kernel - Add many sysctl definitions, sysv, vfs, nfs, etc.
[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
65 #include <sys/signal2.h>
66 #include <sys/mutex2.h>
67 #include <sys/socketvar2.h>
68
69 #include <netinet/in.h>
70 #include <netinet/tcp.h>
71 #include <sys/thread2.h>
72
73 #include "rpcv2.h"
74 #include "nfsproto.h"
75 #include "nfs.h"
76 #include "xdr_subs.h"
77 #include "nfsm_subs.h"
78 #include "nfsmount.h"
79 #include "nfsnode.h"
80 #include "nfsrtt.h"
81
82 #define TRUE    1
83 #define FALSE   0
84
85 /*
86  * RTT calculations are scaled by 256 (8 bits).  A proper fractional
87  * RTT will still be calculated even with a slow NFS timer.
88  */
89 #define NFS_SRTT(r)     (r)->r_nmp->nm_srtt[proct[(r)->r_procnum]]
90 #define NFS_SDRTT(r)    (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum]]
91 #define NFS_RTT_SCALE_BITS      8       /* bits */
92 #define NFS_RTT_SCALE           256     /* value */
93
94 /*
95  * Defines which timer to use for the procnum.
96  * 0 - default
97  * 1 - getattr
98  * 2 - lookup
99  * 3 - read
100  * 4 - write
101  */
102 static int proct[NFS_NPROCS] = {
103         0, 1, 0, 2, 1, 3, 3, 4, 0, 0,   /* 00-09        */
104         0, 0, 0, 0, 0, 0, 3, 3, 0, 0,   /* 10-19        */
105         0, 5, 0, 0, 0, 0,               /* 20-29        */
106 };
107
108 static int multt[NFS_NPROCS] = {
109         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 00-09        */
110         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 10-19        */
111         1, 2, 1, 1, 1, 1,               /* 20-29        */
112 };
113
114 static int nfs_backoff[8] = { 2, 3, 5, 8, 13, 21, 34, 55 };
115 static int nfs_realign_test;
116 static int nfs_realign_count;
117 static int nfs_showrtt;
118 static int nfs_showrexmit;
119 int nfs_maxasyncbio = NFS_MAXASYNCBIO;
120
121 SYSCTL_DECL(_vfs_nfs);
122
123 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0,
124     "Number of times mbufs have been tested for bad alignment");
125 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0,
126     "Number of realignments for badly aligned mbuf data");
127 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrtt, CTLFLAG_RW, &nfs_showrtt, 0,
128     "Show round trip time output");
129 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrexmit, CTLFLAG_RW, &nfs_showrexmit, 0,
130     "Show retransmits info");
131 SYSCTL_INT(_vfs_nfs, OID_AUTO, maxasyncbio, CTLFLAG_RW, &nfs_maxasyncbio, 0,
132     "Max number of asynchronous bio's");
133
134 static int nfs_request_setup(nfsm_info_t info);
135 static int nfs_request_auth(struct nfsreq *rep);
136 static int nfs_request_try(struct nfsreq *rep);
137 static int nfs_request_waitreply(struct nfsreq *rep);
138 static int nfs_request_processreply(nfsm_info_t info, int);
139
140 int nfsrtton = 0;
141 struct nfsrtt nfsrtt;
142 struct callout  nfs_timer_handle;
143
144 static int      nfs_msg (struct thread *,char *,char *);
145 static int      nfs_rcvlock (struct nfsmount *nmp, struct nfsreq *myreq);
146 static void     nfs_rcvunlock (struct nfsmount *nmp);
147 static void     nfs_realign (struct mbuf **pm, int hsiz);
148 static int      nfs_receive (struct nfsmount *nmp, struct nfsreq *rep,
149                                 struct sockaddr **aname, struct mbuf **mp);
150 static void     nfs_softterm (struct nfsreq *rep, int islocked);
151 static void     nfs_hardterm (struct nfsreq *rep, int islocked);
152 static int      nfs_reconnect (struct nfsmount *nmp, struct nfsreq *rep);
153 #ifndef NFS_NOSERVER 
154 static int      nfsrv_getstream (struct nfssvc_sock *, int, int *);
155 static void     nfs_timer_req(struct nfsreq *req);
156 static void     nfs_checkpkt(struct mbuf *m, int len);
157
158 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
159                                     struct nfssvc_sock *slp,
160                                     struct thread *td,
161                                     struct mbuf **mreqp) = {
162         nfsrv_null,
163         nfsrv_getattr,
164         nfsrv_setattr,
165         nfsrv_lookup,
166         nfsrv3_access,
167         nfsrv_readlink,
168         nfsrv_read,
169         nfsrv_write,
170         nfsrv_create,
171         nfsrv_mkdir,
172         nfsrv_symlink,
173         nfsrv_mknod,
174         nfsrv_remove,
175         nfsrv_rmdir,
176         nfsrv_rename,
177         nfsrv_link,
178         nfsrv_readdir,
179         nfsrv_readdirplus,
180         nfsrv_statfs,
181         nfsrv_fsinfo,
182         nfsrv_pathconf,
183         nfsrv_commit,
184         nfsrv_noop,
185         nfsrv_noop,
186         nfsrv_noop,
187         nfsrv_noop
188 };
189 #endif /* NFS_NOSERVER */
190
191 /*
192  * Initialize sockets and congestion for a new NFS connection.
193  * We do not free the sockaddr if error.
194  */
195 int
196 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
197 {
198         struct socket *so;
199         int error;
200         struct sockaddr *saddr;
201         struct sockaddr_in *sin;
202         struct thread *td = &thread0; /* only used for socreate and sobind */
203
204         nmp->nm_so = so = NULL;
205         if (nmp->nm_flag & NFSMNT_FORCE)
206                 return (EINVAL);
207         saddr = nmp->nm_nam;
208         error = socreate(saddr->sa_family, &so, nmp->nm_sotype,
209                 nmp->nm_soproto, td);
210         if (error)
211                 goto bad;
212         nmp->nm_soflags = so->so_proto->pr_flags;
213
214         /*
215          * Some servers require that the client port be a reserved port number.
216          */
217         if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
218                 struct sockopt sopt;
219                 int ip;
220                 struct sockaddr_in ssin;
221
222                 bzero(&sopt, sizeof sopt);
223                 ip = IP_PORTRANGE_LOW;
224                 sopt.sopt_level = IPPROTO_IP;
225                 sopt.sopt_name = IP_PORTRANGE;
226                 sopt.sopt_val = (void *)&ip;
227                 sopt.sopt_valsize = sizeof(ip);
228                 sopt.sopt_td = NULL;
229                 error = sosetopt(so, &sopt);
230                 if (error)
231                         goto bad;
232                 bzero(&ssin, sizeof ssin);
233                 sin = &ssin;
234                 sin->sin_len = sizeof (struct sockaddr_in);
235                 sin->sin_family = AF_INET;
236                 sin->sin_addr.s_addr = INADDR_ANY;
237                 sin->sin_port = htons(0);
238                 error = sobind(so, (struct sockaddr *)sin, td);
239                 if (error)
240                         goto bad;
241                 bzero(&sopt, sizeof sopt);
242                 ip = IP_PORTRANGE_DEFAULT;
243                 sopt.sopt_level = IPPROTO_IP;
244                 sopt.sopt_name = IP_PORTRANGE;
245                 sopt.sopt_val = (void *)&ip;
246                 sopt.sopt_valsize = sizeof(ip);
247                 sopt.sopt_td = NULL;
248                 error = sosetopt(so, &sopt);
249                 if (error)
250                         goto bad;
251         }
252
253         /*
254          * Protocols that do not require connections may be optionally left
255          * unconnected for servers that reply from a port other than NFS_PORT.
256          */
257         if (nmp->nm_flag & NFSMNT_NOCONN) {
258                 if (nmp->nm_soflags & PR_CONNREQUIRED) {
259                         error = ENOTCONN;
260                         goto bad;
261                 }
262         } else {
263                 error = soconnect(so, nmp->nm_nam, td);
264                 if (error)
265                         goto bad;
266
267                 /*
268                  * Wait for the connection to complete. Cribbed from the
269                  * connect system call but with the wait timing out so
270                  * that interruptible mounts don't hang here for a long time.
271                  */
272                 crit_enter();
273                 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
274                         (void) tsleep((caddr_t)&so->so_timeo, 0,
275                                 "nfscon", 2 * hz);
276                         if ((so->so_state & SS_ISCONNECTING) &&
277                             so->so_error == 0 && rep &&
278                             (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
279                                 soclrstate(so, SS_ISCONNECTING);
280                                 crit_exit();
281                                 goto bad;
282                         }
283                 }
284                 if (so->so_error) {
285                         error = so->so_error;
286                         so->so_error = 0;
287                         crit_exit();
288                         goto bad;
289                 }
290                 crit_exit();
291         }
292         so->so_rcv.ssb_timeo = (5 * hz);
293         so->so_snd.ssb_timeo = (5 * hz);
294
295         /*
296          * Get buffer reservation size from sysctl, but impose reasonable
297          * limits.
298          */
299         if (nmp->nm_sotype == SOCK_STREAM) {
300                 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
301                         struct sockopt sopt;
302                         int val;
303
304                         bzero(&sopt, sizeof sopt);
305                         sopt.sopt_level = SOL_SOCKET;
306                         sopt.sopt_name = SO_KEEPALIVE;
307                         sopt.sopt_val = &val;
308                         sopt.sopt_valsize = sizeof val;
309                         val = 1;
310                         sosetopt(so, &sopt);
311                 }
312                 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
313                         struct sockopt sopt;
314                         int val;
315
316                         bzero(&sopt, sizeof sopt);
317                         sopt.sopt_level = IPPROTO_TCP;
318                         sopt.sopt_name = TCP_NODELAY;
319                         sopt.sopt_val = &val;
320                         sopt.sopt_valsize = sizeof val;
321                         val = 1;
322                         sosetopt(so, &sopt);
323
324                         bzero(&sopt, sizeof sopt);
325                         sopt.sopt_level = IPPROTO_TCP;
326                         sopt.sopt_name = TCP_FASTKEEP;
327                         sopt.sopt_val = &val;
328                         sopt.sopt_valsize = sizeof val;
329                         val = 1;
330                         sosetopt(so, &sopt);
331                 }
332         }
333         error = soreserve(so, nfs_soreserve, nfs_soreserve, NULL);
334         if (error)
335                 goto bad;
336         atomic_set_int(&so->so_rcv.ssb_flags, SSB_NOINTR);
337         atomic_set_int(&so->so_snd.ssb_flags, SSB_NOINTR);
338
339         /* Initialize other non-zero congestion variables */
340         nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] = 
341                 nmp->nm_srtt[3] = (NFS_TIMEO << NFS_RTT_SCALE_BITS);
342         nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
343                 nmp->nm_sdrtt[3] = 0;
344         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
345         nmp->nm_timeouts = 0;
346
347         /*
348          * Assign nm_so last.  The moment nm_so is assigned the nfs_timer()
349          * can mess with the socket.
350          */
351         nmp->nm_so = so;
352         return (0);
353
354 bad:
355         if (so) {
356                 soshutdown(so, SHUT_RDWR);
357                 soclose(so, FNONBLOCK);
358         }
359         return (error);
360 }
361
362 /*
363  * Reconnect routine:
364  * Called when a connection is broken on a reliable protocol.
365  * - clean up the old socket
366  * - nfs_connect() again
367  * - set R_NEEDSXMIT for all outstanding requests on mount point
368  * If this fails the mount point is DEAD!
369  * nb: Must be called with the nfs_sndlock() set on the mount point.
370  */
371 static int
372 nfs_reconnect(struct nfsmount *nmp, struct nfsreq *rep)
373 {
374         struct nfsreq *req;
375         int error;
376
377         nfs_disconnect(nmp);
378         if (nmp->nm_rxstate >= NFSSVC_STOPPING)
379                 return (EINTR);
380         while ((error = nfs_connect(nmp, rep)) != 0) {
381                 if (error == EINTR || error == ERESTART)
382                         return (EINTR);
383                 if (error == EINVAL)
384                         return (error);
385                 if (nmp->nm_rxstate >= NFSSVC_STOPPING)
386                         return (EINTR);
387                 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
388         }
389
390         /*
391          * Loop through outstanding request list and fix up all requests
392          * on old socket.
