network - Tokenize NFS, fix MP races
[dragonfly.git] / sys / vfs / nfs / nfs_socket.c
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 #include <sys/socketvar2.h>
69
70 #include <netinet/in.h>
71 #include <netinet/tcp.h>
72 #include <sys/thread2.h>
73
74 #include "rpcv2.h"
75 #include "nfsproto.h"
76 #include "nfs.h"
77 #include "xdr_subs.h"
78 #include "nfsm_subs.h"
79 #include "nfsmount.h"
80 #include "nfsnode.h"
81 #include "nfsrtt.h"
82
83 #define TRUE    1
84 #define FALSE   0
85
86 /*
87  * RTT calculations are scaled by 256 (8 bits).  A proper fractional
88  * RTT will still be calculated even with a slow NFS timer.
89  */
90 #define NFS_SRTT(r)     (r)->r_nmp->nm_srtt[proct[(r)->r_procnum]]
91 #define NFS_SDRTT(r)    (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum]]
92 #define NFS_RTT_SCALE_BITS      8       /* bits */
93 #define NFS_RTT_SCALE           256     /* value */
94
95 /*
96  * Defines which timer to use for the procnum.
97  * 0 - default
98  * 1 - getattr
99  * 2 - lookup
100  * 3 - read
101  * 4 - write
102  */
103 static int proct[NFS_NPROCS] = {
104         0, 1, 0, 2, 1, 3, 3, 4, 0, 0,   /* 00-09        */
105         0, 0, 0, 0, 0, 0, 3, 3, 0, 0,   /* 10-19        */
106         0, 5, 0, 0, 0, 0,               /* 20-29        */
107 };
108
109 static int multt[NFS_NPROCS] = {
110         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 00-09        */
111         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 10-19        */
112         1, 2, 1, 1, 1, 1,               /* 20-29        */
113 };
114
115 static int nfs_backoff[8] = { 2, 3, 5, 8, 13, 21, 34, 55 };
116 static int nfs_realign_test;
117 static int nfs_realign_count;
118 static int nfs_showrtt;
119 static int nfs_showrexmit;
120 int nfs_maxasyncbio = NFS_MAXASYNCBIO;
121
122 SYSCTL_DECL(_vfs_nfs);
123
124 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0, "");
125 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0, "");
126 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrtt, CTLFLAG_RW, &nfs_showrtt, 0, "");
127 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrexmit, CTLFLAG_RW, &nfs_showrexmit, 0, "");
128 SYSCTL_INT(_vfs_nfs, OID_AUTO, maxasyncbio, CTLFLAG_RW, &nfs_maxasyncbio, 0, "");
129
130 static int nfs_request_setup(nfsm_info_t info);
131 static int nfs_request_auth(struct nfsreq *rep);
132 static int nfs_request_try(struct nfsreq *rep);
133 static int nfs_request_waitreply(struct nfsreq *rep);
134 static int nfs_request_processreply(nfsm_info_t info, int);
135
136 int nfsrtton = 0;
137 struct nfsrtt nfsrtt;
138 struct callout  nfs_timer_handle;
139
140 static int      nfs_msg (struct thread *,char *,char *);
141 static int      nfs_rcvlock (struct nfsmount *nmp, struct nfsreq *myreq);
142 static void     nfs_rcvunlock (struct nfsmount *nmp);
143 static void     nfs_realign (struct mbuf **pm, int hsiz);
144 static int      nfs_receive (struct nfsmount *nmp, struct nfsreq *rep,
145                                 struct sockaddr **aname, struct mbuf **mp);
146 static void     nfs_softterm (struct nfsreq *rep, int islocked);
147 static void     nfs_hardterm (struct nfsreq *rep, int islocked);
148 static int      nfs_reconnect (struct nfsmount *nmp, struct nfsreq *rep);
149 #ifndef NFS_NOSERVER 
150 static int      nfsrv_getstream (struct nfssvc_sock *, int, int *);
151 static void     nfs_timer_req(struct nfsreq *req);
152 static void     nfs_checkpkt(struct mbuf *m, int len);
153
154 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
155                                     struct nfssvc_sock *slp,
156                                     struct thread *td,
157                                     struct mbuf **mreqp) = {
158         nfsrv_null,
159         nfsrv_getattr,
160         nfsrv_setattr,
161         nfsrv_lookup,
162         nfsrv3_access,
163         nfsrv_readlink,
164         nfsrv_read,
165         nfsrv_write,
166         nfsrv_create,
167         nfsrv_mkdir,
168         nfsrv_symlink,
169         nfsrv_mknod,
170         nfsrv_remove,
171         nfsrv_rmdir,
172         nfsrv_rename,
173         nfsrv_link,
174         nfsrv_readdir,
175         nfsrv_readdirplus,
176         nfsrv_statfs,
177         nfsrv_fsinfo,
178         nfsrv_pathconf,
179         nfsrv_commit,
180         nfsrv_noop,
181         nfsrv_noop,
182         nfsrv_noop,
183         nfsrv_noop
184 };
185 #endif /* NFS_NOSERVER */
186
187 /*
188  * Initialize sockets and congestion for a new NFS connection.
189  * We do not free the sockaddr if error.
190  */
191 int
192 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
193 {
194         struct socket *so;
195         int error;
196         struct sockaddr *saddr;
197         struct sockaddr_in *sin;
198         struct thread *td = &thread0; /* only used for socreate and sobind */
199
200         nmp->nm_so = so = NULL;
201         if (nmp->nm_flag & NFSMNT_FORCE)
202                 return (EINVAL);
203         saddr = nmp->nm_nam;
204         error = socreate(saddr->sa_family, &so, nmp->nm_sotype,
205                 nmp->nm_soproto, td);
206         if (error)
207                 goto bad;
208         nmp->nm_soflags = so->so_proto->pr_flags;
209
210         /*
211          * Some servers require that the client port be a reserved port number.
212          */
213         if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
214                 struct sockopt sopt;
215                 int ip;
216                 struct sockaddr_in ssin;
217
218                 bzero(&sopt, sizeof sopt);
219                 ip = IP_PORTRANGE_LOW;
220                 sopt.sopt_level = IPPROTO_IP;
221                 sopt.sopt_name = IP_PORTRANGE;
222                 sopt.sopt_val = (void *)&ip;
223                 sopt.sopt_valsize = sizeof(ip);
224                 sopt.sopt_td = NULL;
225                 error = sosetopt(so, &sopt);
226                 if (error)
227                         goto bad;
228                 bzero(&ssin, sizeof ssin);
229                 sin = &ssin;
230                 sin->sin_len = sizeof (struct sockaddr_in);
231                 sin->sin_family = AF_INET;
232                 sin->sin_addr.s_addr = INADDR_ANY;
233                 sin->sin_port = htons(0);
234                 error = sobind(so, (struct sockaddr *)sin, td);
235                 if (error)
236                         goto bad;
237                 bzero(&sopt, sizeof sopt);
238                 ip = IP_PORTRANGE_DEFAULT;
239                 sopt.sopt_level = IPPROTO_IP;
240                 sopt.sopt_name = IP_PORTRANGE;
241                 sopt.sopt_val = (void *)&ip;
242                 sopt.sopt_valsize = sizeof(ip);
243                 sopt.sopt_td = NULL;
244                 error = sosetopt(so, &sopt);
245                 if (error)
246                         goto bad;
247         }
248
249         /*
250          * Protocols that do not require connections may be optionally left
251          * unconnected for servers that reply from a port other than NFS_PORT.
252          */
253         if (nmp->nm_flag & NFSMNT_NOCONN) {
254                 if (nmp->nm_soflags & PR_CONNREQUIRED) {
255                         error = ENOTCONN;
256                         goto bad;
257                 }
258         } else {
259                 error = soconnect(so, nmp->nm_nam, td);
260                 if (error)
261                         goto bad;
262
263                 /*
264                  * Wait for the connection to complete. Cribbed from the
265                  * connect system call but with the wait timing out so
266                  * that interruptible mounts don't hang here for a long time.
267                  */
268                 crit_enter();
269                 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
270                         (void) tsleep((caddr_t)&so->so_timeo, 0,
271                                 "nfscon", 2 * hz);
272                         if ((so->so_state & SS_ISCONNECTING) &&
273                             so->so_error == 0 && rep &&
274                             (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
275                                 soclrstate(so, SS_ISCONNECTING);
276                                 crit_exit();
277                                 goto bad;
278                         }
279                 }
280                 if (so->so_error) {
281                         error = so->so_error;
282                         so->so_error = 0;
283                         crit_exit();
284                         goto bad;
285                 }
286                 crit_exit();
287         }
288         so->so_rcv.ssb_timeo = (5 * hz);
289         so->so_snd.ssb_timeo = (5 * hz);
290
291         /*
292          * Get buffer reservation size from sysctl, but impose reasonable
293          * limits.
294          */
295         if (nmp->nm_sotype == SOCK_STREAM) {
296                 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
297                         struct sockopt sopt;
298                         int val;
299
300                         bzero(&sopt, sizeof sopt);
301                         sopt.sopt_level = SOL_SOCKET;
302                         sopt.sopt_name = SO_KEEPALIVE;
303                         sopt.sopt_val = &val;
304                         sopt.sopt_valsize = sizeof val;
305                         val = 1;
306                         sosetopt(so, &sopt);
307                 }
308                 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
309                         struct sockopt sopt;
310                         int val;
311
312                         bzero(&sopt, sizeof sopt);
313                         sopt.sopt_level = IPPROTO_TCP;
314                         sopt.sopt_name = TCP_NODELAY;
315                         sopt.sopt_val = &val;
316                         sopt.sopt_valsize = sizeof val;
317                         val = 1;
318                         sosetopt(so, &sopt);
319                 }
320         }
321         error = soreserve(so, nfs_soreserve, nfs_soreserve, NULL);
322         if (error)
323                 goto bad;
324         atomic_set_int(&so->so_rcv.ssb_flags, SSB_NOINTR);
325         atomic_set_int(&so->so_snd.ssb_flags, SSB_NOINTR);
326
327         /* Initialize other non-zero congestion variables */
328         nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] = 
329                 nmp->nm_srtt[3] = (NFS_TIMEO << NFS_RTT_SCALE_BITS);
330         nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
331                 nmp->nm_sdrtt[3] = 0;
332         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
333         nmp->nm_timeouts = 0;
334
335         /*
336          * Assign nm_so last.  The moment nm_so is assigned the nfs_timer()
337          * can mess with the socket.
338          */
339         nmp->nm_so = so;
340         return (0);
341
342 bad:
343         if (so) {
344                 soshutdown(so, SHUT_RDWR);
345                 soclose(so, FNONBLOCK);
346         }
347         return (error);
348 }
349
350 /*
351  * Reconnect routine:
352  * Called when a connection is broken on a reliable protocol.
353  * - clean up the old socket
354  * - nfs_connect() again
355  * - set R_NEEDSXMIT for all outstanding requests on mount point
356  * If this fails the mount point is DEAD!
357  * nb: Must be called with the nfs_sndlock() set on the mount point.
358  */
359 static int
360 nfs_reconnect(struct nfsmount *nmp, struct nfsreq *rep)
361 {
362         struct nfsreq *req;
363         int error;
364
365         nfs_disconnect(nmp);
366         if (nmp->nm_rxstate >= NFSSVC_STOPPING)
367                 return (EINTR);
368         while ((error = nfs_connect(nmp, rep)) != 0) {
369                 if (error == EINTR || error == ERESTART)
370                         return (EINTR);
371                 if (error == EINVAL)
372                         return (error);
373                 if (nmp->nm_rxstate >= NFSSVC_STOPPING)
374                         return (EINTR);
375                 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
376         }
377
378         /*
379          * Loop through outstanding request list and fix up all requests
380          * on old socket.
381          */
382         crit_enter();
383         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
384                 KKASSERT(req->r_nmp == nmp);
385                 req->r_flags |= R_NEEDSXMIT;
386         }
387         crit_exit();
388         return (0);
389 }
390
391 /*
392  * NFS disconnect. Clean up and unlink.
