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]
105 * External data, mostly RPC constants in XDR form
107 extern u_int32_t rpc_reply, rpc_msgdenied, rpc_mismatch, rpc_vers,
108 rpc_auth_unix, rpc_msgaccepted, rpc_call, rpc_autherr,
110 extern u_int32_t nfs_prog;
111 extern struct nfsstats nfsstats;
112 extern int nfsv3_procid[NFS_NPROCS];
113 extern int nfs_ticks;
116 * Defines which timer to use for the procnum.
123 static int proct[NFS_NPROCS] = {
124 0, 1, 0, 2, 1, 3, 3, 4, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 0, 0, 0, 0, 0,
128 static int nfs_realign_test;
129 static int nfs_realign_count;
130 static int nfs_bufpackets = 4;
131 static int nfs_timer_raced;
133 SYSCTL_DECL(_vfs_nfs);
135 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0, "");
136 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0, "");
137 SYSCTL_INT(_vfs_nfs, OID_AUTO, bufpackets, CTLFLAG_RW, &nfs_bufpackets, 0, "");
141 * There is a congestion window for outstanding rpcs maintained per mount
142 * point. The cwnd size is adjusted in roughly the way that:
143 * Van Jacobson, Congestion avoidance and Control, In "Proceedings of
144 * SIGCOMM '88". ACM, August 1988.
145 * describes for TCP. The cwnd size is chopped in half on a retransmit timeout
146 * and incremented by 1/cwnd when each rpc reply is received and a full cwnd
147 * of rpcs is in progress.
148 * (The sent count and cwnd are scaled for integer arith.)
149 * Variants of "slow start" were tried and were found to be too much of a
150 * performance hit (ave. rtt 3 times larger),
151 * I suspect due to the large rtt that nfs rpcs have.
153 #define NFS_CWNDSCALE 256
154 #define NFS_MAXCWND (NFS_CWNDSCALE * 32)
155 static int nfs_backoff[8] = { 2, 4, 8, 16, 32, 64, 128, 256, };
157 struct nfsrtt nfsrtt;
158 struct callout nfs_timer_handle;
160 static int nfs_msg (struct thread *,char *,char *);
161 static int nfs_rcvlock (struct nfsreq *);
162 static void nfs_rcvunlock (struct nfsreq *);
163 static void nfs_realign (struct mbuf **pm, int hsiz);
164 static int nfs_receive (struct nfsreq *rep, struct sockaddr **aname,
166 static void nfs_softterm (struct nfsreq *rep);
167 static int nfs_reconnect (struct nfsreq *rep);
169 static int nfsrv_getstream (struct nfssvc_sock *, int, int *);
171 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
172 struct nfssvc_sock *slp,
174 struct mbuf **mreqp) = {
202 #endif /* NFS_NOSERVER */
205 * Initialize sockets and congestion for a new NFS connection.
206 * We do not free the sockaddr if error.
209 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
212 int error, rcvreserve, sndreserve;
214 struct sockaddr *saddr;
215 struct sockaddr_in *sin;
216 struct thread *td = &thread0; /* only used for socreate and sobind */
220 error = socreate(saddr->sa_family, &nmp->nm_so, nmp->nm_sotype,
221 nmp->nm_soproto, td);
225 nmp->nm_soflags = so->so_proto->pr_flags;
228 * Some servers require that the client port be a reserved port number.
230 if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
233 struct sockaddr_in ssin;
235 bzero(&sopt, sizeof sopt);
236 ip = IP_PORTRANGE_LOW;
237 sopt.sopt_level = IPPROTO_IP;
238 sopt.sopt_name = IP_PORTRANGE;
239 sopt.sopt_val = (void *)&ip;
240 sopt.sopt_valsize = sizeof(ip);
242 error = sosetopt(so, &sopt);
245 bzero(&ssin, sizeof ssin);
247 sin->sin_len = sizeof (struct sockaddr_in);
248 sin->sin_family = AF_INET;
249 sin->sin_addr.s_addr = INADDR_ANY;
250 sin->sin_port = htons(0);
251 error = sobind(so, (struct sockaddr *)sin, td);
254 bzero(&sopt, sizeof sopt);
255 ip = IP_PORTRANGE_DEFAULT;
256 sopt.sopt_level = IPPROTO_IP;
257 sopt.sopt_name = IP_PORTRANGE;
258 sopt.sopt_val = (void *)&ip;
259 sopt.sopt_valsize = sizeof(ip);
261 error = sosetopt(so, &sopt);
267 * Protocols that do not require connections may be optionally left
268 * unconnected for servers that reply from a port other than NFS_PORT.
270 if (nmp->nm_flag & NFSMNT_NOCONN) {
271 if (nmp->nm_soflags & PR_CONNREQUIRED) {
276 error = soconnect(so, nmp->nm_nam, td);
281 * Wait for the connection to complete. Cribbed from the
282 * connect system call but with the wait timing out so
283 * that interruptible mounts don't hang here for a long time.
286 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
287 (void) tsleep((caddr_t)&so->so_timeo, 0,
289 if ((so->so_state & SS_ISCONNECTING) &&
290 so->so_error == 0 && rep &&
291 (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
292 so->so_state &= ~SS_ISCONNECTING;
298 error = so->so_error;
305 so->so_rcv.ssb_timeo = (5 * hz);
306 so->so_snd.ssb_timeo = (5 * hz);
309 * Get buffer reservation size from sysctl, but impose reasonable
312 pktscale = nfs_bufpackets;
318 if (nmp->nm_sotype == SOCK_DGRAM) {
319 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
320 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
321 NFS_MAXPKTHDR) * pktscale;
322 } else if (nmp->nm_sotype == SOCK_SEQPACKET) {
323 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
324 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
325 NFS_MAXPKTHDR) * pktscale;
327 if (nmp->nm_sotype != SOCK_STREAM)
328 panic("nfscon sotype");
329 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
333 bzero(&sopt, sizeof sopt);
334 sopt.sopt_level = SOL_SOCKET;
335 sopt.sopt_name = SO_KEEPALIVE;
336 sopt.sopt_val = &val;
337 sopt.sopt_valsize = sizeof val;
341 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
345 bzero(&sopt, sizeof sopt);
346 sopt.sopt_level = IPPROTO_TCP;
347 sopt.sopt_name = TCP_NODELAY;
348 sopt.sopt_val = &val;
349 sopt.sopt_valsize = sizeof val;
353 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR +
354 sizeof (u_int32_t)) * pktscale;
355 rcvreserve = (nmp->nm_rsize + NFS_MAXPKTHDR +
356 sizeof (u_int32_t)) * pktscale;
358 error = soreserve(so, sndreserve, rcvreserve,
359 &td->td_proc->p_rlimit[RLIMIT_SBSIZE]);
362 so->so_rcv.ssb_flags |= SSB_NOINTR;
363 so->so_snd.ssb_flags |= SSB_NOINTR;
365 /* Initialize other non-zero congestion variables */
366 nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] =
367 nmp->nm_srtt[3] = (NFS_TIMEO << 3);
368 nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
369 nmp->nm_sdrtt[3] = 0;
370 nmp->nm_cwnd = NFS_MAXCWND / 2; /* Initial send window */
372 nmp->nm_timeouts = 0;
382 * Called when a connection is broken on a reliable protocol.