393          */
394         crit_enter();
395         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
396                 KKASSERT(req->r_nmp == nmp);
397                 req->r_flags |= R_NEEDSXMIT;
398         }
399         crit_exit();
400         return (0);
401 }
402
403 /*
404  * NFS disconnect. Clean up and unlink.
405  */
406 void
407 nfs_disconnect(struct nfsmount *nmp)
408 {
409         struct socket *so;
410
411         if (nmp->nm_so) {
412                 so = nmp->nm_so;
413                 nmp->nm_so = NULL;
414                 soshutdown(so, SHUT_RDWR);
415                 soclose(so, FNONBLOCK);
416         }
417 }
418
419 void
420 nfs_safedisconnect(struct nfsmount *nmp)
421 {
422         nfs_rcvlock(nmp, NULL);
423         nfs_disconnect(nmp);
424         nfs_rcvunlock(nmp);
425 }
426
427 /*
428  * This is the nfs send routine. For connection based socket types, it
429  * must be called with an nfs_sndlock() on the socket.
430  * "rep == NULL" indicates that it has been called from a server.
431  * For the client side:
432  * - return EINTR if the RPC is terminated, 0 otherwise
433  * - set R_NEEDSXMIT if the send fails for any reason
434  * - do any cleanup required by recoverable socket errors (?)
435  * For the server side:
436  * - return EINTR or ERESTART if interrupted by a signal
437  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
438  * - do any cleanup required by recoverable socket errors (?)
439  */
440 int
441 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
442          struct nfsreq *rep)
443 {
444         struct sockaddr *sendnam;
445         int error, soflags, flags;
446
447         if (rep) {
448                 if (rep->r_flags & R_SOFTTERM) {
449                         m_freem(top);
450                         return (EINTR);
451                 }
452                 if ((so = rep->r_nmp->nm_so) == NULL) {
453                         rep->r_flags |= R_NEEDSXMIT;
454                         m_freem(top);
455                         return (0);
456                 }
457                 rep->r_flags &= ~R_NEEDSXMIT;
458                 soflags = rep->r_nmp->nm_soflags;
459         } else {
460                 soflags = so->so_proto->pr_flags;
461         }
462         if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
463                 sendnam = NULL;
464         else
465                 sendnam = nam;
466         if (so->so_type == SOCK_SEQPACKET)
467                 flags = MSG_EOR;
468         else
469                 flags = 0;
470
471         /*
472          * calls pru_sosend -> sosend -> so_pru_send -> netrpc
473          */
474         error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
475                               curthread /*XXX*/);
476
477         /*
478          * ENOBUFS for dgram sockets is transient and non fatal.
479          * No need to log, and no need to break a soft mount.
480          */
481         if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
482                 error = 0;
483                 /*
484                  * do backoff retransmit on client
485                  */
486                 if (rep) {
487                         if ((rep->r_nmp->nm_state & NFSSTA_SENDSPACE) == 0) {
488                                 rep->r_nmp->nm_state |= NFSSTA_SENDSPACE;
489                                 kprintf("Warning: NFS: Insufficient sendspace "
490                                         "(%lu),\n"
491                                         "\t You must increase vfs.nfs.soreserve"
492                                         "or decrease vfs.nfs.maxasyncbio\n",
493                                         so->so_snd.ssb_hiwat);
494                         }
495                         rep->r_flags |= R_NEEDSXMIT;
496                 }
497         }
498
499         if (error) {
500                 if (rep) {
501                         log(LOG_INFO, "nfs send error %d for server %s\n",error,
502                             rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
503                         /*
504                          * Deal with errors for the client side.
505                          */
506                         if (rep->r_flags & R_SOFTTERM)
507                                 error = EINTR;
508                         else
509                                 rep->r_flags |= R_NEEDSXMIT;
510                 } else {
511                         log(LOG_INFO, "nfsd send error %d\n", error);
512                 }
513
514                 /*
515                  * Handle any recoverable (soft) socket errors here. (?)
516                  */
517                 if (error != EINTR && error != ERESTART &&
518                         error != EWOULDBLOCK && error != EPIPE)
519                         error = 0;
520         }
521         return (error);
522 }
523
524 /*
525  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
526  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
527  * Mark and consolidate the data into a new mbuf list.
528  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
529  *     small mbufs.
530  * For SOCK_STREAM we must be very careful to read an entire record once
531  * we have read any of it, even if the system call has been interrupted.
532  */
533 static int
534 nfs_receive(struct nfsmount *nmp, struct nfsreq *rep,
535             struct sockaddr **aname, struct mbuf **mp)
536 {
537         struct socket *so;
538         struct sockbuf sio;
539         struct uio auio;
540         struct iovec aio;
541         struct mbuf *m;
542         struct mbuf *control;
543         u_int32_t len;
544         struct sockaddr **getnam;
545         int error, sotype, rcvflg;
546         struct thread *td = curthread;  /* XXX */
547
548         /*
549          * Set up arguments for soreceive()
550          */
551         *mp = NULL;
552         *aname = NULL;
553         sotype = nmp->nm_sotype;
554
555         /*
556          * For reliable protocols, lock against other senders/receivers
557          * in case a reconnect is necessary.
558          * For SOCK_STREAM, first get the Record Mark to find out how much
559          * more there is to get.
560          * We must lock the socket against other receivers
561          * until we have an entire rpc request/reply.
562          */
563         if (sotype != SOCK_DGRAM) {
564                 error = nfs_sndlock(nmp, rep);
565                 if (error)
566                         return (error);
567 tryagain:
568                 /*
569                  * Check for fatal errors and resending request.
570                  */
571                 /*
572                  * Ugh: If a reconnect attempt just happened, nm_so
573                  * would have changed. NULL indicates a failed
574                  * attempt that has essentially shut down this
575                  * mount point.
576                  */
577                 if (rep && (rep->r_mrep || (rep->r_flags & R_SOFTTERM))) {
578                         nfs_sndunlock(nmp);
579                         return (EINTR);
580                 }
581                 so = nmp->nm_so;
582                 if (so == NULL) {
583                         error = nfs_reconnect(nmp, rep);
584                         if (error) {
585                                 nfs_sndunlock(nmp);
586                                 return (error);
587                         }
588                         goto tryagain;
589                 }
590                 while (rep && (rep->r_flags & R_NEEDSXMIT)) {
591                         m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
592                         nfsstats.rpcretries++;
593                         error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
594                         if (error) {
595                                 if (error == EINTR || error == ERESTART ||
596                                     (error = nfs_reconnect(nmp, rep)) != 0) {
597                                         nfs_sndunlock(nmp);
598                                         return (error);
599                                 }
600                                 goto tryagain;
601                         }
602                 }
603                 nfs_sndunlock(nmp);
604                 if (sotype == SOCK_STREAM) {
605                         /*
606                          * Get the length marker from the stream
607                          */
608                         aio.iov_base = (caddr_t)&len;
609                         aio.iov_len = sizeof(u_int32_t);
610                         auio.uio_iov = &aio;
611                         auio.uio_iovcnt = 1;
612                         auio.uio_segflg = UIO_SYSSPACE;
613                         auio.uio_rw = UIO_READ;
614                         auio.uio_offset = 0;
615                         auio.uio_resid = sizeof(u_int32_t);
616                         auio.uio_td = td;
617                         do {
618                            rcvflg = MSG_WAITALL;
619                            error = so_pru_soreceive(so, NULL, &auio, NULL,
620                                                     NULL, &rcvflg);
621                            if (error == EWOULDBLOCK && rep) {
622                                 if (rep->r_flags & R_SOFTTERM)
623                                         return (EINTR);
624                            }
625                         } while (error == EWOULDBLOCK);
626
627                         if (error == 0 && auio.uio_resid > 0) {
628                             /*
629                              * Only log short packets if not EOF
630                              */
631                             if (auio.uio_resid != sizeof(u_int32_t))
632                             log(LOG_INFO,
633                                  "short receive (%d/%d) from nfs server %s\n",
634                                  (int)(sizeof(u_int32_t) - auio.uio_resid),
635                                  (int)sizeof(u_int32_t),
636                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
637                             error = EPIPE;
638                         }
639                         if (error)
640                                 goto errout;
641                         len = ntohl(len) & ~0x80000000;
642                         /*
643                          * This is SERIOUS! We are out of sync with the sender
644                          * and forcing a disconnect/reconnect is all I can do.
645                          */
646                         if (len > NFS_MAXPACKET) {
647                             log(LOG_ERR, "%s (%d) from nfs server %s\n",
648                                 "impossible packet length",
649                                 len,
650                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
651                             error = EFBIG;
652                             goto errout;
653                         }
654
655                         /*
656                          * Get the rest of the packet as an mbuf chain
657                          */
658                         sbinit(&sio, len);
659                         do {
660                             rcvflg = MSG_WAITALL;
661                             error = so_pru_soreceive(so, NULL, NULL, &sio,
662                                                      NULL, &rcvflg);
663                         } while (error == EWOULDBLOCK || error == EINTR ||
664                                  error == ERESTART);
665                         if (error == 0 && sio.sb_cc != len) {
666                             if (sio.sb_cc != 0)
667                             log(LOG_INFO,
668                                 "short receive (%zu/%d) from nfs server %s\n",
669                                 (size_t)len - auio.uio_resid, len,
670                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
671                             error = EPIPE;
672                         }
673                         *mp = sio.sb_mb;
674                 } else {
675                         /*
676                          * Non-stream, so get the whole packet by not
677                          * specifying MSG_WAITALL and by specifying a large
678                          * length.
679                          *
680                          * We have no use for control msg., but must grab them
681                          * and then throw them away so we know what is going
682                          * on.
683                          */
684                         sbinit(&sio, 100000000);
685                         do {
686                             rcvflg = 0;
687                             error =  so_pru_soreceive(so, NULL, NULL, &sio,
688                                                       &control, &rcvflg);
689                             if (control)
690                                 m_freem(control);
691                             if (error == EWOULDBLOCK && rep) {
692                                 if (rep->r_flags & R_SOFTTERM) {
693                                         m_freem(sio.sb_mb);
694                                         return (EINTR);
695                                 }
696                             }
697                         } while (error == EWOULDBLOCK ||
698                                  (error == 0 && sio.sb_mb == NULL && control));
699                         if ((rcvflg & MSG_EOR) == 0)
700                                 kprintf("Egad!!\n");
701                         if (error == 0 && sio.sb_mb == NULL)
702                                 error = EPIPE;
703                         len = sio.sb_cc;
704                         *mp = sio.sb_mb;
705                 }
706 errout:
707                 if (error && error != EINTR && error != ERESTART) {
708                         m_freem(*mp);
709                         *mp = NULL;
710                         if (error != EPIPE) {
711                                 log(LOG_INFO,
712                                     "receive error %d from nfs server %s\n",
713                                     error,
714                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
715                         }
716                         error = nfs_sndlock(nmp, rep);
717                         if (!error) {
718                                 error = nfs_reconnect(nmp, rep);
719                                 if (!error)
720                                         goto tryagain;
721                                 else
722                                         nfs_sndunlock(nmp);
723                         }
724                 }
725         } else {
726                 if ((so = nmp->nm_so) == NULL)
727                         return (EACCES);
728                 if (so->so_state & SS_ISCONNECTED)
729                         getnam = NULL;
730                 else
731                         getnam = aname;
732                 sbinit(&sio, 100000000);
733                 do {
734                         rcvflg = 0;
735                         error =  so_pru_soreceive(so, getnam, NULL, &sio,
736                                                   NULL, &rcvflg);
737                         if (error == EWOULDBLOCK && rep &&
738                             (rep->r_flags & R_SOFTTERM)) {
739                                 m_freem(sio.sb_mb);
740                                 return (EINTR);
741                         }
742                 } while (error == EWOULDBLOCK);
743
744                 len = sio.sb_cc;
745                 *mp = sio.sb_mb;
746
747                 /*
748                  * A shutdown may result in no error and no mbuf.
749                  * Convert to EPIPE.
750                  */
751                 if (*mp == NULL && error == 0)
752                         error = EPIPE;
753         }
754         if (error) {
755                 m_freem(*mp);
756                 *mp = NULL;
757         }
758
759         /*
760          * Search for any mbufs that are not a multiple of 4 bytes long
761          * or with m_data not longword aligned.
762          * These could cause pointer alignment problems, so copy them to
763          * well aligned mbufs.
764          */
765         nfs_realign(mp, 5 * NFSX_UNSIGNED);
766         return (error);
767 }
768
769 /*
770  * Implement receipt of reply on a socket.
771  *
772  * We must search through the list of received datagrams matching them
773  * with outstanding requests using the xid, until ours is found.