393  */
394 void
395 nfs_disconnect(struct nfsmount *nmp)
396 {
397         struct socket *so;
398
399         if (nmp->nm_so) {
400                 so = nmp->nm_so;
401                 nmp->nm_so = NULL;
402                 soshutdown(so, SHUT_RDWR);
403                 soclose(so, FNONBLOCK);
404         }
405 }
406
407 void
408 nfs_safedisconnect(struct nfsmount *nmp)
409 {
410         nfs_rcvlock(nmp, NULL);
411         nfs_disconnect(nmp);
412         nfs_rcvunlock(nmp);
413 }
414
415 /*
416  * This is the nfs send routine. For connection based socket types, it
417  * must be called with an nfs_sndlock() on the socket.
418  * "rep == NULL" indicates that it has been called from a server.
419  * For the client side:
420  * - return EINTR if the RPC is terminated, 0 otherwise
421  * - set R_NEEDSXMIT if the send fails for any reason
422  * - do any cleanup required by recoverable socket errors (?)
423  * For the server side:
424  * - return EINTR or ERESTART if interrupted by a signal
425  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
426  * - do any cleanup required by recoverable socket errors (?)
427  */
428 int
429 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
430          struct nfsreq *rep)
431 {
432         struct sockaddr *sendnam;
433         int error, soflags, flags;
434
435         if (rep) {
436                 if (rep->r_flags & R_SOFTTERM) {
437                         m_freem(top);
438                         return (EINTR);
439                 }
440                 if ((so = rep->r_nmp->nm_so) == NULL) {
441                         rep->r_flags |= R_NEEDSXMIT;
442                         m_freem(top);
443                         return (0);
444                 }
445                 rep->r_flags &= ~R_NEEDSXMIT;
446                 soflags = rep->r_nmp->nm_soflags;
447         } else {
448                 soflags = so->so_proto->pr_flags;
449         }
450         if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
451                 sendnam = NULL;
452         else
453                 sendnam = nam;
454         if (so->so_type == SOCK_SEQPACKET)
455                 flags = MSG_EOR;
456         else
457                 flags = 0;
458
459         /*
460          * calls pru_sosend -> sosend -> so_pru_send -> netrpc
461          */
462         error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
463                               curthread /*XXX*/);
464         /*
465          * ENOBUFS for dgram sockets is transient and non fatal.
466          * No need to log, and no need to break a soft mount.
467          */
468         if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
469                 error = 0;
470                 /*
471                  * do backoff retransmit on client
472                  */
473                 if (rep) {
474                         if ((rep->r_nmp->nm_state & NFSSTA_SENDSPACE) == 0) {
475                                 rep->r_nmp->nm_state |= NFSSTA_SENDSPACE;
476                                 kprintf("Warning: NFS: Insufficient sendspace "
477                                         "(%lu),\n"
478                                         "\t You must increase vfs.nfs.soreserve"
479                                         "or decrease vfs.nfs.maxasyncbio\n",
480                                         so->so_snd.ssb_hiwat);
481                         }
482                         rep->r_flags |= R_NEEDSXMIT;
483                 }
484         }
485
486         if (error) {
487                 if (rep) {
488                         log(LOG_INFO, "nfs send error %d for server %s\n",error,
489                             rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
490                         /*
491                          * Deal with errors for the client side.
492                          */
493                         if (rep->r_flags & R_SOFTTERM)
494                                 error = EINTR;
495                         else
496                                 rep->r_flags |= R_NEEDSXMIT;
497                 } else {
498                         log(LOG_INFO, "nfsd send error %d\n", error);
499                 }
500
501                 /*
502                  * Handle any recoverable (soft) socket errors here. (?)
503                  */
504                 if (error != EINTR && error != ERESTART &&
505                         error != EWOULDBLOCK && error != EPIPE)
506                         error = 0;
507         }
508         return (error);
509 }
510
511 /*
512  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
513  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
514  * Mark and consolidate the data into a new mbuf list.
515  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
516  *     small mbufs.
517  * For SOCK_STREAM we must be very careful to read an entire record once
518  * we have read any of it, even if the system call has been interrupted.
519  */
520 static int
521 nfs_receive(struct nfsmount *nmp, struct nfsreq *rep,
522             struct sockaddr **aname, struct mbuf **mp)
523 {
524         struct socket *so;
525         struct sockbuf sio;
526         struct uio auio;
527         struct iovec aio;
528         struct mbuf *m;
529         struct mbuf *control;
530         u_int32_t len;
531         struct sockaddr **getnam;
532         int error, sotype, rcvflg;
533         struct thread *td = curthread;  /* XXX */
534
535         /*
536          * Set up arguments for soreceive()
537          */
538         *mp = NULL;
539         *aname = NULL;
540         sotype = nmp->nm_sotype;
541
542         /*
543          * For reliable protocols, lock against other senders/receivers
544          * in case a reconnect is necessary.
545          * For SOCK_STREAM, first get the Record Mark to find out how much
546          * more there is to get.
547          * We must lock the socket against other receivers
548          * until we have an entire rpc request/reply.
549          */
550         if (sotype != SOCK_DGRAM) {
551                 error = nfs_sndlock(nmp, rep);
552                 if (error)
553                         return (error);
554 tryagain:
555                 /*
556                  * Check for fatal errors and resending request.
557                  */
558                 /*
559                  * Ugh: If a reconnect attempt just happened, nm_so
560                  * would have changed. NULL indicates a failed
561                  * attempt that has essentially shut down this
562                  * mount point.
563                  */
564                 if (rep && (rep->r_mrep || (rep->r_flags & R_SOFTTERM))) {
565                         nfs_sndunlock(nmp);
566                         return (EINTR);
567                 }
568                 so = nmp->nm_so;
569                 if (so == NULL) {
570                         error = nfs_reconnect(nmp, rep);
571                         if (error) {
572                                 nfs_sndunlock(nmp);
573                                 return (error);
574                         }
575                         goto tryagain;
576                 }
577                 while (rep && (rep->r_flags & R_NEEDSXMIT)) {
578                         m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
579                         nfsstats.rpcretries++;
580                         error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
581                         if (error) {
582                                 if (error == EINTR || error == ERESTART ||
583                                     (error = nfs_reconnect(nmp, rep)) != 0) {
584                                         nfs_sndunlock(nmp);
585                                         return (error);
586                                 }
587                                 goto tryagain;
588                         }
589                 }
590                 nfs_sndunlock(nmp);
591                 if (sotype == SOCK_STREAM) {
592                         /*
593                          * Get the length marker from the stream
594                          */
595                         aio.iov_base = (caddr_t)&len;
596                         aio.iov_len = sizeof(u_int32_t);
597                         auio.uio_iov = &aio;
598                         auio.uio_iovcnt = 1;
599                         auio.uio_segflg = UIO_SYSSPACE;
600                         auio.uio_rw = UIO_READ;
601                         auio.uio_offset = 0;
602                         auio.uio_resid = sizeof(u_int32_t);
603                         auio.uio_td = td;
604                         do {
605                            rcvflg = MSG_WAITALL;
606                            error = so_pru_soreceive(so, NULL, &auio, NULL,
607                                                     NULL, &rcvflg);
608                            if (error == EWOULDBLOCK && rep) {
609                                 if (rep->r_flags & R_SOFTTERM)
610                                         return (EINTR);
611                            }
612                         } while (error == EWOULDBLOCK);
613
614                         if (error == 0 && auio.uio_resid > 0) {
615                             /*
616                              * Only log short packets if not EOF
617                              */
618                             if (auio.uio_resid != sizeof(u_int32_t))
619                             log(LOG_INFO,
620                                  "short receive (%d/%d) from nfs server %s\n",
621                                  (int)(sizeof(u_int32_t) - auio.uio_resid),
622                                  (int)sizeof(u_int32_t),
623                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
624                             error = EPIPE;
625                         }
626                         if (error)
627                                 goto errout;
628                         len = ntohl(len) & ~0x80000000;
629                         /*
630                          * This is SERIOUS! We are out of sync with the sender
631                          * and forcing a disconnect/reconnect is all I can do.
632                          */
633                         if (len > NFS_MAXPACKET) {
634                             log(LOG_ERR, "%s (%d) from nfs server %s\n",
635                                 "impossible packet length",
636                                 len,
637                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
638                             error = EFBIG;
639                             goto errout;
640                         }
641
642                         /*
643                          * Get the rest of the packet as an mbuf chain
644                          */
645                         sbinit(&sio, len);
646                         do {
647                             rcvflg = MSG_WAITALL;
648                             error = so_pru_soreceive(so, NULL, NULL, &sio,
649                                                      NULL, &rcvflg);
650                         } while (error == EWOULDBLOCK || error == EINTR ||
651                                  error == ERESTART);
652                         if (error == 0 && sio.sb_cc != len) {
653                             if (sio.sb_cc != 0)
654                             log(LOG_INFO,
655                                 "short receive (%zu/%d) from nfs server %s\n",
656                                 (size_t)len - auio.uio_resid, len,
657                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
658                             error = EPIPE;
659                         }
660                         *mp = sio.sb_mb;
661                 } else {
662                         /*
663                          * Non-stream, so get the whole packet by not
664                          * specifying MSG_WAITALL and by specifying a large
665                          * length.
666                          *
667                          * We have no use for control msg., but must grab them
668                          * and then throw them away so we know what is going
669                          * on.
670                          */
671                         sbinit(&sio, 100000000);
672                         do {
673                             rcvflg = 0;
674                             error =  so_pru_soreceive(so, NULL, NULL, &sio,
675                                                       &control, &rcvflg);
676                             if (control)
677                                 m_freem(control);
678                             if (error == EWOULDBLOCK && rep) {
679                                 if (rep->r_flags & R_SOFTTERM) {
680                                         m_freem(sio.sb_mb);
681                                         return (EINTR);
682                                 }
683                             }
684                         } while (error == EWOULDBLOCK ||
685                                  (error == 0 && sio.sb_mb == NULL && control));
686                         if ((rcvflg & MSG_EOR) == 0)
687                                 kprintf("Egad!!\n");
688                         if (error == 0 && sio.sb_mb == NULL)
689                                 error = EPIPE;
690                         len = sio.sb_cc;
691                         *mp = sio.sb_mb;
692                 }
693 errout:
694                 if (error && error != EINTR && error != ERESTART) {
695                         m_freem(*mp);
696                         *mp = NULL;
697                         if (error != EPIPE) {
698                                 log(LOG_INFO,
699                                     "receive error %d from nfs server %s\n",
700                                     error,
701                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
702                         }
703                         error = nfs_sndlock(nmp, rep);
704                         if (!error) {
705                                 error = nfs_reconnect(nmp, rep);
706                                 if (!error)
707                                         goto tryagain;
708                                 else
709                                         nfs_sndunlock(nmp);
710                         }
711                 }
712         } else {
713                 if ((so = nmp->nm_so) == NULL)
714                         return (EACCES);
715                 if (so->so_state & SS_ISCONNECTED)
716                         getnam = NULL;
717                 else
718                         getnam = aname;
719                 sbinit(&sio, 100000000);
720                 do {
721                         rcvflg = 0;
722                         error =  so_pru_soreceive(so, getnam, NULL, &sio,
723                                                   NULL, &rcvflg);
724                         if (error == EWOULDBLOCK && rep &&
725                             (rep->r_flags & R_SOFTTERM)) {
726                                 m_freem(sio.sb_mb);
727                                 return (EINTR);
728                         }
729                 } while (error == EWOULDBLOCK);
730
731                 len = sio.sb_cc;
732                 *mp = sio.sb_mb;
733
734                 /*
735                  * A shutdown may result in no error and no mbuf.
736                  * Convert to EPIPE.
737                  */
738                 if (*mp == NULL && error == 0)
739                         error = EPIPE;
740         }
741         if (error) {
742                 m_freem(*mp);
743                 *mp = NULL;
744         }
745
746         /*
747          * Search for any mbufs that are not a multiple of 4 bytes long
748          * or with m_data not longword aligned.
749          * These could cause pointer alignment problems, so copy them to
750          * well aligned mbufs.
751          */
752         nfs_realign(mp, 5 * NFSX_UNSIGNED);
753         return (error);
754 }
755
756 /*
757  * Implement receipt of reply on a socket.
758  *
759  * We must search through the list of received datagrams matching them
760  * with outstanding requests using the xid, until ours is found.
