| blob | c45366c0e8da6d2797c47a519fd051d1e646594a |
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 }
2293 /*
2294 * Check here for large file handled
2295 * by LF-unaware process (as
2296 * ufs_create() does)
2297 */
2303 }
2309 /*
2310 * Existing file was truncated;
2311 * emit a create event.
2312 */
2314 }
2315 }
2316 }
2317 }
2323 }
2326 }
2330 /*
2331 * Decide what the group-id of the created file should be.
2332 * Set it in attribute list as advisory...
2333 */
2338 }
2347 }
2349 lfaware);
2350 /*
2351 * If this is not an exclusive create, then the CREATE
2352 * request will be made with the GUARDED mode set. This
2353 * means that the server will return EEXIST if the file
2354 * exists. The file could exist because of a retransmitted
2355 * request. In this case, we recover by starting over and
2356 * checking to see whether the file exists. This second
2357 * time through it should and a CREATE request will not be
2358 * sent.
2359 *
2360 * This handles the problem of a dangling CREATE request
2361 * which contains attributes which indicate that the file
2362 * should be truncated. This retransmitted request could
2363 * possibly truncate valid data in the file if not caught
2364 * by the duplicate request mechanism on the server or if
2365 * not caught by other means. The scenario is:
2366 *
2367 * Client transmits CREATE request with size = 0
2368 * Client times out, retransmits request.
2369 * Response to the first request arrives from the server
2370 * and the client proceeds on.
2371 * Client writes data to the file.
2372 * The server now processes retransmitted CREATE request
2373 * and truncates file.
2374 *
2375 * The use of the GUARDED CREATE request prevents this from
2376 * happening because the retransmitted CREATE would fail
2377 * with EEXIST and would not truncate the file.
2378 */
2380 #ifdef DEBUG
2381 nfs3_create_misses++;
2382 #endif
2384 }
2387 }
2391 }
2393 /* ARGSUSED */
2394 static int
2397 {
2399 CREATE3args args;
2400 CREATE3res res;
2406 timestruc_t now;
2407 hrtime_t t;
2413 /*
2414 * Construct the create verifier. This verifier needs
2415 * to be unique between different clients. It also needs
2416 * to vary for each exclusive create request generated
2417 * from the client to the server.
2418 *
2419 * The first attempt is made to use the hostid and a
2420 * unique number on the client. If the hostid has not
2421 * been set, the high resolution time that the exclusive
2422 * create request is being made is used. This will work
2423 * unless two different clients, both with the hostid
2424 * not set, attempt an exclusive create request on the
2425 * same file, at exactly the same clock time. The
2426 * chances of this happening seem small enough to be
2427 * reasonable.
2428 */
2437 }
2438 /*
2439 * Since the server will use this value for the mtime,
2440 * make sure that it can't overflow. Zero out the MSB.
2441 * The actual value does not matter here, only its uniqeness.
2442 */
2445 /*
2446 * Issue the non-exclusive create in guarded mode. This
2447 * may result in some false EEXIST responses for
2448 * retransmitted requests, but these will be handled at
2449 * a higher level. By using GUARDED, duplicate requests
2450 * to do file truncation and possible access problems
2451 * can be avoided.
2452 */
2457 /* req time field(s) overflow - return immediately */
2459 }
2460 }
2474 }
2482 /*
2483 * On exclusive create the times need to be explicitly
2484 * set to clear any potential verifier that may be stored
2485 * in one of these fields (see comment below). This
2486 * is done here to cover the case where no post op attrs
2487 * were returned or a 'invalid' time was returned in
2488 * the attributes.
2489 */
2506 /*
2507 * On an exclusive create, it is possible
2508 * that attributes were returned but those
2509 * postop attributes failed to decode
2510 * properly. If this is the case,
2511 * then most likely the atime or mtime
2512 * were invalid for our client; this
2513 * is caused by the server storing the
2514 * create verifier in one of the time
2515 * fields(most likely mtime).
2516 * So... we are going to setattr just the
2517 * atime/mtime to clear things up.
2518 */
2523 /*
2524 * Setting the times failed.
2525 * Remove the file and return
2526 * the error.
2527 */
2532 }
2533 }
2535 /*
2536 * This handles the non-exclusive case
2537 * and the exclusive case where no post op
2538 * attrs were returned.
2539 */
2546 }
2548 }
2549 }
2551 }
2555 /*
2556 * Check here for large file handled by
2557 * LF-unaware process (as ufs_create() does)
2558 */
2566 }
2568 }
2572 /*
2573 * If doing an exclusive create, then generate
2574 * a SETATTR to set the initial attributes.
2575 * Try to set the mtime and the atime to the
2576 * server's current time. It is somewhat
2577 * expected that these fields will be used to
2578 * store the exclusive create cookie. If not,
2579 * server implementors will need to know that
2580 * a SETATTR will follow an exclusive create
2581 * and the cookie should be destroyed if
2582 * appropriate. This work may have been done
2583 * earlier in this function if post op attrs
2584 * were not available.
2585 *
2586 * The VATTR_GID and VATTR_SIZE bits are turned off
2587 * so that the SETATTR request will not attempt
2588 * to process these. The gid will be set
2589 * separately if appropriate. The size is turned
2590 * off because it is assumed that a new file will
2591 * be created empty and if the file wasn't empty,
2592 * then the exclusive create will have failed
2593 * because the file must have existed already.
2594 * Therefore, no truncate operation is needed.
2595 */
2599 /*
2600 * Couldn't correct the attributes of
2601 * the newly created file and the
2602 * attributes are wrong. Remove the
2603 * file and return an error to the
2604 * application.
2605 */
2609 }
2610 }
2613 /*
2614 * If the gid on the file isn't right, then
2615 * generate a SETATTR to attempt to change
2616 * it. This may or may not work, depending
2617 * upon the server's semantics for allowing
2618 * file ownership changes.
2619 */
2622 }
2624 /*
2625 * If vnode is a device create special vnode
2626 */
2635 }
2638 }
2640 /*
2641 * Special setattr function to take care of rest of atime/mtime
2642 * after successful exclusive create. This function exists to avoid
2643 * handling attributes from the server; exclusive the atime/mtime fields
2644 * may be 'invalid' in client's view and therefore can not be trusted.
2645 */
2646 static int
2648 {
2650 uint_t mask;
2651 SETATTR3args args;
2652 SETATTR3res res;
2655 hrtime_t t;
2658 /* save the caller's mask so that it can be reset later */
2666 /* Use the mask to initialize the arguments */
2670 /* We want to set just atime/mtime on this request */
2686 }
2690 /*
2691 * It is important to pick up the attributes.
