xref: /dragonfly/sys/vfs/nfs/nfs_socket.c (revision b272101a)
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. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	@(#)nfs_socket.c	8.5 (Berkeley) 3/30/95
33  * $FreeBSD: src/sys/nfs/nfs_socket.c,v 1.60.2.6 2003/03/26 01:44:46 alfred Exp $
34  */
35 
36 /*
37  * Socket operations for use by nfs
38  */
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/proc.h>
43 #include <sys/malloc.h>
44 #include <sys/mount.h>
45 #include <sys/kernel.h>
46 #include <sys/mbuf.h>
47 #include <sys/vnode.h>
48 #include <sys/fcntl.h>
49 #include <sys/protosw.h>
50 #include <sys/resourcevar.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/socketops.h>
54 #include <sys/syslog.h>
55 #include <sys/thread.h>
56 #include <sys/tprintf.h>
57 #include <sys/sysctl.h>
58 #include <sys/signalvar.h>
59 
60 #include <sys/signal2.h>
61 #include <sys/mutex2.h>
62 #include <sys/socketvar2.h>
63 
64 #include <netinet/in.h>
65 #include <netinet/tcp.h>
66 #include <sys/thread2.h>
67 
68 #include "rpcv2.h"
69 #include "nfsproto.h"
70 #include "nfs.h"
71 #include "xdr_subs.h"
72 #include "nfsm_subs.h"
73 #include "nfsmount.h"
74 #include "nfsnode.h"
75 #include "nfsrtt.h"
76 
77 #define	TRUE	1
78 #define	FALSE	0
79 
80 /*
81  * RTT calculations are scaled by 256 (8 bits).  A proper fractional
82  * RTT will still be calculated even with a slow NFS timer.
83  */
84 #define	NFS_SRTT(r)	(r)->r_nmp->nm_srtt[proct[(r)->r_procnum]]
85 #define	NFS_SDRTT(r)	(r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum]]
86 #define NFS_RTT_SCALE_BITS	8	/* bits */
87 #define NFS_RTT_SCALE		256	/* value */
88 
89 /*
90  * Defines which timer to use for the procnum.
91  * 0 - default
92  * 1 - getattr
93  * 2 - lookup
94  * 3 - read
95  * 4 - write
96  */
97 static int proct[NFS_NPROCS] = {
98 	0, 1, 0, 2, 1, 3, 3, 4, 0, 0,	/* 00-09	*/
99 	0, 0, 0, 0, 0, 0, 3, 3, 0, 0,	/* 10-19	*/
100 	0, 5, 0, 0, 0, 0,		/* 20-29	*/
101 };
102 
103 static int multt[NFS_NPROCS] = {
104 	1, 1, 1, 1, 1, 1, 1, 1, 1, 1,	/* 00-09	*/
105 	1, 1, 1, 1, 1, 1, 1, 1, 1, 1,	/* 10-19	*/
106 	1, 2, 1, 1, 1, 1,		/* 20-29	*/
107 };
108 
109 static int nfs_backoff[8] = { 2, 3, 5, 8, 13, 21, 34, 55 };
110 static int nfs_realign_test;
111 static int nfs_realign_count;
112 static int nfs_showrtt;
113 static int nfs_showrexmit;
114 int nfs_maxasyncbio = NFS_MAXASYNCBIO;
115 
116 SYSCTL_DECL(_vfs_nfs);
117 
118 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0,
119     "Number of times mbufs have been tested for bad alignment");
120 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0,
121     "Number of realignments for badly aligned mbuf data");
122 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrtt, CTLFLAG_RW, &nfs_showrtt, 0,
123     "Show round trip time output");
124 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrexmit, CTLFLAG_RW, &nfs_showrexmit, 0,
125     "Show retransmits info");
126 SYSCTL_INT(_vfs_nfs, OID_AUTO, maxasyncbio, CTLFLAG_RW, &nfs_maxasyncbio, 0,
127     "Max number of asynchronous bio's");
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
nfs_connect(struct nfsmount * nmp,struct nfsreq * rep)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, TRUE);
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 				soclrstate(so, 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 			bzero(&sopt, sizeof sopt);
320 			sopt.sopt_level = IPPROTO_TCP;
321 			sopt.sopt_name = TCP_FASTKEEP;
322 			sopt.sopt_val = &val;
323 			sopt.sopt_valsize = sizeof val;
324 			val = 1;
325 			sosetopt(so, &sopt);
326 		}
327 	}
328 	error = soreserve(so, nfs_soreserve, nfs_soreserve, NULL);
329 	if (error)
330 		goto bad;
331 	atomic_set_int(&so->so_rcv.ssb_flags, SSB_NOINTR);
332 	atomic_set_int(&so->so_snd.ssb_flags, SSB_NOINTR);
333 
334 	/*
335 	 * Clear AUTOSIZE, otherwise the socket buffer could be reduced
336 	 * to the point where rpc's cannot be queued using the mbuf
337 	 * interface.
338 	 */
339 	atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
340 	atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
341 
342 	/* Initialize other non-zero congestion variables */
343 	nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] =
344 		nmp->nm_srtt[3] = (NFS_TIMEO << NFS_RTT_SCALE_BITS);
345 	nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
346 		nmp->nm_sdrtt[3] = 0;
347 	nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
348 	nmp->nm_timeouts = 0;
349 
350 	/*
351 	 * Assign nm_so last.  The moment nm_so is assigned the nfs_timer()
352 	 * can mess with the socket.
353 	 */
354 	nmp->nm_so = so;
355 	return (0);
356 
357 bad:
358 	if (so) {
359 		soshutdown(so, SHUT_RDWR);
360 		soclose(so, FNONBLOCK);
361 	}
362 	return (error);
363 }
364 
365 /*
366  * Reconnect routine:
367  * Called when a connection is broken on a reliable protocol.
368  * - clean up the old socket
369  * - nfs_connect() again
370  * - set R_NEEDSXMIT for all outstanding requests on mount point
371  * If this fails the mount point is DEAD!
372  * nb: Must be called with the nfs_sndlock() set on the mount point.
373  */
374 static int
nfs_reconnect(struct nfsmount * nmp,struct nfsreq * rep)375 nfs_reconnect(struct nfsmount *nmp, struct nfsreq *rep)
376 {
377 	struct nfsreq *req;
378 	int error;
379 
380 	nfs_disconnect(nmp);
381 	if (nmp->nm_rxstate >= NFSSVC_STOPPING)
382 		return (EINTR);
383 	while ((error = nfs_connect(nmp, rep)) != 0) {
384 		if (error == EINTR || error == ERESTART)
385 			return (EINTR);
386 		if (error == EINVAL)
387 			return (error);
388 		if (nmp->nm_rxstate >= NFSSVC_STOPPING)
389 			return (EINTR);
390 		(void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
391 	}
392 
393 	/*
394 	 * Loop through outstanding request list and fix up all requests
395 	 * on old socket.
396 	 */
397 	crit_enter();
398 	TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
399 		KKASSERT(req->r_nmp == nmp);
400 		req->r_flags |= R_NEEDSXMIT;
401 	}
402 	crit_exit();
403 	return (0);
404 }
405 
406 /*
407  * NFS disconnect. Clean up and unlink.
408  */
409 void
nfs_disconnect(struct nfsmount * nmp)410 nfs_disconnect(struct nfsmount *nmp)
411 {
412 	struct socket *so;
413 
414 	if (nmp->nm_so) {
415 		so = nmp->nm_so;
416 		nmp->nm_so = NULL;
417 		soshutdown(so, SHUT_RDWR);
418 		soclose(so, FNONBLOCK);
419 	}
420 }
421 
422 void
nfs_safedisconnect(struct nfsmount * nmp)423 nfs_safedisconnect(struct nfsmount *nmp)
424 {
425 	int error;
426 
427 	error = nfs_rcvlock(nmp, NULL);
428 	nfs_disconnect(nmp);
429 	if (error == 0)
430 		nfs_rcvunlock(nmp);
431 }
432 
433 /*
434  * This is the nfs send routine. For connection based socket types, it
435  * must be called with an nfs_sndlock() on the socket.
436  * "rep == NULL" indicates that it has been called from a server.
437  * For the client side:
438  * - return EINTR if the RPC is terminated, 0 otherwise
439  * - set R_NEEDSXMIT if the send fails for any reason
440  * - do any cleanup required by recoverable socket errors (?)
441  * For the server side:
442  * - return EINTR or ERESTART if interrupted by a signal
443  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
444  * - do any cleanup required by recoverable socket errors (?)
445  */
446 int
nfs_send(struct socket * so,struct sockaddr * nam,struct mbuf * top,struct nfsreq * rep)447 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
448 	 struct nfsreq *rep)
449 {
450 	struct sockaddr *sendnam;
451 	int error, soflags, flags;
452 
453 	if (rep) {
454 		if (rep->r_flags & R_SOFTTERM) {
455 			m_freem(top);
456 			return (EINTR);
457 		}
458 		if ((so = rep->r_nmp->nm_so) == NULL) {
459 			rep->r_flags |= R_NEEDSXMIT;
460 			m_freem(top);
461 			return (0);
462 		}
463 		rep->r_flags &= ~R_NEEDSXMIT;
464 		soflags = rep->r_nmp->nm_soflags;
465 	} else {
466 		soflags = so->so_proto->pr_flags;
467 	}
468 	if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
469 		sendnam = NULL;
470 	else
471 		sendnam = nam;
472 	if (so->so_type == SOCK_SEQPACKET)
473 		flags = MSG_EOR;
474 	else
475 		flags = 0;
476 
477 	/*
478 	 * calls pru_sosend -> sosend -> so_pru_send -> netrpc
479 	 */
480 	error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
481 			      curthread /*XXX*/);
482 
483 	/*
484 	 * ENOBUFS for dgram sockets is transient and non fatal.
485 	 * No need to log, and no need to break a soft mount.
486 	 */
487 	if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
488 		error = 0;
489 		/*
490 		 * do backoff retransmit on client
491 		 */
492 		if (rep) {
493 			if ((rep->r_nmp->nm_state & NFSSTA_SENDSPACE) == 0) {
494 				rep->r_nmp->nm_state |= NFSSTA_SENDSPACE;
495 				kprintf("Warning: NFS: Insufficient sendspace "
496 					"(%lu),\n"
497 					"\t You must increase vfs.nfs.soreserve"
498 					"or decrease vfs.nfs.maxasyncbio\n",
499 					so->so_snd.ssb_hiwat);
500 			}
501 			rep->r_flags |= R_NEEDSXMIT;
502 		}
503 	}
504 
505 	if (error) {
506 		if (rep) {
507 			log(LOG_INFO, "nfs send error %d for server %s\n",error,
508 			    rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
509 			/*
510 			 * Deal with errors for the client side.
511 			 */
512 			if (rep->r_flags & R_SOFTTERM)
513 				error = EINTR;
514 			else
515 				rep->r_flags |= R_NEEDSXMIT;
516 		} else {
517 			log(LOG_INFO, "nfsd send error %d\n", error);
518 		}
519 
520 		/*
521 		 * Handle any recoverable (soft) socket errors here. (?)
522 		 */
523 		if (error != EINTR && error != ERESTART &&
524 			error != EWOULDBLOCK && error != EPIPE)
525 			error = 0;
526 	}
527 	return (error);
528 }
529 
530 /*
531  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
532  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
533  * Mark and consolidate the data into a new mbuf list.
534  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
535  *     small mbufs.
536  * For SOCK_STREAM we must be very careful to read an entire record once
537  * we have read any of it, even if the system call has been interrupted.
538  */
539 static int
nfs_receive(struct nfsmount * nmp,struct nfsreq * rep,struct sockaddr ** aname,struct mbuf ** mp)540 nfs_receive(struct nfsmount *nmp, struct nfsreq *rep,
541 	    struct sockaddr **aname, struct mbuf **mp)
542 {
543 	struct socket *so;
544 	struct sockbuf sio;
545 	struct uio auio;
546 	struct iovec aio;
547 	struct mbuf *m;
548 	struct mbuf *control;
549 	u_int32_t len;
550 	struct sockaddr **getnam;
551 	int error, sotype, rcvflg;
552 	struct thread *td = curthread;	/* XXX */
553 
554 	/*
555 	 * Set up arguments for soreceive()
556 	 */
557 	*mp = NULL;
558 	*aname = NULL;
559 	sotype = nmp->nm_sotype;
560 
561 	/*
562 	 * For reliable protocols, lock against other senders/receivers
563 	 * in case a reconnect is necessary.
564 	 * For SOCK_STREAM, first get the Record Mark to find out how much
565 	 * more there is to get.
566 	 * We must lock the socket against other receivers
567 	 * until we have an entire rpc request/reply.