761  *
762  * If myrep is NULL we process packets on the socket until
763  * interrupted or until nm_reqrxq is non-empty.
764  */
765 /* ARGSUSED */
766 int
767 nfs_reply(struct nfsmount *nmp, struct nfsreq *myrep)
768 {
769         struct nfsreq *rep;
770         struct sockaddr *nam;
771         u_int32_t rxid;
772         u_int32_t *tl;
773         int error;
774         struct nfsm_info info;
775
776         /*
777          * Loop around until we get our own reply
778          */
779         for (;;) {
780                 /*
781                  * Lock against other receivers so that I don't get stuck in
782                  * sbwait() after someone else has received my reply for me.
783                  * Also necessary for connection based protocols to avoid
784                  * race conditions during a reconnect.
785                  *
786                  * If nfs_rcvlock() returns EALREADY, that means that
787                  * the reply has already been recieved by another
788                  * process and we can return immediately.  In this
789                  * case, the lock is not taken to avoid races with
790                  * other processes.
791                  */
792                 info.mrep = NULL;
793
794                 error = nfs_rcvlock(nmp, myrep);
795                 if (error == EALREADY)
796                         return (0);
797                 if (error)
798                         return (error);
799
800                 /*
801                  * If myrep is NULL we are the receiver helper thread.
802                  * Stop waiting for incoming replies if there are
803                  * messages sitting on reqrxq that we need to process,
804                  * or if a shutdown request is pending.
805                  */
806                 if (myrep == NULL && (TAILQ_FIRST(&nmp->nm_reqrxq) ||
807                     nmp->nm_rxstate > NFSSVC_PENDING)) {
808                         nfs_rcvunlock(nmp);
809                         return(EWOULDBLOCK);
810                 }
811
812                 /*
813                  * Get the next Rpc reply off the socket
814                  *
815                  * We cannot release the receive lock until we've
816                  * filled in rep->r_mrep, otherwise a waiting
817                  * thread may deadlock in soreceive with no incoming
818                  * packets expected.
819                  */
820                 error = nfs_receive(nmp, myrep, &nam, &info.mrep);
821                 if (error) {
822                         /*
823                          * Ignore routing errors on connectionless protocols??
824                          */
825                         nfs_rcvunlock(nmp);
826                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
827                                 if (nmp->nm_so == NULL)
828                                         return (error);
829                                 nmp->nm_so->so_error = 0;
830                                 continue;
831                         }
832                         return (error);
833                 }
834                 if (nam)
835                         FREE(nam, M_SONAME);
836
837                 /*
838                  * Get the xid and check that it is an rpc reply
839                  */
840                 info.md = info.mrep;
841                 info.dpos = mtod(info.md, caddr_t);
842                 NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
843                 rxid = *tl++;
844                 if (*tl != rpc_reply) {
845                         nfsstats.rpcinvalid++;
846                         m_freem(info.mrep);
847                         info.mrep = NULL;
848 nfsmout:
849                         nfs_rcvunlock(nmp);
850                         continue;
851                 }
852
853                 /*
854                  * Loop through the request list to match up the reply
855                  * Iff no match, just drop the datagram.  On match, set
856                  * r_mrep atomically to prevent the timer from messing
857                  * around with the request after we have exited the critical
858                  * section.
859                  */
860                 crit_enter();
861                 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
862                         if (rep->r_mrep == NULL && rxid == rep->r_xid)
863                                 break;
864                 }
865
866                 /*
867                  * Fill in the rest of the reply if we found a match.
868                  *
869                  * Deal with duplicate responses if there was no match.
870                  */
871                 if (rep) {
872                         rep->r_md = info.md;
873                         rep->r_dpos = info.dpos;
874                         if (nfsrtton) {
875                                 struct rttl *rt;
876
877                                 rt = &nfsrtt.rttl[nfsrtt.pos];
878                                 rt->proc = rep->r_procnum;
879                                 rt->rto = 0;
880                                 rt->sent = 0;
881                                 rt->cwnd = nmp->nm_maxasync_scaled;
882                                 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
883                                 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
884                                 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
885                                 getmicrotime(&rt->tstamp);
886                                 if (rep->r_flags & R_TIMING)
887                                         rt->rtt = rep->r_rtt;
888                                 else
889                                         rt->rtt = 1000000;
890                                 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
891                         }
892
893                         /*
894                          * New congestion control is based only on async
895                          * requests.
896                          */
897                         if (nmp->nm_maxasync_scaled < NFS_MAXASYNC_SCALED)
898                                 ++nmp->nm_maxasync_scaled;
899                         if (rep->r_flags & R_SENT) {
900                                 rep->r_flags &= ~R_SENT;
901                         }
902                         /*
903                          * Update rtt using a gain of 0.125 on the mean
904                          * and a gain of 0.25 on the deviation.
905                          *
906                          * NOTE SRTT/SDRTT are only good if R_TIMING is set.
907                          */
908                         if ((rep->r_flags & R_TIMING) && rep->r_rexmit == 0) {
909                                 /*
910                                  * Since the timer resolution of
911                                  * NFS_HZ is so course, it can often
912                                  * result in r_rtt == 0. Since
913                                  * r_rtt == N means that the actual
914                                  * rtt is between N+dt and N+2-dt ticks,
915                                  * add 1.
916                                  */
917                                 int n;
918                                 int d;
919
920 #define NFSRSB  NFS_RTT_SCALE_BITS
921                                 n = ((NFS_SRTT(rep) * 7) +
922                                      (rep->r_rtt << NFSRSB)) >> 3;
923                                 d = n - NFS_SRTT(rep);
924                                 NFS_SRTT(rep) = n;
925
926                                 /*
927                                  * Don't let the jitter calculation decay
928                                  * too quickly, but we want a fast rampup.
929                                  */
930                                 if (d < 0)
931                                         d = -d;
932                                 d <<= NFSRSB;
933                                 if (d < NFS_SDRTT(rep))
934                                         n = ((NFS_SDRTT(rep) * 15) + d) >> 4;
935                                 else
936                                         n = ((NFS_SDRTT(rep) * 3) + d) >> 2;
937                                 NFS_SDRTT(rep) = n;
938 #undef NFSRSB
939                         }
940                         nmp->nm_timeouts = 0;
941                         rep->r_mrep = info.mrep;
942                         nfs_hardterm(rep, 0);
943                 } else {
944                         /*
945                          * Extract vers, prog, nfsver, procnum.  A duplicate
946                          * response means we didn't wait long enough so
947                          * we increase the SRTT to avoid future spurious
948                          * timeouts.
949                          */
950                         u_int procnum = nmp->nm_lastreprocnum;
951                         int n;
952
953                         if (procnum < NFS_NPROCS && proct[procnum]) {
954                                 if (nfs_showrexmit)
955                                         kprintf("D");
956                                 n = nmp->nm_srtt[proct[procnum]];
957                                 n += NFS_ASYSCALE * NFS_HZ;
958                                 if (n < NFS_ASYSCALE * NFS_HZ * 10)
959                                         n = NFS_ASYSCALE * NFS_HZ * 10;
960                                 nmp->nm_srtt[proct[procnum]] = n;
961                         }
962                 }
963                 nfs_rcvunlock(nmp);
964                 crit_exit();
965
966                 /*
967                  * If not matched to a request, drop it.
968                  * If it's mine, get out.
969                  */
970                 if (rep == NULL) {
971                         nfsstats.rpcunexpected++;
972                         m_freem(info.mrep);
973                         info.mrep = NULL;
974                 } else if (rep == myrep) {
975                         if (rep->r_mrep == NULL)
976                                 panic("nfsreply nil");
977                         return (0);
978                 }
979         }
980 }
981
982 /*
983  * Run the request state machine until the target state is reached
984  * or a fatal error occurs.  The target state is not run.  Specifying
985  * a target of NFSM_STATE_DONE runs the state machine until the rpc
986  * is complete.
987  *
988  * EINPROGRESS is returned for all states other then the DONE state,
989  * indicating that the rpc is still in progress.
990  */
991 int
992 nfs_request(struct nfsm_info *info, nfsm_state_t bstate, nfsm_state_t estate)
993 {
994         struct nfsreq *req;
995
996         while (info->state >= bstate && info->state < estate) {
997                 switch(info->state) {
998                 case NFSM_STATE_SETUP:
999                         /*
1000                          * Setup the nfsreq.  Any error which occurs during
1001                          * this state is fatal.
1002                          */
1003                         info->error = nfs_request_setup(info);
1004                         if (info->error) {
1005                                 info->state = NFSM_STATE_DONE;
1006                                 return (info->error);
1007                         } else {
1008                                 req = info->req;
1009                                 req->r_mrp = &info->mrep;
1010                                 req->r_mdp = &info->md;
1011                                 req->r_dposp = &info->dpos;
1012                                 info->state = NFSM_STATE_AUTH;
1013                         }
1014                         break;
1015                 case NFSM_STATE_AUTH:
1016                         /*
1017                          * Authenticate the nfsreq.  Any error which occurs
1018                          * during this state is fatal.
1019                          */
1020                         info->error = nfs_request_auth(info->req);
1021                         if (info->error) {
1022                                 info->state = NFSM_STATE_DONE;
1023                                 return (info->error);
1024                         } else {
1025                                 info->state = NFSM_STATE_TRY;
1026                         }
1027                         break;
1028                 case NFSM_STATE_TRY:
1029                         /*
1030                          * Transmit or retransmit attempt.  An error in this
1031                          * state is ignored and we always move on to the
1032                          * next state.
1033                          *
1034                          * This can trivially race the receiver if the
1035                          * request is asynchronous.  nfs_request_try()
1036                          * will thus set the state for us and we
1037                          * must also return immediately if we are
1038                          * running an async state machine, because
1039                          * info can become invalid due to races after
1040                          * try() returns.
1041                          */
1042                         if (info->req->r_flags & R_ASYNC) {
1043                                 nfs_request_try(info->req);
1044                                 if (estate == NFSM_STATE_WAITREPLY)
1045                                         return (EINPROGRESS);
1046                         } else {
1047                                 nfs_request_try(info->req);
1048                                 info->state = NFSM_STATE_WAITREPLY;
1049                         }
1050                         break;
1051                 case NFSM_STATE_WAITREPLY:
1052                         /*
1053                          * Wait for a reply or timeout and move on to the
1054                          * next state.  The error returned by this state
1055                          * is passed to the processing code in the next
1056                          * state.
1057                          */
1058                         info->error = nfs_request_waitreply(info->req);
1059                         info->state = NFSM_STATE_PROCESSREPLY;
1060                         break;
1061                 case NFSM_STATE_PROCESSREPLY:
1062                         /*
1063                          * Process the reply or timeout.  Errors which occur
1064                          * in this state may cause the state machine to
1065                          * go back to an earlier state, and are fatal
1066                          * otherwise.
1067                          */
1068                         info->error = nfs_request_processreply(info,
1069                                                                info->error);
1070                         switch(info->error) {
1071                         case ENEEDAUTH:
1072                                 info->state = NFSM_STATE_AUTH;
1073                                 break;
1074                         case EAGAIN:
1075                                 info->state = NFSM_STATE_TRY;
1076                                 break;
1077                         default:
1078                                 /*
1079                                  * Operation complete, with or without an
1080                                  * error.  We are done.
1081                                  */
1082                                 info->req = NULL;
1083                                 info->state = NFSM_STATE_DONE;
1084                                 return (info->error);
1085                         }
1086                         break;
1087                 case NFSM_STATE_DONE:
1088                         /*
1089                          * Shouldn't be reached
1090                          */
1091                         return (info->error);
1092                         /* NOT REACHED */
1093                 }
1094         }
1095
1096         /*
1097          * If we are done return the error code (if any).
1098          * Otherwise return EINPROGRESS.
1099          */
1100         if (info->state == NFSM_STATE_DONE)
1101                 return (info->error);
1102         return (EINPROGRESS);
1103 }
1104
1105 /*
1106  * nfs_request - goes something like this
1107  *      - fill in request struct
1108  *      - links it into list
1109  *      - calls nfs_send() for first transmit
1110  *      - calls nfs_receive() to get reply
1111  *      - break down rpc header and return with nfs reply pointed to
1112  *        by mrep or error
1113  * nb: always frees up mreq mbuf list
1114  */
1115 static int
1116 nfs_request_setup(nfsm_info_t info)
1117 {
1118         struct nfsreq *req;
1119         struct nfsmount *nmp;
1120         struct mbuf *m;
1121         int i;
1122
1123         /*
1124          * Reject requests while attempting a forced unmount.
