Merge branch 'vendor/LIBPCAP'
[dragonfly.git] / sys / vfs / nfs / nfs_socket.c
1 /*
2  * Copyright (c) 1989, 1991, 1993, 1995
3  *      The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Rick Macklem at The University of Guelph.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by the University of
19  *      California, Berkeley and its contributors.
20  * 4. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *      @(#)nfs_socket.c        8.5 (Berkeley) 3/30/95
37  * $FreeBSD: src/sys/nfs/nfs_socket.c,v 1.60.2.6 2003/03/26 01:44:46 alfred Exp $
38  * $DragonFly: src/sys/vfs/nfs/nfs_socket.c,v 1.45 2007/05/18 17:05:13 dillon Exp $
39  */
40
41 /*
42  * Socket operations for use by nfs
43  */
44
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/proc.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/kernel.h>
51 #include <sys/mbuf.h>
52 #include <sys/vnode.h>
53 #include <sys/fcntl.h>
54 #include <sys/protosw.h>
55 #include <sys/resourcevar.h>
56 #include <sys/socket.h>
57 #include <sys/socketvar.h>
58 #include <sys/socketops.h>
59 #include <sys/syslog.h>
60 #include <sys/thread.h>
61 #include <sys/tprintf.h>
62 #include <sys/sysctl.h>
63 #include <sys/signalvar.h>
64 #include <sys/signal2.h>
65
66 #include <netinet/in.h>
67 #include <netinet/tcp.h>
68 #include <sys/thread2.h>
69
70 #include "rpcv2.h"
71 #include "nfsproto.h"
72 #include "nfs.h"
73 #include "xdr_subs.h"
74 #include "nfsm_subs.h"
75 #include "nfsmount.h"
76 #include "nfsnode.h"
77 #include "nfsrtt.h"
78
79 #define TRUE    1
80 #define FALSE   0
81
82 /*
83  * Estimate rto for an nfs rpc sent via. an unreliable datagram.
84  * Use the mean and mean deviation of rtt for the appropriate type of rpc
85  * for the frequent rpcs and a default for the others.
86  * The justification for doing "other" this way is that these rpcs
87  * happen so infrequently that timer est. would probably be stale.
88  * Also, since many of these rpcs are
89  * non-idempotent, a conservative timeout is desired.
90  * getattr, lookup - A+2D
91  * read, write     - A+4D
92  * other           - nm_timeo
93  */
94 #define NFS_RTO(n, t) \
95         ((t) == 0 ? (n)->nm_timeo : \
96          ((t) < 3 ? \
97           (((((n)->nm_srtt[t-1] + 3) >> 2) + (n)->nm_sdrtt[t-1] + 1) >> 1) : \
98           ((((n)->nm_srtt[t-1] + 7) >> 3) + (n)->nm_sdrtt[t-1] + 1)))
99 #define NFS_SRTT(r)     (r)->r_nmp->nm_srtt[proct[(r)->r_procnum] - 1]
100 #define NFS_SDRTT(r)    (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum] - 1]
101 /*
102  * External data, mostly RPC constants in XDR form
103  */
104 extern u_int32_t rpc_reply, rpc_msgdenied, rpc_mismatch, rpc_vers,
105         rpc_auth_unix, rpc_msgaccepted, rpc_call, rpc_autherr,
106         rpc_auth_kerb;
107 extern u_int32_t nfs_prog;
108 extern struct nfsstats nfsstats;
109 extern int nfsv3_procid[NFS_NPROCS];
110 extern int nfs_ticks;
111
112 /*
113  * Defines which timer to use for the procnum.
