kernel - Fix endless nfs error 70 during shutdown
[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  */
39
40 /*
41  * Socket operations for use by nfs
42  */
43
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/proc.h>
47 #include <sys/malloc.h>
48 #include <sys/mount.h>
49 #include <sys/kernel.h>
50 #include <sys/mbuf.h>
51 #include <sys/vnode.h>
52 #include <sys/fcntl.h>
53 #include <sys/protosw.h>
54 #include <sys/resourcevar.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/socketops.h>
58 #include <sys/syslog.h>
59 #include <sys/thread.h>
60 #include <sys/tprintf.h>
61 #include <sys/sysctl.h>
62 #include <sys/signalvar.h>
63
64 #include <sys/signal2.h>
65 #include <sys/mutex2.h>
66 #include <sys/socketvar2.h>
67
68 #include <netinet/in.h>
69 #include <netinet/tcp.h>
70 #include <sys/thread2.h>
71
72 #include "rpcv2.h"
73 #include "nfsproto.h"
74 #include "nfs.h"
75 #include "xdr_subs.h"
76 #include "nfsm_subs.h"
77 #include "nfsmount.h"
78 #include "nfsnode.h"
79 #include "nfsrtt.h"
80
81 #define TRUE    1
82 #define FALSE   0
83
84 /*
85  * RTT calculations are scaled by 256 (8 bits).  A proper fractional
86  * RTT will still be calculated even with a slow NFS timer.
87  */
88 #define NFS_SRTT(r)     (r)->r_nmp->nm_srtt[proct[(r)->r_procnum]]
89 #define NFS_SDRTT(r)    (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum]]
90 #define NFS_RTT_SCALE_BITS      8       /* bits */
91 #define NFS_RTT_SCALE           256     /* value */
92
93 /*
94  * Defines which timer to use for the procnum.
95  * 0 - default
96  * 1 - getattr
97  * 2 - lookup
98  * 3 - read
99  * 4 - write
100  */
101 static int proct[NFS_NPROCS] = {
102         0, 1, 0, 2, 1, 3, 3, 4, 0, 0,   /* 00-09        */
103         0, 0, 0, 0, 0, 0, 3, 3, 0, 0,   /* 10-19        */
104         0, 5, 0, 0, 0, 0,               /* 20-29        */
105 };
106
107 static int multt[NFS_NPROCS] = {
108         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 00-09        */
109         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 10-19        */
110         1, 2, 1, 1, 1, 1,               /* 20-29        */
111 };
112
113 static int nfs_backoff[8] = { 2, 3, 5, 8, 13, 21, 34, 55 };
114 static int nfs_realign_test;
115 static int nfs_realign_count;
116 static int nfs_showrtt;
117 static int nfs_showrexmit;
118 int nfs_maxasyncbio = NFS_MAXASYNCBIO;
119
120 SYSCTL_DECL(_vfs_nfs);
121
122 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0,
123     "Number of times mbufs have been tested for bad alignment");
124 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0,
125     "Number of realignments for badly aligned mbuf data");
126 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrtt, CTLFLAG_RW, &nfs_showrtt, 0,
127     "Show round trip time output");
128 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrexmit, CTLFLAG_RW, &nfs_showrexmit, 0,
129     "Show retransmits info");
130 SYSCTL_INT(_vfs_nfs, OID_AUTO, maxasyncbio, CTLFLAG_RW, &nfs_maxasyncbio, 0,
131     "Max number of asynchronous bio's");
132
133 static int nfs_request_setup(nfsm_info_t info);
134 static int nfs_request_auth(struct nfsreq *rep);
135 static int nfs_request_try(struct nfsreq *rep);
136 static int nfs_request_waitreply(struct nfsreq *rep);
137 static int nfs_request_processreply(nfsm_info_t info, int);
138
139 int nfsrtton = 0;
140 struct nfsrtt nfsrtt;
141 struct callout  nfs_timer_handle;
142
143 static int      nfs_msg (struct thread *,char *,char *);
144 static int      nfs_rcvlock (struct nfsmount *nmp, struct nfsreq *myreq);
145 static void     nfs_rcvunlock (struct nfsmount *nmp);
146 static void     nfs_realign (struct mbuf **pm, int hsiz);
147 static int      nfs_receive (struct nfsmount *nmp, struct nfsreq *rep,
148                                 struct sockaddr **aname, struct mbuf **mp);
149 static void     nfs_softterm (struct nfsreq *rep, int islocked);
150 static void     nfs_hardterm (struct nfsreq *rep, int islocked);
151 static int      nfs_reconnect (struct nfsmount *nmp, struct nfsreq *rep);
152 #ifndef NFS_NOSERVER 
153 static int      nfsrv_getstream (struct nfssvc_sock *, int, int *);
154 static void     nfs_timer_req(struct nfsreq *req);
155 static void     nfs_checkpkt(struct mbuf *m, int len);
156
157 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
158                                     struct nfssvc_sock *slp,
159                                     struct thread *td,
160                                     struct mbuf **mreqp) = {
161         nfsrv_null,
162         nfsrv_getattr,
163         nfsrv_setattr,
164         nfsrv_lookup,
165         nfsrv3_access,
166         nfsrv_readlink,
167         nfsrv_read,
168         nfsrv_write,
169         nfsrv_create,
170         nfsrv_mkdir,
171         nfsrv_symlink,
172         nfsrv_mknod,
173         nfsrv_remove,
174         nfsrv_rmdir,
175         nfsrv_rename,
176         nfsrv_link,
177         nfsrv_readdir,
178         nfsrv_readdirplus,
179         nfsrv_statfs,
180         nfsrv_fsinfo,
181         nfsrv_pathconf,
182         nfsrv_commit,
183         nfsrv_noop,
184         nfsrv_noop,
185         nfsrv_noop,
186         nfsrv_noop
187 };
188 #endif /* NFS_NOSERVER */
189
190 /*
191  * Initialize sockets and congestion for a new NFS connection.
192  * We do not free the sockaddr if error.
193  */
194 int
195 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
196 {
197         struct socket *so;
198         int error;
199         struct sockaddr *saddr;
200         struct sockaddr_in *sin;
201         struct thread *td = &thread0; /* only used for socreate and sobind */
202
203         nmp->nm_so = so = NULL;
204         if (nmp->nm_flag & NFSMNT_FORCE)
205                 return (EINVAL);
206         saddr = nmp->nm_nam;
207         error = socreate(saddr->sa_family, &so, nmp->nm_sotype,
208                 nmp->nm_soproto, td);
209         if (error)
210                 goto bad;
211         nmp->nm_soflags = so->so_proto->pr_flags;
212
213         /*
214          * Some servers require that the client port be a reserved port number.
215          */
216         if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
217                 struct sockopt sopt;
218                 int ip;
219                 struct sockaddr_in ssin;
220
221                 bzero(&sopt, sizeof sopt);
222                 ip = IP_PORTRANGE_LOW;
223                 sopt.sopt_level = IPPROTO_IP;
224                 sopt.sopt_name = IP_PORTRANGE;
225                 sopt.sopt_val = (void *)&ip;
226                 sopt.sopt_valsize = sizeof(ip);
227                 sopt.sopt_td = NULL;
228                 error = sosetopt(so, &sopt);
229                 if (error)
230                         goto bad;
231                 bzero(&ssin, sizeof ssin);
232                 sin = &ssin;
233                 sin->sin_len = sizeof (struct sockaddr_in);
234                 sin->sin_family = AF_INET;
235                 sin->sin_addr.s_addr = INADDR_ANY;
236                 sin->sin_port = htons(0);
237                 error = sobind(so, (struct sockaddr *)sin, td);
238                 if (error)
239                         goto bad;
240                 bzero(&sopt, sizeof sopt);
241                 ip = IP_PORTRANGE_DEFAULT;
242                 sopt.sopt_level = IPPROTO_IP;
243                 sopt.sopt_name = IP_PORTRANGE;
244                 sopt.sopt_val = (void *)&ip;
245                 sopt.sopt_valsize = sizeof(ip);
246                 sopt.sopt_td = NULL;
247                 error = sosetopt(so, &sopt);
248                 if (error)
249                         goto bad;
250         }
251
252         /*
253          * Protocols that do not require connections may be optionally left
254          * unconnected for servers that reply from a port other than NFS_PORT.
255          */
256         if (nmp->nm_flag & NFSMNT_NOCONN) {
257                 if (nmp->nm_soflags & PR_CONNREQUIRED) {
258                         error = ENOTCONN;
259                         goto bad;
260                 }
261         } else {
262                 error = soconnect(so, nmp->nm_nam, td);
263                 if (error)
264                         goto bad;
265
266                 /*
267                  * Wait for the connection to complete. Cribbed from the
268                  * connect system call but with the wait timing out so
269                  * that interruptible mounts don't hang here for a long time.
270                  */
271                 crit_enter();
272                 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
273                         (void) tsleep((caddr_t)&so->so_timeo, 0,
274                                 "nfscon", 2 * hz);
275                         if ((so->so_state & SS_ISCONNECTING) &&
276                             so->so_error == 0 && rep &&
277                             (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
278                                 soclrstate(so, SS_ISCONNECTING);
279                                 crit_exit();
280                                 goto bad;
281                         }
282                 }
283                 if (so->so_error) {
284                         error = so->so_error;
285                         so->so_error = 0;
286                         crit_exit();
287                         goto bad;
288                 }
289                 crit_exit();
290         }
291         so->so_rcv.ssb_timeo = (5 * hz);
292         so->so_snd.ssb_timeo = (5 * hz);
293
294         /*
295          * Get buffer reservation size from sysctl, but impose reasonable
296          * limits.
297          */
298         if (nmp->nm_sotype == SOCK_STREAM) {
299                 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
300                         struct sockopt sopt;
301                         int val;
302
303                         bzero(&sopt, sizeof sopt);
304                         sopt.sopt_level = SOL_SOCKET;
305                         sopt.sopt_name = SO_KEEPALIVE;
306                         sopt.sopt_val = &val;
307                         sopt.sopt_valsize = sizeof val;
308                         val = 1;
309                         sosetopt(so, &sopt);
310                 }
311                 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
312                         struct sockopt sopt;
313                         int val;
314
315                         bzero(&sopt, sizeof sopt);
316                         sopt.sopt_level = IPPROTO_TCP;
317                         sopt.sopt_name = TCP_NODELAY;
318                         sopt.sopt_val = &val;
319                         sopt.sopt_valsize = sizeof val;
320                         val = 1;
321                         sosetopt(so, &sopt);
322
323                         bzero(&sopt, sizeof sopt);
324                         sopt.sopt_level = IPPROTO_TCP;
325                         sopt.sopt_name = TCP_FASTKEEP;
326                         sopt.sopt_val = &val;
327                         sopt.sopt_valsize = sizeof val;
328                         val = 1;
329                         sosetopt(so, &sopt);
330                 }
331         }
332         error = soreserve(so, nfs_soreserve, nfs_soreserve, NULL);
333         if (error)
334                 goto bad;
335         atomic_set_int(&so->so_rcv.ssb_flags, SSB_NOINTR);
336         atomic_set_int(&so->so_snd.ssb_flags, SSB_NOINTR);
337
338         /* Initialize other non-zero congestion variables */
339         nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] = 
340                 nmp->nm_srtt[3] = (NFS_TIMEO << NFS_RTT_SCALE_BITS);
341         nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
342                 nmp->nm_sdrtt[3] = 0;
343         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
344         nmp->nm_timeouts = 0;
345
346         /*
347          * Assign nm_so last.  The moment nm_so is assigned the nfs_timer()
348          * can mess with the socket.
349          */
350         nmp->nm_so = so;
351         return (0);
352
353 bad:
354         if (so) {
355                 soshutdown(so, SHUT_RDWR);
356                 soclose(so, FNONBLOCK);
357         }
358         return (error);
359 }
360
361 /*
362  * Reconnect routine:
363  * Called when a connection is broken on a reliable protocol.
364  * - clean up the old socket
365  * - nfs_connect() again
366  * - set R_NEEDSXMIT for all outstanding requests on mount point
367  * If this fails the mount point is DEAD!
368  * nb: Must be called with the nfs_sndlock() set on the mount point.
369  */
370 static int
371 nfs_reconnect(struct nfsmount *nmp, struct nfsreq *rep)
372 {
373         struct nfsreq *req;
374         int error;
375
376         nfs_disconnect(nmp);
377         if (nmp->nm_rxstate >= NFSSVC_STOPPING)
378                 return (EINTR);
379         while ((error = nfs_connect(nmp, rep)) != 0) {
380                 if (error == EINTR || error == ERESTART)
381                         return (EINTR);
382                 if (error == EINVAL)
383                         return (error);
384                 if (nmp->nm_rxstate >= NFSSVC_STOPPING)
385                         return (EINTR);
386                 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
387         }
388
389         /*
390          * Loop through outstanding request list and fix up all requests
391          * on old socket.