1125          */
1126         if (info->vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
1127                 m_freem(info->mreq);
1128                 info->mreq = NULL;
1129                 return (ESTALE);
1130         }
1131         nmp = VFSTONFS(info->vp->v_mount);
1132         req = kmalloc(sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
1133         req->r_nmp = nmp;
1134         req->r_vp = info->vp;
1135         req->r_td = info->td;
1136         req->r_procnum = info->procnum;
1137         req->r_mreq = NULL;
1138         req->r_cred = info->cred;
1139
1140         i = 0;
1141         m = info->mreq;
1142         while (m) {
1143                 i += m->m_len;
1144                 m = m->m_next;
1145         }
1146         req->r_mrest = info->mreq;
1147         req->r_mrest_len = i;
1148
1149         /*
1150          * The presence of a non-NULL r_info in req indicates
1151          * async completion via our helper threads.  See the receiver
1152          * code.
1153          */
1154         if (info->bio) {
1155                 req->r_info = info;
1156                 req->r_flags = R_ASYNC;
1157         } else {
1158                 req->r_info = NULL;
1159                 req->r_flags = 0;
1160         }
1161         info->req = req;
1162         return(0);
1163 }
1164
1165 static int
1166 nfs_request_auth(struct nfsreq *rep)
1167 {
1168         struct nfsmount *nmp = rep->r_nmp;
1169         struct mbuf *m;
1170         char nickv[RPCX_NICKVERF];
1171         int error = 0, auth_len, auth_type;
1172         int verf_len;
1173         u_int32_t xid;
1174         char *auth_str, *verf_str;
1175         struct ucred *cred;
1176
1177         cred = rep->r_cred;
1178         rep->r_failed_auth = 0;
1179
1180         /*
1181          * Get the RPC header with authorization.
1182          */
1183         verf_str = auth_str = NULL;
1184         if (nmp->nm_flag & NFSMNT_KERB) {
1185                 verf_str = nickv;
1186                 verf_len = sizeof (nickv);
1187                 auth_type = RPCAUTH_KERB4;
1188                 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1189                 if (rep->r_failed_auth ||
1190                     nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1191                                     verf_str, verf_len)) {
1192                         error = nfs_getauth(nmp, rep, cred, &auth_str,
1193                                 &auth_len, verf_str, &verf_len, rep->r_key);
1194                         if (error) {
1195                                 m_freem(rep->r_mrest);
1196                                 rep->r_mrest = NULL;
1197                                 kfree((caddr_t)rep, M_NFSREQ);
1198                                 return (error);
1199                         }
1200                 }
1201         } else {
1202                 auth_type = RPCAUTH_UNIX;
1203                 if (cred->cr_ngroups < 1)
1204                         panic("nfsreq nogrps");
1205                 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1206                         nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1207                         5 * NFSX_UNSIGNED;
1208         }
1209         if (rep->r_mrest)
1210                 nfs_checkpkt(rep->r_mrest, rep->r_mrest_len);
1211         m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1212                         auth_len, auth_str, verf_len, verf_str,
1213                         rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1214         rep->r_mrest = NULL;
1215         if (auth_str)
1216                 kfree(auth_str, M_TEMP);
1217
1218         /*
1219          * For stream protocols, insert a Sun RPC Record Mark.
1220          */
1221         if (nmp->nm_sotype == SOCK_STREAM) {
1222                 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1223                 if (m == NULL) {
1224                         kfree(rep, M_NFSREQ);
1225                         return (ENOBUFS);
1226                 }
1227                 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1228                          (m->m_pkthdr.len - NFSX_UNSIGNED));
1229         }
1230
1231         nfs_checkpkt(m, m->m_pkthdr.len);
1232
1233         rep->r_mreq = m;
1234         rep->r_xid = xid;
1235         return (0);
1236 }
1237
1238 static int
1239 nfs_request_try(struct nfsreq *rep)
1240 {
1241         struct nfsmount *nmp = rep->r_nmp;
1242         struct mbuf *m2;
1243         int error;
1244
1245         /*
1246          * Request is not on any queue, only the owner has access to it
1247          * so it should not be locked by anyone atm.
1248          *
1249          * Interlock to prevent races.  While locked the only remote
1250          * action possible is for r_mrep to be set (once we enqueue it).
1251          */
1252         if (rep->r_flags == 0xdeadc0de) {
1253                 print_backtrace(-1);
1254                 panic("flags nbad\n");
1255         }
1256         KKASSERT((rep->r_flags & (R_LOCKED | R_ONREQQ)) == 0);
1257         if (nmp->nm_flag & NFSMNT_SOFT)
1258                 rep->r_retry = nmp->nm_retry;
1259         else
1260                 rep->r_retry = NFS_MAXREXMIT + 1;       /* past clip limit */
1261         rep->r_rtt = rep->r_rexmit = 0;
1262         if (proct[rep->r_procnum] > 0)
1263                 rep->r_flags |= R_TIMING | R_LOCKED;
1264         else
1265                 rep->r_flags |= R_LOCKED;
1266         rep->r_mrep = NULL;
1267
1268         /*
1269          * Do the client side RPC.
1270          */
1271         nfsstats.rpcrequests++;
1272
1273         if (nmp->nm_flag & NFSMNT_FORCE) {
1274                 rep->r_flags |= R_SOFTTERM;
1275                 rep->r_flags &= ~R_LOCKED;
1276                 return (0);
1277         }
1278
1279         /*
1280          * Chain request into list of outstanding requests. Be sure
1281          * to put it LAST so timer finds oldest requests first.  Note
1282          * that our control of R_LOCKED prevents the request from
1283          * getting ripped out from under us or transmitted by the
1284          * timer code.
1285          *
1286          * For requests with info structures we must atomically set the
1287          * info's state because the structure could become invalid upon
1288          * return due to races (i.e., if async)
1289          */
1290         crit_enter();
1291         mtx_link_init(&rep->r_link);
1292         TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1293         rep->r_flags |= R_ONREQQ;
1294         ++nmp->nm_reqqlen;
1295         if (rep->r_flags & R_ASYNC)
1296                 rep->r_info->state = NFSM_STATE_WAITREPLY;
1297         crit_exit();
1298
1299         error = 0;
1300
1301         /*
1302          * Send if we can.  Congestion control is not handled here any more
1303          * becausing trying to defer the initial send based on the nfs_timer
1304          * requires having a very fast nfs_timer, which is silly.
1305          */
1306         if (nmp->nm_so) {
1307                 if (nmp->nm_soflags & PR_CONNREQUIRED)
1308                         error = nfs_sndlock(nmp, rep);
1309                 if (error == 0) {
1310                         m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1311                         error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1312                         if (nmp->nm_soflags & PR_CONNREQUIRED)
1313                                 nfs_sndunlock(nmp);
1314                         rep->r_flags &= ~R_NEEDSXMIT;
1315                         if ((rep->r_flags & R_SENT) == 0) {
1316                                 rep->r_flags |= R_SENT;
1317                         }
1318                 } else {
1319                         rep->r_flags |= R_NEEDSXMIT;
1320                 }
1321         } else {
1322                 rep->r_flags |= R_NEEDSXMIT;
1323                 rep->r_rtt = -1;
1324         }
1325         if (error == EPIPE)
1326                 error = 0;
1327
1328         /*
1329          * Release the lock.  The only remote action that may have occurred
1330          * would have been the setting of rep->r_mrep.  If this occured
1331          * and the request was async we have to move it to the reader
1332          * thread's queue for action.
1333          *
1334          * For async requests also make sure the reader is woken up so
1335          * it gets on the socket to read responses.
1336          */
1337         crit_enter();
1338         if (rep->r_flags & R_ASYNC) {
1339                 if (rep->r_mrep)
1340                         nfs_hardterm(rep, 1);
1341                 rep->r_flags &= ~R_LOCKED;
1342                 nfssvc_iod_reader_wakeup(nmp);
1343         } else {
1344                 rep->r_flags &= ~R_LOCKED;
1345         }
1346         if (rep->r_flags & R_WANTED) {
1347                 rep->r_flags &= ~R_WANTED;
1348                 wakeup(rep);
1349         }
1350         crit_exit();
1351         return (error);
1352 }
1353
1354 /*
1355  * This code is only called for synchronous requests.  Completed synchronous
1356  * requests are left on reqq and we remove them before moving on to the
1357  * processing state.
1358  */
1359 static int
1360 nfs_request_waitreply(struct nfsreq *rep)
1361 {
1362         struct nfsmount *nmp = rep->r_nmp;
1363         int error;
1364
1365         KKASSERT((rep->r_flags & R_ASYNC) == 0);
1366
1367         /*
1368          * Wait until the request is finished.
1369          */
1370         error = nfs_reply(nmp, rep);
1371
1372         /*
1373          * RPC done, unlink the request, but don't rip it out from under
1374          * the callout timer.
1375          *
1376          * Once unlinked no other receiver or the timer will have
1377          * visibility, so we do not have to set R_LOCKED.
1378          */
1379         crit_enter();
1380         while (rep->r_flags & R_LOCKED) {
1381                 rep->r_flags |= R_WANTED;
1382                 tsleep(rep, 0, "nfstrac", 0);
1383         }
1384         KKASSERT(rep->r_flags & R_ONREQQ);
1385         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1386         rep->r_flags &= ~R_ONREQQ;
1387         --nmp->nm_reqqlen;
1388         if (TAILQ_FIRST(&nmp->nm_bioq) &&
1389             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1390                 nfssvc_iod_writer_wakeup(nmp);
1391         }
1392         crit_exit();
1393
1394         /*
1395          * Decrement the outstanding request count.
1396          */
1397         if (rep->r_flags & R_SENT) {
1398                 rep->r_flags &= ~R_SENT;
1399         }
1400         return (error);
1401 }
1402
1403 /*
1404  * Process reply with error returned from nfs_requet_waitreply().
1405  *
1406  * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1407  * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1408  */
1409 static int
1410 nfs_request_processreply(nfsm_info_t info, int error)
1411 {
1412         struct nfsreq *req = info->req;
1413         struct nfsmount *nmp = req->r_nmp;
1414         u_int32_t *tl;
1415         int verf_type;
1416         int i;
1417
1418         /*
1419          * If there was a successful reply and a tprintf msg.
1420          * tprintf a response.
1421          */
1422         if (error == 0 && (req->r_flags & R_TPRINTFMSG)) {
1423                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1424                     "is alive again");
1425         }
1426         info->mrep = req->r_mrep;
1427         info->md = req->r_md;
1428         info->dpos = req->r_dpos;
1429         if (error) {
1430                 m_freem(req->r_mreq);
1431                 req->r_mreq = NULL;
1432                 kfree(req, M_NFSREQ);
1433                 info->req = NULL;
1434                 return (error);
1435         }
1436
1437         /*
1438          * break down the rpc header and check if ok
1439          */
1440         NULLOUT(tl = nfsm_dissect(info, 3 * NFSX_UNSIGNED));
1441         if (*tl++ == rpc_msgdenied) {
1442                 if (*tl == rpc_mismatch) {
1443                         error = EOPNOTSUPP;
1444                 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1445                            *tl++ == rpc_autherr) {
1446                         if (req->r_failed_auth == 0) {
1447                                 req->r_failed_auth++;
1448                                 req->r_mheadend->m_next = NULL;
1449                                 m_freem(info->mrep);
1450                                 info->mrep = NULL;
1451                                 m_freem(req->r_mreq);
1452                                 req->r_mreq = NULL;
1453                                 return (ENEEDAUTH);
1454                         } else {
1455                                 error = EAUTH;
1456                         }
1457                 } else {
1458                         error = EACCES;
1459                 }
1460                 m_freem(info->mrep);
1461                 info->mrep = NULL;
1462                 m_freem(req->r_mreq);
1463                 req->r_mreq = NULL;
1464                 kfree(req, M_NFSREQ);
1465                 info->req = NULL;
1466                 return (error);
1467         }
1468
1469         /*
1470          * Grab any Kerberos verifier, otherwise just throw it away.
