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 *);
170 static void nfs_timer_req(struct nfsreq *req);
172 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
173 struct nfssvc_sock *slp,
175 struct mbuf **mreqp) = {
203 #endif /* NFS_NOSERVER */
206 * Initialize sockets and congestion for a new NFS connection.
207 * We do not free the sockaddr if error.
210 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
213 int error, rcvreserve, sndreserve;
215 struct sockaddr *saddr;
216 struct sockaddr_in *sin;
217 struct thread *td = &thread0; /* only used for socreate and sobind */
221 error = socreate(saddr->sa_family, &nmp->nm_so, nmp->nm_sotype,
222 nmp->nm_soproto, td);
226 nmp->nm_soflags = so->so_proto->pr_flags;
229 * Some servers require that the client port be a reserved port number.
231 if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
234 struct sockaddr_in ssin;
236 bzero(&sopt, sizeof sopt);
237 ip = IP_PORTRANGE_LOW;
238 sopt.sopt_level = IPPROTO_IP;
239 sopt.sopt_name = IP_PORTRANGE;
240 sopt.sopt_val = (void *)&ip;
241 sopt.sopt_valsize = sizeof(ip);
243 error = sosetopt(so, &sopt);
246 bzero(&ssin, sizeof ssin);
248 sin->sin_len = sizeof (struct sockaddr_in);
249 sin->sin_family = AF_INET;
250 sin->sin_addr.s_addr = INADDR_ANY;
251 sin->sin_port = htons(0);
252 error = sobind(so, (struct sockaddr *)sin, td);
255 bzero(&sopt, sizeof sopt);
256 ip = IP_PORTRANGE_DEFAULT;
257 sopt.sopt_level = IPPROTO_IP;
258 sopt.sopt_name = IP_PORTRANGE;
259 sopt.sopt_val = (void *)&ip;
260 sopt.sopt_valsize = sizeof(ip);
262 error = sosetopt(so, &sopt);
268 * Protocols that do not require connections may be optionally left
269 * unconnected for servers that reply from a port other than NFS_PORT.
271 if (nmp->nm_flag & NFSMNT_NOCONN) {
272 if (nmp->nm_soflags & PR_CONNREQUIRED) {
277 error = soconnect(so, nmp->nm_nam, td);
282 * Wait for the connection to complete. Cribbed from the
283 * connect system call but with the wait timing out so
284 * that interruptible mounts don't hang here for a long time.
287 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
288 (void) tsleep((caddr_t)&so->so_timeo, 0,
290 if ((so->so_state & SS_ISCONNECTING) &&
291 so->so_error == 0 && rep &&
292 (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
293 so->so_state &= ~SS_ISCONNECTING;
299 error = so->so_error;
306 so->so_rcv.ssb_timeo = (5 * hz);
307 so->so_snd.ssb_timeo = (5 * hz);
310 * Get buffer reservation size from sysctl, but impose reasonable
313 pktscale = nfs_bufpackets;
319 if (nmp->nm_sotype == SOCK_DGRAM) {
320 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
321 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
322 NFS_MAXPKTHDR) * pktscale;
323 } else if (nmp->nm_sotype == SOCK_SEQPACKET) {
324 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
325 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
326 NFS_MAXPKTHDR) * pktscale;
328 if (nmp->nm_sotype != SOCK_STREAM)
329 panic("nfscon sotype");
330 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
334 bzero(&sopt, sizeof sopt);
335 sopt.sopt_level = SOL_SOCKET;
336 sopt.sopt_name = SO_KEEPALIVE;
337 sopt.sopt_val = &val;
338 sopt.sopt_valsize = sizeof val;
342 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
346 bzero(&sopt, sizeof sopt);
347 sopt.sopt_level = IPPROTO_TCP;
348 sopt.sopt_name = TCP_NODELAY;
349 sopt.sopt_val = &val;
350 sopt.sopt_valsize = sizeof val;
354 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR +
355 sizeof (u_int32_t)) * pktscale;
356 rcvreserve = (nmp->nm_rsize + NFS_MAXPKTHDR +
357 sizeof (u_int32_t)) * pktscale;
359 error = soreserve(so, sndreserve, rcvreserve,
360 &td->td_proc->p_rlimit[RLIMIT_SBSIZE]);
363 so->so_rcv.ssb_flags |= SSB_NOINTR;
364 so->so_snd.ssb_flags |= SSB_NOINTR;
366 /* Initialize other non-zero congestion variables */
367 nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] =
368 nmp->nm_srtt[3] = (NFS_TIMEO << 3);
369 nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
370 nmp->nm_sdrtt[3] = 0;
371 nmp->nm_cwnd = NFS_MAXCWND / 2; /* Initial send window */
373 nmp->nm_timeouts = 0;
383 * Called when a connection is broken on a reliable protocol.
384 * - clean up the old socket
385 * - nfs_connect() again
386 * - set R_MUSTRESEND for all outstanding requests on mount point
387 * If this fails the mount point is DEAD!
388 * nb: Must be called with the nfs_sndlock() set on the mount point.
391 nfs_reconnect(struct nfsreq *rep)
394 struct nfsmount *nmp = rep->r_nmp;
398 while ((error = nfs_connect(nmp, rep)) != 0) {
399 if (error == EINTR || error == ERESTART)
401 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
405 * Loop through outstanding request list and fix up all requests
409 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
410 KKASSERT(req->r_nmp == nmp);
411 req->r_flags |= R_MUSTRESEND;
418 * NFS disconnect. Clean up and unlink.
421 nfs_disconnect(struct nfsmount *nmp)
428 soshutdown(so, SHUT_RDWR);
429 soclose(so, FNONBLOCK);
434 nfs_safedisconnect(struct nfsmount *nmp)
436 struct nfsreq dummyreq;
438 bzero(&dummyreq, sizeof(dummyreq));
439 dummyreq.r_nmp = nmp;
440 dummyreq.r_td = NULL;
441 mtx_link_init(&dummyreq.r_link);
442 nfs_rcvlock(&dummyreq);
444 nfs_rcvunlock(&dummyreq);
448 * This is the nfs send routine. For connection based socket types, it
449 * must be called with an nfs_sndlock() on the socket.
450 * "rep == NULL" indicates that it has been called from a server.
451 * For the client side:
452 * - return EINTR if the RPC is terminated, 0 otherwise
453 * - set R_MUSTRESEND if the send fails for any reason
454 * - do any cleanup required by recoverable socket errors (?)
455 * For the server side:
456 * - return EINTR or ERESTART if interrupted by a signal
457 * - return EPIPE if a connection is lost for connection based sockets (TCP...)
