2 * Copyright (c) 1989, 1991, 1993, 1995
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * Rick Macklem at The University of Guelph.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * @(#)nfs_socket.c 8.5 (Berkeley) 3/30/95
37 * $FreeBSD: src/sys/nfs/nfs_socket.c,v 1.60.2.6 2003/03/26 01:44:46 alfred Exp $
38 * $DragonFly: src/sys/vfs/nfs/nfs_socket.c,v 1.45 2007/05/18 17:05:13 dillon Exp $
42 * Socket operations for use by nfs
45 #include <sys/param.h>
46 #include <sys/systm.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/kernel.h>
52 #include <sys/vnode.h>
53 #include <sys/fcntl.h>
54 #include <sys/protosw.h>
55 #include <sys/resourcevar.h>
56 #include <sys/socket.h>
57 #include <sys/socketvar.h>
58 #include <sys/socketops.h>
59 #include <sys/syslog.h>
60 #include <sys/thread.h>
61 #include <sys/tprintf.h>
62 #include <sys/sysctl.h>
63 #include <sys/signalvar.h>
64 #include <sys/mutex.h>
66 #include <sys/signal2.h>
67 #include <sys/mutex2.h>
69 #include <netinet/in.h>
70 #include <netinet/tcp.h>
71 #include <sys/thread2.h>
77 #include "nfsm_subs.h"
86 * Estimate rto for an nfs rpc sent via. an unreliable datagram.
87 * Use the mean and mean deviation of rtt for the appropriate type of rpc
88 * for the frequent rpcs and a default for the others.
89 * The justification for doing "other" this way is that these rpcs
90 * happen so infrequently that timer est. would probably be stale.
91 * Also, since many of these rpcs are
92 * non-idempotent, a conservative timeout is desired.
93 * getattr, lookup - A+2D
97 #define NFS_RTO(n, t) \
98 ((t) == 0 ? (n)->nm_timeo : \
100 (((((n)->nm_srtt[t-1] + 3) >> 2) + (n)->nm_sdrtt[t-1] + 1) >> 1) : \
101 ((((n)->nm_srtt[t-1] + 7) >> 3) + (n)->nm_sdrtt[t-1] + 1)))
102 #define NFS_SRTT(r) (r)->r_nmp->nm_srtt[proct[(r)->r_procnum] - 1]
103 #define NFS_SDRTT(r) (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum] - 1]
106 * Defines which timer to use for the procnum.
113 static int proct[NFS_NPROCS] = {
114 0, 1, 0, 2, 1, 3, 3, 4, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 0, 0, 0, 0, 0,
118 static int nfs_realign_test;
119 static int nfs_realign_count;
120 static int nfs_bufpackets = 4;
121 static int nfs_timer_raced;
123 SYSCTL_DECL(_vfs_nfs);
125 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0, "");
126 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0, "");
127 SYSCTL_INT(_vfs_nfs, OID_AUTO, bufpackets, CTLFLAG_RW, &nfs_bufpackets, 0, "");
131 * There is a congestion window for outstanding rpcs maintained per mount
132 * point. The cwnd size is adjusted in roughly the way that:
133 * Van Jacobson, Congestion avoidance and Control, In "Proceedings of
134 * SIGCOMM '88". ACM, August 1988.
135 * describes for TCP. The cwnd size is chopped in half on a retransmit timeout
136 * and incremented by 1/cwnd when each rpc reply is received and a full cwnd
137 * of rpcs is in progress.
138 * (The sent count and cwnd are scaled for integer arith.)
139 * Variants of "slow start" were tried and were found to be too much of a
140 * performance hit (ave. rtt 3 times larger),
141 * I suspect due to the large rtt that nfs rpcs have.
143 #define NFS_CWNDSCALE 256
144 #define NFS_MAXCWND (NFS_CWNDSCALE * 32)
145 static int nfs_backoff[8] = { 2, 4, 8, 16, 32, 64, 128, 256, };
147 struct nfsrtt nfsrtt;
148 struct callout nfs_timer_handle;
150 static int nfs_msg (struct thread *,char *,char *);
151 static int nfs_rcvlock (struct nfsreq *);
152 static void nfs_rcvunlock (struct nfsreq *);
153 static void nfs_realign (struct mbuf **pm, int hsiz);
154 static int nfs_receive (struct nfsreq *rep, struct sockaddr **aname,
156 static void nfs_softterm (struct nfsreq *rep);
157 static int nfs_reconnect (struct nfsreq *rep);
159 static int nfsrv_getstream (struct nfssvc_sock *, int, int *);
160 static void nfs_timer_req(struct nfsreq *req);
162 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
163 struct nfssvc_sock *slp,
165 struct mbuf **mreqp) = {
193 #endif /* NFS_NOSERVER */
196 * Initialize sockets and congestion for a new NFS connection.
197 * We do not free the sockaddr if error.
200 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
203 int error, rcvreserve, sndreserve;
205 struct sockaddr *saddr;
206 struct sockaddr_in *sin;
207 struct thread *td = &thread0; /* only used for socreate and sobind */
211 error = socreate(saddr->sa_family, &nmp->nm_so, nmp->nm_sotype,
212 nmp->nm_soproto, td);
216 nmp->nm_soflags = so->so_proto->pr_flags;
219 * Some servers require that the client port be a reserved port number.
221 if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
224 struct sockaddr_in ssin;
226 bzero(&sopt, sizeof sopt);
227 ip = IP_PORTRANGE_LOW;
228 sopt.sopt_level = IPPROTO_IP;
229 sopt.sopt_name = IP_PORTRANGE;
230 sopt.sopt_val = (void *)&ip;
231 sopt.sopt_valsize = sizeof(ip);
233 error = sosetopt(so, &sopt);
236 bzero(&ssin, sizeof ssin);
238 sin->sin_len = sizeof (struct sockaddr_in);
239 sin->sin_family = AF_INET;
240 sin->sin_addr.s_addr = INADDR_ANY;
241 sin->sin_port = htons(0);
242 error = sobind(so, (struct sockaddr *)sin, td);
245 bzero(&sopt, sizeof sopt);
246 ip = IP_PORTRANGE_DEFAULT;
247 sopt.sopt_level = IPPROTO_IP;
248 sopt.sopt_name = IP_PORTRANGE;
249 sopt.sopt_val = (void *)&ip;
250 sopt.sopt_valsize = sizeof(ip);
252 error = sosetopt(so, &sopt);
258 * Protocols that do not require connections may be optionally left
259 * unconnected for servers that reply from a port other than NFS_PORT.
261 if (nmp->nm_flag & NFSMNT_NOCONN) {
262 if (nmp->nm_soflags & PR_CONNREQUIRED) {
267 error = soconnect(so, nmp->nm_nam, td);
272 * Wait for the connection to complete. Cribbed from the
273 * connect system call but with the wait timing out so
274 * that interruptible mounts don't hang here for a long time.