114  * 0 - default
115  * 1 - getattr
116  * 2 - lookup
117  * 3 - read
118  * 4 - write
119  */
120 static int proct[NFS_NPROCS] = {
121         0, 1, 0, 2, 1, 3, 3, 4, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 0, 0, 0, 0, 0,
122         0, 0, 0,
123 };
124
125 static int nfs_realign_test;
126 static int nfs_realign_count;
127 static int nfs_bufpackets = 4;
128 static int nfs_timer_raced;
129
130 SYSCTL_DECL(_vfs_nfs);
131
132 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0, "");
133 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0, "");
134 SYSCTL_INT(_vfs_nfs, OID_AUTO, bufpackets, CTLFLAG_RW, &nfs_bufpackets, 0, "");
135
136
137 /*
138  * There is a congestion window for outstanding rpcs maintained per mount
139  * point. The cwnd size is adjusted in roughly the way that:
140  * Van Jacobson, Congestion avoidance and Control, In "Proceedings of
141  * SIGCOMM '88". ACM, August 1988.
142  * describes for TCP. The cwnd size is chopped in half on a retransmit timeout
143  * and incremented by 1/cwnd when each rpc reply is received and a full cwnd
144  * of rpcs is in progress.
145  * (The sent count and cwnd are scaled for integer arith.)
146  * Variants of "slow start" were tried and were found to be too much of a
147  * performance hit (ave. rtt 3 times larger),
148  * I suspect due to the large rtt that nfs rpcs have.
149  */
150 #define NFS_CWNDSCALE   256
151 #define NFS_MAXCWND     (NFS_CWNDSCALE * 32)
152 static int nfs_backoff[8] = { 2, 4, 8, 16, 32, 64, 128, 256, };
153 int nfsrtton = 0;
154 struct nfsrtt nfsrtt;
155 struct callout  nfs_timer_handle;
156
157 static int      nfs_msg (struct thread *,char *,char *);
158 static int      nfs_rcvlock (struct nfsreq *);
159 static void     nfs_rcvunlock (struct nfsreq *);
160 static void     nfs_realign (struct mbuf **pm, int hsiz);
161 static int      nfs_receive (struct nfsreq *rep, struct sockaddr **aname,
162                                  struct mbuf **mp);
163 static void     nfs_softterm (struct nfsreq *rep);
164 static int      nfs_reconnect (struct nfsreq *rep);
165 #ifndef NFS_NOSERVER 
166 static int      nfsrv_getstream (struct nfssvc_sock *, int, int *);
167
168 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
169                                     struct nfssvc_sock *slp,
170                                     struct thread *td,
171                                     struct mbuf **mreqp) = {
172         nfsrv_null,
173         nfsrv_getattr,
174         nfsrv_setattr,
175         nfsrv_lookup,
176         nfsrv3_access,
177         nfsrv_readlink,
178         nfsrv_read,
179         nfsrv_write,
180         nfsrv_create,
181         nfsrv_mkdir,
182         nfsrv_symlink,
183         nfsrv_mknod,
184         nfsrv_remove,
185         nfsrv_rmdir,
186         nfsrv_rename,
187         nfsrv_link,
188         nfsrv_readdir,
189         nfsrv_readdirplus,
190         nfsrv_statfs,
191         nfsrv_fsinfo,
192         nfsrv_pathconf,
193         nfsrv_commit,
194         nfsrv_noop,
195         nfsrv_noop,
196         nfsrv_noop,
197         nfsrv_noop
198 };
199 #endif /* NFS_NOSERVER */
200
201 /*
202  * Initialize sockets and congestion for a new NFS connection.
203  * We do not free the sockaddr if error.
204  */
205 int
206 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
207 {
208         struct socket *so;
209         int error, rcvreserve, sndreserve;
210         int pktscale;
211         struct sockaddr *saddr;
212         struct sockaddr_in *sin;
213         struct thread *td = &thread0; /* only used for socreate and sobind */
214
215         nmp->nm_so = NULL;
216         saddr = nmp->nm_nam;
217         error = socreate(saddr->sa_family, &nmp->nm_so, nmp->nm_sotype,
218                 nmp->nm_soproto, td);
219         if (error)
220                 goto bad;
221         so = nmp->nm_so;
222         nmp->nm_soflags = so->so_proto->pr_flags;
223
224         /*
225          * Some servers require that the client port be a reserved port number.
