xref: /dflybsd-src/sys/vfs/nfs/nfs_socket.c (revision c3c96e4421a1087a390825eac6c01c9ed9182387)
1 /*
2  * Copyright (c) 1989, 1991, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Rick Macklem at The University of Guelph.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
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.
23  *
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
34  * SUCH DAMAGE.
35  *
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 $
39  */
40 
41 /*
42  * Socket operations for use by nfs
43  */
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/proc.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/kernel.h>
51 #include <sys/mbuf.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>
65 
66 #include <sys/signal2.h>
67 #include <sys/mutex2.h>
68 
69 #include <netinet/in.h>
70 #include <netinet/tcp.h>
71 #include <sys/thread2.h>
72 
73 #include "rpcv2.h"
74 #include "nfsproto.h"
75 #include "nfs.h"
76 #include "xdr_subs.h"
77 #include "nfsm_subs.h"
78 #include "nfsmount.h"
79 #include "nfsnode.h"
80 #include "nfsrtt.h"
81 
82 #define	TRUE	1
83 #define	FALSE	0
84 
85 /*
86  * RTT calculations are scaled by 256 (8 bits).  A proper fractional
87  * RTT will still be calculated even with a slow NFS timer.
88  */
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 */
93 
94 /*
95  * Defines which timer to use for the procnum.
96  * 0 - default
97  * 1 - getattr
98  * 2 - lookup
99  * 3 - read
100  * 4 - write
101  */
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	*/
106 };
107 
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	*/
112 };
113 
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;
120 
121 SYSCTL_DECL(_vfs_nfs);
122 
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, "");
128 
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);
134 
135 int nfsrtton = 0;
136 struct nfsrtt nfsrtt;
137 struct callout	nfs_timer_handle;
138 
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);
148 #ifndef NFS_NOSERVER
149 static int	nfsrv_getstream (struct nfssvc_sock *, int, int *);
150 static void	nfs_timer_req(struct nfsreq *req);
151 static void	nfs_checkpkt(struct mbuf *m, int len);
152 
153 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
154 				    struct nfssvc_sock *slp,
155 				    struct thread *td,
156 				    struct mbuf **mreqp) = {
157 	nfsrv_null,
158 	nfsrv_getattr,
159 	nfsrv_setattr,
160 	nfsrv_lookup,
161 	nfsrv3_access,
162 	nfsrv_readlink,
163 	nfsrv_read,
164 	nfsrv_write,
165 	nfsrv_create,
166 	nfsrv_mkdir,
167 	nfsrv_symlink,
168 	nfsrv_mknod,
169 	nfsrv_remove,
170 	nfsrv_rmdir,
171 	nfsrv_rename,
172 	nfsrv_link,
173 	nfsrv_readdir,
174 	nfsrv_readdirplus,
175 	nfsrv_statfs,
176 	nfsrv_fsinfo,
177 	nfsrv_pathconf,
178 	nfsrv_commit,
179 	nfsrv_noop,
180 	nfsrv_noop,
181 	nfsrv_noop,
182 	nfsrv_noop
183 };
184 #endif /* NFS_NOSERVER */
185 
186 /*
187  * Initialize sockets and congestion for a new NFS connection.
188  * We do not free the sockaddr if error.
189  */
190 int
191 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
192 {
193 	struct socket *so;
194 	int error;
195 	struct sockaddr *saddr;
196 	struct sockaddr_in *sin;
197 	struct thread *td = &thread0; /* only used for socreate and sobind */
198 
199 	nmp->nm_so = so = NULL;
200 	if (nmp->nm_flag & NFSMNT_FORCE)
201 		return (EINVAL);
202 	saddr = nmp->nm_nam;
203 	error = socreate(saddr->sa_family, &so, nmp->nm_sotype,
204 		nmp->nm_soproto, td);
205 	if (error)
206 		goto bad;
207 	nmp->nm_soflags = so->so_proto->pr_flags;
208 
209 	/*
210 	 * Some servers require that the client port be a reserved port number.
211 	 */
212 	if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
213 		struct sockopt sopt;
214 		int ip;
215 		struct sockaddr_in ssin;
216 
217 		bzero(&sopt, sizeof sopt);
218 		ip = IP_PORTRANGE_LOW;
219 		sopt.sopt_level = IPPROTO_IP;
220 		sopt.sopt_name = IP_PORTRANGE;
221 		sopt.sopt_val = (void *)&ip;
222 		sopt.sopt_valsize = sizeof(ip);
223 		sopt.sopt_td = NULL;
224 		error = sosetopt(so, &sopt);
225 		if (error)
226 			goto bad;
227 		bzero(&ssin, sizeof ssin);
228 		sin = &ssin;
229 		sin->sin_len = sizeof (struct sockaddr_in);
230 		sin->sin_family = AF_INET;
231 		sin->sin_addr.s_addr = INADDR_ANY;
232 		sin->sin_port = htons(0);
233 		error = sobind(so, (struct sockaddr *)sin, td);
234 		if (error)
235 			goto bad;
236 		bzero(&sopt, sizeof sopt);
237 		ip = IP_PORTRANGE_DEFAULT;
238 		sopt.sopt_level = IPPROTO_IP;
239 		sopt.sopt_name = IP_PORTRANGE;
240 		sopt.sopt_val = (void *)&ip;
241 		sopt.sopt_valsize = sizeof(ip);
242 		sopt.sopt_td = NULL;
243 		error = sosetopt(so, &sopt);
244 		if (error)
245 			goto bad;
246 	}
247 
248 	/*
249 	 * Protocols that do not require connections may be optionally left
250 	 * unconnected for servers that reply from a port other than NFS_PORT.
251 	 */
252 	if (nmp->nm_flag & NFSMNT_NOCONN) {
253 		if (nmp->nm_soflags & PR_CONNREQUIRED) {
254 			error = ENOTCONN;
255 			goto bad;
256 		}
257 	} else {
258 		error = soconnect(so, nmp->nm_nam, td);
259 		if (error)
260 			goto bad;
261 
262 		/*
263 		 * Wait for the connection to complete. Cribbed from the
264 		 * connect system call but with the wait timing out so
265 		 * that interruptible mounts don't hang here for a long time.
266 		 */
267 		crit_enter();
268 		while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
269 			(void) tsleep((caddr_t)&so->so_timeo, 0,
270 				"nfscon", 2 * hz);
271 			if ((so->so_state & SS_ISCONNECTING) &&
272 			    so->so_error == 0 && rep &&
273 			    (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
274 				so->so_state &= ~SS_ISCONNECTING;
275 				crit_exit();
276 				goto bad;
277 			}
278 		}
279 		if (so->so_error) {
280 			error = so->so_error;
281 			so->so_error = 0;
282 			crit_exit();
283 			goto bad;
284 		}
285 		crit_exit();
286 	}
287 	so->so_rcv.ssb_timeo = (5 * hz);
288 	so->so_snd.ssb_timeo = (5 * hz);
289 
290 	/*
291 	 * Get buffer reservation size from sysctl, but impose reasonable
292 	 * limits.
293 	 */
294 	if (nmp->nm_sotype == SOCK_STREAM) {
295 		if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
296 			struct sockopt sopt;
297 			int val;
298 
299 			bzero(&sopt, sizeof sopt);
300 			sopt.sopt_level = SOL_SOCKET;
301 			sopt.sopt_name = SO_KEEPALIVE;
302 			sopt.sopt_val = &val;
303 			sopt.sopt_valsize = sizeof val;
304 			val = 1;
305 			sosetopt(so, &sopt);
306 		}
307 		if (so->so_proto->pr_protocol == IPPROTO_TCP) {
308 			struct sockopt sopt;
309 			int val;
310 
311 			bzero(&sopt, sizeof sopt);
312 			sopt.sopt_level = IPPROTO_TCP;
313 			sopt.sopt_name = TCP_NODELAY;
314 			sopt.sopt_val = &val;
315 			sopt.sopt_valsize = sizeof val;
316 			val = 1;
317 			sosetopt(so, &sopt);
318 		}
319 	}
320 	error = soreserve(so, nfs_soreserve, nfs_soreserve, NULL);
321 	if (error)
322 		goto bad;
323 	atomic_set_int(&so->so_rcv.ssb_flags, SSB_NOINTR);
324 	atomic_set_int(&so->so_snd.ssb_flags, SSB_NOINTR);
325 
326 	/* Initialize other non-zero congestion variables */
327 	nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] =
328 		nmp->nm_srtt[3] = (NFS_TIMEO << NFS_RTT_SCALE_BITS);
329 	nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
330 		nmp->nm_sdrtt[3] = 0;
331 	nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
332 	nmp->nm_timeouts = 0;
333 
334 	/*
335 	 * Assign nm_so last.  The moment nm_so is assigned the nfs_timer()
336 	 * can mess with the socket.
337 	 */
338 	nmp->nm_so = so;
339 	return (0);
340 
341 bad:
342 	if (so) {
343 		soshutdown(so, SHUT_RDWR);
344 		soclose(so, FNONBLOCK);
345 	}
346 	return (error);
347 }
348 
349 /*
350  * Reconnect routine:
351  * Called when a connection is broken on a reliable protocol.
352  * - clean up the old socket
353  * - nfs_connect() again
354  * - set R_NEEDSXMIT for all outstanding requests on mount point
355  * If this fails the mount point is DEAD!
356  * nb: Must be called with the nfs_sndlock() set on the mount point.
357  */
358 static int
359 nfs_reconnect(struct nfsmount *nmp, struct nfsreq *rep)
360 {
361 	struct nfsreq *req;
362 	int error;
363 
364 	nfs_disconnect(nmp);
365 	if (nmp->nm_rxstate >= NFSSVC_STOPPING)
366 		return (EINTR);
367 	while ((error = nfs_connect(nmp, rep)) != 0) {
368 		if (error == EINTR || error == ERESTART)
369 			return (EINTR);
370 		if (error == EINVAL)
371 			return (error);
372 		if (nmp->nm_rxstate >= NFSSVC_STOPPING)
373 			return (EINTR);
374 		(void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
375 	}
376 
377 	/*
378 	 * Loop through outstanding request list and fix up all requests
379 	 * on old socket.
380 	 */
381 	crit_enter();
382 	TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
383 		KKASSERT(req->r_nmp == nmp);
384 		req->r_flags |= R_NEEDSXMIT;
385 	}
386 	crit_exit();
387 	return (0);
388 }
389 
390 /*
391  * NFS disconnect. Clean up and unlink.
392  */
393 void
394 nfs_disconnect(struct nfsmount *nmp)
395 {
396 	struct socket *so;
397 
398 	if (nmp->nm_so) {
399 		so = nmp->nm_so;
400 		nmp->nm_so = NULL;
401 		soshutdown(so, SHUT_RDWR);
402 		soclose(so, FNONBLOCK);
403 	}
404 }
405 
406 void
407 nfs_safedisconnect(struct nfsmount *nmp)
408 {
409 	nfs_rcvlock(nmp, NULL);
410 	nfs_disconnect(nmp);
411 	nfs_rcvunlock(nmp);
412 }
413 
414 /*
415  * This is the nfs send routine. For connection based socket types, it
416  * must be called with an nfs_sndlock() on the socket.
417  * "rep == NULL" indicates that it has been called from a server.
418  * For the client side:
419  * - return EINTR if the RPC is terminated, 0 otherwise
420  * - set R_NEEDSXMIT if the send fails for any reason
421  * - do any cleanup required by recoverable socket errors (?)
422  * For the server side:
423  * - return EINTR or ERESTART if interrupted by a signal
424  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
425  * - do any cleanup required by recoverable socket errors (?)
426  */
427 int
428 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
429 	 struct nfsreq *rep)
430 {
431 	struct sockaddr *sendnam;
432 	int error, soflags, flags;
433 
434 	if (rep) {
435 		if (rep->r_flags & R_SOFTTERM) {
436 			m_freem(top);
437 			return (EINTR);
438 		}
439 		if ((so = rep->r_nmp->nm_so) == NULL) {
440 			rep->r_flags |= R_NEEDSXMIT;
441 			m_freem(top);
442 			return (0);
443 		}
444 		rep->r_flags &= ~R_NEEDSXMIT;
445 		soflags = rep->r_nmp->nm_soflags;
446 	} else {
447 		soflags = so->so_proto->pr_flags;
448 	}
449 	if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
450 		sendnam = NULL;
451 	else
452 		sendnam = nam;
453 	if (so->so_type == SOCK_SEQPACKET)
454 		flags = MSG_EOR;
455 	else
456 		flags = 0;
457 
458 	/*
459 	 * calls pru_sosend -> sosend -> so_pru_send -> netrpc
460 	 */
461 	error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
462 			      curthread /*XXX*/);
463 	/*
464 	 * ENOBUFS for dgram sockets is transient and non fatal.