1471          */
1472         verf_type = fxdr_unsigned(int, *tl++);
1473         i = fxdr_unsigned(int32_t, *tl);
1474         if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1475                 error = nfs_savenickauth(nmp, req->r_cred, i, req->r_key,
1476                                          &info->md, &info->dpos, info->mrep);
1477                 if (error)
1478                         goto nfsmout;
1479         } else if (i > 0) {
1480                 ERROROUT(nfsm_adv(info, nfsm_rndup(i)));
1481         }
1482         NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1483         /* 0 == ok */
1484         if (*tl == 0) {
1485                 NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1486                 if (*tl != 0) {
1487                         error = fxdr_unsigned(int, *tl);
1488
1489                         /*
1490                          * Does anyone even implement this?  Just impose
1491                          * a 1-second delay.
1492                          */
1493                         if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1494                                 error == NFSERR_TRYLATER) {
1495                                 m_freem(info->mrep);
1496                                 info->mrep = NULL;
1497                                 error = 0;
1498
1499                                 tsleep((caddr_t)&lbolt, 0, "nqnfstry", 0);
1500                                 return (EAGAIN);        /* goto tryagain */
1501                         }
1502
1503                         /*
1504                          * If the File Handle was stale, invalidate the
1505                          * lookup cache, just in case.
1506                          *
1507                          * To avoid namecache<->vnode deadlocks we must
1508                          * release the vnode lock if we hold it.
1509                          */
1510                         if (error == ESTALE) {
1511                                 struct vnode *vp = req->r_vp;
1512                                 int ltype;
1513
1514                                 ltype = lockstatus(&vp->v_lock, curthread);
1515                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1516                                         lockmgr(&vp->v_lock, LK_RELEASE);
1517                                 cache_inval_vp(vp, CINV_CHILDREN);
1518                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1519                                         lockmgr(&vp->v_lock, ltype);
1520                         }
1521                         if (nmp->nm_flag & NFSMNT_NFSV3) {
1522                                 KKASSERT(*req->r_mrp == info->mrep);
1523                                 KKASSERT(*req->r_mdp == info->md);
1524                                 KKASSERT(*req->r_dposp == info->dpos);
1525                                 error |= NFSERR_RETERR;
1526                         } else {
1527                                 m_freem(info->mrep);
1528                                 info->mrep = NULL;
1529                         }
1530                         m_freem(req->r_mreq);
1531                         req->r_mreq = NULL;
1532                         kfree(req, M_NFSREQ);
1533                         info->req = NULL;
1534                         return (error);
1535                 }
1536
1537                 KKASSERT(*req->r_mrp == info->mrep);
1538                 KKASSERT(*req->r_mdp == info->md);
1539                 KKASSERT(*req->r_dposp == info->dpos);
1540                 m_freem(req->r_mreq);
1541                 req->r_mreq = NULL;
1542                 FREE(req, M_NFSREQ);
1543                 return (0);
1544         }
1545         m_freem(info->mrep);
1546         info->mrep = NULL;
1547         error = EPROTONOSUPPORT;
1548 nfsmout:
1549         m_freem(req->r_mreq);
1550         req->r_mreq = NULL;
1551         kfree(req, M_NFSREQ);
1552         info->req = NULL;
1553         return (error);
1554 }
1555
1556 #ifndef NFS_NOSERVER
1557 /*
1558  * Generate the rpc reply header
1559  * siz arg. is used to decide if adding a cluster is worthwhile
1560  */
1561 int
1562 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1563             int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1564 {
1565         u_int32_t *tl;
1566         struct nfsm_info info;
1567
1568         siz += RPC_REPLYSIZ;
1569         info.mb = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1570         info.mreq = info.mb;
1571         info.mreq->m_pkthdr.len = 0;
1572         /*
1573          * If this is not a cluster, try and leave leading space
1574          * for the lower level headers.
1575          */
1576         if ((max_hdr + siz) < MINCLSIZE)
1577                 info.mreq->m_data += max_hdr;
1578         tl = mtod(info.mreq, u_int32_t *);
1579         info.mreq->m_len = 6 * NFSX_UNSIGNED;
1580         info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1581         *tl++ = txdr_unsigned(nd->nd_retxid);
1582         *tl++ = rpc_reply;
1583         if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1584                 *tl++ = rpc_msgdenied;
1585                 if (err & NFSERR_AUTHERR) {
1586                         *tl++ = rpc_autherr;
1587                         *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1588                         info.mreq->m_len -= NFSX_UNSIGNED;
1589                         info.bpos -= NFSX_UNSIGNED;
1590                 } else {
1591                         *tl++ = rpc_mismatch;
1592                         *tl++ = txdr_unsigned(RPC_VER2);
1593                         *tl = txdr_unsigned(RPC_VER2);
1594                 }
1595         } else {
1596                 *tl++ = rpc_msgaccepted;
1597
1598                 /*
1599                  * For Kerberos authentication, we must send the nickname
1600                  * verifier back, otherwise just RPCAUTH_NULL.
1601                  */
1602                 if (nd->nd_flag & ND_KERBFULL) {
1603                     struct nfsuid *nuidp;
1604                     struct timeval ktvin, ktvout;
1605
1606                     for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1607                         nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1608                         if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1609                             (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1610                              &nuidp->nu_haddr, nd->nd_nam2)))
1611                             break;
1612                     }
1613                     if (nuidp) {
1614                         ktvin.tv_sec =
1615                             txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1616                         ktvin.tv_usec =
1617                             txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1618
1619                         /*
1620                          * Encrypt the timestamp in ecb mode using the
1621                          * session key.
1622                          */
1623 #ifdef NFSKERB
1624                         XXX
1625 #else
1626                         ktvout.tv_sec = 0;
1627                         ktvout.tv_usec = 0;
1628 #endif
1629
1630                         *tl++ = rpc_auth_kerb;
1631                         *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1632                         *tl = ktvout.tv_sec;
1633                         tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1634                         *tl++ = ktvout.tv_usec;
1635                         *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1636                     } else {
1637                         *tl++ = 0;
1638                         *tl++ = 0;
1639                     }
1640                 } else {
1641                         *tl++ = 0;
1642                         *tl++ = 0;
1643                 }
1644                 switch (err) {
1645                 case EPROGUNAVAIL:
1646                         *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1647                         break;
1648                 case EPROGMISMATCH:
1649                         *tl = txdr_unsigned(RPC_PROGMISMATCH);
1650                         tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1651                         *tl++ = txdr_unsigned(2);
1652                         *tl = txdr_unsigned(3);
1653                         break;
1654                 case EPROCUNAVAIL:
1655                         *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1656                         break;
1657                 case EBADRPC:
1658                         *tl = txdr_unsigned(RPC_GARBAGE);
1659                         break;
1660                 default:
1661                         *tl = 0;
1662                         if (err != NFSERR_RETVOID) {
1663                                 tl = nfsm_build(&info, NFSX_UNSIGNED);
1664                                 if (err)
1665                                     *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1666                                 else
1667                                     *tl = 0;
1668                         }
1669                         break;
1670                 };
1671         }
1672
1673         if (mrq != NULL)
1674             *mrq = info.mreq;
1675         *mbp = info.mb;
1676         *bposp = info.bpos;
1677         if (err != 0 && err != NFSERR_RETVOID)
1678                 nfsstats.srvrpc_errs++;
1679         return (0);
1680 }
1681
1682
1683 #endif /* NFS_NOSERVER */
1684
1685 /*
1686  * Nfs timer routine.
1687  *
1688  * Scan the nfsreq list and retranmit any requests that have timed out
1689  * To avoid retransmission attempts on STREAM sockets (in the future) make
1690  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1691  *
1692  * Requests with attached responses, terminated requests, and
1693  * locked requests are ignored.  Locked requests will be picked up
1694  * in a later timer call.
1695  */
1696 void
1697 nfs_timer_callout(void *arg /* never used */)
1698 {
1699         struct nfsmount *nmp;
1700         struct nfsreq *req;
1701 #ifndef NFS_NOSERVER
1702         struct nfssvc_sock *slp;
1703         u_quad_t cur_usec;
1704 #endif /* NFS_NOSERVER */
1705
1706         lwkt_gettoken(&nfs_token);
1707         TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1708                 lwkt_gettoken(&nmp->nm_token);
1709                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1710                         KKASSERT(nmp == req->r_nmp);
1711                         if (req->r_mrep)
1712                                 continue;
1713                         if (req->r_flags & (R_SOFTTERM | R_LOCKED))
1714                                 continue;
1715                         req->r_flags |= R_LOCKED;
1716                         if (nfs_sigintr(nmp, req, req->r_td)) {
1717                                 nfs_softterm(req, 1);
1718                         } else {
1719                                 nfs_timer_req(req);
1720                         }
1721                         req->r_flags &= ~R_LOCKED;
1722                         if (req->r_flags & R_WANTED) {
1723                                 req->r_flags &= ~R_WANTED;
1724                                 wakeup(req);
1725                         }
1726                 }
1727                 lwkt_reltoken(&nmp->nm_token);
1728         }
1729 #ifndef NFS_NOSERVER
1730
1731         /*
1732          * Scan the write gathering queues for writes that need to be
1733          * completed now.
1734          */
1735         cur_usec = nfs_curusec();
1736
1737         TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1738             /* XXX race against removal */
1739             lwkt_gettoken(&slp->ns_token);
1740             if (slp->ns_tq.lh_first && slp->ns_tq.lh_first->nd_time<=cur_usec)
1741                 nfsrv_wakenfsd(slp, 1);
1742             lwkt_reltoken(&slp->ns_token);
1743         }
1744 #endif /* NFS_NOSERVER */
1745
1746         callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer_callout, NULL);
1747         lwkt_reltoken(&nfs_token);
1748 }
1749
1750 static
1751 void
1752 nfs_timer_req(struct nfsreq *req)
1753 {
1754         struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1755         struct nfsmount *nmp = req->r_nmp;
1756         struct mbuf *m;
1757         struct socket *so;
1758         int timeo;
1759         int error;
1760
1761         /*
1762          * rtt ticks and timeout calculation.  Return if the timeout
1763          * has not been reached yet, unless the packet is flagged
1764          * for an immediate send.
1765          *
1766          * The mean rtt doesn't help when we get random I/Os, we have
1767          * to multiply by fairly large numbers.
1768          */
1769         if (req->r_rtt >= 0) {
1770                 /*
1771                  * Calculate the timeout to test against.
1772                  */
1773                 req->r_rtt++;
1774                 if (nmp->nm_flag & NFSMNT_DUMBTIMR) {
1775                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1776                 } else if (req->r_flags & R_TIMING) {
1777                         timeo = NFS_SRTT(req) + NFS_SDRTT(req);
1778                 } else {
1779                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1780                 }
1781                 timeo *= multt[req->r_procnum];
1782                 /* timeo is still scaled by SCALE_BITS */
1783
1784 #define NFSFS   (NFS_RTT_SCALE * NFS_HZ)
1785                 if (req->r_flags & R_TIMING) {
1786                         static long last_time;
1787                         if (nfs_showrtt && last_time != time_second) {
1788                                 kprintf("rpccmd %d NFS SRTT %d SDRTT %d "
1789                                         "timeo %d.%03d\n",
1790                                         proct[req->r_procnum],
1791                                         NFS_SRTT(req), NFS_SDRTT(req),
1792                                         timeo / NFSFS,
1793                                         timeo % NFSFS * 1000 /  NFSFS);
1794                                 last_time = time_second;
1795                         }
1796                 }
1797 #undef NFSFS
1798
1799                 /*
1800                  * deal with nfs_timer jitter.
1801                  */
1802                 timeo = (timeo >> NFS_RTT_SCALE_BITS) + 1;
1803                 if (timeo < 2)
1804                         timeo = 2;
1805
1806                 if (nmp->nm_timeouts > 0)
1807                         timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1808                 if (timeo > NFS_MAXTIMEO)
1809                         timeo = NFS_MAXTIMEO;
1810                 if (req->r_rtt <= timeo) {
1811                         if ((req->r_flags & R_NEEDSXMIT) == 0)
1812                                 return;
1813                 } else if (nmp->nm_timeouts < 8) {
1814                         nmp->nm_timeouts++;
1815                 }
1816         }
1817
1818         /*
1819          * Check for server not responding
1820          */
1821         if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1822              req->r_rexmit > nmp->nm_deadthresh) {
1823                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1824                         "not responding");
1825                 req->r_flags |= R_TPRINTFMSG;
1826         }
1827         if (req->r_rexmit >= req->r_retry) {    /* too many */
1828                 nfsstats.rpctimeouts++;
1829                 nfs_softterm(req, 1);
1830                 return;
1831         }
1832
1833         /*
1834          * Generally disable retransmission on reliable sockets,
1835          * unless the request is flagged for immediate send.