226          */
227         if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
228                 struct sockopt sopt;
229                 int ip;
230                 struct sockaddr_in ssin;
231
232                 bzero(&sopt, sizeof sopt);
233                 ip = IP_PORTRANGE_LOW;
234                 sopt.sopt_level = IPPROTO_IP;
235                 sopt.sopt_name = IP_PORTRANGE;
236                 sopt.sopt_val = (void *)&ip;
237                 sopt.sopt_valsize = sizeof(ip);
238                 sopt.sopt_td = NULL;
239                 error = sosetopt(so, &sopt);
240                 if (error)
241                         goto bad;
242                 bzero(&ssin, sizeof ssin);
243                 sin = &ssin;
244                 sin->sin_len = sizeof (struct sockaddr_in);
245                 sin->sin_family = AF_INET;
246                 sin->sin_addr.s_addr = INADDR_ANY;
247                 sin->sin_port = htons(0);
248                 error = sobind(so, (struct sockaddr *)sin, td);
249                 if (error)
250                         goto bad;
251                 bzero(&sopt, sizeof sopt);
252                 ip = IP_PORTRANGE_DEFAULT;
253                 sopt.sopt_level = IPPROTO_IP;
254                 sopt.sopt_name = IP_PORTRANGE;
255                 sopt.sopt_val = (void *)&ip;
256                 sopt.sopt_valsize = sizeof(ip);
257                 sopt.sopt_td = NULL;
258                 error = sosetopt(so, &sopt);
259                 if (error)
260                         goto bad;
261         }
262
263         /*
264          * Protocols that do not require connections may be optionally left
265          * unconnected for servers that reply from a port other than NFS_PORT.
266          */
267         if (nmp->nm_flag & NFSMNT_NOCONN) {
268                 if (nmp->nm_soflags & PR_CONNREQUIRED) {
269                         error = ENOTCONN;
270                         goto bad;
271                 }
272         } else {
273                 error = soconnect(so, nmp->nm_nam, td);
274                 if (error)
275                         goto bad;
276
277                 /*
278                  * Wait for the connection to complete. Cribbed from the
279                  * connect system call but with the wait timing out so
280                  * that interruptible mounts don't hang here for a long time.
281                  */
282                 crit_enter();
283                 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
284                         (void) tsleep((caddr_t)&so->so_timeo, 0,
285                                 "nfscon", 2 * hz);
286                         if ((so->so_state & SS_ISCONNECTING) &&
287                             so->so_error == 0 && rep &&
288                             (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
289                                 so->so_state &= ~SS_ISCONNECTING;
290                                 crit_exit();
291                                 goto bad;
292                         }
293                 }
294                 if (so->so_error) {
295                         error = so->so_error;
296                         so->so_error = 0;
297                         crit_exit();
298                         goto bad;
299                 }
300                 crit_exit();
301         }
302         so->so_rcv.ssb_timeo = (5 * hz);
303         so->so_snd.ssb_timeo = (5 * hz);
304
305         /*
306          * Get buffer reservation size from sysctl, but impose reasonable
307          * limits.