392          */
393         crit_enter();
394         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
395                 KKASSERT(req->r_nmp == nmp);
396                 req->r_flags |= R_NEEDSXMIT;
397         }
398         crit_exit();
399         return (0);
400 }
401
402 /*
403  * NFS disconnect. Clean up and unlink.
404  */
405 void
406 nfs_disconnect(struct nfsmount *nmp)
407 {
408         struct socket *so;
409
410         if (nmp->nm_so) {
411                 so = nmp->nm_so;
412                 nmp->nm_so = NULL;
413                 soshutdown(so, SHUT_RDWR);
414                 soclose(so, FNONBLOCK);
415         }
416 }
417
418 void
419 nfs_safedisconnect(struct nfsmount *nmp)
420 {
421         nfs_rcvlock(nmp, NULL);
422         nfs_disconnect(nmp);
423         nfs_rcvunlock(nmp);
424 }
425
426 /*
427  * This is the nfs send routine. For connection based socket types, it
428  * must be called with an nfs_sndlock() on the socket.
429  * "rep == NULL" indicates that it has been called from a server.
430  * For the client side:
431  * - return EINTR if the RPC is terminated, 0 otherwise
432  * - set R_NEEDSXMIT if the send fails for any reason
433  * - do any cleanup required by recoverable socket errors (?)
434  * For the server side:
435  * - return EINTR or ERESTART if interrupted by a signal
436  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
437  * - do any cleanup required by recoverable socket errors (?)
438  */
439 int
440 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
441          struct nfsreq *rep)
442 {
443         struct sockaddr *sendnam;
444         int error, soflags, flags;
445
446         if (rep) {
447                 if (rep->r_flags & R_SOFTTERM) {
448                         m_freem(top);
449                         return (EINTR);
450                 }
451                 if ((so = rep->r_nmp->nm_so) == NULL) {
452                         rep->r_flags |= R_NEEDSXMIT;
453                         m_freem(top);
454                         return (0);
455                 }
456                 rep->r_flags &= ~R_NEEDSXMIT;
457                 soflags = rep->r_nmp->nm_soflags;
458         } else {
459                 soflags = so->so_proto->pr_flags;
460         }
461         if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
462                 sendnam = NULL;
463         else
464                 sendnam = nam;
465         if (so->so_type == SOCK_SEQPACKET)
466                 flags = MSG_EOR;
467         else
468                 flags = 0;
469
470         /*
471          * calls pru_sosend -> sosend -> so_pru_send -> netrpc
472          */
473         error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
474                               curthread /*XXX*/);
475
476         /*
477          * ENOBUFS for dgram sockets is transient and non fatal.
478          * No need to log, and no need to break a soft mount.
479          */
480         if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
481                 error = 0;
482                 /*
483                  * do backoff retransmit on client
484                  */
485                 if (rep) {
486                         if ((rep->r_nmp->nm_state & NFSSTA_SENDSPACE) == 0) {
487                                 rep->r_nmp->nm_state |= NFSSTA_SENDSPACE;
488                                 kprintf("Warning: NFS: Insufficient sendspace "
489                                         "(%lu),\n"
490                                         "\t You must increase vfs.nfs.soreserve"
491                                         "or decrease vfs.nfs.maxasyncbio\n",
492                                         so->so_snd.ssb_hiwat);
493                         }
494                         rep->r_flags |= R_NEEDSXMIT;
495                 }
496         }
497
498         if (error) {
499                 if (rep) {
500                         log(LOG_INFO, "nfs send error %d for server %s\n",error,
501                             rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
502                         /*
503                          * Deal with errors for the client side.
504                          */
505                         if (rep->r_flags & R_SOFTTERM)
506                                 error = EINTR;
507                         else
508                                 rep->r_flags |= R_NEEDSXMIT;
509                 } else {
510                         log(LOG_INFO, "nfsd send error %d\n", error);
511                 }
512
513                 /*
514                  * Handle any recoverable (soft) socket errors here. (?)
515                  */
516                 if (error != EINTR && error != ERESTART &&
517                         error != EWOULDBLOCK && error != EPIPE)
518                         error = 0;
519         }
520         return (error);
521 }
522
523 /*
524  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
525  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
526  * Mark and consolidate the data into a new mbuf list.
527  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
528  *     small mbufs.
529  * For SOCK_STREAM we must be very careful to read an entire record once
530  * we have read any of it, even if the system call has been interrupted.
531  */
532 static int
533 nfs_receive(struct nfsmount *nmp, struct nfsreq *rep,
534             struct sockaddr **aname, struct mbuf **mp)
535 {
536         struct socket *so;
537         struct sockbuf sio;
538         struct uio auio;
539         struct iovec aio;
540         struct mbuf *m;
541         struct mbuf *control;
542         u_int32_t len;
543         struct sockaddr **getnam;
544         int error, sotype, rcvflg;
545         struct thread *td = curthread;  /* XXX */
546
547         /*
548          * Set up arguments for soreceive()
549          */
550         *mp = NULL;
551         *aname = NULL;
552         sotype = nmp->nm_sotype;
553
554         /*
555          * For reliable protocols, lock against other senders/receivers
556          * in case a reconnect is necessary.
557          * For SOCK_STREAM, first get the Record Mark to find out how much
558          * more there is to get.
559          * We must lock the socket against other receivers
560          * until we have an entire rpc request/reply.
561          */
562         if (sotype != SOCK_DGRAM) {
563                 error = nfs_sndlock(nmp, rep);
564                 if (error)
565                         return (error);
566 tryagain:
567                 /*
568                  * Check for fatal errors and resending request.
569                  */
570                 /*
571                  * Ugh: If a reconnect attempt just happened, nm_so
572                  * would have changed. NULL indicates a failed
573                  * attempt that has essentially shut down this
574                  * mount point.
575                  */
576                 if (rep && (rep->r_mrep || (rep->r_flags & R_SOFTTERM))) {
577                         nfs_sndunlock(nmp);
578                         return (EINTR);
579                 }
580                 so = nmp->nm_so;
581                 if (so == NULL) {
582                         error = nfs_reconnect(nmp, rep);
583                         if (error) {
584                                 nfs_sndunlock(nmp);
585                                 return (error);
586                         }
587                         goto tryagain;
588                 }
589                 while (rep && (rep->r_flags & R_NEEDSXMIT)) {
590                         m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
591                         nfsstats.rpcretries++;
592                         error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
593                         if (error) {
594                                 if (error == EINTR || error == ERESTART ||
595                                     (error = nfs_reconnect(nmp, rep)) != 0) {
596                                         nfs_sndunlock(nmp);
597                                         return (error);
598                                 }
599                                 goto tryagain;
600                         }
601                 }
602                 nfs_sndunlock(nmp);
603                 if (sotype == SOCK_STREAM) {
604                         /*
605                          * Get the length marker from the stream
606                          */
607                         aio.iov_base = (caddr_t)&len;
608                         aio.iov_len = sizeof(u_int32_t);
609                         auio.uio_iov = &aio;
610                         auio.uio_iovcnt = 1;
611                         auio.uio_segflg = UIO_SYSSPACE;
612                         auio.uio_rw = UIO_READ;
613                         auio.uio_offset = 0;
614                         auio.uio_resid = sizeof(u_int32_t);
615                         auio.uio_td = td;
616                         do {
617                            rcvflg = MSG_WAITALL;
618                            error = so_pru_soreceive(so, NULL, &auio, NULL,
619                                                     NULL, &rcvflg);
620                            if (error == EWOULDBLOCK && rep) {
621                                 if (rep->r_flags & R_SOFTTERM)
622                                         return (EINTR);
623                            }
624                         } while (error == EWOULDBLOCK);
625
626                         if (error == 0 && auio.uio_resid > 0) {
627                             /*
628                              * Only log short packets if not EOF
629                              */
630                             if (auio.uio_resid != sizeof(u_int32_t))
631                             log(LOG_INFO,
632                                  "short receive (%d/%d) from nfs server %s\n",
633                                  (int)(sizeof(u_int32_t) - auio.uio_resid),
634                                  (int)sizeof(u_int32_t),
635                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
636                             error = EPIPE;
637                         }
638                         if (error)
639                                 goto errout;
640                         len = ntohl(len) & ~0x80000000;
641                         /*
642                          * This is SERIOUS! We are out of sync with the sender
643                          * and forcing a disconnect/reconnect is all I can do.
644                          */
645                         if (len > NFS_MAXPACKET) {
646                             log(LOG_ERR, "%s (%d) from nfs server %s\n",
647                                 "impossible packet length",
648                                 len,
649                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
650                             error = EFBIG;
651                             goto errout;
652                         }
653
654                         /*
655                          * Get the rest of the packet as an mbuf chain
656                          */
657                         sbinit(&sio, len);
658                         do {
659                             rcvflg = MSG_WAITALL;
660                             error = so_pru_soreceive(so, NULL, NULL, &sio,
661                                                      NULL, &rcvflg);
662                         } while (error == EWOULDBLOCK || error == EINTR ||
663                                  error == ERESTART);
664                         if (error == 0 && sio.sb_cc != len) {
665                             if (sio.sb_cc != 0)
666                             log(LOG_INFO,
667                                 "short receive (%zu/%d) from nfs server %s\n",
668                                 (size_t)len - auio.uio_resid, len,
669                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
670                             error = EPIPE;
671                         }
672                         *mp = sio.sb_mb;
673                 } else {
674                         /*
675                          * Non-stream, so get the whole packet by not
676                          * specifying MSG_WAITALL and by specifying a large
677                          * length.
678                          *
679                          * We have no use for control msg., but must grab them
680                          * and then throw them away so we know what is going
681                          * on.
682                          */
683                         sbinit(&sio, 100000000);
684                         do {
685                             rcvflg = 0;
686                             error =  so_pru_soreceive(so, NULL, NULL, &sio,
687                                                       &control, &rcvflg);
688                             if (control)
689                                 m_freem(control);
690                             if (error == EWOULDBLOCK && rep) {
691                                 if (rep->r_flags & R_SOFTTERM) {
692                                         m_freem(sio.sb_mb);
693                                         return (EINTR);
694                                 }
695                             }
696                         } while (error == EWOULDBLOCK ||
697                                  (error == 0 && sio.sb_mb == NULL && control));
698                         if ((rcvflg & MSG_EOR) == 0)
699                                 kprintf("Egad!!\n");
700                         if (error == 0 && sio.sb_mb == NULL)
701                                 error = EPIPE;
702                         len = sio.sb_cc;
703                         *mp = sio.sb_mb;
704                 }
705 errout:
706                 if (error && error != EINTR && error != ERESTART) {
707                         m_freem(*mp);
708                         *mp = NULL;
709                         if (error != EPIPE) {
710                                 log(LOG_INFO,
711                                     "receive error %d from nfs server %s\n",
712                                     error,
713                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
714                         }
715                         error = nfs_sndlock(nmp, rep);
716                         if (!error) {
717                                 error = nfs_reconnect(nmp, rep);
718                                 if (!error)
719                                         goto tryagain;
720                                 else
721                                         nfs_sndunlock(nmp);
722                         }
723                 }
724         } else {
725                 if ((so = nmp->nm_so) == NULL)
726                         return (EACCES);
727                 if (so->so_state & SS_ISCONNECTED)
728                         getnam = NULL;
729                 else
730                         getnam = aname;
731                 sbinit(&sio, 100000000);
732                 do {
733                         rcvflg = 0;
734                         error =  so_pru_soreceive(so, getnam, NULL, &sio,
735                                                   NULL, &rcvflg);
736                         if (error == EWOULDBLOCK && rep &&
737                             (rep->r_flags & R_SOFTTERM)) {
738                                 m_freem(sio.sb_mb);
739                                 return (EINTR);
740                         }
741                 } while (error == EWOULDBLOCK);
742
743                 len = sio.sb_cc;
744                 *mp = sio.sb_mb;
745
746                 /*
747                  * A shutdown may result in no error and no mbuf.
748                  * Convert to EPIPE.
749                  */
750                 if (*mp == NULL && error == 0)
751                         error = EPIPE;
752         }
753         if (error) {
754                 m_freem(*mp);
755                 *mp = NULL;
756         }
757
758         /*
759          * Search for any mbufs that are not a multiple of 4 bytes long
760          * or with m_data not longword aligned.