465 	 * No need to log, and no need to break a soft mount.
466 	 */
467 	if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
468 		error = 0;
469 		/*
470 		 * do backoff retransmit on client
471 		 */
472 		if (rep) {
473 			if ((rep->r_nmp->nm_state & NFSSTA_SENDSPACE) == 0) {
474 				rep->r_nmp->nm_state |= NFSSTA_SENDSPACE;
475 				kprintf("Warning: NFS: Insufficient sendspace "
476 					"(%lu),\n"
477 					"\t You must increase vfs.nfs.soreserve"
478 					"or decrease vfs.nfs.maxasyncbio\n",
479 					so->so_snd.ssb_hiwat);
480 			}
481 			rep->r_flags |= R_NEEDSXMIT;
482 		}
483 	}
484 
485 	if (error) {
486 		if (rep) {
487 			log(LOG_INFO, "nfs send error %d for server %s\n",error,
488 			    rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
489 			/*
490 			 * Deal with errors for the client side.
491 			 */
492 			if (rep->r_flags & R_SOFTTERM)
493 				error = EINTR;
494 			else
495 				rep->r_flags |= R_NEEDSXMIT;
496 		} else {
497 			log(LOG_INFO, "nfsd send error %d\n", error);
498 		}
499 
500 		/*
501 		 * Handle any recoverable (soft) socket errors here. (?)
502 		 */
503 		if (error != EINTR && error != ERESTART &&
504 			error != EWOULDBLOCK && error != EPIPE)
505 			error = 0;
506 	}
507 	return (error);
508 }
509 
510 /*
511  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
512  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
513  * Mark and consolidate the data into a new mbuf list.
514  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
515  *     small mbufs.
516  * For SOCK_STREAM we must be very careful to read an entire record once
517  * we have read any of it, even if the system call has been interrupted.
518  */
519 static int
520 nfs_receive(struct nfsmount *nmp, struct nfsreq *rep,
521 	    struct sockaddr **aname, struct mbuf **mp)
522 {
523 	struct socket *so;
524 	struct sockbuf sio;
525 	struct uio auio;
526 	struct iovec aio;
527 	struct mbuf *m;
528 	struct mbuf *control;
529 	u_int32_t len;
530 	struct sockaddr **getnam;
531 	int error, sotype, rcvflg;
532 	struct thread *td = curthread;	/* XXX */
533 
534 	/*
535 	 * Set up arguments for soreceive()
536 	 */
537 	*mp = NULL;
538 	*aname = NULL;
539 	sotype = nmp->nm_sotype;
540 
541 	/*
542 	 * For reliable protocols, lock against other senders/receivers
543 	 * in case a reconnect is necessary.
544 	 * For SOCK_STREAM, first get the Record Mark to find out how much
545 	 * more there is to get.
546 	 * We must lock the socket against other receivers
547 	 * until we have an entire rpc request/reply.
548 	 */
549 	if (sotype != SOCK_DGRAM) {
550 		error = nfs_sndlock(nmp, rep);
551 		if (error)
552 			return (error);
553 tryagain:
554 		/*
555 		 * Check for fatal errors and resending request.
556 		 */
557 		/*
558 		 * Ugh: If a reconnect attempt just happened, nm_so
559 		 * would have changed. NULL indicates a failed
560 		 * attempt that has essentially shut down this
561 		 * mount point.
562 		 */
563 		if (rep && (rep->r_mrep || (rep->r_flags & R_SOFTTERM))) {
564 			nfs_sndunlock(nmp);
565 			return (EINTR);
566 		}
567 		so = nmp->nm_so;
568 		if (so == NULL) {
569 			error = nfs_reconnect(nmp, rep);
570 			if (error) {
571 				nfs_sndunlock(nmp);
572 				return (error);
573 			}
574 			goto tryagain;
575 		}
576 		while (rep && (rep->r_flags & R_NEEDSXMIT)) {
577 			m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
578 			nfsstats.rpcretries++;
579 			error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
580 			if (error) {
581 				if (error == EINTR || error == ERESTART ||
582 				    (error = nfs_reconnect(nmp, rep)) != 0) {
583 					nfs_sndunlock(nmp);
584 					return (error);
585 				}
586 				goto tryagain;
587 			}
588 		}
589 		nfs_sndunlock(nmp);
590 		if (sotype == SOCK_STREAM) {
591 			/*
592 			 * Get the length marker from the stream
593 			 */
594 			aio.iov_base = (caddr_t)&len;
595 			aio.iov_len = sizeof(u_int32_t);
596 			auio.uio_iov = &aio;
597 			auio.uio_iovcnt = 1;
598 			auio.uio_segflg = UIO_SYSSPACE;
599 			auio.uio_rw = UIO_READ;
600 			auio.uio_offset = 0;
601 			auio.uio_resid = sizeof(u_int32_t);
602 			auio.uio_td = td;
603 			do {
604 			   rcvflg = MSG_WAITALL;
605 			   error = so_pru_soreceive(so, NULL, &auio, NULL,
606 						    NULL, &rcvflg);
607 			   if (error == EWOULDBLOCK && rep) {
608 				if (rep->r_flags & R_SOFTTERM)
609 					return (EINTR);
610 			   }
611 			} while (error == EWOULDBLOCK);
612 
613 			if (error == 0 && auio.uio_resid > 0) {
614 			    /*
615 			     * Only log short packets if not EOF
616 			     */
617 			    if (auio.uio_resid != sizeof(u_int32_t))
618 			    log(LOG_INFO,
619 				 "short receive (%d/%d) from nfs server %s\n",
620 				 (int)(sizeof(u_int32_t) - auio.uio_resid),
621 				 (int)sizeof(u_int32_t),
622 				 nmp->nm_mountp->mnt_stat.f_mntfromname);
623 			    error = EPIPE;
624 			}
625 			if (error)
626 				goto errout;
627 			len = ntohl(len) & ~0x80000000;
628 			/*
629 			 * This is SERIOUS! We are out of sync with the sender
630 			 * and forcing a disconnect/reconnect is all I can do.
631 			 */
632 			if (len > NFS_MAXPACKET) {
633 			    log(LOG_ERR, "%s (%d) from nfs server %s\n",
634 				"impossible packet length",
635 				len,
636 				nmp->nm_mountp->mnt_stat.f_mntfromname);
637 			    error = EFBIG;
638 			    goto errout;
639 			}
640 
641 			/*
642 			 * Get the rest of the packet as an mbuf chain
643 			 */
644 			sbinit(&sio, len);
645 			do {
646 			    rcvflg = MSG_WAITALL;
647 			    error = so_pru_soreceive(so, NULL, NULL, &sio,
648 						     NULL, &rcvflg);
649 			} while (error == EWOULDBLOCK || error == EINTR ||
650 				 error == ERESTART);
651 			if (error == 0 && sio.sb_cc != len) {
652 			    if (sio.sb_cc != 0)
653 			    log(LOG_INFO,
654 				"short receive (%zu/%d) from nfs server %s\n",
655 				(size_t)len - auio.uio_resid, len,
656 				nmp->nm_mountp->mnt_stat.f_mntfromname);
657 			    error = EPIPE;
658 			}
659 			*mp = sio.sb_mb;
660 		} else {
661 			/*
662 			 * Non-stream, so get the whole packet by not
663 			 * specifying MSG_WAITALL and by specifying a large
664 			 * length.
665 			 *
666 			 * We have no use for control msg., but must grab them
667 			 * and then throw them away so we know what is going
668 			 * on.
669 			 */
670 			sbinit(&sio, 100000000);
671 			do {
672 			    rcvflg = 0;
673 			    error =  so_pru_soreceive(so, NULL, NULL, &sio,
674 						      &control, &rcvflg);
675 			    if (control)
676 				m_freem(control);
677 			    if (error == EWOULDBLOCK && rep) {
678 				if (rep->r_flags & R_SOFTTERM) {
679 					m_freem(sio.sb_mb);
680 					return (EINTR);
681 				}
682 			    }
683 			} while (error == EWOULDBLOCK ||
684 				 (error == 0 && sio.sb_mb == NULL && control));
685 			if ((rcvflg & MSG_EOR) == 0)
686 				kprintf("Egad!!\n");
687 			if (error == 0 && sio.sb_mb == NULL)
688 				error = EPIPE;
689 			len = sio.sb_cc;
690 			*mp = sio.sb_mb;
691 		}
692 errout:
693 		if (error && error != EINTR && error != ERESTART) {
694 			m_freem(*mp);
695 			*mp = NULL;
696 			if (error != EPIPE) {
697 				log(LOG_INFO,
698 				    "receive error %d from nfs server %s\n",
699 				    error,
700 				 nmp->nm_mountp->mnt_stat.f_mntfromname);
701 			}
702 			error = nfs_sndlock(nmp, rep);
703 			if (!error) {
704 				error = nfs_reconnect(nmp, rep);
705 				if (!error)
706 					goto tryagain;
707 				else
708 					nfs_sndunlock(nmp);
709 			}
710 		}
711 	} else {
712 		if ((so = nmp->nm_so) == NULL)
713 			return (EACCES);
714 		if (so->so_state & SS_ISCONNECTED)
715 			getnam = NULL;
716 		else
717 			getnam = aname;
718 		sbinit(&sio, 100000000);
719 		do {
720 			rcvflg = 0;
721 			error =  so_pru_soreceive(so, getnam, NULL, &sio,
722 						  NULL, &rcvflg);
723 			if (error == EWOULDBLOCK && rep &&
724 			    (rep->r_flags & R_SOFTTERM)) {
725 				m_freem(sio.sb_mb);
726 				return (EINTR);
727 			}
728 		} while (error == EWOULDBLOCK);
729 
730 		len = sio.sb_cc;
731 		*mp = sio.sb_mb;
732 
733 		/*
734 		 * A shutdown may result in no error and no mbuf.
735 		 * Convert to EPIPE.
736 		 */
737 		if (*mp == NULL && error == 0)
738 			error = EPIPE;
739 	}
740 	if (error) {
741 		m_freem(*mp);
742 		*mp = NULL;
743 	}
744 
745 	/*
746 	 * Search for any mbufs that are not a multiple of 4 bytes long
747 	 * or with m_data not longword aligned.
748 	 * These could cause pointer alignment problems, so copy them to
749 	 * well aligned mbufs.
750 	 */
751 	nfs_realign(mp, 5 * NFSX_UNSIGNED);
752 	return (error);
753 }
754 
755 /*
756  * Implement receipt of reply on a socket.
757  *
758  * We must search through the list of received datagrams matching them
759  * with outstanding requests using the xid, until ours is found.
760  *
761  * If myrep is NULL we process packets on the socket until
762  * interrupted or until nm_reqrxq is non-empty.
763  */
764 /* ARGSUSED */
765 int
766 nfs_reply(struct nfsmount *nmp, struct nfsreq *myrep)
767 {
768 	struct nfsreq *rep;
769 	struct sockaddr *nam;
770 	u_int32_t rxid;
771 	u_int32_t *tl;
772 	int error;
773 	struct nfsm_info info;
774 
775 	/*
776 	 * Loop around until we get our own reply
777 	 */
778 	for (;;) {
779 		/*
780 		 * Lock against other receivers so that I don't get stuck in
781 		 * sbwait() after someone else has received my reply for me.
782 		 * Also necessary for connection based protocols to avoid
783 		 * race conditions during a reconnect.
784 		 *
785 		 * If nfs_rcvlock() returns EALREADY, that means that
786 		 * the reply has already been recieved by another
787 		 * process and we can return immediately.  In this
788 		 * case, the lock is not taken to avoid races with
789 		 * other processes.
790 		 */
791 		info.mrep = NULL;
792 
793 		error = nfs_rcvlock(nmp, myrep);
794 		if (error == EALREADY)
795 			return (0);
796 		if (error)
797 			return (error);
798 
799 		/*
800 		 * If myrep is NULL we are the receiver helper thread.