774  *
775  * If myrep is NULL we process packets on the socket until
776  * interrupted or until nm_reqrxq is non-empty.
777  */
778 /* ARGSUSED */
779 int
780 nfs_reply(struct nfsmount *nmp, struct nfsreq *myrep)
781 {
782         struct nfsreq *rep;
783         struct sockaddr *nam;
784         u_int32_t rxid;
785         u_int32_t *tl;
786         int error;
787         struct nfsm_info info;
788
789         /*
790          * Loop around until we get our own reply
791          */
792         for (;;) {
793                 /*
794                  * Lock against other receivers so that I don't get stuck in
795                  * sbwait() after someone else has received my reply for me.
796                  * Also necessary for connection based protocols to avoid
797                  * race conditions during a reconnect.
798                  *
799                  * If nfs_rcvlock() returns EALREADY, that means that
800                  * the reply has already been recieved by another
801                  * process and we can return immediately.  In this
802                  * case, the lock is not taken to avoid races with
803                  * other processes.
804                  */
805                 info.mrep = NULL;
806
807                 error = nfs_rcvlock(nmp, myrep);
808                 if (error == EALREADY)
809                         return (0);
810                 if (error)
811                         return (error);
812
813                 /*
814                  * If myrep is NULL we are the receiver helper thread.
815                  * Stop waiting for incoming replies if there are
816                  * messages sitting on reqrxq that we need to process,
817                  * or if a shutdown request is pending.
818                  */
819                 if (myrep == NULL && (TAILQ_FIRST(&nmp->nm_reqrxq) ||
820                     nmp->nm_rxstate > NFSSVC_PENDING)) {
821                         nfs_rcvunlock(nmp);
822                         return(EWOULDBLOCK);
823                 }
824
825                 /*
826                  * Get the next Rpc reply off the socket
827                  *
828                  * We cannot release the receive lock until we've
829                  * filled in rep->r_mrep, otherwise a waiting
830                  * thread may deadlock in soreceive with no incoming
831                  * packets expected.
832                  */
833                 error = nfs_receive(nmp, myrep, &nam, &info.mrep);
834                 if (error) {
835                         /*
836                          * Ignore routing errors on connectionless protocols??
837                          */
838                         nfs_rcvunlock(nmp);
839                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
840                                 if (nmp->nm_so == NULL)
841                                         return (error);
842                                 nmp->nm_so->so_error = 0;
843                                 continue;
844                         }
845                         return (error);
846                 }
847                 if (nam)
848                         FREE(nam, M_SONAME);
849
850                 /*
851                  * Get the xid and check that it is an rpc reply
852                  */
853                 info.md = info.mrep;
854                 info.dpos = mtod(info.md, caddr_t);
855                 NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
856                 rxid = *tl++;
857                 if (*tl != rpc_reply) {
858                         nfsstats.rpcinvalid++;
859                         m_freem(info.mrep);
860                         info.mrep = NULL;
861 nfsmout:
862                         nfs_rcvunlock(nmp);
863                         continue;
864                 }
865
866                 /*
867                  * Loop through the request list to match up the reply
868                  * Iff no match, just drop the datagram.  On match, set
869                  * r_mrep atomically to prevent the timer from messing
870                  * around with the request after we have exited the critical
871                  * section.
872                  */
873                 crit_enter();
874                 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
875                         if (rep->r_mrep == NULL && rxid == rep->r_xid)
876                                 break;
877                 }
878
879                 /*
880                  * Fill in the rest of the reply if we found a match.
881                  *
882                  * Deal with duplicate responses if there was no match.
883                  */
884                 if (rep) {
885                         rep->r_md = info.md;
886                         rep->r_dpos = info.dpos;
887                         if (nfsrtton) {
888                                 struct rttl *rt;
889
890                                 rt = &nfsrtt.rttl[nfsrtt.pos];
891                                 rt->proc = rep->r_procnum;
892                                 rt->rto = 0;
893                                 rt->sent = 0;
894                                 rt->cwnd = nmp->nm_maxasync_scaled;
895                                 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
896                                 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
897                                 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
898                                 getmicrotime(&rt->tstamp);
899                                 if (rep->r_flags & R_TIMING)
900                                         rt->rtt = rep->r_rtt;
901                                 else
902                                         rt->rtt = 1000000;
903                                 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
904                         }
905
906                         /*
907                          * New congestion control is based only on async
908                          * requests.
909                          */
910                         if (nmp->nm_maxasync_scaled < NFS_MAXASYNC_SCALED)
911                                 ++nmp->nm_maxasync_scaled;
912                         if (rep->r_flags & R_SENT) {
913                                 rep->r_flags &= ~R_SENT;
914                         }
915                         /*
916                          * Update rtt using a gain of 0.125 on the mean
917                          * and a gain of 0.25 on the deviation.
918                          *
919                          * NOTE SRTT/SDRTT are only good if R_TIMING is set.
920                          */
921                         if ((rep->r_flags & R_TIMING) && rep->r_rexmit == 0) {
922                                 /*
923                                  * Since the timer resolution of
924                                  * NFS_HZ is so course, it can often
925                                  * result in r_rtt == 0. Since
926                                  * r_rtt == N means that the actual
927                                  * rtt is between N+dt and N+2-dt ticks,
928                                  * add 1.
929                                  */
930                                 int n;
931                                 int d;
932
933 #define NFSRSB  NFS_RTT_SCALE_BITS
934                                 n = ((NFS_SRTT(rep) * 7) +
935                                      (rep->r_rtt << NFSRSB)) >> 3;
936                                 d = n - NFS_SRTT(rep);
937                                 NFS_SRTT(rep) = n;
938
939                                 /*
940                                  * Don't let the jitter calculation decay
941                                  * too quickly, but we want a fast rampup.
942                                  */
943                                 if (d < 0)
944                                         d = -d;
945                                 d <<= NFSRSB;
946                                 if (d < NFS_SDRTT(rep))
947                                         n = ((NFS_SDRTT(rep) * 15) + d) >> 4;
948                                 else
949                                         n = ((NFS_SDRTT(rep) * 3) + d) >> 2;
950                                 NFS_SDRTT(rep) = n;
951 #undef NFSRSB
952                         }
953                         nmp->nm_timeouts = 0;
954                         rep->r_mrep = info.mrep;
955                         nfs_hardterm(rep, 0);
956                 } else {
957                         /*
958                          * Extract vers, prog, nfsver, procnum.  A duplicate
959                          * response means we didn't wait long enough so
960                          * we increase the SRTT to avoid future spurious
961                          * timeouts.
962                          */
963                         u_int procnum = nmp->nm_lastreprocnum;
964                         int n;
965
966                         if (procnum < NFS_NPROCS && proct[procnum]) {
967                                 if (nfs_showrexmit)
968                                         kprintf("D");
969                                 n = nmp->nm_srtt[proct[procnum]];
970                                 n += NFS_ASYSCALE * NFS_HZ;
971                                 if (n < NFS_ASYSCALE * NFS_HZ * 10)
972                                         n = NFS_ASYSCALE * NFS_HZ * 10;
973                                 nmp->nm_srtt[proct[procnum]] = n;
974                         }
975                 }
976                 nfs_rcvunlock(nmp);
977                 crit_exit();
978
979                 /*
980                  * If not matched to a request, drop it.
981                  * If it's mine, get out.
982                  */
983                 if (rep == NULL) {
984                         nfsstats.rpcunexpected++;
985                         m_freem(info.mrep);
986                         info.mrep = NULL;
987                 } else if (rep == myrep) {
988                         if (rep->r_mrep == NULL)
989                                 panic("nfsreply nil");
990                         return (0);
991                 }
992         }
993 }
994
995 /*
996  * Run the request state machine until the target state is reached
997  * or a fatal error occurs.  The target state is not run.  Specifying
998  * a target of NFSM_STATE_DONE runs the state machine until the rpc
999  * is complete.
1000  *
1001  * EINPROGRESS is returned for all states other then the DONE state,
1002  * indicating that the rpc is still in progress.
1003  */
1004 int
1005 nfs_request(struct nfsm_info *info, nfsm_state_t bstate, nfsm_state_t estate)
1006 {
1007         struct nfsreq *req;
1008
1009         while (info->state >= bstate && info->state < estate) {
1010                 switch(info->state) {
1011                 case NFSM_STATE_SETUP:
1012                         /*
1013                          * Setup the nfsreq.  Any error which occurs during
1014                          * this state is fatal.
1015                          */
1016                         info->error = nfs_request_setup(info);
1017                         if (info->error) {
1018                                 info->state = NFSM_STATE_DONE;
1019                                 return (info->error);
1020                         } else {
1021                                 req = info->req;
1022                                 req->r_mrp = &info->mrep;
1023                                 req->r_mdp = &info->md;
1024                                 req->r_dposp = &info->dpos;
1025                                 info->state = NFSM_STATE_AUTH;
1026                         }
1027                         break;
1028                 case NFSM_STATE_AUTH:
1029                         /*
1030                          * Authenticate the nfsreq.  Any error which occurs
1031                          * during this state is fatal.
1032                          */
1033                         info->error = nfs_request_auth(info->req);
1034                         if (info->error) {
1035                                 info->state = NFSM_STATE_DONE;
1036                                 return (info->error);
1037                         } else {
1038                                 info->state = NFSM_STATE_TRY;
1039                         }
1040                         break;
1041                 case NFSM_STATE_TRY:
1042                         /*
1043                          * Transmit or retransmit attempt.  An error in this
1044                          * state is ignored and we always move on to the
1045                          * next state.
1046                          *
1047                          * This can trivially race the receiver if the
1048                          * request is asynchronous.  nfs_request_try()
1049                          * will thus set the state for us and we
1050                          * must also return immediately if we are
1051                          * running an async state machine, because
1052                          * info can become invalid due to races after
1053                          * try() returns.
1054                          */
1055                         if (info->req->r_flags & R_ASYNC) {
1056                                 nfs_request_try(info->req);
1057                                 if (estate == NFSM_STATE_WAITREPLY)
1058                                         return (EINPROGRESS);
1059                         } else {
1060                                 nfs_request_try(info->req);
1061                                 info->state = NFSM_STATE_WAITREPLY;
1062                         }
1063                         break;
1064                 case NFSM_STATE_WAITREPLY:
1065                         /*
1066                          * Wait for a reply or timeout and move on to the
1067                          * next state.  The error returned by this state
1068                          * is passed to the processing code in the next
1069                          * state.
1070                          */
1071                         info->error = nfs_request_waitreply(info->req);
1072                         info->state = NFSM_STATE_PROCESSREPLY;
1073                         break;
1074                 case NFSM_STATE_PROCESSREPLY:
1075                         /*
1076                          * Process the reply or timeout.  Errors which occur
1077                          * in this state may cause the state machine to
1078                          * go back to an earlier state, and are fatal
1079                          * otherwise.
1080                          */
1081                         info->error = nfs_request_processreply(info,
1082                                                                info->error);
1083                         switch(info->error) {
1084                         case ENEEDAUTH:
1085                                 info->state = NFSM_STATE_AUTH;
1086                                 break;
1087                         case EAGAIN:
1088                                 info->state = NFSM_STATE_TRY;
1089                                 break;
1090                         default:
1091                                 /*
1092                                  * Operation complete, with or without an
1093                                  * error.  We are done.
1094                                  */
1095                                 info->req = NULL;
1096                                 info->state = NFSM_STATE_DONE;
1097                                 return (info->error);
1098                         }
1099                         break;
1100                 case NFSM_STATE_DONE:
1101                         /*
1102                          * Shouldn't be reached
1103                          */
1104                         return (info->error);
1105                         /* NOT REACHED */
1106                 }
1107         }
1108
1109         /*
1110          * If we are done return the error code (if any).
1111          * Otherwise return EINPROGRESS.
1112          */
1113         if (info->state == NFSM_STATE_DONE)
1114                 return (info->error);
1115         return (EINPROGRESS);
1116 }
1117
1118 /*
1119  * nfs_request - goes something like this
1120  *      - fill in request struct
1121  *      - links it into list
1122  *      - calls nfs_send() for first transmit
1123  *      - calls nfs_receive() to get reply
1124  *      - break down rpc header and return with nfs reply pointed to
1125  *        by mrep or error
1126  * nb: always frees up mreq mbuf list
1127  */
1128 static int
1129 nfs_request_setup(nfsm_info_t info)
1130 {
1131         struct nfsreq *req;
1132         struct nfsmount *nmp;
1133         struct mbuf *m;
1134         int i;
1135
1136         /*
1137          * Reject requests while attempting a forced unmount.