2692 * Since this is the exclusive create path, the
2693 * attributes on the initial create were ignored
2694 * and we need these to have the correct info.
2695 */
2697 /*
2698 * No need to do the atime/mtime work again so clear
2699 * the bits.
2700 */
2704 }
2709 }
2711 /* ARGSUSED */
2712 static int
2715 {
2717 MKNOD3args args;
2718 MKNOD3res res;
2722 hrtime_t t;
2733 /* req time field(s) overflow - return immediately */
2735 }
2749 /* req time field(s) overflow - return immediately */
2751 }
2756 }
2770 }
2796 }
2798 }
2800 }
2802 }
2807 }
2809 /*
2810 * If vnode is a device create special vnode
2811 */
2820 }
2822 }
2824 /*
2825 * Weirdness: if the vnode to be removed is open
2826 * we rename it instead of removing it and nfs_inactive
2827 * will remove the new name.
2828 */
2829 /* ARGSUSED */
2830 static int
2832 {
2834 REMOVE3args args;
2835 REMOVE3res res;
2841 hrtime_t t;
2853 }
2859 }
2861 /*
2862 * First just remove the entry from the name cache, as it
2863 * is most likely the only entry for this vp.
2864 */
2867 /*
2868 * If the file has a v_count > 1 then there may be more than one
2869 * entry in the name cache due multiple links or an open file,
2870 * but we don't have the real reference count so flush all
2871 * possible entries.
2872 */
2876 /*
2877 * Now we have the real reference count on the vnode
2878 */
2901 }
2903 }
2906 /*
2907 * We need to flush any dirty pages which happen to
2908 * be hanging around before removing the file. This
2909 * shouldn't happen very often and mostly on file
2910 * systems mounted "nocto".
2911 */
2920 }
2921 }
2934 /*
2935 * The xattr dir may be gone after last attr is removed,
2936 * so flush it from dnlc.
2937 */
2949 cr);
2956 }
2957 }
2958 }
2962 }
2968 }
2970 /* ARGSUSED */
2971 static int
2974 {
2976 LINK3args args;
2977 LINK3res res;
2982 hrtime_t t;
3017 }
3029 cr);
3035 }
3036 }
3041 /*
3042 * Notify the source file of this link operation.
3043 */
3045 }
3047 }
3049 /* ARGSUSED */
3050 static int
3053 {
3062 }
3064 /*
3065 * nfs3rename does the real work of renaming in NFS Version 3.
3066 */
3067 static int
3070 {
3072 RENAME3args args;
3073 RENAME3res res;
3081 hrtime_t t;
3097 }
3104 }
3105 }
3107 /*
3108 * Lookup the target file. If it exists, it needs to be
3109 * checked to see whether it is a mount point and whether
3110 * it is active (open).
3111 */
3114 /*
3115 * If this file has been mounted on, then just
3116 * return busy because renaming to it would remove
3117 * the mounted file system from the name space.
3118 */
3124 }
3126 /*
3127 * Purge the name cache of all references to this vnode
3128 * so that we can check the reference count to infer
3129 * whether it is active or not.
3130 */
3131 /*
3132 * First just remove the entry from the name cache, as it
3133 * is most likely the only entry for this vp.
3134 */
3136 /*
3137 * If the file has a v_count > 1 then there may be more
3138 * than one entry in the name cache due multiple links
3139 * or an open file, but we don't have the real reference
3140 * count so flush all possible entries.
3141 */
3145 /*
3146 * If the vnode is active and is not a directory,
3147 * arrange to rename it to a
3148 * temporary file so that it will continue to be
3149 * accessible. This implements the "unlink-open-file"
3150 * semantics for the target of a rename operation.
3151 * Before doing this though, make sure that the
3152 * source and target files are not already the same.
3153 */
3155 /*
3156 * Lookup the source name.
3157 */
3161 /*
3162 * The source name *should* already exist.
3163 */
3169 }
3171 /*
3172 * Compare the two vnodes. If they are the same,
3173 * just release all held vnodes and return success.
3174 */
3181 }
3183 /*
3184 * Can't mix and match directories and non-
3185 * directories in rename operations. We already
3186 * know that the target is not a directory. If
3187 * the source is a directory, return an error.
3188 */
3195 }
3197 /*
3198 * The target file exists, is not the same as
3199 * the source file, and is active. Link it
3200 * to a temporary filename to avoid having
3201 * the server removing the file completely.
3202 */
3208 }
3216 }
3230 }
3232 }
3233 }
3236 /*
3237 * When renaming directories to be a subdirectory of a
3238 * different parent, the dnlc entry for ".." will no
3239 * longer be valid, so it must be removed.
3240 *
3241 * We do a lookup here to determine whether we are renaming
3242 * a directory and we need to check if we are renaming
3243 * an unlinked file. This might have already been done
3244 * in previous code, so we check ovp == NULL to avoid
3245 * doing it twice.
3246 */
3249 /*
3250 * The source name *should* already exist.
3251 */
3257 }
3259 }
3261 }
3286 }
3288 }
3300 }
3301 /*
3302 * when renaming directories to be a subdirectory of a
3303 * different parent, the dnlc entry for ".." will no
3304 * longer be valid, so it must be removed
3305 */
3312 }
3313 }
3315 /*
3316 * If we are renaming the unlinked file, update the
3317 * r_unldvp and r_unlname as needed.
3318 */
3329 }
3330 }
3331 }
3337 cr);
3338 }
3339 /*
3340 * System V defines rename to return EEXIST, not
3341 * ENOTEMPTY if the target directory is not empty.
3342 * Over the wire, the error is NFSERR_ENOTEMPTY
3343 * which geterrno maps to ENOTEMPTY.
3344 */
3347 }
3357 }
3361 }
3368 }
3370 /* ARGSUSED */
3371 static int
3374 {
3376 MKDIR3args args;
3377 MKDIR3res res;
3382 hrtime_t t;
3388 /*
3389 * Decide what the group-id and set-gid bit of the created directory
3390 * should be. May have to do a setattr to get the gid right.
3391 */
3402 /* req time field(s) overflow - return immediately */
3404 }
3425 }
3438 }
3454 }
3456 }
3457 }
3459 }
3463 }
3468 }
3473 }
3475 /* ARGSUSED */
3476 static int
3479 {
3481 RMDIR3args args;
3482 RMDIR3res res;
3486 hrtime_t t;
3494 /*
3495 * Attempt to prevent a rmdir(".") from succeeding.
