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 * RTT calculations are scaled by 256 (8 bits). A proper fractional
87 * RTT will still be calculated even with a slow NFS timer.
89 #define NFS_SRTT(r) (r)->r_nmp->nm_srtt[proct[(r)->r_procnum]]
90 #define NFS_SDRTT(r) (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum]]
91 #define NFS_RTT_SCALE_BITS 8 /* bits */
92 #define NFS_RTT_SCALE 256 /* value */
95 * Defines which timer to use for the procnum.
102 static int proct[NFS_NPROCS] = {
103 0, 1, 0, 2, 1, 3, 3, 4, 0, 0, /* 00-09 */
104 0, 0, 0, 0, 0, 0, 3, 3, 0, 0, /* 10-19 */
105 0, 5, 0, 0, 0, 0, /* 20-29 */
108 static int multt[NFS_NPROCS] = {
109 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, /* 00-09 */
110 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, /* 10-19 */
111 1, 2, 1, 1, 1, 1, /* 20-29 */
114 static int nfs_backoff[8] = { 2, 3, 5, 8, 13, 21, 34, 55 };
115 static int nfs_realign_test;
116 static int nfs_realign_count;
117 static int nfs_showrtt;
118 static int nfs_showrexmit;
119 int nfs_maxasyncbio = NFS_MAXASYNCBIO;
121 SYSCTL_DECL(_vfs_nfs);
123 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0, "");
124 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0, "");
125 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrtt, CTLFLAG_RW, &nfs_showrtt, 0, "");
126 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrexmit, CTLFLAG_RW, &nfs_showrexmit, 0, "");
127 SYSCTL_INT(_vfs_nfs, OID_AUTO, maxasyncbio, CTLFLAG_RW, &nfs_maxasyncbio, 0, "");
129 static int nfs_request_setup(nfsm_info_t info);
130 static int nfs_request_auth(struct nfsreq *rep);
131 static int nfs_request_try(struct nfsreq *rep);
132 static int nfs_request_waitreply(struct nfsreq *rep);
133 static int nfs_request_processreply(nfsm_info_t info, int);
136 struct nfsrtt nfsrtt;
137 struct callout nfs_timer_handle;
139 static int nfs_msg (struct thread *,char *,char *);
140 static int nfs_rcvlock (struct nfsmount *nmp, struct nfsreq *myreq);
141 static void nfs_rcvunlock (struct nfsmount *nmp);
142 static void nfs_realign (struct mbuf **pm, int hsiz);
143 static int nfs_receive (struct nfsmount *nmp, struct nfsreq *rep,
144 struct sockaddr **aname, struct mbuf **mp);
145 static void nfs_softterm (struct nfsreq *rep, int islocked);
146 static void nfs_hardterm (struct nfsreq *rep, int islocked);
147 static int nfs_reconnect (struct nfsmount *nmp, struct nfsreq *rep);
149 static int nfsrv_getstream (struct nfssvc_sock *, int, int *);
150 static void nfs_timer_req(struct nfsreq *req);
152 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
153 struct nfssvc_sock *slp,
155 struct mbuf **mreqp) = {
183 #endif /* NFS_NOSERVER */
186 * Initialize sockets and congestion for a new NFS connection.
187 * We do not free the sockaddr if error.
190 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
194 struct sockaddr *saddr;
195 struct sockaddr_in *sin;
196 struct thread *td = &thread0; /* only used for socreate and sobind */
200 error = socreate(saddr->sa_family, &nmp->nm_so, nmp->nm_sotype,
201 nmp->nm_soproto, td);
205 nmp->nm_soflags = so->so_proto->pr_flags;
208 * Some servers require that the client port be a reserved port number.
210 if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
213 struct sockaddr_in ssin;
215 bzero(&sopt, sizeof sopt);
216 ip = IP_PORTRANGE_LOW;
217 sopt.sopt_level = IPPROTO_IP;
218 sopt.sopt_name = IP_PORTRANGE;
219 sopt.sopt_val = (void *)&ip;
220 sopt.sopt_valsize = sizeof(ip);
222 error = sosetopt(so, &sopt);
225 bzero(&ssin, sizeof ssin);
227 sin->sin_len = sizeof (struct sockaddr_in);
228 sin->sin_family = AF_INET;
229 sin->sin_addr.s_addr = INADDR_ANY;
230 sin->sin_port = htons(0);
231 error = sobind(so, (struct sockaddr *)sin, td);
234 bzero(&sopt, sizeof sopt);
235 ip = IP_PORTRANGE_DEFAULT;
236 sopt.sopt_level = IPPROTO_IP;
237 sopt.sopt_name = IP_PORTRANGE;
238 sopt.sopt_val = (void *)&ip;
239 sopt.sopt_valsize = sizeof(ip);
241 error = sosetopt(so, &sopt);
247 * Protocols that do not require connections may be optionally left
248 * unconnected for servers that reply from a port other than NFS_PORT.
250 if (nmp->nm_flag & NFSMNT_NOCONN) {
251 if (nmp->nm_soflags & PR_CONNREQUIRED) {
256 error = soconnect(so, nmp->nm_nam, td);
261 * Wait for the connection to complete. Cribbed from the
262 * connect system call but with the wait timing out so
263 * that interruptible mounts don't hang here for a long time.
266 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
267 (void) tsleep((caddr_t)&so->so_timeo, 0,
269 if ((so->so_state & SS_ISCONNECTING) &&
270 so->so_error == 0 && rep &&
271 (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
272 so->so_state &= ~SS_ISCONNECTING;
278 error = so->so_error;
285 so->so_rcv.ssb_timeo = (5 * hz);
286 so->so_snd.ssb_timeo = (5 * hz);
289 * Get buffer reservation size from sysctl, but impose reasonable
292 if (nmp->nm_sotype == SOCK_STREAM) {
293 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
297 bzero(&sopt, sizeof sopt);
298 sopt.sopt_level = SOL_SOCKET;
299 sopt.sopt_name = SO_KEEPALIVE;
300 sopt.sopt_val = &val;
301 sopt.sopt_valsize = sizeof val;
305 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
309 bzero(&sopt, sizeof sopt);
310 sopt.sopt_level = IPPROTO_TCP;
311 sopt.sopt_name = TCP_NODELAY;
312 sopt.sopt_val = &val;
313 sopt.sopt_valsize = sizeof val;
318 error = soreserve(so, nfs_soreserve, nfs_soreserve, NULL);
321 so->so_rcv.ssb_flags |= SSB_NOINTR;
322 so->so_snd.ssb_flags |= SSB_NOINTR;
324 /* Initialize other non-zero congestion variables */
325 nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] =
326 nmp->nm_srtt[3] = (NFS_TIMEO << NFS_RTT_SCALE_BITS);
327 nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
328 nmp->nm_sdrtt[3] = 0;
329 nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
330 nmp->nm_timeouts = 0;
340 * Called when a connection is broken on a reliable protocol.
341 * - clean up the old socket
342 * - nfs_connect() again
343 * - set R_NEEDSXMIT for all outstanding requests on mount point
344 * If this fails the mount point is DEAD!
345 * nb: Must be called with the nfs_sndlock() set on the mount point.
348 nfs_reconnect(struct nfsmount *nmp, struct nfsreq *rep)
354 while ((error = nfs_connect(nmp, rep)) != 0) {
355 if (error == EINTR || error == ERESTART)
357 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
361 * Loop through outstanding request list and fix up all requests
365 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
366 KKASSERT(req->r_nmp == nmp);
367 req->r_flags |= R_NEEDSXMIT;
374 * NFS disconnect. Clean up and unlink.
377 nfs_disconnect(struct nfsmount *nmp)
384 soshutdown(so, SHUT_RDWR);
385 soclose(so, FNONBLOCK);
390 nfs_safedisconnect(struct nfsmount *nmp)
392 nfs_rcvlock(nmp, NULL);
398 * This is the nfs send routine. For connection based socket types, it
399 * must be called with an nfs_sndlock() on the socket.