277 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
278 (void) tsleep((caddr_t)&so->so_timeo, 0,
280 if ((so->so_state & SS_ISCONNECTING) &&
281 so->so_error == 0 && rep &&
282 (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
283 so->so_state &= ~SS_ISCONNECTING;
289 error = so->so_error;
296 so->so_rcv.ssb_timeo = (5 * hz);
297 so->so_snd.ssb_timeo = (5 * hz);
300 * Get buffer reservation size from sysctl, but impose reasonable
303 pktscale = nfs_bufpackets;
309 if (nmp->nm_sotype == SOCK_DGRAM) {
310 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
311 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
312 NFS_MAXPKTHDR) * pktscale;
313 } else if (nmp->nm_sotype == SOCK_SEQPACKET) {
314 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
315 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
316 NFS_MAXPKTHDR) * pktscale;
318 if (nmp->nm_sotype != SOCK_STREAM)
319 panic("nfscon sotype");
320 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
324 bzero(&sopt, sizeof sopt);
325 sopt.sopt_level = SOL_SOCKET;
326 sopt.sopt_name = SO_KEEPALIVE;
327 sopt.sopt_val = &val;
328 sopt.sopt_valsize = sizeof val;
332 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
336 bzero(&sopt, sizeof sopt);
337 sopt.sopt_level = IPPROTO_TCP;
338 sopt.sopt_name = TCP_NODELAY;
339 sopt.sopt_val = &val;
340 sopt.sopt_valsize = sizeof val;
344 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR +
345 sizeof (u_int32_t)) * pktscale;
346 rcvreserve = (nmp->nm_rsize + NFS_MAXPKTHDR +
347 sizeof (u_int32_t)) * pktscale;
349 error = soreserve(so, sndreserve, rcvreserve,
350 &td->td_proc->p_rlimit[RLIMIT_SBSIZE]);
353 so->so_rcv.ssb_flags |= SSB_NOINTR;
354 so->so_snd.ssb_flags |= SSB_NOINTR;
356 /* Initialize other non-zero congestion variables */
357 nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] =
358 nmp->nm_srtt[3] = (NFS_TIMEO << 3);
359 nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
360 nmp->nm_sdrtt[3] = 0;
361 nmp->nm_cwnd = NFS_MAXCWND / 2; /* Initial send window */
363 nmp->nm_timeouts = 0;
373 * Called when a connection is broken on a reliable protocol.
374 * - clean up the old socket
375 * - nfs_connect() again
376 * - set R_MUSTRESEND for all outstanding requests on mount point
377 * If this fails the mount point is DEAD!
378 * nb: Must be called with the nfs_sndlock() set on the mount point.
381 nfs_reconnect(struct nfsreq *rep)
384 struct nfsmount *nmp = rep->r_nmp;
388 while ((error = nfs_connect(nmp, rep)) != 0) {
389 if (error == EINTR || error == ERESTART)
391 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
395 * Loop through outstanding request list and fix up all requests
399 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
400 KKASSERT(req->r_nmp == nmp);
401 req->r_flags |= R_MUSTRESEND;
408 * NFS disconnect. Clean up and unlink.
411 nfs_disconnect(struct nfsmount *nmp)
418 soshutdown(so, SHUT_RDWR);
419 soclose(so, FNONBLOCK);
424 nfs_safedisconnect(struct nfsmount *nmp)
426 struct nfsreq dummyreq;
428 bzero(&dummyreq, sizeof(dummyreq));
429 dummyreq.r_nmp = nmp;
430 dummyreq.r_td = NULL;
431 mtx_link_init(&dummyreq.r_link);
432 nfs_rcvlock(&dummyreq);
434 nfs_rcvunlock(&dummyreq);
438 * This is the nfs send routine. For connection based socket types, it
439 * must be called with an nfs_sndlock() on the socket.
440 * "rep == NULL" indicates that it has been called from a server.
441 * For the client side:
442 * - return EINTR if the RPC is terminated, 0 otherwise
443 * - set R_MUSTRESEND if the send fails for any reason
444 * - do any cleanup required by recoverable socket errors (?)
445 * For the server side:
446 * - return EINTR or ERESTART if interrupted by a signal
447 * - return EPIPE if a connection is lost for connection based sockets (TCP...)
448 * - do any cleanup required by recoverable socket errors (?)
451 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
454 struct sockaddr *sendnam;
455 int error, soflags, flags;
458 if (rep->r_flags & R_SOFTTERM) {
462 if ((so = rep->r_nmp->nm_so) == NULL) {
463 rep->r_flags |= R_MUSTRESEND;
467 rep->r_flags &= ~R_MUSTRESEND;
468 soflags = rep->r_nmp->nm_soflags;
470 soflags = so->so_proto->pr_flags;
471 if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
475 if (so->so_type == SOCK_SEQPACKET)
480 error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
483 * ENOBUFS for dgram sockets is transient and non fatal.
484 * No need to log, and no need to break a soft mount.
486 if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
488 if (rep) /* do backoff retransmit on client */
489 rep->r_flags |= R_MUSTRESEND;
494 log(LOG_INFO, "nfs send error %d for server %s\n",error,
495 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
497 * Deal with errors for the client side.
499 if (rep->r_flags & R_SOFTTERM)
502 rep->r_flags |= R_MUSTRESEND;
504 log(LOG_INFO, "nfsd send error %d\n", error);
507 * Handle any recoverable (soft) socket errors here. (?)
509 if (error != EINTR && error != ERESTART &&
510 error != EWOULDBLOCK && error != EPIPE)
517 * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
518 * done by soreceive(), but for SOCK_STREAM we must deal with the Record
519 * Mark and consolidate the data into a new mbuf list.
520 * nb: Sometimes TCP passes the data up to soreceive() in long lists of
522 * For SOCK_STREAM we must be very careful to read an entire record once
523 * we have read any of it, even if the system call has been interrupted.
526 nfs_receive(struct nfsreq *rep, struct sockaddr **aname, struct mbuf **mp)
533 struct mbuf *control;
535 struct sockaddr **getnam;
536 int error, sotype, rcvflg;
537 struct thread *td = curthread; /* XXX */
540 * Set up arguments for soreceive()
544 sotype = rep->r_nmp->nm_sotype;
547 * For reliable protocols, lock against other senders/receivers
548 * in case a reconnect is necessary.
549 * For SOCK_STREAM, first get the Record Mark to find out how much
550 * more there is to get.
551 * We must lock the socket against other receivers
552 * until we have an entire rpc request/reply.
554 if (sotype != SOCK_DGRAM) {
555 error = nfs_sndlock(rep);
560 * Check for fatal errors and resending request.