568 	 */
569 	if (sotype != SOCK_DGRAM) {
570 		error = nfs_sndlock(nmp, rep);
571 		if (error)
572 			return (error);
573 tryagain:
574 		/*
575 		 * Check for fatal errors and resending request.
576 		 */
577 		/*
578 		 * Ugh: If a reconnect attempt just happened, nm_so
579 		 * would have changed. NULL indicates a failed
580 		 * attempt that has essentially shut down this
581 		 * mount point.
582 		 */
583 		if (rep && (rep->r_mrep || (rep->r_flags & R_SOFTTERM))) {
584 			nfs_sndunlock(nmp);
585 			return (EINTR);
586 		}
587 		so = nmp->nm_so;
588 		if (so == NULL) {
589 			error = nfs_reconnect(nmp, rep);
590 			if (error) {
591 				nfs_sndunlock(nmp);
592 				return (error);
593 			}
594 			goto tryagain;
595 		}
596 		while (rep && (rep->r_flags & R_NEEDSXMIT)) {
597 			m = m_copym(rep->r_mreq, 0, M_COPYALL, M_WAITOK);
598 			nfsstats.rpcretries++;
599 			error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
600 			if (error) {
601 				if (error == EINTR || error == ERESTART ||
602 				    (error = nfs_reconnect(nmp, rep)) != 0) {
603 					nfs_sndunlock(nmp);
604 					return (error);
605 				}
606 				goto tryagain;
607 			}
608 		}
609 		nfs_sndunlock(nmp);
610 		if (sotype == SOCK_STREAM) {
611 			/*
612 			 * Get the length marker from the stream
613 			 */
614 			aio.iov_base = (caddr_t)&len;
615 			aio.iov_len = sizeof(u_int32_t);
616 			auio.uio_iov = &aio;
617 			auio.uio_iovcnt = 1;
618 			auio.uio_segflg = UIO_SYSSPACE;
619 			auio.uio_rw = UIO_READ;
620 			auio.uio_offset = 0;
621 			auio.uio_resid = sizeof(u_int32_t);
622 			auio.uio_td = td;
623 			do {
624 			   rcvflg = MSG_WAITALL;
625 			   error = so_pru_soreceive(so, NULL, &auio, NULL,
626 						    NULL, &rcvflg);
627 			   if (error == EWOULDBLOCK && rep) {
628 				if (rep->r_flags & R_SOFTTERM)
629 					return (EINTR);
630 			   }
631 			} while (error == EWOULDBLOCK);
632 
633 			if (error == 0 && auio.uio_resid > 0) {
634 			    /*
635 			     * Only log short packets if not EOF
636 			     */
637 			    if (auio.uio_resid != sizeof(u_int32_t)) {
638 				log(LOG_INFO,
639 				    "short receive (%d/%d) from nfs server %s\n",
640 				    (int)(sizeof(u_int32_t) - auio.uio_resid),
641 				    (int)sizeof(u_int32_t),
642 				    nmp->nm_mountp->mnt_stat.f_mntfromname);
643 			    }
644 			    error = EPIPE;
645 			}
646 			if (error)
647 				goto errout;
648 			len = ntohl(len) & ~0x80000000;
649 			/*
650 			 * This is SERIOUS! We are out of sync with the sender
651 			 * and forcing a disconnect/reconnect is all I can do.
652 			 */
653 			if (len > NFS_MAXPACKET) {
654 			    log(LOG_ERR, "%s (%d) from nfs server %s\n",
655 				"impossible packet length",
656 				len,
657 				nmp->nm_mountp->mnt_stat.f_mntfromname);
658 			    error = EFBIG;
659 			    goto errout;
660 			}
661 
662 			/*
663 			 * Get the rest of the packet as an mbuf chain
664 			 */
665 			sbinit(&sio, len);
666 			do {
667 			    rcvflg = MSG_WAITALL;
668 			    error = so_pru_soreceive(so, NULL, NULL, &sio,
669 						     NULL, &rcvflg);
670 			} while (error == EWOULDBLOCK || error == EINTR ||
671 				 error == ERESTART);
672 			if (error == 0 && sio.sb_cc != len) {
673 			    if (sio.sb_cc != 0) {
674 				log(LOG_INFO,
675 				    "short receive (%zu/%d) from nfs server %s\n",
676 				    (size_t)len - auio.uio_resid, len,
677 				    nmp->nm_mountp->mnt_stat.f_mntfromname);
678 			    }
679 			    error = EPIPE;
680 			}
681 			*mp = sio.sb_mb;
682 		} else {
683 			/*
684 			 * Non-stream, so get the whole packet by not
685 			 * specifying MSG_WAITALL and by specifying a large
686 			 * length.
687 			 *
688 			 * We have no use for control msg., but must grab them
689 			 * and then throw them away so we know what is going
690 			 * on.
691 			 */
692 			sbinit(&sio, 100000000);
693 			do {
694 			    rcvflg = 0;
695 			    error =  so_pru_soreceive(so, NULL, NULL, &sio,
696 						      &control, &rcvflg);
697 			    if (control)
698 				m_freem(control);
699 			    if (error == EWOULDBLOCK && rep) {
700 				if (rep->r_flags & R_SOFTTERM) {
701 					m_freem(sio.sb_mb);
702 					return (EINTR);
703 				}
704 			    }
705 			} while (error == EWOULDBLOCK ||
706 				 (error == 0 && sio.sb_mb == NULL && control));
707 			if ((rcvflg & MSG_EOR) == 0)
708 				kprintf("Egad!!\n");
709 			if (error == 0 && sio.sb_mb == NULL)
710 				error = EPIPE;
711 			len = sio.sb_cc;
712 			*mp = sio.sb_mb;
713 		}
714 errout:
715 		if (error && error != EINTR && error != ERESTART) {
716 			m_freem(*mp);
717 			*mp = NULL;
718 			if (error != EPIPE) {
719 				log(LOG_INFO,
720 				    "receive error %d from nfs server %s\n",
721 				    error,
722 				 nmp->nm_mountp->mnt_stat.f_mntfromname);
723 			}
724 			error = nfs_sndlock(nmp, rep);
725 			if (!error) {
726 				error = nfs_reconnect(nmp, rep);
727 				if (!error)
728 					goto tryagain;
729 				else
730 					nfs_sndunlock(nmp);
731 			}
732 		}
733 	} else {
734 		if ((so = nmp->nm_so) == NULL)
735 			return (EACCES);
736 		if (so->so_state & SS_ISCONNECTED)
737 			getnam = NULL;
738 		else
739 			getnam = aname;
740 		sbinit(&sio, 100000000);
741 		do {
742 			rcvflg = 0;
743 			error =  so_pru_soreceive(so, getnam, NULL, &sio,
744 						  NULL, &rcvflg);
745 			if (error == EWOULDBLOCK && rep &&
746 			    (rep->r_flags & R_SOFTTERM)) {
747 				m_freem(sio.sb_mb);
748 				return (EINTR);
749 			}
750 		} while (error == EWOULDBLOCK);
751 
752 		len = sio.sb_cc;
753 		*mp = sio.sb_mb;
754 
755 		/*
756 		 * A shutdown may result in no error and no mbuf.
757 		 * Convert to EPIPE.
758 		 */
759 		if (*mp == NULL && error == 0)
760 			error = EPIPE;
761 	}
762 	if (error) {
763 		m_freem(*mp);
764 		*mp = NULL;
765 	}
766 
767 	/*
768 	 * Search for any mbufs that are not a multiple of 4 bytes long
769 	 * or with m_data not longword aligned.
770 	 * These could cause pointer alignment problems, so copy them to
771 	 * well aligned mbufs.
772 	 */
773 	nfs_realign(mp, 5 * NFSX_UNSIGNED);
774 	return (error);
775 }
776 
777 /*
778  * Implement receipt of reply on a socket.
779  *
780  * We must search through the list of received datagrams matching them
781  * with outstanding requests using the xid, until ours is found.
782  *
783  * If myrep is NULL we process packets on the socket until
784  * interrupted or until nm_reqrxq is non-empty.
785  */
786 /* ARGSUSED */
787 int
nfs_reply(struct nfsmount * nmp,struct nfsreq * myrep)788 nfs_reply(struct nfsmount *nmp, struct nfsreq *myrep)
789 {
790 	struct nfsreq *rep;
791 	struct sockaddr *nam;
792 	u_int32_t rxid;
793 	u_int32_t *tl;
794 	int error;
795 	struct nfsm_info info;
796 
797 	/*
798 	 * Loop around until we get our own reply
799 	 */
800 	for (;;) {
801 		/*
802 		 * Lock against other receivers so that I don't get stuck in
803 		 * sbwait() after someone else has received my reply for me.
804 		 * Also necessary for connection based protocols to avoid
805 		 * race conditions during a reconnect.
806 		 *
807 		 * If nfs_rcvlock() returns EALREADY, that means that
808 		 * the reply has already been recieved by another
809 		 * process and we can return immediately.  In this
810 		 * case, the lock is not taken to avoid races with
811 		 * other processes.
812 		 */
813 		info.mrep = NULL;
814 
815 		error = nfs_rcvlock(nmp, myrep);
816 		if (error == EALREADY)
817 			return (0);
818 		if (error)
819 			return (error);
820 
821 		/*
822 		 * If myrep is NULL we are the receiver helper thread.
823 		 * Stop waiting for incoming replies if there are
824 		 * messages sitting on reqrxq that we need to process,
825 		 * or if a shutdown request is pending.
826 		 */
827 		if (myrep == NULL && (TAILQ_FIRST(&nmp->nm_reqrxq) ||
828 		    nmp->nm_rxstate > NFSSVC_PENDING)) {
829 			nfs_rcvunlock(nmp);
830 			return(EWOULDBLOCK);
831 		}
832 
833 		/*
834 		 * Get the next Rpc reply off the socket
835 		 *
836 		 * We cannot release the receive lock until we've
837 		 * filled in rep->r_mrep, otherwise a waiting
838 		 * thread may deadlock in soreceive with no incoming
839 		 * packets expected.
840 		 */
841 		error = nfs_receive(nmp, myrep, &nam, &info.mrep);
842 		if (error) {
843 			/*
844 			 * Ignore routing errors on connectionless protocols??
845 			 */
846 			nfs_rcvunlock(nmp);
847 			if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
848 				if (nmp->nm_so == NULL)
849 					return (error);
850 				nmp->nm_so->so_error = 0;
851 				continue;
852 			}
853 			return (error);
854 		}
855 		if (nam)
856 			kfree(nam, M_SONAME);
857 
858 		/*
859 		 * Get the xid and check that it is an rpc reply
860 		 */
861 		info.md = info.mrep;
862 		info.dpos = mtod(info.md, caddr_t);
863 		NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
864 		rxid = *tl++;
865 		if (*tl != rpc_reply) {
866 			nfsstats.rpcinvalid++;
867 			m_freem(info.mrep);
868 			info.mrep = NULL;
869 nfsmout:
870 			nfs_rcvunlock(nmp);
871 			continue;
872 		}
873 
874 		/*
875 		 * Loop through the request list to match up the reply
876 		 * Iff no match, just drop the datagram.  On match, set
877 		 * r_mrep atomically to prevent the timer from messing
878 		 * around with the request after we have exited the critical
879 		 * section.
880 		 */
881 		crit_enter();
882 		TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
883 			if (rep->r_mrep == NULL && rxid == rep->r_xid)
884 				break;
885 		}
886 
887 		/*
888 		 * Fill in the rest of the reply if we found a match.
889 		 *
890 		 * Deal with duplicate responses if there was no match.
891 		 */
892 		if (rep) {
893 			rep->r_md = info.md;
894 			rep->r_dpos = info.dpos;
895 			if (nfsrtton) {
896 				struct rttl *rt;
897 
898 				rt = &nfsrtt.rttl[nfsrtt.pos];
899 				rt->proc = rep->r_procnum;
900 				rt->rto = 0;
901 				rt->sent = 0;
902 				rt->cwnd = nmp->nm_maxasync_scaled;
903 				rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
904 				rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
905 				rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
906 				getmicrotime(&rt->tstamp);
907 				if (rep->r_flags & R_TIMING)
908 					rt->rtt = rep->r_rtt;
909 				else
910 					rt->rtt = 1000000;
911 				nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
912 			}
913 
914 			/*
915 			 * New congestion control is based only on async
916 			 * requests.
917 			 */
918 			if (nmp->nm_maxasync_scaled < NFS_MAXASYNC_SCALED)
919 				++nmp->nm_maxasync_scaled;
920 			if (rep->r_flags & R_SENT) {
921 				rep->r_flags &= ~R_SENT;
922 			}
923 			/*
924 			 * Update rtt using a gain of 0.125 on the mean
925 			 * and a gain of 0.25 on the deviation.