400 * "rep == NULL" indicates that it has been called from a server.
401 * For the client side:
402 * - return EINTR if the RPC is terminated, 0 otherwise
403 * - set R_NEEDSXMIT if the send fails for any reason
404 * - do any cleanup required by recoverable socket errors (?)
405 * For the server side:
406 * - return EINTR or ERESTART if interrupted by a signal
407 * - return EPIPE if a connection is lost for connection based sockets (TCP...)
408 * - do any cleanup required by recoverable socket errors (?)
411 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
414 struct sockaddr *sendnam;
415 int error, soflags, flags;
418 if (rep->r_flags & R_SOFTTERM) {
422 if ((so = rep->r_nmp->nm_so) == NULL) {
423 rep->r_flags |= R_NEEDSXMIT;
427 rep->r_flags &= ~R_NEEDSXMIT;
428 soflags = rep->r_nmp->nm_soflags;
430 soflags = so->so_proto->pr_flags;
432 if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
436 if (so->so_type == SOCK_SEQPACKET)
441 error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
444 * ENOBUFS for dgram sockets is transient and non fatal.
445 * No need to log, and no need to break a soft mount.
447 if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
450 * do backoff retransmit on client
453 if ((rep->r_nmp->nm_state & NFSSTA_SENDSPACE) == 0) {
454 rep->r_nmp->nm_state |= NFSSTA_SENDSPACE;
455 kprintf("Warning: NFS: Insufficient sendspace "
457 "\t You must increase vfs.nfs.soreserve"
458 "or decrease vfs.nfs.maxasyncbio\n",
459 so->so_snd.ssb_hiwat);
461 rep->r_flags |= R_NEEDSXMIT;
467 log(LOG_INFO, "nfs send error %d for server %s\n",error,
468 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
470 * Deal with errors for the client side.
472 if (rep->r_flags & R_SOFTTERM)
475 rep->r_flags |= R_NEEDSXMIT;
477 log(LOG_INFO, "nfsd send error %d\n", error);
481 * Handle any recoverable (soft) socket errors here. (?)
483 if (error != EINTR && error != ERESTART &&
484 error != EWOULDBLOCK && error != EPIPE)
491 * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
492 * done by soreceive(), but for SOCK_STREAM we must deal with the Record
493 * Mark and consolidate the data into a new mbuf list.
494 * nb: Sometimes TCP passes the data up to soreceive() in long lists of
496 * For SOCK_STREAM we must be very careful to read an entire record once
497 * we have read any of it, even if the system call has been interrupted.
500 nfs_receive(struct nfsmount *nmp, struct nfsreq *rep,
501 struct sockaddr **aname, struct mbuf **mp)
508 struct mbuf *control;
510 struct sockaddr **getnam;
511 int error, sotype, rcvflg;
512 struct thread *td = curthread; /* XXX */
515 * Set up arguments for soreceive()
519 sotype = nmp->nm_sotype;
522 * For reliable protocols, lock against other senders/receivers
523 * in case a reconnect is necessary.
524 * For SOCK_STREAM, first get the Record Mark to find out how much
525 * more there is to get.
526 * We must lock the socket against other receivers
527 * until we have an entire rpc request/reply.
529 if (sotype != SOCK_DGRAM) {
530 error = nfs_sndlock(nmp, rep);
535 * Check for fatal errors and resending request.
538 * Ugh: If a reconnect attempt just happened, nm_so
539 * would have changed. NULL indicates a failed
540 * attempt that has essentially shut down this
543 if (rep && (rep->r_mrep || (rep->r_flags & R_SOFTTERM))) {
549 error = nfs_reconnect(nmp, rep);
556 while (rep && (rep->r_flags & R_NEEDSXMIT)) {
557 m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
558 nfsstats.rpcretries++;
559 error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
561 if (error == EINTR || error == ERESTART ||
562 (error = nfs_reconnect(nmp, rep)) != 0) {
570 if (sotype == SOCK_STREAM) {
572 * Get the length marker from the stream
574 aio.iov_base = (caddr_t)&len;
575 aio.iov_len = sizeof(u_int32_t);
578 auio.uio_segflg = UIO_SYSSPACE;
579 auio.uio_rw = UIO_READ;
581 auio.uio_resid = sizeof(u_int32_t);
584 rcvflg = MSG_WAITALL;
585 error = so_pru_soreceive(so, NULL, &auio, NULL,
587 if (error == EWOULDBLOCK && rep) {
588 if (rep->r_flags & R_SOFTTERM)
591 } while (error == EWOULDBLOCK);
593 if (error == 0 && auio.uio_resid > 0) {
595 * Only log short packets if not EOF
597 if (auio.uio_resid != sizeof(u_int32_t))
599 "short receive (%d/%d) from nfs server %s\n",
600 (int)(sizeof(u_int32_t) - auio.uio_resid),
601 (int)sizeof(u_int32_t),
602 nmp->nm_mountp->mnt_stat.f_mntfromname);
607 len = ntohl(len) & ~0x80000000;
609 * This is SERIOUS! We are out of sync with the sender
610 * and forcing a disconnect/reconnect is all I can do.
612 if (len > NFS_MAXPACKET) {
613 log(LOG_ERR, "%s (%d) from nfs server %s\n",
614 "impossible packet length",
616 nmp->nm_mountp->mnt_stat.f_mntfromname);
622 * Get the rest of the packet as an mbuf chain
626 rcvflg = MSG_WAITALL;
627 error = so_pru_soreceive(so, NULL, NULL, &sio,
629 } while (error == EWOULDBLOCK || error == EINTR ||
631 if (error == 0 && sio.sb_cc != len) {
634 "short receive (%d/%d) from nfs server %s\n",
635 len - auio.uio_resid, len,
636 nmp->nm_mountp->mnt_stat.f_mntfromname);
642 * Non-stream, so get the whole packet by not
643 * specifying MSG_WAITALL and by specifying a large
646 * We have no use for control msg., but must grab them
647 * and then throw them away so we know what is going
650 sbinit(&sio, 100000000);
653 error = so_pru_soreceive(so, NULL, NULL, &sio,
657 if (error == EWOULDBLOCK && rep) {
658 if (rep->r_flags & R_SOFTTERM) {
663 } while (error == EWOULDBLOCK ||
664 (error == 0 && sio.sb_mb == NULL && control));
665 if ((rcvflg & MSG_EOR) == 0)
667 if (error == 0 && sio.sb_mb == NULL)
673 if (error && error != EINTR && error != ERESTART) {
676 if (error != EPIPE) {
678 "receive error %d from nfs server %s\n",
680 nmp->nm_mountp->mnt_stat.f_mntfromname);
682 error = nfs_sndlock(nmp, rep);
684 error = nfs_reconnect(nmp, rep);
692 if ((so = nmp->nm_so) == NULL)
694 if (so->so_state & SS_ISCONNECTED)
698 sbinit(&sio, 100000000);
701 error = so_pru_soreceive(so, getnam, NULL, &sio,
703 if (error == EWOULDBLOCK && rep &&
704 (rep->r_flags & R_SOFTTERM)) {
708 } while (error == EWOULDBLOCK);
714 * A shutdown may result in no error and no mbuf.
717 if (*mp == NULL && error == 0)
726 * Search for any mbufs that are not a multiple of 4 bytes long
727 * or with m_data not longword aligned.
728 * These could cause pointer alignment problems, so copy them to
729 * well aligned mbufs.
731 nfs_realign(mp, 5 * NFSX_UNSIGNED);
736 * Implement receipt of reply on a socket.
738 * We must search through the list of received datagrams matching them
739 * with outstanding requests using the xid, until ours is found.
741 * If myrep is NULL we process packets on the socket until
742 * interrupted or until nm_reqrxq is non-empty.
746 nfs_reply(struct nfsmount *nmp, struct nfsreq *myrep)
749 struct sockaddr *nam;
753 struct nfsm_info info;
756 * Loop around until we get our own reply
760 * Lock against other receivers so that I don't get stuck in
761 * sbwait() after someone else has received my reply for me.