1138          */
1139         if (info->vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
1140                 m_freem(info->mreq);
1141                 info->mreq = NULL;
1142                 return (ESTALE);
1143         }
1144         nmp = VFSTONFS(info->vp->v_mount);
1145         req = kmalloc(sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
1146         req->r_nmp = nmp;
1147         req->r_vp = info->vp;
1148         req->r_td = info->td;
1149         req->r_procnum = info->procnum;
1150         req->r_mreq = NULL;
1151         req->r_cred = info->cred;
1152
1153         i = 0;
1154         m = info->mreq;
1155         while (m) {
1156                 i += m->m_len;
1157                 m = m->m_next;
1158         }
1159         req->r_mrest = info->mreq;
1160         req->r_mrest_len = i;
1161
1162         /*
1163          * The presence of a non-NULL r_info in req indicates
1164          * async completion via our helper threads.  See the receiver
1165          * code.
1166          */
1167         if (info->bio) {
1168                 req->r_info = info;
1169                 req->r_flags = R_ASYNC;
1170         } else {
1171                 req->r_info = NULL;
1172                 req->r_flags = 0;
1173         }
1174         info->req = req;
1175         return(0);
1176 }
1177
1178 static int
1179 nfs_request_auth(struct nfsreq *rep)
1180 {
1181         struct nfsmount *nmp = rep->r_nmp;
1182         struct mbuf *m;
1183         char nickv[RPCX_NICKVERF];
1184         int error = 0, auth_len, auth_type;
1185         int verf_len;
1186         u_int32_t xid;
1187         char *auth_str, *verf_str;
1188         struct ucred *cred;
1189
1190         cred = rep->r_cred;
1191         rep->r_failed_auth = 0;
1192
1193         /*
1194          * Get the RPC header with authorization.
1195          */
1196         verf_str = auth_str = NULL;
1197         if (nmp->nm_flag & NFSMNT_KERB) {
1198                 verf_str = nickv;
1199                 verf_len = sizeof (nickv);
1200                 auth_type = RPCAUTH_KERB4;
1201                 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1202                 if (rep->r_failed_auth ||
1203                     nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1204                                     verf_str, verf_len)) {
1205                         error = nfs_getauth(nmp, rep, cred, &auth_str,
1206                                 &auth_len, verf_str, &verf_len, rep->r_key);
1207                         if (error) {
1208                                 m_freem(rep->r_mrest);
1209                                 rep->r_mrest = NULL;
1210                                 kfree((caddr_t)rep, M_NFSREQ);
1211                                 return (error);
1212                         }
1213                 }
1214         } else {
1215                 auth_type = RPCAUTH_UNIX;
1216                 if (cred->cr_ngroups < 1)
1217                         panic("nfsreq nogrps");
1218                 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1219                         nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1220                         5 * NFSX_UNSIGNED;
1221         }
1222         if (rep->r_mrest)
1223                 nfs_checkpkt(rep->r_mrest, rep->r_mrest_len);
1224         m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1225                         auth_len, auth_str, verf_len, verf_str,
1226                         rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1227         rep->r_mrest = NULL;
1228         if (auth_str)
1229                 kfree(auth_str, M_TEMP);
1230
1231         /*
1232          * For stream protocols, insert a Sun RPC Record Mark.
1233          */
1234         if (nmp->nm_sotype == SOCK_STREAM) {
1235                 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1236                 if (m == NULL) {
1237                         kfree(rep, M_NFSREQ);
1238                         return (ENOBUFS);
1239                 }
1240                 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1241                          (m->m_pkthdr.len - NFSX_UNSIGNED));
1242         }
1243
1244         nfs_checkpkt(m, m->m_pkthdr.len);
1245
1246         rep->r_mreq = m;
1247         rep->r_xid = xid;
1248         return (0);
1249 }
1250
1251 static int
1252 nfs_request_try(struct nfsreq *rep)
1253 {
1254         struct nfsmount *nmp = rep->r_nmp;
1255         struct mbuf *m2;
1256         int error;
1257
1258         /*
1259          * Request is not on any queue, only the owner has access to it
1260          * so it should not be locked by anyone atm.
1261          *
1262          * Interlock to prevent races.  While locked the only remote
1263          * action possible is for r_mrep to be set (once we enqueue it).
1264          */
1265         if (rep->r_flags == 0xdeadc0de) {
1266                 print_backtrace(-1);
1267                 panic("flags nbad\n");
1268         }
1269         KKASSERT((rep->r_flags & (R_LOCKED | R_ONREQQ)) == 0);
1270         if (nmp->nm_flag & NFSMNT_SOFT)
1271                 rep->r_retry = nmp->nm_retry;
1272         else
1273                 rep->r_retry = NFS_MAXREXMIT + 1;       /* past clip limit */
1274         rep->r_rtt = rep->r_rexmit = 0;
1275         if (proct[rep->r_procnum] > 0)
1276                 rep->r_flags |= R_TIMING | R_LOCKED;
1277         else
1278                 rep->r_flags |= R_LOCKED;
1279         rep->r_mrep = NULL;
1280
1281         nfsstats.rpcrequests++;
1282
1283         if (nmp->nm_flag & NFSMNT_FORCE) {
1284                 rep->r_flags |= R_SOFTTERM;
1285                 rep->r_flags &= ~R_LOCKED;
1286                 return (0);
1287         }
1288         rep->r_flags |= R_NEEDSXMIT;    /* in case send lock races us */
1289
1290         /*
1291          * Do the client side RPC.
1292          *
1293          * Chain request into list of outstanding requests. Be sure
1294          * to put it LAST so timer finds oldest requests first.  Note
1295          * that our control of R_LOCKED prevents the request from
1296          * getting ripped out from under us or transmitted by the
1297          * timer code.
1298          *
1299          * For requests with info structures we must atomically set the
1300          * info's state because the structure could become invalid upon
1301          * return due to races (i.e., if async)
1302          */
1303         crit_enter();
1304         mtx_link_init(&rep->r_link);
1305         KKASSERT((rep->r_flags & R_ONREQQ) == 0);
1306         TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1307         rep->r_flags |= R_ONREQQ;
1308         ++nmp->nm_reqqlen;
1309         if (rep->r_flags & R_ASYNC)
1310                 rep->r_info->state = NFSM_STATE_WAITREPLY;
1311         crit_exit();
1312
1313         error = 0;
1314
1315         /*
1316          * Send if we can.  Congestion control is not handled here any more
1317          * becausing trying to defer the initial send based on the nfs_timer
1318          * requires having a very fast nfs_timer, which is silly.
1319          */
1320         if (nmp->nm_so) {
1321                 if (nmp->nm_soflags & PR_CONNREQUIRED)
1322                         error = nfs_sndlock(nmp, rep);
1323                 if (error == 0 && (rep->r_flags & R_NEEDSXMIT)) {
1324                         m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1325                         error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1326                         rep->r_flags &= ~R_NEEDSXMIT;
1327                         if ((rep->r_flags & R_SENT) == 0) {
1328                                 rep->r_flags |= R_SENT;
1329                         }
1330                         if (nmp->nm_soflags & PR_CONNREQUIRED)
1331                                 nfs_sndunlock(nmp);
1332                 }
1333         } else {
1334                 rep->r_rtt = -1;
1335         }
1336         if (error == EPIPE)
1337                 error = 0;
1338
1339         /*
1340          * Release the lock.  The only remote action that may have occurred
1341          * would have been the setting of rep->r_mrep.  If this occured
1342          * and the request was async we have to move it to the reader
1343          * thread's queue for action.
1344          *
1345          * For async requests also make sure the reader is woken up so
1346          * it gets on the socket to read responses.
1347          */
1348         crit_enter();
1349         if (rep->r_flags & R_ASYNC) {
1350                 if (rep->r_mrep)
1351                         nfs_hardterm(rep, 1);
1352                 rep->r_flags &= ~R_LOCKED;
1353                 nfssvc_iod_reader_wakeup(nmp);
1354         } else {
1355                 rep->r_flags &= ~R_LOCKED;
1356         }
1357         if (rep->r_flags & R_WANTED) {
1358                 rep->r_flags &= ~R_WANTED;
1359                 wakeup(rep);
1360         }
1361         crit_exit();
1362         return (error);
1363 }
1364
1365 /*
1366  * This code is only called for synchronous requests.  Completed synchronous
1367  * requests are left on reqq and we remove them before moving on to the
1368  * processing state.
1369  */
1370 static int
1371 nfs_request_waitreply(struct nfsreq *rep)
1372 {
1373         struct nfsmount *nmp = rep->r_nmp;
1374         int error;
1375
1376         KKASSERT((rep->r_flags & R_ASYNC) == 0);
1377
1378         /*
1379          * Wait until the request is finished.
1380          */
1381         error = nfs_reply(nmp, rep);
1382
1383         /*
1384          * RPC done, unlink the request, but don't rip it out from under
1385          * the callout timer.
1386          *
1387          * Once unlinked no other receiver or the timer will have
1388          * visibility, so we do not have to set R_LOCKED.
1389          */
1390         crit_enter();
1391         while (rep->r_flags & R_LOCKED) {
1392                 rep->r_flags |= R_WANTED;
1393                 tsleep(rep, 0, "nfstrac", 0);
1394         }
1395         KKASSERT(rep->r_flags & R_ONREQQ);
1396         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1397         rep->r_flags &= ~R_ONREQQ;
1398         --nmp->nm_reqqlen;
1399         if (TAILQ_FIRST(&nmp->nm_bioq) &&
1400             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1401                 nfssvc_iod_writer_wakeup(nmp);
1402         }
1403         crit_exit();
1404
1405         /*
1406          * Decrement the outstanding request count.
1407          */
1408         if (rep->r_flags & R_SENT) {
1409                 rep->r_flags &= ~R_SENT;
1410         }
1411         return (error);
1412 }
1413
1414 /*
1415  * Process reply with error returned from nfs_requet_waitreply().
1416  *
1417  * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1418  * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1419  */
1420 static int
1421 nfs_request_processreply(nfsm_info_t info, int error)
1422 {
1423         struct nfsreq *req = info->req;
1424         struct nfsmount *nmp = req->r_nmp;
1425         u_int32_t *tl;
1426         int verf_type;
1427         int i;
1428
1429         /*
1430          * If there was a successful reply and a tprintf msg.
1431          * tprintf a response.
1432          */
1433         if (error == 0 && (req->r_flags & R_TPRINTFMSG)) {
1434                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1435                     "is alive again");
1436         }
1437         info->mrep = req->r_mrep;
1438         info->md = req->r_md;
1439         info->dpos = req->r_dpos;
1440         if (error) {
1441                 m_freem(req->r_mreq);
1442                 req->r_mreq = NULL;
1443                 kfree(req, M_NFSREQ);
1444                 info->req = NULL;
1445                 return (error);
1446         }
1447
1448         /*
1449          * break down the rpc header and check if ok
1450          */
1451         NULLOUT(tl = nfsm_dissect(info, 3 * NFSX_UNSIGNED));
1452         if (*tl++ == rpc_msgdenied) {
1453                 if (*tl == rpc_mismatch) {
1454                         error = EOPNOTSUPP;
1455                 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1456                            *tl++ == rpc_autherr) {
1457                         if (req->r_failed_auth == 0) {
1458                                 req->r_failed_auth++;
1459                                 req->r_mheadend->m_next = NULL;
1460                                 m_freem(info->mrep);
1461                                 info->mrep = NULL;
1462                                 m_freem(req->r_mreq);
1463                                 req->r_mreq = NULL;
1464                                 return (ENEEDAUTH);
1465                         } else {
1466                                 error = EAUTH;
1467                         }
1468                 } else {
1469                         error = EACCES;
1470                 }
1471                 m_freem(info->mrep);
1472                 info->mrep = NULL;
1473                 m_freem(req->r_mreq);
1474                 req->r_mreq = NULL;
1475                 kfree(req, M_NFSREQ);
1476                 info->req = NULL;
1477                 return (error);
1478         }
1479
1480         /*
1481          * Grab any Kerberos verifier, otherwise just throw it away.