926 			 *
927 			 * NOTE SRTT/SDRTT are only good if R_TIMING is set.
928 			 */
929 			if ((rep->r_flags & R_TIMING) && rep->r_rexmit == 0) {
930 				/*
931 				 * Since the timer resolution of
932 				 * NFS_HZ is so course, it can often
933 				 * result in r_rtt == 0. Since
934 				 * r_rtt == N means that the actual
935 				 * rtt is between N+dt and N+2-dt ticks,
936 				 * add 1.
937 				 */
938 				int n;
939 				int d;
940 
941 #define NFSRSB	NFS_RTT_SCALE_BITS
942 				n = ((NFS_SRTT(rep) * 7) +
943 				     (rep->r_rtt << NFSRSB)) >> 3;
944 				d = n - NFS_SRTT(rep);
945 				NFS_SRTT(rep) = n;
946 
947 				/*
948 				 * Don't let the jitter calculation decay
949 				 * too quickly, but we want a fast rampup.
950 				 */
951 				if (d < 0)
952 					d = -d;
953 				d <<= NFSRSB;
954 				if (d < NFS_SDRTT(rep))
955 					n = ((NFS_SDRTT(rep) * 15) + d) >> 4;
956 				else
957 					n = ((NFS_SDRTT(rep) * 3) + d) >> 2;
958 				NFS_SDRTT(rep) = n;
959 #undef NFSRSB
960 			}
961 			nmp->nm_timeouts = 0;
962 			rep->r_mrep = info.mrep;
963 			nfs_hardterm(rep, 0);
964 		} else {
965 			/*
966 			 * Extract vers, prog, nfsver, procnum.  A duplicate
967 			 * response means we didn't wait long enough so
968 			 * we increase the SRTT to avoid future spurious
969 			 * timeouts.
970 			 */
971 			u_int procnum = nmp->nm_lastreprocnum;
972 			int n;
973 
974 			if (procnum < NFS_NPROCS && proct[procnum]) {
975 				if (nfs_showrexmit)
976 					kprintf("D");
977 				n = nmp->nm_srtt[proct[procnum]];
978 				n += NFS_ASYSCALE * NFS_HZ;
979 				if (n < NFS_ASYSCALE * NFS_HZ * 10)
980 					n = NFS_ASYSCALE * NFS_HZ * 10;
981 				nmp->nm_srtt[proct[procnum]] = n;
982 			}
983 		}
984 		nfs_rcvunlock(nmp);
985 		crit_exit();
986 
987 		/*
988 		 * If not matched to a request, drop it.
989 		 * If it's mine, get out.
990 		 */
991 		if (rep == NULL) {
992 			nfsstats.rpcunexpected++;
993 			m_freem(info.mrep);
994 			info.mrep = NULL;
995 		} else if (rep == myrep) {
996 			if (rep->r_mrep == NULL)
997 				panic("nfsreply nil");
998 			return (0);
999 		}
1000 	}
1001 }
1002 
1003 /*
1004  * Run the request state machine until the target state is reached
1005  * or a fatal error occurs.  The target state is not run.  Specifying
1006  * a target of NFSM_STATE_DONE runs the state machine until the rpc
1007  * is complete.
1008  *
1009  * EINPROGRESS is returned for all states other then the DONE state,
1010  * indicating that the rpc is still in progress.
1011  */
1012 int
nfs_request(struct nfsm_info * info,nfsm_state_t bstate,nfsm_state_t estate)1013 nfs_request(struct nfsm_info *info, nfsm_state_t bstate, nfsm_state_t estate)
1014 {
1015 	struct nfsreq *req;
1016 
1017 	while (info->state >= bstate && info->state < estate) {
1018 		switch(info->state) {
1019 		case NFSM_STATE_SETUP:
1020 			/*
1021 			 * Setup the nfsreq.  Any error which occurs during
1022 			 * this state is fatal.
1023 			 */
1024 			info->error = nfs_request_setup(info);
1025 			if (info->error) {
1026 				info->state = NFSM_STATE_DONE;
1027 				return (info->error);
1028 			} else {
1029 				req = info->req;
1030 				req->r_mrp = &info->mrep;
1031 				req->r_mdp = &info->md;
1032 				req->r_dposp = &info->dpos;
1033 				info->state = NFSM_STATE_AUTH;
1034 			}
1035 			break;
1036 		case NFSM_STATE_AUTH:
1037 			/*
1038 			 * Authenticate the nfsreq.  Any error which occurs
1039 			 * during this state is fatal.
1040 			 */
1041 			info->error = nfs_request_auth(info->req);
1042 			if (info->error) {
1043 				info->state = NFSM_STATE_DONE;
1044 				return (info->error);
1045 			} else {
1046 				info->state = NFSM_STATE_TRY;
1047 			}
1048 			break;
1049 		case NFSM_STATE_TRY:
1050 			/*
1051 			 * Transmit or retransmit attempt.  An error in this
1052 			 * state is ignored and we always move on to the
1053 			 * next state.
1054 			 *
1055 			 * This can trivially race the receiver if the
1056 			 * request is asynchronous.  nfs_request_try()
1057 			 * will thus set the state for us and we
1058 			 * must also return immediately if we are
1059 			 * running an async state machine, because
1060 			 * info can become invalid due to races after
1061 			 * try() returns.
1062 			 */
1063 			if (info->req->r_flags & R_ASYNC) {
1064 				nfs_request_try(info->req);
1065 				if (estate == NFSM_STATE_WAITREPLY)
1066 					return (EINPROGRESS);
1067 			} else {
1068 				nfs_request_try(info->req);
1069 				info->state = NFSM_STATE_WAITREPLY;
1070 			}
1071 			break;
1072 		case NFSM_STATE_WAITREPLY:
1073 			/*
1074 			 * Wait for a reply or timeout and move on to the
1075 			 * next state.  The error returned by this state
1076 			 * is passed to the processing code in the next
1077 			 * state.
1078 			 */
1079 			info->error = nfs_request_waitreply(info->req);
1080 			info->state = NFSM_STATE_PROCESSREPLY;
1081 			break;
1082 		case NFSM_STATE_PROCESSREPLY:
1083 			/*
1084 			 * Process the reply or timeout.  Errors which occur
1085 			 * in this state may cause the state machine to
1086 			 * go back to an earlier state, and are fatal
1087 			 * otherwise.
1088 			 */
1089 			info->error = nfs_request_processreply(info,
1090 							       info->error);
1091 			switch(info->error) {
1092 			case ENEEDAUTH:
1093 				info->state = NFSM_STATE_AUTH;
1094 				break;
1095 			case EAGAIN:
1096 				info->state = NFSM_STATE_TRY;
1097 				break;
1098 			default:
1099 				/*
1100 				 * Operation complete, with or without an
1101 				 * error.  We are done.
1102 				 */
1103 				info->req = NULL;
1104 				info->state = NFSM_STATE_DONE;
1105 				return (info->error);
1106 			}
1107 			break;
1108 		case NFSM_STATE_DONE:
1109 			/*
1110 			 * Shouldn't be reached
1111 			 */
1112 			return (info->error);
1113 			/* NOT REACHED */
1114 		}
1115 	}
1116 
1117 	/*
1118 	 * If we are done return the error code (if any).
1119 	 * Otherwise return EINPROGRESS.
1120 	 */
1121 	if (info->state == NFSM_STATE_DONE)
1122 		return (info->error);
1123 	return (EINPROGRESS);
1124 }
1125 
1126 /*
1127  * nfs_request - goes something like this
1128  *	- fill in request struct
1129  *	- links it into list
1130  *	- calls nfs_send() for first transmit
1131  *	- calls nfs_receive() to get reply
1132  *	- break down rpc header and return with nfs reply pointed to
1133  *	  by mrep or error
1134  * nb: always frees up mreq mbuf list
1135  */
1136 static int
nfs_request_setup(nfsm_info_t info)1137 nfs_request_setup(nfsm_info_t info)
1138 {
1139 	struct nfsreq *req;
1140 	struct nfsmount *nmp;
1141 	struct mbuf *m;
1142 	int i;
1143 
1144 	/*
1145 	 * Reject requests while attempting a forced unmount.
1146 	 */
1147 	if (info->vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
1148 		m_freem(info->mreq);
1149 		info->mreq = NULL;
1150 		return (EIO);
1151 	}
1152 	nmp = VFSTONFS(info->vp->v_mount);
1153 	req = kmalloc(sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
1154 	req->r_nmp = nmp;
1155 	req->r_vp = info->vp;
1156 	req->r_td = info->td;
1157 	req->r_procnum = info->procnum;
1158 	req->r_mreq = NULL;
1159 	req->r_cred = info->cred;
1160 
1161 	i = 0;
1162 	m = info->mreq;
1163 	while (m) {
1164 		i += m->m_len;
1165 		m = m->m_next;
1166 	}
1167 	req->r_mrest = info->mreq;
1168 	req->r_mrest_len = i;
1169 
1170 	/*
1171 	 * The presence of a non-NULL r_info in req indicates
1172 	 * async completion via our helper threads.  See the receiver
1173 	 * code.
1174 	 */
1175 	if (info->bio) {
1176 		req->r_info = info;
1177 		req->r_flags = R_ASYNC;
1178 	} else {
1179 		req->r_info = NULL;
1180 		req->r_flags = 0;
1181 	}
1182 	info->req = req;
1183 	return(0);
1184 }
1185 
1186 static int
nfs_request_auth(struct nfsreq * rep)1187 nfs_request_auth(struct nfsreq *rep)
1188 {
1189 	struct nfsmount *nmp = rep->r_nmp;
1190 	struct mbuf *m;
1191 	char nickv[RPCX_NICKVERF];
1192 	int error = 0, auth_len, auth_type;
1193 	int verf_len;
1194 	u_int32_t xid;
1195 	char *auth_str, *verf_str;
1196 	struct ucred *cred;
1197 
1198 	cred = rep->r_cred;
1199 	rep->r_failed_auth = 0;
1200 
1201 	/*
1202 	 * Get the RPC header with authorization.
1203 	 */
1204 	verf_str = auth_str = NULL;
1205 	verf_len = 0;
1206 	if (nmp->nm_flag & NFSMNT_KERB) {
1207 		verf_str = nickv;
1208 		verf_len = sizeof (nickv);
1209 		auth_type = RPCAUTH_KERB4;
1210 		bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1211 		if (rep->r_failed_auth ||
1212 		    nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1213 				    verf_str, verf_len)) {
1214 			error = nfs_getauth(nmp, rep, cred, &auth_str,
1215 				&auth_len, verf_str, &verf_len, rep->r_key);
1216 			if (error) {
1217 				m_freem(rep->r_mrest);
1218 				rep->r_mrest = NULL;
1219 				kfree((caddr_t)rep, M_NFSREQ);
1220 				return (error);
1221 			}
1222 		}
1223 	} else {
1224 		auth_type = RPCAUTH_UNIX;
1225 		if (cred->cr_ngroups < 1)
1226 			panic("nfsreq nogrps");
1227 		auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1228 			nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1229 			5 * NFSX_UNSIGNED;
1230 	}
1231 	if (rep->r_mrest)
1232 		nfs_checkpkt(rep->r_mrest, rep->r_mrest_len);
1233 	m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1234 			auth_len, auth_str, verf_len, verf_str,
1235 			rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1236 	rep->r_mrest = NULL;
1237 	if (auth_str)
1238 		kfree(auth_str, M_TEMP);
1239 
1240 	/*
1241 	 * For stream protocols, insert a Sun RPC Record Mark.
1242 	 */
1243 	if (nmp->nm_sotype == SOCK_STREAM) {
1244 		M_PREPEND(m, NFSX_UNSIGNED, M_WAITOK);
1245 		*mtod(m, u_int32_t *) = htonl(0x80000000 |
1246 			 (m->m_pkthdr.len - NFSX_UNSIGNED));
1247 	}
1248 
1249 	nfs_checkpkt(m, m->m_pkthdr.len);
1250 
1251 	rep->r_mreq = m;
1252 	rep->r_xid = xid;
1253 	return (0);
1254 }
1255 
1256 static int
nfs_request_try(struct nfsreq * rep)1257 nfs_request_try(struct nfsreq *rep)
1258 {
1259 	struct nfsmount *nmp = rep->r_nmp;
1260 	struct mbuf *m2;
1261 	int error;
1262 
1263 	/*
1264 	 * Request is not on any queue, only the owner has access to it
1265 	 * so it should not be locked by anyone atm.
1266 	 *
1267 	 * Interlock to prevent races.  While locked the only remote
1268 	 * action possible is for r_mrep to be set (once we enqueue it).