762 * Also necessary for connection based protocols to avoid
763 * race conditions during a reconnect.
765 * If nfs_rcvlock() returns EALREADY, that means that
766 * the reply has already been recieved by another
767 * process and we can return immediately. In this
768 * case, the lock is not taken to avoid races with
773 error = nfs_rcvlock(nmp, myrep);
774 if (error == EALREADY)
780 * If myrep is NULL we are the receiver helper thread.
781 * Stop waiting for incoming replies if there are
782 * messages sitting on reqrxq that we need to process,
783 * or if a shutdown request is pending.
785 if (myrep == NULL && (TAILQ_FIRST(&nmp->nm_reqrxq) ||
786 nmp->nm_rxstate > NFSSVC_PENDING)) {
792 * Get the next Rpc reply off the socket
794 * We cannot release the receive lock until we've
795 * filled in rep->r_mrep, otherwise a waiting
796 * thread may deadlock in soreceive with no incoming
799 error = nfs_receive(nmp, myrep, &nam, &info.mrep);
802 * Ignore routing errors on connectionless protocols??
805 if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
806 if (nmp->nm_so == NULL)
808 nmp->nm_so->so_error = 0;
817 * Get the xid and check that it is an rpc reply
820 info.dpos = mtod(info.md, caddr_t);
821 NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
823 if (*tl != rpc_reply) {
824 nfsstats.rpcinvalid++;
833 * Loop through the request list to match up the reply
834 * Iff no match, just drop the datagram. On match, set
835 * r_mrep atomically to prevent the timer from messing
836 * around with the request after we have exited the critical
840 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
841 if (rep->r_mrep == NULL && rxid == rep->r_xid)
846 * Fill in the rest of the reply if we found a match.
848 * Deal with duplicate responses if there was no match.
852 rep->r_dpos = info.dpos;
856 rt = &nfsrtt.rttl[nfsrtt.pos];
857 rt->proc = rep->r_procnum;
860 rt->cwnd = nmp->nm_maxasync_scaled;
861 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
862 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
863 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
864 getmicrotime(&rt->tstamp);
865 if (rep->r_flags & R_TIMING)
866 rt->rtt = rep->r_rtt;
869 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
873 * New congestion control is based only on async
876 if (nmp->nm_maxasync_scaled < NFS_MAXASYNC_SCALED)
877 ++nmp->nm_maxasync_scaled;
878 if (rep->r_flags & R_SENT) {
879 rep->r_flags &= ~R_SENT;
882 * Update rtt using a gain of 0.125 on the mean
883 * and a gain of 0.25 on the deviation.
885 * NOTE SRTT/SDRTT are only good if R_TIMING is set.
887 if ((rep->r_flags & R_TIMING) && rep->r_rexmit == 0) {
889 * Since the timer resolution of
890 * NFS_HZ is so course, it can often
891 * result in r_rtt == 0. Since
892 * r_rtt == N means that the actual
893 * rtt is between N+dt and N+2-dt ticks,
899 #define NFSRSB NFS_RTT_SCALE_BITS
900 n = ((NFS_SRTT(rep) * 7) +
901 (rep->r_rtt << NFSRSB)) >> 3;
902 d = n - NFS_SRTT(rep);
906 * Don't let the jitter calculation decay
907 * too quickly, but we want a fast rampup.
912 if (d < NFS_SDRTT(rep))
913 n = ((NFS_SDRTT(rep) * 15) + d) >> 4;
915 n = ((NFS_SDRTT(rep) * 3) + d) >> 2;
919 nmp->nm_timeouts = 0;
920 rep->r_mrep = info.mrep;
921 nfs_hardterm(rep, 0);
924 * Extract vers, prog, nfsver, procnum. A duplicate
925 * response means we didn't wait long enough so
926 * we increase the SRTT to avoid future spurious
929 u_int procnum = nmp->nm_lastreprocnum;
932 if (procnum < NFS_NPROCS && proct[procnum]) {
935 n = nmp->nm_srtt[proct[procnum]];
936 n += NFS_ASYSCALE * NFS_HZ;
937 if (n < NFS_ASYSCALE * NFS_HZ * 10)
938 n = NFS_ASYSCALE * NFS_HZ * 10;
939 nmp->nm_srtt[proct[procnum]] = n;
946 * If not matched to a request, drop it.
947 * If it's mine, get out.
950 nfsstats.rpcunexpected++;
953 } else if (rep == myrep) {
954 if (rep->r_mrep == NULL)
955 panic("nfsreply nil");
962 * Run the request state machine until the target state is reached
963 * or a fatal error occurs. The target state is not run. Specifying
964 * a target of NFSM_STATE_DONE runs the state machine until the rpc
967 * EINPROGRESS is returned for all states other then the DONE state,
968 * indicating that the rpc is still in progress.
971 nfs_request(struct nfsm_info *info, nfsm_state_t bstate, nfsm_state_t estate)
975 while (info->state >= bstate && info->state < estate) {
976 switch(info->state) {
977 case NFSM_STATE_SETUP:
979 * Setup the nfsreq. Any error which occurs during
980 * this state is fatal.
982 info->error = nfs_request_setup(info);
984 info->state = NFSM_STATE_DONE;
985 return (info->error);
988 req->r_mrp = &info->mrep;
989 req->r_mdp = &info->md;
990 req->r_dposp = &info->dpos;
991 info->state = NFSM_STATE_AUTH;
994 case NFSM_STATE_AUTH:
996 * Authenticate the nfsreq. Any error which occurs
997 * during this state is fatal.
999 info->error = nfs_request_auth(info->req);
1001 info->state = NFSM_STATE_DONE;
1002 return (info->error);
1004 info->state = NFSM_STATE_TRY;
1007 case NFSM_STATE_TRY:
1009 * Transmit or retransmit attempt. An error in this
1010 * state is ignored and we always move on to the
1013 * This can trivially race the receiver if the
1014 * request is asynchronous. nfs_request_try()
1015 * will thus set the state for us and we
1016 * must also return immediately if we are
1017 * running an async state machine, because
1018 * info can become invalid due to races after
1021 if (info->req->r_flags & R_ASYNC) {
1022 nfs_request_try(info->req);
1023 if (estate == NFSM_STATE_WAITREPLY)
1024 return (EINPROGRESS);
1026 nfs_request_try(info->req);
1027 info->state = NFSM_STATE_WAITREPLY;
1030 case NFSM_STATE_WAITREPLY:
1032 * Wait for a reply or timeout and move on to the
1033 * next state. The error returned by this state
1034 * is passed to the processing code in the next
1037 info->error = nfs_request_waitreply(info->req);
1038 info->state = NFSM_STATE_PROCESSREPLY;
1040 case NFSM_STATE_PROCESSREPLY:
1042 * Process the reply or timeout. Errors which occur
1043 * in this state may cause the state machine to
1044 * go back to an earlier state, and are fatal
1047 info->error = nfs_request_processreply(info,
1049 switch(info->error) {
1051 info->state = NFSM_STATE_AUTH;
1054 info->state = NFSM_STATE_TRY;
1058 * Operation complete, with or without an
1059 * error. We are done.
1062 info->state = NFSM_STATE_DONE;
1063 return (info->error);
1066 case NFSM_STATE_DONE:
1068 * Shouldn't be reached
1070 return (info->error);
1076 * If we are done return the error code (if any).
1077 * Otherwise return EINPROGRESS.
1079 if (info->state == NFSM_STATE_DONE)
1080 return (info->error);
1081 return (EINPROGRESS);
1085 * nfs_request - goes something like this
1086 * - fill in request struct
1087 * - links it into list
1088 * - calls nfs_send() for first transmit
1089 * - calls nfs_receive() to get reply
1090 * - break down rpc header and return with nfs reply pointed to
1092 * nb: always frees up mreq mbuf list
1095 nfs_request_setup(nfsm_info_t info)
1098 struct nfsmount *nmp;
1103 * Reject requests while attempting a forced unmount.