308          */
309         pktscale = nfs_bufpackets;
310         if (pktscale < 2)
311                 pktscale = 2;
312         if (pktscale > 64)
313                 pktscale = 64;
314
315         if (nmp->nm_sotype == SOCK_DGRAM) {
316                 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
317                 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
318                     NFS_MAXPKTHDR) * pktscale;
319         } else if (nmp->nm_sotype == SOCK_SEQPACKET) {
320                 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
321                 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
322                     NFS_MAXPKTHDR) * pktscale;
323         } else {
324                 if (nmp->nm_sotype != SOCK_STREAM)
325                         panic("nfscon sotype");
326                 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
327                         struct sockopt sopt;
328                         int val;
329
330                         bzero(&sopt, sizeof sopt);
331                         sopt.sopt_level = SOL_SOCKET;
332                         sopt.sopt_name = SO_KEEPALIVE;
333                         sopt.sopt_val = &val;
334                         sopt.sopt_valsize = sizeof val;
335                         val = 1;
336                         sosetopt(so, &sopt);
337                 }
338                 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
339                         struct sockopt sopt;
340                         int val;
341
342                         bzero(&sopt, sizeof sopt);
343                         sopt.sopt_level = IPPROTO_TCP;
344                         sopt.sopt_name = TCP_NODELAY;
345                         sopt.sopt_val = &val;
346                         sopt.sopt_valsize = sizeof val;
347                         val = 1;
348                         sosetopt(so, &sopt);
349                 }
350                 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR +
351                     sizeof (u_int32_t)) * pktscale;
352                 rcvreserve = (nmp->nm_rsize + NFS_MAXPKTHDR +
353                     sizeof (u_int32_t)) * pktscale;
354         }
355         error = soreserve(so, sndreserve, rcvreserve,
356                           &td->td_proc->p_rlimit[RLIMIT_SBSIZE]);
357         if (error)
358                 goto bad;
359         so->so_rcv.ssb_flags |= SSB_NOINTR;
360         so->so_snd.ssb_flags |= SSB_NOINTR;
361
362         /* Initialize other non-zero congestion variables */
363         nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] = 
364                 nmp->nm_srtt[3] = (NFS_TIMEO << 3);
365         nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
366                 nmp->nm_sdrtt[3] = 0;
367         nmp->nm_cwnd = NFS_MAXCWND / 2;     /* Initial send window */
368         nmp->nm_sent = 0;
369         nmp->nm_timeouts = 0;
370         return (0);
371
372 bad:
373         nfs_disconnect(nmp);
374         return (error);
375 }
376
377 /*
378  * Reconnect routine:
379  * Called when a connection is broken on a reliable protocol.
380  * - clean up the old socket
381  * - nfs_connect() again
382  * - set R_MUSTRESEND for all outstanding requests on mount point
383  * If this fails the mount point is DEAD!
384  * nb: Must be called with the nfs_sndlock() set on the mount point.
385  */
386 static int
387 nfs_reconnect(struct nfsreq *rep)
388 {
389         struct nfsreq *rp;
390         struct nfsmount *nmp = rep->r_nmp;
391         int error;
392
393         nfs_disconnect(nmp);
394         while ((error = nfs_connect(nmp, rep)) != 0) {
395                 if (error == EINTR || error == ERESTART)
396                         return (EINTR);
397                 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
398         }
399
400         /*
401          * Loop through outstanding request list and fix up all requests
402          * on old socket.
403          */
404         crit_enter();
405         TAILQ_FOREACH(rp, &nfs_reqq, r_chain) {
406                 if (rp->r_nmp == nmp)
407                         rp->r_flags |= R_MUSTRESEND;
408         }
409         crit_exit();
410         return (0);
411 }
412
413 /*
414  * NFS disconnect. Clean up and unlink.
415  */
416 void
417 nfs_disconnect(struct nfsmount *nmp)
418 {
419         struct socket *so;
420
421         if (nmp->nm_so) {
422                 so = nmp->nm_so;
423                 nmp->nm_so = NULL;
424                 soshutdown(so, SHUT_RDWR);
425                 soclose(so, FNONBLOCK);
426         }
427 }
428
429 void
430 nfs_safedisconnect(struct nfsmount *nmp)
431 {
432         struct nfsreq dummyreq;
433
434         bzero(&dummyreq, sizeof(dummyreq));
435         dummyreq.r_nmp = nmp;
436         dummyreq.r_td = NULL;
437         nfs_rcvlock(&dummyreq);
438         nfs_disconnect(nmp);
439         nfs_rcvunlock(&dummyreq);
440 }
441
442 /*
443  * This is the nfs send routine. For connection based socket types, it
444  * must be called with an nfs_sndlock() on the socket.
445  * "rep == NULL" indicates that it has been called from a server.
446  * For the client side:
447  * - return EINTR if the RPC is terminated, 0 otherwise
448  * - set R_MUSTRESEND if the send fails for any reason
449  * - do any cleanup required by recoverable socket errors (?)