1482          */
1483         verf_type = fxdr_unsigned(int, *tl++);
1484         i = fxdr_unsigned(int32_t, *tl);
1485         if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1486                 error = nfs_savenickauth(nmp, req->r_cred, i, req->r_key,
1487                                          &info->md, &info->dpos, info->mrep);
1488                 if (error)
1489                         goto nfsmout;
1490         } else if (i > 0) {
1491                 ERROROUT(nfsm_adv(info, nfsm_rndup(i)));
1492         }
1493         NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1494         /* 0 == ok */
1495         if (*tl == 0) {
1496                 NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1497                 if (*tl != 0) {
1498                         error = fxdr_unsigned(int, *tl);
1499
1500                         /*
1501                          * Does anyone even implement this?  Just impose
1502                          * a 1-second delay.
1503                          */
1504                         if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1505                                 error == NFSERR_TRYLATER) {
1506                                 m_freem(info->mrep);
1507                                 info->mrep = NULL;
1508                                 error = 0;
1509
1510                                 tsleep((caddr_t)&lbolt, 0, "nqnfstry", 0);
1511                                 return (EAGAIN);        /* goto tryagain */
1512                         }
1513
1514                         /*
1515                          * If the File Handle was stale, invalidate the
1516                          * lookup cache, just in case.
1517                          *
1518                          * To avoid namecache<->vnode deadlocks we must
1519                          * release the vnode lock if we hold it.
1520                          */
1521                         if (error == ESTALE) {
1522                                 struct vnode *vp = req->r_vp;
1523                                 int ltype;
1524
1525                                 ltype = lockstatus(&vp->v_lock, curthread);
1526                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1527                                         lockmgr(&vp->v_lock, LK_RELEASE);
1528                                 cache_inval_vp(vp, CINV_CHILDREN);
1529                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1530                                         lockmgr(&vp->v_lock, ltype);
1531                         }
1532                         if (nmp->nm_flag & NFSMNT_NFSV3) {
1533                                 KKASSERT(*req->r_mrp == info->mrep);
1534                                 KKASSERT(*req->r_mdp == info->md);
1535                                 KKASSERT(*req->r_dposp == info->dpos);
1536                                 error |= NFSERR_RETERR;
1537                         } else {
1538                                 m_freem(info->mrep);
1539                                 info->mrep = NULL;
1540                         }
1541                         m_freem(req->r_mreq);
1542                         req->r_mreq = NULL;
1543                         kfree(req, M_NFSREQ);
1544                         info->req = NULL;
1545                         return (error);
1546                 }
1547
1548                 KKASSERT(*req->r_mrp == info->mrep);
1549                 KKASSERT(*req->r_mdp == info->md);
1550                 KKASSERT(*req->r_dposp == info->dpos);
1551                 m_freem(req->r_mreq);
1552                 req->r_mreq = NULL;
1553                 FREE(req, M_NFSREQ);
1554                 return (0);
1555         }
1556         m_freem(info->mrep);
1557         info->mrep = NULL;
1558         error = EPROTONOSUPPORT;
1559 nfsmout:
1560         m_freem(req->r_mreq);
1561         req->r_mreq = NULL;
1562         kfree(req, M_NFSREQ);
1563         info->req = NULL;
1564         return (error);
1565 }
1566
1567 #ifndef NFS_NOSERVER
1568 /*
1569  * Generate the rpc reply header
1570  * siz arg. is used to decide if adding a cluster is worthwhile
1571  */
1572 int
1573 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1574             int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1575 {
1576         u_int32_t *tl;
1577         struct nfsm_info info;
1578
1579         siz += RPC_REPLYSIZ;
1580         info.mb = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1581         info.mreq = info.mb;
1582         info.mreq->m_pkthdr.len = 0;
1583         /*
1584          * If this is not a cluster, try and leave leading space
1585          * for the lower level headers.
1586          */
1587         if ((max_hdr + siz) < MINCLSIZE)
1588                 info.mreq->m_data += max_hdr;
1589         tl = mtod(info.mreq, u_int32_t *);
1590         info.mreq->m_len = 6 * NFSX_UNSIGNED;
1591         info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1592         *tl++ = txdr_unsigned(nd->nd_retxid);
1593         *tl++ = rpc_reply;
1594         if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1595                 *tl++ = rpc_msgdenied;
1596                 if (err & NFSERR_AUTHERR) {
1597                         *tl++ = rpc_autherr;
1598                         *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1599                         info.mreq->m_len -= NFSX_UNSIGNED;
1600                         info.bpos -= NFSX_UNSIGNED;
1601                 } else {
1602                         *tl++ = rpc_mismatch;
1603                         *tl++ = txdr_unsigned(RPC_VER2);
1604                         *tl = txdr_unsigned(RPC_VER2);
1605                 }
1606         } else {
1607                 *tl++ = rpc_msgaccepted;
1608
1609                 /*
1610                  * For Kerberos authentication, we must send the nickname
1611                  * verifier back, otherwise just RPCAUTH_NULL.
1612                  */
1613                 if (nd->nd_flag & ND_KERBFULL) {
1614                     struct nfsuid *nuidp;
1615                     struct timeval ktvin, ktvout;
1616
1617                     for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1618                         nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1619                         if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1620                             (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1621                              &nuidp->nu_haddr, nd->nd_nam2)))
1622                             break;
1623                     }
1624                     if (nuidp) {
1625                         ktvin.tv_sec =
1626                             txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1627                         ktvin.tv_usec =
1628                             txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1629
1630                         /*
1631                          * Encrypt the timestamp in ecb mode using the
1632                          * session key.
1633                          */
1634 #ifdef NFSKERB
1635                         XXX
1636 #else
1637                         ktvout.tv_sec = 0;
1638                         ktvout.tv_usec = 0;
1639 #endif
1640
1641                         *tl++ = rpc_auth_kerb;
1642                         *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1643                         *tl = ktvout.tv_sec;
1644                         tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1645                         *tl++ = ktvout.tv_usec;
1646                         *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1647                     } else {
1648                         *tl++ = 0;
1649                         *tl++ = 0;
1650                     }
1651                 } else {
1652                         *tl++ = 0;
1653                         *tl++ = 0;
1654                 }
1655                 switch (err) {
1656                 case EPROGUNAVAIL:
1657                         *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1658                         break;
1659                 case EPROGMISMATCH:
1660                         *tl = txdr_unsigned(RPC_PROGMISMATCH);
1661                         tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1662                         *tl++ = txdr_unsigned(2);
1663                         *tl = txdr_unsigned(3);
1664                         break;
1665                 case EPROCUNAVAIL:
1666                         *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1667                         break;
1668                 case EBADRPC:
1669                         *tl = txdr_unsigned(RPC_GARBAGE);
1670                         break;
1671                 default:
1672                         *tl = 0;
1673                         if (err != NFSERR_RETVOID) {
1674                                 tl = nfsm_build(&info, NFSX_UNSIGNED);
1675                                 if (err)
1676                                     *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1677                                 else
1678                                     *tl = 0;
1679                         }
1680                         break;
1681                 };
1682         }
1683
1684         if (mrq != NULL)
1685             *mrq = info.mreq;
1686         *mbp = info.mb;
1687         *bposp = info.bpos;
1688         if (err != 0 && err != NFSERR_RETVOID)
1689                 nfsstats.srvrpc_errs++;
1690         return (0);
1691 }
1692
1693
1694 #endif /* NFS_NOSERVER */
1695
1696 /*
1697  * Nfs timer routine.
1698  *
1699  * Scan the nfsreq list and retranmit any requests that have timed out
1700  * To avoid retransmission attempts on STREAM sockets (in the future) make
1701  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1702  *
1703  * Requests with attached responses, terminated requests, and
1704  * locked requests are ignored.  Locked requests will be picked up
1705  * in a later timer call.
1706  */
1707 void
1708 nfs_timer_callout(void *arg /* never used */)
1709 {
1710         struct nfsmount *nmp;
1711         struct nfsreq *req;
1712 #ifndef NFS_NOSERVER
1713         struct nfssvc_sock *slp;
1714         u_quad_t cur_usec;
1715 #endif /* NFS_NOSERVER */
1716
1717         lwkt_gettoken(&nfs_token);
1718         TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1719                 lwkt_gettoken(&nmp->nm_token);
1720                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1721                         KKASSERT(nmp == req->r_nmp);
1722                         if (req->r_mrep)
1723                                 continue;
1724                         if (req->r_flags & (R_SOFTTERM | R_LOCKED))
1725                                 continue;
1726
1727                         /*
1728                          * Handle timeout/retry.  Be sure to process r_mrep
1729                          * for async requests that completed while we had
1730                          * the request locked or they will hang in the reqq
1731                          * forever.
1732                          */
1733                         req->r_flags |= R_LOCKED;
1734                         if (nfs_sigintr(nmp, req, req->r_td)) {
1735                                 nfs_softterm(req, 1);
1736                                 req->r_flags &= ~R_LOCKED;
1737                         } else {
1738                                 nfs_timer_req(req);
1739                                 if (req->r_flags & R_ASYNC) {
1740                                         if (req->r_mrep)
1741                                                 nfs_hardterm(req, 1);
1742                                         req->r_flags &= ~R_LOCKED;
1743                                         nfssvc_iod_reader_wakeup(nmp);
1744                                 } else {
1745                                         req->r_flags &= ~R_LOCKED;
1746                                 }
1747                         }
1748                         if (req->r_flags & R_WANTED) {
1749                                 req->r_flags &= ~R_WANTED;
1750                                 wakeup(req);
1751                         }
1752                 }
1753                 lwkt_reltoken(&nmp->nm_token);
1754         }
1755 #ifndef NFS_NOSERVER
1756
1757         /*
1758          * Scan the write gathering queues for writes that need to be
1759          * completed now.
1760          */
1761         cur_usec = nfs_curusec();
1762
1763         TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1764                 /* XXX race against removal */
1765                 if (lwkt_trytoken(&slp->ns_token)) {
1766                         if (slp->ns_tq.lh_first &&
1767                             (slp->ns_tq.lh_first->nd_time <= cur_usec)) {
1768                                 nfsrv_wakenfsd(slp, 1);
1769                         }
1770                         lwkt_reltoken(&slp->ns_token);
1771                 }
1772         }
1773 #endif /* NFS_NOSERVER */
1774
1775         callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer_callout, NULL);
1776         lwkt_reltoken(&nfs_token);
1777 }
1778
1779 static
1780 void
1781 nfs_timer_req(struct nfsreq *req)
1782 {
1783         struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1784         struct nfsmount *nmp = req->r_nmp;
1785         struct mbuf *m;
1786         struct socket *so;
1787         int timeo;
1788         int error;
1789
1790         /*
1791          * rtt ticks and timeout calculation.  Return if the timeout
1792          * has not been reached yet, unless the packet is flagged
1793          * for an immediate send.
1794          *
1795          * The mean rtt doesn't help when we get random I/Os, we have
1796          * to multiply by fairly large numbers.
1797          */
1798         if (req->r_rtt >= 0) {
1799                 /*
1800                  * Calculate the timeout to test against.
1801                  */
1802                 req->r_rtt++;
1803                 if (nmp->nm_flag & NFSMNT_DUMBTIMR) {
1804                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1805                 } else if (req->r_flags & R_TIMING) {
1806                         timeo = NFS_SRTT(req) + NFS_SDRTT(req);
1807                 } else {
1808                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1809                 }
1810                 timeo *= multt[req->r_procnum];
1811                 /* timeo is still scaled by SCALE_BITS */
1812
1813 #define NFSFS   (NFS_RTT_SCALE * NFS_HZ)
1814                 if (req->r_flags & R_TIMING) {
1815                         static long last_time;
1816                         if (nfs_showrtt && last_time != time_second) {
1817                                 kprintf("rpccmd %d NFS SRTT %d SDRTT %d "
1818                                         "timeo %d.%03d\n",
1819                                         proct[req->r_procnum],
1820                                         NFS_SRTT(req), NFS_SDRTT(req),
1821                                         timeo / NFSFS,
1822                                         timeo % NFSFS * 1000 /  NFSFS);
1823                                 last_time = time_second;
1824                         }
1825                 }
1826 #undef NFSFS
1827
1828                 /*
1829                  * deal with nfs_timer jitter.