3496 */
3501 }
3507 }
3511 /*
3512 * First just remove the entry from the name cache, as it
3513 * is most likely an entry for this vp.
3514 */
3517 /*
3518 * If there vnode reference count is greater than one, then
3519 * there may be additional references in the DNLC which will
3520 * need to be purged. First, trying removing the entry for
3521 * the parent directory and see if that removes the additional
3522 * reference(s). If that doesn't do it, then use dnlc_purge_vp
3523 * to completely remove any references to the directory which
3524 * might still exist in the DNLC.
3525 */
3530 }
3548 }
3560 /*
3561 * System V defines rmdir to return EEXIST, not
3562 * ENOTEMPTY if the directory is not empty. Over
3563 * the wire, the error is NFSERR_ENOTEMPTY which
3564 * geterrno maps to ENOTEMPTY.
3565 */
3568 }
3572 }
3578 }
3580 /* ARGSUSED */
3581 static int
3584 {
3586 SYMLINK3args args;
3587 SYMLINK3res res;
3594 hrtime_t t;
3606 /* req time field(s) overflow - return immediately */
3608 }
3630 }
3648 }
3655 KM_NOSLEEP);
3664 contents;
3670 MAXPATHLEN);
3671 }
3672 }
3673 }
3675 }
3683 }
3684 }
3689 }
3691 #ifdef DEBUG
3697 #endif
3701 /*
3702 * Read directory entries.
3703 * There are some weird things to look out for here. The uio_loffset
3704 * field is either 0 or it is the offset returned from a previous
3705 * readdir. It is an opaque value used by the server to find the
3706 * correct directory block to read. The count field is the number
3707 * of blocks to read on the server. This is advisory only, the server
3708 * may return only one block's worth of entries. Entries may be compressed
3709 * on the server.
3710 */
3711 /* ARGSUSED */
3712 static int
3715 {
3722 #ifdef DEBUG
3724 #endif
3726 rddir_cache srdc;
3727 avl_index_t where;
3735 /*
3736 * Make sure that the directory cache is valid.
3737 */
3740 /*
3741 * Setting nfs_disable_rddir_cache in /etc/system
3742 * allows interoperability with servers that do not
3743 * properly update the attributes of directories.
3744 * Any cached information gets purged before an
3745 * access is made to it.
3746 */
3752 }
3753 }
3755 /*
3756 * It is possible that some servers may not be able to correctly
3757 * handle a large READDIR or READDIRPLUS request due to bugs in
3758 * their implementation. In order to continue to interoperate
3759 * with them, this workaround is provided to limit the maximum
3760 * size of a READDIRPLUS request to 1024. In any case, the request
3761 * size is limited to MAXBSIZE.
3762 */
3767 #ifdef DEBUG
3769 #endif
3770 top:
3771 /*
3772 * Short circuit last readdir which always returns 0 bytes.
3773 * This can be done after the directory has been read through
3774 * completely at least once. This will set r_direof which
3775 * can be used to find the value of the last cookie.
3776 */
3781 #ifdef DEBUG
3782 nfs3_readdir_cache_shorts++;
3783 #endif
3789 }
3790 /*
3791 * Look for a cache entry. Cache entries are identified
3792 * by the NFS cookie value and the byte count requested.
3793 */
3799 /*
3800 * If the cache entry is in the process of being
3801 * filled in, wait until this completes. The
3802 * RDDIRWAIT bit is set to indicate that someone
3803 * is waiting and then the thread currently
3804 * filling the entry is done, it should do a
3805 * cv_broadcast to wakeup all of the threads
3806 * waiting for it to finish.
3807 */
3811 #ifdef DEBUG
3812 nfs3_readdir_cache_waits++;
3813 #endif
3815 /*
3816 * We got interrupted, probably
3817 * the user typed ^C or an alarm
3818 * fired. We free the new entry
3819 * if we allocated one.
3820 */
3828 }
3834 }
3835 /*
3836 * Check to see if a readdir is required to
3837 * fill the entry. If so, mark this entry
3838 * as being filled, remove our reference,
3839 * and branch to the code to fill the entry.
3840 */
3849 }
3850 #ifdef DEBUG
3852 nfs3_readdir_cache_hits++;
3853 #endif
3854 /*
3855 * If an error occurred while attempting
3856 * to fill the cache entry, just return it.
3857 */
3865 }
3867 /*
3868 * The cache entry is complete and good,
3869 * copyout the dirent structs to the calling
3870 * thread.
3871 */
3874 /*
3875 * If no error occurred during the copyout,
3876 * update the offset in the uio struct to
3877 * contain the value of the next cookie
3878 * and set the eof value appropriately.
3879 */
3884 }
3886 /*
3887 * Decide whether to do readahead.
3888 *
3889 * Don't if have already read to the end of
3890 * directory. There is nothing more to read.
3891 *
3892 * Don't if the application is not doing
3893 * lookups in the directory. The readahead
3894 * is only effective if the application can
3895 * be doing work while an async thread is
3896 * handling the over the wire request.
3897 */
3903 else
3912 }
3914 /*
3915 * Check to see whether we found an entry
3916 * for the readahead. If so, we don't need
3917 * to do anything further, so free the new
3918 * entry if one was allocated. Otherwise,
3919 * allocate a new entry, add it to the cache,
3920 * and then initiate an asynchronous readdir
3921 * operation to fill it.
3922 */
3934 }
3942 #ifdef DEBUG
3943 nfs3_readdir_readahead++;
3944 #endif
3947 }
3948 }
3953 }
3955 /*
3956 * Didn't find an entry in the cache. Construct a new empty
3957 * entry and link it into the cache. Other processes attempting
3958 * to access this entry will need to wait until it is filled in.
3959 *
3960 * Since kmem_alloc may block, another pass through the cache
3961 * will need to be taken to make sure that another process
3962 * hasn't already added an entry to the cache for this request.
3963 */
3970 }
3972 /*
3973 * Add this entry to the cache.
3974 */
3979 bottom:
3980 #ifdef DEBUG
3982 nfs3_readdir_cache_misses++;
3983 #endif
3984 /*
3985 * Do the readdir. This routine decides whether to use
3986 * READDIR or READDIRPLUS.
3987 */
3990 /*
3991 * If this operation failed, just return the error which occurred.