450  * For the server side:
451  * - return EINTR or ERESTART if interrupted by a signal
452  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
453  * - do any cleanup required by recoverable socket errors (?)
454  */
455 int
456 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
457          struct nfsreq *rep)
458 {
459         struct sockaddr *sendnam;
460         int error, soflags, flags;
461
462         if (rep) {
463                 if (rep->r_flags & R_SOFTTERM) {
464                         m_freem(top);
465                         return (EINTR);
466                 }
467                 if ((so = rep->r_nmp->nm_so) == NULL) {
468                         rep->r_flags |= R_MUSTRESEND;
469                         m_freem(top);
470                         return (0);
471                 }
472                 rep->r_flags &= ~R_MUSTRESEND;
473                 soflags = rep->r_nmp->nm_soflags;
474         } else
475                 soflags = so->so_proto->pr_flags;
476         if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
477                 sendnam = NULL;
478         else
479                 sendnam = nam;
480         if (so->so_type == SOCK_SEQPACKET)
481                 flags = MSG_EOR;
482         else
483                 flags = 0;
484
485         error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
486             curthread /*XXX*/);
487         /*
488          * ENOBUFS for dgram sockets is transient and non fatal.
489          * No need to log, and no need to break a soft mount.
490          */
491         if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
492                 error = 0;
493                 if (rep)                /* do backoff retransmit on client */
494                         rep->r_flags |= R_MUSTRESEND;
495         }
496
497         if (error) {
498                 if (rep) {
499                         log(LOG_INFO, "nfs send error %d for server %s\n",error,
500                             rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
501                         /*
502                          * Deal with errors for the client side.
503                          */
504                         if (rep->r_flags & R_SOFTTERM)
505                                 error = EINTR;
506                         else
507                                 rep->r_flags |= R_MUSTRESEND;
508                 } else
509                         log(LOG_INFO, "nfsd send error %d\n", error);
510
511                 /*
512                  * Handle any recoverable (soft) socket errors here. (?)
513                  */
514                 if (error != EINTR && error != ERESTART &&
515                         error != EWOULDBLOCK && error != EPIPE)
516                         error = 0;
517         }
518         return (error);
519 }
520
521 /*
522  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
523  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
524  * Mark and consolidate the data into a new mbuf list.
525  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
526  *     small mbufs.
527  * For SOCK_STREAM we must be very careful to read an entire record once
528  * we have read any of it, even if the system call has been interrupted.
529  */
530 static int
531 nfs_receive(struct nfsreq *rep, struct sockaddr **aname, struct mbuf **mp)
532 {
533         struct socket *so;
534         struct sockbuf sio;
535         struct uio auio;
536         struct iovec aio;
537         struct mbuf *m;
538         struct mbuf *control;
539         u_int32_t len;
540         struct sockaddr **getnam;
541         int error, sotype, rcvflg;
542         struct thread *td = curthread;  /* XXX */
543
544         /*
545          * Set up arguments for soreceive()
546          */
547         *mp = NULL;
548         *aname = NULL;
549         sotype = rep->r_nmp->nm_sotype;
550
551         /*
552          * For reliable protocols, lock against other senders/receivers
553          * in case a reconnect is necessary.
554          * For SOCK_STREAM, first get the Record Mark to find out how much
555          * more there is to get.
556          * We must lock the socket against other receivers
557          * until we have an entire rpc request/reply.
558          */
559         if (sotype != SOCK_DGRAM) {
560                 error = nfs_sndlock(rep);
561                 if (error)
562                         return (error);
563 tryagain:
564                 /*
565                  * Check for fatal errors and resending request.
566                  */
567                 /*
568                  * Ugh: If a reconnect attempt just happened, nm_so
569                  * would have changed. NULL indicates a failed
570                  * attempt that has essentially shut down this
571                  * mount point.