563 * Ugh: If a reconnect attempt just happened, nm_so
564 * would have changed. NULL indicates a failed
565 * attempt that has essentially shut down this
568 if (rep->r_mrep || (rep->r_flags & R_SOFTTERM)) {
572 so = rep->r_nmp->nm_so;
574 error = nfs_reconnect(rep);
581 while (rep->r_flags & R_MUSTRESEND) {
582 m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
583 nfsstats.rpcretries++;
584 error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
586 if (error == EINTR || error == ERESTART ||
587 (error = nfs_reconnect(rep)) != 0) {
595 if (sotype == SOCK_STREAM) {
597 * Get the length marker from the stream
599 aio.iov_base = (caddr_t)&len;
600 aio.iov_len = sizeof(u_int32_t);
603 auio.uio_segflg = UIO_SYSSPACE;
604 auio.uio_rw = UIO_READ;
606 auio.uio_resid = sizeof(u_int32_t);
609 rcvflg = MSG_WAITALL;
610 error = so_pru_soreceive(so, NULL, &auio, NULL,
612 if (error == EWOULDBLOCK && rep) {
613 if (rep->r_flags & R_SOFTTERM)
616 } while (error == EWOULDBLOCK);
618 if (error == 0 && auio.uio_resid > 0) {
620 * Only log short packets if not EOF
622 if (auio.uio_resid != sizeof(u_int32_t))
624 "short receive (%d/%d) from nfs server %s\n",
625 (int)(sizeof(u_int32_t) - auio.uio_resid),
626 (int)sizeof(u_int32_t),
627 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
632 len = ntohl(len) & ~0x80000000;
634 * This is SERIOUS! We are out of sync with the sender
635 * and forcing a disconnect/reconnect is all I can do.
637 if (len > NFS_MAXPACKET) {
638 log(LOG_ERR, "%s (%d) from nfs server %s\n",
639 "impossible packet length",
641 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
647 * Get the rest of the packet as an mbuf chain
651 rcvflg = MSG_WAITALL;
652 error = so_pru_soreceive(so, NULL, NULL, &sio,
654 } while (error == EWOULDBLOCK || error == EINTR ||
656 if (error == 0 && sio.sb_cc != len) {
659 "short receive (%d/%d) from nfs server %s\n",
660 len - auio.uio_resid, len,
661 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
667 * Non-stream, so get the whole packet by not
668 * specifying MSG_WAITALL and by specifying a large
671 * We have no use for control msg., but must grab them
672 * and then throw them away so we know what is going
675 sbinit(&sio, 100000000);
678 error = so_pru_soreceive(so, NULL, NULL, &sio,
682 if (error == EWOULDBLOCK && rep) {
683 if (rep->r_flags & R_SOFTTERM) {
688 } while (error == EWOULDBLOCK ||
689 (error == 0 && sio.sb_mb == NULL && control));
690 if ((rcvflg & MSG_EOR) == 0)
692 if (error == 0 && sio.sb_mb == NULL)
698 if (error && error != EINTR && error != ERESTART) {
701 if (error != EPIPE) {
703 "receive error %d from nfs server %s\n",
705 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
707 error = nfs_sndlock(rep);
709 error = nfs_reconnect(rep);
717 if ((so = rep->r_nmp->nm_so) == NULL)
719 if (so->so_state & SS_ISCONNECTED)
723 sbinit(&sio, 100000000);
726 error = so_pru_soreceive(so, getnam, NULL, &sio,
728 if (error == EWOULDBLOCK &&
729 (rep->r_flags & R_SOFTTERM)) {
733 } while (error == EWOULDBLOCK);
742 * Search for any mbufs that are not a multiple of 4 bytes long
743 * or with m_data not longword aligned.
744 * These could cause pointer alignment problems, so copy them to
745 * well aligned mbufs.
747 nfs_realign(mp, 5 * NFSX_UNSIGNED);
752 * Implement receipt of reply on a socket.
753 * We must search through the list of received datagrams matching them
754 * with outstanding requests using the xid, until ours is found.
758 nfs_reply(struct nfsreq *myrep)
761 struct nfsmount *nmp = myrep->r_nmp;
762 struct sockaddr *nam;
766 struct nfsm_info info;
770 * Loop around until we get our own reply
774 * Lock against other receivers so that I don't get stuck in
775 * sbwait() after someone else has received my reply for me.
776 * Also necessary for connection based protocols to avoid
777 * race conditions during a reconnect.
779 * If nfs_rcvlock() returns EALREADY, that means that
780 * the reply has already been recieved by another
781 * process and we can return immediately. In this
782 * case, the lock is not taken to avoid races with
787 error = nfs_rcvlock(myrep);
788 if (error == EALREADY)
793 * Get the next Rpc reply off the socket
795 error = nfs_receive(myrep, &nam, &info.mrep);
796 nfs_rcvunlock(myrep);
799 * Ignore routing errors on connectionless protocols??
801 if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
802 nmp->nm_so->so_error = 0;
803 if (myrep->r_flags & R_GETONEREP)
813 * Get the xid and check that it is an rpc reply
816 info.dpos = mtod(info.md, caddr_t);
817 NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
819 if (*tl != rpc_reply) {
820 nfsstats.rpcinvalid++;
824 if (myrep->r_flags & R_GETONEREP)
830 * Loop through the request list to match up the reply
831 * Iff no match, just drop the datagram. On match, set
832 * r_mrep atomically to prevent the timer from messing
833 * around with the request after we have exited the critical
837 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
838 if (rep->r_mrep == NULL && rxid == rep->r_xid)
844 * Fill in the rest of the reply if we found a match.
848 rep->r_dpos = info.dpos;
852 rt = &nfsrtt.rttl[nfsrtt.pos];
853 rt->proc = rep->r_procnum;
854 rt->rto = NFS_RTO(nmp, proct[rep->r_procnum]);
855 rt->sent = nmp->nm_sent;
856 rt->cwnd = nmp->nm_cwnd;
857 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
858 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
859 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
860 getmicrotime(&rt->tstamp);
861 if (rep->r_flags & R_TIMING)
862 rt->rtt = rep->r_rtt;
865 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
868 * Update congestion window.
869 * Do the additive increase of
872 if (nmp->nm_cwnd <= nmp->nm_sent) {
874 (NFS_CWNDSCALE * NFS_CWNDSCALE +
875 (nmp->nm_cwnd >> 1)) / nmp->nm_cwnd;
876 if (nmp->nm_cwnd > NFS_MAXCWND)
877 nmp->nm_cwnd = NFS_MAXCWND;
879 crit_enter(); /* nfs_timer interlock for nm_sent */
880 if (rep->r_flags & R_SENT) {
881 rep->r_flags &= ~R_SENT;
882 nmp->nm_sent -= NFS_CWNDSCALE;
886 * Update rtt using a gain of 0.125 on the mean
887 * and a gain of 0.25 on the deviation.
889 if (rep->r_flags & R_TIMING) {
891 * Since the timer resolution of
892 * NFS_HZ is so course, it can often
893 * result in r_rtt == 0. Since
894 * r_rtt == N means that the actual
895 * rtt is between N+dt and N+2-dt ticks,
899 t1 -= (NFS_SRTT(rep) >> 3);
903 t1 -= (NFS_SDRTT(rep) >> 2);
904 NFS_SDRTT(rep) += t1;
906 nmp->nm_timeouts = 0;
907 rep->r_mrep = info.mrep;
908 mtx_abort_ex_link(&rep->r_nmp->nm_rxlock, &rep->r_link);
911 * If not matched to a request, drop it.
912 * If it's mine, get out.