3992 */
3996 /*
3997 * Since the RPC operation will have taken sometime and blocked
3998 * this process, another pass through the cache will need to be
3999 * taken to find the correct cache entry. It is possible that
4000 * the correct cache entry will not be there (although one was
4001 * added) because the directory changed during the RPC operation
4002 * and the readdir cache was flushed. In this case, just start
4003 * over. It is hoped that this will not happen too often... :-)
4004 */
4007 /* NOTREACHED */
4008 }
4010 static int
4012 {
4020 /*
4021 * Issue the proper request.
4022 *
4023 * If the server does not support READDIRPLUS, then use READDIR.
4024 *
4025 * Otherwise --
4026 * Issue a READDIRPLUS if reading to fill an empty cache or if
4027 * an application has performed a lookup in the directory which
4028 * required an over the wire lookup. The use of READDIRPLUS
4029 * will help to (re)populate the DNLC.
4030 */
4037 }
4049 }
4058 }
4060 static void
4062 {
4064 READDIR3args args;
4065 READDIR3vres res;
4066 vattr_t dva;
4071 hrtime_t t;
4082 /*
4083 * NFS client failover support
4084 * suppress failover unless we have a zero cookie
4085 */
4093 }
4095 #ifdef DEBUG
4097 #else
4099 #endif
4113 }
4126 }
4137 }
4144 }
4154 err:
4158 }
4160 /*
4161 * Read directory entries.
4162 * There are some weird things to look out for here. The uio_loffset
4163 * field is either 0 or it is the offset returned from a previous
4164 * readdir. It is an opaque value used by the server to find the
4165 * correct directory block to read. The count field is the number
4166 * of blocks to read on the server. This is advisory only, the server
4167 * may return only one block's worth of entries. Entries may be compressed
4168 * on the server.
4169 */
4170 static void
4172 {
4174 READDIRPLUS3args args;
4175 READDIRPLUS3vres res;
4176 vattr_t dva;
4192 /*
4193 * NFS client failover support
4194 * suppress failover unless we have a zero cookie
4195 */
4203 }
4205 #ifdef DEBUG
4207 #else
4209 #endif
4224 }
4237 }
4241 }
4252 }
4254 }
4261 }
4272 err:
4276 }
4278 #ifdef DEBUG
4280 #endif
4282 static int
4284 {
4289 offset_t offset;
4306 }
4308 read_again:
4314 /*
4315 * Didn't get it all because we hit EOF,
4316 * zero all the memory beyond the EOF.
4317 */
4318 /* bzero(rdaddr + */
4321 }
4325 /*
4326 * We didn't read anything at all as we are
4327 * past EOF. Return an error indicator back
4328 * but don't destroy the pages (yet).
4329 */
4331 }
4344 }
4346 }
4358 }
4360 write_again:
4366 #ifdef DEBUG
4370 offset);
4374 }
4375 #endif
4390 }
4392 }
4395 /*
4396 * Don't print EDQUOT errors on the console.
4397 * Don't print asynchronous EACCES errors.
4398 * Don't print EFBIG errors.
4399 * Print all other write errors.
4400 */
4405 /*
4406 * Update r_error and r_flags as appropriate.
4407 * If the error was ESTALE, then mark the
4408 * rnode as not being writeable and save
4409 * the error status. Otherwise, save any
4410 * errors which occur from asynchronous
4411 * page invalidations. Any errors occurring
4412 * from other operations should be saved
4413 * by the caller.
4414 */
4425 }
4427 }
4431 /*
4432 * A close may have cleared r_error, if so,
4433 * propagate ESTALE error return properly
4434 */
4437 }
4438 }
4446 }
4448 /* ARGSUSED */
4449 static int
4451 {
4461 }
4465 }
4467 /* ARGSUSED2 */
4468 static int
4470 {
4476 }
4483 }
4487 }
4489 /* ARGSUSED */
4490 static void
4492 {
4496 }
4498 /* ARGSUSED */
4499 static int
4501 {
4503 /*
4504 * Because we stuff the readdir cookie into the offset field
4505 * someone may attempt to do an lseek with the cookie which
4506 * we want to succeed.
4507 */
4513 }
4515 /*
4516 * number of nfs3_bsize blocks to read ahead.
4517 */
4520 #ifdef DEBUG
4522 #endif
4524 /*
4525 * Return all the pages from [off..off+len) in file
4526 */
4527 /* ARGSUSED */
4528 static int
4532 {
4545 /*
4546 * Now valididate that the caches are up to date.
4547 */
4554 retry:
4557 /*
4558 * Don't create dirty pages faster than they
4559 * can be cleaned so that the system doesn't
4560 * get imbalanced. If the async queue is
4561 * maxed out, then wait for it to drain before
4562 * creating more dirty pages. Also, wait for
4563 * any threads doing pagewalks in the vop_getattr
4564 * entry points so that they don't block for
4565 * long periods.
4566 */
4572 }
4574 /*
4575 * If we are getting called as a side effect of an nfs_write()
4576 * operation the local file size might not be extended yet.
4577 * In this case we want to be able to return pages of zeroes.
4578 */
4582 }
4595 }
4598 }
4600 /*
4601 * Called from pvn_getpages to get a particular page.
4602 */
4603 /* ARGSUSED */
4604 static int
4608 {
4610 uint_t bsize;
4613 uoff_t lbn;
4614 uoff_t io_off;
4615 uoff_t blkoff;
4616 uoff_t rablkoff;
4618 uint_t blksize;
4631 reread:
4643 /*
4644 * Queueing up the readahead before doing the synchronous read
4645 * results in a significant increase in read throughput because
4646 * of the increased parallelism between the async threads and
4647 * the process context.
4648 */
4654 /*
4655 * Calculate the number of readaheads to do.
4656 * a) No readaheads at offset = 0.
4657 * b) Do maximum(nfs3_nra) readaheads when the readahead
4658 * window is closed.
4659 * c) Do readaheads between 1 to (nfs3_nra - 1) depending
4660 * upon how far the readahead window is open or close.
4661 * d) No readaheads if rp->r_nextr is not within the scope
4662 * of the readahead window (random i/o).
4663 */
4673 else
4684 }
4685 readahead--;
4687 /*
4688 * Indicate that we did a readahead so
4689 * readahead offset is not updated
4690 * by the synchronous read below.
4691 */
4694 /*
4695 * set readahead offset to
4696 * offset of last async readahead
4697 * request.