383 * - clean up the old socket
384 * - nfs_connect() again
385 * - set R_MUSTRESEND for all outstanding requests on mount point
386 * If this fails the mount point is DEAD!
387 * nb: Must be called with the nfs_sndlock() set on the mount point.
390 nfs_reconnect(struct nfsreq *rep)
393 struct nfsmount *nmp = rep->r_nmp;
397 while ((error = nfs_connect(nmp, rep)) != 0) {
398 if (error == EINTR || error == ERESTART)
400 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
404 * Loop through outstanding request list and fix up all requests
408 TAILQ_FOREACH(rp, &nfs_reqq, r_chain) {
409 if (rp->r_nmp == nmp)
410 rp->r_flags |= R_MUSTRESEND;
417 * NFS disconnect. Clean up and unlink.
420 nfs_disconnect(struct nfsmount *nmp)
427 soshutdown(so, SHUT_RDWR);
428 soclose(so, FNONBLOCK);
433 nfs_safedisconnect(struct nfsmount *nmp)
435 struct nfsreq dummyreq;
437 bzero(&dummyreq, sizeof(dummyreq));
438 dummyreq.r_nmp = nmp;
439 dummyreq.r_td = NULL;
440 mtx_link_init(&dummyreq.r_link);
441 nfs_rcvlock(&dummyreq);
443 nfs_rcvunlock(&dummyreq);
447 * This is the nfs send routine. For connection based socket types, it
448 * must be called with an nfs_sndlock() on the socket.
449 * "rep == NULL" indicates that it has been called from a server.
450 * For the client side:
451 * - return EINTR if the RPC is terminated, 0 otherwise
452 * - set R_MUSTRESEND if the send fails for any reason
453 * - do any cleanup required by recoverable socket errors (?)
454 * For the server side:
455 * - return EINTR or ERESTART if interrupted by a signal
456 * - return EPIPE if a connection is lost for connection based sockets (TCP...)
457 * - do any cleanup required by recoverable socket errors (?)
460 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
463 struct sockaddr *sendnam;
464 int error, soflags, flags;
467 if (rep->r_flags & R_SOFTTERM) {
471 if ((so = rep->r_nmp->nm_so) == NULL) {
472 rep->r_flags |= R_MUSTRESEND;
476 rep->r_flags &= ~R_MUSTRESEND;
477 soflags = rep->r_nmp->nm_soflags;
479 soflags = so->so_proto->pr_flags;
480 if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
484 if (so->so_type == SOCK_SEQPACKET)
489 error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
492 * ENOBUFS for dgram sockets is transient and non fatal.
493 * No need to log, and no need to break a soft mount.
495 if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
497 if (rep) /* do backoff retransmit on client */
498 rep->r_flags |= R_MUSTRESEND;
503 log(LOG_INFO, "nfs send error %d for server %s\n",error,
504 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
506 * Deal with errors for the client side.
508 if (rep->r_flags & R_SOFTTERM)
511 rep->r_flags |= R_MUSTRESEND;
513 log(LOG_INFO, "nfsd send error %d\n", error);
516 * Handle any recoverable (soft) socket errors here. (?)
518 if (error != EINTR && error != ERESTART &&
519 error != EWOULDBLOCK && error != EPIPE)
526 * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
527 * done by soreceive(), but for SOCK_STREAM we must deal with the Record
528 * Mark and consolidate the data into a new mbuf list.
529 * nb: Sometimes TCP passes the data up to soreceive() in long lists of
531 * For SOCK_STREAM we must be very careful to read an entire record once
532 * we have read any of it, even if the system call has been interrupted.
535 nfs_receive(struct nfsreq *rep, struct sockaddr **aname, struct mbuf **mp)
542 struct mbuf *control;
544 struct sockaddr **getnam;
545 int error, sotype, rcvflg;
546 struct thread *td = curthread; /* XXX */
549 * Set up arguments for soreceive()
553 sotype = rep->r_nmp->nm_sotype;
556 * For reliable protocols, lock against other senders/receivers
557 * in case a reconnect is necessary.
558 * For SOCK_STREAM, first get the Record Mark to find out how much
559 * more there is to get.
560 * We must lock the socket against other receivers
561 * until we have an entire rpc request/reply.
563 if (sotype != SOCK_DGRAM) {
564 error = nfs_sndlock(rep);
569 * Check for fatal errors and resending request.
572 * Ugh: If a reconnect attempt just happened, nm_so
573 * would have changed. NULL indicates a failed
574 * attempt that has essentially shut down this
577 if (rep->r_mrep || (rep->r_flags & R_SOFTTERM)) {
581 so = rep->r_nmp->nm_so;
583 error = nfs_reconnect(rep);
590 while (rep->r_flags & R_MUSTRESEND) {
591 m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
592 nfsstats.rpcretries++;
593 error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
595 if (error == EINTR || error == ERESTART ||
596 (error = nfs_reconnect(rep)) != 0) {
604 if (sotype == SOCK_STREAM) {
606 * Get the length marker from the stream
608 aio.iov_base = (caddr_t)&len;
609 aio.iov_len = sizeof(u_int32_t);
612 auio.uio_segflg = UIO_SYSSPACE;
613 auio.uio_rw = UIO_READ;
615 auio.uio_resid = sizeof(u_int32_t);
618 rcvflg = MSG_WAITALL;
619 error = so_pru_soreceive(so, NULL, &auio, NULL,
621 if (error == EWOULDBLOCK && rep) {
622 if (rep->r_flags & R_SOFTTERM)
625 } while (error == EWOULDBLOCK);
627 if (error == 0 && auio.uio_resid > 0) {
629 * Only log short packets if not EOF
631 if (auio.uio_resid != sizeof(u_int32_t))
633 "short receive (%d/%d) from nfs server %s\n",
634 (int)(sizeof(u_int32_t) - auio.uio_resid),
635 (int)sizeof(u_int32_t),
636 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
641 len = ntohl(len) & ~0x80000000;
643 * This is SERIOUS! We are out of sync with the sender
644 * and forcing a disconnect/reconnect is all I can do.