915 nfsstats.rpcunexpected++;
918 } else if (rep == myrep) {
919 if (rep->r_mrep == NULL)
920 panic("nfsreply nil");
923 if (myrep->r_flags & R_GETONEREP)
929 nfs_request(struct vnode *vp, struct mbuf *mrest, int procnum,
930 struct thread *td, struct ucred *cred, struct mbuf **mrp,
931 struct mbuf **mdp, caddr_t *dposp)
933 struct nfsreq *rep = NULL;
936 error = nfs_request_setup(vp, mrest, procnum, td, cred, &rep);
941 rep->r_dposp = dposp;
943 error = nfs_request_auth(rep);
947 error = nfs_request_try(rep); /* error ignored */
948 error = nfs_request_waitreply(rep); /* pass to process */
949 error = nfs_request_processreply(rep, error);
950 if (error == ENEEDAUTH)
958 * nfs_request - goes something like this
959 * - fill in request struct
960 * - links it into list
961 * - calls nfs_send() for first transmit
962 * - calls nfs_receive() to get reply
963 * - break down rpc header and return with nfs reply pointed to
965 * nb: always frees up mreq mbuf list
968 nfs_request_setup(struct vnode *vp, struct mbuf *mrest, int procnum,
969 struct thread *td, struct ucred *cred,
970 struct nfsreq **repp)
973 struct nfsmount *nmp;
977 /* Reject requests while attempting a forced unmount. */
978 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
982 nmp = VFSTONFS(vp->v_mount);
983 MALLOC(rep, struct nfsreq *, sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
987 rep->r_procnum = procnum;
995 rep->r_mrest = mrest;
996 rep->r_mrest_len = i;
1003 nfs_request_auth(struct nfsreq *rep)
1005 struct nfsmount *nmp = rep->r_nmp;
1007 char nickv[RPCX_NICKVERF];
1008 int error = 0, auth_len, auth_type;
1011 char *auth_str, *verf_str;
1015 rep->r_failed_auth = 0;
1018 * Get the RPC header with authorization.
1020 verf_str = auth_str = NULL;
1021 if (nmp->nm_flag & NFSMNT_KERB) {
1023 verf_len = sizeof (nickv);
1024 auth_type = RPCAUTH_KERB4;
1025 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1026 if (rep->r_failed_auth ||
1027 nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1028 verf_str, verf_len)) {
1029 error = nfs_getauth(nmp, rep, cred, &auth_str,
1030 &auth_len, verf_str, &verf_len, rep->r_key);
1032 m_freem(rep->r_mrest);
1033 rep->r_mrest = NULL;
1034 kfree((caddr_t)rep, M_NFSREQ);
1039 auth_type = RPCAUTH_UNIX;
1040 if (cred->cr_ngroups < 1)
1041 panic("nfsreq nogrps");
1042 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1043 nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1046 m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1047 auth_len, auth_str, verf_len, verf_str,
1048 rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1049 rep->r_mrest = NULL;
1051 kfree(auth_str, M_TEMP);
1054 * For stream protocols, insert a Sun RPC Record Mark.
1056 if (nmp->nm_sotype == SOCK_STREAM) {
1057 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1059 kfree(rep, M_NFSREQ);
1062 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1063 (m->m_pkthdr.len - NFSX_UNSIGNED));
1071 nfs_request_try(struct nfsreq *rep)
1073 struct nfsmount *nmp = rep->r_nmp;
1077 if (nmp->nm_flag & NFSMNT_SOFT)
1078 rep->r_retry = nmp->nm_retry;
1080 rep->r_retry = NFS_MAXREXMIT + 1; /* past clip limit */
1081 rep->r_rtt = rep->r_rexmit = 0;
1082 if (proct[rep->r_procnum] > 0)
1083 rep->r_flags = R_TIMING | R_MASKTIMER;
1085 rep->r_flags = R_MASKTIMER;
1089 * Do the client side RPC.
1091 nfsstats.rpcrequests++;
1094 * Chain request into list of outstanding requests. Be sure
1095 * to put it LAST so timer finds oldest requests first. Note
1096 * that R_MASKTIMER is set at the moment to prevent any timer
1097 * action on this request while we are still doing processing on
1098 * it below. splsoftclock() primarily protects nm_sent. Note
1099 * that we may block in this code so there is no atomicy guarentee.
1102 TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1103 mtx_link_init(&rep->r_link);
1108 * If backing off another request or avoiding congestion, don't
1109 * send this one now but let timer do it. If not timing a request,
1112 * Even though the timer will not mess with our request there is
1113 * still the possibility that we will race a reply (which clears
1114 * R_SENT), especially on localhost connections, so be very careful
1115 * when setting R_SENT. We could set R_SENT prior to calling
1116 * nfs_send() but why bother if the response occurs that quickly?
1118 if (nmp->nm_so && (nmp->nm_sotype != SOCK_DGRAM ||
1119 (nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1120 nmp->nm_sent < nmp->nm_cwnd)) {
1121 if (nmp->nm_soflags & PR_CONNREQUIRED)
1122 error = nfs_sndlock(rep);
1124 m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1125 error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1126 if (nmp->nm_soflags & PR_CONNREQUIRED)
1129 if (!error && (rep->r_flags & R_MUSTRESEND) == 0 &&
1130 rep->r_mrep == NULL) {
1131 KASSERT((rep->r_flags & R_SENT) == 0,
1132 ("R_SENT ASSERT %p", rep));
1133 nmp->nm_sent += NFS_CWNDSCALE;
1134 rep->r_flags |= R_SENT;
1142 * Let the timer do what it will with the request, then
1143 * wait for the reply from our send or the timer's.
1146 rep->r_flags &= ~R_MASKTIMER;
1152 nfs_request_waitreply(struct nfsreq *rep)
1154 struct nfsmount *nmp = rep->r_nmp;
1158 error = nfs_reply(rep);
1162 * RPC done, unlink the request, but don't rip it out from under
1163 * the callout timer.
1165 while (rep->r_flags & R_LOCKED) {
1166 nfs_timer_raced = 1;
1167 tsleep(&nfs_timer_raced, 0, "nfstrac", 0);
1169 TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1172 * Decrement the outstanding request count.
1174 if (rep->r_flags & R_SENT) {
1175 rep->r_flags &= ~R_SENT;
1176 nmp->nm_sent -= NFS_CWNDSCALE;
1184 * Process reply with error returned from nfs_requet_waitreply().
1186 * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1187 * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1190 nfs_request_processreply(struct nfsreq *rep, int error)
1192 struct nfsmount *nmp = rep->r_nmp;
1195 int trylater_delay = 15, trylater_cnt = 0;
1198 struct nfsm_info info;
1201 * If there was a successful reply and a tprintf msg.
1202 * tprintf a response.
1204 if (!error && (rep->r_flags & R_TPRINTFMSG))
1205 nfs_msg(rep->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1207 info.mrep = rep->r_mrep;
1208 info.md = rep->r_md;
1209 info.dpos = rep->r_dpos;
1211 m_freem(rep->r_mreq);
1212 kfree((caddr_t)rep, M_NFSREQ);
1217 * break down the rpc header and check if ok
1219 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
1220 if (*tl++ == rpc_msgdenied) {
1221 if (*tl == rpc_mismatch) {
1223 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1224 *tl++ == rpc_autherr) {
1225 if (!rep->r_failed_auth) {
1226 rep->r_failed_auth++;
1227 rep->r_mheadend->m_next = NULL;
1230 m_freem(rep->r_mreq);
1240 m_freem(rep->r_mreq);
1241 kfree((caddr_t)rep, M_NFSREQ);
1246 * Grab any Kerberos verifier, otherwise just throw it away.