4698 */
4700 }
4702 }
4704 again:
4708 nfs3_readahead);
4710 /*
4711 * Block for this page is not allocated, or the offset
4712 * is beyond the current allocation size, or we're
4713 * allocating a swap slot and the page was not found,
4714 * so allocate it and return a zero page.
4715 */
4724 /*
4725 * Need to go to server to get a BLOCK, exception to
4726 * that being while reading at offset = 0 or doing
4727 * random i/o, in that case read only a PAGE.
4728 */
4732 /*
4733 * If only a block or less is left in
4734 * the file, read all that is remaining.
4735 */
4737 /*
4738 * Trying to access beyond EOF,
4739 * set up to get at least one page.
4740 */
4755 /*
4756 * Some other thread has entered the page,
4757 * so just use it.
4758 */
4762 /*
4763 * Now round the request size up to page boundaries.
4764 * This ensures that the entire page will be
4765 * initialized to zeroes if EOF is encountered.
4766 */
4772 /*
4773 * pageio_setup should have set b_addr to 0. This
4774 * is correct since we want to do I/O on a page
4775 * boundary. bp_mapin will use this addr to calculate
4776 * an offset, and then set b_addr to the kernel virtual
4777 * address it allocated for us.
4778 */
4788 /*
4789 * If doing a write beyond what we believe is EOF,
4790 * don't bother trying to read the pages from the
4791 * server, we'll just zero the pages here. We
4792 * don't check that the rw flag is S_WRITE here
4793 * because some implementations may attempt a
4794 * read access to the buffer before copying data.
4795 */
4803 }
4805 /*
4806 * Unmap the buffer before freeing it.
4807 */
4817 /*
4818 * If doing a write system call just return
4819 * zeroed pages, else user tried to get pages
4820 * beyond EOF, return error. We don't check
4821 * that the rw flag is S_WRITE here because
4822 * some implementations may attempt a read
4823 * access to the buffer before copying data.
4824 */
4827 else
4829 }
4835 }
4836 }
4837 }
4839 out:
4847 }
4852 /*
4853 * Page exists in the cache, acquire the appropriate lock.
4854 * If this fails, start all over again.
4855 */
4857 #ifdef DEBUG
4858 nfs3_lostpage++;
4859 #endif
4861 }
4865 }
4871 }
4873 static void
4876 {
4879 uoff_t io_off;
4891 /*
4892 * If less than a block left in file read less
4893 * than a block.
4894 */
4902 /*
4903 * The isra flag passed to the kluster function is 1, we may have
4904 * gotten a return value of NULL for a variety of reasons (# of free
4905 * pages < minfree, someone entered the page on the vnode etc). In all
4906 * cases, we want to punt on the readahead.
4907 */
4911 /*
4912 * Now round the request size up to page boundaries.
4913 * This ensures that the entire page will be
4914 * initialized to zeroes if EOF is encountered.
4915 */
4921 /*
4922 * pageio_setup should have set b_addr to 0. This is correct since
4923 * we want to do I/O on a page boundary. bp_mapin() will use this addr
4924 * to calculate an offset, and then set b_addr to the kernel virtual
4925 * address it allocated for us.
4926 */
4936 /*
4937 * If doing a write beyond what we believe is EOF, don't bother trying
4938 * to read the pages from the server, we'll just zero the pages here.
4939 * We don't check that the rw flag is S_WRITE here because some
4940 * implementations may attempt a read access to the buffer before
4941 * copying data.
4942 */
4953 }
4955 /*
4956 * Unmap the buffer before freeing it.
4957 */
4968 /*
4969 * In case of error set readahead offset
4970 * to the lowest offset.
4971 * pvn_read_done() calls VN_DISPOSE to destroy the pages
4972 */
4978 }
4979 }
4981 /*
4982 * Flags are composed of {B_INVAL, B_FREE, B_DONTNEED, B_FORCE}
4983 * If len == 0, do from off to EOF.
4984 *
4985 * The normal cases should be len == 0 && off == 0 (entire vp list),
4986 * len == MAXBSIZE (from segmap_release actions), and len == PAGESIZE
4987 * (from pageout).
4988 */
4989 /* ARGSUSED */
4990 static int
4993 {
4999 /*
5000 * XXX - Why should this check be made here?
5001 */
5020 }
5022 /*
5023 * Write out a single page, possibly klustering adjacent dirty pages.
5024 */
5025 int
5028 {
5029 uoff_t io_off;
5030 uoff_t lbn_off;
5031 uoff_t lbn;
5033 uint_t bsize;
5049 /*
5050 * Find a kluster that fits in one block, or in
5051 * one page if pages are bigger than blocks. If
5052 * there is less file space allocated than a whole
5053 * page, we'll shorten the i/o request below.
5054 */
5058 /*
5059 * pvn_write_kluster shouldn't have returned a page with offset
5060 * behind the original page we were given. Verify that.
5061 */
5064 /*
5065 * Now pp will have the list of kept dirty pages marked for
5066 * write back. It will also handle invalidation and freeing
5067 * of pages that are not dirty. Check for page length rounding
5068 * problems.
5069 */
5073 }
5074 /*
5075 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
5076 * consistent value of r_size. RMODINPROGRESS is set in writerp().
5077 * When RMODINPROGRESS is set it indicates that a uiomove() is in
5078 * progress and the r_size has not been made consistent with the
5079 * new size of the file. When the uiomove() completes the r_size is
5080 * updated and the RMODINPROGRESS flag is cleared.
5081 *
5082 * The RMODINPROGRESS flag makes sure that nfs(3)_bio() sees a
5083 * consistent value of r_size. Without this handshaking, it is
5084 * possible that nfs(3)_bio() picks up the old value of r_size
5085 * before the uiomove() in writerp() completes. This will result
5086 * in the write through nfs(3)_bio() being dropped.
5087 *
5088 * More precisely, there is a window between the time the uiomove()
5089 * completes and the time the r_size is updated. If a fop_putpage()
5090 * operation intervenes in this window, the page will be picked up,
5091 * because it is dirty (it will be unlocked, unless it was
5092 * pagecreate'd). When the page is picked up as dirty, the dirty
5093 * bit is reset (pvn_getdirty()). In nfs(3)write(), r_size is
5094 * checked. This will still be the old size. Therefore the page will
5095 * not be written out. When segmap_release() calls fop_putpage(),
5096 * the page will be found to be clean and the write will be dropped.
5097 */
5104 /*
5105 * A write is in progress for this region of the file.