572                  */
573                 if (rep->r_mrep || (rep->r_flags & R_SOFTTERM)) {
574                         nfs_sndunlock(rep);
575                         return (EINTR);
576                 }
577                 so = rep->r_nmp->nm_so;
578                 if (!so) {
579                         error = nfs_reconnect(rep);
580                         if (error) {
581                                 nfs_sndunlock(rep);
582                                 return (error);
583                         }
584                         goto tryagain;
585                 }
586                 while (rep->r_flags & R_MUSTRESEND) {
587                         m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
588                         nfsstats.rpcretries++;
589                         error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
590                         if (error) {
591                                 if (error == EINTR || error == ERESTART ||
592                                     (error = nfs_reconnect(rep)) != 0) {
593                                         nfs_sndunlock(rep);
594                                         return (error);
595                                 }
596                                 goto tryagain;
597                         }
598                 }
599                 nfs_sndunlock(rep);
600                 if (sotype == SOCK_STREAM) {
601                         /*
602                          * Get the length marker from the stream
603                          */
604                         aio.iov_base = (caddr_t)&len;
605                         aio.iov_len = sizeof(u_int32_t);
606                         auio.uio_iov = &aio;
607                         auio.uio_iovcnt = 1;
608                         auio.uio_segflg = UIO_SYSSPACE;
609                         auio.uio_rw = UIO_READ;
610                         auio.uio_offset = 0;
611                         auio.uio_resid = sizeof(u_int32_t);
612                         auio.uio_td = td;
613                         do {
614                            rcvflg = MSG_WAITALL;
615                            error = so_pru_soreceive(so, NULL, &auio, NULL,
616                                                     NULL, &rcvflg);
617                            if (error == EWOULDBLOCK && rep) {
618                                 if (rep->r_flags & R_SOFTTERM)
619                                         return (EINTR);
620                            }
621                         } while (error == EWOULDBLOCK);
622
623                         if (error == 0 && auio.uio_resid > 0) {
624                             /*
625                              * Only log short packets if not EOF
626                              */
627                             if (auio.uio_resid != sizeof(u_int32_t))
628                             log(LOG_INFO,
629                                  "short receive (%d/%d) from nfs server %s\n",
630                                  (int)(sizeof(u_int32_t) - auio.uio_resid),
631                                  (int)sizeof(u_int32_t),
632                                  rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
633                             error = EPIPE;
634                         }
635                         if (error)
636                                 goto errout;
637                         len = ntohl(len) & ~0x80000000;
638                         /*
639                          * This is SERIOUS! We are out of sync with the sender
640                          * and forcing a disconnect/reconnect is all I can do.
641                          */
642                         if (len > NFS_MAXPACKET) {
643                             log(LOG_ERR, "%s (%d) from nfs server %s\n",
644                                 "impossible packet length",
645                                 len,
646                                 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
647                             error = EFBIG;
648                             goto errout;
649                         }
650
651                         /*
652                          * Get the rest of the packet as an mbuf chain
653                          */
654                         sbinit(&sio, len);
655                         do {
656                             rcvflg = MSG_WAITALL;
657                             error = so_pru_soreceive(so, NULL, NULL, &sio,
658                                                      NULL, &rcvflg);
659                         } while (error == EWOULDBLOCK || error == EINTR ||
660                                  error == ERESTART);
661                         if (error == 0 && sio.sb_cc != len) {
662                             if (sio.sb_cc != 0)
663                             log(LOG_INFO,
664                                 "short receive (%d/%d) from nfs server %s\n",
665                                 len - auio.uio_resid, len,
666                                 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
667                             error = EPIPE;
668                         }
669                         *mp = sio.sb_mb;
670                 } else {
671                         /*
672                          * Non-stream, so get the whole packet by not
673                          * specifying MSG_WAITALL and by specifying a large
674                          * length.
675                          *
676                          * We have no use for control msg., but must grab them
677                          * and then throw them away so we know what is going
678                          * on.