1248 verf_type = fxdr_unsigned(int, *tl++);
1249 i = fxdr_unsigned(int32_t, *tl);
1250 if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1251 error = nfs_savenickauth(nmp, rep->r_cred, i, rep->r_key,
1252 &info.md, &info.dpos, info.mrep);
1256 ERROROUT(nfsm_adv(&info, nfsm_rndup(i)));
1258 NULLOUT(tl = nfsm_dissect(&info, NFSX_UNSIGNED));
1261 NULLOUT(tl = nfsm_dissect(&info, NFSX_UNSIGNED));
1263 error = fxdr_unsigned(int, *tl);
1264 if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1265 error == NFSERR_TRYLATER) {
1269 waituntil = time_second + trylater_delay;
1270 while (time_second < waituntil)
1271 (void) tsleep((caddr_t)&lbolt,
1273 trylater_delay *= nfs_backoff[trylater_cnt];
1274 if (trylater_cnt < 7)
1276 rep->r_flags &= ~R_MASKTIMER;
1277 return (EAGAIN); /* goto tryagain */
1281 * If the File Handle was stale, invalidate the
1282 * lookup cache, just in case.
1284 * To avoid namecache<->vnode deadlocks we must
1285 * release the vnode lock if we hold it.
1287 if (error == ESTALE) {
1288 struct vnode *vp = rep->r_vp;
1291 ltype = lockstatus(&vp->v_lock, curthread);
1292 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1293 lockmgr(&vp->v_lock, LK_RELEASE);
1294 cache_inval_vp(vp, CINV_CHILDREN);
1295 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1296 lockmgr(&vp->v_lock, ltype);
1298 if (nmp->nm_flag & NFSMNT_NFSV3) {
1299 *rep->r_mrp = info.mrep;
1300 *rep->r_mdp = info.md;
1301 *rep->r_dposp = info.dpos;
1302 error |= NFSERR_RETERR;
1307 m_freem(rep->r_mreq);
1308 kfree((caddr_t)rep, M_NFSREQ);
1312 *rep->r_mrp = info.mrep;
1313 *rep->r_mdp = info.md;
1314 *rep->r_dposp = info.dpos;
1315 m_freem(rep->r_mreq);
1316 FREE((caddr_t)rep, M_NFSREQ);
1321 error = EPROTONOSUPPORT;
1323 m_freem(rep->r_mreq);
1324 kfree((caddr_t)rep, M_NFSREQ);
1328 #ifndef NFS_NOSERVER
1330 * Generate the rpc reply header
1331 * siz arg. is used to decide if adding a cluster is worthwhile
1334 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1335 int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1338 struct nfsm_info info;
1340 siz += RPC_REPLYSIZ;
1341 info.mb = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1342 info.mreq = info.mb;
1343 info.mreq->m_pkthdr.len = 0;
1345 * If this is not a cluster, try and leave leading space
1346 * for the lower level headers.
1348 if ((max_hdr + siz) < MINCLSIZE)
1349 info.mreq->m_data += max_hdr;
1350 tl = mtod(info.mreq, u_int32_t *);
1351 info.mreq->m_len = 6 * NFSX_UNSIGNED;
1352 info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1353 *tl++ = txdr_unsigned(nd->nd_retxid);
1355 if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1356 *tl++ = rpc_msgdenied;
1357 if (err & NFSERR_AUTHERR) {
1358 *tl++ = rpc_autherr;
1359 *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1360 info.mreq->m_len -= NFSX_UNSIGNED;
1361 info.bpos -= NFSX_UNSIGNED;
1363 *tl++ = rpc_mismatch;
1364 *tl++ = txdr_unsigned(RPC_VER2);
1365 *tl = txdr_unsigned(RPC_VER2);
1368 *tl++ = rpc_msgaccepted;
1371 * For Kerberos authentication, we must send the nickname
1372 * verifier back, otherwise just RPCAUTH_NULL.
1374 if (nd->nd_flag & ND_KERBFULL) {
1375 struct nfsuid *nuidp;
1376 struct timeval ktvin, ktvout;
1378 for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1379 nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1380 if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1381 (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1382 &nuidp->nu_haddr, nd->nd_nam2)))
1387 txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1389 txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1392 * Encrypt the timestamp in ecb mode using the
1399 *tl++ = rpc_auth_kerb;
1400 *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1401 *tl = ktvout.tv_sec;
1402 tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1403 *tl++ = ktvout.tv_usec;
1404 *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1415 *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1418 *tl = txdr_unsigned(RPC_PROGMISMATCH);
1419 tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1420 *tl++ = txdr_unsigned(2);
1421 *tl = txdr_unsigned(3);
1424 *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1427 *tl = txdr_unsigned(RPC_GARBAGE);
1431 if (err != NFSERR_RETVOID) {
1432 tl = nfsm_build(&info, NFSX_UNSIGNED);
1434 *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1446 if (err != 0 && err != NFSERR_RETVOID)
1447 nfsstats.srvrpc_errs++;
1452 #endif /* NFS_NOSERVER */
1455 * Scan the nfsreq list and retranmit any requests that have timed out
1456 * To avoid retransmission attempts on STREAM sockets (in the future) make
1457 * sure to set the r_retry field to 0 (implies nm_retry == 0).
1460 nfs_timer(void *arg /* never used */)
1462 struct nfsmount *nmp;
1464 #ifndef NFS_NOSERVER
1465 struct nfssvc_sock *slp;
1467 #endif /* NFS_NOSERVER */
1470 TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1471 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1472 KKASSERT(nmp == req->r_nmp);
1474 (req->r_flags & (R_SOFTTERM|R_MASKTIMER))) {
1477 req->r_flags |= R_LOCKED;
1478 if (nfs_sigintr(nmp, req, req->r_td)) {
1483 req->r_flags &= ~R_LOCKED;
1486 #ifndef NFS_NOSERVER
1489 * Scan the write gathering queues for writes that need to be
1492 cur_usec = nfs_curusec();
1493 TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1494 if (slp->ns_tq.lh_first && slp->ns_tq.lh_first->nd_time<=cur_usec)
1495 nfsrv_wakenfsd(slp, 1);
1497 #endif /* NFS_NOSERVER */
1500 * Due to possible blocking, a client operation may be waiting for
1501 * us to finish processing this request so it can remove it.