1105 if (info->vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
1106 m_freem(info->mreq);
1110 nmp = VFSTONFS(info->vp->v_mount);
1111 req = kmalloc(sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
1113 req->r_vp = info->vp;
1114 req->r_td = info->td;
1115 req->r_procnum = info->procnum;
1117 req->r_cred = info->cred;
1125 req->r_mrest = info->mreq;
1126 req->r_mrest_len = i;
1129 * The presence of a non-NULL r_info in req indicates
1130 * async completion via our helper threads. See the receiver
1135 req->r_flags = R_ASYNC;
1145 nfs_request_auth(struct nfsreq *rep)
1147 struct nfsmount *nmp = rep->r_nmp;
1149 char nickv[RPCX_NICKVERF];
1150 int error = 0, auth_len, auth_type;
1153 char *auth_str, *verf_str;
1157 rep->r_failed_auth = 0;
1160 * Get the RPC header with authorization.
1162 verf_str = auth_str = NULL;
1163 if (nmp->nm_flag & NFSMNT_KERB) {
1165 verf_len = sizeof (nickv);
1166 auth_type = RPCAUTH_KERB4;
1167 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1168 if (rep->r_failed_auth ||
1169 nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1170 verf_str, verf_len)) {
1171 error = nfs_getauth(nmp, rep, cred, &auth_str,
1172 &auth_len, verf_str, &verf_len, rep->r_key);
1174 m_freem(rep->r_mrest);
1175 rep->r_mrest = NULL;
1176 kfree((caddr_t)rep, M_NFSREQ);
1181 auth_type = RPCAUTH_UNIX;
1182 if (cred->cr_ngroups < 1)
1183 panic("nfsreq nogrps");
1184 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1185 nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1188 m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1189 auth_len, auth_str, verf_len, verf_str,
1190 rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1191 rep->r_mrest = NULL;
1193 kfree(auth_str, M_TEMP);
1196 * For stream protocols, insert a Sun RPC Record Mark.
1198 if (nmp->nm_sotype == SOCK_STREAM) {
1199 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1201 kfree(rep, M_NFSREQ);
1204 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1205 (m->m_pkthdr.len - NFSX_UNSIGNED));
1213 nfs_request_try(struct nfsreq *rep)
1215 struct nfsmount *nmp = rep->r_nmp;
1220 * Request is not on any queue, only the owner has access to it
1221 * so it should not be locked by anyone atm.
1223 * Interlock to prevent races. While locked the only remote
1224 * action possible is for r_mrep to be set (once we enqueue it).
1226 if (rep->r_flags == 0xdeadc0de) {
1228 panic("flags nbad\n");
1230 KKASSERT((rep->r_flags & (R_LOCKED | R_ONREQQ)) == 0);
1231 if (nmp->nm_flag & NFSMNT_SOFT)
1232 rep->r_retry = nmp->nm_retry;
1234 rep->r_retry = NFS_MAXREXMIT + 1; /* past clip limit */
1235 rep->r_rtt = rep->r_rexmit = 0;
1236 if (proct[rep->r_procnum] > 0)
1237 rep->r_flags |= R_TIMING | R_LOCKED;
1239 rep->r_flags |= R_LOCKED;
1243 * Do the client side RPC.
1245 nfsstats.rpcrequests++;
1248 * Chain request into list of outstanding requests. Be sure
1249 * to put it LAST so timer finds oldest requests first. Note
1250 * that our control of R_LOCKED prevents the request from
1251 * getting ripped out from under us or transmitted by the
1254 * For requests with info structures we must atomically set the
1255 * info's state because the structure could become invalid upon
1256 * return due to races (i.e., if async)
1259 mtx_link_init(&rep->r_link);
1260 TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1261 rep->r_flags |= R_ONREQQ;
1263 if (rep->r_flags & R_ASYNC)
1264 rep->r_info->state = NFSM_STATE_WAITREPLY;
1270 * Send if we can. Congestion control is not handled here any more
1271 * becausing trying to defer the initial send based on the nfs_timer
1272 * requires having a very fast nfs_timer, which is silly.
1275 if (nmp->nm_soflags & PR_CONNREQUIRED)
1276 error = nfs_sndlock(nmp, rep);
1278 m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1279 error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1280 if (nmp->nm_soflags & PR_CONNREQUIRED)
1282 rep->r_flags &= ~R_NEEDSXMIT;
1283 if ((rep->r_flags & R_SENT) == 0) {
1284 rep->r_flags |= R_SENT;
1287 rep->r_flags |= R_NEEDSXMIT;
1290 rep->r_flags |= R_NEEDSXMIT;
1297 * Release the lock. The only remote action that may have occurred
1298 * would have been the setting of rep->r_mrep. If this occured
1299 * and the request was async we have to move it to the reader
1300 * thread's queue for action.
1302 * For async requests also make sure the reader is woken up so
1303 * it gets on the socket to read responses.
1306 if (rep->r_flags & R_ASYNC) {
1308 nfs_hardterm(rep, 1);
1309 rep->r_flags &= ~R_LOCKED;
1310 nfssvc_iod_reader_wakeup(nmp);
1312 rep->r_flags &= ~R_LOCKED;
1314 if (rep->r_flags & R_WANTED) {
1315 rep->r_flags &= ~R_WANTED;
1323 * This code is only called for synchronous requests. Completed synchronous
1324 * requests are left on reqq and we remove them before moving on to the
1328 nfs_request_waitreply(struct nfsreq *rep)
1330 struct nfsmount *nmp = rep->r_nmp;
1333 KKASSERT((rep->r_flags & R_ASYNC) == 0);
1336 * Wait until the request is finished.
1338 error = nfs_reply(nmp, rep);
1341 * RPC done, unlink the request, but don't rip it out from under
1342 * the callout timer.
1344 * Once unlinked no other receiver or the timer will have
1345 * visibility, so we do not have to set R_LOCKED.
1348 while (rep->r_flags & R_LOCKED) {
1349 rep->r_flags |= R_WANTED;
1350 tsleep(rep, 0, "nfstrac", 0);
1352 KKASSERT(rep->r_flags & R_ONREQQ);
1353 TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1354 rep->r_flags &= ~R_ONREQQ;
1356 if (TAILQ_FIRST(&nmp->nm_bioq) &&
1357 nmp->nm_reqqlen == NFS_MAXASYNCBIO * 2 / 3) {
1358 nfssvc_iod_writer_wakeup(nmp);
1363 * Decrement the outstanding request count.
1365 if (rep->r_flags & R_SENT) {
1366 rep->r_flags &= ~R_SENT;
1372 * Process reply with error returned from nfs_requet_waitreply().
1374 * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1375 * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1378 nfs_request_processreply(nfsm_info_t info, int error)
1380 struct nfsreq *req = info->req;
1381 struct nfsmount *nmp = req->r_nmp;
1387 * If there was a successful reply and a tprintf msg.
1388 * tprintf a response.
1390 if (error == 0 && (req->r_flags & R_TPRINTFMSG)) {
1391 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1394 info->mrep = req->r_mrep;
1395 info->md = req->r_md;
1396 info->dpos = req->r_dpos;
1398 m_freem(req->r_mreq);
1400 kfree(req, M_NFSREQ);
1406 * break down the rpc header and check if ok
1408 NULLOUT(tl = nfsm_dissect(info, 3 * NFSX_UNSIGNED));
1409 if (*tl++ == rpc_msgdenied) {
1410 if (*tl == rpc_mismatch) {
1412 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1413 *tl++ == rpc_autherr) {
1414 if (req->r_failed_auth == 0) {
1415 req->r_failed_auth++;
1416 req->r_mheadend->m_next = NULL;
1417 m_freem(info->mrep);
1419 m_freem(req->r_mreq);
1427 m_freem(info->mrep);
1429 m_freem(req->r_mreq);
1431 kfree(req, M_NFSREQ);
1437 * Grab any Kerberos verifier, otherwise just throw it away.