761          * These could cause pointer alignment problems, so copy them to
762          * well aligned mbufs.
763          */
764         nfs_realign(mp, 5 * NFSX_UNSIGNED);
765         return (error);
766 }
767
768 /*
769  * Implement receipt of reply on a socket.
770  *
771  * We must search through the list of received datagrams matching them
772  * with outstanding requests using the xid, until ours is found.
773  *
774  * If myrep is NULL we process packets on the socket until
775  * interrupted or until nm_reqrxq is non-empty.
776  */
777 /* ARGSUSED */
778 int
779 nfs_reply(struct nfsmount *nmp, struct nfsreq *myrep)
780 {
781         struct nfsreq *rep;
782         struct sockaddr *nam;
783         u_int32_t rxid;
784         u_int32_t *tl;
785         int error;
786         struct nfsm_info info;
787
788         /*
789          * Loop around until we get our own reply
790          */
791         for (;;) {
792                 /*
793                  * Lock against other receivers so that I don't get stuck in
794                  * sbwait() after someone else has received my reply for me.
795                  * Also necessary for connection based protocols to avoid
796                  * race conditions during a reconnect.
797                  *
798                  * If nfs_rcvlock() returns EALREADY, that means that
799                  * the reply has already been recieved by another
800                  * process and we can return immediately.  In this
801                  * case, the lock is not taken to avoid races with
802                  * other processes.
803                  */
804                 info.mrep = NULL;
805
806                 error = nfs_rcvlock(nmp, myrep);
807                 if (error == EALREADY)
808                         return (0);
809                 if (error)
810                         return (error);
811
812                 /*
813                  * If myrep is NULL we are the receiver helper thread.
814                  * Stop waiting for incoming replies if there are
815                  * messages sitting on reqrxq that we need to process,
816                  * or if a shutdown request is pending.
817                  */
818                 if (myrep == NULL && (TAILQ_FIRST(&nmp->nm_reqrxq) ||
819                     nmp->nm_rxstate > NFSSVC_PENDING)) {
820                         nfs_rcvunlock(nmp);
821                         return(EWOULDBLOCK);
822                 }
823
824                 /*
825                  * Get the next Rpc reply off the socket
826                  *
827                  * We cannot release the receive lock until we've
828                  * filled in rep->r_mrep, otherwise a waiting
829                  * thread may deadlock in soreceive with no incoming
830                  * packets expected.
831                  */
832                 error = nfs_receive(nmp, myrep, &nam, &info.mrep);
833                 if (error) {
834                         /*
835                          * Ignore routing errors on connectionless protocols??
836                          */
837                         nfs_rcvunlock(nmp);
838                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
839                                 if (nmp->nm_so == NULL)
840                                         return (error);
841                                 nmp->nm_so->so_error = 0;
842                                 continue;
843                         }
844                         return (error);
845                 }
846                 if (nam)
847                         kfree(nam, M_SONAME);
848
849                 /*
850                  * Get the xid and check that it is an rpc reply
851                  */
852                 info.md = info.mrep;
853                 info.dpos = mtod(info.md, caddr_t);
854                 NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
855                 rxid = *tl++;
856                 if (*tl != rpc_reply) {
857                         nfsstats.rpcinvalid++;
858                         m_freem(info.mrep);
859                         info.mrep = NULL;
860 nfsmout:
861                         nfs_rcvunlock(nmp);
862                         continue;
863                 }
864
865                 /*
866                  * Loop through the request list to match up the reply
867                  * Iff no match, just drop the datagram.  On match, set
868                  * r_mrep atomically to prevent the timer from messing
869                  * around with the request after we have exited the critical
870                  * section.
871                  */
872                 crit_enter();
873                 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
874                         if (rep->r_mrep == NULL && rxid == rep->r_xid)
875                                 break;
876                 }
877
878                 /*
879                  * Fill in the rest of the reply if we found a match.
880                  *
881                  * Deal with duplicate responses if there was no match.
882                  */
883                 if (rep) {
884                         rep->r_md = info.md;
885                         rep->r_dpos = info.dpos;
886                         if (nfsrtton) {
887                                 struct rttl *rt;
888
889                                 rt = &nfsrtt.rttl[nfsrtt.pos];
890                                 rt->proc = rep->r_procnum;
891                                 rt->rto = 0;
892                                 rt->sent = 0;
893                                 rt->cwnd = nmp->nm_maxasync_scaled;
894                                 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
895                                 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
896                                 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
897                                 getmicrotime(&rt->tstamp);
898                                 if (rep->r_flags & R_TIMING)
899                                         rt->rtt = rep->r_rtt;
900                                 else
901                                         rt->rtt = 1000000;
902                                 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
903                         }
904
905                         /*
906                          * New congestion control is based only on async
907                          * requests.
908                          */
909                         if (nmp->nm_maxasync_scaled < NFS_MAXASYNC_SCALED)
910                                 ++nmp->nm_maxasync_scaled;
911                         if (rep->r_flags & R_SENT) {
912                                 rep->r_flags &= ~R_SENT;
913                         }
914                         /*
915                          * Update rtt using a gain of 0.125 on the mean
916                          * and a gain of 0.25 on the deviation.
917                          *
918                          * NOTE SRTT/SDRTT are only good if R_TIMING is set.
919                          */
920                         if ((rep->r_flags & R_TIMING) && rep->r_rexmit == 0) {
921                                 /*
922                                  * Since the timer resolution of
923                                  * NFS_HZ is so course, it can often
924                                  * result in r_rtt == 0. Since
925                                  * r_rtt == N means that the actual
926                                  * rtt is between N+dt and N+2-dt ticks,
927                                  * add 1.
928                                  */
929                                 int n;
930                                 int d;
931
932 #define NFSRSB  NFS_RTT_SCALE_BITS
933                                 n = ((NFS_SRTT(rep) * 7) +
934                                      (rep->r_rtt << NFSRSB)) >> 3;
935                                 d = n - NFS_SRTT(rep);
936                                 NFS_SRTT(rep) = n;
937
938                                 /*
939                                  * Don't let the jitter calculation decay
940                                  * too quickly, but we want a fast rampup.
941                                  */
942                                 if (d < 0)
943                                         d = -d;
944                                 d <<= NFSRSB;
945                                 if (d < NFS_SDRTT(rep))
946                                         n = ((NFS_SDRTT(rep) * 15) + d) >> 4;
947                                 else
948                                         n = ((NFS_SDRTT(rep) * 3) + d) >> 2;
949                                 NFS_SDRTT(rep) = n;
950 #undef NFSRSB
951                         }
952                         nmp->nm_timeouts = 0;
953                         rep->r_mrep = info.mrep;
954                         nfs_hardterm(rep, 0);
955                 } else {
956                         /*
957                          * Extract vers, prog, nfsver, procnum.  A duplicate
958                          * response means we didn't wait long enough so
959                          * we increase the SRTT to avoid future spurious
960                          * timeouts.
961                          */
962                         u_int procnum = nmp->nm_lastreprocnum;
963                         int n;
964
965                         if (procnum < NFS_NPROCS && proct[procnum]) {
966                                 if (nfs_showrexmit)
967                                         kprintf("D");
968                                 n = nmp->nm_srtt[proct[procnum]];
969                                 n += NFS_ASYSCALE * NFS_HZ;
970                                 if (n < NFS_ASYSCALE * NFS_HZ * 10)
971                                         n = NFS_ASYSCALE * NFS_HZ * 10;
972                                 nmp->nm_srtt[proct[procnum]] = n;
973                         }
974                 }
975                 nfs_rcvunlock(nmp);
976                 crit_exit();
977
978                 /*
979                  * If not matched to a request, drop it.
980                  * If it's mine, get out.
981                  */
982                 if (rep == NULL) {
983                         nfsstats.rpcunexpected++;
984                         m_freem(info.mrep);
985                         info.mrep = NULL;
986                 } else if (rep == myrep) {
987                         if (rep->r_mrep == NULL)
988                                 panic("nfsreply nil");
989                         return (0);
990                 }
991         }
992 }
993
994 /*
995  * Run the request state machine until the target state is reached
996  * or a fatal error occurs.  The target state is not run.  Specifying
997  * a target of NFSM_STATE_DONE runs the state machine until the rpc
998  * is complete.
999  *
1000  * EINPROGRESS is returned for all states other then the DONE state,
1001  * indicating that the rpc is still in progress.
1002  */
1003 int
1004 nfs_request(struct nfsm_info *info, nfsm_state_t bstate, nfsm_state_t estate)
1005 {
1006         struct nfsreq *req;
1007
1008         while (info->state >= bstate && info->state < estate) {
1009                 switch(info->state) {
1010                 case NFSM_STATE_SETUP:
1011                         /*
1012                          * Setup the nfsreq.  Any error which occurs during
1013                          * this state is fatal.
1014                          */
1015                         info->error = nfs_request_setup(info);
1016                         if (info->error) {
1017                                 info->state = NFSM_STATE_DONE;
1018                                 return (info->error);
1019                         } else {
1020                                 req = info->req;
1021                                 req->r_mrp = &info->mrep;
1022                                 req->r_mdp = &info->md;
1023                                 req->r_dposp = &info->dpos;
1024                                 info->state = NFSM_STATE_AUTH;
1025                         }
1026                         break;
1027                 case NFSM_STATE_AUTH:
1028                         /*
1029                          * Authenticate the nfsreq.  Any error which occurs
1030                          * during this state is fatal.
1031                          */
1032                         info->error = nfs_request_auth(info->req);
1033                         if (info->error) {
1034                                 info->state = NFSM_STATE_DONE;
1035                                 return (info->error);
1036                         } else {
1037                                 info->state = NFSM_STATE_TRY;
1038                         }
1039                         break;
1040                 case NFSM_STATE_TRY:
1041                         /*
1042                          * Transmit or retransmit attempt.  An error in this
1043                          * state is ignored and we always move on to the
1044                          * next state.
1045                          *
1046                          * This can trivially race the receiver if the
1047                          * request is asynchronous.  nfs_request_try()
1048                          * will thus set the state for us and we
1049                          * must also return immediately if we are
1050                          * running an async state machine, because
1051                          * info can become invalid due to races after
1052                          * try() returns.
1053                          */
1054                         if (info->req->r_flags & R_ASYNC) {
1055                                 nfs_request_try(info->req);
1056                                 if (estate == NFSM_STATE_WAITREPLY)
1057                                         return (EINPROGRESS);
1058                         } else {
1059                                 nfs_request_try(info->req);
1060                                 info->state = NFSM_STATE_WAITREPLY;
1061                         }
1062                         break;
1063                 case NFSM_STATE_WAITREPLY:
1064                         /*
1065                          * Wait for a reply or timeout and move on to the
1066                          * next state.  The error returned by this state
1067                          * is passed to the processing code in the next
1068                          * state.
1069                          */
1070                         info->error = nfs_request_waitreply(info->req);
1071                         info->state = NFSM_STATE_PROCESSREPLY;
1072                         break;
1073                 case NFSM_STATE_PROCESSREPLY:
1074                         /*
1075                          * Process the reply or timeout.  Errors which occur
1076                          * in this state may cause the state machine to
1077                          * go back to an earlier state, and are fatal
1078                          * otherwise.
1079                          */
1080                         info->error = nfs_request_processreply(info,
1081                                                                info->error);
1082                         switch(info->error) {
1083                         case ENEEDAUTH:
1084                                 info->state = NFSM_STATE_AUTH;
1085                                 break;
1086                         case EAGAIN:
1087                                 info->state = NFSM_STATE_TRY;
1088                                 break;
1089                         default:
1090                                 /*
1091                                  * Operation complete, with or without an
1092                                  * error.  We are done.
1093                                  */
1094                                 info->req = NULL;
1095                                 info->state = NFSM_STATE_DONE;
1096                                 return (info->error);
1097                         }
1098                         break;
1099                 case NFSM_STATE_DONE:
1100                         /*
1101                          * Shouldn't be reached
1102                          */
1103                         return (info->error);
1104                         /* NOT REACHED */
1105                 }
1106         }
1107
1108         /*
1109          * If we are done return the error code (if any).
1110          * Otherwise return EINPROGRESS.