801 		 * Stop waiting for incoming replies if there are
802 		 * messages sitting on reqrxq that we need to process,
803 		 * or if a shutdown request is pending.
804 		 */
805 		if (myrep == NULL && (TAILQ_FIRST(&nmp->nm_reqrxq) ||
806 		    nmp->nm_rxstate > NFSSVC_PENDING)) {
807 			nfs_rcvunlock(nmp);
808 			return(EWOULDBLOCK);
809 		}
810 
811 		/*
812 		 * Get the next Rpc reply off the socket
813 		 *
814 		 * We cannot release the receive lock until we've
815 		 * filled in rep->r_mrep, otherwise a waiting
816 		 * thread may deadlock in soreceive with no incoming
817 		 * packets expected.
818 		 */
819 		error = nfs_receive(nmp, myrep, &nam, &info.mrep);
820 		if (error) {
821 			/*
822 			 * Ignore routing errors on connectionless protocols??
823 			 */
824 			nfs_rcvunlock(nmp);
825 			if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
826 				if (nmp->nm_so == NULL)
827 					return (error);
828 				nmp->nm_so->so_error = 0;
829 				continue;
830 			}
831 			return (error);
832 		}
833 		if (nam)
834 			FREE(nam, M_SONAME);
835 
836 		/*
837 		 * Get the xid and check that it is an rpc reply
838 		 */
839 		info.md = info.mrep;
840 		info.dpos = mtod(info.md, caddr_t);
841 		NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
842 		rxid = *tl++;
843 		if (*tl != rpc_reply) {
844 			nfsstats.rpcinvalid++;
845 			m_freem(info.mrep);
846 			info.mrep = NULL;
847 nfsmout:
848 			nfs_rcvunlock(nmp);
849 			continue;
850 		}
851 
852 		/*
853 		 * Loop through the request list to match up the reply
854 		 * Iff no match, just drop the datagram.  On match, set
855 		 * r_mrep atomically to prevent the timer from messing
856 		 * around with the request after we have exited the critical
857 		 * section.
858 		 */
859 		crit_enter();
860 		TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
861 			if (rep->r_mrep == NULL && rxid == rep->r_xid)
862 				break;
863 		}
864 
865 		/*
866 		 * Fill in the rest of the reply if we found a match.
867 		 *
868 		 * Deal with duplicate responses if there was no match.
869 		 */
870 		if (rep) {
871 			rep->r_md = info.md;
872 			rep->r_dpos = info.dpos;
873 			if (nfsrtton) {
874 				struct rttl *rt;
875 
876 				rt = &nfsrtt.rttl[nfsrtt.pos];
877 				rt->proc = rep->r_procnum;
878 				rt->rto = 0;
879 				rt->sent = 0;
880 				rt->cwnd = nmp->nm_maxasync_scaled;
881 				rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
882 				rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
883 				rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
884 				getmicrotime(&rt->tstamp);
885 				if (rep->r_flags & R_TIMING)
886 					rt->rtt = rep->r_rtt;
887 				else
888 					rt->rtt = 1000000;
889 				nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
890 			}
891 
892 			/*
893 			 * New congestion control is based only on async
894 			 * requests.
895 			 */
896 			if (nmp->nm_maxasync_scaled < NFS_MAXASYNC_SCALED)
897 				++nmp->nm_maxasync_scaled;
898 			if (rep->r_flags & R_SENT) {
899 				rep->r_flags &= ~R_SENT;
900 			}
901 			/*
902 			 * Update rtt using a gain of 0.125 on the mean
903 			 * and a gain of 0.25 on the deviation.
904 			 *
905 			 * NOTE SRTT/SDRTT are only good if R_TIMING is set.
906 			 */
907 			if ((rep->r_flags & R_TIMING) && rep->r_rexmit == 0) {
908 				/*
909 				 * Since the timer resolution of
910 				 * NFS_HZ is so course, it can often
911 				 * result in r_rtt == 0. Since
912 				 * r_rtt == N means that the actual
913 				 * rtt is between N+dt and N+2-dt ticks,
914 				 * add 1.
915 				 */
916 				int n;
917 				int d;
918 
919 #define NFSRSB	NFS_RTT_SCALE_BITS
920 				n = ((NFS_SRTT(rep) * 7) +
921 				     (rep->r_rtt << NFSRSB)) >> 3;
922 				d = n - NFS_SRTT(rep);
923 				NFS_SRTT(rep) = n;
924 
925 				/*
926 				 * Don't let the jitter calculation decay
927 				 * too quickly, but we want a fast rampup.
928 				 */
929 				if (d < 0)
930 					d = -d;
931 				d <<= NFSRSB;
932 				if (d < NFS_SDRTT(rep))
933 					n = ((NFS_SDRTT(rep) * 15) + d) >> 4;
934 				else
935 					n = ((NFS_SDRTT(rep) * 3) + d) >> 2;
936 				NFS_SDRTT(rep) = n;
937 #undef NFSRSB
938 			}
939 			nmp->nm_timeouts = 0;
940 			rep->r_mrep = info.mrep;
941 			nfs_hardterm(rep, 0);
942 		} else {
943 			/*
944 			 * Extract vers, prog, nfsver, procnum.  A duplicate
945 			 * response means we didn't wait long enough so
946 			 * we increase the SRTT to avoid future spurious
947 			 * timeouts.
948 			 */
949 			u_int procnum = nmp->nm_lastreprocnum;
950 			int n;
951 
952 			if (procnum < NFS_NPROCS && proct[procnum]) {
953 				if (nfs_showrexmit)
954 					kprintf("D");
955 				n = nmp->nm_srtt[proct[procnum]];
956 				n += NFS_ASYSCALE * NFS_HZ;
957 				if (n < NFS_ASYSCALE * NFS_HZ * 10)
958 					n = NFS_ASYSCALE * NFS_HZ * 10;
959 				nmp->nm_srtt[proct[procnum]] = n;
960 			}
961 		}
962 		nfs_rcvunlock(nmp);
963 		crit_exit();
964 
965 		/*
966 		 * If not matched to a request, drop it.
967 		 * If it's mine, get out.
968 		 */
969 		if (rep == NULL) {
970 			nfsstats.rpcunexpected++;
971 			m_freem(info.mrep);
972 			info.mrep = NULL;
973 		} else if (rep == myrep) {
974 			if (rep->r_mrep == NULL)
975 				panic("nfsreply nil");
976 			return (0);
977 		}
978 	}
979 }
980 
981 /*
982  * Run the request state machine until the target state is reached
983  * or a fatal error occurs.  The target state is not run.  Specifying
984  * a target of NFSM_STATE_DONE runs the state machine until the rpc
985  * is complete.
986  *
987  * EINPROGRESS is returned for all states other then the DONE state,
988  * indicating that the rpc is still in progress.
989  */
990 int
991 nfs_request(struct nfsm_info *info, nfsm_state_t bstate, nfsm_state_t estate)
992 {
993 	struct nfsreq *req;
994 
995 	while (info->state >= bstate && info->state < estate) {
996 		switch(info->state) {
997 		case NFSM_STATE_SETUP:
998 			/*
999 			 * Setup the nfsreq.  Any error which occurs during
1000 			 * this state is fatal.
1001 			 */
1002 			info->error = nfs_request_setup(info);
1003 			if (info->error) {
1004 				info->state = NFSM_STATE_DONE;
1005 				return (info->error);
1006 			} else {
1007 				req = info->req;
1008 				req->r_mrp = &info->mrep;
1009 				req->r_mdp = &info->md;
1010 				req->r_dposp = &info->dpos;
1011 				info->state = NFSM_STATE_AUTH;
1012 			}
1013 			break;
1014 		case NFSM_STATE_AUTH:
1015 			/*
1016 			 * Authenticate the nfsreq.  Any error which occurs
1017 			 * during this state is fatal.
1018 			 */
1019 			info->error = nfs_request_auth(info->req);
1020 			if (info->error) {
1021 				info->state = NFSM_STATE_DONE;
1022 				return (info->error);
1023 			} else {
1024 				info->state = NFSM_STATE_TRY;
1025 			}
1026 			break;
1027 		case NFSM_STATE_TRY:
1028 			/*
1029 			 * Transmit or retransmit attempt.  An error in this
1030 			 * state is ignored and we always move on to the
1031 			 * next state.
1032 			 *
1033 			 * This can trivially race the receiver if the
1034 			 * request is asynchronous.  nfs_request_try()
1035 			 * will thus set the state for us and we
1036 			 * must also return immediately if we are
1037 			 * running an async state machine, because
1038 			 * info can become invalid due to races after
1039 			 * try() returns.
1040 			 */
1041 			if (info->req->r_flags & R_ASYNC) {
1042 				nfs_request_try(info->req);
1043 				if (estate == NFSM_STATE_WAITREPLY)
1044 					return (EINPROGRESS);
1045 			} else {
1046 				nfs_request_try(info->req);
1047 				info->state = NFSM_STATE_WAITREPLY;
1048 			}
1049 			break;
1050 		case NFSM_STATE_WAITREPLY:
1051 			/*
1052 			 * Wait for a reply or timeout and move on to the
1053 			 * next state.  The error returned by this state
1054 			 * is passed to the processing code in the next
1055 			 * state.
1056 			 */
1057 			info->error = nfs_request_waitreply(info->req);
1058 			info->state = NFSM_STATE_PROCESSREPLY;
1059 			break;
1060 		case NFSM_STATE_PROCESSREPLY:
1061 			/*
1062 			 * Process the reply or timeout.  Errors which occur
1063 			 * in this state may cause the state machine to
1064 			 * go back to an earlier state, and are fatal
1065 			 * otherwise.
1066 			 */
1067 			info->error = nfs_request_processreply(info,
1068 							       info->error);
1069 			switch(info->error) {
1070 			case ENEEDAUTH:
1071 				info->state = NFSM_STATE_AUTH;
1072 				break;
1073 			case EAGAIN:
1074 				info->state = NFSM_STATE_TRY;
1075 				break;
1076 			default:
1077 				/*
1078 				 * Operation complete, with or without an
1079 				 * error.  We are done.
1080 				 */
1081 				info->req = NULL;
1082 				info->state = NFSM_STATE_DONE;
1083 				return (info->error);
1084 			}
1085 			break;
1086 		case NFSM_STATE_DONE:
1087 			/*
1088 			 * Shouldn't be reached
1089 			 */
1090 			return (info->error);
1091 			/* NOT REACHED */
1092 		}
1093 	}
1094 
1095 	/*
1096 	 * If we are done return the error code (if any).
1097 	 * Otherwise return EINPROGRESS.
1098 	 */
1099 	if (info->state == NFSM_STATE_DONE)
1100 		return (info->error);
1101 	return (EINPROGRESS);
1102 }
1103 
1104 /*
1105  * nfs_request - goes something like this
1106  *	- fill in request struct
1107  *	- links it into list
1108  *	- calls nfs_send() for first transmit
1109  *	- calls nfs_receive() to get reply
1110  *	- break down rpc header and return with nfs reply pointed to
1111  *	  by mrep or error
1112  * nb: always frees up mreq mbuf list
1113  */
1114 static int
1115 nfs_request_setup(nfsm_info_t info)
1116 {
1117 	struct nfsreq *req;
1118 	struct nfsmount *nmp;
1119 	struct mbuf *m;
1120 	int i;
1121 
1122 	/*
1123 	 * Reject requests while attempting a forced unmount.
1124 	 */
1125 	if (info->vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
1126 		m_freem(info->mreq);
1127 		info->mreq = NULL;
1128 		return (ESTALE);
1129 	}
1130 	nmp = VFSTONFS(info->vp->v_mount);
1131 	req = kmalloc(sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
1132 	req->r_nmp = nmp;
1133 	req->r_vp = info->vp;
1134 	req->r_td = info->td;
1135 	req->r_procnum = info->procnum;
1136 	req->r_mreq = NULL;
1137 	req->r_cred = info->cred;
1138 
1139 	i = 0;
1140 	m = info->mreq;
1141 	while (m) {
1142 		i += m->m_len;
1143 		m = m->m_next;
1144 	}
1145 	req->r_mrest = info->mreq;
1146 	req->r_mrest_len = i;
1147 
1148 	/*
1149 	 * The presence of a non-NULL r_info in req indicates
1150 	 * async completion via our helper threads.  See the receiver
1151 	 * code.