5106 * If we did not detect RMODINPROGRESS here then this
5107 * path through nfs_putapage() would eventually go to
5108 * nfs(3)_bio() and may not write out all of the data
5109 * in the pages. We end up losing data. So we decide
5110 * to set the modified bit on each page in the page
5111 * list and mark the rnode with RDIRTY. This write
5112 * will be restarted at some later time.
5113 */
5121 }
5129 }
5131 }
5135 nfs3_sync_putapage);
5144 }
5146 static int
5149 {
5168 }
5171 /*
5172 * If this was not an async thread, then try again to
5173 * write out the pages, but this time, also destroy
5174 * them whether or not the write is successful. This
5175 * will prevent memory from filling up with these
5176 * pages and destroying them is the only alternative
5177 * if they can't be written out.
5178 *
5179 * Don't do this if this is an async thread because
5180 * when the pages are unlocked in pvn_write_done,
5181 * some other thread could have come along, locked
5182 * them, and queued for an async thread. It would be
5183 * possible for all of the async threads to be tied
5184 * up waiting to lock the pages again and they would
5185 * all already be locked and waiting for an async
5186 * thread to handle them. Deadlock.
5187 */
5191 }
5199 }
5203 }
5206 }
5208 /* ARGSUSED */
5209 static int
5213 {
5231 /*
5232 * If there is cached data and if close-to-open consistency
5233 * checking is not turned off and if the file system is not
5234 * mounted readonly, then force an over the wire getattr.
5235 * Otherwise, just invoke nfs3getattr to get a copy of the
5236 * attributes. The attribute cache will be used unless it
5237 * is timed out and if it is, then an over the wire getattr
5238 * will be issued.
5239 */
5244 else
5249 /*
5250 * Check to see if the vnode is currently marked as not cachable.
5251 * This means portions of the file are locked (through fop_frlock).
5252 * In this case the map request must be refused. We use
5253 * rp->r_lkserlock to avoid a race with concurrent lock requests.
5254 */
5257 /*
5258 * Atomically increment r_inmap after acquiring r_rwlock. The
5259 * idea here is to acquire r_rwlock to block read/write and
5260 * not to protect r_inmap. r_inmap will inform nfs3_read/write()
5261 * that we are in nfs3_map(). Now, r_rwlock is acquired in order
5262 * and we can prevent the deadlock that would have occurred
5263 * when nfs3_addmap() would have acquired it out of order.
5264 *
5265 * Since we are not protecting r_inmap by any lock, we do not
5266 * hold any lock when we decrement it. We atomically decrement
5267 * r_inmap after we release r_lkserlock.
5268 */
5278 }
5283 }
5285 /*
5286 * Don't allow concurrent locks and mapping if mandatory locking is
5287 * enabled.
5288 */
5293 }
5300 }
5316 done:
5320 }
5322 /* ARGSUSED */
5323 static int
5327 {
5339 }
5341 /* ARGSUSED */
5342 static int
5346 {
5347 netobj lm_fh3;
5355 /* check for valid cmd parameter */
5359 /* Verify l_type. */
5375 }
5377 /* check the validity of the lock range */
5383 /*
5384 * If the filesystem is mounted using local locking, pass the
5385 * request off to the local locking code.
5386 */
5389 /*
5390 * For complete safety, we should be holding
5391 * r_lkserlock. However, we can't call
5392 * lm_safelock and then fs_frlock while
5393 * holding r_lkserlock, so just invoke
5394 * lm_safelock and expect that this will
5395 * catch enough of the cases.
5396 */
5399 }
5401 }
5405 /*
5406 * Check whether the given lock request can proceed, given the
5407 * current file mappings.
5408 */
5415 }
5416 }
5418 /*
5419 * Flush the cache after waiting for async I/O to finish. For new
5420 * locks, this is so that the process gets the latest bits from the
5421 * server. For unlocks, this is so that other clients see the
5422 * latest bits once the file has been unlocked. If currently dirty
5423 * pages can't be flushed, then don't allow a lock to be set. But
5424 * allow unlocks to succeed, to avoid having orphan locks on the
5425 * server.
5426 */
5441 }
5446 }
5457 }
5461 }
5462 }
5463 }
5468 /*
5469 * Call the lock manager to do the real work of contacting
5470 * the server and obtaining the lock.
5471 */
5477 done:
5480 }
5482 /*
5483 * Free storage space associated with the specified vnode. The portion
5484 * to be freed is specified by bfp->l_start and bfp->l_len (already
5485 * normalized to a "whence" of 0).
5486 *
5487 * This is an experimental facility whose continued existence is not
5488 * guaranteed. Currently, we only support the special case
5489 * of l_len == 0, meaning free to end of file.
5490 */
5491 /* ARGSUSED */
5492 static int
5495 {
5510 /*
5511 * ftruncate should not change the ctime and
5512 * mtime if we truncate the file to its
5513 * previous size.
5514 */
5527 }
5530 }
5532 /* ARGSUSED */
5533 static int
5535 {
5538 }
5540 /*
5541 * Setup and add an address space callback to do the work of the delmap call.
5542 * The callback will (and must be) deleted in the actual callback function.
5543 *
5544 * This is done in order to take care of the problem that we have with holding
5545 * the address space's a_lock for a long period of time (e.g. if the NFS server
5546 * is down). Callbacks will be executed in the address space code while the
5547 * a_lock is not held. Holding the address space's a_lock causes things such
5548 * as ps and fork to hang because they are trying to acquire this lock as well.
5549 */
5550 /* ARGSUSED */
5551 static int
5555 {
5564 /*
5565 * A process may not change zones if it has NFS pages mmap'ed
5566 * in, so we can't legitimately get here from the wrong zone.
5567 */
5572 /*
5573 * The way that the address space of this process deletes its mapping
5574 * of this file is via the following call chains:
5575 * - as_free()->segop_unmap()/segvn_unmap()->fop_delmap()/nfs3_delmap()
5576 * - as_unmap()->segop_unmap()/segvn_unmap()->fop_delmap()/nfs3_delmap()
5577 *
5578 * With the use of address space callbacks we are allowed to drop the
5579 * address space lock, a_lock, while executing the NFS operations that
5580 * need to go over the wire. Returning EAGAIN to the caller of this
5581 * function is what drives the execution of the callback that we add
5582 * below. The callback will be executed by the address space code
5583 * after dropping the a_lock. When the callback is finished, since
5584 * we dropped the a_lock, it must be re-acquired and segvn_unmap()
5585 * is called again on the same segment to finish the rest of the work
5586 * that needs to happen during unmapping.