1830                  */
1831                 timeo = (timeo >> NFS_RTT_SCALE_BITS) + 1;
1832                 if (timeo < 2)
1833                         timeo = 2;
1834
1835                 if (nmp->nm_timeouts > 0)
1836                         timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1837                 if (timeo > NFS_MAXTIMEO)
1838                         timeo = NFS_MAXTIMEO;
1839                 if (req->r_rtt <= timeo) {
1840                         if ((req->r_flags & R_NEEDSXMIT) == 0)
1841                                 return;
1842                 } else if (nmp->nm_timeouts < 8) {
1843                         nmp->nm_timeouts++;
1844                 }
1845         }
1846
1847         /*
1848          * Check for server not responding
1849          */
1850         if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1851              req->r_rexmit > nmp->nm_deadthresh) {
1852                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1853                         "not responding");
1854                 req->r_flags |= R_TPRINTFMSG;
1855         }
1856         if (req->r_rexmit >= req->r_retry) {    /* too many */
1857                 nfsstats.rpctimeouts++;
1858                 nfs_softterm(req, 1);
1859                 return;
1860         }
1861
1862         /*
1863          * Generally disable retransmission on reliable sockets,
1864          * unless the request is flagged for immediate send.
1865          */
1866         if (nmp->nm_sotype != SOCK_DGRAM) {
1867                 if (++req->r_rexmit > NFS_MAXREXMIT)
1868                         req->r_rexmit = NFS_MAXREXMIT;
1869                 if ((req->r_flags & R_NEEDSXMIT) == 0)
1870                         return;
1871         }
1872
1873         /*
1874          * Stop here if we do not have a socket!
1875          */
1876         if ((so = nmp->nm_so) == NULL)
1877                 return;
1878
1879         /*
1880          * If there is enough space and the window allows.. resend it.
1881          *
1882          * r_rtt is left intact in case we get an answer after the
1883          * retry that was a reply to the original packet.
1884          *
1885          * NOTE: so_pru_send()
1886          */
1887         if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1888             (req->r_flags & (R_SENT | R_NEEDSXMIT)) &&
1889            (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1890                 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1891                     error = so_pru_send(so, 0, m, NULL, NULL, td);
1892                 else
1893                     error = so_pru_send(so, 0, m, nmp->nm_nam, NULL, td);
1894                 if (error) {
1895                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1896                                 so->so_error = 0;
1897                         req->r_flags |= R_NEEDSXMIT;
1898                 } else if (req->r_mrep == NULL) {
1899                         /*
1900                          * Iff first send, start timing
1901                          * else turn timing off, backoff timer
1902                          * and divide congestion window by 2.
1903                          *
1904                          * It is possible for the so_pru_send() to
1905                          * block and for us to race a reply so we
1906                          * only do this if the reply field has not
1907                          * been filled in.  R_LOCKED will prevent
1908                          * the request from being ripped out from under
1909                          * us entirely.
1910                          *
1911                          * Record the last resent procnum to aid us
1912                          * in duplicate detection on receive.
1913                          */
1914                         if ((req->r_flags & R_NEEDSXMIT) == 0) {
1915                                 if (nfs_showrexmit)
1916                                         kprintf("X");
1917                                 if (++req->r_rexmit > NFS_MAXREXMIT)
1918                                         req->r_rexmit = NFS_MAXREXMIT;
1919                                 nmp->nm_maxasync_scaled >>= 1;
1920                                 if (nmp->nm_maxasync_scaled < NFS_MINASYNC_SCALED)
1921                                         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
1922                                 nfsstats.rpcretries++;
1923                                 nmp->nm_lastreprocnum = req->r_procnum;
1924                         } else {
1925                                 req->r_flags |= R_SENT;
1926                                 req->r_flags &= ~R_NEEDSXMIT;
1927                         }
1928                 }
1929         }
1930 }
1931
1932 /*
1933  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1934  * wait for all requests to complete. This is used by forced unmounts
1935  * to terminate any outstanding RPCs.
1936  *
1937  * Locked requests cannot be canceled but will be marked for
1938  * soft-termination.
1939  */
1940 int
1941 nfs_nmcancelreqs(struct nfsmount *nmp)
1942 {
1943         struct nfsreq *req;
1944         int i;
1945
1946         crit_enter();
1947         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1948                 if (req->r_mrep != NULL || (req->r_flags & R_SOFTTERM))
1949                         continue;
1950                 nfs_softterm(req, 0);
1951         }
1952         /* XXX  the other two queues as well */
1953         crit_exit();
1954
1955         for (i = 0; i < 30; i++) {
1956                 crit_enter();
1957                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1958                         if (nmp == req->r_nmp)
1959                                 break;
1960                 }
1961                 crit_exit();
1962                 if (req == NULL)
1963                         return (0);
1964                 tsleep(&lbolt, 0, "nfscancel", 0);
1965         }
1966         return (EBUSY);
1967 }
1968
1969 /*
1970  * Soft-terminate a request, effectively marking it as failed.
1971  *
1972  * Must be called from within a critical section.
1973  */
1974 static void
1975 nfs_softterm(struct nfsreq *rep, int islocked)
1976 {
1977         rep->r_flags |= R_SOFTTERM;
1978         nfs_hardterm(rep, islocked);
1979 }
1980
1981 /*
1982  * Hard-terminate a request, typically after getting a response.
1983  *
1984  * The state machine can still decide to re-issue it later if necessary.
1985  *
1986  * Must be called from within a critical section.
1987  */
1988 static void
1989 nfs_hardterm(struct nfsreq *rep, int islocked)
1990 {
1991         struct nfsmount *nmp = rep->r_nmp;
1992
1993         /*
1994          * The nm_send count is decremented now to avoid deadlocks
1995          * when the process in soreceive() hasn't yet managed to send
1996          * its own request.
1997          */
1998         if (rep->r_flags & R_SENT) {
1999                 rep->r_flags &= ~R_SENT;
2000         }
2001
2002         /*
2003          * If we locked the request or nobody else has locked the request,
2004          * and the request is async, we can move it to the reader thread's
2005          * queue now and fix up the state.
2006          *
2007          * If we locked the request or nobody else has locked the request,
2008          * we can wake up anyone blocked waiting for a response on the
2009          * request.
2010          */
2011         if (islocked || (rep->r_flags & R_LOCKED) == 0) {
2012                 if ((rep->r_flags & (R_ONREQQ | R_ASYNC)) ==
2013                     (R_ONREQQ | R_ASYNC)) {
2014                         rep->r_flags &= ~R_ONREQQ;
2015                         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
2016                         --nmp->nm_reqqlen;
2017                         TAILQ_INSERT_TAIL(&nmp->nm_reqrxq, rep, r_chain);
2018                         KKASSERT(rep->r_info->state == NFSM_STATE_TRY ||
2019                                  rep->r_info->state == NFSM_STATE_WAITREPLY);
2020                         rep->r_info->state = NFSM_STATE_PROCESSREPLY;
2021                         nfssvc_iod_reader_wakeup(nmp);
2022                         if (TAILQ_FIRST(&nmp->nm_bioq) &&
2023                             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
2024                                 nfssvc_iod_writer_wakeup(nmp);
2025                         }
2026                 }
2027                 mtx_abort_ex_link(&nmp->nm_rxlock, &rep->r_link);
2028         }
2029 }
2030
2031 /*
2032  * Test for a termination condition pending on the process.
2033  * This is used for NFSMNT_INT mounts.
2034  */
2035 int
2036 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
2037 {
2038         sigset_t tmpset;
2039         struct proc *p;
2040         struct lwp *lp;
2041
2042         if (rep && (rep->r_flags & R_SOFTTERM))
2043                 return (EINTR);
2044         /* Terminate all requests while attempting a forced unmount. */
2045         if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
2046                 return (EINTR);
2047         if (!(nmp->nm_flag & NFSMNT_INT))
2048                 return (0);
2049         /* td might be NULL YYY */
2050         if (td == NULL || (p = td->td_proc) == NULL)
2051                 return (0);
2052
2053         lp = td->td_lwp;
2054         tmpset = lwp_sigpend(lp);
2055         SIGSETNAND(tmpset, lp->lwp_sigmask);
2056         SIGSETNAND(tmpset, p->p_sigignore);
2057         if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
2058                 return (EINTR);
2059
2060         return (0);
2061 }
2062
2063 /*
2064  * Lock a socket against others.
2065  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
2066  * and also to avoid race conditions between the processes with nfs requests
2067  * in progress when a reconnect is necessary.
2068  */
2069 int
2070 nfs_sndlock(struct nfsmount *nmp, struct nfsreq *rep)
2071 {
2072         mtx_t mtx = &nmp->nm_txlock;
2073         struct thread *td;
2074         int slptimeo;
2075         int slpflag;
2076         int error;
2077
2078         slpflag = 0;
2079         slptimeo = 0;
2080         td = rep ? rep->r_td : NULL;
2081         if (nmp->nm_flag & NFSMNT_INT)
2082                 slpflag = PCATCH;
2083
2084         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2085                 if (nfs_sigintr(nmp, rep, td)) {
2086                         error = EINTR;
2087                         break;
2088                 }
2089                 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
2090                 if (error == 0)
2091                         break;
2092                 if (slpflag == PCATCH) {
2093                         slpflag = 0;
2094                         slptimeo = 2 * hz;
2095                 }
2096         }
2097         /* Always fail if our request has been cancelled. */
2098         if (rep && (rep->r_flags & R_SOFTTERM)) {
2099                 if (error == 0)
2100                         mtx_unlock(mtx);
2101                 error = EINTR;
2102         }
2103         return (error);
2104 }
2105
2106 /*
2107  * Unlock the stream socket for others.
2108  */
2109 void
2110 nfs_sndunlock(struct nfsmount *nmp)
2111 {
2112         mtx_unlock(&nmp->nm_txlock);
2113 }
2114
2115 /*
2116  * Lock the receiver side of the socket.
2117  *
2118  * rep may be NULL.
2119  */
2120 static int
2121 nfs_rcvlock(struct nfsmount *nmp, struct nfsreq *rep)
2122 {
2123         mtx_t mtx = &nmp->nm_rxlock;
2124         int slpflag;
2125         int slptimeo;
2126         int error;
2127
2128         /*
2129          * Unconditionally check for completion in case another nfsiod
2130          * get the packet while the caller was blocked, before the caller
2131          * called us.  Packet reception is handled by mainline code which
2132          * is protected by the BGL at the moment.
2133          *
2134          * We do not strictly need the second check just before the
2135          * tsleep(), but it's good defensive programming.
2136          */
2137         if (rep && rep->r_mrep != NULL)
2138                 return (EALREADY);
2139
2140         if (nmp->nm_flag & NFSMNT_INT)
2141                 slpflag = PCATCH;
2142         else
2143                 slpflag = 0;
2144         slptimeo = 0;
2145
2146         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2147                 if (nfs_sigintr(nmp, rep, (rep ? rep->r_td : NULL))) {
2148                         error = EINTR;
2149                         break;
2150                 }
2151                 if (rep && rep->r_mrep != NULL) {
2152                         error = EALREADY;
2153                         break;
2154                 }
2155
2156                 /*
2157                  * NOTE: can return ENOLCK, but in that case rep->r_mrep
2158                  *       will already be set.
2159                  */
2160                 if (rep) {
2161                         error = mtx_lock_ex_link(mtx, &rep->r_link,
2162                                                  "nfsrcvlk",
2163                                                  slpflag, slptimeo);
2164                 } else {
2165                         error = mtx_lock_ex(mtx, "nfsrcvlk", slpflag, slptimeo);
2166                 }
2167                 if (error == 0)
2168                         break;
2169
2170                 /*
2171                  * If our reply was recieved while we were sleeping,
2172                  * then just return without taking the lock to avoid a
2173                  * situation where a single iod could 'capture' the
2174                  * recieve lock.
2175                  */
2176                 if (rep && rep->r_mrep != NULL) {
2177                         error = EALREADY;
2178                         break;
2179                 }
2180                 if (slpflag == PCATCH) {
2181                         slpflag = 0;
2182                         slptimeo = 2 * hz;
2183                 }
2184         }
2185         if (error == 0) {
2186                 if (rep && rep->r_mrep != NULL) {
2187                         error = EALREADY;
2188                         mtx_unlock(mtx);
2189                 }
2190         }
2191         return (error);
2192 }
2193
2194 /*
2195  * Unlock the stream socket for others.
2196  */
2197 static void
2198 nfs_rcvunlock(struct nfsmount *nmp)
2199 {
2200         mtx_unlock(&nmp->nm_rxlock);
2201 }
2202
2203 /*
2204  * nfs_realign:
2205  *
2206  * Check for badly aligned mbuf data and realign by copying the unaligned
2207  * portion of the data into a new mbuf chain and freeing the portions
2208  * of the old chain that were replaced.
2209  *
2210  * We cannot simply realign the data within the existing mbuf chain
2211  * because the underlying buffers may contain other rpc commands and
2212  * we cannot afford to overwrite them.