458 * - do any cleanup required by recoverable socket errors (?)
461 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
464 struct sockaddr *sendnam;
465 int error, soflags, flags;
468 if (rep->r_flags & R_SOFTTERM) {
472 if ((so = rep->r_nmp->nm_so) == NULL) {
473 rep->r_flags |= R_MUSTRESEND;
477 rep->r_flags &= ~R_MUSTRESEND;
478 soflags = rep->r_nmp->nm_soflags;
480 soflags = so->so_proto->pr_flags;
481 if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
485 if (so->so_type == SOCK_SEQPACKET)
490 error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
493 * ENOBUFS for dgram sockets is transient and non fatal.
494 * No need to log, and no need to break a soft mount.
496 if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
498 if (rep) /* do backoff retransmit on client */
499 rep->r_flags |= R_MUSTRESEND;
504 log(LOG_INFO, "nfs send error %d for server %s\n",error,
505 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
507 * Deal with errors for the client side.
509 if (rep->r_flags & R_SOFTTERM)
512 rep->r_flags |= R_MUSTRESEND;
514 log(LOG_INFO, "nfsd send error %d\n", error);
517 * Handle any recoverable (soft) socket errors here. (?)
519 if (error != EINTR && error != ERESTART &&
520 error != EWOULDBLOCK && error != EPIPE)
527 * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
528 * done by soreceive(), but for SOCK_STREAM we must deal with the Record
529 * Mark and consolidate the data into a new mbuf list.
530 * nb: Sometimes TCP passes the data up to soreceive() in long lists of
532 * For SOCK_STREAM we must be very careful to read an entire record once
533 * we have read any of it, even if the system call has been interrupted.
536 nfs_receive(struct nfsreq *rep, struct sockaddr **aname, struct mbuf **mp)
543 struct mbuf *control;
545 struct sockaddr **getnam;
546 int error, sotype, rcvflg;
547 struct thread *td = curthread; /* XXX */
550 * Set up arguments for soreceive()
554 sotype = rep->r_nmp->nm_sotype;
557 * For reliable protocols, lock against other senders/receivers
558 * in case a reconnect is necessary.
559 * For SOCK_STREAM, first get the Record Mark to find out how much
560 * more there is to get.
561 * We must lock the socket against other receivers
562 * until we have an entire rpc request/reply.
564 if (sotype != SOCK_DGRAM) {
565 error = nfs_sndlock(rep);
570 * Check for fatal errors and resending request.
573 * Ugh: If a reconnect attempt just happened, nm_so
574 * would have changed. NULL indicates a failed
575 * attempt that has essentially shut down this
578 if (rep->r_mrep || (rep->r_flags & R_SOFTTERM)) {
582 so = rep->r_nmp->nm_so;
584 error = nfs_reconnect(rep);
591 while (rep->r_flags & R_MUSTRESEND) {
592 m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
593 nfsstats.rpcretries++;
594 error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
596 if (error == EINTR || error == ERESTART ||
597 (error = nfs_reconnect(rep)) != 0) {
605 if (sotype == SOCK_STREAM) {
607 * Get the length marker from the stream
609 aio.iov_base = (caddr_t)&len;
610 aio.iov_len = sizeof(u_int32_t);
613 auio.uio_segflg = UIO_SYSSPACE;
614 auio.uio_rw = UIO_READ;
616 auio.uio_resid = sizeof(u_int32_t);
619 rcvflg = MSG_WAITALL;
620 error = so_pru_soreceive(so, NULL, &auio, NULL,
622 if (error == EWOULDBLOCK && rep) {
623 if (rep->r_flags & R_SOFTTERM)
626 } while (error == EWOULDBLOCK);
628 if (error == 0 && auio.uio_resid > 0) {
630 * Only log short packets if not EOF
632 if (auio.uio_resid != sizeof(u_int32_t))
634 "short receive (%d/%d) from nfs server %s\n",
635 (int)(sizeof(u_int32_t) - auio.uio_resid),
636 (int)sizeof(u_int32_t),
637 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
642 len = ntohl(len) & ~0x80000000;
644 * This is SERIOUS! We are out of sync with the sender
645 * and forcing a disconnect/reconnect is all I can do.
647 if (len > NFS_MAXPACKET) {
648 log(LOG_ERR, "%s (%d) from nfs server %s\n",
649 "impossible packet length",
651 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
657 * Get the rest of the packet as an mbuf chain
661 rcvflg = MSG_WAITALL;
662 error = so_pru_soreceive(so, NULL, NULL, &sio,
664 } while (error == EWOULDBLOCK || error == EINTR ||
666 if (error == 0 && sio.sb_cc != len) {
669 "short receive (%d/%d) from nfs server %s\n",
670 len - auio.uio_resid, len,
671 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
677 * Non-stream, so get the whole packet by not
678 * specifying MSG_WAITALL and by specifying a large
681 * We have no use for control msg., but must grab them
682 * and then throw them away so we know what is going
685 sbinit(&sio, 100000000);
688 error = so_pru_soreceive(so, NULL, NULL, &sio,
692 if (error == EWOULDBLOCK && rep) {
693 if (rep->r_flags & R_SOFTTERM) {
698 } while (error == EWOULDBLOCK ||
699 (error == 0 && sio.sb_mb == NULL && control));
700 if ((rcvflg & MSG_EOR) == 0)
702 if (error == 0 && sio.sb_mb == NULL)
708 if (error && error != EINTR && error != ERESTART) {
711 if (error != EPIPE) {
713 "receive error %d from nfs server %s\n",
715 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
717 error = nfs_sndlock(rep);
719 error = nfs_reconnect(rep);
727 if ((so = rep->r_nmp->nm_so) == NULL)
729 if (so->so_state & SS_ISCONNECTED)
733 sbinit(&sio, 100000000);
736 error = so_pru_soreceive(so, getnam, NULL, &sio,
738 if (error == EWOULDBLOCK &&
739 (rep->r_flags & R_SOFTTERM)) {
743 } while (error == EWOULDBLOCK);
752 * Search for any mbufs that are not a multiple of 4 bytes long
753 * or with m_data not longword aligned.
754 * These could cause pointer alignment problems, so copy them to
755 * well aligned mbufs.
757 nfs_realign(mp, 5 * NFSX_UNSIGNED);
762 * Implement receipt of reply on a socket.
763 * We must search through the list of received datagrams matching them
764 * with outstanding requests using the xid, until ours is found.