679                          */
680                         sbinit(&sio, 100000000);
681                         do {
682                             rcvflg = 0;
683                             error =  so_pru_soreceive(so, NULL, NULL, &sio,
684                                                       &control, &rcvflg);
685                             if (control)
686                                 m_freem(control);
687                             if (error == EWOULDBLOCK && rep) {
688                                 if (rep->r_flags & R_SOFTTERM) {
689                                         m_freem(sio.sb_mb);
690                                         return (EINTR);
691                                 }
692                             }
693                         } while (error == EWOULDBLOCK ||
694                                  (error == 0 && sio.sb_mb == NULL && control));
695                         if ((rcvflg & MSG_EOR) == 0)
696                                 kprintf("Egad!!\n");
697                         if (error == 0 && sio.sb_mb == NULL)
698                                 error = EPIPE;
699                         len = sio.sb_cc;
700                         *mp = sio.sb_mb;
701                 }
702 errout:
703                 if (error && error != EINTR && error != ERESTART) {
704                         m_freem(*mp);
705                         *mp = NULL;
706                         if (error != EPIPE) {
707                                 log(LOG_INFO,
708                                     "receive error %d from nfs server %s\n",
709                                     error,
710                                  rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
711                         }
712                         error = nfs_sndlock(rep);
713                         if (!error) {
714                                 error = nfs_reconnect(rep);
715                                 if (!error)
716                                         goto tryagain;
717                                 else
718                                         nfs_sndunlock(rep);
719                         }
720                 }
721         } else {
722                 if ((so = rep->r_nmp->nm_so) == NULL)
723                         return (EACCES);
724                 if (so->so_state & SS_ISCONNECTED)
725                         getnam = NULL;
726                 else
727                         getnam = aname;
728                 sbinit(&sio, 100000000);
729                 do {
730                         rcvflg = 0;
731                         error =  so_pru_soreceive(so, getnam, NULL, &sio,
732                                                   NULL, &rcvflg);
733                         if (error == EWOULDBLOCK &&
734                             (rep->r_flags & R_SOFTTERM)) {
735                                 m_freem(sio.sb_mb);
736                                 return (EINTR);
737                         }
738                 } while (error == EWOULDBLOCK);
739                 len = sio.sb_cc;
740                 *mp = sio.sb_mb;
741         }
742         if (error) {
743                 m_freem(*mp);
744                 *mp = NULL;
745         }
746         /*
747          * Search for any mbufs that are not a multiple of 4 bytes long
748          * or with m_data not longword aligned.
749          * These could cause pointer alignment problems, so copy them to
750          * well aligned mbufs.
751          */
752         nfs_realign(mp, 5 * NFSX_UNSIGNED);
753         return (error);
754 }
755
756 /*
757  * Implement receipt of reply on a socket.
758  * We must search through the list of received datagrams matching them
759  * with outstanding requests using the xid, until ours is found.
760  */
761 /* ARGSUSED */
762 int
763 nfs_reply(struct nfsreq *myrep)
764 {
765         struct nfsreq *rep;
766         struct nfsmount *nmp = myrep->r_nmp;
767         int32_t t1;
768         struct mbuf *mrep, *md;
769         struct sockaddr *nam;
770         u_int32_t rxid, *tl;
771         caddr_t dpos, cp2;
772         int error;
773
774         /*
775          * Loop around until we get our own reply
776          */
777         for (;;) {
778                 /*
779                  * Lock against other receivers so that I don't get stuck in
780                  * sbwait() after someone else has received my reply for me.
781                  * Also necessary for connection based protocols to avoid
782                  * race conditions during a reconnect.
783                  * If nfs_rcvlock() returns EALREADY, that means that
784                  * the reply has already been recieved by another
785                  * process and we can return immediately.  In this
786                  * case, the lock is not taken to avoid races with
787                  * other processes.
788                  */
789                 error = nfs_rcvlock(myrep);
790                 if (error == EALREADY)
791                         return (0);
792                 if (error)
793                         return (error);
794                 /*
795                  * Get the next Rpc reply off the socket
796                  */
797                 error = nfs_receive(myrep, &nam, &mrep);
798                 nfs_rcvunlock(myrep);
799                 if (error) {
800                         /*
801                          * Ignore routing errors on connectionless protocols??