1503 if (nfs_timer_raced) {
1504 nfs_timer_raced = 0;
1505 wakeup(&nfs_timer_raced);
1508 callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer, NULL);
1513 nfs_timer_req(struct nfsreq *req)
1515 struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1516 struct nfsmount *nmp = req->r_nmp;
1522 if (req->r_rtt >= 0) {
1524 if (nmp->nm_flag & NFSMNT_DUMBTIMR)
1525 timeo = nmp->nm_timeo;
1527 timeo = NFS_RTO(nmp, proct[req->r_procnum]);
1528 if (nmp->nm_timeouts > 0)
1529 timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1530 if (req->r_rtt <= timeo)
1532 if (nmp->nm_timeouts < 8)
1536 * Check for server not responding
1538 if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1539 req->r_rexmit > nmp->nm_deadthresh) {
1541 nmp->nm_mountp->mnt_stat.f_mntfromname,
1543 req->r_flags |= R_TPRINTFMSG;
1545 if (req->r_rexmit >= req->r_retry) { /* too many */
1546 nfsstats.rpctimeouts++;
1550 if (nmp->nm_sotype != SOCK_DGRAM) {
1551 if (++req->r_rexmit > NFS_MAXREXMIT)
1552 req->r_rexmit = NFS_MAXREXMIT;
1555 if ((so = nmp->nm_so) == NULL)
1559 * If there is enough space and the window allows..
1561 * Set r_rtt to -1 in case we fail to send it now.
1564 if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1565 ((nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1566 (req->r_flags & R_SENT) ||
1567 nmp->nm_sent < nmp->nm_cwnd) &&
1568 (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1569 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1570 error = so_pru_send(so, 0, m, NULL, NULL, td);
1572 error = so_pru_send(so, 0, m, nmp->nm_nam,
1575 if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1577 } else if (req->r_mrep == NULL) {
1579 * Iff first send, start timing
1580 * else turn timing off, backoff timer
1581 * and divide congestion window by 2.
1583 * It is possible for the so_pru_send() to
1584 * block and for us to race a reply so we
1585 * only do this if the reply field has not
1586 * been filled in. R_LOCKED will prevent
1587 * the request from being ripped out from under
1590 if (req->r_flags & R_SENT) {
1591 req->r_flags &= ~R_TIMING;
1592 if (++req->r_rexmit > NFS_MAXREXMIT)
1593 req->r_rexmit = NFS_MAXREXMIT;
1595 if (nmp->nm_cwnd < NFS_CWNDSCALE)
1596 nmp->nm_cwnd = NFS_CWNDSCALE;
1597 nfsstats.rpcretries++;
1599 req->r_flags |= R_SENT;
1600 nmp->nm_sent += NFS_CWNDSCALE;
1608 * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1609 * wait for all requests to complete. This is used by forced unmounts
1610 * to terminate any outstanding RPCs.
1613 nfs_nmcancelreqs(struct nfsmount *nmp)
1619 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1620 if (nmp != req->r_nmp || req->r_mrep != NULL ||
1621 (req->r_flags & R_SOFTTERM)) {
1628 for (i = 0; i < 30; i++) {
1630 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1631 if (nmp == req->r_nmp)
1637 tsleep(&lbolt, 0, "nfscancel", 0);
1643 * Flag a request as being about to terminate (due to NFSMNT_INT/NFSMNT_SOFT).
1644 * The nm_send count is decremented now to avoid deadlocks when the process in
1645 * soreceive() hasn't yet managed to send its own request.
1647 * This routine must be called at splsoftclock() to protect r_flags and
1652 nfs_softterm(struct nfsreq *rep)
1654 rep->r_flags |= R_SOFTTERM;
1656 if (rep->r_flags & R_SENT) {
1657 rep->r_nmp->nm_sent -= NFS_CWNDSCALE;
1658 rep->r_flags &= ~R_SENT;
1663 * Test for a termination condition pending on the process.
1664 * This is used for NFSMNT_INT mounts.
1667 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
1673 if (rep && (rep->r_flags & R_SOFTTERM))
1675 /* Terminate all requests while attempting a forced unmount. */
1676 if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
1678 if (!(nmp->nm_flag & NFSMNT_INT))
1680 /* td might be NULL YYY */
1681 if (td == NULL || (p = td->td_proc) == NULL)
1685 tmpset = lwp_sigpend(lp);
1686 SIGSETNAND(tmpset, lp->lwp_sigmask);
1687 SIGSETNAND(tmpset, p->p_sigignore);
1688 if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
1695 * Lock a socket against others.
1696 * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
1697 * and also to avoid race conditions between the processes with nfs requests
1698 * in progress when a reconnect is necessary.
1701 nfs_sndlock(struct nfsreq *rep)
1703 mtx_t mtx = &rep->r_nmp->nm_txlock;
1712 if (rep->r_nmp->nm_flag & NFSMNT_INT)
1715 while ((error = mtx_lock_ex_try(mtx)) != 0) {
1716 if (nfs_sigintr(rep->r_nmp, rep, td)) {
1720 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
1723 if (slpflag == PCATCH) {
1728 /* Always fail if our request has been cancelled. */
1729 if (rep->r_flags & R_SOFTTERM) {
1738 * Unlock the stream socket for others.
1741 nfs_sndunlock(struct nfsreq *rep)
1743 mtx_t mtx = &rep->r_nmp->nm_txlock;
1749 nfs_rcvlock(struct nfsreq *rep)
1751 mtx_t mtx = &rep->r_nmp->nm_rxlock;
1757 * Unconditionally check for completion in case another nfsiod
1758 * get the packet while the caller was blocked, before the caller
1759 * called us. Packet reception is handled by mainline code which
1760 * is protected by the BGL at the moment.
1762 * We do not strictly need the second check just before the
1763 * tsleep(), but it's good defensive programming.
1765 if (rep->r_mrep != NULL)
1768 if (rep->r_nmp->nm_flag & NFSMNT_INT)
1774 while ((error = mtx_lock_ex_try(mtx)) != 0) {
1775 if (nfs_sigintr(rep->r_nmp, rep, rep->r_td)) {
1779 if (rep->r_mrep != NULL) {
1785 * NOTE: can return ENOLCK, but in that case rep->r_mrep
1786 * will already be set.
1788 error = mtx_lock_ex_link(mtx, &rep->r_link, "nfsrcvlk",
1794 * If our reply was recieved while we were sleeping,
1795 * then just return without taking the lock to avoid a
1796 * situation where a single iod could 'capture' the
1799 if (rep->r_mrep != NULL) {
1803 if (slpflag == PCATCH) {
1809 if (rep->r_mrep != NULL) {
1818 * Unlock the stream socket for others.
1821 nfs_rcvunlock(struct nfsreq *rep)
1823 mtx_t mtx = &rep->r_nmp->nm_rxlock;
1831 * Check for badly aligned mbuf data and realign by copying the unaligned
1832 * portion of the data into a new mbuf chain and freeing the portions
1833 * of the old chain that were replaced.
1835 * We cannot simply realign the data within the existing mbuf chain
1836 * because the underlying buffers may contain other rpc commands and
1837 * we cannot afford to overwrite them.
1839 * We would prefer to avoid this situation entirely. The situation does
1840 * not occur with NFS/UDP and is supposed to only occassionally occur
1841 * with TCP. Use vfs.nfs.realign_count and realign_test to check this.
1844 nfs_realign(struct mbuf **pm, int hsiz)
1847 struct mbuf *n = NULL;
1852 while ((m = *pm) != NULL) {
1853 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3)) {
1854 n = m_getl(m->m_len, MB_WAIT, MT_DATA, 0, NULL);
1862 * If n is non-NULL, loop on m copying data, then replace the
1863 * portion of the chain that had to be realigned.