768 nfs_reply(struct nfsreq *myrep)
771 struct nfsmount *nmp = myrep->r_nmp;
773 struct mbuf *mrep, *md;
774 struct sockaddr *nam;
780 * Loop around until we get our own reply
784 * Lock against other receivers so that I don't get stuck in
785 * sbwait() after someone else has received my reply for me.
786 * Also necessary for connection based protocols to avoid
787 * race conditions during a reconnect.
789 * If nfs_rcvlock() returns EALREADY, that means that
790 * the reply has already been recieved by another
791 * process and we can return immediately. In this
792 * case, the lock is not taken to avoid races with
795 error = nfs_rcvlock(myrep);
796 if (error == EALREADY)
801 * Get the next Rpc reply off the socket
803 error = nfs_receive(myrep, &nam, &mrep);
804 nfs_rcvunlock(myrep);
807 * Ignore routing errors on connectionless protocols??
809 if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
810 nmp->nm_so->so_error = 0;
811 if (myrep->r_flags & R_GETONEREP)
821 * Get the xid and check that it is an rpc reply
824 dpos = mtod(md, caddr_t);
825 nfsm_dissect(tl, u_int32_t *, 2*NFSX_UNSIGNED);
827 if (*tl != rpc_reply) {
828 nfsstats.rpcinvalid++;
831 if (myrep->r_flags & R_GETONEREP)
837 * Loop through the request list to match up the reply
838 * Iff no match, just drop the datagram. On match, set
839 * r_mrep atomically to prevent the timer from messing
840 * around with the request after we have exited the critical
844 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
845 if (rep->r_mrep == NULL && rxid == rep->r_xid)
851 * Fill in the rest of the reply if we found a match.
859 rt = &nfsrtt.rttl[nfsrtt.pos];
860 rt->proc = rep->r_procnum;
861 rt->rto = NFS_RTO(nmp, proct[rep->r_procnum]);
862 rt->sent = nmp->nm_sent;
863 rt->cwnd = nmp->nm_cwnd;
864 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
865 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
866 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
867 getmicrotime(&rt->tstamp);
868 if (rep->r_flags & R_TIMING)
869 rt->rtt = rep->r_rtt;
872 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
875 * Update congestion window.
876 * Do the additive increase of
879 if (nmp->nm_cwnd <= nmp->nm_sent) {
881 (NFS_CWNDSCALE * NFS_CWNDSCALE +
882 (nmp->nm_cwnd >> 1)) / nmp->nm_cwnd;
883 if (nmp->nm_cwnd > NFS_MAXCWND)
884 nmp->nm_cwnd = NFS_MAXCWND;
886 crit_enter(); /* nfs_timer interlock for nm_sent */
887 if (rep->r_flags & R_SENT) {
888 rep->r_flags &= ~R_SENT;
889 nmp->nm_sent -= NFS_CWNDSCALE;
893 * Update rtt using a gain of 0.125 on the mean
894 * and a gain of 0.25 on the deviation.
896 if (rep->r_flags & R_TIMING) {
898 * Since the timer resolution of
899 * NFS_HZ is so course, it can often
900 * result in r_rtt == 0. Since
901 * r_rtt == N means that the actual
902 * rtt is between N+dt and N+2-dt ticks,
906 t1 -= (NFS_SRTT(rep) >> 3);
910 t1 -= (NFS_SDRTT(rep) >> 2);
911 NFS_SDRTT(rep) += t1;
913 nmp->nm_timeouts = 0;
915 mtx_abort_ex_link(&rep->r_nmp->nm_rxlock, &rep->r_link);
918 * If not matched to a request, drop it.
919 * If it's mine, get out.
922 nfsstats.rpcunexpected++;
924 } else if (rep == myrep) {
925 if (rep->r_mrep == NULL)
926 panic("nfsreply nil");
929 if (myrep->r_flags & R_GETONEREP)
935 nfs_request(struct vnode *vp, struct mbuf *mrest, int procnum,
936 struct thread *td, struct ucred *cred, struct mbuf **mrp,
937 struct mbuf **mdp, caddr_t *dposp)
939 struct nfsreq *rep = NULL;
942 error = nfs_request_setup(vp, mrest, procnum, td, cred, &rep);
947 rep->r_dposp = dposp;
949 error = nfs_request_auth(rep);
953 error = nfs_request_try(rep); /* error ignored */
954 error = nfs_request_waitreply(rep); /* pass to process */
955 error = nfs_request_processreply(rep, error);
956 if (error == ENEEDAUTH)
964 * nfs_request - goes something like this
965 * - fill in request struct
966 * - links it into list
967 * - calls nfs_send() for first transmit
968 * - calls nfs_receive() to get reply
969 * - break down rpc header and return with nfs reply pointed to
971 * nb: always frees up mreq mbuf list
974 nfs_request_setup(struct vnode *vp, struct mbuf *mrest, int procnum,
975 struct thread *td, struct ucred *cred,
976 struct nfsreq **repp)
979 struct nfsmount *nmp;
983 /* Reject requests while attempting a forced unmount. */
984 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
988 nmp = VFSTONFS(vp->v_mount);
989 MALLOC(rep, struct nfsreq *, sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
993 rep->r_procnum = procnum;
1001 rep->r_mrest = mrest;
1002 rep->r_mrest_len = i;
1009 nfs_request_auth(struct nfsreq *rep)
1011 struct nfsmount *nmp = rep->r_nmp;
1013 char nickv[RPCX_NICKVERF];
1014 int error = 0, auth_len, auth_type;
1017 char *auth_str, *verf_str;
1021 rep->r_failed_auth = 0;
1024 * Get the RPC header with authorization.
1026 verf_str = auth_str = NULL;
1027 if (nmp->nm_flag & NFSMNT_KERB) {
1029 verf_len = sizeof (nickv);
1030 auth_type = RPCAUTH_KERB4;
1031 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1032 if (rep->r_failed_auth ||
1033 nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1034 verf_str, verf_len)) {
1035 error = nfs_getauth(nmp, rep, cred, &auth_str,
1036 &auth_len, verf_str, &verf_len, rep->r_key);
1038 m_freem(rep->r_mrest);
1039 rep->r_mrest = NULL;
1040 kfree((caddr_t)rep, M_NFSREQ);
1045 auth_type = RPCAUTH_UNIX;
1046 if (cred->cr_ngroups < 1)
1047 panic("nfsreq nogrps");
1048 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1049 nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1052 m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1053 auth_len, auth_str, verf_len, verf_str,
1054 rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1055 rep->r_mrest = NULL;
1057 kfree(auth_str, M_TEMP);
1060 * For stream protocols, insert a Sun RPC Record Mark.