1152 	 */
1153 	if (info->bio) {
1154 		req->r_info = info;
1155 		req->r_flags = R_ASYNC;
1156 	} else {
1157 		req->r_info = NULL;
1158 		req->r_flags = 0;
1159 	}
1160 	info->req = req;
1161 	return(0);
1162 }
1163 
1164 static int
1165 nfs_request_auth(struct nfsreq *rep)
1166 {
1167 	struct nfsmount *nmp = rep->r_nmp;
1168 	struct mbuf *m;
1169 	char nickv[RPCX_NICKVERF];
1170 	int error = 0, auth_len, auth_type;
1171 	int verf_len;
1172 	u_int32_t xid;
1173 	char *auth_str, *verf_str;
1174 	struct ucred *cred;
1175 
1176 	cred = rep->r_cred;
1177 	rep->r_failed_auth = 0;
1178 
1179 	/*
1180 	 * Get the RPC header with authorization.
1181 	 */
1182 	verf_str = auth_str = NULL;
1183 	if (nmp->nm_flag & NFSMNT_KERB) {
1184 		verf_str = nickv;
1185 		verf_len = sizeof (nickv);
1186 		auth_type = RPCAUTH_KERB4;
1187 		bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1188 		if (rep->r_failed_auth ||
1189 		    nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1190 				    verf_str, verf_len)) {
1191 			error = nfs_getauth(nmp, rep, cred, &auth_str,
1192 				&auth_len, verf_str, &verf_len, rep->r_key);
1193 			if (error) {
1194 				m_freem(rep->r_mrest);
1195 				rep->r_mrest = NULL;
1196 				kfree((caddr_t)rep, M_NFSREQ);
1197 				return (error);
1198 			}
1199 		}
1200 	} else {
1201 		auth_type = RPCAUTH_UNIX;
1202 		if (cred->cr_ngroups < 1)
1203 			panic("nfsreq nogrps");
1204 		auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1205 			nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1206 			5 * NFSX_UNSIGNED;
1207 	}
1208 	if (rep->r_mrest)
1209 		nfs_checkpkt(rep->r_mrest, rep->r_mrest_len);
1210 	m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1211 			auth_len, auth_str, verf_len, verf_str,
1212 			rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1213 	rep->r_mrest = NULL;
1214 	if (auth_str)
1215 		kfree(auth_str, M_TEMP);
1216 
1217 	/*
1218 	 * For stream protocols, insert a Sun RPC Record Mark.
1219 	 */
1220 	if (nmp->nm_sotype == SOCK_STREAM) {
1221 		M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1222 		if (m == NULL) {
1223 			kfree(rep, M_NFSREQ);
1224 			return (ENOBUFS);
1225 		}
1226 		*mtod(m, u_int32_t *) = htonl(0x80000000 |
1227 			 (m->m_pkthdr.len - NFSX_UNSIGNED));
1228 	}
1229 
1230 	nfs_checkpkt(m, m->m_pkthdr.len);
1231 
1232 	rep->r_mreq = m;
1233 	rep->r_xid = xid;
1234 	return (0);
1235 }
1236 
1237 static int
1238 nfs_request_try(struct nfsreq *rep)
1239 {
1240 	struct nfsmount *nmp = rep->r_nmp;
1241 	struct mbuf *m2;
1242 	int error;
1243 
1244 	/*
1245 	 * Request is not on any queue, only the owner has access to it
1246 	 * so it should not be locked by anyone atm.
1247 	 *
1248 	 * Interlock to prevent races.  While locked the only remote
1249 	 * action possible is for r_mrep to be set (once we enqueue it).
1250 	 */
1251 	if (rep->r_flags == 0xdeadc0de) {
1252 		print_backtrace(-1);
1253 		panic("flags nbad\n");
1254 	}
1255 	KKASSERT((rep->r_flags & (R_LOCKED | R_ONREQQ)) == 0);
1256 	if (nmp->nm_flag & NFSMNT_SOFT)
1257 		rep->r_retry = nmp->nm_retry;
1258 	else
1259 		rep->r_retry = NFS_MAXREXMIT + 1;	/* past clip limit */
1260 	rep->r_rtt = rep->r_rexmit = 0;
1261 	if (proct[rep->r_procnum] > 0)
1262 		rep->r_flags |= R_TIMING | R_LOCKED;
1263 	else
1264 		rep->r_flags |= R_LOCKED;
1265 	rep->r_mrep = NULL;
1266 
1267 	/*
1268 	 * Do the client side RPC.
1269 	 */
1270 	nfsstats.rpcrequests++;
1271 
1272 	if (nmp->nm_flag & NFSMNT_FORCE) {
1273 		rep->r_flags |= R_SOFTTERM;
1274 		rep->r_flags &= ~R_LOCKED;
1275 		return (0);
1276 	}
1277 
1278 	/*
1279 	 * Chain request into list of outstanding requests. Be sure
1280 	 * to put it LAST so timer finds oldest requests first.  Note
1281 	 * that our control of R_LOCKED prevents the request from
1282 	 * getting ripped out from under us or transmitted by the
1283 	 * timer code.
1284 	 *
1285 	 * For requests with info structures we must atomically set the
1286 	 * info's state because the structure could become invalid upon
1287 	 * return due to races (i.e., if async)
1288 	 */
1289 	crit_enter();
1290 	mtx_link_init(&rep->r_link);
1291 	TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1292 	rep->r_flags |= R_ONREQQ;
1293 	++nmp->nm_reqqlen;
1294 	if (rep->r_flags & R_ASYNC)
1295 		rep->r_info->state = NFSM_STATE_WAITREPLY;
1296 	crit_exit();
1297 
1298 	error = 0;
1299 
1300 	/*
1301 	 * Send if we can.  Congestion control is not handled here any more
1302 	 * becausing trying to defer the initial send based on the nfs_timer
1303 	 * requires having a very fast nfs_timer, which is silly.
1304 	 */
1305 	if (nmp->nm_so) {
1306 		if (nmp->nm_soflags & PR_CONNREQUIRED)
1307 			error = nfs_sndlock(nmp, rep);
1308 		if (error == 0) {
1309 			m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1310 			error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1311 			if (nmp->nm_soflags & PR_CONNREQUIRED)
1312 				nfs_sndunlock(nmp);
1313 			rep->r_flags &= ~R_NEEDSXMIT;
1314 			if ((rep->r_flags & R_SENT) == 0) {
1315 				rep->r_flags |= R_SENT;
1316 			}
1317 		} else {
1318 			rep->r_flags |= R_NEEDSXMIT;
1319 		}
1320 	} else {
1321 		rep->r_flags |= R_NEEDSXMIT;
1322 		rep->r_rtt = -1;
1323 	}
1324 	if (error == EPIPE)
1325 		error = 0;
1326 
1327 	/*
1328 	 * Release the lock.  The only remote action that may have occurred
1329 	 * would have been the setting of rep->r_mrep.  If this occured
1330 	 * and the request was async we have to move it to the reader
1331 	 * thread's queue for action.
1332 	 *
1333 	 * For async requests also make sure the reader is woken up so
1334 	 * it gets on the socket to read responses.
1335 	 */
1336 	crit_enter();
1337 	if (rep->r_flags & R_ASYNC) {
1338 		if (rep->r_mrep)
1339 			nfs_hardterm(rep, 1);
1340 		rep->r_flags &= ~R_LOCKED;
1341 		nfssvc_iod_reader_wakeup(nmp);
1342 	} else {
1343 		rep->r_flags &= ~R_LOCKED;
1344 	}
1345 	if (rep->r_flags & R_WANTED) {
1346 		rep->r_flags &= ~R_WANTED;
1347 		wakeup(rep);
1348 	}
1349 	crit_exit();
1350 	return (error);
1351 }
1352 
1353 /*
1354  * This code is only called for synchronous requests.  Completed synchronous
1355  * requests are left on reqq and we remove them before moving on to the
1356  * processing state.
1357  */
1358 static int
1359 nfs_request_waitreply(struct nfsreq *rep)
1360 {
1361 	struct nfsmount *nmp = rep->r_nmp;
1362 	int error;
1363 
1364 	KKASSERT((rep->r_flags & R_ASYNC) == 0);
1365 
1366 	/*
1367 	 * Wait until the request is finished.
1368 	 */
1369 	error = nfs_reply(nmp, rep);
1370 
1371 	/*
1372 	 * RPC done, unlink the request, but don't rip it out from under
1373 	 * the callout timer.
1374 	 *
1375 	 * Once unlinked no other receiver or the timer will have
1376 	 * visibility, so we do not have to set R_LOCKED.
1377 	 */
1378 	crit_enter();
1379 	while (rep->r_flags & R_LOCKED) {
1380 		rep->r_flags |= R_WANTED;
1381 		tsleep(rep, 0, "nfstrac", 0);
1382 	}
1383 	KKASSERT(rep->r_flags & R_ONREQQ);
1384 	TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1385 	rep->r_flags &= ~R_ONREQQ;
1386 	--nmp->nm_reqqlen;
1387 	if (TAILQ_FIRST(&nmp->nm_bioq) &&
1388 	    nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1389 		nfssvc_iod_writer_wakeup(nmp);
1390 	}
1391 	crit_exit();
1392 
1393 	/*
1394 	 * Decrement the outstanding request count.
1395 	 */
1396 	if (rep->r_flags & R_SENT) {
1397 		rep->r_flags &= ~R_SENT;
1398 	}
1399 	return (error);
1400 }
1401 
1402 /*
1403  * Process reply with error returned from nfs_requet_waitreply().
1404  *
1405  * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1406  * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1407  */
1408 static int
1409 nfs_request_processreply(nfsm_info_t info, int error)
1410 {
1411 	struct nfsreq *req = info->req;
1412 	struct nfsmount *nmp = req->r_nmp;
1413 	u_int32_t *tl;
1414 	int verf_type;
1415 	int i;
1416 
1417 	/*
1418 	 * If there was a successful reply and a tprintf msg.
1419 	 * tprintf a response.
1420 	 */
1421 	if (error == 0 && (req->r_flags & R_TPRINTFMSG)) {
1422 		nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1423 		    "is alive again");
1424 	}
1425 	info->mrep = req->r_mrep;
1426 	info->md = req->r_md;
1427 	info->dpos = req->r_dpos;
1428 	if (error) {
1429 		m_freem(req->r_mreq);
1430 		req->r_mreq = NULL;
1431 		kfree(req, M_NFSREQ);
1432 		info->req = NULL;
1433 		return (error);
1434 	}
1435 
1436 	/*
1437 	 * break down the rpc header and check if ok
1438 	 */
1439 	NULLOUT(tl = nfsm_dissect(info, 3 * NFSX_UNSIGNED));
1440 	if (*tl++ == rpc_msgdenied) {
1441 		if (*tl == rpc_mismatch) {
1442 			error = EOPNOTSUPP;
1443 		} else if ((nmp->nm_flag & NFSMNT_KERB) &&
1444 			   *tl++ == rpc_autherr) {
1445 			if (req->r_failed_auth == 0) {
1446 				req->r_failed_auth++;
1447 				req->r_mheadend->m_next = NULL;
1448 				m_freem(info->mrep);
1449 				info->mrep = NULL;
1450 				m_freem(req->r_mreq);
1451 				req->r_mreq = NULL;
1452 				return (ENEEDAUTH);
1453 			} else {
1454 				error = EAUTH;
1455 			}
1456 		} else {
1457 			error = EACCES;
1458 		}
1459 		m_freem(info->mrep);
1460 		info->mrep = NULL;
1461 		m_freem(req->r_mreq);
1462 		req->r_mreq = NULL;
1463 		kfree(req, M_NFSREQ);
1464 		info->req = NULL;
1465 		return (error);
1466 	}
1467 
1468 	/*
1469 	 * Grab any Kerberos verifier, otherwise just throw it away.
1470 	 */
1471 	verf_type = fxdr_unsigned(int, *tl++);
1472 	i = fxdr_unsigned(int32_t, *tl);
1473 	if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1474 		error = nfs_savenickauth(nmp, req->r_cred, i, req->r_key,
1475 					 &info->md, &info->dpos, info->mrep);
1476 		if (error)
1477 			goto nfsmout;
1478 	} else if (i > 0) {
1479 		ERROROUT(nfsm_adv(info, nfsm_rndup(i)));
1480 	}
1481 	NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1482 	/* 0 == ok */
1483 	if (*tl == 0) {
1484 		NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1485 		if (*tl != 0) {
1486 			error = fxdr_unsigned(int, *tl);
1487 
1488 			/*
1489 			 * Does anyone even implement this?  Just impose
1490 			 * a 1-second delay.