1111          */
1112         if (info->state == NFSM_STATE_DONE)
1113                 return (info->error);
1114         return (EINPROGRESS);
1115 }
1116
1117 /*
1118  * nfs_request - goes something like this
1119  *      - fill in request struct
1120  *      - links it into list
1121  *      - calls nfs_send() for first transmit
1122  *      - calls nfs_receive() to get reply
1123  *      - break down rpc header and return with nfs reply pointed to
1124  *        by mrep or error
1125  * nb: always frees up mreq mbuf list
1126  */
1127 static int
1128 nfs_request_setup(nfsm_info_t info)
1129 {
1130         struct nfsreq *req;
1131         struct nfsmount *nmp;
1132         struct mbuf *m;
1133         int i;
1134
1135         /*
1136          * Reject requests while attempting a forced unmount.
1137          */
1138         if (info->vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
1139                 m_freem(info->mreq);
1140                 info->mreq = NULL;
1141                 return (EIO);
1142         }
1143         nmp = VFSTONFS(info->vp->v_mount);
1144         req = kmalloc(sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
1145         req->r_nmp = nmp;
1146         req->r_vp = info->vp;
1147         req->r_td = info->td;
1148         req->r_procnum = info->procnum;
1149         req->r_mreq = NULL;
1150         req->r_cred = info->cred;
1151
1152         i = 0;
1153         m = info->mreq;
1154         while (m) {
1155                 i += m->m_len;
1156                 m = m->m_next;
1157         }
1158         req->r_mrest = info->mreq;
1159         req->r_mrest_len = i;
1160
1161         /*
1162          * The presence of a non-NULL r_info in req indicates
1163          * async completion via our helper threads.  See the receiver
1164          * code.
1165          */
1166         if (info->bio) {
1167                 req->r_info = info;
1168                 req->r_flags = R_ASYNC;
1169         } else {
1170                 req->r_info = NULL;
1171                 req->r_flags = 0;
1172         }
1173         info->req = req;
1174         return(0);
1175 }
1176
1177 static int
1178 nfs_request_auth(struct nfsreq *rep)
1179 {
1180         struct nfsmount *nmp = rep->r_nmp;
1181         struct mbuf *m;
1182         char nickv[RPCX_NICKVERF];
1183         int error = 0, auth_len, auth_type;
1184         int verf_len;
1185         u_int32_t xid;
1186         char *auth_str, *verf_str;
1187         struct ucred *cred;
1188
1189         cred = rep->r_cred;
1190         rep->r_failed_auth = 0;
1191
1192         /*
1193          * Get the RPC header with authorization.
1194          */
1195         verf_str = auth_str = NULL;
1196         if (nmp->nm_flag & NFSMNT_KERB) {
1197                 verf_str = nickv;
1198                 verf_len = sizeof (nickv);
1199                 auth_type = RPCAUTH_KERB4;
1200                 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1201                 if (rep->r_failed_auth ||
1202                     nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1203                                     verf_str, verf_len)) {
1204                         error = nfs_getauth(nmp, rep, cred, &auth_str,
1205                                 &auth_len, verf_str, &verf_len, rep->r_key);
1206                         if (error) {
1207                                 m_freem(rep->r_mrest);
1208                                 rep->r_mrest = NULL;
1209                                 kfree((caddr_t)rep, M_NFSREQ);
1210                                 return (error);
1211                         }
1212                 }
1213         } else {
1214                 auth_type = RPCAUTH_UNIX;
1215                 if (cred->cr_ngroups < 1)
1216                         panic("nfsreq nogrps");
1217                 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1218                         nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1219                         5 * NFSX_UNSIGNED;
1220         }
1221         if (rep->r_mrest)
1222                 nfs_checkpkt(rep->r_mrest, rep->r_mrest_len);
1223         m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1224                         auth_len, auth_str, verf_len, verf_str,
1225                         rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1226         rep->r_mrest = NULL;
1227         if (auth_str)
1228                 kfree(auth_str, M_TEMP);
1229
1230         /*
1231          * For stream protocols, insert a Sun RPC Record Mark.
1232          */
1233         if (nmp->nm_sotype == SOCK_STREAM) {
1234                 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1235                 if (m == NULL) {
1236                         kfree(rep, M_NFSREQ);
1237                         return (ENOBUFS);
1238                 }
1239                 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1240                          (m->m_pkthdr.len - NFSX_UNSIGNED));
1241         }
1242
1243         nfs_checkpkt(m, m->m_pkthdr.len);
1244
1245         rep->r_mreq = m;
1246         rep->r_xid = xid;
1247         return (0);
1248 }
1249
1250 static int
1251 nfs_request_try(struct nfsreq *rep)
1252 {
1253         struct nfsmount *nmp = rep->r_nmp;
1254         struct mbuf *m2;
1255         int error;
1256
1257         /*
1258          * Request is not on any queue, only the owner has access to it
1259          * so it should not be locked by anyone atm.
1260          *
1261          * Interlock to prevent races.  While locked the only remote
1262          * action possible is for r_mrep to be set (once we enqueue it).
1263          */
1264         if (rep->r_flags == 0xdeadc0de) {
1265                 print_backtrace(-1);
1266                 panic("flags nbad\n");
1267         }
1268         KKASSERT((rep->r_flags & (R_LOCKED | R_ONREQQ)) == 0);
1269         if (nmp->nm_flag & NFSMNT_SOFT)
1270                 rep->r_retry = nmp->nm_retry;
1271         else
1272                 rep->r_retry = NFS_MAXREXMIT + 1;       /* past clip limit */
1273         rep->r_rtt = rep->r_rexmit = 0;
1274         if (proct[rep->r_procnum] > 0)
1275                 rep->r_flags |= R_TIMING | R_LOCKED;
1276         else
1277                 rep->r_flags |= R_LOCKED;
1278         rep->r_mrep = NULL;
1279
1280         nfsstats.rpcrequests++;
1281
1282         if (nmp->nm_flag & NFSMNT_FORCE) {
1283                 rep->r_flags |= R_SOFTTERM;
1284                 rep->r_flags &= ~R_LOCKED;
1285                 return (0);
1286         }
1287         rep->r_flags |= R_NEEDSXMIT;    /* in case send lock races us */
1288
1289         /*
1290          * Do the client side RPC.
1291          *
1292          * Chain request into list of outstanding requests. Be sure
1293          * to put it LAST so timer finds oldest requests first.  Note
1294          * that our control of R_LOCKED prevents the request from
1295          * getting ripped out from under us or transmitted by the
1296          * timer code.
1297          *
1298          * For requests with info structures we must atomically set the
1299          * info's state because the structure could become invalid upon
1300          * return due to races (i.e., if async)
1301          */
1302         crit_enter();
1303         mtx_link_init(&rep->r_link);
1304         KKASSERT((rep->r_flags & R_ONREQQ) == 0);
1305         TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1306         rep->r_flags |= R_ONREQQ;
1307         ++nmp->nm_reqqlen;
1308         if (rep->r_flags & R_ASYNC)
1309                 rep->r_info->state = NFSM_STATE_WAITREPLY;
1310         crit_exit();
1311
1312         error = 0;
1313
1314         /*
1315          * Send if we can.  Congestion control is not handled here any more
1316          * becausing trying to defer the initial send based on the nfs_timer
1317          * requires having a very fast nfs_timer, which is silly.
1318          */
1319         if (nmp->nm_so) {
1320                 if (nmp->nm_soflags & PR_CONNREQUIRED)
1321                         error = nfs_sndlock(nmp, rep);
1322                 if (error == 0 && (rep->r_flags & R_NEEDSXMIT)) {
1323                         m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1324                         error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1325                         rep->r_flags &= ~R_NEEDSXMIT;
1326                         if ((rep->r_flags & R_SENT) == 0) {
1327                                 rep->r_flags |= R_SENT;
1328                         }
1329                         if (nmp->nm_soflags & PR_CONNREQUIRED)
1330                                 nfs_sndunlock(nmp);
1331                 }
1332         } else {
1333                 rep->r_rtt = -1;
1334         }
1335         if (error == EPIPE)
1336                 error = 0;
1337
1338         /*
1339          * Release the lock.  The only remote action that may have occurred
1340          * would have been the setting of rep->r_mrep.  If this occured
1341          * and the request was async we have to move it to the reader
1342          * thread's queue for action.
1343          *
1344          * For async requests also make sure the reader is woken up so
1345          * it gets on the socket to read responses.
1346          */
1347         crit_enter();
1348         if (rep->r_flags & R_ASYNC) {
1349                 if (rep->r_mrep)
1350                         nfs_hardterm(rep, 1);
1351                 rep->r_flags &= ~R_LOCKED;
1352                 nfssvc_iod_reader_wakeup(nmp);
1353         } else {
1354                 rep->r_flags &= ~R_LOCKED;
1355         }
1356         if (rep->r_flags & R_WANTED) {
1357                 rep->r_flags &= ~R_WANTED;
1358                 wakeup(rep);
1359         }
1360         crit_exit();
1361         return (error);
1362 }
1363
1364 /*
1365  * This code is only called for synchronous requests.  Completed synchronous
1366  * requests are left on reqq and we remove them before moving on to the
1367  * processing state.
1368  */
1369 static int
1370 nfs_request_waitreply(struct nfsreq *rep)
1371 {
1372         struct nfsmount *nmp = rep->r_nmp;
1373         int error;
1374
1375         KKASSERT((rep->r_flags & R_ASYNC) == 0);
1376
1377         /*
1378          * Wait until the request is finished.
1379          */
1380         error = nfs_reply(nmp, rep);
1381
1382         /*
1383          * RPC done, unlink the request, but don't rip it out from under
1384          * the callout timer.
1385          *
1386          * Once unlinked no other receiver or the timer will have
1387          * visibility, so we do not have to set R_LOCKED.
1388          */
1389         crit_enter();
1390         while (rep->r_flags & R_LOCKED) {
1391                 rep->r_flags |= R_WANTED;
1392                 tsleep(rep, 0, "nfstrac", 0);
1393         }
1394         KKASSERT(rep->r_flags & R_ONREQQ);
1395         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1396         rep->r_flags &= ~R_ONREQQ;
1397         --nmp->nm_reqqlen;
1398         if (TAILQ_FIRST(&nmp->nm_bioq) &&
1399             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1400                 nfssvc_iod_writer_wakeup(nmp);
1401         }
1402         crit_exit();
1403
1404         /*
1405          * Decrement the outstanding request count.
1406          */
1407         if (rep->r_flags & R_SENT) {
1408                 rep->r_flags &= ~R_SENT;
1409         }
1410         return (error);
1411 }
1412
1413 /*
1414  * Process reply with error returned from nfs_requet_waitreply().
1415  *
1416  * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1417  * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1418  */
1419 static int
1420 nfs_request_processreply(nfsm_info_t info, int error)
1421 {
1422         struct nfsreq *req = info->req;
1423         struct nfsmount *nmp = req->r_nmp;
1424         u_int32_t *tl;
1425         int verf_type;
1426         int i;
1427
1428         /*
1429          * If there was a successful reply and a tprintf msg.
1430          * tprintf a response.
1431          */
1432         if (error == 0 && (req->r_flags & R_TPRINTFMSG)) {
1433                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1434                     "is alive again");
1435         }
1436         info->mrep = req->r_mrep;
1437         info->md = req->r_md;
1438         info->dpos = req->r_dpos;
1439         if (error) {
1440                 m_freem(req->r_mreq);
1441                 req->r_mreq = NULL;
1442                 kfree(req, M_NFSREQ);
1443                 info->req = NULL;
1444                 return (error);
1445         }
1446
1447         /*
1448          * break down the rpc header and check if ok
1449          */
1450         NULLOUT(tl = nfsm_dissect(info, 3 * NFSX_UNSIGNED));
1451         if (*tl++ == rpc_msgdenied) {
1452                 if (*tl == rpc_mismatch) {
1453                         error = EOPNOTSUPP;
1454                 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1455                            *tl++ == rpc_autherr) {
1456                         if (req->r_failed_auth == 0) {
1457                                 req->r_failed_auth++;
1458                                 req->r_mheadend->m_next = NULL;
1459                                 m_freem(info->mrep);
1460                                 info->mrep = NULL;
1461                                 m_freem(req->r_mreq);
1462                                 req->r_mreq = NULL;
1463                                 return (ENEEDAUTH);
1464                         } else {
1465                                 error = EAUTH;
1466                         }
1467                 } else {
1468                         error = EACCES;
1469                 }
1470                 m_freem(info->mrep);
1471                 info->mrep = NULL;
1472                 m_freem(req->r_mreq);
1473                 req->r_mreq = NULL;
1474                 kfree(req, M_NFSREQ);
1475                 info->req = NULL;
1476                 return (error);
1477         }
1478
1479         /*
1480          * Grab any Kerberos verifier, otherwise just throw it away.