1062 if (nmp->nm_sotype == SOCK_STREAM) {
1063 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1065 kfree(rep, M_NFSREQ);
1068 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1069 (m->m_pkthdr.len - NFSX_UNSIGNED));
1077 nfs_request_try(struct nfsreq *rep)
1079 struct nfsmount *nmp = rep->r_nmp;
1083 if (nmp->nm_flag & NFSMNT_SOFT)
1084 rep->r_retry = nmp->nm_retry;
1086 rep->r_retry = NFS_MAXREXMIT + 1; /* past clip limit */
1087 rep->r_rtt = rep->r_rexmit = 0;
1088 if (proct[rep->r_procnum] > 0)
1089 rep->r_flags = R_TIMING | R_MASKTIMER;
1091 rep->r_flags = R_MASKTIMER;
1095 * Do the client side RPC.
1097 nfsstats.rpcrequests++;
1100 * Chain request into list of outstanding requests. Be sure
1101 * to put it LAST so timer finds oldest requests first. Note
1102 * that R_MASKTIMER is set at the moment to prevent any timer
1103 * action on this request while we are still doing processing on
1104 * it below. splsoftclock() primarily protects nm_sent. Note
1105 * that we may block in this code so there is no atomicy guarentee.
1108 TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1109 mtx_link_init(&rep->r_link);
1114 * If backing off another request or avoiding congestion, don't
1115 * send this one now but let timer do it. If not timing a request,
1118 * Even though the timer will not mess with our request there is
1119 * still the possibility that we will race a reply (which clears
1120 * R_SENT), especially on localhost connections, so be very careful
1121 * when setting R_SENT. We could set R_SENT prior to calling
1122 * nfs_send() but why bother if the response occurs that quickly?
1124 if (nmp->nm_so && (nmp->nm_sotype != SOCK_DGRAM ||
1125 (nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1126 nmp->nm_sent < nmp->nm_cwnd)) {
1127 if (nmp->nm_soflags & PR_CONNREQUIRED)
1128 error = nfs_sndlock(rep);
1130 m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1131 error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1132 if (nmp->nm_soflags & PR_CONNREQUIRED)
1135 if (!error && (rep->r_flags & R_MUSTRESEND) == 0 &&
1136 rep->r_mrep == NULL) {
1137 KASSERT((rep->r_flags & R_SENT) == 0,
1138 ("R_SENT ASSERT %p", rep));
1139 nmp->nm_sent += NFS_CWNDSCALE;
1140 rep->r_flags |= R_SENT;
1148 * Let the timer do what it will with the request, then
1149 * wait for the reply from our send or the timer's.
1152 rep->r_flags &= ~R_MASKTIMER;
1158 nfs_request_waitreply(struct nfsreq *rep)
1160 struct nfsmount *nmp = rep->r_nmp;
1164 error = nfs_reply(rep);
1168 * RPC done, unlink the request, but don't rip it out from under
1169 * the callout timer.
1171 while (rep->r_flags & R_LOCKED) {
1172 nfs_timer_raced = 1;
1173 tsleep(&nfs_timer_raced, 0, "nfstrac", 0);
1175 TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1178 * Decrement the outstanding request count.
1180 if (rep->r_flags & R_SENT) {
1181 rep->r_flags &= ~R_SENT;
1182 nmp->nm_sent -= NFS_CWNDSCALE;
1190 * Process reply with error returned from nfs_requet_waitreply().
1192 * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1193 * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1196 nfs_request_processreply(struct nfsreq *rep, int error)
1198 struct nfsmount *nmp = rep->r_nmp;
1204 int trylater_delay = 15, trylater_cnt = 0;
1210 * If there was a successful reply and a tprintf msg.
1211 * tprintf a response.
1213 if (!error && (rep->r_flags & R_TPRINTFMSG))
1214 nfs_msg(rep->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1220 m_freem(rep->r_mreq);
1221 kfree((caddr_t)rep, M_NFSREQ);
1226 * break down the rpc header and check if ok
1228 nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1229 if (*tl++ == rpc_msgdenied) {
1230 if (*tl == rpc_mismatch) {
1232 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1233 *tl++ == rpc_autherr) {
1234 if (!rep->r_failed_auth) {
1235 rep->r_failed_auth++;
1236 rep->r_mheadend->m_next = NULL;
1238 m_freem(rep->r_mreq);
1247 m_freem(rep->r_mreq);
1248 kfree((caddr_t)rep, M_NFSREQ);
1253 * Grab any Kerberos verifier, otherwise just throw it away.
1255 verf_type = fxdr_unsigned(int, *tl++);
1256 i = fxdr_unsigned(int32_t, *tl);
1257 if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1258 error = nfs_savenickauth(nmp, rep->r_cred, i,
1259 rep->r_key, &md, &dpos, mrep);
1263 nfsm_adv(nfsm_rndup(i));
1264 nfsm_dissect(tl, u_int32_t *, NFSX_UNSIGNED);
1267 nfsm_dissect(tl, u_int32_t *, NFSX_UNSIGNED);
1269 error = fxdr_unsigned(int, *tl);
1270 if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1271 error == NFSERR_TRYLATER) {
1274 waituntil = time_second + trylater_delay;
1275 while (time_second < waituntil)
1276 (void) tsleep((caddr_t)&lbolt,
1278 trylater_delay *= nfs_backoff[trylater_cnt];
1279 if (trylater_cnt < 7)
1281 rep->r_flags &= ~R_MASKTIMER;
1282 return (EAGAIN); /* goto tryagain */
1286 * If the File Handle was stale, invalidate the
1287 * lookup cache, just in case.
1289 * To avoid namecache<->vnode deadlocks we must
1290 * release the vnode lock if we hold it.
1292 if (error == ESTALE) {
1293 struct vnode *vp = rep->r_vp;
1296 ltype = lockstatus(&vp->v_lock, curthread);
1297 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1298 lockmgr(&vp->v_lock, LK_RELEASE);
1299 cache_inval_vp(vp, CINV_CHILDREN);
1300 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1301 lockmgr(&vp->v_lock, ltype);
1303 if (nmp->nm_flag & NFSMNT_NFSV3) {
1306 *rep->r_dposp = dpos;
1307 error |= NFSERR_RETERR;
1310 m_freem(rep->r_mreq);
1311 kfree((caddr_t)rep, M_NFSREQ);
1317 *rep->r_dposp = dpos;
1318 m_freem(rep->r_mreq);
1319 FREE((caddr_t)rep, M_NFSREQ);
1323 error = EPROTONOSUPPORT;
1325 m_freem(rep->r_mreq);
1326 kfree((caddr_t)rep, M_NFSREQ);
1330 #ifndef NFS_NOSERVER
1332 * Generate the rpc reply header
1333 * siz arg. is used to decide if adding a cluster is worthwhile
1336 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1337 int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1342 struct mbuf *mb, *mb2;
1344 siz += RPC_REPLYSIZ;
1345 mb = mreq = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1346 mreq->m_pkthdr.len = 0;
1348 * If this is not a cluster, try and leave leading space
1349 * for the lower level headers.
