let the page be committed at some later time.
6014 */
6021 }
6024 /*
6025 * We are starting to need to commit pages, so let's try
6026 * to commit as many as possible at once to reduce the
6027 * overhead.
6028 *
6029 * Set the `commit inprogress' state bit. We must
6030 * first wait until any current one finishes. Then
6031 * we initialize the c_pages list with this page.
6032 */
6037 }
6045 /*
6046 * Gather together all other pages which can be committed.
6047 * They will all be chained off r_commit.c_pages.
6048 */
6051 /*
6052 * Clear the `commit inprogress' status and disconnect
6053 * the list of pages to be committed from the rnode.
6054 * At this same time, we also save the starting offset
6055 * and length of data to be committed on the server.
6056 */
6070 }
6072 /*
6073 * Actually generate the COMMIT3 over the wire operation.
6074 */
6077 /*
6078 * If we got an error during the commit, just unlock all
6079 * of the pages. The pages will get retransmitted to the
6080 * server during a putpage operation.
6081 */
6087 }
6089 }
6091 /*
6092 * We've tried as hard as we can to commit the data to stable
6093 * storage on the server. We release the rest of the pages
6094 * and clear the commit required state. They will be put
6095 * onto the tail of the cachelist if they are nolonger
6096 * mapped.
6097 */
6103 }
6105 /*
6106 * It is possible that nfs3_commit didn't return error but
6107 * some other thread has modified the page we are going
6108 * to free/destroy.
6109 * In this case we need to rewrite the page. Do an explicit check
6110 * before attempting to free/destroy the page. If modified, needs to
6111 * be rewritten so unlock the page and return.
6112 */
6117 }
6119 /*
6120 * Now, as appropriate, either free or destroy the page
6121 * that we were called with.
6122 */
6126 else
6128 }
6130 static int
6132 {
6135 COMMIT3args args;
6136 COMMIT3res res;
6152 }
6159 doitagain:
6178 }
6195 }
6197 }
6198 /*
6199 * Can't do a PURGE_STALE_FH here because this
6200 * can cause a deadlock. nfs3_commit can
6201 * be called from nfs3_dispose which can be called
6202 * indirectly via pvn_vplist_dirty. PURGE_STALE_FH
6203 * can call back to pvn_vplist_dirty.
6204 */
6217 }
6218 }
6221 }
6223 static void
6225 {
6229 }
6231 /*
6232 * This routine is used to gather together a page list of the pages
6233 * which are to be committed on the server. This routine must not
6234 * be called if the calling thread holds any locked pages.
6235 *
6236 * The calling thread must have set RCOMMIT. This bit is used to
6237 * serialize access to the commit structure in the rnode. As long
6238 * as the thread has set RCOMMIT, then it can manipulate the commit
6239 * structure without requiring any other locks.
6240 */
6241 static void
6243 {
6253 /*
6254 * Step through all of the pages associated with this vnode
6255 * looking for pages which need to be committed.
6256 */
6260 /* Skip marker pages. */
6264 /*
6265 * If this page does not need to be committed or is
6266 * modified, then just skip it.
6267 */
6271 /*
6272 * Attempt to lock the page. If we can't, then
6273 * someone else is messing with it and we will
6274 * just skip it.
6275 */
6279 /*
6280 * If this page does not need to be committed or is
6281 * modified, then just skip it. Recheck now that
6282 * the page is locked.
6283 */
6287 }
6292 }
6294 /*
6295 * The page needs to be committed and we locked it.
6296 * Update the base and length parameters and add it
6297 * to r_pages.
6298 */
6310 }
6312 }
6315 }
6317 /*
6318 * This routine is used to gather together a page list of the pages
6319 * which are to be committed on the server. This routine must not
6320 * be called if the calling thread holds any locked pages.
6321 *
6322 * The calling thread must have set RCOMMIT. This bit is used to
6323 * serialize access to the commit structure in the rnode. As long
6324 * as the thread has set RCOMMIT, then it can manipulate the commit
6325 * structure without requiring any other locks.