646 if (len > NFS_MAXPACKET) {
647 log(LOG_ERR, "%s (%d) from nfs server %s\n",
648 "impossible packet length",
650 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
656 * Get the rest of the packet as an mbuf chain
660 rcvflg = MSG_WAITALL;
661 error = so_pru_soreceive(so, NULL, NULL, &sio,
663 } while (error == EWOULDBLOCK || error == EINTR ||
665 if (error == 0 && sio.sb_cc != len) {
668 "short receive (%d/%d) from nfs server %s\n",
669 len - auio.uio_resid, len,
670 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
676 * Non-stream, so get the whole packet by not
677 * specifying MSG_WAITALL and by specifying a large
680 * We have no use for control msg., but must grab them
681 * and then throw them away so we know what is going
684 sbinit(&sio, 100000000);
687 error = so_pru_soreceive(so, NULL, NULL, &sio,
691 if (error == EWOULDBLOCK && rep) {
692 if (rep->r_flags & R_SOFTTERM) {
697 } while (error == EWOULDBLOCK ||
698 (error == 0 && sio.sb_mb == NULL && control));
699 if ((rcvflg & MSG_EOR) == 0)
701 if (error == 0 && sio.sb_mb == NULL)
707 if (error && error != EINTR && error != ERESTART) {
710 if (error != EPIPE) {
712 "receive error %d from nfs server %s\n",
714 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
716 error = nfs_sndlock(rep);
718 error = nfs_reconnect(rep);
726 if ((so = rep->r_nmp->nm_so) == NULL)
728 if (so->so_state & SS_ISCONNECTED)
732 sbinit(&sio, 100000000);
735 error = so_pru_soreceive(so, getnam, NULL, &sio,
737 if (error == EWOULDBLOCK &&
738 (rep->r_flags & R_SOFTTERM)) {
742 } while (error == EWOULDBLOCK);
751 * Search for any mbufs that are not a multiple of 4 bytes long
752 * or with m_data not longword aligned.
753 * These could cause pointer alignment problems, so copy them to
754 * well aligned mbufs.
756 nfs_realign(mp, 5 * NFSX_UNSIGNED);
761 * Implement receipt of reply on a socket.
762 * We must search through the list of received datagrams matching them
763 * with outstanding requests using the xid, until ours is found.
767 nfs_reply(struct nfsreq *myrep)
770 struct nfsmount *nmp = myrep->r_nmp;
772 struct mbuf *mrep, *md;
773 struct sockaddr *nam;
779 * Loop around until we get our own reply
783 * Lock against other receivers so that I don't get stuck in
784 * sbwait() after someone else has received my reply for me.
785 * Also necessary for connection based protocols to avoid
786 * race conditions during a reconnect.
788 * If nfs_rcvlock() returns EALREADY, that means that
789 * the reply has already been recieved by another
790 * process and we can return immediately. In this
791 * case, the lock is not taken to avoid races with
794 error = nfs_rcvlock(myrep);
795 if (error == EALREADY)
800 * Get the next Rpc reply off the socket
802 error = nfs_receive(myrep, &nam, &mrep);
803 nfs_rcvunlock(myrep);
806 * Ignore routing errors on connectionless protocols??
808 if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
809 nmp->nm_so->so_error = 0;
810 if (myrep->r_flags & R_GETONEREP)
820 * Get the xid and check that it is an rpc reply
823 dpos = mtod(md, caddr_t);
824 nfsm_dissect(tl, u_int32_t *, 2*NFSX_UNSIGNED);
826 if (*tl != rpc_reply) {
827 nfsstats.rpcinvalid++;
830 if (myrep->r_flags & R_GETONEREP)
836 * Loop through the request list to match up the reply
837 * Iff no match, just drop the datagram. On match, set
838 * r_mrep atomically to prevent the timer from messing
839 * around with the request after we have exited the critical
843 TAILQ_FOREACH(rep, &nfs_reqq, r_chain) {
844 if (rep->r_mrep == NULL && rxid == rep->r_xid)
850 * Fill in the rest of the reply if we found a match.
858 rt = &nfsrtt.rttl[nfsrtt.pos];
859 rt->proc = rep->r_procnum;
860 rt->rto = NFS_RTO(nmp, proct[rep->r_procnum]);
861 rt->sent = nmp->nm_sent;
862 rt->cwnd = nmp->nm_cwnd;
863 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
864 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
865 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
866 getmicrotime(&rt->tstamp);
867 if (rep->r_flags & R_TIMING)
868 rt->rtt = rep->r_rtt;
871 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
874 * Update congestion window.
875 * Do the additive increase of
878 if (nmp->nm_cwnd <= nmp->nm_sent) {
880 (NFS_CWNDSCALE * NFS_CWNDSCALE +
881 (nmp->nm_cwnd >> 1)) / nmp->nm_cwnd;
882 if (nmp->nm_cwnd > NFS_MAXCWND)
883 nmp->nm_cwnd = NFS_MAXCWND;
885 crit_enter(); /* nfs_timer interlock for nm_sent */
886 if (rep->r_flags & R_SENT) {
887 rep->r_flags &= ~R_SENT;
888 nmp->nm_sent -= NFS_CWNDSCALE;
892 * Update rtt using a gain of 0.125 on the mean
893 * and a gain of 0.25 on the deviation.
895 if (rep->r_flags & R_TIMING) {
897 * Since the timer resolution of
898 * NFS_HZ is so course, it can often
899 * result in r_rtt == 0. Since
900 * r_rtt == N means that the actual
901 * rtt is between N+dt and N+2-dt ticks,
905 t1 -= (NFS_SRTT(rep) >> 3);
909 t1 -= (NFS_SDRTT(rep) >> 2);
910 NFS_SDRTT(rep) += t1;
912 nmp->nm_timeouts = 0;
914 mtx_abort_ex_link(&rep->r_nmp->nm_rxlock, &rep->r_link);
917 * If not matched to a request, drop it.
918 * If it's mine, get out.
921 nfsstats.rpcunexpected++;
923 } else if (rep == myrep) {
924 if (rep->r_mrep == NULL)
925 panic("nfsreply nil");
928 if (myrep->r_flags & R_GETONEREP)
934 nfs_request(struct vnode *vp, struct mbuf *mrest, int procnum,
935 struct thread *td, struct ucred *cred, struct mbuf **mrp,
936 struct mbuf **mdp, caddr_t *dposp)
938 struct nfsreq *rep = NULL;
941 error = nfs_request_setup(vp, mrest, procnum, td, cred, &rep);
946 rep->r_dposp = dposp;
948 error = nfs_request_auth(rep);
952 error = nfs_request_try(rep); /* error ignored */
953 error = nfs_request_waitreply(rep); /* pass to process */
954 error = nfs_request_processreply(rep, error);
955 if (error == ENEEDAUTH)
963 * nfs_request - goes something like this
964 * - fill in request struct
965 * - links it into list
966 * - calls nfs_send() for first transmit
967 * - calls nfs_receive() to get reply
968 * - break down rpc header and return with nfs reply pointed to
970 * nb: always frees up mreq mbuf list
973 nfs_request_setup(struct vnode *vp, struct mbuf *mrest, int procnum,
974 struct thread *td, struct ucred *cred,
975 struct nfsreq **repp)
978 struct nfsmount *nmp;
982 /* Reject requests while attempting a forced unmount. */
983 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
987 nmp = VFSTONFS(vp->v_mount);
988 MALLOC(rep, struct nfsreq *, sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
992 rep->r_procnum = procnum;
1000 rep->r_mrest = mrest;
1001 rep->r_mrest_len = i;
1008 nfs_request_auth(struct nfsreq *rep)
1010 struct nfsmount *nmp = rep->r_nmp;
1012 char nickv[RPCX_NICKVERF];
1013 int error = 0, auth_len, auth_type;
1016 char *auth_str, *verf_str;
1020 rep->r_failed_auth = 0;
1023 * Get the RPC header with authorization.