1836          */
1837         if (nmp->nm_sotype != SOCK_DGRAM) {
1838                 if (++req->r_rexmit > NFS_MAXREXMIT)
1839                         req->r_rexmit = NFS_MAXREXMIT;
1840                 if ((req->r_flags & R_NEEDSXMIT) == 0)
1841                         return;
1842         }
1843
1844         /*
1845          * Stop here if we do not have a socket!
1846          */
1847         if ((so = nmp->nm_so) == NULL)
1848                 return;
1849
1850         /*
1851          * If there is enough space and the window allows.. resend it.
1852          *
1853          * r_rtt is left intact in case we get an answer after the
1854          * retry that was a reply to the original packet.
1855          *
1856          * NOTE: so_pru_send()
1857          */
1858         if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1859             (req->r_flags & (R_SENT | R_NEEDSXMIT)) &&
1860            (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1861                 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1862                     error = so_pru_send(so, 0, m, NULL, NULL, td);
1863                 else
1864                     error = so_pru_send(so, 0, m, nmp->nm_nam, NULL, td);
1865                 if (error) {
1866                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1867                                 so->so_error = 0;
1868                         req->r_flags |= R_NEEDSXMIT;
1869                 } else if (req->r_mrep == NULL) {
1870                         /*
1871                          * Iff first send, start timing
1872                          * else turn timing off, backoff timer
1873                          * and divide congestion window by 2.
1874                          *
1875                          * It is possible for the so_pru_send() to
1876                          * block and for us to race a reply so we
1877                          * only do this if the reply field has not
1878                          * been filled in.  R_LOCKED will prevent
1879                          * the request from being ripped out from under
1880                          * us entirely.
1881                          *
1882                          * Record the last resent procnum to aid us
1883                          * in duplicate detection on receive.
1884                          */
1885                         if ((req->r_flags & R_NEEDSXMIT) == 0) {
1886                                 if (nfs_showrexmit)
1887                                         kprintf("X");
1888                                 if (++req->r_rexmit > NFS_MAXREXMIT)
1889                                         req->r_rexmit = NFS_MAXREXMIT;
1890                                 nmp->nm_maxasync_scaled >>= 1;
1891                                 if (nmp->nm_maxasync_scaled < NFS_MINASYNC_SCALED)
1892                                         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
1893                                 nfsstats.rpcretries++;
1894                                 nmp->nm_lastreprocnum = req->r_procnum;
1895                         } else {
1896                                 req->r_flags |= R_SENT;
1897                                 req->r_flags &= ~R_NEEDSXMIT;
1898                         }
1899                 }
1900         }
1901 }
1902
1903 /*
1904  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1905  * wait for all requests to complete. This is used by forced unmounts
1906  * to terminate any outstanding RPCs.
1907  *
1908  * Locked requests cannot be canceled but will be marked for
1909  * soft-termination.
1910  */
1911 int
1912 nfs_nmcancelreqs(struct nfsmount *nmp)
1913 {
1914         struct nfsreq *req;
1915         int i;
1916
1917         crit_enter();
1918         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1919                 if (req->r_mrep != NULL || (req->r_flags & R_SOFTTERM))
1920                         continue;
1921                 nfs_softterm(req, 0);
1922         }
1923         /* XXX  the other two queues as well */
1924         crit_exit();
1925
1926         for (i = 0; i < 30; i++) {
1927                 crit_enter();
1928                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1929                         if (nmp == req->r_nmp)
1930                                 break;
1931                 }
1932                 crit_exit();
1933                 if (req == NULL)
1934                         return (0);
1935                 tsleep(&lbolt, 0, "nfscancel", 0);
1936         }
1937         return (EBUSY);
1938 }
1939
1940 /*
1941  * Soft-terminate a request, effectively marking it as failed.
1942  *
1943  * Must be called from within a critical section.
1944  */
1945 static void
1946 nfs_softterm(struct nfsreq *rep, int islocked)
1947 {
1948         rep->r_flags |= R_SOFTTERM;
1949         nfs_hardterm(rep, islocked);
1950 }
1951
1952 /*
1953  * Hard-terminate a request, typically after getting a response.
1954  *
1955  * The state machine can still decide to re-issue it later if necessary.
1956  *
1957  * Must be called from within a critical section.
1958  */
1959 static void
1960 nfs_hardterm(struct nfsreq *rep, int islocked)
1961 {
1962         struct nfsmount *nmp = rep->r_nmp;
1963
1964         /*
1965          * The nm_send count is decremented now to avoid deadlocks
1966          * when the process in soreceive() hasn't yet managed to send
1967          * its own request.
1968          */
1969         if (rep->r_flags & R_SENT) {
1970                 rep->r_flags &= ~R_SENT;
1971         }
1972
1973         /*
1974          * If we locked the request or nobody else has locked the request,
1975          * and the request is async, we can move it to the reader thread's
1976          * queue now and fix up the state.
1977          *
1978          * If we locked the request or nobody else has locked the request,
1979          * we can wake up anyone blocked waiting for a response on the
1980          * request.
1981          */
1982         if (islocked || (rep->r_flags & R_LOCKED) == 0) {
1983                 if ((rep->r_flags & (R_ONREQQ | R_ASYNC)) ==
1984                     (R_ONREQQ | R_ASYNC)) {
1985                         rep->r_flags &= ~R_ONREQQ;
1986                         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1987                         --nmp->nm_reqqlen;
1988                         TAILQ_INSERT_TAIL(&nmp->nm_reqrxq, rep, r_chain);
1989                         KKASSERT(rep->r_info->state == NFSM_STATE_TRY ||
1990                                  rep->r_info->state == NFSM_STATE_WAITREPLY);
1991                         rep->r_info->state = NFSM_STATE_PROCESSREPLY;
1992                         nfssvc_iod_reader_wakeup(nmp);
1993                         if (TAILQ_FIRST(&nmp->nm_bioq) &&
1994                             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1995                                 nfssvc_iod_writer_wakeup(nmp);
1996                         }
1997                 }
1998                 mtx_abort_ex_link(&nmp->nm_rxlock, &rep->r_link);
1999         }
2000 }
2001
2002 /*
2003  * Test for a termination condition pending on the process.
2004  * This is used for NFSMNT_INT mounts.
2005  */
2006 int
2007 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
2008 {
2009         sigset_t tmpset;
2010         struct proc *p;
2011         struct lwp *lp;
2012
2013         if (rep && (rep->r_flags & R_SOFTTERM))
2014                 return (EINTR);
2015         /* Terminate all requests while attempting a forced unmount. */
2016         if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
2017                 return (EINTR);
2018         if (!(nmp->nm_flag & NFSMNT_INT))
2019                 return (0);
2020         /* td might be NULL YYY */
2021         if (td == NULL || (p = td->td_proc) == NULL)
2022                 return (0);
2023
2024         lp = td->td_lwp;
2025         tmpset = lwp_sigpend(lp);
2026         SIGSETNAND(tmpset, lp->lwp_sigmask);
2027         SIGSETNAND(tmpset, p->p_sigignore);
2028         if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
2029                 return (EINTR);
2030
2031         return (0);
2032 }
2033
2034 /*
2035  * Lock a socket against others.
2036  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
2037  * and also to avoid race conditions between the processes with nfs requests
2038  * in progress when a reconnect is necessary.
2039  */
2040 int
2041 nfs_sndlock(struct nfsmount *nmp, struct nfsreq *rep)
2042 {
2043         mtx_t mtx = &nmp->nm_txlock;
2044         struct thread *td;
2045         int slptimeo;
2046         int slpflag;
2047         int error;
2048
2049         slpflag = 0;
2050         slptimeo = 0;
2051         td = rep ? rep->r_td : NULL;
2052         if (nmp->nm_flag & NFSMNT_INT)
2053                 slpflag = PCATCH;
2054
2055         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2056                 if (nfs_sigintr(nmp, rep, td)) {
2057                         error = EINTR;
2058                         break;
2059                 }
2060                 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
2061                 if (error == 0)
2062                         break;
2063                 if (slpflag == PCATCH) {
2064                         slpflag = 0;
2065                         slptimeo = 2 * hz;
2066                 }
2067         }
2068         /* Always fail if our request has been cancelled. */
2069         if (rep && (rep->r_flags & R_SOFTTERM)) {
2070                 if (error == 0)
2071                         mtx_unlock(mtx);
2072                 error = EINTR;
2073         }
2074         return (error);
2075 }
2076
2077 /*
2078  * Unlock the stream socket for others.
2079  */
2080 void
2081 nfs_sndunlock(struct nfsmount *nmp)
2082 {
2083         mtx_unlock(&nmp->nm_txlock);
2084 }
2085
2086 /*
2087  * Lock the receiver side of the socket.
2088  *
2089  * rep may be NULL.
2090  */
2091 static int
2092 nfs_rcvlock(struct nfsmount *nmp, struct nfsreq *rep)
2093 {
2094         mtx_t mtx = &nmp->nm_rxlock;
2095         int slpflag;
2096         int slptimeo;
2097         int error;
2098
2099         /*
2100          * Unconditionally check for completion in case another nfsiod
2101          * get the packet while the caller was blocked, before the caller
2102          * called us.  Packet reception is handled by mainline code which
2103          * is protected by the BGL at the moment.
2104          *
2105          * We do not strictly need the second check just before the
2106          * tsleep(), but it's good defensive programming.
2107          */
2108         if (rep && rep->r_mrep != NULL)
2109                 return (EALREADY);
2110
2111         if (nmp->nm_flag & NFSMNT_INT)
2112                 slpflag = PCATCH;
2113         else
2114                 slpflag = 0;
2115         slptimeo = 0;
2116
2117         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2118                 if (nfs_sigintr(nmp, rep, (rep ? rep->r_td : NULL))) {
2119                         error = EINTR;
2120                         break;
2121                 }
2122                 if (rep && rep->r_mrep != NULL) {
2123                         error = EALREADY;
2124                         break;
2125                 }
2126
2127                 /*
2128                  * NOTE: can return ENOLCK, but in that case rep->r_mrep
2129                  *       will already be set.
2130                  */
2131                 if (rep) {
2132                         error = mtx_lock_ex_link(mtx, &rep->r_link,
2133                                                  "nfsrcvlk",
2134                                                  slpflag, slptimeo);
2135                 } else {
2136                         error = mtx_lock_ex(mtx, "nfsrcvlk", slpflag, slptimeo);
2137                 }
2138                 if (error == 0)
2139                         break;
2140
2141                 /*
2142                  * If our reply was recieved while we were sleeping,
2143                  * then just return without taking the lock to avoid a
2144                  * situation where a single iod could 'capture' the
2145                  * recieve lock.
2146                  */
2147                 if (rep && rep->r_mrep != NULL) {
2148                         error = EALREADY;
2149                         break;
2150                 }
2151                 if (slpflag == PCATCH) {
2152                         slpflag = 0;
2153                         slptimeo = 2 * hz;
2154                 }
2155         }
2156         if (error == 0) {
2157                 if (rep && rep->r_mrep != NULL) {
2158                         error = EALREADY;
2159                         mtx_unlock(mtx);
2160                 }
2161         }
2162         return (error);
2163 }
2164
2165 /*
2166  * Unlock the stream socket for others.
2167  */
2168 static void
2169 nfs_rcvunlock(struct nfsmount *nmp)
2170 {
2171         mtx_unlock(&nmp->nm_rxlock);
2172 }
2173
2174 /*
2175  * nfs_realign:
2176  *
2177  * Check for badly aligned mbuf data and realign by copying the unaligned
2178  * portion of the data into a new mbuf chain and freeing the portions
2179  * of the old chain that were replaced.