1491 			 */
1492 			if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1493 				error == NFSERR_TRYLATER) {
1494 				m_freem(info->mrep);
1495 				info->mrep = NULL;
1496 				error = 0;
1497 
1498 				tsleep((caddr_t)&lbolt, 0, "nqnfstry", 0);
1499 				return (EAGAIN);	/* goto tryagain */
1500 			}
1501 
1502 			/*
1503 			 * If the File Handle was stale, invalidate the
1504 			 * lookup cache, just in case.
1505 			 *
1506 			 * To avoid namecache<->vnode deadlocks we must
1507 			 * release the vnode lock if we hold it.
1508 			 */
1509 			if (error == ESTALE) {
1510 				struct vnode *vp = req->r_vp;
1511 				int ltype;
1512 
1513 				ltype = lockstatus(&vp->v_lock, curthread);
1514 				if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1515 					lockmgr(&vp->v_lock, LK_RELEASE);
1516 				cache_inval_vp(vp, CINV_CHILDREN);
1517 				if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1518 					lockmgr(&vp->v_lock, ltype);
1519 			}
1520 			if (nmp->nm_flag & NFSMNT_NFSV3) {
1521 				KKASSERT(*req->r_mrp == info->mrep);
1522 				KKASSERT(*req->r_mdp == info->md);
1523 				KKASSERT(*req->r_dposp == info->dpos);
1524 				error |= NFSERR_RETERR;
1525 			} else {
1526 				m_freem(info->mrep);
1527 				info->mrep = NULL;
1528 			}
1529 			m_freem(req->r_mreq);
1530 			req->r_mreq = NULL;
1531 			kfree(req, M_NFSREQ);
1532 			info->req = NULL;
1533 			return (error);
1534 		}
1535 
1536 		KKASSERT(*req->r_mrp == info->mrep);
1537 		KKASSERT(*req->r_mdp == info->md);
1538 		KKASSERT(*req->r_dposp == info->dpos);
1539 		m_freem(req->r_mreq);
1540 		req->r_mreq = NULL;
1541 		FREE(req, M_NFSREQ);
1542 		return (0);
1543 	}
1544 	m_freem(info->mrep);
1545 	info->mrep = NULL;
1546 	error = EPROTONOSUPPORT;
1547 nfsmout:
1548 	m_freem(req->r_mreq);
1549 	req->r_mreq = NULL;
1550 	kfree(req, M_NFSREQ);
1551 	info->req = NULL;
1552 	return (error);
1553 }
1554 
1555 #ifndef NFS_NOSERVER
1556 /*
1557  * Generate the rpc reply header
1558  * siz arg. is used to decide if adding a cluster is worthwhile
1559  */
1560 int
1561 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1562 	    int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1563 {
1564 	u_int32_t *tl;
1565 	struct nfsm_info info;
1566 
1567 	siz += RPC_REPLYSIZ;
1568 	info.mb = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1569 	info.mreq = info.mb;
1570 	info.mreq->m_pkthdr.len = 0;
1571 	/*
1572 	 * If this is not a cluster, try and leave leading space
1573 	 * for the lower level headers.
1574 	 */
1575 	if ((max_hdr + siz) < MINCLSIZE)
1576 		info.mreq->m_data += max_hdr;
1577 	tl = mtod(info.mreq, u_int32_t *);
1578 	info.mreq->m_len = 6 * NFSX_UNSIGNED;
1579 	info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1580 	*tl++ = txdr_unsigned(nd->nd_retxid);
1581 	*tl++ = rpc_reply;
1582 	if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1583 		*tl++ = rpc_msgdenied;
1584 		if (err & NFSERR_AUTHERR) {
1585 			*tl++ = rpc_autherr;
1586 			*tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1587 			info.mreq->m_len -= NFSX_UNSIGNED;
1588 			info.bpos -= NFSX_UNSIGNED;
1589 		} else {
1590 			*tl++ = rpc_mismatch;
1591 			*tl++ = txdr_unsigned(RPC_VER2);
1592 			*tl = txdr_unsigned(RPC_VER2);
1593 		}
1594 	} else {
1595 		*tl++ = rpc_msgaccepted;
1596 
1597 		/*
1598 		 * For Kerberos authentication, we must send the nickname
1599 		 * verifier back, otherwise just RPCAUTH_NULL.
1600 		 */
1601 		if (nd->nd_flag & ND_KERBFULL) {
1602 		    struct nfsuid *nuidp;
1603 		    struct timeval ktvin, ktvout;
1604 
1605 		    for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1606 			nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1607 			if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1608 			    (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1609 			     &nuidp->nu_haddr, nd->nd_nam2)))
1610 			    break;
1611 		    }
1612 		    if (nuidp) {
1613 			ktvin.tv_sec =
1614 			    txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1615 			ktvin.tv_usec =
1616 			    txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1617 
1618 			/*
1619 			 * Encrypt the timestamp in ecb mode using the
1620 			 * session key.
1621 			 */
1622 #ifdef NFSKERB
1623 			XXX
1624 #else
1625 			ktvout.tv_sec = 0;
1626 			ktvout.tv_usec = 0;
1627 #endif
1628 
1629 			*tl++ = rpc_auth_kerb;
1630 			*tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1631 			*tl = ktvout.tv_sec;
1632 			tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1633 			*tl++ = ktvout.tv_usec;
1634 			*tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1635 		    } else {
1636 			*tl++ = 0;
1637 			*tl++ = 0;
1638 		    }
1639 		} else {
1640 			*tl++ = 0;
1641 			*tl++ = 0;
1642 		}
1643 		switch (err) {
1644 		case EPROGUNAVAIL:
1645 			*tl = txdr_unsigned(RPC_PROGUNAVAIL);
1646 			break;
1647 		case EPROGMISMATCH:
1648 			*tl = txdr_unsigned(RPC_PROGMISMATCH);
1649 			tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1650 			*tl++ = txdr_unsigned(2);
1651 			*tl = txdr_unsigned(3);
1652 			break;
1653 		case EPROCUNAVAIL:
1654 			*tl = txdr_unsigned(RPC_PROCUNAVAIL);
1655 			break;
1656 		case EBADRPC:
1657 			*tl = txdr_unsigned(RPC_GARBAGE);
1658 			break;
1659 		default:
1660 			*tl = 0;
1661 			if (err != NFSERR_RETVOID) {
1662 				tl = nfsm_build(&info, NFSX_UNSIGNED);
1663 				if (err)
1664 				    *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1665 				else
1666 				    *tl = 0;
1667 			}
1668 			break;
1669 		};
1670 	}
1671 
1672 	if (mrq != NULL)
1673 	    *mrq = info.mreq;
1674 	*mbp = info.mb;
1675 	*bposp = info.bpos;
1676 	if (err != 0 && err != NFSERR_RETVOID)
1677 		nfsstats.srvrpc_errs++;
1678 	return (0);
1679 }
1680 
1681 
1682 #endif /* NFS_NOSERVER */
1683 
1684 /*
1685  * Nfs timer routine.
1686  *
1687  * Scan the nfsreq list and retranmit any requests that have timed out
1688  * To avoid retransmission attempts on STREAM sockets (in the future) make
1689  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1690  *
1691  * Requests with attached responses, terminated requests, and
1692  * locked requests are ignored.  Locked requests will be picked up
1693  * in a later timer call.
1694  */
1695 void
1696 nfs_timer(void *arg /* never used */)
1697 {
1698 	struct nfsmount *nmp;
1699 	struct nfsreq *req;
1700 #ifndef NFS_NOSERVER
1701 	struct nfssvc_sock *slp;
1702 	u_quad_t cur_usec;
1703 #endif /* NFS_NOSERVER */
1704 
1705 	crit_enter();
1706 	TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1707 		TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1708 			KKASSERT(nmp == req->r_nmp);
1709 			if (req->r_mrep)
1710 				continue;
1711 			if (req->r_flags & (R_SOFTTERM | R_LOCKED))
1712 				continue;
1713 			req->r_flags |= R_LOCKED;
1714 			if (nfs_sigintr(nmp, req, req->r_td)) {
1715 				nfs_softterm(req, 1);
1716 			} else {
1717 				nfs_timer_req(req);
1718 			}
1719 			req->r_flags &= ~R_LOCKED;
1720 			if (req->r_flags & R_WANTED) {
1721 				req->r_flags &= ~R_WANTED;
1722 				wakeup(req);
1723 			}
1724 		}
1725 	}
1726 #ifndef NFS_NOSERVER
1727 
1728 	/*
1729 	 * Scan the write gathering queues for writes that need to be
1730 	 * completed now.
1731 	 */
1732 	cur_usec = nfs_curusec();
1733 	TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1734 	    if (slp->ns_tq.lh_first && slp->ns_tq.lh_first->nd_time<=cur_usec)
1735 		nfsrv_wakenfsd(slp, 1);
1736 	}
1737 #endif /* NFS_NOSERVER */
1738 	crit_exit();
1739 	callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer, NULL);
1740 }
1741 
1742 static
1743 void
1744 nfs_timer_req(struct nfsreq *req)
1745 {
1746 	struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1747 	struct nfsmount *nmp = req->r_nmp;
1748 	struct mbuf *m;
1749 	struct socket *so;
1750 	int timeo;
1751 	int error;
1752 
1753 	/*
1754 	 * rtt ticks and timeout calculation.  Return if the timeout
1755 	 * has not been reached yet, unless the packet is flagged
1756 	 * for an immediate send.
1757 	 *
1758 	 * The mean rtt doesn't help when we get random I/Os, we have
1759 	 * to multiply by fairly large numbers.
1760 	 */
1761 	if (req->r_rtt >= 0) {
1762 		/*
1763 		 * Calculate the timeout to test against.
1764 		 */
1765 		req->r_rtt++;
1766 		if (nmp->nm_flag & NFSMNT_DUMBTIMR) {
1767 			timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1768 		} else if (req->r_flags & R_TIMING) {
1769 			timeo = NFS_SRTT(req) + NFS_SDRTT(req);
1770 		} else {
1771 			timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1772 		}
1773 		timeo *= multt[req->r_procnum];
1774 		/* timeo is still scaled by SCALE_BITS */
1775 
1776 #define NFSFS	(NFS_RTT_SCALE * NFS_HZ)
1777 		if (req->r_flags & R_TIMING) {
1778 			static long last_time;
1779 			if (nfs_showrtt && last_time != time_second) {
1780 				kprintf("rpccmd %d NFS SRTT %d SDRTT %d "
1781 					"timeo %d.%03d\n",
1782 					proct[req->r_procnum],
1783 					NFS_SRTT(req), NFS_SDRTT(req),
1784 					timeo / NFSFS,
1785 					timeo % NFSFS * 1000 /  NFSFS);
1786 				last_time = time_second;
1787 			}
1788 		}
1789 #undef NFSFS
1790 
1791 		/*
1792 		 * deal with nfs_timer jitter.
1793 		 */
1794 		timeo = (timeo >> NFS_RTT_SCALE_BITS) + 1;
1795 		if (timeo < 2)
1796 			timeo = 2;
1797 
1798 		if (nmp->nm_timeouts > 0)
1799 			timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1800 		if (timeo > NFS_MAXTIMEO)
1801 			timeo = NFS_MAXTIMEO;
1802 		if (req->r_rtt <= timeo) {
1803 			if ((req->r_flags & R_NEEDSXMIT) == 0)
1804 				return;
1805 		} else if (nmp->nm_timeouts < 8) {
1806 			nmp->nm_timeouts++;
1807 		}
1808 	}
1809 
1810 	/*
1811 	 * Check for server not responding
1812 	 */
1813 	if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1814 	     req->r_rexmit > nmp->nm_deadthresh) {
1815 		nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1816 			"not responding");
1817 		req->r_flags |= R_TPRINTFMSG;
1818 	}
1819 	if (req->r_rexmit >= req->r_retry) {	/* too many */
1820 		nfsstats.rpctimeouts++;
1821 		nfs_softterm(req, 1);
1822 		return;
1823 	}
1824 
1825 	/*
1826 	 * Generally disable retransmission on reliable sockets,
1827 	 * unless the request is flagged for immediate send.