2213  *
2214  * We would prefer to avoid this situation entirely.  The situation does
2215  * not occur with NFS/UDP and is supposed to only occassionally occur
2216  * with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
2217  *
2218  * NOTE!  MB_DONTWAIT cannot be used here.  The mbufs must be acquired
2219  *        because the rpc request OR reply cannot be thrown away.  TCP NFS
2220  *        mounts do not retry their RPCs unless the TCP connection itself
2221  *        is dropped so throwing away a RPC will basically cause the NFS
2222  *        operation to lockup indefinitely.
2223  */
2224 static void
2225 nfs_realign(struct mbuf **pm, int hsiz)
2226 {
2227         struct mbuf *m;
2228         struct mbuf *n = NULL;
2229
2230         /*
2231          * Check for misalignemnt
2232          */
2233         ++nfs_realign_test;
2234         while ((m = *pm) != NULL) {
2235                 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3))
2236                         break;
2237                 pm = &m->m_next;
2238         }
2239
2240         /*
2241          * If misalignment found make a completely new copy.
2242          */
2243         if (m) {
2244                 ++nfs_realign_count;
2245                 n = m_dup_data(m, MB_WAIT);
2246                 m_freem(*pm);
2247                 *pm = n;
2248         }
2249 }
2250
2251 #ifndef NFS_NOSERVER
2252
2253 /*
2254  * Parse an RPC request
2255  * - verify it
2256  * - fill in the cred struct.
2257  */
2258 int
2259 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
2260 {
2261         int len, i;
2262         u_int32_t *tl;
2263         struct uio uio;
2264         struct iovec iov;
2265         caddr_t cp;
2266         u_int32_t nfsvers, auth_type;
2267         uid_t nickuid;
2268         int error = 0, ticklen;
2269         struct nfsuid *nuidp;
2270         struct timeval tvin, tvout;
2271         struct nfsm_info info;
2272 #if 0                           /* until encrypted keys are implemented */
2273         NFSKERBKEYSCHED_T keys; /* stores key schedule */
2274 #endif
2275
2276         info.mrep = nd->nd_mrep;
2277         info.md = nd->nd_md;
2278         info.dpos = nd->nd_dpos;
2279
2280         if (has_header) {
2281                 NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
2282                 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
2283                 if (*tl++ != rpc_call) {
2284                         m_freem(info.mrep);
2285                         return (EBADRPC);
2286                 }
2287         } else {
2288                 NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
2289         }
2290         nd->nd_repstat = 0;
2291         nd->nd_flag = 0;
2292         if (*tl++ != rpc_vers) {
2293                 nd->nd_repstat = ERPCMISMATCH;
2294                 nd->nd_procnum = NFSPROC_NOOP;
2295                 return (0);
2296         }
2297         if (*tl != nfs_prog) {
2298                 nd->nd_repstat = EPROGUNAVAIL;
2299                 nd->nd_procnum = NFSPROC_NOOP;
2300                 return (0);
2301         }
2302         tl++;
2303         nfsvers = fxdr_unsigned(u_int32_t, *tl++);
2304         if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
2305                 nd->nd_repstat = EPROGMISMATCH;
2306                 nd->nd_procnum = NFSPROC_NOOP;
2307                 return (0);
2308         }
2309         if (nfsvers == NFS_VER3)
2310                 nd->nd_flag = ND_NFSV3;
2311         nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
2312         if (nd->nd_procnum == NFSPROC_NULL)
2313                 return (0);
2314         if (nd->nd_procnum >= NFS_NPROCS ||
2315                 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
2316                 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
2317                 nd->nd_repstat = EPROCUNAVAIL;
2318                 nd->nd_procnum = NFSPROC_NOOP;
2319                 return (0);
2320         }
2321         if ((nd->nd_flag & ND_NFSV3) == 0)
2322                 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
2323         auth_type = *tl++;
2324         len = fxdr_unsigned(int, *tl++);
2325         if (len < 0 || len > RPCAUTH_MAXSIZ) {
2326                 m_freem(info.mrep);
2327                 return (EBADRPC);
2328         }
2329
2330         nd->nd_flag &= ~ND_KERBAUTH;
2331         /*
2332          * Handle auth_unix or auth_kerb.
2333          */
2334         if (auth_type == rpc_auth_unix) {
2335                 len = fxdr_unsigned(int, *++tl);
2336                 if (len < 0 || len > NFS_MAXNAMLEN) {
2337                         m_freem(info.mrep);
2338                         return (EBADRPC);
2339                 }
2340                 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2341                 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2342                 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
2343                 nd->nd_cr.cr_ref = 1;
2344                 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
2345                 nd->nd_cr.cr_ruid = nd->nd_cr.cr_svuid = nd->nd_cr.cr_uid;
2346                 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
2347                 nd->nd_cr.cr_rgid = nd->nd_cr.cr_svgid = nd->nd_cr.cr_gid;
2348                 len = fxdr_unsigned(int, *tl);
2349                 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
2350                         m_freem(info.mrep);
2351                         return (EBADRPC);
2352                 }
2353                 NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
2354                 for (i = 1; i <= len; i++)
2355                     if (i < NGROUPS)
2356                         nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
2357                     else
2358                         tl++;
2359                 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
2360                 if (nd->nd_cr.cr_ngroups > 1)
2361                     nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
2362                 len = fxdr_unsigned(int, *++tl);
2363                 if (len < 0 || len > RPCAUTH_MAXSIZ) {
2364                         m_freem(info.mrep);
2365                         return (EBADRPC);
2366                 }
2367                 if (len > 0) {
2368                         ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2369                 }
2370         } else if (auth_type == rpc_auth_kerb) {
2371                 switch (fxdr_unsigned(int, *tl++)) {
2372                 case RPCAKN_FULLNAME:
2373                         ticklen = fxdr_unsigned(int, *tl);
2374                         *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2375                         uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2376                         nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2377                         if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2378                                 m_freem(info.mrep);
2379                                 return (EBADRPC);
2380                         }
2381                         uio.uio_offset = 0;
2382                         uio.uio_iov = &iov;
2383                         uio.uio_iovcnt = 1;
2384                         uio.uio_segflg = UIO_SYSSPACE;
2385                         iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2386                         iov.iov_len = RPCAUTH_MAXSIZ - 4;
2387                         ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2388                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2389                         if (*tl++ != rpc_auth_kerb ||
2390                                 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2391                                 kprintf("Bad kerb verifier\n");
2392                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2393                                 nd->nd_procnum = NFSPROC_NOOP;
2394                                 return (0);
2395                         }
2396                         NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2397                         tl = (u_int32_t *)cp;
2398                         if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2399                                 kprintf("Not fullname kerb verifier\n");
2400                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2401                                 nd->nd_procnum = NFSPROC_NOOP;
2402                                 return (0);
2403                         }
2404                         cp += NFSX_UNSIGNED;
2405                         bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2406                         nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2407                         nd->nd_flag |= ND_KERBFULL;
2408                         nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2409                         break;
2410                 case RPCAKN_NICKNAME:
2411                         if (len != 2 * NFSX_UNSIGNED) {
2412                                 kprintf("Kerb nickname short\n");
2413                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2414                                 nd->nd_procnum = NFSPROC_NOOP;
2415                                 return (0);
2416                         }
2417                         nickuid = fxdr_unsigned(uid_t, *tl);
2418                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2419                         if (*tl++ != rpc_auth_kerb ||
2420                                 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2421                                 kprintf("Kerb nick verifier bad\n");
2422                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2423                                 nd->nd_procnum = NFSPROC_NOOP;
2424                                 return (0);
2425                         }
2426                         NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2427                         tvin.tv_sec = *tl++;
2428                         tvin.tv_usec = *tl;
2429
2430                         for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2431                             nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2432                                 if (nuidp->nu_cr.cr_uid == nickuid &&
2433                                     (!nd->nd_nam2 ||
2434                                      netaddr_match(NU_NETFAM(nuidp),
2435                                       &nuidp->nu_haddr, nd->nd_nam2)))
2436                                         break;
2437                         }
2438                         if (!nuidp) {
2439                                 nd->nd_repstat =
2440                                         (NFSERR_AUTHERR|AUTH_REJECTCRED);
2441                                 nd->nd_procnum = NFSPROC_NOOP;
2442                                 return (0);
2443                         }
2444
2445                         /*
2446                          * Now, decrypt the timestamp using the session key
2447                          * and validate it.
2448                          */
2449 #ifdef NFSKERB
2450                         XXX
2451 #else
2452                         tvout.tv_sec = 0;
2453                         tvout.tv_usec = 0;
2454 #endif
2455
2456                         tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2457                         tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2458                         if (nuidp->nu_expire < time_second ||
2459                             nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2460                             (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2461                              nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2462                                 nuidp->nu_expire = 0;
2463                                 nd->nd_repstat =
2464                                     (NFSERR_AUTHERR|AUTH_REJECTVERF);
2465                                 nd->nd_procnum = NFSPROC_NOOP;
2466                                 return (0);
2467                         }
2468                         nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2469                         nd->nd_flag |= ND_KERBNICK;
2470                 };
2471         } else {
2472                 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2473                 nd->nd_procnum = NFSPROC_NOOP;
2474                 return (0);
2475         }
2476
2477         nd->nd_md = info.md;
2478         nd->nd_dpos = info.dpos;
2479         return (0);
2480 nfsmout:
2481         return (error);
2482 }
2483
2484 #endif
2485
2486 /*
2487  * Send a message to the originating process's terminal.  The thread and/or
2488  * process may be NULL.  YYY the thread should not be NULL but there may
2489  * still be some uio_td's that are still being passed as NULL through to
2490  * nfsm_request().
2491  */
2492 static int
2493 nfs_msg(struct thread *td, char *server, char *msg)
2494 {
2495         tpr_t tpr;
2496
2497         if (td && td->td_proc)
2498                 tpr = tprintf_open(td->td_proc);
2499         else
2500                 tpr = NULL;
2501         tprintf(tpr, "nfs server %s: %s\n", server, msg);
2502         tprintf_close(tpr);
2503         return (0);
2504 }
2505
2506 #ifndef NFS_NOSERVER
2507
2508 /*
2509  * Socket upcall routine for nfsd sockets.  This runs in the protocol
2510  * thread and passes waitflag == MB_DONTWAIT.
2511  */
2512 void
2513 nfsrv_rcv_upcall(struct socket *so, void *arg, int waitflag)
2514 {
2515         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2516
2517         if (slp->ns_needq_upcall == 0) {
2518                 slp->ns_needq_upcall = 1;       /* ok to race */
2519                 lwkt_gettoken(&nfs_token);
2520                 nfsrv_wakenfsd(slp, 1);
2521                 lwkt_reltoken(&nfs_token);
2522         }
2523 #if 0
2524         lwkt_gettoken(&slp->ns_token);
2525         slp->ns_flag |= SLP_NEEDQ;
2526         nfsrv_rcv(so, arg, waitflag);
2527         lwkt_reltoken(&slp->ns_token);
2528 #endif
2529 }
2530
2531 /*
2532  * Process new data on a receive socket.  Essentially do as much as we can
2533  * non-blocking, else punt and it will be called with MB_WAIT from an nfsd.
2534  *
2535  * slp->ns_token is held on call
2536  */
2537 void
2538 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2539 {
2540         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2541         struct mbuf *m;
2542         struct sockaddr *nam;
2543         struct sockbuf sio;
2544         int flags, error;
2545         int nparallel_wakeup = 0;
2546
2547         ASSERT_LWKT_TOKEN_HELD(&slp->ns_token);
2548
2549         if ((slp->ns_flag & SLP_VALID) == 0)
2550                 return;
2551
2552         /*
2553          * Do not allow an infinite number of completed RPC records to build 
2554          * up before we stop reading data from the socket.  Otherwise we could
2555          * end up holding onto an unreasonable number of mbufs for requests
2556          * waiting for service.
2557          *
2558          * This should give pretty good feedback to the TCP layer and
2559          * prevents a memory crunch for other protocols.
2560          *
2561          * Note that the same service socket can be dispatched to several
2562          * nfs servers simultaniously.  The tcp protocol callback calls us
2563          * with MB_DONTWAIT.  nfsd calls us with MB_WAIT (typically).
2564          */
2565         if (NFSRV_RECLIMIT(slp))
2566                 return;
2567
2568         /*
2569          * Handle protocol specifics to parse an RPC request.  We always
2570          * pull from the socket using non-blocking I/O.