1481          */
1482         verf_type = fxdr_unsigned(int, *tl++);
1483         i = fxdr_unsigned(int32_t, *tl);
1484         if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1485                 error = nfs_savenickauth(nmp, req->r_cred, i, req->r_key,
1486                                          &info->md, &info->dpos, info->mrep);
1487                 if (error)
1488                         goto nfsmout;
1489         } else if (i > 0) {
1490                 ERROROUT(nfsm_adv(info, nfsm_rndup(i)));
1491         }
1492         NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1493         /* 0 == ok */
1494         if (*tl == 0) {
1495                 NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1496                 if (*tl != 0) {
1497                         error = fxdr_unsigned(int, *tl);
1498
1499                         /*
1500                          * Does anyone even implement this?  Just impose
1501                          * a 1-second delay.
1502                          */
1503                         if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1504                                 error == NFSERR_TRYLATER) {
1505                                 m_freem(info->mrep);
1506                                 info->mrep = NULL;
1507                                 error = 0;
1508
1509                                 tsleep((caddr_t)&lbolt, 0, "nqnfstry", 0);
1510                                 return (EAGAIN);        /* goto tryagain */
1511                         }
1512
1513                         /*
1514                          * If the File Handle was stale, invalidate the
1515                          * lookup cache, just in case.
1516                          *
1517                          * To avoid namecache<->vnode deadlocks we must
1518                          * release the vnode lock if we hold it.
1519                          */
1520                         if (error == ESTALE) {
1521                                 struct vnode *vp = req->r_vp;
1522                                 int ltype;
1523
1524                                 ltype = lockstatus(&vp->v_lock, curthread);
1525                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1526                                         lockmgr(&vp->v_lock, LK_RELEASE);
1527                                 cache_inval_vp(vp, CINV_CHILDREN);
1528                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1529                                         lockmgr(&vp->v_lock, ltype);
1530                         }
1531                         if (nmp->nm_flag & NFSMNT_NFSV3) {
1532                                 KKASSERT(*req->r_mrp == info->mrep);
1533                                 KKASSERT(*req->r_mdp == info->md);
1534                                 KKASSERT(*req->r_dposp == info->dpos);
1535                                 error |= NFSERR_RETERR;
1536                         } else {
1537                                 m_freem(info->mrep);
1538                                 info->mrep = NULL;
1539                         }
1540                         m_freem(req->r_mreq);
1541                         req->r_mreq = NULL;
1542                         kfree(req, M_NFSREQ);
1543                         info->req = NULL;
1544                         return (error);
1545                 }
1546
1547                 KKASSERT(*req->r_mrp == info->mrep);
1548                 KKASSERT(*req->r_mdp == info->md);
1549                 KKASSERT(*req->r_dposp == info->dpos);
1550                 m_freem(req->r_mreq);
1551                 req->r_mreq = NULL;
1552                 kfree(req, M_NFSREQ);
1553                 return (0);
1554         }
1555         m_freem(info->mrep);
1556         info->mrep = NULL;
1557         error = EPROTONOSUPPORT;
1558 nfsmout:
1559         m_freem(req->r_mreq);
1560         req->r_mreq = NULL;
1561         kfree(req, M_NFSREQ);
1562         info->req = NULL;
1563         return (error);
1564 }
1565
1566 #ifndef NFS_NOSERVER
1567 /*
1568  * Generate the rpc reply header
1569  * siz arg. is used to decide if adding a cluster is worthwhile
1570  */
1571 int
1572 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1573             int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1574 {
1575         u_int32_t *tl;
1576         struct nfsm_info info;
1577
1578         siz += RPC_REPLYSIZ;
1579         info.mb = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1580         info.mreq = info.mb;
1581         info.mreq->m_pkthdr.len = 0;
1582         /*
1583          * If this is not a cluster, try and leave leading space
1584          * for the lower level headers.
1585          */
1586         if ((max_hdr + siz) < MINCLSIZE)
1587                 info.mreq->m_data += max_hdr;
1588         tl = mtod(info.mreq, u_int32_t *);
1589         info.mreq->m_len = 6 * NFSX_UNSIGNED;
1590         info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1591         *tl++ = txdr_unsigned(nd->nd_retxid);
1592         *tl++ = rpc_reply;
1593         if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1594                 *tl++ = rpc_msgdenied;
1595                 if (err & NFSERR_AUTHERR) {
1596                         *tl++ = rpc_autherr;
1597                         *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1598                         info.mreq->m_len -= NFSX_UNSIGNED;
1599                         info.bpos -= NFSX_UNSIGNED;
1600                 } else {
1601                         *tl++ = rpc_mismatch;
1602                         *tl++ = txdr_unsigned(RPC_VER2);
1603                         *tl = txdr_unsigned(RPC_VER2);
1604                 }
1605         } else {
1606                 *tl++ = rpc_msgaccepted;
1607
1608                 /*
1609                  * For Kerberos authentication, we must send the nickname
1610                  * verifier back, otherwise just RPCAUTH_NULL.
1611                  */
1612                 if (nd->nd_flag & ND_KERBFULL) {
1613                     struct nfsuid *nuidp;
1614                     struct timeval ktvin, ktvout;
1615
1616                     for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1617                         nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1618                         if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1619                             (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1620                              &nuidp->nu_haddr, nd->nd_nam2)))
1621                             break;
1622                     }
1623                     if (nuidp) {
1624                         ktvin.tv_sec =
1625                             txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1626                         ktvin.tv_usec =
1627                             txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1628
1629                         /*
1630                          * Encrypt the timestamp in ecb mode using the
1631                          * session key.
1632                          */
1633 #ifdef NFSKERB
1634                         XXX
1635 #else
1636                         ktvout.tv_sec = 0;
1637                         ktvout.tv_usec = 0;
1638 #endif
1639
1640                         *tl++ = rpc_auth_kerb;
1641                         *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1642                         *tl = ktvout.tv_sec;
1643                         tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1644                         *tl++ = ktvout.tv_usec;
1645                         *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1646                     } else {
1647                         *tl++ = 0;
1648                         *tl++ = 0;
1649                     }
1650                 } else {
1651                         *tl++ = 0;
1652                         *tl++ = 0;
1653                 }
1654                 switch (err) {
1655                 case EPROGUNAVAIL:
1656                         *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1657                         break;
1658                 case EPROGMISMATCH:
1659                         *tl = txdr_unsigned(RPC_PROGMISMATCH);
1660                         tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1661                         *tl++ = txdr_unsigned(2);
1662                         *tl = txdr_unsigned(3);
1663                         break;
1664                 case EPROCUNAVAIL:
1665                         *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1666                         break;
1667                 case EBADRPC:
1668                         *tl = txdr_unsigned(RPC_GARBAGE);
1669                         break;
1670                 default:
1671                         *tl = 0;
1672                         if (err != NFSERR_RETVOID) {
1673                                 tl = nfsm_build(&info, NFSX_UNSIGNED);
1674                                 if (err)
1675                                     *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1676                                 else
1677                                     *tl = 0;
1678                         }
1679                         break;
1680                 };
1681         }
1682
1683         if (mrq != NULL)
1684             *mrq = info.mreq;
1685         *mbp = info.mb;
1686         *bposp = info.bpos;
1687         if (err != 0 && err != NFSERR_RETVOID)
1688                 nfsstats.srvrpc_errs++;
1689         return (0);
1690 }
1691
1692
1693 #endif /* NFS_NOSERVER */
1694
1695 /*
1696  * Nfs timer routine.
1697  *
1698  * Scan the nfsreq list and retranmit any requests that have timed out
1699  * To avoid retransmission attempts on STREAM sockets (in the future) make
1700  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1701  *
1702  * Requests with attached responses, terminated requests, and
1703  * locked requests are ignored.  Locked requests will be picked up
1704  * in a later timer call.
1705  */
1706 void
1707 nfs_timer_callout(void *arg /* never used */)
1708 {
1709         struct nfsmount *nmp;
1710         struct nfsreq *req;
1711 #ifndef NFS_NOSERVER
1712         struct nfssvc_sock *slp;
1713         u_quad_t cur_usec;
1714 #endif /* NFS_NOSERVER */
1715
1716         lwkt_gettoken(&nfs_token);
1717         TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1718                 lwkt_gettoken(&nmp->nm_token);
1719                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1720                         KKASSERT(nmp == req->r_nmp);
1721                         if (req->r_mrep)
1722                                 continue;
1723                         if (req->r_flags & (R_SOFTTERM | R_LOCKED))
1724                                 continue;
1725
1726                         /*
1727                          * Handle timeout/retry.  Be sure to process r_mrep
1728                          * for async requests that completed while we had
1729                          * the request locked or they will hang in the reqq
1730                          * forever.
1731                          */
1732                         req->r_flags |= R_LOCKED;
1733                         if (nfs_sigintr(nmp, req, req->r_td)) {
1734                                 nfs_softterm(req, 1);
1735                                 req->r_flags &= ~R_LOCKED;
1736                         } else {
1737                                 nfs_timer_req(req);
1738                                 if (req->r_flags & R_ASYNC) {
1739                                         if (req->r_mrep)
1740                                                 nfs_hardterm(req, 1);
1741                                         req->r_flags &= ~R_LOCKED;
1742                                         nfssvc_iod_reader_wakeup(nmp);
1743                                 } else {
1744                                         req->r_flags &= ~R_LOCKED;
1745                                 }
1746                         }
1747                         if (req->r_flags & R_WANTED) {
1748                                 req->r_flags &= ~R_WANTED;
1749                                 wakeup(req);
1750                         }
1751                 }
1752                 lwkt_reltoken(&nmp->nm_token);
1753         }
1754 #ifndef NFS_NOSERVER
1755
1756         /*
1757          * Scan the write gathering queues for writes that need to be
1758          * completed now.
1759          */
1760         cur_usec = nfs_curusec();
1761
1762         TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1763                 /* XXX race against removal */
1764                 if (lwkt_trytoken(&slp->ns_token)) {
1765                         if (slp->ns_tq.lh_first &&
1766                             (slp->ns_tq.lh_first->nd_time <= cur_usec)) {
1767                                 nfsrv_wakenfsd(slp, 1);
1768                         }
1769                         lwkt_reltoken(&slp->ns_token);
1770                 }
1771         }
1772 #endif /* NFS_NOSERVER */
1773
1774         callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer_callout, NULL);
1775         lwkt_reltoken(&nfs_token);
1776 }
1777
1778 static
1779 void
1780 nfs_timer_req(struct nfsreq *req)
1781 {
1782         struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1783         struct nfsmount *nmp = req->r_nmp;
1784         struct mbuf *m;
1785         struct socket *so;
1786         int timeo;
1787         int error;
1788
1789         /*
1790          * rtt ticks and timeout calculation.  Return if the timeout
1791          * has not been reached yet, unless the packet is flagged
1792          * for an immediate send.
1793          *
1794          * The mean rtt doesn't help when we get random I/Os, we have
1795          * to multiply by fairly large numbers.
1796          */
1797         if (req->r_rtt >= 0) {
1798                 /*
1799                  * Calculate the timeout to test against.
1800                  */
1801                 req->r_rtt++;
1802                 if (nmp->nm_flag & NFSMNT_DUMBTIMR) {
1803                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1804                 } else if (req->r_flags & R_TIMING) {
1805                         timeo = NFS_SRTT(req) + NFS_SDRTT(req);
1806                 } else {
1807                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1808                 }
1809                 timeo *= multt[req->r_procnum];
1810                 /* timeo is still scaled by SCALE_BITS */
1811
1812 #define NFSFS   (NFS_RTT_SCALE * NFS_HZ)
1813                 if (req->r_flags & R_TIMING) {
1814                         static long last_time;
1815                         if (nfs_showrtt && last_time != time_second) {
1816                                 kprintf("rpccmd %d NFS SRTT %d SDRTT %d "
1817                                         "timeo %d.%03d\n",
1818                                         proct[req->r_procnum],
1819                                         NFS_SRTT(req), NFS_SDRTT(req),
1820                                         timeo / NFSFS,
1821                                         timeo % NFSFS * 1000 /  NFSFS);
1822                                 last_time = time_second;
1823                         }
1824                 }
1825 #undef NFSFS
1826
1827                 /*
1828                  * deal with nfs_timer jitter.