1828 	 */
1829 	if (nmp->nm_sotype != SOCK_DGRAM) {
1830 		if (++req->r_rexmit > NFS_MAXREXMIT)
1831 			req->r_rexmit = NFS_MAXREXMIT;
1832 		if ((req->r_flags & R_NEEDSXMIT) == 0)
1833 			return;
1834 	}
1835 
1836 	/*
1837 	 * Stop here if we do not have a socket!
1838 	 */
1839 	if ((so = nmp->nm_so) == NULL)
1840 		return;
1841 
1842 	/*
1843 	 * If there is enough space and the window allows.. resend it.
1844 	 *
1845 	 * r_rtt is left intact in case we get an answer after the
1846 	 * retry that was a reply to the original packet.
1847 	 *
1848 	 * NOTE: so_pru_send()
1849 	 */
1850 	if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1851 	    (req->r_flags & (R_SENT | R_NEEDSXMIT)) &&
1852 	   (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1853 		if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1854 		    error = so_pru_send(so, 0, m, NULL, NULL, td);
1855 		else
1856 		    error = so_pru_send(so, 0, m, nmp->nm_nam, NULL, td);
1857 		if (error) {
1858 			if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1859 				so->so_error = 0;
1860 			req->r_flags |= R_NEEDSXMIT;
1861 		} else if (req->r_mrep == NULL) {
1862 			/*
1863 			 * Iff first send, start timing
1864 			 * else turn timing off, backoff timer
1865 			 * and divide congestion window by 2.
1866 			 *
1867 			 * It is possible for the so_pru_send() to
1868 			 * block and for us to race a reply so we
1869 			 * only do this if the reply field has not
1870 			 * been filled in.  R_LOCKED will prevent
1871 			 * the request from being ripped out from under
1872 			 * us entirely.
1873 			 *
1874 			 * Record the last resent procnum to aid us
1875 			 * in duplicate detection on receive.
1876 			 */
1877 			if ((req->r_flags & R_NEEDSXMIT) == 0) {
1878 				if (nfs_showrexmit)
1879 					kprintf("X");
1880 				if (++req->r_rexmit > NFS_MAXREXMIT)
1881 					req->r_rexmit = NFS_MAXREXMIT;
1882 				nmp->nm_maxasync_scaled >>= 1;
1883 				if (nmp->nm_maxasync_scaled < NFS_MINASYNC_SCALED)
1884 					nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
1885 				nfsstats.rpcretries++;
1886 				nmp->nm_lastreprocnum = req->r_procnum;
1887 			} else {
1888 				req->r_flags |= R_SENT;
1889 				req->r_flags &= ~R_NEEDSXMIT;
1890 			}
1891 		}
1892 	}
1893 }
1894 
1895 /*
1896  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1897  * wait for all requests to complete. This is used by forced unmounts
1898  * to terminate any outstanding RPCs.
1899  *
1900  * Locked requests cannot be canceled but will be marked for
1901  * soft-termination.
1902  */
1903 int
1904 nfs_nmcancelreqs(struct nfsmount *nmp)
1905 {
1906 	struct nfsreq *req;
1907 	int i;
1908 
1909 	crit_enter();
1910 	TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1911 		if (req->r_mrep != NULL || (req->r_flags & R_SOFTTERM))
1912 			continue;
1913 		nfs_softterm(req, 0);
1914 	}
1915 	/* XXX  the other two queues as well */
1916 	crit_exit();
1917 
1918 	for (i = 0; i < 30; i++) {
1919 		crit_enter();
1920 		TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1921 			if (nmp == req->r_nmp)
1922 				break;
1923 		}
1924 		crit_exit();
1925 		if (req == NULL)
1926 			return (0);
1927 		tsleep(&lbolt, 0, "nfscancel", 0);
1928 	}
1929 	return (EBUSY);
1930 }
1931 
1932 /*
1933  * Soft-terminate a request, effectively marking it as failed.
1934  *
1935  * Must be called from within a critical section.
1936  */
1937 static void
1938 nfs_softterm(struct nfsreq *rep, int islocked)
1939 {
1940 	rep->r_flags |= R_SOFTTERM;
1941 	nfs_hardterm(rep, islocked);
1942 }
1943 
1944 /*
1945  * Hard-terminate a request, typically after getting a response.
1946  *
1947  * The state machine can still decide to re-issue it later if necessary.
1948  *
1949  * Must be called from within a critical section.
1950  */
1951 static void
1952 nfs_hardterm(struct nfsreq *rep, int islocked)
1953 {
1954 	struct nfsmount *nmp = rep->r_nmp;
1955 
1956 	/*
1957 	 * The nm_send count is decremented now to avoid deadlocks
1958 	 * when the process in soreceive() hasn't yet managed to send
1959 	 * its own request.
1960 	 */
1961 	if (rep->r_flags & R_SENT) {
1962 		rep->r_flags &= ~R_SENT;
1963 	}
1964 
1965 	/*
1966 	 * If we locked the request or nobody else has locked the request,
1967 	 * and the request is async, we can move it to the reader thread's
1968 	 * queue now and fix up the state.
1969 	 *
1970 	 * If we locked the request or nobody else has locked the request,
1971 	 * we can wake up anyone blocked waiting for a response on the
1972 	 * request.
1973 	 */
1974 	if (islocked || (rep->r_flags & R_LOCKED) == 0) {
1975 		if ((rep->r_flags & (R_ONREQQ | R_ASYNC)) ==
1976 		    (R_ONREQQ | R_ASYNC)) {
1977 			rep->r_flags &= ~R_ONREQQ;
1978 			TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1979 			--nmp->nm_reqqlen;
1980 			TAILQ_INSERT_TAIL(&nmp->nm_reqrxq, rep, r_chain);
1981 			KKASSERT(rep->r_info->state == NFSM_STATE_TRY ||
1982 				 rep->r_info->state == NFSM_STATE_WAITREPLY);
1983 			rep->r_info->state = NFSM_STATE_PROCESSREPLY;
1984 			nfssvc_iod_reader_wakeup(nmp);
1985 			if (TAILQ_FIRST(&nmp->nm_bioq) &&
1986 			    nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1987 				nfssvc_iod_writer_wakeup(nmp);
1988 			}
1989 		}
1990 		mtx_abort_ex_link(&nmp->nm_rxlock, &rep->r_link);
1991 	}
1992 }
1993 
1994 /*
1995  * Test for a termination condition pending on the process.
1996  * This is used for NFSMNT_INT mounts.
1997  */
1998 int
1999 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
2000 {
2001 	sigset_t tmpset;
2002 	struct proc *p;
2003 	struct lwp *lp;
2004 
2005 	if (rep && (rep->r_flags & R_SOFTTERM))
2006 		return (EINTR);
2007 	/* Terminate all requests while attempting a forced unmount. */
2008 	if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
2009 		return (EINTR);
2010 	if (!(nmp->nm_flag & NFSMNT_INT))
2011 		return (0);
2012 	/* td might be NULL YYY */
2013 	if (td == NULL || (p = td->td_proc) == NULL)
2014 		return (0);
2015 
2016 	lp = td->td_lwp;
2017 	tmpset = lwp_sigpend(lp);
2018 	SIGSETNAND(tmpset, lp->lwp_sigmask);
2019 	SIGSETNAND(tmpset, p->p_sigignore);
2020 	if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
2021 		return (EINTR);
2022 
2023 	return (0);
2024 }
2025 
2026 /*
2027  * Lock a socket against others.
2028  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
2029  * and also to avoid race conditions between the processes with nfs requests
2030  * in progress when a reconnect is necessary.
2031  */
2032 int
2033 nfs_sndlock(struct nfsmount *nmp, struct nfsreq *rep)
2034 {
2035 	mtx_t mtx = &nmp->nm_txlock;
2036 	struct thread *td;
2037 	int slptimeo;
2038 	int slpflag;
2039 	int error;
2040 
2041 	slpflag = 0;
2042 	slptimeo = 0;
2043 	td = rep ? rep->r_td : NULL;
2044 	if (nmp->nm_flag & NFSMNT_INT)
2045 		slpflag = PCATCH;
2046 
2047 	while ((error = mtx_lock_ex_try(mtx)) != 0) {
2048 		if (nfs_sigintr(nmp, rep, td)) {
2049 			error = EINTR;
2050 			break;
2051 		}
2052 		error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
2053 		if (error == 0)
2054 			break;
2055 		if (slpflag == PCATCH) {
2056 			slpflag = 0;
2057 			slptimeo = 2 * hz;
2058 		}
2059 	}
2060 	/* Always fail if our request has been cancelled. */
2061 	if (rep && (rep->r_flags & R_SOFTTERM)) {
2062 		if (error == 0)
2063 			mtx_unlock(mtx);
2064 		error = EINTR;
2065 	}
2066 	return (error);
2067 }
2068 
2069 /*
2070  * Unlock the stream socket for others.
2071  */
2072 void
2073 nfs_sndunlock(struct nfsmount *nmp)
2074 {
2075 	mtx_unlock(&nmp->nm_txlock);
2076 }
2077 
2078 /*
2079  * Lock the receiver side of the socket.
2080  *
2081  * rep may be NULL.
2082  */
2083 static int
2084 nfs_rcvlock(struct nfsmount *nmp, struct nfsreq *rep)
2085 {
2086 	mtx_t mtx = &nmp->nm_rxlock;
2087 	int slpflag;
2088 	int slptimeo;
2089 	int error;
2090 
2091 	/*
2092 	 * Unconditionally check for completion in case another nfsiod
2093 	 * get the packet while the caller was blocked, before the caller
2094 	 * called us.  Packet reception is handled by mainline code which
2095 	 * is protected by the BGL at the moment.
2096 	 *
2097 	 * We do not strictly need the second check just before the
2098 	 * tsleep(), but it's good defensive programming.
2099 	 */
2100 	if (rep && rep->r_mrep != NULL)
2101 		return (EALREADY);
2102 
2103 	if (nmp->nm_flag & NFSMNT_INT)
2104 		slpflag = PCATCH;
2105 	else
2106 		slpflag = 0;
2107 	slptimeo = 0;
2108 
2109 	while ((error = mtx_lock_ex_try(mtx)) != 0) {
2110 		if (nfs_sigintr(nmp, rep, (rep ? rep->r_td : NULL))) {
2111 			error = EINTR;
2112 			break;
2113 		}
2114 		if (rep && rep->r_mrep != NULL) {
2115 			error = EALREADY;
2116 			break;
2117 		}
2118 
2119 		/*
2120 		 * NOTE: can return ENOLCK, but in that case rep->r_mrep
2121 		 *       will already be set.
2122 		 */
2123 		if (rep) {
2124 			error = mtx_lock_ex_link(mtx, &rep->r_link,
2125 						 "nfsrcvlk",
2126 						 slpflag, slptimeo);
2127 		} else {
2128 			error = mtx_lock_ex(mtx, "nfsrcvlk", slpflag, slptimeo);
2129 		}
2130 		if (error == 0)
2131 			break;
2132 
2133 		/*
2134 		 * If our reply was recieved while we were sleeping,
2135 		 * then just return without taking the lock to avoid a
2136 		 * situation where a single iod could 'capture' the
2137 		 * recieve lock.
2138 		 */
2139 		if (rep && rep->r_mrep != NULL) {
2140 			error = EALREADY;
2141 			break;
2142 		}
2143 		if (slpflag == PCATCH) {
2144 			slpflag = 0;
2145 			slptimeo = 2 * hz;
2146 		}
2147 	}
2148 	if (error == 0) {
2149 		if (rep && rep->r_mrep != NULL) {
2150 			error = EALREADY;
2151 			mtx_unlock(mtx);
2152 		}
2153 	}
2154 	return (error);
2155 }
2156 
2157 /*
2158  * Unlock the stream socket for others.
2159  */
2160 static void
2161 nfs_rcvunlock(struct nfsmount *nmp)
2162 {
2163 	mtx_unlock(&nmp->nm_rxlock);
2164 }
2165 
2166 /*
2167  * nfs_realign:
2168  *
2169  * Check for badly aligned mbuf data and realign by copying the unaligned
2170  * portion of the data into a new mbuf chain and freeing the portions
2171  * of the old chain that were replaced.