2571          */
2572         if (so->so_type == SOCK_STREAM) {
2573                 /*
2574                  * The data has to be read in an orderly fashion from a TCP
2575                  * stream, unlike a UDP socket.  It is possible for soreceive
2576                  * and/or nfsrv_getstream() to block, so make sure only one
2577                  * entity is messing around with the TCP stream at any given
2578                  * moment.  The receive sockbuf's lock in soreceive is not
2579                  * sufficient.
2580                  */
2581                 if (slp->ns_flag & SLP_GETSTREAM)
2582                         return;
2583                 slp->ns_flag |= SLP_GETSTREAM;
2584
2585                 /*
2586                  * Do soreceive().  Pull out as much data as possible without
2587                  * blocking.
2588                  */
2589                 sbinit(&sio, 1000000000);
2590                 flags = MSG_DONTWAIT;
2591                 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2592                 if (error || sio.sb_mb == NULL) {
2593                         if (error != EWOULDBLOCK)
2594                                 slp->ns_flag |= SLP_DISCONN;
2595                         slp->ns_flag &= ~(SLP_GETSTREAM | SLP_NEEDQ);
2596                         goto done;
2597                 }
2598                 m = sio.sb_mb;
2599                 if (slp->ns_rawend) {
2600                         slp->ns_rawend->m_next = m;
2601                         slp->ns_cc += sio.sb_cc;
2602                 } else {
2603                         slp->ns_raw = m;
2604                         slp->ns_cc = sio.sb_cc;
2605                 }
2606                 while (m->m_next)
2607                         m = m->m_next;
2608                 slp->ns_rawend = m;
2609
2610                 /*
2611                  * Now try and parse as many record(s) as we can out of the
2612                  * raw stream data.  This will set SLP_DOREC.
2613                  */
2614                 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2615                 if (error && error != EWOULDBLOCK)
2616                         slp->ns_flag |= SLP_DISCONN;
2617                 slp->ns_flag &= ~SLP_GETSTREAM;
2618         } else {
2619                 /*
2620                  * For UDP soreceive typically pulls just one packet, loop
2621                  * to get the whole batch.
2622                  */
2623                 do {
2624                         sbinit(&sio, 1000000000);
2625                         flags = MSG_DONTWAIT;
2626                         error = so_pru_soreceive(so, &nam, NULL, &sio,
2627                                                  NULL, &flags);
2628                         if (sio.sb_mb) {
2629                                 struct nfsrv_rec *rec;
2630                                 int mf = (waitflag & MB_DONTWAIT) ?
2631                                             M_NOWAIT : M_WAITOK;
2632                                 rec = kmalloc(sizeof(struct nfsrv_rec),
2633                                              M_NFSRVDESC, mf);
2634                                 if (!rec) {
2635                                         if (nam)
2636                                                 FREE(nam, M_SONAME);
2637                                         m_freem(sio.sb_mb);
2638                                         continue;
2639                                 }
2640                                 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2641                                 rec->nr_address = nam;
2642                                 rec->nr_packet = sio.sb_mb;
2643                                 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2644                                 ++slp->ns_numrec;
2645                                 slp->ns_flag |= SLP_DOREC;
2646                                 ++nparallel_wakeup;
2647                         } else {
2648                                 slp->ns_flag &= ~SLP_NEEDQ;
2649                         }
2650                         if (error) {
2651                                 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2652                                     && error != EWOULDBLOCK) {
2653                                         slp->ns_flag |= SLP_DISCONN;
2654                                         break;
2655                                 }
2656                         }
2657                         if (NFSRV_RECLIMIT(slp))
2658                                 break;
2659                 } while (sio.sb_mb);
2660         }
2661
2662         /*
2663          * If we were upcalled from the tcp protocol layer and we have
2664          * fully parsed records ready to go, or there is new data pending,
2665          * or something went wrong, try to wake up a nfsd thread to deal
2666          * with it.
2667          */
2668 done:
2669         /* XXX this code is currently not executed (nfsrv_rcv_upcall) */
2670         if (waitflag == MB_DONTWAIT && (slp->ns_flag & SLP_ACTION_MASK)) {
2671                 lwkt_gettoken(&nfs_token);
2672                 nfsrv_wakenfsd(slp, nparallel_wakeup);
2673                 lwkt_reltoken(&nfs_token);
2674         }
2675 }
2676
2677 /*
2678  * Try and extract an RPC request from the mbuf data list received on a
2679  * stream socket. The "waitflag" argument indicates whether or not it
2680  * can sleep.
2681  */
2682 static int
2683 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2684 {
2685         struct mbuf *m, **mpp;
2686         char *cp1, *cp2;
2687         int len;
2688         struct mbuf *om, *m2, *recm;
2689         u_int32_t recmark;
2690
2691         for (;;) {
2692             if (slp->ns_reclen == 0) {
2693                 if (slp->ns_cc < NFSX_UNSIGNED)
2694                         return (0);
2695                 m = slp->ns_raw;
2696                 if (m->m_len >= NFSX_UNSIGNED) {
2697                         bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2698                         m->m_data += NFSX_UNSIGNED;
2699                         m->m_len -= NFSX_UNSIGNED;
2700                 } else {
2701                         cp1 = (caddr_t)&recmark;
2702                         cp2 = mtod(m, caddr_t);
2703                         while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2704                                 while (m->m_len == 0) {
2705                                         m = m->m_next;
2706                                         cp2 = mtod(m, caddr_t);
2707                                 }
2708                                 *cp1++ = *cp2++;
2709                                 m->m_data++;
2710                                 m->m_len--;
2711                         }
2712                 }
2713                 slp->ns_cc -= NFSX_UNSIGNED;
2714                 recmark = ntohl(recmark);
2715                 slp->ns_reclen = recmark & ~0x80000000;
2716                 if (recmark & 0x80000000)
2717                         slp->ns_flag |= SLP_LASTFRAG;
2718                 else
2719                         slp->ns_flag &= ~SLP_LASTFRAG;
2720                 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2721                         log(LOG_ERR, "%s (%d) from nfs client\n",
2722                             "impossible packet length",
2723                             slp->ns_reclen);
2724                         return (EPERM);
2725                 }
2726             }
2727
2728             /*
2729              * Now get the record part.
2730              *
2731              * Note that slp->ns_reclen may be 0.  Linux sometimes
2732              * generates 0-length RPCs
2733              */
2734             recm = NULL;
2735             if (slp->ns_cc == slp->ns_reclen) {
2736                 recm = slp->ns_raw;
2737                 slp->ns_raw = slp->ns_rawend = NULL;
2738                 slp->ns_cc = slp->ns_reclen = 0;
2739             } else if (slp->ns_cc > slp->ns_reclen) {
2740                 len = 0;
2741                 m = slp->ns_raw;
2742                 om = NULL;
2743
2744                 while (len < slp->ns_reclen) {
2745                         if ((len + m->m_len) > slp->ns_reclen) {
2746                                 m2 = m_copym(m, 0, slp->ns_reclen - len,
2747                                         waitflag);
2748                                 if (m2) {
2749                                         if (om) {
2750                                                 om->m_next = m2;
2751                                                 recm = slp->ns_raw;
2752                                         } else
2753                                                 recm = m2;
2754                                         m->m_data += slp->ns_reclen - len;
2755                                         m->m_len -= slp->ns_reclen - len;
2756                                         len = slp->ns_reclen;
2757                                 } else {
2758                                         return (EWOULDBLOCK);
2759                                 }
2760                         } else if ((len + m->m_len) == slp->ns_reclen) {
2761                                 om = m;
2762                                 len += m->m_len;
2763                                 m = m->m_next;
2764                                 recm = slp->ns_raw;
2765                                 om->m_next = NULL;
2766                         } else {
2767                                 om = m;
2768                                 len += m->m_len;
2769                                 m = m->m_next;
2770                         }
2771                 }
2772                 slp->ns_raw = m;
2773                 slp->ns_cc -= len;
2774                 slp->ns_reclen = 0;
2775             } else {
2776                 return (0);
2777             }
2778
2779             /*
2780              * Accumulate the fragments into a record.
2781              */
2782             mpp = &slp->ns_frag;
2783             while (*mpp)
2784                 mpp = &((*mpp)->m_next);
2785             *mpp = recm;
2786             if (slp->ns_flag & SLP_LASTFRAG) {
2787                 struct nfsrv_rec *rec;
2788                 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2789                 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2790                 if (!rec) {
2791                     m_freem(slp->ns_frag);
2792                 } else {
2793                     nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2794                     rec->nr_address = NULL;
2795                     rec->nr_packet = slp->ns_frag;
2796                     STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2797                     ++slp->ns_numrec;
2798                     slp->ns_flag |= SLP_DOREC;
2799                     ++*countp;
2800                 }
2801                 slp->ns_frag = NULL;
2802             }
2803         }
2804 }
2805
2806 #ifdef INVARIANTS
2807
2808 /*
2809  * Sanity check our mbuf chain.
2810  */
2811 static void
2812 nfs_checkpkt(struct mbuf *m, int len)
2813 {
2814         int xlen = 0;
2815         while (m) {
2816                 xlen += m->m_len;
2817                 m = m->m_next;
2818         }
2819         if (xlen != len) {
2820                 panic("nfs_checkpkt: len mismatch %d/%d mbuf %p\n",
2821                         xlen, len, m);
2822         }
2823 }
2824
2825 #else
2826
2827 static void
2828 nfs_checkpkt(struct mbuf *m __unused, int len __unused)
2829 {
2830 }
2831
2832 #endif
2833
2834 /*
2835  * Parse an RPC header.
2836  *
2837  * If the socket is invalid or no records are pending we return ENOBUFS.
2838  * The caller must deal with NEEDQ races.
2839  */
2840 int
2841 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2842             struct nfsrv_descript **ndp)
2843 {
2844         struct nfsrv_rec *rec;
2845         struct mbuf *m;
2846         struct sockaddr *nam;
2847         struct nfsrv_descript *nd;
2848         int error;
2849
2850         *ndp = NULL;
2851         if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2852                 return (ENOBUFS);
2853         rec = STAILQ_FIRST(&slp->ns_rec);
2854         STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2855         KKASSERT(slp->ns_numrec > 0);
2856         if (--slp->ns_numrec == 0)
2857                 slp->ns_flag &= ~SLP_DOREC;
2858         nam = rec->nr_address;
2859         m = rec->nr_packet;
2860         kfree(rec, M_NFSRVDESC);
2861         MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2862                 M_NFSRVDESC, M_WAITOK);
2863         nd->nd_md = nd->nd_mrep = m;
2864         nd->nd_nam2 = nam;
2865         nd->nd_dpos = mtod(m, caddr_t);
2866         error = nfs_getreq(nd, nfsd, TRUE);
2867         if (error) {
2868                 if (nam) {
2869                         FREE(nam, M_SONAME);
2870                 }
2871                 kfree((caddr_t)nd, M_NFSRVDESC);
2872                 return (error);
2873         }
2874         *ndp = nd;
2875         nfsd->nfsd_nd = nd;
2876         return (0);
2877 }
2878
2879 /*
2880  * Try to assign service sockets to nfsd threads based on the number
2881  * of new rpc requests that have been queued on the service socket.
2882  *
2883  * If no nfsd's are available or additonal requests are pending, set the
2884  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2885  * the work in the nfssvc_sock list when it is finished processing its
2886  * current work.  This flag is only cleared when an nfsd can not find
2887  * any new work to perform.
2888  */
2889 void
2890 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2891 {
2892         struct nfsd *nd;
2893
2894         if ((slp->ns_flag & SLP_VALID) == 0)
2895                 return;
2896         if (nparallel <= 1)
2897                 nparallel = 1;
2898         TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2899                 if (nd->nfsd_flag & NFSD_WAITING) {
2900                         nd->nfsd_flag &= ~NFSD_WAITING;
2901                         if (nd->nfsd_slp)
2902                                 panic("nfsd wakeup");
2903                         nfsrv_slpref(slp);
2904                         nd->nfsd_slp = slp;
2905                         wakeup((caddr_t)nd);
2906                         if (--nparallel == 0)
2907                                 break;
2908                 }
2909         }
2910
2911         /*
2912          * If we couldn't assign slp then the NFSDs are all busy and
2913          * we set a flag indicating that there is pending work.
2914          */
2915         if (nparallel)
2916                 nfsd_head_flag |= NFSD_CHECKSLP;
2917 }
2918 #endif /* NFS_NOSERVER */