1351 if ((max_hdr + siz) < MINCLSIZE)
1352 mreq->m_data += max_hdr;
1353 tl = mtod(mreq, u_int32_t *);
1354 mreq->m_len = 6 * NFSX_UNSIGNED;
1355 bpos = ((caddr_t)tl) + mreq->m_len;
1356 *tl++ = txdr_unsigned(nd->nd_retxid);
1358 if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1359 *tl++ = rpc_msgdenied;
1360 if (err & NFSERR_AUTHERR) {
1361 *tl++ = rpc_autherr;
1362 *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1363 mreq->m_len -= NFSX_UNSIGNED;
1364 bpos -= NFSX_UNSIGNED;
1366 *tl++ = rpc_mismatch;
1367 *tl++ = txdr_unsigned(RPC_VER2);
1368 *tl = txdr_unsigned(RPC_VER2);
1371 *tl++ = rpc_msgaccepted;
1374 * For Kerberos authentication, we must send the nickname
1375 * verifier back, otherwise just RPCAUTH_NULL.
1377 if (nd->nd_flag & ND_KERBFULL) {
1378 struct nfsuid *nuidp;
1379 struct timeval ktvin, ktvout;
1381 for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1382 nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1383 if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1384 (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1385 &nuidp->nu_haddr, nd->nd_nam2)))
1390 txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1392 txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1395 * Encrypt the timestamp in ecb mode using the
1402 *tl++ = rpc_auth_kerb;
1403 *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1404 *tl = ktvout.tv_sec;
1405 nfsm_build(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1406 *tl++ = ktvout.tv_usec;
1407 *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1418 *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1421 *tl = txdr_unsigned(RPC_PROGMISMATCH);
1422 nfsm_build(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
1423 *tl++ = txdr_unsigned(2);
1424 *tl = txdr_unsigned(3);
1427 *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1430 *tl = txdr_unsigned(RPC_GARBAGE);
1434 if (err != NFSERR_RETVOID) {
1435 nfsm_build(tl, u_int32_t *, NFSX_UNSIGNED);
1437 *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1449 if (err != 0 && err != NFSERR_RETVOID)
1450 nfsstats.srvrpc_errs++;
1455 #endif /* NFS_NOSERVER */
1458 * Scan the nfsreq list and retranmit any requests that have timed out
1459 * To avoid retransmission attempts on STREAM sockets (in the future) make
1460 * sure to set the r_retry field to 0 (implies nm_retry == 0).
1463 nfs_timer(void *arg /* never used */)
1465 struct nfsmount *nmp;
1467 #ifndef NFS_NOSERVER
1468 struct nfssvc_sock *slp;
1470 #endif /* NFS_NOSERVER */
1473 TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1474 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1475 KKASSERT(nmp == req->r_nmp);
1477 (req->r_flags & (R_SOFTTERM|R_MASKTIMER))) {
1480 req->r_flags |= R_LOCKED;
1481 if (nfs_sigintr(nmp, req, req->r_td)) {
1486 req->r_flags &= ~R_LOCKED;
1489 #ifndef NFS_NOSERVER
1492 * Scan the write gathering queues for writes that need to be
1495 cur_usec = nfs_curusec();
1496 TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1497 if (slp->ns_tq.lh_first && slp->ns_tq.lh_first->nd_time<=cur_usec)
1498 nfsrv_wakenfsd(slp, 1);
1500 #endif /* NFS_NOSERVER */
1503 * Due to possible blocking, a client operation may be waiting for
1504 * us to finish processing this request so it can remove it.
1506 if (nfs_timer_raced) {
1507 nfs_timer_raced = 0;
1508 wakeup(&nfs_timer_raced);
1511 callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer, NULL);
1516 nfs_timer_req(struct nfsreq *req)
1518 struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1519 struct nfsmount *nmp = req->r_nmp;
1525 if (req->r_rtt >= 0) {
1527 if (nmp->nm_flag & NFSMNT_DUMBTIMR)
1528 timeo = nmp->nm_timeo;
1530 timeo = NFS_RTO(nmp, proct[req->r_procnum]);
1531 if (nmp->nm_timeouts > 0)
1532 timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1533 if (req->r_rtt <= timeo)
1535 if (nmp->nm_timeouts < 8)
1539 * Check for server not responding
1541 if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1542 req->r_rexmit > nmp->nm_deadthresh) {
1544 nmp->nm_mountp->mnt_stat.f_mntfromname,
1546 req->r_flags |= R_TPRINTFMSG;
1548 if (req->r_rexmit >= req->r_retry) { /* too many */
1549 nfsstats.rpctimeouts++;
1553 if (nmp->nm_sotype != SOCK_DGRAM) {
1554 if (++req->r_rexmit > NFS_MAXREXMIT)
1555 req->r_rexmit = NFS_MAXREXMIT;
1558 if ((so = nmp->nm_so) == NULL)
1562 * If there is enough space and the window allows..
1564 * Set r_rtt to -1 in case we fail to send it now.
1567 if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1568 ((nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1569 (req->r_flags & R_SENT) ||
1570 nmp->nm_sent < nmp->nm_cwnd) &&
1571 (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1572 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1573 error = so_pru_send(so, 0, m, NULL, NULL, td);
1575 error = so_pru_send(so, 0, m, nmp->nm_nam,
1578 if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1580 } else if (req->r_mrep == NULL) {
1582 * Iff first send, start timing
1583 * else turn timing off, backoff timer
1584 * and divide congestion window by 2.
1586 * It is possible for the so_pru_send() to
1587 * block and for us to race a reply so we
1588 * only do this if the reply field has not
1589 * been filled in. R_LOCKED will prevent
1590 * the request from being ripped out from under
1593 if (req->r_flags & R_SENT) {
1594 req->r_flags &= ~R_TIMING;
1595 if (++req->r_rexmit > NFS_MAXREXMIT)
1596 req->r_rexmit = NFS_MAXREXMIT;
1598 if (nmp->nm_cwnd < NFS_CWNDSCALE)
1599 nmp->nm_cwnd = NFS_CWNDSCALE;
1600 nfsstats.rpcretries++;
1602 req->r_flags |= R_SENT;
1603 nmp->nm_sent += NFS_CWNDSCALE;
1611 * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1612 * wait for all requests to complete. This is used by forced unmounts
1613 * to terminate any outstanding RPCs.