1829                  */
1830                 timeo = (timeo >> NFS_RTT_SCALE_BITS) + 1;
1831                 if (timeo < 2)
1832                         timeo = 2;
1833
1834                 if (nmp->nm_timeouts > 0)
1835                         timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1836                 if (timeo > NFS_MAXTIMEO)
1837                         timeo = NFS_MAXTIMEO;
1838                 if (req->r_rtt <= timeo) {
1839                         if ((req->r_flags & R_NEEDSXMIT) == 0)
1840                                 return;
1841                 } else if (nmp->nm_timeouts < 8) {
1842                         nmp->nm_timeouts++;
1843                 }
1844         }
1845
1846         /*
1847          * Check for server not responding
1848          */
1849         if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1850              req->r_rexmit > nmp->nm_deadthresh) {
1851                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1852                         "not responding");
1853                 req->r_flags |= R_TPRINTFMSG;
1854         }
1855         if (req->r_rexmit >= req->r_retry) {    /* too many */
1856                 nfsstats.rpctimeouts++;
1857                 nfs_softterm(req, 1);
1858                 return;
1859         }
1860
1861         /*
1862          * Generally disable retransmission on reliable sockets,
1863          * unless the request is flagged for immediate send.
1864          */
1865         if (nmp->nm_sotype != SOCK_DGRAM) {
1866                 if (++req->r_rexmit > NFS_MAXREXMIT)
1867                         req->r_rexmit = NFS_MAXREXMIT;
1868                 if ((req->r_flags & R_NEEDSXMIT) == 0)
1869                         return;
1870         }
1871
1872         /*
1873          * Stop here if we do not have a socket!
1874          */
1875         if ((so = nmp->nm_so) == NULL)
1876                 return;
1877
1878         /*
1879          * If there is enough space and the window allows.. resend it.
1880          *
1881          * r_rtt is left intact in case we get an answer after the
1882          * retry that was a reply to the original packet.
1883          *
1884          * NOTE: so_pru_send()
1885          */
1886         if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1887             (req->r_flags & (R_SENT | R_NEEDSXMIT)) &&
1888            (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1889                 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1890                     error = so_pru_send(so, 0, m, NULL, NULL, td);
1891                 else
1892                     error = so_pru_send(so, 0, m, nmp->nm_nam, NULL, td);
1893                 if (error) {
1894                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1895                                 so->so_error = 0;
1896                         req->r_flags |= R_NEEDSXMIT;
1897                 } else if (req->r_mrep == NULL) {
1898                         /*
1899                          * Iff first send, start timing
1900                          * else turn timing off, backoff timer
1901                          * and divide congestion window by 2.
1902                          *
1903                          * It is possible for the so_pru_send() to
1904                          * block and for us to race a reply so we
1905                          * only do this if the reply field has not
1906                          * been filled in.  R_LOCKED will prevent
1907                          * the request from being ripped out from under
1908                          * us entirely.
1909                          *
1910                          * Record the last resent procnum to aid us
1911                          * in duplicate detection on receive.
1912                          */
1913                         if ((req->r_flags & R_NEEDSXMIT) == 0) {
1914                                 if (nfs_showrexmit)
1915                                         kprintf("X");
1916                                 if (++req->r_rexmit > NFS_MAXREXMIT)
1917                                         req->r_rexmit = NFS_MAXREXMIT;
1918                                 nmp->nm_maxasync_scaled >>= 1;
1919                                 if (nmp->nm_maxasync_scaled < NFS_MINASYNC_SCALED)
1920                                         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
1921                                 nfsstats.rpcretries++;
1922                                 nmp->nm_lastreprocnum = req->r_procnum;
1923                         } else {
1924                                 req->r_flags |= R_SENT;
1925                                 req->r_flags &= ~R_NEEDSXMIT;
1926                         }
1927                 }
1928         }
1929 }
1930
1931 /*
1932  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1933  * wait for all requests to complete. This is used by forced unmounts
1934  * to terminate any outstanding RPCs.
1935  *
1936  * Locked requests cannot be canceled but will be marked for
1937  * soft-termination.
1938  */
1939 int
1940 nfs_nmcancelreqs(struct nfsmount *nmp)
1941 {
1942         struct nfsreq *req;
1943         int i;
1944
1945         crit_enter();
1946         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1947                 if (req->r_mrep != NULL || (req->r_flags & R_SOFTTERM))
1948                         continue;
1949                 nfs_softterm(req, 0);
1950         }
1951         /* XXX  the other two queues as well */
1952         crit_exit();
1953
1954         for (i = 0; i < 30; i++) {
1955                 crit_enter();
1956                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1957                         if (nmp == req->r_nmp)
1958                                 break;
1959                 }
1960                 crit_exit();
1961                 if (req == NULL)
1962                         return (0);
1963                 tsleep(&lbolt, 0, "nfscancel", 0);
1964         }
1965         return (EBUSY);
1966 }
1967
1968 /*
1969  * Soft-terminate a request, effectively marking it as failed.
1970  *
1971  * Must be called from within a critical section.
1972  */
1973 static void
1974 nfs_softterm(struct nfsreq *rep, int islocked)
1975 {
1976         rep->r_flags |= R_SOFTTERM;
1977         nfs_hardterm(rep, islocked);
1978 }
1979
1980 /*
1981  * Hard-terminate a request, typically after getting a response.
1982  *
1983  * The state machine can still decide to re-issue it later if necessary.
1984  *
1985  * Must be called from within a critical section.
1986  */
1987 static void
1988 nfs_hardterm(struct nfsreq *rep, int islocked)
1989 {
1990         struct nfsmount *nmp = rep->r_nmp;
1991
1992         /*
1993          * The nm_send count is decremented now to avoid deadlocks
1994          * when the process in soreceive() hasn't yet managed to send
1995          * its own request.
1996          */
1997         if (rep->r_flags & R_SENT) {
1998                 rep->r_flags &= ~R_SENT;
1999         }
2000
2001         /*
2002          * If we locked the request or nobody else has locked the request,
2003          * and the request is async, we can move it to the reader thread's
2004          * queue now and fix up the state.
2005          *
2006          * If we locked the request or nobody else has locked the request,
2007          * we can wake up anyone blocked waiting for a response on the
2008          * request.
2009          */
2010         if (islocked || (rep->r_flags & R_LOCKED) == 0) {
2011                 if ((rep->r_flags & (R_ONREQQ | R_ASYNC)) ==
2012                     (R_ONREQQ | R_ASYNC)) {
2013                         rep->r_flags &= ~R_ONREQQ;
2014                         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
2015                         --nmp->nm_reqqlen;
2016                         TAILQ_INSERT_TAIL(&nmp->nm_reqrxq, rep, r_chain);
2017                         KKASSERT(rep->r_info->state == NFSM_STATE_TRY ||
2018                                  rep->r_info->state == NFSM_STATE_WAITREPLY);
2019                         rep->r_info->state = NFSM_STATE_PROCESSREPLY;
2020                         nfssvc_iod_reader_wakeup(nmp);
2021                         if (TAILQ_FIRST(&nmp->nm_bioq) &&
2022                             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
2023                                 nfssvc_iod_writer_wakeup(nmp);
2024                         }
2025                 }
2026                 mtx_abort_ex_link(&nmp->nm_rxlock, &rep->r_link);
2027         }
2028 }
2029
2030 /*
2031  * Test for a termination condition pending on the process.
2032  * This is used for NFSMNT_INT mounts.
2033  */
2034 int
2035 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
2036 {
2037         sigset_t tmpset;
2038         struct proc *p;
2039         struct lwp *lp;
2040
2041         if (rep && (rep->r_flags & R_SOFTTERM))
2042                 return (EINTR);
2043         /* Terminate all requests while attempting a forced unmount. */
2044         if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
2045                 return (EINTR);
2046         if (!(nmp->nm_flag & NFSMNT_INT))
2047                 return (0);
2048         /* td might be NULL YYY */
2049         if (td == NULL || (p = td->td_proc) == NULL)
2050                 return (0);
2051
2052         lp = td->td_lwp;
2053         tmpset = lwp_sigpend(lp);
2054         SIGSETNAND(tmpset, lp->lwp_sigmask);
2055         SIGSETNAND(tmpset, p->p_sigignore);
2056         if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
2057                 return (EINTR);
2058
2059         return (0);
2060 }
2061
2062 /*
2063  * Lock a socket against others.
2064  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
2065  * and also to avoid race conditions between the processes with nfs requests
2066  * in progress when a reconnect is necessary.
2067  */
2068 int
2069 nfs_sndlock(struct nfsmount *nmp, struct nfsreq *rep)
2070 {
2071         mtx_t mtx = &nmp->nm_txlock;
2072         struct thread *td;
2073         int slptimeo;
2074         int slpflag;
2075         int error;
2076
2077         slpflag = 0;
2078         slptimeo = 0;
2079         td = rep ? rep->r_td : NULL;
2080         if (nmp->nm_flag & NFSMNT_INT)
2081                 slpflag = PCATCH;
2082
2083         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2084                 if (nfs_sigintr(nmp, rep, td)) {
2085                         error = EINTR;
2086                         break;
2087                 }
2088                 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
2089                 if (error == 0)
2090                         break;
2091                 if (slpflag == PCATCH) {
2092                         slpflag = 0;
2093                         slptimeo = 2 * hz;
2094                 }
2095         }
2096         /* Always fail if our request has been cancelled. */
2097         if (rep && (rep->r_flags & R_SOFTTERM)) {
2098                 if (error == 0)
2099                         mtx_unlock(mtx);
2100                 error = EINTR;
2101         }
2102         return (error);
2103 }
2104
2105 /*
2106  * Unlock the stream socket for others.
2107  */
2108 void
2109 nfs_sndunlock(struct nfsmount *nmp)
2110 {
2111         mtx_unlock(&nmp->nm_txlock);
2112 }
2113
2114 /*
2115  * Lock the receiver side of the socket.
2116  *
2117  * rep may be NULL.
2118  */
2119 static int
2120 nfs_rcvlock(struct nfsmount *nmp, struct nfsreq *rep)
2121 {
2122         mtx_t mtx = &nmp->nm_rxlock;
2123         int slpflag;
2124         int slptimeo;
2125         int error;
2126
2127         /*
2128          * Unconditionally check for completion in case another nfsiod
2129          * get the packet while the caller was blocked, before the caller
2130          * called us.  Packet reception is handled by mainline code which
2131          * is protected by the BGL at the moment.
2132          *
2133          * We do not strictly need the second check just before the
2134          * tsleep(), but it's good defensive programming.
2135          */
2136         if (rep && rep->r_mrep != NULL)
2137                 return (EALREADY);
2138
2139         if (nmp->nm_flag & NFSMNT_INT)
2140                 slpflag = PCATCH;
2141         else
2142                 slpflag = 0;
2143         slptimeo = 0;
2144
2145         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2146                 if (nfs_sigintr(nmp, rep, (rep ? rep->r_td : NULL))) {
2147                         error = EINTR;
2148                         break;
2149                 }
2150                 if (rep && rep->r_mrep != NULL) {
2151                         error = EALREADY;
2152                         break;
2153                 }
2154
2155                 /*
2156                  * NOTE: can return ENOLCK, but in that case rep->r_mrep
2157                  *       will already be set.
2158                  */
2159                 if (rep) {
2160                         error = mtx_lock_ex_link(mtx, &rep->r_link,
2161                                                  "nfsrcvlk",
2162                                                  slpflag, slptimeo);
2163                 } else {
2164                         error = mtx_lock_ex(mtx, "nfsrcvlk", slpflag, slptimeo);
2165                 }
2166                 if (error == 0)
2167                         break;
2168
2169                 /*
2170                  * If our reply was recieved while we were sleeping,
2171                  * then just return without taking the lock to avoid a
2172                  * situation where a single iod could 'capture' the
2173                  * recieve lock.
2174                  */
2175                 if (rep && rep->r_mrep != NULL) {
2176                         error = EALREADY;
2177                         break;
2178                 }
2179                 if (slpflag == PCATCH) {
2180                         slpflag = 0;
2181                         slptimeo = 2 * hz;
2182                 }
2183         }
2184         if (error == 0) {
2185                 if (rep && rep->r_mrep != NULL) {
2186                         error = EALREADY;
2187                         mtx_unlock(mtx);
2188                 }
2189         }
2190         return (error);
2191 }
2192
2193 /*
2194  * Unlock the stream socket for others.
2195  */
2196 static void
2197 nfs_rcvunlock(struct nfsmount *nmp)
2198 {
2199         mtx_unlock(&nmp->nm_rxlock);
2200 }
2201
2202 /*
2203  * nfs_realign:
2204  *
2205  * Check for badly aligned mbuf data and realign by copying the unaligned
2206  * portion of the data into a new mbuf chain and freeing the portions
2207  * of the old chain that were replaced.