1025 verf_str = auth_str = NULL;
1026 if (nmp->nm_flag & NFSMNT_KERB) {
1028 verf_len = sizeof (nickv);
1029 auth_type = RPCAUTH_KERB4;
1030 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1031 if (rep->r_failed_auth ||
1032 nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1033 verf_str, verf_len)) {
1034 error = nfs_getauth(nmp, rep, cred, &auth_str,
1035 &auth_len, verf_str, &verf_len, rep->r_key);
1037 m_freem(rep->r_mrest);
1038 rep->r_mrest = NULL;
1039 kfree((caddr_t)rep, M_NFSREQ);
1044 auth_type = RPCAUTH_UNIX;
1045 if (cred->cr_ngroups < 1)
1046 panic("nfsreq nogrps");
1047 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1048 nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1051 m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1052 auth_len, auth_str, verf_len, verf_str,
1053 rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1054 rep->r_mrest = NULL;
1056 kfree(auth_str, M_TEMP);
1059 * For stream protocols, insert a Sun RPC Record Mark.
1061 if (nmp->nm_sotype == SOCK_STREAM) {
1062 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1064 kfree(rep, M_NFSREQ);
1067 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1068 (m->m_pkthdr.len - NFSX_UNSIGNED));
1076 nfs_request_try(struct nfsreq *rep)
1078 struct nfsmount *nmp = rep->r_nmp;
1082 if (nmp->nm_flag & NFSMNT_SOFT)
1083 rep->r_retry = nmp->nm_retry;
1085 rep->r_retry = NFS_MAXREXMIT + 1; /* past clip limit */
1086 rep->r_rtt = rep->r_rexmit = 0;
1087 if (proct[rep->r_procnum] > 0)
1088 rep->r_flags = R_TIMING | R_MASKTIMER;
1090 rep->r_flags = R_MASKTIMER;
1094 * Do the client side RPC.
1096 nfsstats.rpcrequests++;
1099 * Chain request into list of outstanding requests. Be sure
1100 * to put it LAST so timer finds oldest requests first. Note
1101 * that R_MASKTIMER is set at the moment to prevent any timer
1102 * action on this request while we are still doing processing on
1103 * it below. splsoftclock() primarily protects nm_sent. Note
1104 * that we may block in this code so there is no atomicy guarentee.
1107 TAILQ_INSERT_TAIL(&nfs_reqq, rep, r_chain);
1108 mtx_link_init(&rep->r_link);
1113 * If backing off another request or avoiding congestion, don't
1114 * send this one now but let timer do it. If not timing a request,
1117 * Even though the timer will not mess with our request there is
1118 * still the possibility that we will race a reply (which clears
1119 * R_SENT), especially on localhost connections, so be very careful
1120 * when setting R_SENT. We could set R_SENT prior to calling
1121 * nfs_send() but why bother if the response occurs that quickly?
1123 if (nmp->nm_so && (nmp->nm_sotype != SOCK_DGRAM ||
1124 (nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1125 nmp->nm_sent < nmp->nm_cwnd)) {
1126 if (nmp->nm_soflags & PR_CONNREQUIRED)
1127 error = nfs_sndlock(rep);
1129 m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1130 error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1131 if (nmp->nm_soflags & PR_CONNREQUIRED)
1134 if (!error && (rep->r_flags & R_MUSTRESEND) == 0 &&
1135 rep->r_mrep == NULL) {
1136 KASSERT((rep->r_flags & R_SENT) == 0,
1137 ("R_SENT ASSERT %p", rep));
1138 nmp->nm_sent += NFS_CWNDSCALE;
1139 rep->r_flags |= R_SENT;
1147 * Let the timer do what it will with the request, then
1148 * wait for the reply from our send or the timer's.
1151 rep->r_flags &= ~R_MASKTIMER;
1157 nfs_request_waitreply(struct nfsreq *rep)
1159 struct nfsmount *nmp = rep->r_nmp;
1163 error = nfs_reply(rep);
1167 * RPC done, unlink the request, but don't rip it out from under
1168 * the callout timer.
1170 while (rep->r_flags & R_LOCKED) {
1171 nfs_timer_raced = 1;
1172 tsleep(&nfs_timer_raced, 0, "nfstrac", 0);
1174 TAILQ_REMOVE(&nfs_reqq, rep, r_chain);
1177 * Decrement the outstanding request count.
1179 if (rep->r_flags & R_SENT) {
1180 rep->r_flags &= ~R_SENT;
1181 nmp->nm_sent -= NFS_CWNDSCALE;
1189 * Process reply with error returned from nfs_requet_waitreply().
1191 * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1192 * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1195 nfs_request_processreply(struct nfsreq *rep, int error)
1197 struct nfsmount *nmp = rep->r_nmp;
1203 int trylater_delay = 15, trylater_cnt = 0;
1209 * If there was a successful reply and a tprintf msg.
1210 * tprintf a response.
1212 if (!error && (rep->r_flags & R_TPRINTFMSG))
1213 nfs_msg(rep->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1219 m_freem(rep->r_mreq);
1220 kfree((caddr_t)rep, M_NFSREQ);
1225 * break down the rpc header and check if ok
1227 nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1228 if (*tl++ == rpc_msgdenied) {
1229 if (*tl == rpc_mismatch) {
1231 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1232 *tl++ == rpc_autherr) {
1233 if (!rep->r_failed_auth) {
1234 rep->r_failed_auth++;
1235 rep->r_mheadend->m_next = NULL;
1237 m_freem(rep->r_mreq);
1246 m_freem(rep->r_mreq);
1247 kfree((caddr_t)rep, M_NFSREQ);
1252 * Grab any Kerberos verifier, otherwise just throw it away.
1254 verf_type = fxdr_unsigned(int, *tl++);
1255 i = fxdr_unsigned(int32_t, *tl);
1256 if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1257 error = nfs_savenickauth(nmp, rep->r_cred, i,
1258 rep->r_key, &md, &dpos, mrep);
1262 nfsm_adv(nfsm_rndup(i));
1263 nfsm_dissect(tl, u_int32_t *, NFSX_UNSIGNED);
1266 nfsm_dissect(tl, u_int32_t *, NFSX_UNSIGNED);
1268 error = fxdr_unsigned(int, *tl);
1269 if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1270 error == NFSERR_TRYLATER) {
1273 waituntil = time_second + trylater_delay;
1274 while (time_second < waituntil)
1275 (void) tsleep((caddr_t)&lbolt,
1277 trylater_delay *= nfs_backoff[trylater_cnt];
1278 if (trylater_cnt < 7)
1280 rep->r_flags &= ~R_MASKTIMER;
1281 return (EAGAIN); /* goto tryagain */
1285 * If the File Handle was stale, invalidate the
1286 * lookup cache, just in case.