1439 verf_type = fxdr_unsigned(int, *tl++);
1440 i = fxdr_unsigned(int32_t, *tl);
1441 if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1442 error = nfs_savenickauth(nmp, req->r_cred, i, req->r_key,
1443 &info->md, &info->dpos, info->mrep);
1447 ERROROUT(nfsm_adv(info, nfsm_rndup(i)));
1449 NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1452 NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1454 error = fxdr_unsigned(int, *tl);
1457 * Does anyone even implement this? Just impose
1460 if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1461 error == NFSERR_TRYLATER) {
1462 m_freem(info->mrep);
1466 tsleep((caddr_t)&lbolt, 0, "nqnfstry", 0);
1467 return (EAGAIN); /* goto tryagain */
1471 * If the File Handle was stale, invalidate the
1472 * lookup cache, just in case.
1474 * To avoid namecache<->vnode deadlocks we must
1475 * release the vnode lock if we hold it.
1477 if (error == ESTALE) {
1478 struct vnode *vp = req->r_vp;
1481 ltype = lockstatus(&vp->v_lock, curthread);
1482 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1483 lockmgr(&vp->v_lock, LK_RELEASE);
1484 cache_inval_vp(vp, CINV_CHILDREN);
1485 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1486 lockmgr(&vp->v_lock, ltype);
1488 if (nmp->nm_flag & NFSMNT_NFSV3) {
1489 KKASSERT(*req->r_mrp == info->mrep);
1490 KKASSERT(*req->r_mdp == info->md);
1491 KKASSERT(*req->r_dposp == info->dpos);
1492 error |= NFSERR_RETERR;
1494 m_freem(info->mrep);
1497 m_freem(req->r_mreq);
1499 kfree(req, M_NFSREQ);
1504 KKASSERT(*req->r_mrp == info->mrep);
1505 KKASSERT(*req->r_mdp == info->md);
1506 KKASSERT(*req->r_dposp == info->dpos);
1507 m_freem(req->r_mreq);
1509 FREE(req, M_NFSREQ);
1512 m_freem(info->mrep);
1514 error = EPROTONOSUPPORT;
1516 m_freem(req->r_mreq);
1518 kfree(req, M_NFSREQ);
1523 #ifndef NFS_NOSERVER
1525 * Generate the rpc reply header
1526 * siz arg. is used to decide if adding a cluster is worthwhile
1529 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1530 int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1533 struct nfsm_info info;
1535 siz += RPC_REPLYSIZ;
1536 info.mb = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1537 info.mreq = info.mb;
1538 info.mreq->m_pkthdr.len = 0;
1540 * If this is not a cluster, try and leave leading space
1541 * for the lower level headers.
1543 if ((max_hdr + siz) < MINCLSIZE)
1544 info.mreq->m_data += max_hdr;
1545 tl = mtod(info.mreq, u_int32_t *);
1546 info.mreq->m_len = 6 * NFSX_UNSIGNED;
1547 info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1548 *tl++ = txdr_unsigned(nd->nd_retxid);
1550 if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1551 *tl++ = rpc_msgdenied;
1552 if (err & NFSERR_AUTHERR) {
1553 *tl++ = rpc_autherr;
1554 *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1555 info.mreq->m_len -= NFSX_UNSIGNED;
1556 info.bpos -= NFSX_UNSIGNED;
1558 *tl++ = rpc_mismatch;
1559 *tl++ = txdr_unsigned(RPC_VER2);
1560 *tl = txdr_unsigned(RPC_VER2);
1563 *tl++ = rpc_msgaccepted;
1566 * For Kerberos authentication, we must send the nickname
1567 * verifier back, otherwise just RPCAUTH_NULL.
1569 if (nd->nd_flag & ND_KERBFULL) {
1570 struct nfsuid *nuidp;
1571 struct timeval ktvin, ktvout;
1573 for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1574 nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1575 if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1576 (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1577 &nuidp->nu_haddr, nd->nd_nam2)))
1582 txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1584 txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1587 * Encrypt the timestamp in ecb mode using the
1594 *tl++ = rpc_auth_kerb;
1595 *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1596 *tl = ktvout.tv_sec;
1597 tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1598 *tl++ = ktvout.tv_usec;
1599 *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1610 *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1613 *tl = txdr_unsigned(RPC_PROGMISMATCH);
1614 tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1615 *tl++ = txdr_unsigned(2);
1616 *tl = txdr_unsigned(3);
1619 *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1622 *tl = txdr_unsigned(RPC_GARBAGE);
1626 if (err != NFSERR_RETVOID) {
1627 tl = nfsm_build(&info, NFSX_UNSIGNED);
1629 *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1641 if (err != 0 && err != NFSERR_RETVOID)
1642 nfsstats.srvrpc_errs++;
1647 #endif /* NFS_NOSERVER */
1650 * Nfs timer routine.
1652 * Scan the nfsreq list and retranmit any requests that have timed out
1653 * To avoid retransmission attempts on STREAM sockets (in the future) make
1654 * sure to set the r_retry field to 0 (implies nm_retry == 0).
1656 * Requests with attached responses, terminated requests, and
1657 * locked requests are ignored. Locked requests will be picked up
1658 * in a later timer call.
1661 nfs_timer(void *arg /* never used */)
1663 struct nfsmount *nmp;
1665 #ifndef NFS_NOSERVER
1666 struct nfssvc_sock *slp;
1668 #endif /* NFS_NOSERVER */
1671 TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1672 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1673 KKASSERT(nmp == req->r_nmp);
1676 if (req->r_flags & (R_SOFTTERM | R_LOCKED))
1678 req->r_flags |= R_LOCKED;
1679 if (nfs_sigintr(nmp, req, req->r_td)) {
1680 nfs_softterm(req, 1);
1684 req->r_flags &= ~R_LOCKED;
1685 if (req->r_flags & R_WANTED) {
1686 req->r_flags &= ~R_WANTED;
1691 #ifndef NFS_NOSERVER
1694 * Scan the write gathering queues for writes that need to be
1697 cur_usec = nfs_curusec();
1698 TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1699 if (slp->ns_tq.lh_first && slp->ns_tq.lh_first->nd_time<=cur_usec)
1700 nfsrv_wakenfsd(slp, 1);
1702 #endif /* NFS_NOSERVER */
1704 callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer, NULL);
1709 nfs_timer_req(struct nfsreq *req)
1711 struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1712 struct nfsmount *nmp = req->r_nmp;
1719 * rtt ticks and timeout calculation. Return if the timeout
1720 * has not been reached yet, unless the packet is flagged
1721 * for an immediate send.
1723 * The mean rtt doesn't help when we get random I/Os, we have
1724 * to multiply by fairly large numbers.
1726 if (req->r_rtt >= 0) {
1728 * Calculate the timeout to test against.
1731 if (nmp->nm_flag & NFSMNT_DUMBTIMR) {
1732 timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1733 } else if (req->r_flags & R_TIMING) {
1734 timeo = NFS_SRTT(req) + NFS_SDRTT(req);
1736 timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1738 timeo *= multt[req->r_procnum];
1739 /* timeo is still scaled by SCALE_BITS */
1741 #define NFSFS (NFS_RTT_SCALE * NFS_HZ)
1742 if (req->r_flags & R_TIMING) {
1743 static long last_time;
1744 if (nfs_showrtt && last_time != time_second) {
1745 kprintf("rpccmd %d NFS SRTT %d SDRTT %d "
1747 proct[req->r_procnum],
1748 NFS_SRTT(req), NFS_SDRTT(req),
1750 timeo % NFSFS * 1000 / NFSFS);
1751 last_time = time_second;
1757 * deal with nfs_timer jitter.