1269 	 */
1270 	if (rep->r_flags == 0xdeadc0de) {
1271 		print_backtrace(-1);
1272 		panic("flags nbad");
1273 	}
1274 	KKASSERT((rep->r_flags & (R_LOCKED | R_ONREQQ)) == 0);
1275 	if (nmp->nm_flag & NFSMNT_SOFT)
1276 		rep->r_retry = nmp->nm_retry;
1277 	else
1278 		rep->r_retry = NFS_MAXREXMIT + 1;	/* past clip limit */
1279 	rep->r_rtt = rep->r_rexmit = 0;
1280 	if (proct[rep->r_procnum] > 0)
1281 		rep->r_flags |= R_TIMING | R_LOCKED;
1282 	else
1283 		rep->r_flags |= R_LOCKED;
1284 	rep->r_mrep = NULL;
1285 
1286 	nfsstats.rpcrequests++;
1287 
1288 	if (nmp->nm_flag & NFSMNT_FORCE) {
1289 		rep->r_flags |= R_SOFTTERM;
1290 		rep->r_flags &= ~R_LOCKED;
1291 		if (rep->r_info)
1292 			rep->r_info->error = EINTR;
1293 		return (0);
1294 	}
1295 	rep->r_flags |= R_NEEDSXMIT;	/* in case send lock races us */
1296 
1297 	/*
1298 	 * Do the client side RPC.
1299 	 *
1300 	 * Chain request into list of outstanding requests. Be sure
1301 	 * to put it LAST so timer finds oldest requests first.  Note
1302 	 * that our control of R_LOCKED prevents the request from
1303 	 * getting ripped out from under us or transmitted by the
1304 	 * timer code.
1305 	 *
1306 	 * For requests with info structures we must atomically set the
1307 	 * info's state because the structure could become invalid upon
1308 	 * return due to races (i.e., if async)
1309 	 */
1310 	crit_enter();
1311 	mtx_link_init(&rep->r_link);
1312 	KKASSERT((rep->r_flags & R_ONREQQ) == 0);
1313 	TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1314 	rep->r_flags |= R_ONREQQ;
1315 	++nmp->nm_reqqlen;
1316 	if (rep->r_flags & R_ASYNC)
1317 		rep->r_info->state = NFSM_STATE_WAITREPLY;
1318 	crit_exit();
1319 
1320 	error = 0;
1321 
1322 	/*
1323 	 * Send if we can.  Congestion control is not handled here any more
1324 	 * becausing trying to defer the initial send based on the nfs_timer
1325 	 * requires having a very fast nfs_timer, which is silly.
1326 	 */
1327 	if (nmp->nm_so) {
1328 		if (nmp->nm_soflags & PR_CONNREQUIRED)
1329 			error = nfs_sndlock(nmp, rep);
1330 		if (error == 0 && (rep->r_flags & R_NEEDSXMIT)) {
1331 			m2 = m_copym(rep->r_mreq, 0, M_COPYALL, M_WAITOK);
1332 			error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1333 			rep->r_flags &= ~R_NEEDSXMIT;
1334 			if ((rep->r_flags & R_SENT) == 0) {
1335 				rep->r_flags |= R_SENT;
1336 			}
1337 			if (nmp->nm_soflags & PR_CONNREQUIRED)
1338 				nfs_sndunlock(nmp);
1339 		}
1340 	} else {
1341 		rep->r_rtt = -1;
1342 	}
1343 	if (error == EPIPE)
1344 		error = 0;
1345 
1346 	/*
1347 	 * Release the lock.  The only remote action that may have occurred
1348 	 * would have been the setting of rep->r_mrep.  If this occured
1349 	 * and the request was async we have to move it to the reader
1350 	 * thread's queue for action.
1351 	 *
1352 	 * For async requests also make sure the reader is woken up so
1353 	 * it gets on the socket to read responses.
1354 	 */
1355 	crit_enter();
1356 	if (rep->r_flags & R_ASYNC) {
1357 		if (rep->r_mrep)
1358 			nfs_hardterm(rep, 1);
1359 		rep->r_flags &= ~R_LOCKED;
1360 		nfssvc_iod_reader_wakeup(nmp);
1361 	} else {
1362 		rep->r_flags &= ~R_LOCKED;
1363 	}
1364 	if (rep->r_flags & R_WANTED) {
1365 		rep->r_flags &= ~R_WANTED;
1366 		wakeup(rep);
1367 	}
1368 	crit_exit();
1369 	return (error);
1370 }
1371 
1372 /*
1373  * This code is only called for synchronous requests.  Completed synchronous
1374  * requests are left on reqq and we remove them before moving on to the
1375  * processing state.
1376  */
1377 static int
nfs_request_waitreply(struct nfsreq * rep)1378 nfs_request_waitreply(struct nfsreq *rep)
1379 {
1380 	struct nfsmount *nmp = rep->r_nmp;
1381 	int error;
1382 
1383 	KKASSERT((rep->r_flags & R_ASYNC) == 0);
1384 
1385 	/*
1386 	 * Wait until the request is finished.
1387 	 */
1388 	error = nfs_reply(nmp, rep);
1389 
1390 	/*
1391 	 * RPC done, unlink the request, but don't rip it out from under
1392 	 * the callout timer.
1393 	 *
1394 	 * Once unlinked no other receiver or the timer will have
1395 	 * visibility, so we do not have to set R_LOCKED.
1396 	 */
1397 	crit_enter();
1398 	while (rep->r_flags & R_LOCKED) {
1399 		rep->r_flags |= R_WANTED;
1400 		tsleep(rep, 0, "nfstrac", 0);
1401 	}
1402 	KKASSERT(rep->r_flags & R_ONREQQ);
1403 	TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1404 	rep->r_flags &= ~R_ONREQQ;
1405 	--nmp->nm_reqqlen;
1406 	if (TAILQ_FIRST(&nmp->nm_bioq) &&
1407 	    nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1408 		nfssvc_iod_writer_wakeup(nmp);
1409 	}
1410 	crit_exit();
1411 
1412 	/*
1413 	 * Decrement the outstanding request count.
1414 	 */
1415 	if (rep->r_flags & R_SENT) {
1416 		rep->r_flags &= ~R_SENT;
1417 	}
1418 	return (error);
1419 }
1420 
1421 /*
1422  * Process reply with error returned from nfs_requet_waitreply().
1423  *
1424  * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1425  * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1426  */
1427 static int
nfs_request_processreply(nfsm_info_t info,int error)1428 nfs_request_processreply(nfsm_info_t info, int error)
1429 {
1430 	struct nfsreq *req = info->req;
1431 	struct nfsmount *nmp = req->r_nmp;
1432 	u_int32_t *tl;
1433 	int verf_type;
1434 	int i;
1435 
1436 	/*
1437 	 * If there was a successful reply and a tprintf msg.
1438 	 * tprintf a response.
1439 	 */
1440 	if (error == 0 && (req->r_flags & R_TPRINTFMSG)) {
1441 		nfs_msg(req->r_td,
1442 			nmp->nm_mountp->mnt_stat.f_mntfromname,
1443 			"is alive again");
1444 	}
1445 
1446 	/*
1447 	 * Assign response and handle any pre-process error.  Response
1448 	 * fields can be NULL if an error is already pending.
1449 	 */
1450 	info->mrep = req->r_mrep;
1451 	info->md = req->r_md;
1452 	info->dpos = req->r_dpos;
1453 
1454 	if (error) {
1455 		m_freem(req->r_mreq);
1456 		req->r_mreq = NULL;
1457 		kfree(req, M_NFSREQ);
1458 		info->req = NULL;
1459 		return (error);
1460 	}
1461 
1462 	/*
1463 	 * break down the rpc header and check if ok
1464 	 */
1465 	NULLOUT(tl = nfsm_dissect(info, 3 * NFSX_UNSIGNED));
1466 	if (*tl++ == rpc_msgdenied) {
1467 		if (*tl == rpc_mismatch) {
1468 			error = EOPNOTSUPP;
1469 		} else if ((nmp->nm_flag & NFSMNT_KERB) &&
1470 			   *tl++ == rpc_autherr) {
1471 			if (req->r_failed_auth == 0) {
1472 				req->r_failed_auth++;
1473 				req->r_mheadend->m_next = NULL;
1474 				m_freem(info->mrep);
1475 				info->mrep = NULL;
1476 				m_freem(req->r_mreq);
1477 				req->r_mreq = NULL;
1478 				return (ENEEDAUTH);
1479 			} else {
1480 				error = EAUTH;
1481 			}
1482 		} else {
1483 			error = EACCES;
1484 		}
1485 		m_freem(info->mrep);
1486 		info->mrep = NULL;
1487 		m_freem(req->r_mreq);
1488 		req->r_mreq = NULL;
1489 		kfree(req, M_NFSREQ);
1490 		info->req = NULL;
1491 		return (error);
1492 	}
1493 
1494 	/*
1495 	 * Grab any Kerberos verifier, otherwise just throw it away.
1496 	 */
1497 	verf_type = fxdr_unsigned(int, *tl++);
1498 	i = fxdr_unsigned(int32_t, *tl);
1499 	if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1500 		error = nfs_savenickauth(nmp, req->r_cred, i, req->r_key,
1501 					 &info->md, &info->dpos, info->mrep);
1502 		if (error)
1503 			goto nfsmout;
1504 	} else if (i > 0) {
1505 		ERROROUT(nfsm_adv(info, nfsm_rndup(i)));
1506 	}
1507 	NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1508 	/* 0 == ok */
1509 	if (*tl == 0) {
1510 		NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1511 		if (*tl != 0) {
1512 			error = fxdr_unsigned(int, *tl);
1513 
1514 			/*
1515 			 * Does anyone even implement this?  Just impose
1516 			 * a 1-second delay.
1517 			 */
1518 			if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1519 				error == NFSERR_TRYLATER) {
1520 				m_freem(info->mrep);
1521 				info->mrep = NULL;
1522 				error = 0;
1523 
1524 				tsleep((caddr_t)&lbolt, 0, "nqnfstry", 0);
1525 				return (EAGAIN);	/* goto tryagain */
1526 			}
1527 
1528 #if 0
1529 			/*
1530 			 * XXX We can't do this here any more because the
1531 			 *     caller may be holding a shared lock on the
1532 			 *     namecache entry.
1533 			 *
1534 			 * If the File Handle was stale, invalidate the
1535 			 * lookup cache, just in case.
1536 			 *
1537 			 * To avoid namecache<->vnode deadlocks we must
1538 			 * release the vnode lock if we hold it.
1539 			 */
1540 			if (error == ESTALE) {
1541 				struct vnode *vp = req->r_vp;
1542 				int ltype;
1543 
1544 				ltype = lockstatus(&vp->v_lock, curthread);
1545 				if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1546 					lockmgr(&vp->v_lock, LK_RELEASE);
1547 				cache_inval_vp(vp, CINV_CHILDREN);
1548 				if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1549 					lockmgr(&vp->v_lock, ltype);
1550 			}
1551 #endif
1552 			if (nmp->nm_flag & NFSMNT_NFSV3) {
1553 				KKASSERT(*req->r_mrp == info->mrep);
1554 				KKASSERT(*req->r_mdp == info->md);
1555 				KKASSERT(*req->r_dposp == info->dpos);
1556 				error |= NFSERR_RETERR;
1557 			} else {
1558 				m_freem(info->mrep);
1559 				info->mrep = NULL;
1560 			}
1561 			m_freem(req->r_mreq);
1562 			req->r_mreq = NULL;
1563 			kfree(req, M_NFSREQ);
1564 			info->req = NULL;
1565 			return (error);
1566 		}
1567 
1568 		KKASSERT(*req->r_mrp == info->mrep);
1569 		KKASSERT(*req->r_mdp == info->md);
1570 		KKASSERT(*req->r_dposp == info->dpos);
1571 		m_freem(req->r_mreq);
1572 		req->r_mreq = NULL;
1573 		kfree(req, M_NFSREQ);
1574 		return (0);
1575 	}
1576 	m_freem(info->mrep);
1577 	info->mrep = NULL;
1578 	error = EPROTONOSUPPORT;
1579 nfsmout:
1580 	m_freem(req->r_mreq);
1581 	req->r_mreq = NULL;
1582 	kfree(req, M_NFSREQ);
1583 	info->req = NULL;
1584 	return (error);
1585 }
1586 
1587 #ifndef NFS_NOSERVER
1588 /*
1589  * Generate the rpc reply header
1590  * siz arg. is used to decide if adding a cluster is worthwhile
1591  */
1592 int
nfs_rephead(int siz,struct nfsrv_descript * nd,struct nfssvc_sock * slp,int err,struct mbuf ** mrq,struct mbuf ** mbp,caddr_t * bposp)1593 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1594 	    int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1595 {
1596 	u_int32_t *tl;
1597 	struct nfsm_info info;
1598 
1599 	siz += RPC_REPLYSIZ;
1600 	info.mb = m_getl(max_hdr + siz, M_WAITOK, MT_DATA, M_PKTHDR, NULL);
1601 	info.mreq = info.mb;
1602 	info.mreq->m_pkthdr.len = 0;
1603 	/*
1604 	 * If this is not a cluster, try and leave leading space
1605 	 * for the lower level headers.