1288 * To avoid namecache<->vnode deadlocks we must
1289 * release the vnode lock if we hold it.
1291 if (error == ESTALE) {
1292 struct vnode *vp = rep->r_vp;
1295 ltype = lockstatus(&vp->v_lock, curthread);
1296 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1297 lockmgr(&vp->v_lock, LK_RELEASE);
1298 cache_inval_vp(vp, CINV_CHILDREN);
1299 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1300 lockmgr(&vp->v_lock, ltype);
1302 if (nmp->nm_flag & NFSMNT_NFSV3) {
1305 *rep->r_dposp = dpos;
1306 error |= NFSERR_RETERR;
1309 m_freem(rep->r_mreq);
1310 kfree((caddr_t)rep, M_NFSREQ);
1316 *rep->r_dposp = dpos;
1317 m_freem(rep->r_mreq);
1318 FREE((caddr_t)rep, M_NFSREQ);
1322 error = EPROTONOSUPPORT;
1324 m_freem(rep->r_mreq);
1325 kfree((caddr_t)rep, M_NFSREQ);
1329 #ifndef NFS_NOSERVER
1331 * Generate the rpc reply header
1332 * siz arg. is used to decide if adding a cluster is worthwhile
1335 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1336 int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1341 struct mbuf *mb, *mb2;
1343 siz += RPC_REPLYSIZ;
1344 mb = mreq = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1345 mreq->m_pkthdr.len = 0;
1347 * If this is not a cluster, try and leave leading space
1348 * for the lower level headers.
1350 if ((max_hdr + siz) < MINCLSIZE)
1351 mreq->m_data += max_hdr;
1352 tl = mtod(mreq, u_int32_t *);
1353 mreq->m_len = 6 * NFSX_UNSIGNED;
1354 bpos = ((caddr_t)tl) + mreq->m_len;
1355 *tl++ = txdr_unsigned(nd->nd_retxid);
1357 if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1358 *tl++ = rpc_msgdenied;
1359 if (err & NFSERR_AUTHERR) {
1360 *tl++ = rpc_autherr;
1361 *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1362 mreq->m_len -= NFSX_UNSIGNED;
1363 bpos -= NFSX_UNSIGNED;
1365 *tl++ = rpc_mismatch;
1366 *tl++ = txdr_unsigned(RPC_VER2);
1367 *tl = txdr_unsigned(RPC_VER2);
1370 *tl++ = rpc_msgaccepted;
1373 * For Kerberos authentication, we must send the nickname
1374 * verifier back, otherwise just RPCAUTH_NULL.
1376 if (nd->nd_flag & ND_KERBFULL) {
1377 struct nfsuid *nuidp;
1378 struct timeval ktvin, ktvout;
1380 for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1381 nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1382 if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1383 (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1384 &nuidp->nu_haddr, nd->nd_nam2)))
1389 txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1391 txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1394 * Encrypt the timestamp in ecb mode using the
1401 *tl++ = rpc_auth_kerb;
1402 *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1403 *tl = ktvout.tv_sec;
1404 nfsm_build(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1405 *tl++ = ktvout.tv_usec;
1406 *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1417 *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1420 *tl = txdr_unsigned(RPC_PROGMISMATCH);
1421 nfsm_build(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
1422 *tl++ = txdr_unsigned(2);
1423 *tl = txdr_unsigned(3);
1426 *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1429 *tl = txdr_unsigned(RPC_GARBAGE);
1433 if (err != NFSERR_RETVOID) {
1434 nfsm_build(tl, u_int32_t *, NFSX_UNSIGNED);
1436 *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1448 if (err != 0 && err != NFSERR_RETVOID)
1449 nfsstats.srvrpc_errs++;
1454 #endif /* NFS_NOSERVER */
1457 * Scan the nfsreq list and retranmit any requests that have timed out
1458 * To avoid retransmission attempts on STREAM sockets (in the future) make
1459 * sure to set the r_retry field to 0 (implies nm_retry == 0).
1462 nfs_timer(void *arg /* never used */)
1467 struct nfsmount *nmp;
1470 #ifndef NFS_NOSERVER
1471 struct nfssvc_sock *slp;
1473 #endif /* NFS_NOSERVER */
1474 struct thread *td = &thread0; /* XXX for credentials, will break if sleep */
1477 TAILQ_FOREACH(rep, &nfs_reqq, r_chain) {
1479 if (rep->r_mrep || (rep->r_flags & (R_SOFTTERM|R_MASKTIMER)))
1481 rep->r_flags |= R_LOCKED;
1482 if (nfs_sigintr(nmp, rep, rep->r_td)) {
1486 if (rep->r_rtt >= 0) {
1488 if (nmp->nm_flag & NFSMNT_DUMBTIMR)
1489 timeo = nmp->nm_timeo;
1491 timeo = NFS_RTO(nmp, proct[rep->r_procnum]);
1492 if (nmp->nm_timeouts > 0)
1493 timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1494 if (rep->r_rtt <= timeo)
1496 if (nmp->nm_timeouts < 8)
1500 * Check for server not responding
1502 if ((rep->r_flags & R_TPRINTFMSG) == 0 &&
1503 rep->r_rexmit > nmp->nm_deadthresh) {
1505 nmp->nm_mountp->mnt_stat.f_mntfromname,
1507 rep->r_flags |= R_TPRINTFMSG;
1509 if (rep->r_rexmit >= rep->r_retry) { /* too many */
1510 nfsstats.rpctimeouts++;
1514 if (nmp->nm_sotype != SOCK_DGRAM) {
1515 if (++rep->r_rexmit > NFS_MAXREXMIT)
1516 rep->r_rexmit = NFS_MAXREXMIT;
1519 if ((so = nmp->nm_so) == NULL)
1523 * If there is enough space and the window allows..
1525 * Set r_rtt to -1 in case we fail to send it now.
1528 if (ssb_space(&so->so_snd) >= rep->r_mreq->m_pkthdr.len &&
1529 ((nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1530 (rep->r_flags & R_SENT) ||
1531 nmp->nm_sent < nmp->nm_cwnd) &&
1532 (m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1533 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1534 error = so_pru_send(so, 0, m, NULL, NULL, td);
1536 error = so_pru_send(so, 0, m, nmp->nm_nam,
1539 if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1541 } else if (rep->r_mrep == NULL) {
1543 * Iff first send, start timing
1544 * else turn timing off, backoff timer
1545 * and divide congestion window by 2.