1759 timeo = (timeo >> NFS_RTT_SCALE_BITS) + 1;
1763 if (nmp->nm_timeouts > 0)
1764 timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1765 if (timeo > NFS_MAXTIMEO)
1766 timeo = NFS_MAXTIMEO;
1767 if (req->r_rtt <= timeo) {
1768 if ((req->r_flags & R_NEEDSXMIT) == 0)
1770 } else if (nmp->nm_timeouts < 8) {
1776 * Check for server not responding
1778 if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1779 req->r_rexmit > nmp->nm_deadthresh) {
1780 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1782 req->r_flags |= R_TPRINTFMSG;
1784 if (req->r_rexmit >= req->r_retry) { /* too many */
1785 nfsstats.rpctimeouts++;
1786 nfs_softterm(req, 1);
1791 * Generally disable retransmission on reliable sockets,
1792 * unless the request is flagged for immediate send.
1794 if (nmp->nm_sotype != SOCK_DGRAM) {
1795 if (++req->r_rexmit > NFS_MAXREXMIT)
1796 req->r_rexmit = NFS_MAXREXMIT;
1797 if ((req->r_flags & R_NEEDSXMIT) == 0)
1802 * Stop here if we do not have a socket!
1804 if ((so = nmp->nm_so) == NULL)
1808 * If there is enough space and the window allows.. resend it.
1810 * r_rtt is left intact in case we get an answer after the
1811 * retry that was a reply to the original packet.
1813 if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1814 (req->r_flags & (R_SENT | R_NEEDSXMIT)) &&
1815 (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1816 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1817 error = so_pru_send(so, 0, m, NULL, NULL, td);
1819 error = so_pru_send(so, 0, m, nmp->nm_nam,
1822 if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1824 req->r_flags |= R_NEEDSXMIT;
1825 } else if (req->r_mrep == NULL) {
1827 * Iff first send, start timing
1828 * else turn timing off, backoff timer
1829 * and divide congestion window by 2.
1831 * It is possible for the so_pru_send() to
1832 * block and for us to race a reply so we
1833 * only do this if the reply field has not
1834 * been filled in. R_LOCKED will prevent
1835 * the request from being ripped out from under
1838 * Record the last resent procnum to aid us
1839 * in duplicate detection on receive.
1841 if ((req->r_flags & R_NEEDSXMIT) == 0) {
1844 if (++req->r_rexmit > NFS_MAXREXMIT)
1845 req->r_rexmit = NFS_MAXREXMIT;
1846 nmp->nm_maxasync_scaled >>= 1;
1847 if (nmp->nm_maxasync_scaled < NFS_MINASYNC_SCALED)
1848 nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
1849 nfsstats.rpcretries++;
1850 nmp->nm_lastreprocnum = req->r_procnum;
1852 req->r_flags |= R_SENT;
1853 req->r_flags &= ~R_NEEDSXMIT;
1860 * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1861 * wait for all requests to complete. This is used by forced unmounts
1862 * to terminate any outstanding RPCs.
1864 * Locked requests cannot be canceled but will be marked for
1868 nfs_nmcancelreqs(struct nfsmount *nmp)
1874 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1875 if (req->r_mrep != NULL || (req->r_flags & R_SOFTTERM))
1877 nfs_softterm(req, 0);
1879 /* XXX the other two queues as well */
1882 for (i = 0; i < 30; i++) {
1884 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1885 if (nmp == req->r_nmp)
1891 tsleep(&lbolt, 0, "nfscancel", 0);
1897 * Soft-terminate a request, effectively marking it as failed.
1899 * Must be called from within a critical section.
1902 nfs_softterm(struct nfsreq *rep, int islocked)
1904 rep->r_flags |= R_SOFTTERM;
1905 nfs_hardterm(rep, islocked);
1909 * Hard-terminate a request, typically after getting a response.
1911 * The state machine can still decide to re-issue it later if necessary.
1913 * Must be called from within a critical section.
1916 nfs_hardterm(struct nfsreq *rep, int islocked)
1918 struct nfsmount *nmp = rep->r_nmp;
1921 * The nm_send count is decremented now to avoid deadlocks
1922 * when the process in soreceive() hasn't yet managed to send
1925 if (rep->r_flags & R_SENT) {
1926 rep->r_flags &= ~R_SENT;
1930 * If we locked the request or nobody else has locked the request,
1931 * and the request is async, we can move it to the reader thread's
1932 * queue now and fix up the state.
1934 * If we locked the request or nobody else has locked the request,
1935 * we can wake up anyone blocked waiting for a response on the
1938 if (islocked || (rep->r_flags & R_LOCKED) == 0) {
1939 if ((rep->r_flags & (R_ONREQQ | R_ASYNC)) ==
1940 (R_ONREQQ | R_ASYNC)) {
1941 rep->r_flags &= ~R_ONREQQ;
1942 TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1944 TAILQ_INSERT_TAIL(&nmp->nm_reqrxq, rep, r_chain);
1945 KKASSERT(rep->r_info->state == NFSM_STATE_TRY ||
1946 rep->r_info->state == NFSM_STATE_WAITREPLY);
1947 rep->r_info->state = NFSM_STATE_PROCESSREPLY;
1948 nfssvc_iod_reader_wakeup(nmp);
1949 if (TAILQ_FIRST(&nmp->nm_bioq) &&
1950 nmp->nm_reqqlen == NFS_MAXASYNCBIO * 2 / 3) {
1951 nfssvc_iod_writer_wakeup(nmp);
1954 mtx_abort_ex_link(&nmp->nm_rxlock, &rep->r_link);
1959 * Test for a termination condition pending on the process.
1960 * This is used for NFSMNT_INT mounts.
1963 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
1969 if (rep && (rep->r_flags & R_SOFTTERM))
1971 /* Terminate all requests while attempting a forced unmount. */
1972 if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
1974 if (!(nmp->nm_flag & NFSMNT_INT))
1976 /* td might be NULL YYY */
1977 if (td == NULL || (p = td->td_proc) == NULL)
1981 tmpset = lwp_sigpend(lp);
1982 SIGSETNAND(tmpset, lp->lwp_sigmask);
1983 SIGSETNAND(tmpset, p->p_sigignore);
1984 if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
1991 * Lock a socket against others.
1992 * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
1993 * and also to avoid race conditions between the processes with nfs requests
1994 * in progress when a reconnect is necessary.
1997 nfs_sndlock(struct nfsmount *nmp, struct nfsreq *rep)
1999 mtx_t mtx = &nmp->nm_txlock;
2007 td = rep ? rep->r_td : NULL;
2008 if (nmp->nm_flag & NFSMNT_INT)
2011 while ((error = mtx_lock_ex_try(mtx)) != 0) {
2012 if (nfs_sigintr(nmp, rep, td)) {
2016 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
2019 if (slpflag == PCATCH) {
2024 /* Always fail if our request has been cancelled. */
2025 if (rep && (rep->r_flags & R_SOFTTERM)) {
2034 * Unlock the stream socket for others.
2037 nfs_sndunlock(struct nfsmount *nmp)
2039 mtx_unlock(&nmp->nm_txlock);
2043 * Lock the receiver side of the socket.
2048 nfs_rcvlock(struct nfsmount *nmp, struct nfsreq *rep)
2050 mtx_t mtx = &nmp->nm_rxlock;
2056 * Unconditionally check for completion in case another nfsiod
2057 * get the packet while the caller was blocked, before the caller
2058 * called us. Packet reception is handled by mainline code which
2059 * is protected by the BGL at the moment.
2061 * We do not strictly need the second check just before the
2062 * tsleep(), but it's good defensive programming.
2064 if (rep && rep->r_mrep != NULL)
2067 if (nmp->nm_flag & NFSMNT_INT)
2073 while ((error = mtx_lock_ex_try(mtx)) != 0) {
2074 if (nfs_sigintr(nmp, rep, (rep ? rep->r_td : NULL))) {
2078 if (rep && rep->r_mrep != NULL) {
2084 * NOTE: can return ENOLCK, but in that case rep->r_mrep
2085 * will already be set.