5587 *
5588 * This action of calling back into the segment driver causes
5589 * nfs3_delmap() to get called again, but since the callback was
5590 * already executed at this point, it already did the work and there
5591 * is nothing left for us to do.
5592 *
5593 * To Summarize:
5594 * - The first time nfs3_delmap is called by the current thread is when
5595 * we add the caller associated with this delmap to the delmap caller
5596 * list, add the callback, and return EAGAIN.
5597 * - The second time in this call chain when nfs3_delmap is called we
5598 * will find this caller in the delmap caller list and realize there
5599 * is no more work to do thus removing this caller from the list and
5600 * returning the error that was set in the callback execution.
5601 */
5604 /*
5605 * 'error' is from the actual delmap operations. To avoid
5606 * hangs, we need to handle the return of EAGAIN differently
5607 * since this is what drives the callback execution.
5608 * In this case, we don't want to return EAGAIN and do the
5609 * callback execution because there are none to execute.
5610 */
5613 else
5615 }
5617 /* current caller was not in the list */
5640 }
5642 /*
5643 * Remove some pages from an mmap'd vnode. Just update the
5644 * count of pages. If doing close-to-open, then flush and
5645 * commit all of the pages associated with this file.
5646 * Otherwise, start an asynchronous page flush to write out
5647 * any dirty pages. This will also associate a credential
5648 * with the rnode which can be used to write the pages.
5649 */
5650 /* ARGSUSED */
5651 static void
5653 {
5665 /*
5666 * Initiate a page flush and potential commit if there are
5667 * pages, the file system was not mounted readonly, the segment
5668 * was mapped shared, and the pages themselves were writeable.
5669 */
5675 /*
5676 * If this is a cross-zone access a sync putpage won't work, so
5677 * the best we can do is try an async putpage. That seems
5678 * better than something more draconian such as discarding the
5679 * dirty pages.
5680 */
5685 else
5693 }
5704 }
5708 #ifdef DEBUG
5711 #endif
5713 /* ARGSUSED */
5714 static int
5717 {
5719 PATHCONF3args args;
5720 PATHCONF3res res;
5722 failinfo_t fi;
5724 hrtime_t t;
5728 /*
5729 * Large file spec - need to base answer on info stored
5730 * on original FSINFO response.
5731 */
5741 }
5745 }
5748 }
5751 }
5754 }
5757 }
5764 }
5769 }
5800 }
5801 }
5803 }
5811 }
5834 }
5836 #ifdef DEBUG
5837 nfs3_pathconf_cache_hits++;
5838 #endif
5840 }
5842 }
5843 #ifdef DEBUG
5844 nfs3_pathconf_cache_misses++;
5845 #endif
5877 }
5879 }
5899 }
5903 }
5906 }
5908 /*
5909 * Called by async thread to do synchronous pageio. Do the i/o, wait
5910 * for it to complete, and cleanup the page list when done.
5911 */
5912 static int
5915 {
5922 else
5925 }
5927 /* ARGSUSED */
5928 static int
5931 {
5947 nfs3_sync_pageio);
5955 }
5957 /* ARGSUSED */
5958 static void
5961 {
5966 offset3 offset;
5967 count3 len;
5968 k_sigset_t smask;
5970 /*
5971 * We should get called with fl equal to either B_FREE or
5972 * B_INVAL. Any other value is illegal.
5973 *
5974 * The page that we are either supposed to free or destroy
5975 * should be exclusive locked and its io lock should not
5976 * be held.
5977 */
5982 /*
5983 * If the page doesn't need to be committed or we shouldn't
5984 * even bother attempting to commit it, then just make sure
5985 * that the p_fsdata byte is clear and then either free or
5986 * destroy the page as appropriate.
5987 */
5992 else
5995 }
5997 /*
5998 * If there is a page invalidation operation going on, then
5999 * if this is one of the pages being destroyed, then just
6000 * clear the p_fsdata byte and then either free or destroy
6001 * the page as appropriate.
6002 */
6009 else
6012 }
6014 /*
6015 * If we are freeing this page and someone else is already
6016 * waiting to do a commit, then just unlock the page and
6017 * return. That other thread will take care of commiting
6018 * this page. The page can be freed sometime after the
6019 * commit has finished. Otherwise, if the page is marked
6020 * as delay commit, then we may be getting called from
6021 * pvn_write_done, one page at a time. This could result
6022 * in one commit per page, so we end up doing lots of small
6023 * commits instead of fewer larger commits. This is bad,
6024 * we want do as few commits as possible.
6025 */
6031 }
6037 }
6038 }
6040 /*
6041 * Check to see if there is a signal which would prevent an
6042 * attempt to commit the pages from being successful. If so,
6043 * then don't bother with all of the work to gather pages and
6044 * generate the unsuccessful RPC. Just return from here and
6045 * let the page be committed at some later time.
6046 */
6053 }
6056 /*
6057 * We are starting to need to commit pages, so let's try
6058 * to commit as many as possible at once to reduce the
6059 * overhead.
6060 *
6061 * Set the `commit inprogress' state bit. We must
6062 * first wait until any current one finishes. Then
6063 * we initialize the c_pages list with this page.
6064 */
6069 }
6077 /*
6078 * Gather together all other pages which can be committed.
6079 * They will all be chained off r_commit.c_pages.
6080 */
6083 /*
6084 * Clear the `commit inprogress' status and disconnect
6085 * the list of pages to be committed from the rnode.
6086 * At this same time, we also save the starting offset
6087 * and length of data to be committed on the server.
6088 */
6102 }
6104 /*
6105 * Actually generate the COMMIT3 over the wire operation.
6106 */
6109 /*
6110 * If we got an error during the commit, just unlock all
6111 * of the pages. The pages will get retransmitted to the
6112 * server during a putpage operation.
6113 */
6119 }
6121 }
6123 /*
6124 * We've tried as hard as we can to commit the data to stable
6125 * storage on the server. We release the rest of the pages
6126 * and clear the commit required state. They will be put
6127 * onto the tail of the cachelist if they are nolonger
6128 * mapped.
6129 */
6135 }
6137 /*
6138 * It is possible that nfs3_commit didn't return error but
6139 * some other thread has modified the page we are going
6140 * to free/destroy.
6141 * In this case we need to rewrite the page. Do an explicit check
6142 * before attempting to free/destroy the page. If modified, needs to
6143 * be rewritten so unlock the page and return.