1547 * It is possible for the so_pru_send() to
1548 * block and for us to race a reply so we
1549 * only do this if the reply field has not
1550 * been filled in. R_LOCKED will prevent
1551 * the request from being ripped out from under
1554 if (rep->r_flags & R_SENT) {
1555 rep->r_flags &= ~R_TIMING;
1556 if (++rep->r_rexmit > NFS_MAXREXMIT)
1557 rep->r_rexmit = NFS_MAXREXMIT;
1559 if (nmp->nm_cwnd < NFS_CWNDSCALE)
1560 nmp->nm_cwnd = NFS_CWNDSCALE;
1561 nfsstats.rpcretries++;
1563 rep->r_flags |= R_SENT;
1564 nmp->nm_sent += NFS_CWNDSCALE;
1570 rep->r_flags &= ~R_LOCKED;
1572 #ifndef NFS_NOSERVER
1575 * Scan the write gathering queues for writes that need to be
1578 cur_usec = nfs_curusec();
1579 TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1580 if (slp->ns_tq.lh_first && slp->ns_tq.lh_first->nd_time<=cur_usec)
1581 nfsrv_wakenfsd(slp, 1);
1583 #endif /* NFS_NOSERVER */
1586 * Due to possible blocking, a client operation may be waiting for
1587 * us to finish processing this request so it can remove it.
1589 if (nfs_timer_raced) {
1590 nfs_timer_raced = 0;
1591 wakeup(&nfs_timer_raced);
1594 callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer, NULL);
1598 * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1599 * wait for all requests to complete. This is used by forced unmounts
1600 * to terminate any outstanding RPCs.
1603 nfs_nmcancelreqs(struct nfsmount *nmp)
1609 TAILQ_FOREACH(req, &nfs_reqq, r_chain) {
1610 if (nmp != req->r_nmp || req->r_mrep != NULL ||
1611 (req->r_flags & R_SOFTTERM)) {
1618 for (i = 0; i < 30; i++) {
1620 TAILQ_FOREACH(req, &nfs_reqq, r_chain) {
1621 if (nmp == req->r_nmp)
1627 tsleep(&lbolt, 0, "nfscancel", 0);
1633 * Flag a request as being about to terminate (due to NFSMNT_INT/NFSMNT_SOFT).
1634 * The nm_send count is decremented now to avoid deadlocks when the process in
1635 * soreceive() hasn't yet managed to send its own request.
1637 * This routine must be called at splsoftclock() to protect r_flags and
1642 nfs_softterm(struct nfsreq *rep)
1644 rep->r_flags |= R_SOFTTERM;
1646 if (rep->r_flags & R_SENT) {
1647 rep->r_nmp->nm_sent -= NFS_CWNDSCALE;
1648 rep->r_flags &= ~R_SENT;
1653 * Test for a termination condition pending on the process.
1654 * This is used for NFSMNT_INT mounts.
1657 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
1663 if (rep && (rep->r_flags & R_SOFTTERM))
1665 /* Terminate all requests while attempting a forced unmount. */
1666 if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
1668 if (!(nmp->nm_flag & NFSMNT_INT))
1670 /* td might be NULL YYY */
1671 if (td == NULL || (p = td->td_proc) == NULL)
1675 tmpset = lwp_sigpend(lp);
1676 SIGSETNAND(tmpset, lp->lwp_sigmask);
1677 SIGSETNAND(tmpset, p->p_sigignore);
1678 if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
1685 * Lock a socket against others.
1686 * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
1687 * and also to avoid race conditions between the processes with nfs requests
1688 * in progress when a reconnect is necessary.
1691 nfs_sndlock(struct nfsreq *rep)
1693 mtx_t mtx = &rep->r_nmp->nm_txlock;
1702 if (rep->r_nmp->nm_flag & NFSMNT_INT)
1705 while ((error = mtx_lock_ex_try(mtx)) != 0) {
1706 if (nfs_sigintr(rep->r_nmp, rep, td)) {
1710 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
1713 if (slpflag == PCATCH) {
1718 /* Always fail if our request has been cancelled. */
1719 if (rep->r_flags & R_SOFTTERM) {
1728 * Unlock the stream socket for others.
1731 nfs_sndunlock(struct nfsreq *rep)
1733 mtx_t mtx = &rep->r_nmp->nm_txlock;
1739 nfs_rcvlock(struct nfsreq *rep)
1741 mtx_t mtx = &rep->r_nmp->nm_rxlock;
1747 * Unconditionally check for completion in case another nfsiod
1748 * get the packet while the caller was blocked, before the caller
1749 * called us. Packet reception is handled by mainline code which
1750 * is protected by the BGL at the moment.
1752 * We do not strictly need the second check just before the
1753 * tsleep(), but it's good defensive programming.
1755 if (rep->r_mrep != NULL)
1758 if (rep->r_nmp->nm_flag & NFSMNT_INT)
1764 while ((error = mtx_lock_ex_try(mtx)) != 0) {
1765 if (nfs_sigintr(rep->r_nmp, rep, rep->r_td)) {
1769 if (rep->r_mrep != NULL) {
1775 * NOTE: can return ENOLCK, but in that case rep->r_mrep
1776 * will already be set.
1778 error = mtx_lock_ex_link(mtx, &rep->r_link, "nfsrcvlk",
1784 * If our reply was recieved while we were sleeping,
1785 * then just return without taking the lock to avoid a
1786 * situation where a single iod could 'capture' the
1789 if (rep->r_mrep != NULL) {
1793 if (slpflag == PCATCH) {
1799 if (rep->r_mrep != NULL) {
1808 * Unlock the stream socket for others.
1811 nfs_rcvunlock(struct nfsreq *rep)
1813 mtx_t mtx = &rep->r_nmp->nm_rxlock;
1821 * Check for badly aligned mbuf data and realign by copying the unaligned
1822 * portion of the data into a new mbuf chain and freeing the portions
1823 * of the old chain that were replaced.
1825 * We cannot simply realign the data within the existing mbuf chain
1826 * because the underlying buffers may contain other rpc commands and
1827 * we cannot afford to overwrite them.
1829 * We would prefer to avoid this situation entirely. The situation does
1830 * not occur with NFS/UDP and is supposed to only occassionally occur
1831 * with TCP. Use vfs.nfs.realign_count and realign_test to check this.
1834 nfs_realign(struct mbuf **pm, int hsiz)
1837 struct mbuf *n = NULL;
1842 while ((m = *pm) != NULL) {
1843 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3)) {
1844 n = m_getl(m->m_len, MB_WAIT, MT_DATA, 0, NULL);
1852 * If n is non-NULL, loop on m copying data, then replace the
1853 * portion of the chain that had to be realigned.
1856 ++nfs_realign_count;
1858 m_copyback(n, off, m->m_len, mtod(m, caddr_t));
1867 #ifndef NFS_NOSERVER
1870 * Parse an RPC request
1872 * - fill in the cred struct.