1616 nfs_nmcancelreqs(struct nfsmount *nmp)
1622 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1623 if (nmp != req->r_nmp || req->r_mrep != NULL ||
1624 (req->r_flags & R_SOFTTERM)) {
1631 for (i = 0; i < 30; i++) {
1633 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1634 if (nmp == req->r_nmp)
1640 tsleep(&lbolt, 0, "nfscancel", 0);
1646 * Flag a request as being about to terminate (due to NFSMNT_INT/NFSMNT_SOFT).
1647 * The nm_send count is decremented now to avoid deadlocks when the process in
1648 * soreceive() hasn't yet managed to send its own request.
1650 * This routine must be called at splsoftclock() to protect r_flags and
1655 nfs_softterm(struct nfsreq *rep)
1657 rep->r_flags |= R_SOFTTERM;
1659 if (rep->r_flags & R_SENT) {
1660 rep->r_nmp->nm_sent -= NFS_CWNDSCALE;
1661 rep->r_flags &= ~R_SENT;
1666 * Test for a termination condition pending on the process.
1667 * This is used for NFSMNT_INT mounts.
1670 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
1676 if (rep && (rep->r_flags & R_SOFTTERM))
1678 /* Terminate all requests while attempting a forced unmount. */
1679 if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
1681 if (!(nmp->nm_flag & NFSMNT_INT))
1683 /* td might be NULL YYY */
1684 if (td == NULL || (p = td->td_proc) == NULL)
1688 tmpset = lwp_sigpend(lp);
1689 SIGSETNAND(tmpset, lp->lwp_sigmask);
1690 SIGSETNAND(tmpset, p->p_sigignore);
1691 if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
1698 * Lock a socket against others.
1699 * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
1700 * and also to avoid race conditions between the processes with nfs requests
1701 * in progress when a reconnect is necessary.
1704 nfs_sndlock(struct nfsreq *rep)
1706 mtx_t mtx = &rep->r_nmp->nm_txlock;
1715 if (rep->r_nmp->nm_flag & NFSMNT_INT)
1718 while ((error = mtx_lock_ex_try(mtx)) != 0) {
1719 if (nfs_sigintr(rep->r_nmp, rep, td)) {
1723 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
1726 if (slpflag == PCATCH) {
1731 /* Always fail if our request has been cancelled. */
1732 if (rep->r_flags & R_SOFTTERM) {
1741 * Unlock the stream socket for others.
1744 nfs_sndunlock(struct nfsreq *rep)
1746 mtx_t mtx = &rep->r_nmp->nm_txlock;
1752 nfs_rcvlock(struct nfsreq *rep)
1754 mtx_t mtx = &rep->r_nmp->nm_rxlock;
1760 * Unconditionally check for completion in case another nfsiod
1761 * get the packet while the caller was blocked, before the caller
1762 * called us. Packet reception is handled by mainline code which
1763 * is protected by the BGL at the moment.
1765 * We do not strictly need the second check just before the
1766 * tsleep(), but it's good defensive programming.
1768 if (rep->r_mrep != NULL)
1771 if (rep->r_nmp->nm_flag & NFSMNT_INT)
1777 while ((error = mtx_lock_ex_try(mtx)) != 0) {
1778 if (nfs_sigintr(rep->r_nmp, rep, rep->r_td)) {
1782 if (rep->r_mrep != NULL) {
1788 * NOTE: can return ENOLCK, but in that case rep->r_mrep
1789 * will already be set.
1791 error = mtx_lock_ex_link(mtx, &rep->r_link, "nfsrcvlk",
1797 * If our reply was recieved while we were sleeping,
1798 * then just return without taking the lock to avoid a
1799 * situation where a single iod could 'capture' the
1802 if (rep->r_mrep != NULL) {
1806 if (slpflag == PCATCH) {
1812 if (rep->r_mrep != NULL) {
1821 * Unlock the stream socket for others.
1824 nfs_rcvunlock(struct nfsreq *rep)
1826 mtx_t mtx = &rep->r_nmp->nm_rxlock;
1834 * Check for badly aligned mbuf data and realign by copying the unaligned
1835 * portion of the data into a new mbuf chain and freeing the portions
1836 * of the old chain that were replaced.
1838 * We cannot simply realign the data within the existing mbuf chain
1839 * because the underlying buffers may contain other rpc commands and
1840 * we cannot afford to overwrite them.
1842 * We would prefer to avoid this situation entirely. The situation does
1843 * not occur with NFS/UDP and is supposed to only occassionally occur
1844 * with TCP. Use vfs.nfs.realign_count and realign_test to check this.
1847 nfs_realign(struct mbuf **pm, int hsiz)
1850 struct mbuf *n = NULL;
1855 while ((m = *pm) != NULL) {
1856 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3)) {
1857 n = m_getl(m->m_len, MB_WAIT, MT_DATA, 0, NULL);
1865 * If n is non-NULL, loop on m copying data, then replace the
1866 * portion of the chain that had to be realigned.
1869 ++nfs_realign_count;
1871 m_copyback(n, off, m->m_len, mtod(m, caddr_t));
1880 #ifndef NFS_NOSERVER
1883 * Parse an RPC request
1885 * - fill in the cred struct.
1888 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
1895 caddr_t dpos, cp2, cp;
1896 u_int32_t nfsvers, auth_type;
1898 int error = 0, ticklen;
1899 struct mbuf *mrep, *md;
1900 struct nfsuid *nuidp;
1901 struct timeval tvin, tvout;
1902 #if 0 /* until encrypted keys are implemented */
1903 NFSKERBKEYSCHED_T keys; /* stores key schedule */
1910 nfsm_dissect(tl, u_int32_t *, 10 * NFSX_UNSIGNED);
1911 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
1912 if (*tl++ != rpc_call) {
1917 nfsm_dissect(tl, u_int32_t *, 8 * NFSX_UNSIGNED);
1920 if (*tl++ != rpc_vers) {
1921 nd->nd_repstat = ERPCMISMATCH;
1922 nd->nd_procnum = NFSPROC_NOOP;
1925 if (*tl != nfs_prog) {
1926 nd->nd_repstat = EPROGUNAVAIL;
1927 nd->nd_procnum = NFSPROC_NOOP;
1931 nfsvers = fxdr_unsigned(u_int32_t, *tl++);
1932 if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
1933 nd->nd_repstat = EPROGMISMATCH;
1934 nd->nd_procnum = NFSPROC_NOOP;
1937 if (nfsvers == NFS_VER3)
1938 nd->nd_flag = ND_NFSV3;
1939 nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
1940 if (nd->nd_procnum == NFSPROC_NULL)
1942 if (nd->nd_procnum >= NFS_NPROCS ||
1943 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
1944 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
1945 nd->nd_repstat = EPROCUNAVAIL;
1946 nd->nd_procnum = NFSPROC_NOOP;
1949 if ((nd->nd_flag & ND_NFSV3) == 0)
1950 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
1952 len = fxdr_unsigned(int, *tl++);
1953 if (len < 0 || len > RPCAUTH_MAXSIZ) {
1958 nd->nd_flag &= ~ND_KERBAUTH;
1960 * Handle auth_unix or auth_kerb.