1866 ++nfs_realign_count;
1868 m_copyback(n, off, m->m_len, mtod(m, caddr_t));
1877 #ifndef NFS_NOSERVER
1880 * Parse an RPC request
1882 * - fill in the cred struct.
1885 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
1892 u_int32_t nfsvers, auth_type;
1894 int error = 0, ticklen;
1895 struct nfsuid *nuidp;
1896 struct timeval tvin, tvout;
1897 struct nfsm_info info;
1898 #if 0 /* until encrypted keys are implemented */
1899 NFSKERBKEYSCHED_T keys; /* stores key schedule */
1902 info.mrep = nd->nd_mrep;
1903 info.md = nd->nd_md;
1904 info.dpos = nd->nd_dpos;
1907 NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
1908 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
1909 if (*tl++ != rpc_call) {
1914 NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
1918 if (*tl++ != rpc_vers) {
1919 nd->nd_repstat = ERPCMISMATCH;
1920 nd->nd_procnum = NFSPROC_NOOP;
1923 if (*tl != nfs_prog) {
1924 nd->nd_repstat = EPROGUNAVAIL;
1925 nd->nd_procnum = NFSPROC_NOOP;
1929 nfsvers = fxdr_unsigned(u_int32_t, *tl++);
1930 if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
1931 nd->nd_repstat = EPROGMISMATCH;
1932 nd->nd_procnum = NFSPROC_NOOP;
1935 if (nfsvers == NFS_VER3)
1936 nd->nd_flag = ND_NFSV3;
1937 nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
1938 if (nd->nd_procnum == NFSPROC_NULL)
1940 if (nd->nd_procnum >= NFS_NPROCS ||
1941 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
1942 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
1943 nd->nd_repstat = EPROCUNAVAIL;
1944 nd->nd_procnum = NFSPROC_NOOP;
1947 if ((nd->nd_flag & ND_NFSV3) == 0)
1948 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
1950 len = fxdr_unsigned(int, *tl++);
1951 if (len < 0 || len > RPCAUTH_MAXSIZ) {
1956 nd->nd_flag &= ~ND_KERBAUTH;
1958 * Handle auth_unix or auth_kerb.
1960 if (auth_type == rpc_auth_unix) {
1961 len = fxdr_unsigned(int, *++tl);
1962 if (len < 0 || len > NFS_MAXNAMLEN) {
1966 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
1967 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
1968 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
1969 nd->nd_cr.cr_ref = 1;
1970 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
1971 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
1972 len = fxdr_unsigned(int, *tl);
1973 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
1977 NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
1978 for (i = 1; i <= len; i++)
1980 nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
1983 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
1984 if (nd->nd_cr.cr_ngroups > 1)
1985 nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
1986 len = fxdr_unsigned(int, *++tl);
1987 if (len < 0 || len > RPCAUTH_MAXSIZ) {
1992 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
1994 } else if (auth_type == rpc_auth_kerb) {
1995 switch (fxdr_unsigned(int, *tl++)) {
1996 case RPCAKN_FULLNAME:
1997 ticklen = fxdr_unsigned(int, *tl);
1998 *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
1999 uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2000 nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2001 if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2008 uio.uio_segflg = UIO_SYSSPACE;
2009 iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2010 iov.iov_len = RPCAUTH_MAXSIZ - 4;
2011 ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2012 NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2013 if (*tl++ != rpc_auth_kerb ||
2014 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2015 kprintf("Bad kerb verifier\n");
2016 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2017 nd->nd_procnum = NFSPROC_NOOP;
2020 NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2021 tl = (u_int32_t *)cp;
2022 if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2023 kprintf("Not fullname kerb verifier\n");
2024 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2025 nd->nd_procnum = NFSPROC_NOOP;
2028 cp += NFSX_UNSIGNED;
2029 bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2030 nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2031 nd->nd_flag |= ND_KERBFULL;
2032 nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2034 case RPCAKN_NICKNAME:
2035 if (len != 2 * NFSX_UNSIGNED) {
2036 kprintf("Kerb nickname short\n");
2037 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2038 nd->nd_procnum = NFSPROC_NOOP;
2041 nickuid = fxdr_unsigned(uid_t, *tl);
2042 NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2043 if (*tl++ != rpc_auth_kerb ||
2044 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2045 kprintf("Kerb nick verifier bad\n");
2046 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2047 nd->nd_procnum = NFSPROC_NOOP;
2050 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2051 tvin.tv_sec = *tl++;
2054 for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2055 nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2056 if (nuidp->nu_cr.cr_uid == nickuid &&
2058 netaddr_match(NU_NETFAM(nuidp),
2059 &nuidp->nu_haddr, nd->nd_nam2)))
2064 (NFSERR_AUTHERR|AUTH_REJECTCRED);
2065 nd->nd_procnum = NFSPROC_NOOP;
2070 * Now, decrypt the timestamp using the session key
2077 tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2078 tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2079 if (nuidp->nu_expire < time_second ||
2080 nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2081 (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2082 nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2083 nuidp->nu_expire = 0;
2085 (NFSERR_AUTHERR|AUTH_REJECTVERF);
2086 nd->nd_procnum = NFSPROC_NOOP;
2089 nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2090 nd->nd_flag |= ND_KERBNICK;
2093 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2094 nd->nd_procnum = NFSPROC_NOOP;
2098 nd->nd_md = info.md;
2099 nd->nd_dpos = info.dpos;
2108 * Send a message to the originating process's terminal. The thread and/or
2109 * process may be NULL. YYY the thread should not be NULL but there may
2110 * still be some uio_td's that are still being passed as NULL through to
2114 nfs_msg(struct thread *td, char *server, char *msg)
2118 if (td && td->td_proc)
2119 tpr = tprintf_open(td->td_proc);
2122 tprintf(tpr, "nfs server %s: %s\n", server, msg);
2127 #ifndef NFS_NOSERVER
2129 * Socket upcall routine for the nfsd sockets.
2130 * The caddr_t arg is a pointer to the "struct nfssvc_sock".
2131 * Essentially do as much as possible non-blocking, else punt and it will
2132 * be called with MB_WAIT from an nfsd.
2135 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2137 struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2139 struct sockaddr *nam;
2142 int nparallel_wakeup = 0;
2144 if ((slp->ns_flag & SLP_VALID) == 0)
2148 * Do not allow an infinite number of completed RPC records to build
2149 * up before we stop reading data from the socket. Otherwise we could
2150 * end up holding onto an unreasonable number of mbufs for requests
2151 * waiting for service.
2153 * This should give pretty good feedback to the TCP
2154 * layer and prevents a memory crunch for other protocols.
2156 * Note that the same service socket can be dispatched to several
2157 * nfs servers simultaniously.
2159 * the tcp protocol callback calls us with MB_DONTWAIT.
2160 * nfsd calls us with MB_WAIT (typically).
2162 if (waitflag == MB_DONTWAIT && slp->ns_numrec >= nfsd_waiting / 2 + 1) {
2163 slp->ns_flag |= SLP_NEEDQ;
2168 * Handle protocol specifics to parse an RPC request. We always
2169 * pull from the socket using non-blocking I/O.