6144 */
6149 }
6151 /*
6152 * Now, as appropriate, either free or destroy the page
6153 * that we were called with.
6154 */
6158 else
6160 }
6162 static int
6164 {
6167 COMMIT3args args;
6168 COMMIT3res res;
6184 }
6191 doitagain:
6210 }
6227 }
6229 }
6230 /*
6231 * Can't do a PURGE_STALE_FH here because this
6232 * can cause a deadlock. nfs3_commit can
6233 * be called from nfs3_dispose which can be called
6234 * indirectly via pvn_vplist_dirty. PURGE_STALE_FH
6235 * can call back to pvn_vplist_dirty.
6236 */
6249 }
6250 }
6253 }
6255 static void
6257 {
6261 }
6263 /*
6264 * This routine is used to gather together a page list of the pages
6265 * which are to be committed on the server. This routine must not
6266 * be called if the calling thread holds any locked pages.
6267 *
6268 * The calling thread must have set RCOMMIT. This bit is used to
6269 * serialize access to the commit structure in the rnode. As long
6270 * as the thread has set RCOMMIT, then it can manipulate the commit
6271 * structure without requiring any other locks.
6272 */
6273 static void
6275 {
6285 /*
6286 * Step through all of the pages associated with this vnode
6287 * looking for pages which need to be committed.
6288 */
6292 /* Skip marker pages. */
6296 /*
6297 * If this page does not need to be committed or is
6298 * modified, then just skip it.
6299 */
6303 /*
6304 * Attempt to lock the page. If we can't, then
6305 * someone else is messing with it and we will
6306 * just skip it.
6307 */
6311 /*
6312 * If this page does not need to be committed or is
6313 * modified, then just skip it. Recheck now that
6314 * the page is locked.
6315 */
6319 }
6324 }
6326 /*
6327 * The page needs to be committed and we locked it.
6328 * Update the base and length parameters and add it
6329 * to r_pages.
6330 */
6342 }
6344 }
6347 }
6349 /*
6350 * This routine is used to gather together a page list of the pages
6351 * which are to be committed on the server. This routine must not
6352 * be called if the calling thread holds any locked pages.
6353 *
6354 * The calling thread must have set RCOMMIT. This bit is used to
6355 * serialize access to the commit structure in the rnode. As long
6356 * as the thread has set RCOMMIT, then it can manipulate the commit
6357 * structure without requiring any other locks.
6358 */
6359 static void
6361 {
6365 uoff_t end;
6366 uoff_t off;
6375 /*
6376 * If there are no pages associated with this vnode, then
6377 * just return.
6378 */
6382 /*
6383 * Calculate the ending offset.
6384 */
6388 /*
6389 * Lookup each page by vp, offset.
6390 */
6394 /*
6395 * If this page does not need to be committed or is
6396 * modified, then just skip it.
6397 */
6401 }
6405 /*
6406 * The page needs to be committed and we locked it.
6407 * Update the base and length parameters and add it
6408 * to r_pages.
6409 */
6416 }
6418 }
6419 }
6421 static int
6423 {
6425 writeverf3 write_verf;
6429 /*
6430 * Flush the data portion of the file and then commit any
6431 * portions which need to be committed. This may need to
6432 * be done twice if the server has changed state since
6433 * data was last written. The data will need to be
6434 * rewritten to the server and then a new commit done.
6435 *
6436 * In fact, this may need to be done several times if the
6437 * server is having problems and crashing while we are
6438 * attempting to do this.
6439 */
6441 top:
6442 /*
6443 * Do a flush based on the poff and plen arguments. This
6444 * will asynchronously write out any modified pages in the
6445 * range specified by (poff, plen). This starts all of the
6446 * i/o operations which will be waited for in the next
6447 * call to nfs3_putpage
6448 */
6458 /*
6459 * Do a flush based on the poff and plen arguments. This
6460 * will synchronously write out any modified pages in the
6461 * range specified by (poff, plen) and wait until all of
6462 * the asynchronous i/o's in that range are done as well.
6463 */
6474 }
6477 /*
6478 * Now commit any pages which might need to be committed.
6479 * If the error, NFS_VERF_MISMATCH, is returned, then
6480 * start over with the flush operation.
6481 */
6489 }
6491 static int
6493 {
6496 offset3 offset;
6497 count3 len;
6504 /*
6505 * Set the `commit inprogress' state bit. We must
6506 * first wait until any current one finishes.
6507 */
6513 }
6517 /*
6518 * Gather together all of the pages which need to be
6519 * committed.
6520 */
6523 else
6526 /*
6527 * Clear the `commit inprogress' bit and disconnect the
6528 * page list which was gathered together in nfs3_get_commit.
6529 */
6539 /*
6540 * If any pages need to be committed, commit them and
6541 * then unlock them so that they can be freed some
6542 * time later.
6543 */
6545 /*
6546 * No error occurred during the flush portion
6547 * of this operation, so now attempt to commit
6548 * the data to stable storage on the server.
6549 *
6550 * This will unlock all of the pages on the list.
6551 */
6553 }
6555 }
6557 static int
6560 {
6567 /*
6568 * If we got an error, then just unlock all of the pages
6569 * on the list.
6570 */
6576 }
6578 }
6579 /*
6580 * We've tried as hard as we can to commit the data to stable
6581 * storage on the server. We just unlock the pages and clear
6582 * the commit required state. They will get freed later.
6583 */
6589 }
6592 }
6594 static void
6597 {
6600 }
6602 /* ARGSUSED */
6603 static int
6606 {
6619 }
6622 }
6624 /* ARGSUSED */
6625 static int
6628 {
6641 }
6644 }
6646 /* ARGSUSED */
6647 static int
6650 {
6654 netobj lm_fh3;
6659 /*
6660 * check for valid cmd parameter
6661 */
6665 /*
6666 * Check access permissions
6667 */
6673 /*
6674 * If the filesystem is mounted using local locking, pass the
6675 * request off to the local share code.
6676 */
6686 /*
6687 * If passed an owner that is too large to fit in an
6688 * nfs_owner it is likely a recursive call from the
6689 * lock manager client and pass it straight through. If
6690 * it is not a nfs_owner then simply return an error.
6691 */
6694 NFS_OWNER_MAGIC)
6699 }
6701 }
6702 /*
6703 * Remote share reservations owner is a combination of
6704 * a magic number, hostname, and the local owner
6705 */
6720 }
6725 /*
6726 * NFS client can't store remote locks itself
6727 */
6735 }
6738 }