1962 if (auth_type == rpc_auth_unix) {
1963 len = fxdr_unsigned(int, *++tl);
1964 if (len < 0 || len > NFS_MAXNAMLEN) {
1968 nfsm_adv(nfsm_rndup(len));
1969 nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1970 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
1971 nd->nd_cr.cr_ref = 1;
1972 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
1973 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
1974 len = fxdr_unsigned(int, *tl);
1975 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
1979 nfsm_dissect(tl, u_int32_t *, (len + 2) * NFSX_UNSIGNED);
1980 for (i = 1; i <= len; i++)
1982 nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
1985 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
1986 if (nd->nd_cr.cr_ngroups > 1)
1987 nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
1988 len = fxdr_unsigned(int, *++tl);
1989 if (len < 0 || len > RPCAUTH_MAXSIZ) {
1994 nfsm_adv(nfsm_rndup(len));
1995 } else if (auth_type == rpc_auth_kerb) {
1996 switch (fxdr_unsigned(int, *tl++)) {
1997 case RPCAKN_FULLNAME:
1998 ticklen = fxdr_unsigned(int, *tl);
1999 *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2000 uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2001 nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2002 if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2009 uio.uio_segflg = UIO_SYSSPACE;
2010 iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2011 iov.iov_len = RPCAUTH_MAXSIZ - 4;
2012 nfsm_mtouio(&uio, uio.uio_resid);
2013 nfsm_dissect(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
2014 if (*tl++ != rpc_auth_kerb ||
2015 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2016 kprintf("Bad kerb verifier\n");
2017 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2018 nd->nd_procnum = NFSPROC_NOOP;
2021 nfsm_dissect(cp, caddr_t, 4 * NFSX_UNSIGNED);
2022 tl = (u_int32_t *)cp;
2023 if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2024 kprintf("Not fullname kerb verifier\n");
2025 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2026 nd->nd_procnum = NFSPROC_NOOP;
2029 cp += NFSX_UNSIGNED;
2030 bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2031 nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2032 nd->nd_flag |= ND_KERBFULL;
2033 nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2035 case RPCAKN_NICKNAME:
2036 if (len != 2 * NFSX_UNSIGNED) {
2037 kprintf("Kerb nickname short\n");
2038 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2039 nd->nd_procnum = NFSPROC_NOOP;
2042 nickuid = fxdr_unsigned(uid_t, *tl);
2043 nfsm_dissect(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
2044 if (*tl++ != rpc_auth_kerb ||
2045 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2046 kprintf("Kerb nick verifier bad\n");
2047 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2048 nd->nd_procnum = NFSPROC_NOOP;
2051 nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
2052 tvin.tv_sec = *tl++;
2055 for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2056 nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2057 if (nuidp->nu_cr.cr_uid == nickuid &&
2059 netaddr_match(NU_NETFAM(nuidp),
2060 &nuidp->nu_haddr, nd->nd_nam2)))
2065 (NFSERR_AUTHERR|AUTH_REJECTCRED);
2066 nd->nd_procnum = NFSPROC_NOOP;
2071 * Now, decrypt the timestamp using the session key
2078 tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2079 tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2080 if (nuidp->nu_expire < time_second ||
2081 nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2082 (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2083 nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2084 nuidp->nu_expire = 0;
2086 (NFSERR_AUTHERR|AUTH_REJECTVERF);
2087 nd->nd_procnum = NFSPROC_NOOP;
2090 nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2091 nd->nd_flag |= ND_KERBNICK;
2094 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2095 nd->nd_procnum = NFSPROC_NOOP;
2109 * Send a message to the originating process's terminal. The thread and/or
2110 * process may be NULL. YYY the thread should not be NULL but there may
2111 * still be some uio_td's that are still being passed as NULL through to
2115 nfs_msg(struct thread *td, char *server, char *msg)
2119 if (td && td->td_proc)
2120 tpr = tprintf_open(td->td_proc);
2123 tprintf(tpr, "nfs server %s: %s\n", server, msg);
2128 #ifndef NFS_NOSERVER
2130 * Socket upcall routine for the nfsd sockets.
2131 * The caddr_t arg is a pointer to the "struct nfssvc_sock".
2132 * Essentially do as much as possible non-blocking, else punt and it will
2133 * be called with MB_WAIT from an nfsd.
2136 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2138 struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2140 struct sockaddr *nam;
2143 int nparallel_wakeup = 0;
2145 if ((slp->ns_flag & SLP_VALID) == 0)
2149 * Do not allow an infinite number of completed RPC records to build
2150 * up before we stop reading data from the socket. Otherwise we could
2151 * end up holding onto an unreasonable number of mbufs for requests
2152 * waiting for service.
2154 * This should give pretty good feedback to the TCP
2155 * layer and prevents a memory crunch for other protocols.
2157 * Note that the same service socket can be dispatched to several
2158 * nfs servers simultaniously.
2160 * the tcp protocol callback calls us with MB_DONTWAIT.
2161 * nfsd calls us with MB_WAIT (typically).
2163 if (waitflag == MB_DONTWAIT && slp->ns_numrec >= nfsd_waiting / 2 + 1) {
2164 slp->ns_flag |= SLP_NEEDQ;
2169 * Handle protocol specifics to parse an RPC request. We always
2170 * pull from the socket using non-blocking I/O.