2171 if (so->so_type == SOCK_STREAM) {
2173 * The data has to be read in an orderly fashion from a TCP
2174 * stream, unlike a UDP socket. It is possible for soreceive
2175 * and/or nfsrv_getstream() to block, so make sure only one
2176 * entity is messing around with the TCP stream at any given
2177 * moment. The receive sockbuf's lock in soreceive is not
2180 * Note that this procedure can be called from any number of
2181 * NFS severs *OR* can be upcalled directly from a TCP
2184 if (slp->ns_flag & SLP_GETSTREAM) {
2185 slp->ns_flag |= SLP_NEEDQ;
2188 slp->ns_flag |= SLP_GETSTREAM;
2191 * Do soreceive(). Pull out as much data as possible without
2194 sbinit(&sio, 1000000000);
2195 flags = MSG_DONTWAIT;
2196 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2197 if (error || sio.sb_mb == NULL) {
2198 if (error == EWOULDBLOCK)
2199 slp->ns_flag |= SLP_NEEDQ;
2201 slp->ns_flag |= SLP_DISCONN;
2202 slp->ns_flag &= ~SLP_GETSTREAM;
2206 if (slp->ns_rawend) {
2207 slp->ns_rawend->m_next = m;
2208 slp->ns_cc += sio.sb_cc;
2211 slp->ns_cc = sio.sb_cc;
2218 * Now try and parse as many record(s) as we can out of the
2221 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2224 slp->ns_flag |= SLP_DISCONN;
2226 slp->ns_flag |= SLP_NEEDQ;
2228 slp->ns_flag &= ~SLP_GETSTREAM;
2231 * For UDP soreceive typically pulls just one packet, loop
2232 * to get the whole batch.
2235 sbinit(&sio, 1000000000);
2236 flags = MSG_DONTWAIT;
2237 error = so_pru_soreceive(so, &nam, NULL, &sio,
2240 struct nfsrv_rec *rec;
2241 int mf = (waitflag & MB_DONTWAIT) ?
2242 M_NOWAIT : M_WAITOK;
2243 rec = kmalloc(sizeof(struct nfsrv_rec),
2247 FREE(nam, M_SONAME);
2251 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2252 rec->nr_address = nam;
2253 rec->nr_packet = sio.sb_mb;
2254 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2259 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2260 && error != EWOULDBLOCK) {
2261 slp->ns_flag |= SLP_DISCONN;
2265 } while (sio.sb_mb);
2269 * If we were upcalled from the tcp protocol layer and we have
2270 * fully parsed records ready to go, or there is new data pending,
2271 * or something went wrong, try to wake up an nfsd thread to deal
2275 if (waitflag == MB_DONTWAIT && (slp->ns_numrec > 0
2276 || (slp->ns_flag & (SLP_NEEDQ | SLP_DISCONN)))) {
2277 nfsrv_wakenfsd(slp, nparallel_wakeup);
2282 * Try and extract an RPC request from the mbuf data list received on a
2283 * stream socket. The "waitflag" argument indicates whether or not it
2287 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2289 struct mbuf *m, **mpp;
2292 struct mbuf *om, *m2, *recm;
2296 if (slp->ns_reclen == 0) {
2297 if (slp->ns_cc < NFSX_UNSIGNED)
2300 if (m->m_len >= NFSX_UNSIGNED) {
2301 bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2302 m->m_data += NFSX_UNSIGNED;
2303 m->m_len -= NFSX_UNSIGNED;
2305 cp1 = (caddr_t)&recmark;
2306 cp2 = mtod(m, caddr_t);
2307 while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2308 while (m->m_len == 0) {
2310 cp2 = mtod(m, caddr_t);
2317 slp->ns_cc -= NFSX_UNSIGNED;
2318 recmark = ntohl(recmark);
2319 slp->ns_reclen = recmark & ~0x80000000;
2320 if (recmark & 0x80000000)
2321 slp->ns_flag |= SLP_LASTFRAG;
2323 slp->ns_flag &= ~SLP_LASTFRAG;
2324 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2325 log(LOG_ERR, "%s (%d) from nfs client\n",
2326 "impossible packet length",
2333 * Now get the record part.
2335 * Note that slp->ns_reclen may be 0. Linux sometimes
2336 * generates 0-length RPCs
2339 if (slp->ns_cc == slp->ns_reclen) {
2341 slp->ns_raw = slp->ns_rawend = NULL;
2342 slp->ns_cc = slp->ns_reclen = 0;
2343 } else if (slp->ns_cc > slp->ns_reclen) {
2348 while (len < slp->ns_reclen) {
2349 if ((len + m->m_len) > slp->ns_reclen) {
2350 m2 = m_copym(m, 0, slp->ns_reclen - len,
2358 m->m_data += slp->ns_reclen - len;
2359 m->m_len -= slp->ns_reclen - len;
2360 len = slp->ns_reclen;
2362 return (EWOULDBLOCK);
2364 } else if ((len + m->m_len) == slp->ns_reclen) {
2384 * Accumulate the fragments into a record.
2386 mpp = &slp->ns_frag;
2388 mpp = &((*mpp)->m_next);
2390 if (slp->ns_flag & SLP_LASTFRAG) {
2391 struct nfsrv_rec *rec;
2392 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2393 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2395 m_freem(slp->ns_frag);
2397 nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2398 rec->nr_address = NULL;
2399 rec->nr_packet = slp->ns_frag;
2400 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2404 slp->ns_frag = NULL;
2410 * Parse an RPC header.
2413 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2414 struct nfsrv_descript **ndp)
2416 struct nfsrv_rec *rec;
2418 struct sockaddr *nam;
2419 struct nfsrv_descript *nd;
2423 if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2425 rec = STAILQ_FIRST(&slp->ns_rec);
2426 STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2427 KKASSERT(slp->ns_numrec > 0);
2429 nam = rec->nr_address;
2431 kfree(rec, M_NFSRVDESC);
2432 MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2433 M_NFSRVDESC, M_WAITOK);
2434 nd->nd_md = nd->nd_mrep = m;
2436 nd->nd_dpos = mtod(m, caddr_t);
2437 error = nfs_getreq(nd, nfsd, TRUE);
2440 FREE(nam, M_SONAME);
2442 kfree((caddr_t)nd, M_NFSRVDESC);
2451 * Try to assign service sockets to nfsd threads based on the number
2452 * of new rpc requests that have been queued on the service socket.
2454 * If no nfsd's are available or additonal requests are pending, set the
2455 * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2456 * the work in the nfssvc_sock list when it is finished processing its
2457 * current work. This flag is only cleared when an nfsd can not find
2458 * any new work to perform.
2461 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2465 if ((slp->ns_flag & SLP_VALID) == 0)
2469 TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2470 if (nd->nfsd_flag & NFSD_WAITING) {
2471 nd->nfsd_flag &= ~NFSD_WAITING;
2473 panic("nfsd wakeup");
2476 wakeup((caddr_t)nd);
2477 if (--nparallel == 0)
2482 slp->ns_flag |= SLP_DOREC;
2483 nfsd_head_flag |= NFSD_CHECKSLP;
2486 #endif /* NFS_NOSERVER */