2180  *
2181  * We cannot simply realign the data within the existing mbuf chain
2182  * because the underlying buffers may contain other rpc commands and
2183  * we cannot afford to overwrite them.
2184  *
2185  * We would prefer to avoid this situation entirely.  The situation does
2186  * not occur with NFS/UDP and is supposed to only occassionally occur
2187  * with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
2188  *
2189  * NOTE!  MB_DONTWAIT cannot be used here.  The mbufs must be acquired
2190  *        because the rpc request OR reply cannot be thrown away.  TCP NFS
2191  *        mounts do not retry their RPCs unless the TCP connection itself
2192  *        is dropped so throwing away a RPC will basically cause the NFS
2193  *        operation to lockup indefinitely.
2194  */
2195 static void
2196 nfs_realign(struct mbuf **pm, int hsiz)
2197 {
2198         struct mbuf *m;
2199         struct mbuf *n = NULL;
2200
2201         /*
2202          * Check for misalignemnt
2203          */
2204         ++nfs_realign_test;
2205         while ((m = *pm) != NULL) {
2206                 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3))
2207                         break;
2208                 pm = &m->m_next;
2209         }
2210
2211         /*
2212          * If misalignment found make a completely new copy.
2213          */
2214         if (m) {
2215                 ++nfs_realign_count;
2216                 n = m_dup_data(m, MB_WAIT);
2217                 m_freem(*pm);
2218                 *pm = n;
2219         }
2220 }
2221
2222 #ifndef NFS_NOSERVER
2223
2224 /*
2225  * Parse an RPC request
2226  * - verify it
2227  * - fill in the cred struct.
2228  */
2229 int
2230 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
2231 {
2232         int len, i;
2233         u_int32_t *tl;
2234         struct uio uio;
2235         struct iovec iov;
2236         caddr_t cp;
2237         u_int32_t nfsvers, auth_type;
2238         uid_t nickuid;
2239         int error = 0, ticklen;
2240         struct nfsuid *nuidp;
2241         struct timeval tvin, tvout;
2242         struct nfsm_info info;
2243 #if 0                           /* until encrypted keys are implemented */
2244         NFSKERBKEYSCHED_T keys; /* stores key schedule */
2245 #endif
2246
2247         info.mrep = nd->nd_mrep;
2248         info.md = nd->nd_md;
2249         info.dpos = nd->nd_dpos;
2250
2251         if (has_header) {
2252                 NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
2253                 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
2254                 if (*tl++ != rpc_call) {
2255                         m_freem(info.mrep);
2256                         return (EBADRPC);
2257                 }
2258         } else {
2259                 NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
2260         }
2261         nd->nd_repstat = 0;
2262         nd->nd_flag = 0;
2263         if (*tl++ != rpc_vers) {
2264                 nd->nd_repstat = ERPCMISMATCH;
2265                 nd->nd_procnum = NFSPROC_NOOP;
2266                 return (0);
2267         }
2268         if (*tl != nfs_prog) {
2269                 nd->nd_repstat = EPROGUNAVAIL;
2270                 nd->nd_procnum = NFSPROC_NOOP;
2271                 return (0);
2272         }
2273         tl++;
2274         nfsvers = fxdr_unsigned(u_int32_t, *tl++);
2275         if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
2276                 nd->nd_repstat = EPROGMISMATCH;
2277                 nd->nd_procnum = NFSPROC_NOOP;
2278                 return (0);
2279         }
2280         if (nfsvers == NFS_VER3)
2281                 nd->nd_flag = ND_NFSV3;
2282         nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
2283         if (nd->nd_procnum == NFSPROC_NULL)
2284                 return (0);
2285         if (nd->nd_procnum >= NFS_NPROCS ||
2286                 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
2287                 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
2288                 nd->nd_repstat = EPROCUNAVAIL;
2289                 nd->nd_procnum = NFSPROC_NOOP;
2290                 return (0);
2291         }
2292         if ((nd->nd_flag & ND_NFSV3) == 0)
2293                 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
2294         auth_type = *tl++;
2295         len = fxdr_unsigned(int, *tl++);
2296         if (len < 0 || len > RPCAUTH_MAXSIZ) {
2297                 m_freem(info.mrep);
2298                 return (EBADRPC);
2299         }
2300
2301         nd->nd_flag &= ~ND_KERBAUTH;
2302         /*
2303          * Handle auth_unix or auth_kerb.
2304          */
2305         if (auth_type == rpc_auth_unix) {
2306                 len = fxdr_unsigned(int, *++tl);
2307                 if (len < 0 || len > NFS_MAXNAMLEN) {
2308                         m_freem(info.mrep);
2309                         return (EBADRPC);
2310                 }
2311                 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2312                 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2313                 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
2314                 nd->nd_cr.cr_ref = 1;
2315                 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
2316                 nd->nd_cr.cr_ruid = nd->nd_cr.cr_svuid = nd->nd_cr.cr_uid;
2317                 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
2318                 nd->nd_cr.cr_rgid = nd->nd_cr.cr_svgid = nd->nd_cr.cr_gid;
2319                 len = fxdr_unsigned(int, *tl);
2320                 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
2321                         m_freem(info.mrep);
2322                         return (EBADRPC);
2323                 }
2324                 NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
2325                 for (i = 1; i <= len; i++)
2326                     if (i < NGROUPS)
2327                         nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
2328                     else
2329                         tl++;
2330                 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
2331                 if (nd->nd_cr.cr_ngroups > 1)
2332                     nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
2333                 len = fxdr_unsigned(int, *++tl);
2334                 if (len < 0 || len > RPCAUTH_MAXSIZ) {
2335                         m_freem(info.mrep);
2336                         return (EBADRPC);
2337                 }
2338                 if (len > 0) {
2339                         ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2340                 }
2341         } else if (auth_type == rpc_auth_kerb) {
2342                 switch (fxdr_unsigned(int, *tl++)) {
2343                 case RPCAKN_FULLNAME:
2344                         ticklen = fxdr_unsigned(int, *tl);
2345                         *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2346                         uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2347                         nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2348                         if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2349                                 m_freem(info.mrep);
2350                                 return (EBADRPC);
2351                         }
2352                         uio.uio_offset = 0;
2353                         uio.uio_iov = &iov;
2354                         uio.uio_iovcnt = 1;
2355                         uio.uio_segflg = UIO_SYSSPACE;
2356                         iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2357                         iov.iov_len = RPCAUTH_MAXSIZ - 4;
2358                         ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2359                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2360                         if (*tl++ != rpc_auth_kerb ||
2361                                 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2362                                 kprintf("Bad kerb verifier\n");
2363                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2364                                 nd->nd_procnum = NFSPROC_NOOP;
2365                                 return (0);
2366                         }
2367                         NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2368                         tl = (u_int32_t *)cp;
2369                         if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2370                                 kprintf("Not fullname kerb verifier\n");
2371                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2372                                 nd->nd_procnum = NFSPROC_NOOP;
2373                                 return (0);
2374                         }
2375                         cp += NFSX_UNSIGNED;
2376                         bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2377                         nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2378                         nd->nd_flag |= ND_KERBFULL;
2379                         nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2380                         break;
2381                 case RPCAKN_NICKNAME:
2382                         if (len != 2 * NFSX_UNSIGNED) {
2383                                 kprintf("Kerb nickname short\n");
2384                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2385                                 nd->nd_procnum = NFSPROC_NOOP;
2386                                 return (0);
2387                         }
2388                         nickuid = fxdr_unsigned(uid_t, *tl);
2389                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2390                         if (*tl++ != rpc_auth_kerb ||
2391                                 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2392                                 kprintf("Kerb nick verifier bad\n");
2393                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2394                                 nd->nd_procnum = NFSPROC_NOOP;
2395                                 return (0);
2396                         }
2397                         NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2398                         tvin.tv_sec = *tl++;
2399                         tvin.tv_usec = *tl;
2400
2401                         for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2402                             nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2403                                 if (nuidp->nu_cr.cr_uid == nickuid &&
2404                                     (!nd->nd_nam2 ||
2405                                      netaddr_match(NU_NETFAM(nuidp),
2406                                       &nuidp->nu_haddr, nd->nd_nam2)))
2407                                         break;
2408                         }
2409                         if (!nuidp) {
2410                                 nd->nd_repstat =
2411                                         (NFSERR_AUTHERR|AUTH_REJECTCRED);
2412                                 nd->nd_procnum = NFSPROC_NOOP;
2413                                 return (0);
2414                         }
2415
2416                         /*
2417                          * Now, decrypt the timestamp using the session key
2418                          * and validate it.
2419                          */
2420 #ifdef NFSKERB
2421                         XXX
2422 #else
2423                         tvout.tv_sec = 0;
2424                         tvout.tv_usec = 0;
2425 #endif
2426
2427                         tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2428                         tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2429                         if (nuidp->nu_expire < time_second ||
2430                             nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2431                             (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2432                              nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2433                                 nuidp->nu_expire = 0;
2434                                 nd->nd_repstat =
2435                                     (NFSERR_AUTHERR|AUTH_REJECTVERF);
2436                                 nd->nd_procnum = NFSPROC_NOOP;
2437                                 return (0);
2438                         }
2439                         nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2440                         nd->nd_flag |= ND_KERBNICK;
2441                 };
2442         } else {
2443                 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2444                 nd->nd_procnum = NFSPROC_NOOP;
2445                 return (0);
2446         }
2447
2448         nd->nd_md = info.md;
2449         nd->nd_dpos = info.dpos;
2450         return (0);
2451 nfsmout:
2452         return (error);
2453 }
2454
2455 #endif
2456
2457 /*
2458  * Send a message to the originating process's terminal.  The thread and/or
2459  * process may be NULL.  YYY the thread should not be NULL but there may
2460  * still be some uio_td's that are still being passed as NULL through to
2461  * nfsm_request().
2462  */
2463 static int
2464 nfs_msg(struct thread *td, char *server, char *msg)
2465 {
2466         tpr_t tpr;
2467
2468         if (td && td->td_proc)
2469                 tpr = tprintf_open(td->td_proc);
2470         else
2471                 tpr = NULL;
2472         tprintf(tpr, "nfs server %s: %s\n", server, msg);
2473         tprintf_close(tpr);
2474         return (0);
2475 }
2476
2477 #ifndef NFS_NOSERVER
2478
2479 void
2480 nfsrv_rcv_upcall(struct socket *so, void *arg, int waitflag)
2481 {
2482         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2483
2484         lwkt_gettoken(&slp->ns_token);
2485         nfsrv_rcv(so, arg, waitflag);
2486         lwkt_reltoken(&slp->ns_token);
2487 }
2488
2489 /*
2490  * Socket upcall routine for the nfsd sockets.
2491  * The caddr_t arg is a pointer to the "struct nfssvc_sock".
2492  * Essentially do as much as possible non-blocking, else punt and it will
2493  * be called with MB_WAIT from an nfsd.
2494  *
2495  * slp->ns_token is held on call
2496  */
2497 void
2498 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2499 {
2500         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2501         struct mbuf *m;
2502         struct sockaddr *nam;
2503         struct sockbuf sio;
2504         int flags, error;
2505         int nparallel_wakeup = 0;
2506
2507         ASSERT_LWKT_TOKEN_HELD(&slp->ns_token);
2508
2509         if ((slp->ns_flag & SLP_VALID) == 0)
2510                 return;
2511
2512         /*
2513          * Do not allow an infinite number of completed RPC records to build 
2514          * up before we stop reading data from the socket.  Otherwise we could
2515          * end up holding onto an unreasonable number of mbufs for requests
2516          * waiting for service.
2517          *
2518          * This should give pretty good feedback to the TCP
2519          * layer and prevents a memory crunch for other protocols.
2520          *
2521          * Note that the same service socket can be dispatched to several
2522          * nfs servers simultaniously.
2523          *
2524          * the tcp protocol callback calls us with MB_DONTWAIT.  
2525          * nfsd calls us with MB_WAIT (typically).
2526          */
2527         if (waitflag == MB_DONTWAIT && slp->ns_numrec >= nfsd_waiting / 2 + 1) {
2528                 slp->ns_flag |= SLP_NEEDQ;
2529                 goto dorecs;
2530         }
2531
2532         /*
2533          * Handle protocol specifics to parse an RPC request.  We always
2534          * pull from the socket using non-blocking I/O.