1606 	 */
1607 	if ((max_hdr + siz) < MINCLSIZE)
1608 		info.mreq->m_data += max_hdr;
1609 	tl = mtod(info.mreq, u_int32_t *);
1610 	info.mreq->m_len = 6 * NFSX_UNSIGNED;
1611 	info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1612 	*tl++ = txdr_unsigned(nd->nd_retxid);
1613 	*tl++ = rpc_reply;
1614 	if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1615 		*tl++ = rpc_msgdenied;
1616 		if (err & NFSERR_AUTHERR) {
1617 			*tl++ = rpc_autherr;
1618 			*tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1619 			info.mreq->m_len -= NFSX_UNSIGNED;
1620 			info.bpos -= NFSX_UNSIGNED;
1621 		} else {
1622 			*tl++ = rpc_mismatch;
1623 			*tl++ = txdr_unsigned(RPC_VER2);
1624 			*tl = txdr_unsigned(RPC_VER2);
1625 		}
1626 	} else {
1627 		*tl++ = rpc_msgaccepted;
1628 
1629 		/*
1630 		 * For Kerberos authentication, we must send the nickname
1631 		 * verifier back, otherwise just RPCAUTH_NULL.
1632 		 */
1633 		if (nd->nd_flag & ND_KERBFULL) {
1634 		    struct nfsuid *nuidp;
1635 		    struct timeval ktvout;
1636 
1637 		    for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1638 			nuidp != NULL; nuidp = nuidp->nu_hash.le_next) {
1639 			if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1640 			    (!nd->nd_nam2 || netaddr_match(AF_INET,
1641 			     &nuidp->nu_haddr, nd->nd_nam2)))
1642 			    break;
1643 		    }
1644 		    if (nuidp) {
1645 			/*
1646 			 * Encrypt the timestamp in ecb mode using the
1647 			 * session key.
1648 			 */
1649 #ifdef NFSKERB
1650 			XXX
1651 #else
1652 			ktvout.tv_sec = 0;
1653 			ktvout.tv_usec = 0;
1654 #endif
1655 
1656 			*tl++ = rpc_auth_kerb;
1657 			*tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1658 			*tl = ktvout.tv_sec;
1659 			tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1660 			*tl++ = ktvout.tv_usec;
1661 			*tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1662 		    } else {
1663 			*tl++ = 0;
1664 			*tl++ = 0;
1665 		    }
1666 		} else {
1667 			*tl++ = 0;
1668 			*tl++ = 0;
1669 		}
1670 		switch (err) {
1671 		case EPROGUNAVAIL:
1672 			*tl = txdr_unsigned(RPC_PROGUNAVAIL);
1673 			break;
1674 		case EPROGMISMATCH:
1675 			*tl = txdr_unsigned(RPC_PROGMISMATCH);
1676 			tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1677 			*tl++ = txdr_unsigned(2);
1678 			*tl = txdr_unsigned(3);
1679 			break;
1680 		case EPROCUNAVAIL:
1681 			*tl = txdr_unsigned(RPC_PROCUNAVAIL);
1682 			break;
1683 		case EBADRPC:
1684 			*tl = txdr_unsigned(RPC_GARBAGE);
1685 			break;
1686 		default:
1687 			*tl = 0;
1688 			if (err != NFSERR_RETVOID) {
1689 				tl = nfsm_build(&info, NFSX_UNSIGNED);
1690 				if (err)
1691 				    *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1692 				else
1693 				    *tl = 0;
1694 			}
1695 			break;
1696 		}
1697 	}
1698 
1699 	if (mrq != NULL)
1700 	    *mrq = info.mreq;
1701 	*mbp = info.mb;
1702 	*bposp = info.bpos;
1703 	if (err != 0 && err != NFSERR_RETVOID)
1704 		nfsstats.srvrpc_errs++;
1705 	return (0);
1706 }
1707 
1708 
1709 #endif /* NFS_NOSERVER */
1710 
1711 /*
1712  * Nfs timer routine.
1713  *
1714  * Scan the nfsreq list and retranmit any requests that have timed out
1715  * To avoid retransmission attempts on STREAM sockets (in the future) make
1716  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1717  *
1718  * Requests with attached responses, terminated requests, and
1719  * locked requests are ignored.  Locked requests will be picked up
1720  * in a later timer call.
1721  */
1722 void
nfs_timer_callout(void * arg)1723 nfs_timer_callout(void *arg /* never used */)
1724 {
1725 	struct nfsmount *nmp;
1726 	struct nfsreq *req;
1727 #ifndef NFS_NOSERVER
1728 	struct nfssvc_sock *slp;
1729 	u_quad_t cur_usec;
1730 #endif /* NFS_NOSERVER */
1731 
1732 	lwkt_gettoken(&nfs_token);
1733 	TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1734 		lwkt_gettoken(&nmp->nm_token);
1735 		TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1736 			KKASSERT(nmp == req->r_nmp);
1737 			if (req->r_mrep)
1738 				continue;
1739 			if (req->r_flags & (R_SOFTTERM | R_LOCKED))
1740 				continue;
1741 
1742 			/*
1743 			 * Handle timeout/retry.  Be sure to process r_mrep
1744 			 * for async requests that completed while we had
1745 			 * the request locked or they will hang in the reqq
1746 			 * forever.
1747 			 */
1748 			req->r_flags |= R_LOCKED;
1749 			if (nfs_sigintr(nmp, req, req->r_td)) {
1750 				nfs_softterm(req, 1);
1751 				req->r_flags &= ~R_LOCKED;
1752 			} else {
1753 				nfs_timer_req(req);
1754 				if (req->r_flags & R_ASYNC) {
1755 					if (req->r_mrep)
1756 						nfs_hardterm(req, 1);
1757 					req->r_flags &= ~R_LOCKED;
1758 					nfssvc_iod_reader_wakeup(nmp);
1759 				} else {
1760 					req->r_flags &= ~R_LOCKED;
1761 				}
1762 			}
1763 			if (req->r_flags & R_WANTED) {
1764 				req->r_flags &= ~R_WANTED;
1765 				wakeup(req);
1766 			}
1767 		}
1768 		lwkt_reltoken(&nmp->nm_token);
1769 	}
1770 #ifndef NFS_NOSERVER
1771 
1772 	/*
1773 	 * Scan the write gathering queues for writes that need to be
1774 	 * completed now.
1775 	 */
1776 	cur_usec = nfs_curusec();
1777 
1778 	TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1779 		/* XXX race against removal */
1780 		if (lwkt_trytoken(&slp->ns_token)) {
1781 			if (slp->ns_tq.lh_first &&
1782 			    (slp->ns_tq.lh_first->nd_time <= cur_usec)) {
1783 				nfsrv_wakenfsd(slp, 1);
1784 			}
1785 			lwkt_reltoken(&slp->ns_token);
1786 		}
1787 	}
1788 #endif /* NFS_NOSERVER */
1789 
1790 	callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer_callout, NULL);
1791 	lwkt_reltoken(&nfs_token);
1792 }
1793 
1794 static
1795 void
nfs_timer_req(struct nfsreq * req)1796 nfs_timer_req(struct nfsreq *req)
1797 {
1798 	struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1799 	struct nfsmount *nmp = req->r_nmp;
1800 	struct mbuf *m;
1801 	struct socket *so;
1802 	int timeo;
1803 	int error;
1804 
1805 	/*
1806 	 * rtt ticks and timeout calculation.  Return if the timeout
1807 	 * has not been reached yet, unless the packet is flagged
1808 	 * for an immediate send.
1809 	 *
1810 	 * The mean rtt doesn't help when we get random I/Os, we have
1811 	 * to multiply by fairly large numbers.
1812 	 */
1813 	if (req->r_rtt >= 0) {
1814 		/*
1815 		 * Calculate the timeout to test against.
1816 		 */
1817 		req->r_rtt++;
1818 		if (nmp->nm_flag & NFSMNT_DUMBTIMR) {
1819 			timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1820 		} else if (req->r_flags & R_TIMING) {
1821 			timeo = NFS_SRTT(req) + NFS_SDRTT(req);
1822 		} else {
1823 			timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1824 		}
1825 		timeo *= multt[req->r_procnum];
1826 		/* timeo is still scaled by SCALE_BITS */
1827 
1828 #define NFSFS	(NFS_RTT_SCALE * NFS_HZ)
1829 		if (req->r_flags & R_TIMING) {
1830 			static long last_time;
1831 			if (nfs_showrtt && last_time != time_uptime) {
1832 				kprintf("rpccmd %d NFS SRTT %d SDRTT %d "
1833 					"timeo %d.%03d\n",
1834 					proct[req->r_procnum],
1835 					NFS_SRTT(req), NFS_SDRTT(req),
1836 					timeo / NFSFS,
1837 					timeo % NFSFS * 1000 /  NFSFS);
1838 				last_time = time_uptime;
1839 			}
1840 		}
1841 #undef NFSFS
1842 
1843 		/*
1844 		 * deal with nfs_timer jitter.
1845 		 */
1846 		timeo = (timeo >> NFS_RTT_SCALE_BITS) + 1;
1847 		if (timeo < 2)
1848 			timeo = 2;
1849 
1850 		if (nmp->nm_timeouts > 0)
1851 			timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1852 		if (timeo > NFS_MAXTIMEO)
1853 			timeo = NFS_MAXTIMEO;
1854 		if (req->r_rtt <= timeo) {
1855 			if ((req->r_flags & R_NEEDSXMIT) == 0)
1856 				return;
1857 		} else if (nmp->nm_timeouts < 8) {
1858 			nmp->nm_timeouts++;
1859 		}
1860 	}
1861 
1862 	/*
1863 	 * Check for server not responding
1864 	 */
1865 	if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1866 	     req->r_rexmit > nmp->nm_deadthresh) {
1867 		nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1868 			"not responding");
1869 		req->r_flags |= R_TPRINTFMSG;
1870 	}
1871 	if (req->r_rexmit >= req->r_retry) {	/* too many */
1872 		nfsstats.rpctimeouts++;
1873 		nfs_softterm(req, 1);
1874 		return;
1875 	}
1876 
1877 	/*
1878 	 * Generally disable retransmission on reliable sockets,
1879 	 * unless the request is flagged for immediate send.
1880 	 */
1881 	if (nmp->nm_sotype != SOCK_DGRAM) {
1882 		if (++req->r_rexmit > NFS_MAXREXMIT)
1883 			req->r_rexmit = NFS_MAXREXMIT;
1884 		if ((req->r_flags & R_NEEDSXMIT) == 0)
1885 			return;
1886 	}
1887 
1888 	/*
1889 	 * Stop here if we do not have a socket!
1890 	 */
1891 	if ((so = nmp->nm_so) == NULL)
1892 		return;
1893 
1894 	/*
1895 	 * If there is enough space and the window allows.. resend it.
1896 	 *
1897 	 * r_rtt is left intact in case we get an answer after the
1898 	 * retry that was a reply to the original packet.
1899 	 *
1900 	 * NOTE: so_pru_send()
1901 	 */
1902 	if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1903 	    (req->r_flags & (R_SENT | R_NEEDSXMIT)) &&
1904 	   (m = m_copym(req->r_mreq, 0, M_COPYALL, M_NOWAIT))){
1905 		if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1906 		    error = so_pru_send(so, 0, m, NULL, NULL, td);
1907 		else
1908 		    error = so_pru_send(so, 0, m, nmp->nm_nam, NULL, td);
1909 		if (error) {
1910 			if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1911 				so->so_error = 0;
1912 			req->r_flags |= R_NEEDSXMIT;
1913 		} else if (req->r_mrep == NULL) {
1914 			/*
1915 			 * Iff first send, start timing
1916 			 * else turn timing off, backoff timer
1917 			 * and divide congestion window by 2.
1918 			 *
1919 			 * It is possible for the so_pru_send() to
1920 			 * block and for us to race a reply so we
1921 			 * only do this if the reply field has not
1922 			 * been filled in.  R_LOCKED will prevent
1923 			 * the request from being ripped out from under
1924 			 * us entirely.