2088 error = mtx_lock_ex_link(mtx, &rep->r_link,
2092 error = mtx_lock_ex(mtx, "nfsrcvlk", slpflag, slptimeo);
2098 * If our reply was recieved while we were sleeping,
2099 * then just return without taking the lock to avoid a
2100 * situation where a single iod could 'capture' the
2103 if (rep && rep->r_mrep != NULL) {
2107 if (slpflag == PCATCH) {
2113 if (rep && rep->r_mrep != NULL) {
2122 * Unlock the stream socket for others.
2125 nfs_rcvunlock(struct nfsmount *nmp)
2127 mtx_unlock(&nmp->nm_rxlock);
2133 * Check for badly aligned mbuf data and realign by copying the unaligned
2134 * portion of the data into a new mbuf chain and freeing the portions
2135 * of the old chain that were replaced.
2137 * We cannot simply realign the data within the existing mbuf chain
2138 * because the underlying buffers may contain other rpc commands and
2139 * we cannot afford to overwrite them.
2141 * We would prefer to avoid this situation entirely. The situation does
2142 * not occur with NFS/UDP and is supposed to only occassionally occur
2143 * with TCP. Use vfs.nfs.realign_count and realign_test to check this.
2146 nfs_realign(struct mbuf **pm, int hsiz)
2149 struct mbuf *n = NULL;
2154 while ((m = *pm) != NULL) {
2155 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3)) {
2156 n = m_getl(m->m_len, MB_WAIT, MT_DATA, 0, NULL);
2164 * If n is non-NULL, loop on m copying data, then replace the
2165 * portion of the chain that had to be realigned.
2168 ++nfs_realign_count;
2170 m_copyback(n, off, m->m_len, mtod(m, caddr_t));
2179 #ifndef NFS_NOSERVER
2182 * Parse an RPC request
2184 * - fill in the cred struct.
2187 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
2194 u_int32_t nfsvers, auth_type;
2196 int error = 0, ticklen;
2197 struct nfsuid *nuidp;
2198 struct timeval tvin, tvout;
2199 struct nfsm_info info;
2200 #if 0 /* until encrypted keys are implemented */
2201 NFSKERBKEYSCHED_T keys; /* stores key schedule */
2204 info.mrep = nd->nd_mrep;
2205 info.md = nd->nd_md;
2206 info.dpos = nd->nd_dpos;
2209 NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
2210 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
2211 if (*tl++ != rpc_call) {
2216 NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
2220 if (*tl++ != rpc_vers) {
2221 nd->nd_repstat = ERPCMISMATCH;
2222 nd->nd_procnum = NFSPROC_NOOP;
2225 if (*tl != nfs_prog) {
2226 nd->nd_repstat = EPROGUNAVAIL;
2227 nd->nd_procnum = NFSPROC_NOOP;
2231 nfsvers = fxdr_unsigned(u_int32_t, *tl++);
2232 if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
2233 nd->nd_repstat = EPROGMISMATCH;
2234 nd->nd_procnum = NFSPROC_NOOP;
2237 if (nfsvers == NFS_VER3)
2238 nd->nd_flag = ND_NFSV3;
2239 nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
2240 if (nd->nd_procnum == NFSPROC_NULL)
2242 if (nd->nd_procnum >= NFS_NPROCS ||
2243 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
2244 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
2245 nd->nd_repstat = EPROCUNAVAIL;
2246 nd->nd_procnum = NFSPROC_NOOP;
2249 if ((nd->nd_flag & ND_NFSV3) == 0)
2250 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
2252 len = fxdr_unsigned(int, *tl++);
2253 if (len < 0 || len > RPCAUTH_MAXSIZ) {
2258 nd->nd_flag &= ~ND_KERBAUTH;
2260 * Handle auth_unix or auth_kerb.
2262 if (auth_type == rpc_auth_unix) {
2263 len = fxdr_unsigned(int, *++tl);
2264 if (len < 0 || len > NFS_MAXNAMLEN) {
2268 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2269 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2270 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
2271 nd->nd_cr.cr_ref = 1;
2272 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
2273 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
2274 len = fxdr_unsigned(int, *tl);
2275 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
2279 NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
2280 for (i = 1; i <= len; i++)
2282 nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
2285 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
2286 if (nd->nd_cr.cr_ngroups > 1)
2287 nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
2288 len = fxdr_unsigned(int, *++tl);
2289 if (len < 0 || len > RPCAUTH_MAXSIZ) {
2294 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2296 } else if (auth_type == rpc_auth_kerb) {
2297 switch (fxdr_unsigned(int, *tl++)) {
2298 case RPCAKN_FULLNAME:
2299 ticklen = fxdr_unsigned(int, *tl);
2300 *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2301 uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2302 nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2303 if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2310 uio.uio_segflg = UIO_SYSSPACE;
2311 iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2312 iov.iov_len = RPCAUTH_MAXSIZ - 4;
2313 ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2314 NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2315 if (*tl++ != rpc_auth_kerb ||
2316 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2317 kprintf("Bad kerb verifier\n");
2318 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2319 nd->nd_procnum = NFSPROC_NOOP;
2322 NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2323 tl = (u_int32_t *)cp;
2324 if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2325 kprintf("Not fullname kerb verifier\n");
2326 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2327 nd->nd_procnum = NFSPROC_NOOP;
2330 cp += NFSX_UNSIGNED;
2331 bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2332 nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2333 nd->nd_flag |= ND_KERBFULL;
2334 nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2336 case RPCAKN_NICKNAME:
2337 if (len != 2 * NFSX_UNSIGNED) {
2338 kprintf("Kerb nickname short\n");
2339 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2340 nd->nd_procnum = NFSPROC_NOOP;
2343 nickuid = fxdr_unsigned(uid_t, *tl);
2344 NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2345 if (*tl++ != rpc_auth_kerb ||
2346 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2347 kprintf("Kerb nick verifier bad\n");
2348 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2349 nd->nd_procnum = NFSPROC_NOOP;
2352 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2353 tvin.tv_sec = *tl++;
2356 for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2357 nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2358 if (nuidp->nu_cr.cr_uid == nickuid &&
2360 netaddr_match(NU_NETFAM(nuidp),
2361 &nuidp->nu_haddr, nd->nd_nam2)))
2366 (NFSERR_AUTHERR|AUTH_REJECTCRED);
2367 nd->nd_procnum = NFSPROC_NOOP;
2372 * Now, decrypt the timestamp using the session key
2379 tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2380 tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2381 if (nuidp->nu_expire < time_second ||
2382 nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2383 (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2384 nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2385 nuidp->nu_expire = 0;
2387 (NFSERR_AUTHERR|AUTH_REJECTVERF);
2388 nd->nd_procnum = NFSPROC_NOOP;
2391 nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2392 nd->nd_flag |= ND_KERBNICK;
2395 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2396 nd->nd_procnum = NFSPROC_NOOP;
2400 nd->nd_md = info.md;
2401 nd->nd_dpos = info.dpos;
2410 * Send a message to the originating process's terminal. The thread and/or
2411 * process may be NULL. YYY the thread should not be NULL but there may
2412 * still be some uio_td's that are still being passed as NULL through to
2416 nfs_msg(struct thread *td, char *server, char *msg)
2420 if (td && td->td_proc)
2421 tpr = tprintf_open(td->td_proc);
2424 tprintf(tpr, "nfs server %s: %s\n", server, msg);
2429 #ifndef NFS_NOSERVER
2431 * Socket upcall routine for the nfsd sockets.
2432 * The caddr_t arg is a pointer to the "struct nfssvc_sock".
2433 * Essentially do as much as possible non-blocking, else punt and it will
2434 * be called with MB_WAIT from an nfsd.
2437 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2439 struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2441 struct sockaddr *nam;
2444 int nparallel_wakeup = 0;
2446 if ((slp->ns_flag & SLP_VALID) == 0)
2450 * Do not allow an infinite number of completed RPC records to build
2451 * up before we stop reading data from the socket. Otherwise we could
2452 * end up holding onto an unreasonable number of mbufs for requests
2453 * waiting for service.