2535          */
2536         if (so->so_type == SOCK_STREAM) {
2537                 /*
2538                  * The data has to be read in an orderly fashion from a TCP
2539                  * stream, unlike a UDP socket.  It is possible for soreceive
2540                  * and/or nfsrv_getstream() to block, so make sure only one
2541                  * entity is messing around with the TCP stream at any given
2542                  * moment.  The receive sockbuf's lock in soreceive is not
2543                  * sufficient.
2544                  *
2545                  * Note that this procedure can be called from any number of
2546                  * NFS severs *OR* can be upcalled directly from a TCP
2547                  * protocol thread without the lock.
2548                  */
2549                 if (slp->ns_flag & SLP_GETSTREAM) {
2550                         slp->ns_flag |= SLP_NEEDQ;
2551                         goto dorecs;
2552                 }
2553                 slp->ns_flag |= SLP_GETSTREAM;
2554
2555                 /*
2556                  * Do soreceive().  Pull out as much data as possible without
2557                  * blocking.
2558                  */
2559                 sbinit(&sio, 1000000000);
2560                 flags = MSG_DONTWAIT;
2561                 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2562                 if (error || sio.sb_mb == NULL) {
2563                         if (error == EWOULDBLOCK)
2564                                 slp->ns_flag |= SLP_NEEDQ;
2565                         else
2566                                 slp->ns_flag |= SLP_DISCONN;
2567                         slp->ns_flag &= ~SLP_GETSTREAM;
2568                         goto dorecs;
2569                 }
2570                 m = sio.sb_mb;
2571                 if (slp->ns_rawend) {
2572                         slp->ns_rawend->m_next = m;
2573                         slp->ns_cc += sio.sb_cc;
2574                 } else {
2575                         slp->ns_raw = m;
2576                         slp->ns_cc = sio.sb_cc;
2577                 }
2578                 while (m->m_next)
2579                         m = m->m_next;
2580                 slp->ns_rawend = m;
2581
2582                 /*
2583                  * Now try and parse as many record(s) as we can out of the
2584                  * raw stream data.
2585                  */
2586                 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2587                 if (error) {
2588                         if (error == EPERM)
2589                                 slp->ns_flag |= SLP_DISCONN;
2590                         else
2591                                 slp->ns_flag |= SLP_NEEDQ;
2592                 }
2593                 slp->ns_flag &= ~SLP_GETSTREAM;
2594         } else {
2595                 /*
2596                  * For UDP soreceive typically pulls just one packet, loop
2597                  * to get the whole batch.
2598                  */
2599                 do {
2600                         sbinit(&sio, 1000000000);
2601                         flags = MSG_DONTWAIT;
2602                         error = so_pru_soreceive(so, &nam, NULL, &sio,
2603                                                  NULL, &flags);
2604                         if (sio.sb_mb) {
2605                                 struct nfsrv_rec *rec;
2606                                 int mf = (waitflag & MB_DONTWAIT) ?
2607                                             M_NOWAIT : M_WAITOK;
2608                                 rec = kmalloc(sizeof(struct nfsrv_rec),
2609                                              M_NFSRVDESC, mf);
2610                                 if (!rec) {
2611                                         if (nam)
2612                                                 FREE(nam, M_SONAME);
2613                                         m_freem(sio.sb_mb);
2614                                         continue;
2615                                 }
2616                                 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2617                                 rec->nr_address = nam;
2618                                 rec->nr_packet = sio.sb_mb;
2619                                 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2620                                 ++slp->ns_numrec;
2621                                 ++nparallel_wakeup;
2622                         }
2623                         if (error) {
2624                                 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2625                                         && error != EWOULDBLOCK) {
2626                                         slp->ns_flag |= SLP_DISCONN;
2627                                         goto dorecs;
2628                                 }
2629                         }
2630                 } while (sio.sb_mb);
2631         }
2632
2633         /*
2634          * If we were upcalled from the tcp protocol layer and we have
2635          * fully parsed records ready to go, or there is new data pending,
2636          * or something went wrong, try to wake up an nfsd thread to deal
2637          * with it.
2638          */
2639 dorecs:
2640         if (waitflag == MB_DONTWAIT && (slp->ns_numrec > 0
2641              || (slp->ns_flag & (SLP_NEEDQ | SLP_DISCONN)))) {
2642                 nfsrv_wakenfsd(slp, nparallel_wakeup);
2643         }
2644 }
2645
2646 /*
2647  * Try and extract an RPC request from the mbuf data list received on a
2648  * stream socket. The "waitflag" argument indicates whether or not it
2649  * can sleep.
2650  */
2651 static int
2652 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2653 {
2654         struct mbuf *m, **mpp;
2655         char *cp1, *cp2;
2656         int len;
2657         struct mbuf *om, *m2, *recm;
2658         u_int32_t recmark;
2659
2660         for (;;) {
2661             if (slp->ns_reclen == 0) {
2662                 if (slp->ns_cc < NFSX_UNSIGNED)
2663                         return (0);
2664                 m = slp->ns_raw;
2665                 if (m->m_len >= NFSX_UNSIGNED) {
2666                         bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2667                         m->m_data += NFSX_UNSIGNED;
2668                         m->m_len -= NFSX_UNSIGNED;
2669                 } else {
2670                         cp1 = (caddr_t)&recmark;
2671                         cp2 = mtod(m, caddr_t);
2672                         while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2673                                 while (m->m_len == 0) {
2674                                         m = m->m_next;
2675                                         cp2 = mtod(m, caddr_t);
2676                                 }
2677                                 *cp1++ = *cp2++;
2678                                 m->m_data++;
2679                                 m->m_len--;
2680                         }
2681                 }
2682                 slp->ns_cc -= NFSX_UNSIGNED;
2683                 recmark = ntohl(recmark);
2684                 slp->ns_reclen = recmark & ~0x80000000;
2685                 if (recmark & 0x80000000)
2686                         slp->ns_flag |= SLP_LASTFRAG;
2687                 else
2688                         slp->ns_flag &= ~SLP_LASTFRAG;
2689                 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2690                         log(LOG_ERR, "%s (%d) from nfs client\n",
2691                             "impossible packet length",
2692                             slp->ns_reclen);
2693                         return (EPERM);
2694                 }
2695             }
2696
2697             /*
2698              * Now get the record part.
2699              *
2700              * Note that slp->ns_reclen may be 0.  Linux sometimes
2701              * generates 0-length RPCs
2702              */
2703             recm = NULL;
2704             if (slp->ns_cc == slp->ns_reclen) {
2705                 recm = slp->ns_raw;
2706                 slp->ns_raw = slp->ns_rawend = NULL;
2707                 slp->ns_cc = slp->ns_reclen = 0;
2708             } else if (slp->ns_cc > slp->ns_reclen) {
2709                 len = 0;
2710                 m = slp->ns_raw;
2711                 om = NULL;
2712
2713                 while (len < slp->ns_reclen) {
2714                         if ((len + m->m_len) > slp->ns_reclen) {
2715                                 m2 = m_copym(m, 0, slp->ns_reclen - len,
2716                                         waitflag);
2717                                 if (m2) {
2718                                         if (om) {
2719                                                 om->m_next = m2;
2720                                                 recm = slp->ns_raw;
2721                                         } else
2722                                                 recm = m2;
2723                                         m->m_data += slp->ns_reclen - len;
2724                                         m->m_len -= slp->ns_reclen - len;
2725                                         len = slp->ns_reclen;
2726                                 } else {
2727                                         return (EWOULDBLOCK);
2728                                 }
2729                         } else if ((len + m->m_len) == slp->ns_reclen) {
2730                                 om = m;
2731                                 len += m->m_len;
2732                                 m = m->m_next;
2733                                 recm = slp->ns_raw;
2734                                 om->m_next = NULL;
2735                         } else {
2736                                 om = m;
2737                                 len += m->m_len;
2738                                 m = m->m_next;
2739                         }
2740                 }
2741                 slp->ns_raw = m;
2742                 slp->ns_cc -= len;
2743                 slp->ns_reclen = 0;
2744             } else {
2745                 return (0);
2746             }
2747
2748             /*
2749              * Accumulate the fragments into a record.
2750              */
2751             mpp = &slp->ns_frag;
2752             while (*mpp)
2753                 mpp = &((*mpp)->m_next);
2754             *mpp = recm;
2755             if (slp->ns_flag & SLP_LASTFRAG) {
2756                 struct nfsrv_rec *rec;
2757                 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2758                 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2759                 if (!rec) {
2760                     m_freem(slp->ns_frag);
2761                 } else {
2762                     nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2763                     rec->nr_address = NULL;
2764                     rec->nr_packet = slp->ns_frag;
2765                     STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2766                     ++slp->ns_numrec;
2767                     ++*countp;
2768                 }
2769                 slp->ns_frag = NULL;
2770             }
2771         }
2772 }
2773
2774 #ifdef INVARIANTS
2775
2776 /*
2777  * Sanity check our mbuf chain.
2778  */
2779 static void
2780 nfs_checkpkt(struct mbuf *m, int len)
2781 {
2782         int xlen = 0;
2783         while (m) {
2784                 xlen += m->m_len;
2785                 m = m->m_next;
2786         }
2787         if (xlen != len) {
2788                 panic("nfs_checkpkt: len mismatch %d/%d mbuf %p\n",
2789                         xlen, len, m);
2790         }
2791 }
2792
2793 #else
2794
2795 static void
2796 nfs_checkpkt(struct mbuf *m __unused, int len __unused)
2797 {
2798 }
2799
2800 #endif
2801
2802 /*
2803  * Parse an RPC header.
2804  */
2805 int
2806 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2807             struct nfsrv_descript **ndp)
2808 {
2809         struct nfsrv_rec *rec;
2810         struct mbuf *m;
2811         struct sockaddr *nam;
2812         struct nfsrv_descript *nd;
2813         int error;
2814
2815         *ndp = NULL;
2816         if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2817                 return (ENOBUFS);
2818         rec = STAILQ_FIRST(&slp->ns_rec);
2819         STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2820         KKASSERT(slp->ns_numrec > 0);
2821         --slp->ns_numrec;
2822         nam = rec->nr_address;
2823         m = rec->nr_packet;
2824         kfree(rec, M_NFSRVDESC);
2825         MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2826                 M_NFSRVDESC, M_WAITOK);
2827         nd->nd_md = nd->nd_mrep = m;
2828         nd->nd_nam2 = nam;
2829         nd->nd_dpos = mtod(m, caddr_t);
2830         error = nfs_getreq(nd, nfsd, TRUE);
2831         if (error) {
2832                 if (nam) {
2833                         FREE(nam, M_SONAME);
2834                 }
2835                 kfree((caddr_t)nd, M_NFSRVDESC);
2836                 return (error);
2837         }
2838         *ndp = nd;
2839         nfsd->nfsd_nd = nd;
2840         return (0);
2841 }
2842
2843 /*
2844  * Try to assign service sockets to nfsd threads based on the number
2845  * of new rpc requests that have been queued on the service socket.
2846  *
2847  * If no nfsd's are available or additonal requests are pending, set the
2848  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2849  * the work in the nfssvc_sock list when it is finished processing its
2850  * current work.  This flag is only cleared when an nfsd can not find
2851  * any new work to perform.
2852  */
2853 void
2854 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2855 {
2856         struct nfsd *nd;
2857
2858         if ((slp->ns_flag & SLP_VALID) == 0)
2859                 return;
2860         if (nparallel <= 1)
2861                 nparallel = 1;
2862         TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2863                 if (nd->nfsd_flag & NFSD_WAITING) {
2864                         nd->nfsd_flag &= ~NFSD_WAITING;
2865                         if (nd->nfsd_slp)
2866                                 panic("nfsd wakeup");
2867                         slp->ns_sref++;
2868                         nd->nfsd_slp = slp;
2869                         wakeup((caddr_t)nd);
2870                         if (--nparallel == 0)
2871                                 break;
2872                 }
2873         }
2874         if (nparallel) {
2875                 slp->ns_flag |= SLP_DOREC;
2876                 nfsd_head_flag |= NFSD_CHECKSLP;
2877         }
2878 }
2879 #endif /* NFS_NOSERVER */