1925 			 *
1926 			 * Record the last resent procnum to aid us
1927 			 * in duplicate detection on receive.
1928 			 */
1929 			if ((req->r_flags & R_NEEDSXMIT) == 0) {
1930 				if (nfs_showrexmit)
1931 					kprintf("X");
1932 				if (++req->r_rexmit > NFS_MAXREXMIT)
1933 					req->r_rexmit = NFS_MAXREXMIT;
1934 				nmp->nm_maxasync_scaled >>= 1;
1935 				if (nmp->nm_maxasync_scaled < NFS_MINASYNC_SCALED)
1936 					nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
1937 				nfsstats.rpcretries++;
1938 				nmp->nm_lastreprocnum = req->r_procnum;
1939 			} else {
1940 				req->r_flags |= R_SENT;
1941 				req->r_flags &= ~R_NEEDSXMIT;
1942 			}
1943 		}
1944 	}
1945 }
1946 
1947 /*
1948  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1949  * wait for all requests to complete. This is used by forced unmounts
1950  * to terminate any outstanding RPCs.
1951  *
1952  * Locked requests cannot be canceled but will be marked for
1953  * soft-termination.
1954  */
1955 int
nfs_nmcancelreqs(struct nfsmount * nmp)1956 nfs_nmcancelreqs(struct nfsmount *nmp)
1957 {
1958 	struct nfsreq *req;
1959 	int i;
1960 
1961 	crit_enter();
1962 	TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1963 		if (req->r_mrep != NULL || (req->r_flags & R_SOFTTERM))
1964 			continue;
1965 		nfs_softterm(req, 0);
1966 	}
1967 	/* XXX  the other two queues as well */
1968 	crit_exit();
1969 
1970 	for (i = 0; i < 30; i++) {
1971 		crit_enter();
1972 		TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1973 			if (nmp == req->r_nmp)
1974 				break;
1975 		}
1976 		crit_exit();
1977 		if (req == NULL)
1978 			return (0);
1979 		tsleep(&lbolt, 0, "nfscancel", 0);
1980 	}
1981 	return (EBUSY);
1982 }
1983 
1984 /*
1985  * Soft-terminate a request, effectively marking it as failed.
1986  *
1987  * Must be called from within a critical section.
1988  */
1989 static void
nfs_softterm(struct nfsreq * rep,int islocked)1990 nfs_softterm(struct nfsreq *rep, int islocked)
1991 {
1992 	rep->r_flags |= R_SOFTTERM;
1993 	if (rep->r_info)
1994 		rep->r_info->error = EINTR;
1995 	nfs_hardterm(rep, islocked);
1996 }
1997 
1998 /*
1999  * Hard-terminate a request, typically after getting a response.
2000  *
2001  * The state machine can still decide to re-issue it later if necessary.
2002  *
2003  * Must be called from within a critical section.
2004  */
2005 static void
nfs_hardterm(struct nfsreq * rep,int islocked)2006 nfs_hardterm(struct nfsreq *rep, int islocked)
2007 {
2008 	struct nfsmount *nmp = rep->r_nmp;
2009 
2010 	/*
2011 	 * The nm_send count is decremented now to avoid deadlocks
2012 	 * when the process in soreceive() hasn't yet managed to send
2013 	 * its own request.
2014 	 */
2015 	if (rep->r_flags & R_SENT) {
2016 		rep->r_flags &= ~R_SENT;
2017 	}
2018 
2019 	/*
2020 	 * If we locked the request or nobody else has locked the request,
2021 	 * and the request is async, we can move it to the reader thread's
2022 	 * queue now and fix up the state.
2023 	 *
2024 	 * If we locked the request or nobody else has locked the request,
2025 	 * we can wake up anyone blocked waiting for a response on the
2026 	 * request.
2027 	 */
2028 	if (islocked || (rep->r_flags & R_LOCKED) == 0) {
2029 		if ((rep->r_flags & (R_ONREQQ | R_ASYNC)) ==
2030 		    (R_ONREQQ | R_ASYNC)) {
2031 			rep->r_flags &= ~R_ONREQQ;
2032 			TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
2033 			--nmp->nm_reqqlen;
2034 			TAILQ_INSERT_TAIL(&nmp->nm_reqrxq, rep, r_chain);
2035 			KKASSERT(rep->r_info->state == NFSM_STATE_TRY ||
2036 				 rep->r_info->state == NFSM_STATE_WAITREPLY);
2037 
2038 			/*
2039 			 * When setting the state to PROCESSREPLY we must
2040 			 * roll-up any error not related to the contents of
2041 			 * the reply (i.e. if there is no contents).
2042 			 */
2043 			rep->r_info->state = NFSM_STATE_PROCESSREPLY;
2044 			nfssvc_iod_reader_wakeup(nmp);
2045 			if (TAILQ_FIRST(&nmp->nm_bioq) &&
2046 			    nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
2047 				nfssvc_iod_writer_wakeup(nmp);
2048 			}
2049 		}
2050 		mtx_abort_link(&nmp->nm_rxlock, &rep->r_link);
2051 	}
2052 }
2053 
2054 /*
2055  * Test for a termination condition pending on the process.
2056  * This is used for NFSMNT_INT mounts.
2057  */
2058 int
nfs_sigintr(struct nfsmount * nmp,struct nfsreq * rep,struct thread * td)2059 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
2060 {
2061 	sigset_t tmpset;
2062 	struct proc *p;
2063 	struct lwp *lp;
2064 
2065 	if (rep && (rep->r_flags & R_SOFTTERM))
2066 		return (EINTR);
2067 	/* Terminate all requests while attempting a forced unmount. */
2068 	if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
2069 		return (EINTR);
2070 	if (!(nmp->nm_flag & NFSMNT_INT))
2071 		return (0);
2072 	/* td might be NULL YYY */
2073 	if (td == NULL || (p = td->td_proc) == NULL)
2074 		return (0);
2075 
2076 	lp = td->td_lwp;
2077 	tmpset = lwp_sigpend(lp);
2078 	SIGSETNAND(tmpset, lp->lwp_sigmask);
2079 	SIGSETNAND(tmpset, p->p_sigignore);
2080 	if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
2081 		return (EINTR);
2082 
2083 	return (0);
2084 }
2085 
2086 /*
2087  * Lock a socket against others.
2088  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
2089  * and also to avoid race conditions between the processes with nfs requests
2090  * in progress when a reconnect is necessary.
2091  */
2092 int
nfs_sndlock(struct nfsmount * nmp,struct nfsreq * rep)2093 nfs_sndlock(struct nfsmount *nmp, struct nfsreq *rep)
2094 {
2095 	mtx_t *mtx = &nmp->nm_txlock;
2096 	struct thread *td;
2097 	int slptimeo;
2098 	int slpflag;
2099 	int error;
2100 
2101 	slpflag = 0;
2102 	slptimeo = 0;
2103 	td = rep ? rep->r_td : NULL;
2104 	if (nmp->nm_flag & NFSMNT_INT)
2105 		slpflag = PCATCH;
2106 
2107 	while ((error = mtx_lock_ex_try(mtx)) != 0) {
2108 		if (nfs_sigintr(nmp, rep, td)) {
2109 			error = EINTR;
2110 			break;
2111 		}
2112 		error = mtx_lock_ex(mtx, slpflag, slptimeo);
2113 		if (error == 0)
2114 			break;
2115 		if (slpflag == PCATCH) {
2116 			slpflag = 0;
2117 			slptimeo = 2 * hz;
2118 		}
2119 	}
2120 	/* Always fail if our request has been cancelled. */
2121 	if (rep && (rep->r_flags & R_SOFTTERM)) {
2122 		if (error == 0)
2123 			mtx_unlock(mtx);
2124 		error = EINTR;
2125 	}
2126 	return (error);
2127 }
2128 
2129 /*
2130  * Unlock the stream socket for others.
2131  */
2132 void
nfs_sndunlock(struct nfsmount * nmp)2133 nfs_sndunlock(struct nfsmount *nmp)
2134 {
2135 	mtx_unlock(&nmp->nm_txlock);
2136 }
2137 
2138 /*
2139  * Lock the receiver side of the socket.
2140  *
2141  * rep may be NULL.
2142  */
2143 static int
nfs_rcvlock(struct nfsmount * nmp,struct nfsreq * rep)2144 nfs_rcvlock(struct nfsmount *nmp, struct nfsreq *rep)
2145 {
2146 	mtx_t *mtx = &nmp->nm_rxlock;
2147 	int slpflag;
2148 	int slptimeo;
2149 	int error;
2150 
2151 	/*
2152 	 * Unconditionally check for completion in case another nfsiod
2153 	 * get the packet while the caller was blocked, before the caller
2154 	 * called us.  Packet reception is handled by mainline code which
2155 	 * is protected by the BGL at the moment.
2156 	 *
2157 	 * We do not strictly need the second check just before the
2158 	 * tsleep(), but it's good defensive programming.
2159 	 */
2160 	if (rep && rep->r_mrep != NULL)
2161 		return (EALREADY);
2162 
2163 	if (nmp->nm_flag & NFSMNT_INT)
2164 		slpflag = PCATCH;
2165 	else
2166 		slpflag = 0;
2167 	slptimeo = 0;
2168 
2169 	while ((error = mtx_lock_ex_try(mtx)) != 0) {
2170 		if (nfs_sigintr(nmp, rep, (rep ? rep->r_td : NULL))) {
2171 			error = EINTR;
2172 			break;
2173 		}
2174 		if (rep && rep->r_mrep != NULL) {
2175 			error = EALREADY;
2176 			break;
2177 		}
2178 
2179 		/*
2180 		 * NOTE: can return ENOLCK, but in that case rep->r_mrep
2181 		 *       will already be set.
2182 		 */
2183 		if (rep) {
2184 			error = mtx_lock_ex_link(mtx, &rep->r_link,
2185 						 slpflag, slptimeo);
2186 		} else {
2187 			error = mtx_lock_ex(mtx, slpflag, slptimeo);
2188 		}
2189 		if (error == 0)
2190 			break;
2191 
2192 		/*
2193 		 * If our reply was recieved while we were sleeping,
2194 		 * then just return without taking the lock to avoid a
2195 		 * situation where a single iod could 'capture' the
2196 		 * recieve lock.
2197 		 */
2198 		if (rep && rep->r_mrep != NULL) {
2199 			error = EALREADY;
2200 			break;
2201 		}
2202 		if (slpflag == PCATCH) {
2203 			slpflag = 0;
2204 			slptimeo = 2 * hz;
2205 		}
2206 	}
2207 	if (error == 0) {
2208 		if (rep && rep->r_mrep != NULL) {
2209 			error = EALREADY;
2210 			mtx_unlock(mtx);
2211 		}
2212 	}
2213 	return (error);
2214 }
2215 
2216 /*
2217  * Unlock the stream socket for others.
2218  */
2219 static void
nfs_rcvunlock(struct nfsmount * nmp)2220 nfs_rcvunlock(struct nfsmount *nmp)
2221 {
2222 	mtx_unlock(&nmp->nm_rxlock);
2223 }
2224 
2225 /*
2226  * nfs_realign:
2227  *
2228  * Check for badly aligned mbuf data and realign by copying the unaligned
2229  * portion of the data into a new mbuf chain and freeing the portions
2230  * of the old chain that were replaced.
2231  *
2232  * We cannot simply realign the data within the existing mbuf chain
2233  * because the underlying buffers may contain other rpc commands and
2234  * we cannot afford to overwrite them.
2235  *
2236  * We would prefer to avoid this situation entirely.  The situation does
2237  * not occur with NFS/UDP and is supposed to only occassionally occur
2238  * with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
2239  *
2240  * NOTE!  M_NOWAIT cannot be used here.  The mbufs must be acquired
2241  *	  because the rpc request OR reply cannot be thrown away.  TCP NFS
2242  *	  mounts do not retry their RPCs unless the TCP connection itself
2243  *	  is dropped so throwing away a RPC will basically cause the NFS
2244  *	  operation to lockup indefinitely.
2245  */
2246 static void
nfs_realign(struct mbuf ** pm,int hsiz)2247 nfs_realign(struct mbuf **pm, int hsiz)
2248 {
2249 	struct mbuf *m;
2250 	struct mbuf *n = NULL;
2251 
2252 	/*
2253 	 * Check for misalignemnt
2254 	 */
2255 	++nfs_realign_test;
2256 	while ((m = *pm) != NULL) {
2257 		if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3))
2258 			break;
2259 		pm = &m->m_next;
2260 	}
2261 
2262 	/*
2263 	 * If misalignment found make a completely new copy.