1875 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
1882 caddr_t dpos, cp2, cp;
1883 u_int32_t nfsvers, auth_type;
1885 int error = 0, ticklen;
1886 struct mbuf *mrep, *md;
1887 struct nfsuid *nuidp;
1888 struct timeval tvin, tvout;
1889 #if 0 /* until encrypted keys are implemented */
1890 NFSKERBKEYSCHED_T keys; /* stores key schedule */
1897 nfsm_dissect(tl, u_int32_t *, 10 * NFSX_UNSIGNED);
1898 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
1899 if (*tl++ != rpc_call) {
1904 nfsm_dissect(tl, u_int32_t *, 8 * NFSX_UNSIGNED);
1907 if (*tl++ != rpc_vers) {
1908 nd->nd_repstat = ERPCMISMATCH;
1909 nd->nd_procnum = NFSPROC_NOOP;
1912 if (*tl != nfs_prog) {
1913 nd->nd_repstat = EPROGUNAVAIL;
1914 nd->nd_procnum = NFSPROC_NOOP;
1918 nfsvers = fxdr_unsigned(u_int32_t, *tl++);
1919 if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
1920 nd->nd_repstat = EPROGMISMATCH;
1921 nd->nd_procnum = NFSPROC_NOOP;
1924 if (nfsvers == NFS_VER3)
1925 nd->nd_flag = ND_NFSV3;
1926 nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
1927 if (nd->nd_procnum == NFSPROC_NULL)
1929 if (nd->nd_procnum >= NFS_NPROCS ||
1930 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
1931 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
1932 nd->nd_repstat = EPROCUNAVAIL;
1933 nd->nd_procnum = NFSPROC_NOOP;
1936 if ((nd->nd_flag & ND_NFSV3) == 0)
1937 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
1939 len = fxdr_unsigned(int, *tl++);
1940 if (len < 0 || len > RPCAUTH_MAXSIZ) {
1945 nd->nd_flag &= ~ND_KERBAUTH;
1947 * Handle auth_unix or auth_kerb.
1949 if (auth_type == rpc_auth_unix) {
1950 len = fxdr_unsigned(int, *++tl);
1951 if (len < 0 || len > NFS_MAXNAMLEN) {
1955 nfsm_adv(nfsm_rndup(len));
1956 nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1957 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
1958 nd->nd_cr.cr_ref = 1;
1959 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
1960 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
1961 len = fxdr_unsigned(int, *tl);
1962 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
1966 nfsm_dissect(tl, u_int32_t *, (len + 2) * NFSX_UNSIGNED);
1967 for (i = 1; i <= len; i++)
1969 nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
1972 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
1973 if (nd->nd_cr.cr_ngroups > 1)
1974 nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
1975 len = fxdr_unsigned(int, *++tl);
1976 if (len < 0 || len > RPCAUTH_MAXSIZ) {
1981 nfsm_adv(nfsm_rndup(len));
1982 } else if (auth_type == rpc_auth_kerb) {
1983 switch (fxdr_unsigned(int, *tl++)) {
1984 case RPCAKN_FULLNAME:
1985 ticklen = fxdr_unsigned(int, *tl);
1986 *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
1987 uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
1988 nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
1989 if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
1996 uio.uio_segflg = UIO_SYSSPACE;
1997 iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
1998 iov.iov_len = RPCAUTH_MAXSIZ - 4;
1999 nfsm_mtouio(&uio, uio.uio_resid);
2000 nfsm_dissect(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
2001 if (*tl++ != rpc_auth_kerb ||
2002 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2003 kprintf("Bad kerb verifier\n");
2004 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2005 nd->nd_procnum = NFSPROC_NOOP;
2008 nfsm_dissect(cp, caddr_t, 4 * NFSX_UNSIGNED);
2009 tl = (u_int32_t *)cp;
2010 if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2011 kprintf("Not fullname kerb verifier\n");
2012 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2013 nd->nd_procnum = NFSPROC_NOOP;
2016 cp += NFSX_UNSIGNED;
2017 bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2018 nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2019 nd->nd_flag |= ND_KERBFULL;
2020 nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2022 case RPCAKN_NICKNAME:
2023 if (len != 2 * NFSX_UNSIGNED) {
2024 kprintf("Kerb nickname short\n");
2025 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2026 nd->nd_procnum = NFSPROC_NOOP;
2029 nickuid = fxdr_unsigned(uid_t, *tl);
2030 nfsm_dissect(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
2031 if (*tl++ != rpc_auth_kerb ||
2032 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2033 kprintf("Kerb nick verifier bad\n");
2034 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2035 nd->nd_procnum = NFSPROC_NOOP;
2038 nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
2039 tvin.tv_sec = *tl++;
2042 for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2043 nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2044 if (nuidp->nu_cr.cr_uid == nickuid &&
2046 netaddr_match(NU_NETFAM(nuidp),
2047 &nuidp->nu_haddr, nd->nd_nam2)))
2052 (NFSERR_AUTHERR|AUTH_REJECTCRED);
2053 nd->nd_procnum = NFSPROC_NOOP;
2058 * Now, decrypt the timestamp using the session key
2065 tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2066 tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2067 if (nuidp->nu_expire < time_second ||
2068 nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2069 (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2070 nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2071 nuidp->nu_expire = 0;
2073 (NFSERR_AUTHERR|AUTH_REJECTVERF);
2074 nd->nd_procnum = NFSPROC_NOOP;
2077 nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2078 nd->nd_flag |= ND_KERBNICK;
2081 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2082 nd->nd_procnum = NFSPROC_NOOP;
2096 * Send a message to the originating process's terminal. The thread and/or
2097 * process may be NULL. YYY the thread should not be NULL but there may
2098 * still be some uio_td's that are still being passed as NULL through to
2102 nfs_msg(struct thread *td, char *server, char *msg)
2106 if (td && td->td_proc)
2107 tpr = tprintf_open(td->td_proc);
2110 tprintf(tpr, "nfs server %s: %s\n", server, msg);
2115 #ifndef NFS_NOSERVER
2117 * Socket upcall routine for the nfsd sockets.
2118 * The caddr_t arg is a pointer to the "struct nfssvc_sock".
2119 * Essentially do as much as possible non-blocking, else punt and it will
2120 * be called with MB_WAIT from an nfsd.
2123 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2125 struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2127 struct sockaddr *nam;
2130 int nparallel_wakeup = 0;
2132 if ((slp->ns_flag & SLP_VALID) == 0)
2136 * Do not allow an infinite number of completed RPC records to build
2137 * up before we stop reading data from the socket. Otherwise we could
2138 * end up holding onto an unreasonable number of mbufs for requests
2139 * waiting for service.
2141 * This should give pretty good feedback to the TCP
2142 * layer and prevents a memory crunch for other protocols.
2144 * Note that the same service socket can be dispatched to several
2145 * nfs servers simultaniously.
2147 * the tcp protocol callback calls us with MB_DONTWAIT.
2148 * nfsd calls us with MB_WAIT (typically).
2150 if (waitflag == MB_DONTWAIT && slp->ns_numrec >= nfsd_waiting / 2 + 1) {
2151 slp->ns_flag |= SLP_NEEDQ;
2156 * Handle protocol specifics to parse an RPC request. We always
2157 * pull from the socket using non-blocking I/O.