2208  *
2209  * We cannot simply realign the data within the existing mbuf chain
2210  * because the underlying buffers may contain other rpc commands and
2211  * we cannot afford to overwrite them.
2212  *
2213  * We would prefer to avoid this situation entirely.  The situation does
2214  * not occur with NFS/UDP and is supposed to only occassionally occur
2215  * with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
2216  *
2217  * NOTE!  MB_DONTWAIT cannot be used here.  The mbufs must be acquired
2218  *        because the rpc request OR reply cannot be thrown away.  TCP NFS
2219  *        mounts do not retry their RPCs unless the TCP connection itself
2220  *        is dropped so throwing away a RPC will basically cause the NFS
2221  *        operation to lockup indefinitely.
2222  */
2223 static void
2224 nfs_realign(struct mbuf **pm, int hsiz)
2225 {
2226         struct mbuf *m;
2227         struct mbuf *n = NULL;
2228
2229         /*
2230          * Check for misalignemnt
2231          */
2232         ++nfs_realign_test;
2233         while ((m = *pm) != NULL) {
2234                 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3))
2235                         break;
2236                 pm = &m->m_next;
2237         }
2238
2239         /*
2240          * If misalignment found make a completely new copy.
2241          */
2242         if (m) {
2243                 ++nfs_realign_count;
2244                 n = m_dup_data(m, MB_WAIT);
2245                 m_freem(*pm);
2246                 *pm = n;
2247         }
2248 }
2249
2250 #ifndef NFS_NOSERVER
2251
2252 /*
2253  * Parse an RPC request
2254  * - verify it
2255  * - fill in the cred struct.
2256  */
2257 int
2258 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
2259 {
2260         int len, i;
2261         u_int32_t *tl;
2262         struct uio uio;
2263         struct iovec iov;
2264         caddr_t cp;
2265         u_int32_t nfsvers, auth_type;
2266         uid_t nickuid;
2267         int error = 0, ticklen;
2268         struct nfsuid *nuidp;
2269         struct timeval tvin, tvout;
2270         struct nfsm_info info;
2271 #if 0                           /* until encrypted keys are implemented */
2272         NFSKERBKEYSCHED_T keys; /* stores key schedule */
2273 #endif
2274
2275         info.mrep = nd->nd_mrep;
2276         info.md = nd->nd_md;
2277         info.dpos = nd->nd_dpos;
2278
2279         if (has_header) {
2280                 NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
2281                 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
2282                 if (*tl++ != rpc_call) {
2283                         m_freem(info.mrep);
2284                         return (EBADRPC);
2285                 }
2286         } else {
2287                 NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
2288         }
2289         nd->nd_repstat = 0;
2290         nd->nd_flag = 0;
2291         if (*tl++ != rpc_vers) {
2292                 nd->nd_repstat = ERPCMISMATCH;
2293                 nd->nd_procnum = NFSPROC_NOOP;
2294                 return (0);
2295         }
2296         if (*tl != nfs_prog) {
2297                 nd->nd_repstat = EPROGUNAVAIL;
2298                 nd->nd_procnum = NFSPROC_NOOP;
2299                 return (0);
2300         }
2301         tl++;
2302         nfsvers = fxdr_unsigned(u_int32_t, *tl++);
2303         if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
2304                 nd->nd_repstat = EPROGMISMATCH;
2305                 nd->nd_procnum = NFSPROC_NOOP;
2306                 return (0);
2307         }
2308         if (nfsvers == NFS_VER3)
2309                 nd->nd_flag = ND_NFSV3;
2310         nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
2311         if (nd->nd_procnum == NFSPROC_NULL)
2312                 return (0);
2313         if (nd->nd_procnum >= NFS_NPROCS ||
2314                 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
2315                 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
2316                 nd->nd_repstat = EPROCUNAVAIL;
2317                 nd->nd_procnum = NFSPROC_NOOP;
2318                 return (0);
2319         }
2320         if ((nd->nd_flag & ND_NFSV3) == 0)
2321                 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
2322         auth_type = *tl++;
2323         len = fxdr_unsigned(int, *tl++);
2324         if (len < 0 || len > RPCAUTH_MAXSIZ) {
2325                 m_freem(info.mrep);
2326                 return (EBADRPC);
2327         }
2328
2329         nd->nd_flag &= ~ND_KERBAUTH;
2330         /*
2331          * Handle auth_unix or auth_kerb.
2332          */
2333         if (auth_type == rpc_auth_unix) {
2334                 len = fxdr_unsigned(int, *++tl);
2335                 if (len < 0 || len > NFS_MAXNAMLEN) {
2336                         m_freem(info.mrep);
2337                         return (EBADRPC);
2338                 }
2339                 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2340                 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2341                 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
2342                 nd->nd_cr.cr_ref = 1;
2343                 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
2344                 nd->nd_cr.cr_ruid = nd->nd_cr.cr_svuid = nd->nd_cr.cr_uid;
2345                 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
2346                 nd->nd_cr.cr_rgid = nd->nd_cr.cr_svgid = nd->nd_cr.cr_gid;
2347                 len = fxdr_unsigned(int, *tl);
2348                 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
2349                         m_freem(info.mrep);
2350                         return (EBADRPC);
2351                 }
2352                 NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
2353                 for (i = 1; i <= len; i++)
2354                     if (i < NGROUPS)
2355                         nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
2356                     else
2357                         tl++;
2358                 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
2359                 if (nd->nd_cr.cr_ngroups > 1)
2360                     nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
2361                 len = fxdr_unsigned(int, *++tl);
2362                 if (len < 0 || len > RPCAUTH_MAXSIZ) {
2363                         m_freem(info.mrep);
2364                         return (EBADRPC);
2365                 }
2366                 if (len > 0) {
2367                         ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2368                 }
2369         } else if (auth_type == rpc_auth_kerb) {
2370                 switch (fxdr_unsigned(int, *tl++)) {
2371                 case RPCAKN_FULLNAME:
2372                         ticklen = fxdr_unsigned(int, *tl);
2373                         *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2374                         uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2375                         nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2376                         if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2377                                 m_freem(info.mrep);
2378                                 return (EBADRPC);
2379                         }
2380                         uio.uio_offset = 0;
2381                         uio.uio_iov = &iov;
2382                         uio.uio_iovcnt = 1;
2383                         uio.uio_segflg = UIO_SYSSPACE;
2384                         iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2385                         iov.iov_len = RPCAUTH_MAXSIZ - 4;
2386                         ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2387                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2388                         if (*tl++ != rpc_auth_kerb ||
2389                                 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2390                                 kprintf("Bad kerb verifier\n");
2391                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2392                                 nd->nd_procnum = NFSPROC_NOOP;
2393                                 return (0);
2394                         }
2395                         NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2396                         tl = (u_int32_t *)cp;
2397                         if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2398                                 kprintf("Not fullname kerb verifier\n");
2399                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2400                                 nd->nd_procnum = NFSPROC_NOOP;
2401                                 return (0);
2402                         }
2403                         cp += NFSX_UNSIGNED;
2404                         bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2405                         nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2406                         nd->nd_flag |= ND_KERBFULL;
2407                         nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2408                         break;
2409                 case RPCAKN_NICKNAME:
2410                         if (len != 2 * NFSX_UNSIGNED) {
2411                                 kprintf("Kerb nickname short\n");
2412                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2413                                 nd->nd_procnum = NFSPROC_NOOP;
2414                                 return (0);
2415                         }
2416                         nickuid = fxdr_unsigned(uid_t, *tl);
2417                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2418                         if (*tl++ != rpc_auth_kerb ||
2419                                 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2420                                 kprintf("Kerb nick verifier bad\n");
2421                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2422                                 nd->nd_procnum = NFSPROC_NOOP;
2423                                 return (0);
2424                         }
2425                         NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2426                         tvin.tv_sec = *tl++;
2427                         tvin.tv_usec = *tl;
2428
2429                         for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2430                             nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2431                                 if (nuidp->nu_cr.cr_uid == nickuid &&
2432                                     (!nd->nd_nam2 ||
2433                                      netaddr_match(NU_NETFAM(nuidp),
2434                                       &nuidp->nu_haddr, nd->nd_nam2)))
2435                                         break;
2436                         }
2437                         if (!nuidp) {
2438                                 nd->nd_repstat =
2439                                         (NFSERR_AUTHERR|AUTH_REJECTCRED);
2440                                 nd->nd_procnum = NFSPROC_NOOP;
2441                                 return (0);
2442                         }
2443
2444                         /*
2445                          * Now, decrypt the timestamp using the session key
2446                          * and validate it.
2447                          */
2448 #ifdef NFSKERB
2449                         XXX
2450 #else
2451                         tvout.tv_sec = 0;
2452                         tvout.tv_usec = 0;
2453 #endif
2454
2455                         tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2456                         tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2457                         if (nuidp->nu_expire < time_second ||
2458                             nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2459                             (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2460                              nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2461                                 nuidp->nu_expire = 0;
2462                                 nd->nd_repstat =
2463                                     (NFSERR_AUTHERR|AUTH_REJECTVERF);
2464                                 nd->nd_procnum = NFSPROC_NOOP;
2465                                 return (0);
2466                         }
2467                         nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2468                         nd->nd_flag |= ND_KERBNICK;
2469                 };
2470         } else {
2471                 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2472                 nd->nd_procnum = NFSPROC_NOOP;
2473                 return (0);
2474         }
2475
2476         nd->nd_md = info.md;
2477         nd->nd_dpos = info.dpos;
2478         return (0);
2479 nfsmout:
2480         return (error);
2481 }
2482
2483 #endif
2484
2485 /*
2486  * Send a message to the originating process's terminal.  The thread and/or
2487  * process may be NULL.  YYY the thread should not be NULL but there may
2488  * still be some uio_td's that are still being passed as NULL through to
2489  * nfsm_request().
2490  */
2491 static int
2492 nfs_msg(struct thread *td, char *server, char *msg)
2493 {
2494         tpr_t tpr;
2495
2496         if (td && td->td_proc)
2497                 tpr = tprintf_open(td->td_proc);
2498         else
2499                 tpr = NULL;
2500         tprintf(tpr, "nfs server %s: %s\n", server, msg);
2501         tprintf_close(tpr);
2502         return (0);
2503 }
2504
2505 #ifndef NFS_NOSERVER
2506
2507 /*
2508  * Socket upcall routine for nfsd sockets.  This runs in the protocol
2509  * thread and passes waitflag == MB_DONTWAIT.
2510  */
2511 void
2512 nfsrv_rcv_upcall(struct socket *so, void *arg, int waitflag)
2513 {
2514         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2515
2516         if (slp->ns_needq_upcall == 0) {
2517                 slp->ns_needq_upcall = 1;       /* ok to race */
2518                 lwkt_gettoken(&nfs_token);
2519                 nfsrv_wakenfsd(slp, 1);
2520                 lwkt_reltoken(&nfs_token);
2521         }
2522 #if 0
2523         lwkt_gettoken(&slp->ns_token);
2524         slp->ns_flag |= SLP_NEEDQ;
2525         nfsrv_rcv(so, arg, waitflag);
2526         lwkt_reltoken(&slp->ns_token);
2527 #endif
2528 }
2529
2530 /*
2531  * Process new data on a receive socket.  Essentially do as much as we can
2532  * non-blocking, else punt and it will be called with MB_WAIT from an nfsd.
2533  *
2534  * slp->ns_token is held on call
2535  */
2536 void
2537 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2538 {
2539         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2540         struct mbuf *m;
2541         struct sockaddr *nam;
2542         struct sockbuf sio;
2543         int flags, error;
2544         int nparallel_wakeup = 0;
2545
2546         ASSERT_LWKT_TOKEN_HELD(&slp->ns_token);
2547
2548         if ((slp->ns_flag & SLP_VALID) == 0)
2549                 return;
2550
2551         /*
2552          * Do not allow an infinite number of completed RPC records to build 
2553          * up before we stop reading data from the socket.  Otherwise we could
2554          * end up holding onto an unreasonable number of mbufs for requests
2555          * waiting for service.
2556          *
2557          * This should give pretty good feedback to the TCP layer and
2558          * prevents a memory crunch for other protocols.
2559          *
2560          * Note that the same service socket can be dispatched to several
2561          * nfs servers simultaniously.  The tcp protocol callback calls us
2562          * with MB_DONTWAIT.  nfsd calls us with MB_WAIT (typically).
2563          */
2564         if (NFSRV_RECLIMIT(slp))
2565                 return;
2566
2567         /*
2568          * Handle protocol specifics to parse an RPC request.  We always
2569          * pull from the socket using non-blocking I/O.