2264 	 */
2265 	if (m) {
2266 		++nfs_realign_count;
2267 		n = m_dup_data(m, M_WAITOK);
2268 		m_freem(*pm);
2269 		*pm = n;
2270 	}
2271 }
2272 
2273 #ifndef NFS_NOSERVER
2274 
2275 /*
2276  * Parse an RPC request
2277  * - verify it
2278  * - fill in the cred struct.
2279  */
2280 int
nfs_getreq(struct nfsrv_descript * nd,struct nfsd * nfsd,int has_header)2281 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
2282 {
2283 	int len, i;
2284 	u_int32_t *tl;
2285 	struct uio uio;
2286 	struct iovec iov;
2287 	caddr_t cp;
2288 	u_int32_t nfsvers, auth_type;
2289 	uid_t nickuid;
2290 	int error = 0, ticklen;
2291 	struct nfsuid *nuidp;
2292 	struct timeval tvin, tvout;
2293 	struct nfsm_info info;
2294 #if 0				/* until encrypted keys are implemented */
2295 	NFSKERBKEYSCHED_T keys;	/* stores key schedule */
2296 #endif
2297 
2298 	info.mrep = nd->nd_mrep;
2299 	info.md = nd->nd_md;
2300 	info.dpos = nd->nd_dpos;
2301 
2302 	if (has_header) {
2303 		NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
2304 		nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
2305 		if (*tl++ != rpc_call) {
2306 			m_freem(info.mrep);
2307 			return (EBADRPC);
2308 		}
2309 	} else {
2310 		NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
2311 	}
2312 	nd->nd_repstat = 0;
2313 	nd->nd_flag = 0;
2314 	if (*tl++ != rpc_vers) {
2315 		nd->nd_repstat = ERPCMISMATCH;
2316 		nd->nd_procnum = NFSPROC_NOOP;
2317 		return (0);
2318 	}
2319 	if (*tl != nfs_prog) {
2320 		nd->nd_repstat = EPROGUNAVAIL;
2321 		nd->nd_procnum = NFSPROC_NOOP;
2322 		return (0);
2323 	}
2324 	tl++;
2325 	nfsvers = fxdr_unsigned(u_int32_t, *tl++);
2326 	if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
2327 		nd->nd_repstat = EPROGMISMATCH;
2328 		nd->nd_procnum = NFSPROC_NOOP;
2329 		return (0);
2330 	}
2331 	if (nfsvers == NFS_VER3)
2332 		nd->nd_flag = ND_NFSV3;
2333 	nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
2334 	if (nd->nd_procnum == NFSPROC_NULL)
2335 		return (0);
2336 	if (nd->nd_procnum >= NFS_NPROCS ||
2337 		(nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
2338 		(!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
2339 		nd->nd_repstat = EPROCUNAVAIL;
2340 		nd->nd_procnum = NFSPROC_NOOP;
2341 		return (0);
2342 	}
2343 	if ((nd->nd_flag & ND_NFSV3) == 0)
2344 		nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
2345 	auth_type = *tl++;
2346 	len = fxdr_unsigned(int, *tl++);
2347 	if (len < 0 || len > RPCAUTH_MAXSIZ) {
2348 		m_freem(info.mrep);
2349 		return (EBADRPC);
2350 	}
2351 
2352 	nd->nd_flag &= ~ND_KERBAUTH;
2353 	/*
2354 	 * Handle auth_unix or auth_kerb.
2355 	 */
2356 	if (auth_type == rpc_auth_unix) {
2357 		len = fxdr_unsigned(int, *++tl);
2358 		if (len < 0 || len > NFS_MAXNAMLEN) {
2359 			m_freem(info.mrep);
2360 			return (EBADRPC);
2361 		}
2362 		ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2363 		NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2364 		bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
2365 		nd->nd_cr.cr_ref = 1;
2366 		nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
2367 		nd->nd_cr.cr_ruid = nd->nd_cr.cr_svuid = nd->nd_cr.cr_uid;
2368 		nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
2369 		nd->nd_cr.cr_rgid = nd->nd_cr.cr_svgid = nd->nd_cr.cr_gid;
2370 		len = fxdr_unsigned(int, *tl);
2371 		if (len < 0 || len > RPCAUTH_UNIXGIDS) {
2372 			m_freem(info.mrep);
2373 			return (EBADRPC);
2374 		}
2375 		NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
2376 		for (i = 1; i <= len; i++)
2377 		    if (i < NGROUPS)
2378 			nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
2379 		    else
2380 			tl++;
2381 		nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
2382 		if (nd->nd_cr.cr_ngroups > 1)
2383 		    nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
2384 		len = fxdr_unsigned(int, *++tl);
2385 		if (len < 0 || len > RPCAUTH_MAXSIZ) {
2386 			m_freem(info.mrep);
2387 			return (EBADRPC);
2388 		}
2389 		if (len > 0) {
2390 			ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2391 		}
2392 	} else if (auth_type == rpc_auth_kerb) {
2393 		switch (fxdr_unsigned(int, *tl++)) {
2394 		case RPCAKN_FULLNAME:
2395 			ticklen = fxdr_unsigned(int, *tl);
2396 			*((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2397 			uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2398 			nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2399 			if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2400 				m_freem(info.mrep);
2401 				return (EBADRPC);
2402 			}
2403 			uio.uio_offset = 0;
2404 			uio.uio_iov = &iov;
2405 			uio.uio_iovcnt = 1;
2406 			uio.uio_segflg = UIO_SYSSPACE;
2407 			iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2408 			iov.iov_len = RPCAUTH_MAXSIZ - 4;
2409 			ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2410 			NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2411 			if (*tl++ != rpc_auth_kerb ||
2412 				fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2413 				kprintf("Bad kerb verifier\n");
2414 				nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2415 				nd->nd_procnum = NFSPROC_NOOP;
2416 				return (0);
2417 			}
2418 			NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2419 			tl = (u_int32_t *)cp;
2420 			if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2421 				kprintf("Not fullname kerb verifier\n");
2422 				nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2423 				nd->nd_procnum = NFSPROC_NOOP;
2424 				return (0);
2425 			}
2426 			cp += NFSX_UNSIGNED;
2427 			bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2428 			nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2429 			nd->nd_flag |= ND_KERBFULL;
2430 			nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2431 			break;
2432 		case RPCAKN_NICKNAME:
2433 			if (len != 2 * NFSX_UNSIGNED) {
2434 				kprintf("Kerb nickname short\n");
2435 				nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2436 				nd->nd_procnum = NFSPROC_NOOP;
2437 				return (0);
2438 			}
2439 			nickuid = fxdr_unsigned(uid_t, *tl);
2440 			NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2441 			if (*tl++ != rpc_auth_kerb ||
2442 				fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2443 				kprintf("Kerb nick verifier bad\n");
2444 				nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2445 				nd->nd_procnum = NFSPROC_NOOP;
2446 				return (0);
2447 			}
2448 			NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2449 			tvin.tv_sec = *tl++;
2450 			tvin.tv_usec = *tl;
2451 
2452 			for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2453 			    nuidp != NULL; nuidp = nuidp->nu_hash.le_next) {
2454 				if (nuidp->nu_cr.cr_uid == nickuid &&
2455 				    (!nd->nd_nam2 ||
2456 				     netaddr_match(AF_INET,
2457 				      &nuidp->nu_haddr, nd->nd_nam2)))
2458 					break;
2459 			}
2460 			if (!nuidp) {
2461 				nd->nd_repstat =
2462 					(NFSERR_AUTHERR|AUTH_REJECTCRED);
2463 				nd->nd_procnum = NFSPROC_NOOP;
2464 				return (0);
2465 			}
2466 
2467 			/*
2468 			 * Now, decrypt the timestamp using the session key
2469 			 * and validate it.
2470 			 */
2471 #ifdef NFSKERB
2472 			XXX
2473 #else
2474 			tvout.tv_sec = 0;
2475 			tvout.tv_usec = 0;
2476 #endif
2477 
2478 			tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2479 			tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2480 			if (nuidp->nu_expire != time_uptime ||
2481 			    nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2482 			    (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2483 			     nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2484 				nuidp->nu_expire = 0;
2485 				nd->nd_repstat =
2486 				    (NFSERR_AUTHERR|AUTH_REJECTVERF);
2487 				nd->nd_procnum = NFSPROC_NOOP;
2488 				return (0);
2489 			}
2490 			nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2491 			nd->nd_flag |= ND_KERBNICK;
2492 			break;
2493 		}
2494 	} else {
2495 		nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2496 		nd->nd_procnum = NFSPROC_NOOP;
2497 		return (0);
2498 	}
2499 
2500 	nd->nd_md = info.md;
2501 	nd->nd_dpos = info.dpos;
2502 	return (0);
2503 nfsmout:
2504 	return (error);
2505 }
2506 
2507 #endif
2508 
2509 /*
2510  * Send a message to the originating process's terminal.  The thread and/or
2511  * process may be NULL.  YYY the thread should not be NULL but there may
2512  * still be some uio_td's that are still being passed as NULL through to
2513  * nfsm_request().
2514  */
2515 static int
nfs_msg(struct thread * td,char * server,char * msg)2516 nfs_msg(struct thread *td, char *server, char *msg)
2517 {
2518 	tpr_t tpr;
2519 
2520 	if (td && td->td_proc)
2521 		tpr = tprintf_open(td->td_proc);
2522 	else
2523 		tpr = NULL;
2524 	tprintf(tpr, "nfs server %s: %s\n", server, msg);
2525 	tprintf_close(tpr);
2526 	return (0);
2527 }
2528 
2529 #ifndef NFS_NOSERVER
2530 
2531 /*
2532  * Socket upcall routine for nfsd sockets.  This runs in the protocol
2533  * thread and passes waitflag == M_NOWAIT.
2534  */
2535 void
nfsrv_rcv_upcall(struct socket * so,void * arg,int waitflag)2536 nfsrv_rcv_upcall(struct socket *so, void *arg, int waitflag)
2537 {
2538 	struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2539 
2540 	if (slp->ns_needq_upcall == 0) {
2541 		slp->ns_needq_upcall = 1;	/* ok to race */
2542 		lwkt_gettoken(&nfs_token);
2543 		nfsrv_wakenfsd(slp, 1);
2544 		lwkt_reltoken(&nfs_token);
2545 	}
2546 #if 0
2547 	lwkt_gettoken(&slp->ns_token);
2548 	slp->ns_flag |= SLP_NEEDQ;
2549 	nfsrv_rcv(so, arg, waitflag);
2550 	lwkt_reltoken(&slp->ns_token);
2551 #endif
2552 }
2553 
2554 /*
2555  * Process new data on a receive socket.  Essentially do as much as we can
2556  * non-blocking, else punt and it will be called with M_WAITOK from an nfsd.
2557  *
2558  * slp->ns_token is held on call
2559  */
2560 void
nfsrv_rcv(struct socket * so,void * arg,int waitflag)2561 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2562 {
2563 	struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2564 	struct mbuf *m;
2565 	struct sockaddr *nam;
2566 	struct sockbuf sio;
2567 	int flags, error;
2568 	int nparallel_wakeup = 0;
2569 
2570 	ASSERT_LWKT_TOKEN_HELD(&slp->ns_token);
2571 
2572 	if ((slp->ns_flag & SLP_VALID) == 0)
2573 		return;
2574 
2575 	/*
2576 	 * Do not allow an infinite number of completed RPC records to build
2577 	 * up before we stop reading data from the socket.  Otherwise we could
2578 	 * end up holding onto an unreasonable number of mbufs for requests
2579 	 * waiting for service.
2580 	 *
2581 	 * This should give pretty good feedback to the TCP layer and
2582 	 * prevents a memory crunch for other protocols.
2583 	 *
2584 	 * Note that the same service socket can be dispatched to several
2585 	 * nfs servers simultaniously.  The tcp protocol callback calls us
2586 	 * with M_NOWAIT.  nfsd calls us with M_WAITOK (typically).
2587 	 */
2588 	if (NFSRV_RECLIMIT(slp))
2589 		return;
2590 
2591 	/*
2592 	 * Handle protocol specifics to parse an RPC request.  We always
2593 	 * pull from the socket using non-blocking I/O.
2594 	 */
2595 	if (so->so_type == SOCK_STREAM) {
2596 		/*
2597 		 * The data has to be read in an orderly fashion from a TCP
2598 		 * stream, unlike a UDP socket.  It is possible for soreceive
2599 		 * and/or nfsrv_getstream() to block, so make sure only one
2600 		 * entity is messing around with the TCP stream at any given
2601 		 * moment.  The receive sockbuf's lock in soreceive is not
2602 		 * sufficient.
2603 		 */
2604 		if (slp->ns_flag & SLP_GETSTREAM)
2605 			return;
2606 		slp->ns_flag |= SLP_GETSTREAM;
2607 
2608 		/*
2609 		 * Do soreceive().  Pull out as much data as possible without
2610 		 * blocking.