2159 if (so->so_type == SOCK_STREAM) {
2161 * The data has to be read in an orderly fashion from a TCP
2162 * stream, unlike a UDP socket. It is possible for soreceive
2163 * and/or nfsrv_getstream() to block, so make sure only one
2164 * entity is messing around with the TCP stream at any given
2165 * moment. The receive sockbuf's lock in soreceive is not
2168 * Note that this procedure can be called from any number of
2169 * NFS severs *OR* can be upcalled directly from a TCP
2172 if (slp->ns_flag & SLP_GETSTREAM) {
2173 slp->ns_flag |= SLP_NEEDQ;
2176 slp->ns_flag |= SLP_GETSTREAM;
2179 * Do soreceive(). Pull out as much data as possible without
2182 sbinit(&sio, 1000000000);
2183 flags = MSG_DONTWAIT;
2184 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2185 if (error || sio.sb_mb == NULL) {
2186 if (error == EWOULDBLOCK)
2187 slp->ns_flag |= SLP_NEEDQ;
2189 slp->ns_flag |= SLP_DISCONN;
2190 slp->ns_flag &= ~SLP_GETSTREAM;
2194 if (slp->ns_rawend) {
2195 slp->ns_rawend->m_next = m;
2196 slp->ns_cc += sio.sb_cc;
2199 slp->ns_cc = sio.sb_cc;
2206 * Now try and parse as many record(s) as we can out of the
2209 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2212 slp->ns_flag |= SLP_DISCONN;
2214 slp->ns_flag |= SLP_NEEDQ;
2216 slp->ns_flag &= ~SLP_GETSTREAM;
2219 * For UDP soreceive typically pulls just one packet, loop
2220 * to get the whole batch.
2223 sbinit(&sio, 1000000000);
2224 flags = MSG_DONTWAIT;
2225 error = so_pru_soreceive(so, &nam, NULL, &sio,
2228 struct nfsrv_rec *rec;
2229 int mf = (waitflag & MB_DONTWAIT) ?
2230 M_NOWAIT : M_WAITOK;
2231 rec = kmalloc(sizeof(struct nfsrv_rec),
2235 FREE(nam, M_SONAME);
2239 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2240 rec->nr_address = nam;
2241 rec->nr_packet = sio.sb_mb;
2242 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2247 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2248 && error != EWOULDBLOCK) {
2249 slp->ns_flag |= SLP_DISCONN;
2253 } while (sio.sb_mb);
2257 * If we were upcalled from the tcp protocol layer and we have
2258 * fully parsed records ready to go, or there is new data pending,
2259 * or something went wrong, try to wake up an nfsd thread to deal
2263 if (waitflag == MB_DONTWAIT && (slp->ns_numrec > 0
2264 || (slp->ns_flag & (SLP_NEEDQ | SLP_DISCONN)))) {
2265 nfsrv_wakenfsd(slp, nparallel_wakeup);
2270 * Try and extract an RPC request from the mbuf data list received on a
2271 * stream socket. The "waitflag" argument indicates whether or not it
2275 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2277 struct mbuf *m, **mpp;
2280 struct mbuf *om, *m2, *recm;
2284 if (slp->ns_reclen == 0) {
2285 if (slp->ns_cc < NFSX_UNSIGNED)
2288 if (m->m_len >= NFSX_UNSIGNED) {
2289 bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2290 m->m_data += NFSX_UNSIGNED;
2291 m->m_len -= NFSX_UNSIGNED;
2293 cp1 = (caddr_t)&recmark;
2294 cp2 = mtod(m, caddr_t);
2295 while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2296 while (m->m_len == 0) {
2298 cp2 = mtod(m, caddr_t);
2305 slp->ns_cc -= NFSX_UNSIGNED;
2306 recmark = ntohl(recmark);
2307 slp->ns_reclen = recmark & ~0x80000000;
2308 if (recmark & 0x80000000)
2309 slp->ns_flag |= SLP_LASTFRAG;
2311 slp->ns_flag &= ~SLP_LASTFRAG;
2312 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2313 log(LOG_ERR, "%s (%d) from nfs client\n",
2314 "impossible packet length",
2321 * Now get the record part.
2323 * Note that slp->ns_reclen may be 0. Linux sometimes
2324 * generates 0-length RPCs
2327 if (slp->ns_cc == slp->ns_reclen) {
2329 slp->ns_raw = slp->ns_rawend = NULL;
2330 slp->ns_cc = slp->ns_reclen = 0;
2331 } else if (slp->ns_cc > slp->ns_reclen) {
2336 while (len < slp->ns_reclen) {
2337 if ((len + m->m_len) > slp->ns_reclen) {
2338 m2 = m_copym(m, 0, slp->ns_reclen - len,
2346 m->m_data += slp->ns_reclen - len;
2347 m->m_len -= slp->ns_reclen - len;
2348 len = slp->ns_reclen;
2350 return (EWOULDBLOCK);
2352 } else if ((len + m->m_len) == slp->ns_reclen) {
2372 * Accumulate the fragments into a record.
2374 mpp = &slp->ns_frag;
2376 mpp = &((*mpp)->m_next);
2378 if (slp->ns_flag & SLP_LASTFRAG) {
2379 struct nfsrv_rec *rec;
2380 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2381 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2383 m_freem(slp->ns_frag);
2385 nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2386 rec->nr_address = NULL;
2387 rec->nr_packet = slp->ns_frag;
2388 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2392 slp->ns_frag = NULL;
2398 * Parse an RPC header.
2401 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2402 struct nfsrv_descript **ndp)
2404 struct nfsrv_rec *rec;
2406 struct sockaddr *nam;
2407 struct nfsrv_descript *nd;
2411 if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2413 rec = STAILQ_FIRST(&slp->ns_rec);
2414 STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2415 KKASSERT(slp->ns_numrec > 0);
2417 nam = rec->nr_address;
2419 kfree(rec, M_NFSRVDESC);
2420 MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2421 M_NFSRVDESC, M_WAITOK);
2422 nd->nd_md = nd->nd_mrep = m;
2424 nd->nd_dpos = mtod(m, caddr_t);
2425 error = nfs_getreq(nd, nfsd, TRUE);
2428 FREE(nam, M_SONAME);
2430 kfree((caddr_t)nd, M_NFSRVDESC);
2439 * Try to assign service sockets to nfsd threads based on the number
2440 * of new rpc requests that have been queued on the service socket.
2442 * If no nfsd's are available or additonal requests are pending, set the
2443 * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2444 * the work in the nfssvc_sock list when it is finished processing its
2445 * current work. This flag is only cleared when an nfsd can not find
2446 * any new work to perform.
2449 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2453 if ((slp->ns_flag & SLP_VALID) == 0)
2457 TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2458 if (nd->nfsd_flag & NFSD_WAITING) {
2459 nd->nfsd_flag &= ~NFSD_WAITING;
2461 panic("nfsd wakeup");
2464 wakeup((caddr_t)nd);
2465 if (--nparallel == 0)
2470 slp->ns_flag |= SLP_DOREC;
2471 nfsd_head_flag |= NFSD_CHECKSLP;
2474 #endif /* NFS_NOSERVER */