2172  *
2173  * We cannot simply realign the data within the existing mbuf chain
2174  * because the underlying buffers may contain other rpc commands and
2175  * we cannot afford to overwrite them.
2176  *
2177  * We would prefer to avoid this situation entirely.  The situation does
2178  * not occur with NFS/UDP and is supposed to only occassionally occur
2179  * with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
2180  *
2181  * NOTE!  MB_DONTWAIT cannot be used here.  The mbufs must be acquired
2182  *	  because the rpc request OR reply cannot be thrown away.  TCP NFS
2183  *	  mounts do not retry their RPCs unless the TCP connection itself
2184  *	  is dropped so throwing away a RPC will basically cause the NFS
2185  *	  operation to lockup indefinitely.
2186  */
2187 static void
2188 nfs_realign(struct mbuf **pm, int hsiz)
2189 {
2190 	struct mbuf *m;
2191 	struct mbuf *n = NULL;
2192 
2193 	/*
2194 	 * Check for misalignemnt
2195 	 */
2196 	++nfs_realign_test;
2197 	while ((m = *pm) != NULL) {
2198 		if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3))
2199 			break;
2200 		pm = &m->m_next;
2201 	}
2202 
2203 	/*
2204 	 * If misalignment found make a completely new copy.
2205 	 */
2206 	if (m) {
2207 		++nfs_realign_count;
2208 		n = m_dup_data(m, MB_WAIT);
2209 		m_freem(*pm);
2210 		*pm = n;
2211 	}
2212 }
2213 
2214 #ifndef NFS_NOSERVER
2215 
2216 /*
2217  * Parse an RPC request
2218  * - verify it
2219  * - fill in the cred struct.
2220  */
2221 int
2222 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
2223 {
2224 	int len, i;
2225 	u_int32_t *tl;
2226 	struct uio uio;
2227 	struct iovec iov;
2228 	caddr_t cp;
2229 	u_int32_t nfsvers, auth_type;
2230 	uid_t nickuid;
2231 	int error = 0, ticklen;
2232 	struct nfsuid *nuidp;
2233 	struct timeval tvin, tvout;
2234 	struct nfsm_info info;
2235 #if 0				/* until encrypted keys are implemented */
2236 	NFSKERBKEYSCHED_T keys;	/* stores key schedule */
2237 #endif
2238 
2239 	info.mrep = nd->nd_mrep;
2240 	info.md = nd->nd_md;
2241 	info.dpos = nd->nd_dpos;
2242 
2243 	if (has_header) {
2244 		NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
2245 		nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
2246 		if (*tl++ != rpc_call) {
2247 			m_freem(info.mrep);
2248 			return (EBADRPC);
2249 		}
2250 	} else {
2251 		NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
2252 	}
2253 	nd->nd_repstat = 0;
2254 	nd->nd_flag = 0;
2255 	if (*tl++ != rpc_vers) {
2256 		nd->nd_repstat = ERPCMISMATCH;
2257 		nd->nd_procnum = NFSPROC_NOOP;
2258 		return (0);
2259 	}
2260 	if (*tl != nfs_prog) {
2261 		nd->nd_repstat = EPROGUNAVAIL;
2262 		nd->nd_procnum = NFSPROC_NOOP;
2263 		return (0);
2264 	}
2265 	tl++;
2266 	nfsvers = fxdr_unsigned(u_int32_t, *tl++);
2267 	if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
2268 		nd->nd_repstat = EPROGMISMATCH;
2269 		nd->nd_procnum = NFSPROC_NOOP;
2270 		return (0);
2271 	}
2272 	if (nfsvers == NFS_VER3)
2273 		nd->nd_flag = ND_NFSV3;
2274 	nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
2275 	if (nd->nd_procnum == NFSPROC_NULL)
2276 		return (0);
2277 	if (nd->nd_procnum >= NFS_NPROCS ||
2278 		(nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
2279 		(!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
2280 		nd->nd_repstat = EPROCUNAVAIL;
2281 		nd->nd_procnum = NFSPROC_NOOP;
2282 		return (0);
2283 	}
2284 	if ((nd->nd_flag & ND_NFSV3) == 0)
2285 		nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
2286 	auth_type = *tl++;
2287 	len = fxdr_unsigned(int, *tl++);
2288 	if (len < 0 || len > RPCAUTH_MAXSIZ) {
2289 		m_freem(info.mrep);
2290 		return (EBADRPC);
2291 	}
2292 
2293 	nd->nd_flag &= ~ND_KERBAUTH;
2294 	/*
2295 	 * Handle auth_unix or auth_kerb.
2296 	 */
2297 	if (auth_type == rpc_auth_unix) {
2298 		len = fxdr_unsigned(int, *++tl);
2299 		if (len < 0 || len > NFS_MAXNAMLEN) {
2300 			m_freem(info.mrep);
2301 			return (EBADRPC);
2302 		}
2303 		ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2304 		NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2305 		bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
2306 		nd->nd_cr.cr_ref = 1;
2307 		nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
2308 		nd->nd_cr.cr_ruid = nd->nd_cr.cr_svuid = nd->nd_cr.cr_uid;
2309 		nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
2310 		nd->nd_cr.cr_rgid = nd->nd_cr.cr_svgid = nd->nd_cr.cr_gid;
2311 		len = fxdr_unsigned(int, *tl);
2312 		if (len < 0 || len > RPCAUTH_UNIXGIDS) {
2313 			m_freem(info.mrep);
2314 			return (EBADRPC);
2315 		}
2316 		NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
2317 		for (i = 1; i <= len; i++)
2318 		    if (i < NGROUPS)
2319 			nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
2320 		    else
2321 			tl++;
2322 		nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
2323 		if (nd->nd_cr.cr_ngroups > 1)
2324 		    nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
2325 		len = fxdr_unsigned(int, *++tl);
2326 		if (len < 0 || len > RPCAUTH_MAXSIZ) {
2327 			m_freem(info.mrep);
2328 			return (EBADRPC);
2329 		}
2330 		if (len > 0) {
2331 			ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2332 		}
2333 	} else if (auth_type == rpc_auth_kerb) {
2334 		switch (fxdr_unsigned(int, *tl++)) {
2335 		case RPCAKN_FULLNAME:
2336 			ticklen = fxdr_unsigned(int, *tl);
2337 			*((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2338 			uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2339 			nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2340 			if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2341 				m_freem(info.mrep);
2342 				return (EBADRPC);
2343 			}
2344 			uio.uio_offset = 0;
2345 			uio.uio_iov = &iov;
2346 			uio.uio_iovcnt = 1;
2347 			uio.uio_segflg = UIO_SYSSPACE;
2348 			iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2349 			iov.iov_len = RPCAUTH_MAXSIZ - 4;
2350 			ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2351 			NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2352 			if (*tl++ != rpc_auth_kerb ||
2353 				fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2354 				kprintf("Bad kerb verifier\n");
2355 				nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2356 				nd->nd_procnum = NFSPROC_NOOP;
2357 				return (0);
2358 			}
2359 			NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2360 			tl = (u_int32_t *)cp;
2361 			if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2362 				kprintf("Not fullname kerb verifier\n");
2363 				nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2364 				nd->nd_procnum = NFSPROC_NOOP;
2365 				return (0);
2366 			}
2367 			cp += NFSX_UNSIGNED;
2368 			bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2369 			nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2370 			nd->nd_flag |= ND_KERBFULL;
2371 			nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2372 			break;
2373 		case RPCAKN_NICKNAME:
2374 			if (len != 2 * NFSX_UNSIGNED) {
2375 				kprintf("Kerb nickname short\n");
2376 				nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2377 				nd->nd_procnum = NFSPROC_NOOP;
2378 				return (0);
2379 			}
2380 			nickuid = fxdr_unsigned(uid_t, *tl);
2381 			NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2382 			if (*tl++ != rpc_auth_kerb ||
2383 				fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2384 				kprintf("Kerb nick verifier bad\n");
2385 				nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2386 				nd->nd_procnum = NFSPROC_NOOP;
2387 				return (0);
2388 			}
2389 			NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2390 			tvin.tv_sec = *tl++;
2391 			tvin.tv_usec = *tl;
2392 
2393 			for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2394 			    nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2395 				if (nuidp->nu_cr.cr_uid == nickuid &&
2396 				    (!nd->nd_nam2 ||
2397 				     netaddr_match(NU_NETFAM(nuidp),
2398 				      &nuidp->nu_haddr, nd->nd_nam2)))
2399 					break;
2400 			}
2401 			if (!nuidp) {
2402 				nd->nd_repstat =
2403 					(NFSERR_AUTHERR|AUTH_REJECTCRED);
2404 				nd->nd_procnum = NFSPROC_NOOP;
2405 				return (0);
2406 			}
2407 
2408 			/*
2409 			 * Now, decrypt the timestamp using the session key
2410 			 * and validate it.
2411 			 */
2412 #ifdef NFSKERB
2413 			XXX
2414 #else
2415 			tvout.tv_sec = 0;
2416 			tvout.tv_usec = 0;
2417 #endif
2418 
2419 			tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2420 			tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2421 			if (nuidp->nu_expire < time_second ||
2422 			    nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2423 			    (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2424 			     nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2425 				nuidp->nu_expire = 0;
2426 				nd->nd_repstat =
2427 				    (NFSERR_AUTHERR|AUTH_REJECTVERF);
2428 				nd->nd_procnum = NFSPROC_NOOP;
2429 				return (0);
2430 			}
2431 			nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2432 			nd->nd_flag |= ND_KERBNICK;
2433 		};
2434 	} else {
2435 		nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2436 		nd->nd_procnum = NFSPROC_NOOP;
2437 		return (0);
2438 	}
2439 
2440 	nd->nd_md = info.md;
2441 	nd->nd_dpos = info.dpos;
2442 	return (0);
2443 nfsmout:
2444 	return (error);
2445 }
2446 
2447 #endif
2448 
2449 /*
2450  * Send a message to the originating process's terminal.  The thread and/or
2451  * process may be NULL.  YYY the thread should not be NULL but there may
2452  * still be some uio_td's that are still being passed as NULL through to
2453  * nfsm_request().
2454  */
2455 static int
2456 nfs_msg(struct thread *td, char *server, char *msg)
2457 {
2458 	tpr_t tpr;
2459 
2460 	if (td && td->td_proc)
2461 		tpr = tprintf_open(td->td_proc);
2462 	else
2463 		tpr = NULL;
2464 	tprintf(tpr, "nfs server %s: %s\n", server, msg);
2465 	tprintf_close(tpr);
2466 	return (0);
2467 }
2468 
2469 #ifndef NFS_NOSERVER
2470 /*
2471  * Socket upcall routine for the nfsd sockets.
2472  * The caddr_t arg is a pointer to the "struct nfssvc_sock".
2473  * Essentially do as much as possible non-blocking, else punt and it will
2474  * be called with MB_WAIT from an nfsd.
2475  */
2476 void
2477 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2478 {
2479 	struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2480 	struct mbuf *m;
2481 	struct sockaddr *nam;
2482 	struct sockbuf sio;
2483 	int flags, error;
2484 	int nparallel_wakeup = 0;
2485 
2486 	if ((slp->ns_flag & SLP_VALID) == 0)
2487 		return;
2488 
2489 	/*
2490 	 * Do not allow an infinite number of completed RPC records to build
2491 	 * up before we stop reading data from the socket.  Otherwise we could
2492 	 * end up holding onto an unreasonable number of mbufs for requests
2493 	 * waiting for service.
2494 	 *
2495 	 * This should give pretty good feedback to the TCP
2496 	 * layer and prevents a memory crunch for other protocols.
2497 	 *
2498 	 * Note that the same service socket can be dispatched to several
2499 	 * nfs servers simultaniously.
2500 	 *
2501 	 * the tcp protocol callback calls us with MB_DONTWAIT.
2502 	 * nfsd calls us with MB_WAIT (typically).
2503 	 */
2504 	if (waitflag == MB_DONTWAIT && slp->ns_numrec >= nfsd_waiting / 2 + 1) {
2505 		slp->ns_flag |= SLP_NEEDQ;
2506 		goto dorecs;
2507 	}
2508 
2509 	/*
2510 	 * Handle protocol specifics to parse an RPC request.  We always
2511 	 * pull from the socket using non-blocking I/O.