2172 if (so->so_type == SOCK_STREAM) {
2174 * The data has to be read in an orderly fashion from a TCP
2175 * stream, unlike a UDP socket. It is possible for soreceive
2176 * and/or nfsrv_getstream() to block, so make sure only one
2177 * entity is messing around with the TCP stream at any given
2178 * moment. The receive sockbuf's lock in soreceive is not
2181 * Note that this procedure can be called from any number of
2182 * NFS severs *OR* can be upcalled directly from a TCP
2185 if (slp->ns_flag & SLP_GETSTREAM) {
2186 slp->ns_flag |= SLP_NEEDQ;
2189 slp->ns_flag |= SLP_GETSTREAM;
2192 * Do soreceive(). Pull out as much data as possible without
2195 sbinit(&sio, 1000000000);
2196 flags = MSG_DONTWAIT;
2197 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2198 if (error || sio.sb_mb == NULL) {
2199 if (error == EWOULDBLOCK)
2200 slp->ns_flag |= SLP_NEEDQ;
2202 slp->ns_flag |= SLP_DISCONN;
2203 slp->ns_flag &= ~SLP_GETSTREAM;
2207 if (slp->ns_rawend) {
2208 slp->ns_rawend->m_next = m;
2209 slp->ns_cc += sio.sb_cc;
2212 slp->ns_cc = sio.sb_cc;
2219 * Now try and parse as many record(s) as we can out of the
2222 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2225 slp->ns_flag |= SLP_DISCONN;
2227 slp->ns_flag |= SLP_NEEDQ;
2229 slp->ns_flag &= ~SLP_GETSTREAM;
2232 * For UDP soreceive typically pulls just one packet, loop
2233 * to get the whole batch.
2236 sbinit(&sio, 1000000000);
2237 flags = MSG_DONTWAIT;
2238 error = so_pru_soreceive(so, &nam, NULL, &sio,
2241 struct nfsrv_rec *rec;
2242 int mf = (waitflag & MB_DONTWAIT) ?
2243 M_NOWAIT : M_WAITOK;
2244 rec = kmalloc(sizeof(struct nfsrv_rec),
2248 FREE(nam, M_SONAME);
2252 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2253 rec->nr_address = nam;
2254 rec->nr_packet = sio.sb_mb;
2255 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2260 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2261 && error != EWOULDBLOCK) {
2262 slp->ns_flag |= SLP_DISCONN;
2266 } while (sio.sb_mb);
2270 * If we were upcalled from the tcp protocol layer and we have
2271 * fully parsed records ready to go, or there is new data pending,
2272 * or something went wrong, try to wake up an nfsd thread to deal
2276 if (waitflag == MB_DONTWAIT && (slp->ns_numrec > 0
2277 || (slp->ns_flag & (SLP_NEEDQ | SLP_DISCONN)))) {
2278 nfsrv_wakenfsd(slp, nparallel_wakeup);
2283 * Try and extract an RPC request from the mbuf data list received on a
2284 * stream socket. The "waitflag" argument indicates whether or not it
2288 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2290 struct mbuf *m, **mpp;
2293 struct mbuf *om, *m2, *recm;
2297 if (slp->ns_reclen == 0) {
2298 if (slp->ns_cc < NFSX_UNSIGNED)
2301 if (m->m_len >= NFSX_UNSIGNED) {
2302 bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2303 m->m_data += NFSX_UNSIGNED;
2304 m->m_len -= NFSX_UNSIGNED;
2306 cp1 = (caddr_t)&recmark;
2307 cp2 = mtod(m, caddr_t);
2308 while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2309 while (m->m_len == 0) {
2311 cp2 = mtod(m, caddr_t);
2318 slp->ns_cc -= NFSX_UNSIGNED;
2319 recmark = ntohl(recmark);
2320 slp->ns_reclen = recmark & ~0x80000000;
2321 if (recmark & 0x80000000)
2322 slp->ns_flag |= SLP_LASTFRAG;
2324 slp->ns_flag &= ~SLP_LASTFRAG;
2325 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2326 log(LOG_ERR, "%s (%d) from nfs client\n",
2327 "impossible packet length",
2334 * Now get the record part.
2336 * Note that slp->ns_reclen may be 0. Linux sometimes
2337 * generates 0-length RPCs
2340 if (slp->ns_cc == slp->ns_reclen) {
2342 slp->ns_raw = slp->ns_rawend = NULL;
2343 slp->ns_cc = slp->ns_reclen = 0;
2344 } else if (slp->ns_cc > slp->ns_reclen) {
2349 while (len < slp->ns_reclen) {
2350 if ((len + m->m_len) > slp->ns_reclen) {
2351 m2 = m_copym(m, 0, slp->ns_reclen - len,
2359 m->m_data += slp->ns_reclen - len;
2360 m->m_len -= slp->ns_reclen - len;
2361 len = slp->ns_reclen;
2363 return (EWOULDBLOCK);
2365 } else if ((len + m->m_len) == slp->ns_reclen) {
2385 * Accumulate the fragments into a record.
2387 mpp = &slp->ns_frag;
2389 mpp = &((*mpp)->m_next);
2391 if (slp->ns_flag & SLP_LASTFRAG) {
2392 struct nfsrv_rec *rec;
2393 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2394 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2396 m_freem(slp->ns_frag);
2398 nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2399 rec->nr_address = NULL;
2400 rec->nr_packet = slp->ns_frag;
2401 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2405 slp->ns_frag = NULL;
2411 * Parse an RPC header.
2414 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2415 struct nfsrv_descript **ndp)
2417 struct nfsrv_rec *rec;
2419 struct sockaddr *nam;
2420 struct nfsrv_descript *nd;
2424 if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2426 rec = STAILQ_FIRST(&slp->ns_rec);
2427 STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2428 KKASSERT(slp->ns_numrec > 0);
2430 nam = rec->nr_address;
2432 kfree(rec, M_NFSRVDESC);
2433 MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2434 M_NFSRVDESC, M_WAITOK);
2435 nd->nd_md = nd->nd_mrep = m;
2437 nd->nd_dpos = mtod(m, caddr_t);
2438 error = nfs_getreq(nd, nfsd, TRUE);
2441 FREE(nam, M_SONAME);
2443 kfree((caddr_t)nd, M_NFSRVDESC);
2452 * Try to assign service sockets to nfsd threads based on the number
2453 * of new rpc requests that have been queued on the service socket.
2455 * If no nfsd's are available or additonal requests are pending, set the
2456 * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2457 * the work in the nfssvc_sock list when it is finished processing its
2458 * current work. This flag is only cleared when an nfsd can not find
2459 * any new work to perform.
2462 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2466 if ((slp->ns_flag & SLP_VALID) == 0)
2470 TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2471 if (nd->nfsd_flag & NFSD_WAITING) {
2472 nd->nfsd_flag &= ~NFSD_WAITING;
2474 panic("nfsd wakeup");
2477 wakeup((caddr_t)nd);
2478 if (--nparallel == 0)
2483 slp->ns_flag |= SLP_DOREC;
2484 nfsd_head_flag |= NFSD_CHECKSLP;
2487 #endif /* NFS_NOSERVER */