6326 */
6327 static void
6329 {
6333 uoff_t end;
6334 uoff_t off;
6343 /*
6344 * If there are no pages associated with this vnode, then
6345 * just return.
6346 */
6350 /*
6351 * Calculate the ending offset.
6352 */
6356 /*
6357 * Lookup each page by vp, offset.
6358 */
6362 /*
6363 * If this page does not need to be committed or is
6364 * modified, then just skip it.
6365 */
6369 }
6373 /*
6374 * The page needs to be committed and we locked it.
6375 * Update the base and length parameters and add it
6376 * to r_pages.
6377 */
6384 }
6386 }
6387 }
6389 static int
6391 {
6393 writeverf3 write_verf;
6397 /*
6398 * Flush the data portion of the file and then commit any
6399 * portions which need to be committed. This may need to
6400 * be done twice if the server has changed state since
6401 * data was last written. The data will need to be
6402 * rewritten to the server and then a new commit done.
6403 *
6404 * In fact, this may need to be done several times if the
6405 * server is having problems and crashing while we are
6406 * attempting to do this.
6407 */
6409 top:
6410 /*
6411 * Do a flush based on the poff and plen arguments. This
6412 * will asynchronously write out any modified pages in the
6413 * range specified by (poff, plen). This starts all of the
6414 * i/o operations which will be waited for in the next
6415 * call to nfs3_putpage
6416 */
6426 /*
6427 * Do a flush based on the poff and plen arguments. This
6428 * will synchronously write out any modified pages in the
6429 * range specified by (poff, plen) and wait until all of
6430 * the asynchronous i/o's in that range are done as well.
6431 */
6442 }
6445 /*
6446 * Now commit any pages which might need to be committed.
6447 * If the error, NFS_VERF_MISMATCH, is returned, then
6448 * start over with the flush operation.
6449 */
6457 }
6459 static int
6461 {
6464 offset3 offset;
6465 count3 len;
6472 /*
6473 * Set the `commit inprogress' state bit. We must
6474 * first wait until any current one finishes.
6475 */
6481 }
6485 /*
6486 * Gather together all of the pages which need to be
6487 * committed.
6488 */
6491 else
6494 /*
6495 * Clear the `commit inprogress' bit and disconnect the
6496 * page list which was gathered together in nfs3_get_commit.
6497 */
6507 /*
6508 * If any pages need to be committed, commit them and
6509 * then unlock them so that they can be freed some
6510 * time later.
6511 */
6513 /*
6514 * No error occurred during the flush portion
6515 * of this operation, so now attempt to commit
6516 * the data to stable storage on the server.
6517 *
6518 * This will unlock all of the pages on the list.
6519 */
6521 }
6523 }
6525 static int
6528 {
6535 /*
6536 * If we got an error, then just unlock all of the pages
6537 * on the list.
6538 */
6544 }
6546 }
6547 /*
6548 * We've tried as hard as we can to commit the data to stable
6549 * storage on the server. We just unlock the pages and clear
6550 * the commit required state. They will get freed later.
6551 */
6557 }
6560 }
6562 static void
6565 {
6568 }
6570 /* ARGSUSED */
6571 static int
6574 {
6587 }
6590 }
6592 /* ARGSUSED */
6593 static int
6596 {
6609 }
6612 }
6614 /* ARGSUSED */
6615 static int
6618 {
6622 netobj lm_fh3;
6627 /*
6628 * check for valid cmd parameter
6629 */
6633 /*
6634 * Check access permissions
6635 */
6641 /*
6642 * If the filesystem is mounted using local locking, pass the
6643 * request off to the local share code.
6644 */
6654 /*
6655 * If passed an owner that is too large to fit in an
6656 * nfs_owner it is likely a recursive call from the
6657 * lock manager client and pass it straight through. If
6658 * it is not a nfs_owner then simply return an error.
6659 */
6662 NFS_OWNER_MAGIC)
6667 }
6669 }
6670 /*
6671 * Remote share reservations owner is a combination of
6672 * a magic number, hostname, and the local owner
6673 */
6688 }
6693 /*
6694 * NFS client can't store remote locks itself
6695 */
6703 }
6706 }