2512 	 */
2513 	if (so->so_type == SOCK_STREAM) {
2514 		/*
2515 		 * The data has to be read in an orderly fashion from a TCP
2516 		 * stream, unlike a UDP socket.  It is possible for soreceive
2517 		 * and/or nfsrv_getstream() to block, so make sure only one
2518 		 * entity is messing around with the TCP stream at any given
2519 		 * moment.  The receive sockbuf's lock in soreceive is not
2520 		 * sufficient.
2521 		 *
2522 		 * Note that this procedure can be called from any number of
2523 		 * NFS severs *OR* can be upcalled directly from a TCP
2524 		 * protocol thread.
2525 		 */
2526 		if (slp->ns_flag & SLP_GETSTREAM) {
2527 			slp->ns_flag |= SLP_NEEDQ;
2528 			goto dorecs;
2529 		}
2530 		slp->ns_flag |= SLP_GETSTREAM;
2531 
2532 		/*
2533 		 * Do soreceive().  Pull out as much data as possible without
2534 		 * blocking.
2535 		 */
2536 		sbinit(&sio, 1000000000);
2537 		flags = MSG_DONTWAIT;
2538 		error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2539 		if (error || sio.sb_mb == NULL) {
2540 			if (error == EWOULDBLOCK)
2541 				slp->ns_flag |= SLP_NEEDQ;
2542 			else
2543 				slp->ns_flag |= SLP_DISCONN;
2544 			slp->ns_flag &= ~SLP_GETSTREAM;
2545 			goto dorecs;
2546 		}
2547 		m = sio.sb_mb;
2548 		if (slp->ns_rawend) {
2549 			slp->ns_rawend->m_next = m;
2550 			slp->ns_cc += sio.sb_cc;
2551 		} else {
2552 			slp->ns_raw = m;
2553 			slp->ns_cc = sio.sb_cc;
2554 		}
2555 		while (m->m_next)
2556 			m = m->m_next;
2557 		slp->ns_rawend = m;
2558 
2559 		/*
2560 		 * Now try and parse as many record(s) as we can out of the
2561 		 * raw stream data.
2562 		 */
2563 		error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2564 		if (error) {
2565 			if (error == EPERM)
2566 				slp->ns_flag |= SLP_DISCONN;
2567 			else
2568 				slp->ns_flag |= SLP_NEEDQ;
2569 		}
2570 		slp->ns_flag &= ~SLP_GETSTREAM;
2571 	} else {
2572 		/*
2573 		 * For UDP soreceive typically pulls just one packet, loop
2574 		 * to get the whole batch.
2575 		 */
2576 		do {
2577 			sbinit(&sio, 1000000000);
2578 			flags = MSG_DONTWAIT;
2579 			error = so_pru_soreceive(so, &nam, NULL, &sio,
2580 						 NULL, &flags);
2581 			if (sio.sb_mb) {
2582 				struct nfsrv_rec *rec;
2583 				int mf = (waitflag & MB_DONTWAIT) ?
2584 					    M_NOWAIT : M_WAITOK;
2585 				rec = kmalloc(sizeof(struct nfsrv_rec),
2586 					     M_NFSRVDESC, mf);
2587 				if (!rec) {
2588 					if (nam)
2589 						FREE(nam, M_SONAME);
2590 					m_freem(sio.sb_mb);
2591 					continue;
2592 				}
2593 				nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2594 				rec->nr_address = nam;
2595 				rec->nr_packet = sio.sb_mb;
2596 				STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2597 				++slp->ns_numrec;
2598 				++nparallel_wakeup;
2599 			}
2600 			if (error) {
2601 				if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2602 					&& error != EWOULDBLOCK) {
2603 					slp->ns_flag |= SLP_DISCONN;
2604 					goto dorecs;
2605 				}
2606 			}
2607 		} while (sio.sb_mb);
2608 	}
2609 
2610 	/*
2611 	 * If we were upcalled from the tcp protocol layer and we have
2612 	 * fully parsed records ready to go, or there is new data pending,
2613 	 * or something went wrong, try to wake up an nfsd thread to deal
2614 	 * with it.
2615 	 */
2616 dorecs:
2617 	if (waitflag == MB_DONTWAIT && (slp->ns_numrec > 0
2618 	     || (slp->ns_flag & (SLP_NEEDQ | SLP_DISCONN)))) {
2619 		nfsrv_wakenfsd(slp, nparallel_wakeup);
2620 	}
2621 }
2622 
2623 /*
2624  * Try and extract an RPC request from the mbuf data list received on a
2625  * stream socket. The "waitflag" argument indicates whether or not it
2626  * can sleep.
2627  */
2628 static int
2629 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2630 {
2631 	struct mbuf *m, **mpp;
2632 	char *cp1, *cp2;
2633 	int len;
2634 	struct mbuf *om, *m2, *recm;
2635 	u_int32_t recmark;
2636 
2637 	for (;;) {
2638 	    if (slp->ns_reclen == 0) {
2639 		if (slp->ns_cc < NFSX_UNSIGNED)
2640 			return (0);
2641 		m = slp->ns_raw;
2642 		if (m->m_len >= NFSX_UNSIGNED) {
2643 			bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2644 			m->m_data += NFSX_UNSIGNED;
2645 			m->m_len -= NFSX_UNSIGNED;
2646 		} else {
2647 			cp1 = (caddr_t)&recmark;
2648 			cp2 = mtod(m, caddr_t);
2649 			while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2650 				while (m->m_len == 0) {
2651 					m = m->m_next;
2652 					cp2 = mtod(m, caddr_t);
2653 				}
2654 				*cp1++ = *cp2++;
2655 				m->m_data++;
2656 				m->m_len--;
2657 			}
2658 		}
2659 		slp->ns_cc -= NFSX_UNSIGNED;
2660 		recmark = ntohl(recmark);
2661 		slp->ns_reclen = recmark & ~0x80000000;
2662 		if (recmark & 0x80000000)
2663 			slp->ns_flag |= SLP_LASTFRAG;
2664 		else
2665 			slp->ns_flag &= ~SLP_LASTFRAG;
2666 		if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2667 			log(LOG_ERR, "%s (%d) from nfs client\n",
2668 			    "impossible packet length",
2669 			    slp->ns_reclen);
2670 			return (EPERM);
2671 		}
2672 	    }
2673 
2674 	    /*
2675 	     * Now get the record part.
2676 	     *
2677 	     * Note that slp->ns_reclen may be 0.  Linux sometimes
2678 	     * generates 0-length RPCs
2679 	     */
2680 	    recm = NULL;
2681 	    if (slp->ns_cc == slp->ns_reclen) {
2682 		recm = slp->ns_raw;
2683 		slp->ns_raw = slp->ns_rawend = NULL;
2684 		slp->ns_cc = slp->ns_reclen = 0;
2685 	    } else if (slp->ns_cc > slp->ns_reclen) {
2686 		len = 0;
2687 		m = slp->ns_raw;
2688 		om = NULL;
2689 
2690 		while (len < slp->ns_reclen) {
2691 			if ((len + m->m_len) > slp->ns_reclen) {
2692 				m2 = m_copym(m, 0, slp->ns_reclen - len,
2693 					waitflag);
2694 				if (m2) {
2695 					if (om) {
2696 						om->m_next = m2;
2697 						recm = slp->ns_raw;
2698 					} else
2699 						recm = m2;
2700 					m->m_data += slp->ns_reclen - len;
2701 					m->m_len -= slp->ns_reclen - len;
2702 					len = slp->ns_reclen;
2703 				} else {
2704 					return (EWOULDBLOCK);
2705 				}
2706 			} else if ((len + m->m_len) == slp->ns_reclen) {
2707 				om = m;
2708 				len += m->m_len;
2709 				m = m->m_next;
2710 				recm = slp->ns_raw;
2711 				om->m_next = NULL;
2712 			} else {
2713 				om = m;
2714 				len += m->m_len;
2715 				m = m->m_next;
2716 			}
2717 		}
2718 		slp->ns_raw = m;
2719 		slp->ns_cc -= len;
2720 		slp->ns_reclen = 0;
2721 	    } else {
2722 		return (0);
2723 	    }
2724 
2725 	    /*
2726 	     * Accumulate the fragments into a record.
2727 	     */
2728 	    mpp = &slp->ns_frag;
2729 	    while (*mpp)
2730 		mpp = &((*mpp)->m_next);
2731 	    *mpp = recm;
2732 	    if (slp->ns_flag & SLP_LASTFRAG) {
2733 		struct nfsrv_rec *rec;
2734 		int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2735 		rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2736 		if (!rec) {
2737 		    m_freem(slp->ns_frag);
2738 		} else {
2739 		    nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2740 		    rec->nr_address = NULL;
2741 		    rec->nr_packet = slp->ns_frag;
2742 		    STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2743 		    ++slp->ns_numrec;
2744 		    ++*countp;
2745 		}
2746 		slp->ns_frag = NULL;
2747 	    }
2748 	}
2749 }
2750 
2751 #ifdef INVARIANTS
2752 
2753 /*
2754  * Sanity check our mbuf chain.
2755  */
2756 static void
2757 nfs_checkpkt(struct mbuf *m, int len)
2758 {
2759 	int xlen = 0;
2760 	while (m) {
2761 		xlen += m->m_len;
2762 		m = m->m_next;
2763 	}
2764 	if (xlen != len) {
2765 		panic("nfs_checkpkt: len mismatch %d/%d mbuf %p\n",
2766 			xlen, len, m);
2767 	}
2768 }
2769 
2770 #else
2771 
2772 static void
2773 nfs_checkpkt(struct mbuf *m __unused, int len __unused)
2774 {
2775 }
2776 
2777 #endif
2778 
2779 /*
2780  * Parse an RPC header.
2781  */
2782 int
2783 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2784 	    struct nfsrv_descript **ndp)
2785 {
2786 	struct nfsrv_rec *rec;
2787 	struct mbuf *m;
2788 	struct sockaddr *nam;
2789 	struct nfsrv_descript *nd;
2790 	int error;
2791 
2792 	*ndp = NULL;
2793 	if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2794 		return (ENOBUFS);
2795 	rec = STAILQ_FIRST(&slp->ns_rec);
2796 	STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2797 	KKASSERT(slp->ns_numrec > 0);
2798 	--slp->ns_numrec;
2799 	nam = rec->nr_address;
2800 	m = rec->nr_packet;
2801 	kfree(rec, M_NFSRVDESC);
2802 	MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2803 		M_NFSRVDESC, M_WAITOK);
2804 	nd->nd_md = nd->nd_mrep = m;
2805 	nd->nd_nam2 = nam;
2806 	nd->nd_dpos = mtod(m, caddr_t);
2807 	error = nfs_getreq(nd, nfsd, TRUE);
2808 	if (error) {
2809 		if (nam) {
2810 			FREE(nam, M_SONAME);
2811 		}
2812 		kfree((caddr_t)nd, M_NFSRVDESC);
2813 		return (error);
2814 	}
2815 	*ndp = nd;
2816 	nfsd->nfsd_nd = nd;
2817 	return (0);
2818 }
2819 
2820 /*
2821  * Try to assign service sockets to nfsd threads based on the number
2822  * of new rpc requests that have been queued on the service socket.
2823  *
2824  * If no nfsd's are available or additonal requests are pending, set the
2825  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2826  * the work in the nfssvc_sock list when it is finished processing its
2827  * current work.  This flag is only cleared when an nfsd can not find
2828  * any new work to perform.
2829  */
2830 void
2831 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2832 {
2833 	struct nfsd *nd;
2834 
2835 	if ((slp->ns_flag & SLP_VALID) == 0)
2836 		return;
2837 	if (nparallel <= 1)
2838 		nparallel = 1;
2839 	TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2840 		if (nd->nfsd_flag & NFSD_WAITING) {
2841 			nd->nfsd_flag &= ~NFSD_WAITING;
2842 			if (nd->nfsd_slp)
2843 				panic("nfsd wakeup");
2844 			slp->ns_sref++;
2845 			nd->nfsd_slp = slp;
2846 			wakeup((caddr_t)nd);
2847 			if (--nparallel == 0)
2848 				break;
2849 		}
2850 	}
2851 	if (nparallel) {
2852 		slp->ns_flag |= SLP_DOREC;
2853 		nfsd_head_flag |= NFSD_CHECKSLP;
2854 	}
2855 }
2856 #endif /* NFS_NOSERVER */
2857