2455 * This should give pretty good feedback to the TCP
2456 * layer and prevents a memory crunch for other protocols.
2458 * Note that the same service socket can be dispatched to several
2459 * nfs servers simultaniously.
2461 * the tcp protocol callback calls us with MB_DONTWAIT.
2462 * nfsd calls us with MB_WAIT (typically).
2464 if (waitflag == MB_DONTWAIT && slp->ns_numrec >= nfsd_waiting / 2 + 1) {
2465 slp->ns_flag |= SLP_NEEDQ;
2470 * Handle protocol specifics to parse an RPC request. We always
2471 * pull from the socket using non-blocking I/O.
2473 if (so->so_type == SOCK_STREAM) {
2475 * The data has to be read in an orderly fashion from a TCP
2476 * stream, unlike a UDP socket. It is possible for soreceive
2477 * and/or nfsrv_getstream() to block, so make sure only one
2478 * entity is messing around with the TCP stream at any given
2479 * moment. The receive sockbuf's lock in soreceive is not
2482 * Note that this procedure can be called from any number of
2483 * NFS severs *OR* can be upcalled directly from a TCP
2486 if (slp->ns_flag & SLP_GETSTREAM) {
2487 slp->ns_flag |= SLP_NEEDQ;
2490 slp->ns_flag |= SLP_GETSTREAM;
2493 * Do soreceive(). Pull out as much data as possible without
2496 sbinit(&sio, 1000000000);
2497 flags = MSG_DONTWAIT;
2498 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2499 if (error || sio.sb_mb == NULL) {
2500 if (error == EWOULDBLOCK)
2501 slp->ns_flag |= SLP_NEEDQ;
2503 slp->ns_flag |= SLP_DISCONN;
2504 slp->ns_flag &= ~SLP_GETSTREAM;
2508 if (slp->ns_rawend) {
2509 slp->ns_rawend->m_next = m;
2510 slp->ns_cc += sio.sb_cc;
2513 slp->ns_cc = sio.sb_cc;
2520 * Now try and parse as many record(s) as we can out of the
2523 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2526 slp->ns_flag |= SLP_DISCONN;
2528 slp->ns_flag |= SLP_NEEDQ;
2530 slp->ns_flag &= ~SLP_GETSTREAM;
2533 * For UDP soreceive typically pulls just one packet, loop
2534 * to get the whole batch.
2537 sbinit(&sio, 1000000000);
2538 flags = MSG_DONTWAIT;
2539 error = so_pru_soreceive(so, &nam, NULL, &sio,
2542 struct nfsrv_rec *rec;
2543 int mf = (waitflag & MB_DONTWAIT) ?
2544 M_NOWAIT : M_WAITOK;
2545 rec = kmalloc(sizeof(struct nfsrv_rec),
2549 FREE(nam, M_SONAME);
2553 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2554 rec->nr_address = nam;
2555 rec->nr_packet = sio.sb_mb;
2556 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2561 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2562 && error != EWOULDBLOCK) {
2563 slp->ns_flag |= SLP_DISCONN;
2567 } while (sio.sb_mb);
2571 * If we were upcalled from the tcp protocol layer and we have
2572 * fully parsed records ready to go, or there is new data pending,
2573 * or something went wrong, try to wake up an nfsd thread to deal
2577 if (waitflag == MB_DONTWAIT && (slp->ns_numrec > 0
2578 || (slp->ns_flag & (SLP_NEEDQ | SLP_DISCONN)))) {
2579 nfsrv_wakenfsd(slp, nparallel_wakeup);
2584 * Try and extract an RPC request from the mbuf data list received on a
2585 * stream socket. The "waitflag" argument indicates whether or not it
2589 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2591 struct mbuf *m, **mpp;
2594 struct mbuf *om, *m2, *recm;
2598 if (slp->ns_reclen == 0) {
2599 if (slp->ns_cc < NFSX_UNSIGNED)
2602 if (m->m_len >= NFSX_UNSIGNED) {
2603 bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2604 m->m_data += NFSX_UNSIGNED;
2605 m->m_len -= NFSX_UNSIGNED;
2607 cp1 = (caddr_t)&recmark;
2608 cp2 = mtod(m, caddr_t);
2609 while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2610 while (m->m_len == 0) {
2612 cp2 = mtod(m, caddr_t);
2619 slp->ns_cc -= NFSX_UNSIGNED;
2620 recmark = ntohl(recmark);
2621 slp->ns_reclen = recmark & ~0x80000000;
2622 if (recmark & 0x80000000)
2623 slp->ns_flag |= SLP_LASTFRAG;
2625 slp->ns_flag &= ~SLP_LASTFRAG;
2626 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2627 log(LOG_ERR, "%s (%d) from nfs client\n",
2628 "impossible packet length",
2635 * Now get the record part.
2637 * Note that slp->ns_reclen may be 0. Linux sometimes
2638 * generates 0-length RPCs
2641 if (slp->ns_cc == slp->ns_reclen) {
2643 slp->ns_raw = slp->ns_rawend = NULL;
2644 slp->ns_cc = slp->ns_reclen = 0;
2645 } else if (slp->ns_cc > slp->ns_reclen) {
2650 while (len < slp->ns_reclen) {
2651 if ((len + m->m_len) > slp->ns_reclen) {
2652 m2 = m_copym(m, 0, slp->ns_reclen - len,
2660 m->m_data += slp->ns_reclen - len;
2661 m->m_len -= slp->ns_reclen - len;
2662 len = slp->ns_reclen;
2664 return (EWOULDBLOCK);
2666 } else if ((len + m->m_len) == slp->ns_reclen) {
2686 * Accumulate the fragments into a record.
2688 mpp = &slp->ns_frag;
2690 mpp = &((*mpp)->m_next);
2692 if (slp->ns_flag & SLP_LASTFRAG) {
2693 struct nfsrv_rec *rec;
2694 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2695 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2697 m_freem(slp->ns_frag);
2699 nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2700 rec->nr_address = NULL;
2701 rec->nr_packet = slp->ns_frag;
2702 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2706 slp->ns_frag = NULL;
2712 * Parse an RPC header.
2715 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2716 struct nfsrv_descript **ndp)
2718 struct nfsrv_rec *rec;
2720 struct sockaddr *nam;
2721 struct nfsrv_descript *nd;
2725 if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2727 rec = STAILQ_FIRST(&slp->ns_rec);
2728 STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2729 KKASSERT(slp->ns_numrec > 0);
2731 nam = rec->nr_address;
2733 kfree(rec, M_NFSRVDESC);
2734 MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2735 M_NFSRVDESC, M_WAITOK);
2736 nd->nd_md = nd->nd_mrep = m;
2738 nd->nd_dpos = mtod(m, caddr_t);
2739 error = nfs_getreq(nd, nfsd, TRUE);
2742 FREE(nam, M_SONAME);
2744 kfree((caddr_t)nd, M_NFSRVDESC);
2753 * Try to assign service sockets to nfsd threads based on the number
2754 * of new rpc requests that have been queued on the service socket.
2756 * If no nfsd's are available or additonal requests are pending, set the
2757 * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2758 * the work in the nfssvc_sock list when it is finished processing its
2759 * current work. This flag is only cleared when an nfsd can not find
2760 * any new work to perform.
2763 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2767 if ((slp->ns_flag & SLP_VALID) == 0)
2771 TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2772 if (nd->nfsd_flag & NFSD_WAITING) {
2773 nd->nfsd_flag &= ~NFSD_WAITING;
2775 panic("nfsd wakeup");
2778 wakeup((caddr_t)nd);
2779 if (--nparallel == 0)
2784 slp->ns_flag |= SLP_DOREC;
2785 nfsd_head_flag |= NFSD_CHECKSLP;
2788 #endif /* NFS_NOSERVER */