2611 		 */
2612 		sbinit(&sio, 1000000000);
2613 		flags = MSG_DONTWAIT;
2614 		error = so_pru_soreceive(so, NULL, NULL, &sio, NULL, &flags);
2615 		if (error || sio.sb_mb == NULL) {
2616 			if (error != EWOULDBLOCK)
2617 				slp->ns_flag |= SLP_DISCONN;
2618 			slp->ns_flag &= ~(SLP_GETSTREAM | SLP_NEEDQ);
2619 			goto done;
2620 		}
2621 		m = sio.sb_mb;
2622 		if (slp->ns_rawend) {
2623 			slp->ns_rawend->m_next = m;
2624 			slp->ns_cc += sio.sb_cc;
2625 		} else {
2626 			slp->ns_raw = m;
2627 			slp->ns_cc = sio.sb_cc;
2628 		}
2629 		while (m->m_next)
2630 			m = m->m_next;
2631 		slp->ns_rawend = m;
2632 
2633 		/*
2634 		 * Now try and parse as many record(s) as we can out of the
2635 		 * raw stream data.  This will set SLP_DOREC.
2636 		 */
2637 		error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2638 		if (error && error != EWOULDBLOCK)
2639 			slp->ns_flag |= SLP_DISCONN;
2640 		slp->ns_flag &= ~SLP_GETSTREAM;
2641 	} else {
2642 		/*
2643 		 * For UDP soreceive typically pulls just one packet, loop
2644 		 * to get the whole batch.
2645 		 */
2646 		do {
2647 			sbinit(&sio, 1000000000);
2648 			flags = MSG_DONTWAIT;
2649 			error = so_pru_soreceive(so, &nam, NULL, &sio,
2650 						 NULL, &flags);
2651 			if (sio.sb_mb) {
2652 				struct nfsrv_rec *rec;
2653 				int mf = (waitflag & M_NOWAIT) ?
2654 					    M_NOWAIT : M_WAITOK;
2655 				rec = kmalloc(sizeof(struct nfsrv_rec),
2656 					     M_NFSRVDESC, mf);
2657 				if (!rec) {
2658 					if (nam)
2659 						kfree(nam, M_SONAME);
2660 					m_freem(sio.sb_mb);
2661 					continue;
2662 				}
2663 				nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2664 				rec->nr_address = nam;
2665 				rec->nr_packet = sio.sb_mb;
2666 				STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2667 				++slp->ns_numrec;
2668 				slp->ns_flag |= SLP_DOREC;
2669 				++nparallel_wakeup;
2670 			} else {
2671 				slp->ns_flag &= ~SLP_NEEDQ;
2672 			}
2673 			if (error) {
2674 				if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2675 				    && error != EWOULDBLOCK) {
2676 					slp->ns_flag |= SLP_DISCONN;
2677 					break;
2678 				}
2679 			}
2680 			if (NFSRV_RECLIMIT(slp))
2681 				break;
2682 		} while (sio.sb_mb);
2683 	}
2684 
2685 	/*
2686 	 * If we were upcalled from the tcp protocol layer and we have
2687 	 * fully parsed records ready to go, or there is new data pending,
2688 	 * or something went wrong, try to wake up a nfsd thread to deal
2689 	 * with it.
2690 	 */
2691 done:
2692 	/* XXX this code is currently not executed (nfsrv_rcv_upcall) */
2693 	if (waitflag == M_NOWAIT && (slp->ns_flag & SLP_ACTION_MASK)) {
2694 		lwkt_gettoken(&nfs_token);
2695 		nfsrv_wakenfsd(slp, nparallel_wakeup);
2696 		lwkt_reltoken(&nfs_token);
2697 	}
2698 }
2699 
2700 /*
2701  * Try and extract an RPC request from the mbuf data list received on a
2702  * stream socket. The "waitflag" argument indicates whether or not it
2703  * can sleep.
2704  */
2705 static int
nfsrv_getstream(struct nfssvc_sock * slp,int waitflag,int * countp)2706 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2707 {
2708 	struct mbuf *m, **mpp;
2709 	char *cp1, *cp2;
2710 	int len;
2711 	struct mbuf *om, *m2, *recm;
2712 	u_int32_t recmark;
2713 
2714 	for (;;) {
2715 	    if (slp->ns_reclen == 0) {
2716 		if (slp->ns_cc < NFSX_UNSIGNED)
2717 			return (0);
2718 		m = slp->ns_raw;
2719 		if (m->m_len >= NFSX_UNSIGNED) {
2720 			bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2721 			m->m_data += NFSX_UNSIGNED;
2722 			m->m_len -= NFSX_UNSIGNED;
2723 		} else {
2724 			cp1 = (caddr_t)&recmark;
2725 			cp2 = mtod(m, caddr_t);
2726 			while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2727 				while (m->m_len == 0) {
2728 					m = m->m_next;
2729 					cp2 = mtod(m, caddr_t);
2730 				}
2731 				*cp1++ = *cp2++;
2732 				m->m_data++;
2733 				m->m_len--;
2734 			}
2735 		}
2736 		slp->ns_cc -= NFSX_UNSIGNED;
2737 		recmark = ntohl(recmark);
2738 		slp->ns_reclen = recmark & ~0x80000000;
2739 		if (recmark & 0x80000000)
2740 			slp->ns_flag |= SLP_LASTFRAG;
2741 		else
2742 			slp->ns_flag &= ~SLP_LASTFRAG;
2743 		if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2744 			log(LOG_ERR, "%s (%d) from nfs client\n",
2745 			    "impossible packet length",
2746 			    slp->ns_reclen);
2747 			return (EPERM);
2748 		}
2749 	    }
2750 
2751 	    /*
2752 	     * Now get the record part.
2753 	     *
2754 	     * Note that slp->ns_reclen may be 0.  Linux sometimes
2755 	     * generates 0-length RPCs
2756 	     */
2757 	    recm = NULL;
2758 	    if (slp->ns_cc == slp->ns_reclen) {
2759 		recm = slp->ns_raw;
2760 		slp->ns_raw = slp->ns_rawend = NULL;
2761 		slp->ns_cc = slp->ns_reclen = 0;
2762 	    } else if (slp->ns_cc > slp->ns_reclen) {
2763 		len = 0;
2764 		m = slp->ns_raw;
2765 		om = NULL;
2766 
2767 		while (len < slp->ns_reclen) {
2768 			if ((len + m->m_len) > slp->ns_reclen) {
2769 				m2 = m_copym(m, 0, slp->ns_reclen - len,
2770 					waitflag);
2771 				if (m2) {
2772 					if (om) {
2773 						om->m_next = m2;
2774 						recm = slp->ns_raw;
2775 					} else
2776 						recm = m2;
2777 					m->m_data += slp->ns_reclen - len;
2778 					m->m_len -= slp->ns_reclen - len;
2779 					len = slp->ns_reclen;
2780 				} else {
2781 					return (EWOULDBLOCK);
2782 				}
2783 			} else if ((len + m->m_len) == slp->ns_reclen) {
2784 				om = m;
2785 				len += m->m_len;
2786 				m = m->m_next;
2787 				recm = slp->ns_raw;
2788 				om->m_next = NULL;
2789 			} else {
2790 				om = m;
2791 				len += m->m_len;
2792 				m = m->m_next;
2793 			}
2794 		}
2795 		slp->ns_raw = m;
2796 		slp->ns_cc -= len;
2797 		slp->ns_reclen = 0;
2798 	    } else {
2799 		return (0);
2800 	    }
2801 
2802 	    /*
2803 	     * Accumulate the fragments into a record.
2804 	     */
2805 	    mpp = &slp->ns_frag;
2806 	    while (*mpp)
2807 		mpp = &((*mpp)->m_next);
2808 	    *mpp = recm;
2809 	    if (slp->ns_flag & SLP_LASTFRAG) {
2810 		struct nfsrv_rec *rec;
2811 		int mf = (waitflag & M_NOWAIT) ? M_NOWAIT : M_WAITOK;
2812 		rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2813 		if (!rec) {
2814 		    m_freem(slp->ns_frag);
2815 		} else {
2816 		    nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2817 		    rec->nr_address = NULL;
2818 		    rec->nr_packet = slp->ns_frag;
2819 		    STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2820 		    ++slp->ns_numrec;
2821 		    slp->ns_flag |= SLP_DOREC;
2822 		    ++*countp;
2823 		}
2824 		slp->ns_frag = NULL;
2825 	    }
2826 	}
2827 }
2828 
2829 #ifdef INVARIANTS
2830 
2831 /*
2832  * Sanity check our mbuf chain.
2833  */
2834 static void
nfs_checkpkt(struct mbuf * m,int len)2835 nfs_checkpkt(struct mbuf *m, int len)
2836 {
2837 	int xlen = 0;
2838 	while (m) {
2839 		xlen += m->m_len;
2840 		m = m->m_next;
2841 	}
2842 	if (xlen != len) {
2843 		panic("nfs_checkpkt: len mismatch %d/%d mbuf %p",
2844 			xlen, len, m);
2845 	}
2846 }
2847 
2848 #else
2849 
2850 static void
nfs_checkpkt(struct mbuf * m __unused,int len __unused)2851 nfs_checkpkt(struct mbuf *m __unused, int len __unused)
2852 {
2853 }
2854 
2855 #endif
2856 
2857 /*
2858  * Parse an RPC header.
2859  *
2860  * If the socket is invalid or no records are pending we return ENOBUFS.
2861  * The caller must deal with NEEDQ races.
2862  */
2863 int
nfsrv_dorec(struct nfssvc_sock * slp,struct nfsd * nfsd,struct nfsrv_descript ** ndp)2864 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2865 	    struct nfsrv_descript **ndp)
2866 {
2867 	struct nfsrv_rec *rec;
2868 	struct mbuf *m;
2869 	struct sockaddr *nam;
2870 	struct nfsrv_descript *nd;
2871 	int error;
2872 
2873 	*ndp = NULL;
2874 	if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2875 		return (ENOBUFS);
2876 	rec = STAILQ_FIRST(&slp->ns_rec);
2877 	STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2878 	KKASSERT(slp->ns_numrec > 0);
2879 	if (--slp->ns_numrec == 0)
2880 		slp->ns_flag &= ~SLP_DOREC;
2881 	nam = rec->nr_address;
2882 	m = rec->nr_packet;
2883 	kfree(rec, M_NFSRVDESC);
2884 	nd = kmalloc(sizeof(struct nfsrv_descript), M_NFSRVDESC, M_WAITOK);
2885 	nd->nd_md = nd->nd_mrep = m;
2886 	nd->nd_nam2 = nam;
2887 	nd->nd_dpos = mtod(m, caddr_t);
2888 	error = nfs_getreq(nd, nfsd, TRUE);
2889 	if (error) {
2890 		if (nam) {
2891 			kfree(nam, M_SONAME);
2892 		}
2893 		kfree((caddr_t)nd, M_NFSRVDESC);
2894 		return (error);
2895 	}
2896 	*ndp = nd;
2897 	nfsd->nfsd_nd = nd;
2898 	return (0);
2899 }
2900 
2901 /*
2902  * Try to assign service sockets to nfsd threads based on the number
2903  * of new rpc requests that have been queued on the service socket.
2904  *
2905  * If no nfsd's are available or additonal requests are pending, set the
2906  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2907  * the work in the nfssvc_sock list when it is finished processing its
2908  * current work.  This flag is only cleared when an nfsd can not find
2909  * any new work to perform.
2910  */
2911 void
nfsrv_wakenfsd(struct nfssvc_sock * slp,int nparallel)2912 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2913 {
2914 	struct nfsd *nd;
2915 
2916 	if ((slp->ns_flag & SLP_VALID) == 0)
2917 		return;
2918 	if (nparallel <= 1)
2919 		nparallel = 1;
2920 	TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2921 		if (nd->nfsd_flag & NFSD_WAITING) {
2922 			nd->nfsd_flag &= ~NFSD_WAITING;
2923 			if (nd->nfsd_slp)
2924 				panic("nfsd wakeup");
2925 			nfsrv_slpref(slp);
2926 			nd->nfsd_slp = slp;
2927 			wakeup((caddr_t)nd);
2928 			if (--nparallel == 0)
2929 				break;
2930 		}
2931 	}
2932 
2933 	/*
2934 	 * If we couldn't assign slp then the NFSDs are all busy and
2935 	 * we set a flag indicating that there is pending work.
2936 	 */
2937 	if (nparallel)
2938 		nfsd_head_flag |= NFSD_CHECKSLP;
2939 }
2940 #endif /* NFS_NOSERVER */
2941