2570          */
2571         if (so->so_type == SOCK_STREAM) {
2572                 /*
2573                  * The data has to be read in an orderly fashion from a TCP
2574                  * stream, unlike a UDP socket.  It is possible for soreceive
2575                  * and/or nfsrv_getstream() to block, so make sure only one
2576                  * entity is messing around with the TCP stream at any given
2577                  * moment.  The receive sockbuf's lock in soreceive is not
2578                  * sufficient.
2579                  */
2580                 if (slp->ns_flag & SLP_GETSTREAM)
2581                         return;
2582                 slp->ns_flag |= SLP_GETSTREAM;
2583
2584                 /*
2585                  * Do soreceive().  Pull out as much data as possible without
2586                  * blocking.
2587                  */
2588                 sbinit(&sio, 1000000000);
2589                 flags = MSG_DONTWAIT;
2590                 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2591                 if (error || sio.sb_mb == NULL) {
2592                         if (error != EWOULDBLOCK)
2593                                 slp->ns_flag |= SLP_DISCONN;
2594                         slp->ns_flag &= ~(SLP_GETSTREAM | SLP_NEEDQ);
2595                         goto done;
2596                 }
2597                 m = sio.sb_mb;
2598                 if (slp->ns_rawend) {
2599                         slp->ns_rawend->m_next = m;
2600                         slp->ns_cc += sio.sb_cc;
2601                 } else {
2602                         slp->ns_raw = m;
2603                         slp->ns_cc = sio.sb_cc;
2604                 }
2605                 while (m->m_next)
2606                         m = m->m_next;
2607                 slp->ns_rawend = m;
2608
2609                 /*
2610                  * Now try and parse as many record(s) as we can out of the
2611                  * raw stream data.  This will set SLP_DOREC.
2612                  */
2613                 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2614                 if (error && error != EWOULDBLOCK)
2615                         slp->ns_flag |= SLP_DISCONN;
2616                 slp->ns_flag &= ~SLP_GETSTREAM;
2617         } else {
2618                 /*
2619                  * For UDP soreceive typically pulls just one packet, loop
2620                  * to get the whole batch.
2621                  */
2622                 do {
2623                         sbinit(&sio, 1000000000);
2624                         flags = MSG_DONTWAIT;
2625                         error = so_pru_soreceive(so, &nam, NULL, &sio,
2626                                                  NULL, &flags);
2627                         if (sio.sb_mb) {
2628                                 struct nfsrv_rec *rec;
2629                                 int mf = (waitflag & MB_DONTWAIT) ?
2630                                             M_NOWAIT : M_WAITOK;
2631                                 rec = kmalloc(sizeof(struct nfsrv_rec),
2632                                              M_NFSRVDESC, mf);
2633                                 if (!rec) {
2634                                         if (nam)
2635                                                 kfree(nam, M_SONAME);
2636                                         m_freem(sio.sb_mb);
2637                                         continue;
2638                                 }
2639                                 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2640                                 rec->nr_address = nam;
2641                                 rec->nr_packet = sio.sb_mb;
2642                                 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2643                                 ++slp->ns_numrec;
2644                                 slp->ns_flag |= SLP_DOREC;
2645                                 ++nparallel_wakeup;
2646                         } else {
2647                                 slp->ns_flag &= ~SLP_NEEDQ;
2648                         }
2649                         if (error) {
2650                                 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2651                                     && error != EWOULDBLOCK) {
2652                                         slp->ns_flag |= SLP_DISCONN;
2653                                         break;
2654                                 }
2655                         }
2656                         if (NFSRV_RECLIMIT(slp))
2657                                 break;
2658                 } while (sio.sb_mb);
2659         }
2660
2661         /*
2662          * If we were upcalled from the tcp protocol layer and we have
2663          * fully parsed records ready to go, or there is new data pending,
2664          * or something went wrong, try to wake up a nfsd thread to deal
2665          * with it.
2666          */
2667 done:
2668         /* XXX this code is currently not executed (nfsrv_rcv_upcall) */
2669         if (waitflag == MB_DONTWAIT && (slp->ns_flag & SLP_ACTION_MASK)) {
2670                 lwkt_gettoken(&nfs_token);
2671                 nfsrv_wakenfsd(slp, nparallel_wakeup);
2672                 lwkt_reltoken(&nfs_token);
2673         }
2674 }
2675
2676 /*
2677  * Try and extract an RPC request from the mbuf data list received on a
2678  * stream socket. The "waitflag" argument indicates whether or not it
2679  * can sleep.
2680  */
2681 static int
2682 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2683 {
2684         struct mbuf *m, **mpp;
2685         char *cp1, *cp2;
2686         int len;
2687         struct mbuf *om, *m2, *recm;
2688         u_int32_t recmark;
2689
2690         for (;;) {
2691             if (slp->ns_reclen == 0) {
2692                 if (slp->ns_cc < NFSX_UNSIGNED)
2693                         return (0);
2694                 m = slp->ns_raw;
2695                 if (m->m_len >= NFSX_UNSIGNED) {
2696                         bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2697                         m->m_data += NFSX_UNSIGNED;
2698                         m->m_len -= NFSX_UNSIGNED;
2699                 } else {
2700                         cp1 = (caddr_t)&recmark;
2701                         cp2 = mtod(m, caddr_t);
2702                         while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2703                                 while (m->m_len == 0) {
2704                                         m = m->m_next;
2705                                         cp2 = mtod(m, caddr_t);
2706                                 }
2707                                 *cp1++ = *cp2++;
2708                                 m->m_data++;
2709                                 m->m_len--;
2710                         }
2711                 }
2712                 slp->ns_cc -= NFSX_UNSIGNED;
2713                 recmark = ntohl(recmark);
2714                 slp->ns_reclen = recmark & ~0x80000000;
2715                 if (recmark & 0x80000000)
2716                         slp->ns_flag |= SLP_LASTFRAG;
2717                 else
2718                         slp->ns_flag &= ~SLP_LASTFRAG;
2719                 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2720                         log(LOG_ERR, "%s (%d) from nfs client\n",
2721                             "impossible packet length",
2722                             slp->ns_reclen);
2723                         return (EPERM);
2724                 }
2725             }
2726
2727             /*
2728              * Now get the record part.
2729              *
2730              * Note that slp->ns_reclen may be 0.  Linux sometimes
2731              * generates 0-length RPCs
2732              */
2733             recm = NULL;
2734             if (slp->ns_cc == slp->ns_reclen) {
2735                 recm = slp->ns_raw;
2736                 slp->ns_raw = slp->ns_rawend = NULL;
2737                 slp->ns_cc = slp->ns_reclen = 0;
2738             } else if (slp->ns_cc > slp->ns_reclen) {
2739                 len = 0;
2740                 m = slp->ns_raw;
2741                 om = NULL;
2742
2743                 while (len < slp->ns_reclen) {
2744                         if ((len + m->m_len) > slp->ns_reclen) {
2745                                 m2 = m_copym(m, 0, slp->ns_reclen - len,
2746                                         waitflag);
2747                                 if (m2) {
2748                                         if (om) {
2749                                                 om->m_next = m2;
2750                                                 recm = slp->ns_raw;
2751                                         } else
2752                                                 recm = m2;
2753                                         m->m_data += slp->ns_reclen - len;
2754                                         m->m_len -= slp->ns_reclen - len;
2755                                         len = slp->ns_reclen;
2756                                 } else {
2757                                         return (EWOULDBLOCK);
2758                                 }
2759                         } else if ((len + m->m_len) == slp->ns_reclen) {
2760                                 om = m;
2761                                 len += m->m_len;
2762                                 m = m->m_next;
2763                                 recm = slp->ns_raw;
2764                                 om->m_next = NULL;
2765                         } else {
2766                                 om = m;
2767                                 len += m->m_len;
2768                                 m = m->m_next;
2769                         }
2770                 }
2771                 slp->ns_raw = m;
2772                 slp->ns_cc -= len;
2773                 slp->ns_reclen = 0;
2774             } else {
2775                 return (0);
2776             }
2777
2778             /*
2779              * Accumulate the fragments into a record.
2780              */
2781             mpp = &slp->ns_frag;
2782             while (*mpp)
2783                 mpp = &((*mpp)->m_next);
2784             *mpp = recm;
2785             if (slp->ns_flag & SLP_LASTFRAG) {
2786                 struct nfsrv_rec *rec;
2787                 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2788                 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2789                 if (!rec) {
2790                     m_freem(slp->ns_frag);
2791                 } else {
2792                     nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2793                     rec->nr_address = NULL;
2794                     rec->nr_packet = slp->ns_frag;
2795                     STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2796                     ++slp->ns_numrec;
2797                     slp->ns_flag |= SLP_DOREC;
2798                     ++*countp;
2799                 }
2800                 slp->ns_frag = NULL;
2801             }
2802         }
2803 }
2804
2805 #ifdef INVARIANTS
2806
2807 /*
2808  * Sanity check our mbuf chain.
2809  */
2810 static void
2811 nfs_checkpkt(struct mbuf *m, int len)
2812 {
2813         int xlen = 0;
2814         while (m) {
2815                 xlen += m->m_len;
2816                 m = m->m_next;
2817         }
2818         if (xlen != len) {
2819                 panic("nfs_checkpkt: len mismatch %d/%d mbuf %p\n",
2820                         xlen, len, m);
2821         }
2822 }
2823
2824 #else
2825
2826 static void
2827 nfs_checkpkt(struct mbuf *m __unused, int len __unused)
2828 {
2829 }
2830
2831 #endif
2832
2833 /*
2834  * Parse an RPC header.
2835  *
2836  * If the socket is invalid or no records are pending we return ENOBUFS.
2837  * The caller must deal with NEEDQ races.
2838  */
2839 int
2840 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2841             struct nfsrv_descript **ndp)
2842 {
2843         struct nfsrv_rec *rec;
2844         struct mbuf *m;
2845         struct sockaddr *nam;
2846         struct nfsrv_descript *nd;
2847         int error;
2848
2849         *ndp = NULL;
2850         if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2851                 return (ENOBUFS);
2852         rec = STAILQ_FIRST(&slp->ns_rec);
2853         STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2854         KKASSERT(slp->ns_numrec > 0);
2855         if (--slp->ns_numrec == 0)
2856                 slp->ns_flag &= ~SLP_DOREC;
2857         nam = rec->nr_address;
2858         m = rec->nr_packet;
2859         kfree(rec, M_NFSRVDESC);
2860         nd = kmalloc(sizeof(struct nfsrv_descript), M_NFSRVDESC, M_WAITOK);
2861         nd->nd_md = nd->nd_mrep = m;
2862         nd->nd_nam2 = nam;
2863         nd->nd_dpos = mtod(m, caddr_t);
2864         error = nfs_getreq(nd, nfsd, TRUE);
2865         if (error) {
2866                 if (nam) {
2867                         kfree(nam, M_SONAME);
2868                 }
2869                 kfree((caddr_t)nd, M_NFSRVDESC);
2870                 return (error);
2871         }
2872         *ndp = nd;
2873         nfsd->nfsd_nd = nd;
2874         return (0);
2875 }
2876
2877 /*
2878  * Try to assign service sockets to nfsd threads based on the number
2879  * of new rpc requests that have been queued on the service socket.
2880  *
2881  * If no nfsd's are available or additonal requests are pending, set the
2882  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2883  * the work in the nfssvc_sock list when it is finished processing its
2884  * current work.  This flag is only cleared when an nfsd can not find
2885  * any new work to perform.
2886  */
2887 void
2888 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2889 {
2890         struct nfsd *nd;
2891
2892         if ((slp->ns_flag & SLP_VALID) == 0)
2893                 return;
2894         if (nparallel <= 1)
2895                 nparallel = 1;
2896         TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2897                 if (nd->nfsd_flag & NFSD_WAITING) {
2898                         nd->nfsd_flag &= ~NFSD_WAITING;
2899                         if (nd->nfsd_slp)
2900                                 panic("nfsd wakeup");
2901                         nfsrv_slpref(slp);
2902                         nd->nfsd_slp = slp;
2903                         wakeup((caddr_t)nd);
2904                         if (--nparallel == 0)
2905                                 break;
2906                 }
2907         }
2908
2909         /*
2910          * If we couldn't assign slp then the NFSDs are all busy and
2911          * we set a flag indicating that there is pending work.
2912          */
2913         if (nparallel)
2914                 nfsd_head_flag |= NFSD_CHECKSLP;
2915 }
2916 #endif /* NFS_NOSERVER */