802                          */
803                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
804                                 nmp->nm_so->so_error = 0;
805                                 if (myrep->r_flags & R_GETONEREP)
806                                         return (0);
807                                 continue;
808                         }
809                         return (error);
810                 }
811                 if (nam)
812                         FREE(nam, M_SONAME);
813
814                 /*
815                  * Get the xid and check that it is an rpc reply
816                  */
817                 md = mrep;
818                 dpos = mtod(md, caddr_t);
819                 nfsm_dissect(tl, u_int32_t *, 2*NFSX_UNSIGNED);
820                 rxid = *tl++;
821                 if (*tl != rpc_reply) {
822                         nfsstats.rpcinvalid++;
823                         m_freem(mrep);
824 nfsmout:
825                         if (myrep->r_flags & R_GETONEREP)
826                                 return (0);
827                         continue;
828                 }
829
830                 /*
831                  * Loop through the request list to match up the reply
832                  * Iff no match, just drop the datagram.  On match, set
833                  * r_mrep atomically to prevent the timer from messing
834                  * around with the request after we have exited the critical
835                  * section.
836                  */
837                 crit_enter();
838                 TAILQ_FOREACH(rep, &nfs_reqq, r_chain) {
839                         if (rep->r_mrep == NULL && rxid == rep->r_xid) {
840                                 rep->r_mrep = mrep;
841                                 break;
842                         }
843                 }
844                 crit_exit();
845
846                 /*
847                  * Fill in the rest of the reply if we found a match.
848                  */
849                 if (rep) {
850                         rep->r_md = md;
851                         rep->r_dpos = dpos;
852                         if (nfsrtton) {
853                                 struct rttl *rt;
854
855                                 rt = &nfsrtt.rttl[nfsrtt.pos];
856                                 rt->proc = rep->r_procnum;
857                                 rt->rto = NFS_RTO(nmp, proct[rep->r_procnum]);
858                                 rt->sent = nmp->nm_sent;
859                                 rt->cwnd = nmp->nm_cwnd;
860                                 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
861                                 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
862                                 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
863                                 getmicrotime(&rt->tstamp);
864                                 if (rep->r_flags & R_TIMING)
865                                         rt->rtt = rep->r_rtt;
866                                 else
867                                         rt->rtt = 1000000;
868                                 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
869                         }
870                         /*
871                          * Update congestion window.
872                          * Do the additive increase of
873                          * one rpc/rtt.
874                          */
875                         if (nmp->nm_cwnd <= nmp->nm_sent) {
876                                 nmp->nm_cwnd +=
877                                    (NFS_CWNDSCALE * NFS_CWNDSCALE +
878                                    (nmp->nm_cwnd >> 1)) / nmp->nm_cwnd;
879                                 if (nmp->nm_cwnd > NFS_MAXCWND)
880                                         nmp->nm_cwnd = NFS_MAXCWND;
881                         }
882                         crit_enter();   /* nfs_timer interlock for nm_sent */
883                         if (rep->r_flags & R_SENT) {
884                                 rep->r_flags &= ~R_SENT;
885                                 nmp->nm_sent -= NFS_CWNDSCALE;
886                         }
887                         crit_exit();
888                         /*
889                          * Update rtt using a gain of 0.125 on the mean
890                          * and a gain of 0.25 on the deviation.
891                          */
892                         if (rep->r_flags & R_TIMING) {
893                                 /*
894                                  * Since the timer resolution of
895                                  * NFS_HZ is so course, it can often
896                                  * result in r_rtt == 0. Since
897                                  * r_rtt == N means that the actual
898                                  * rtt is between N+dt and N+2-dt ticks,
899                                  * add 1.
900                                  */
901                                 t1 = rep->r_rtt + 1;
902                                 t1 -= (NFS_SRTT(rep) >> 3);
903                                 NFS_SRTT(rep) += t1;
904                                 if (t1 < 0)
905                                         t1 = -t1;
906                                 t1 -= (NFS_SDRTT(rep) >> 2);
907                                 NFS_SDRTT(rep) += t1;
908                         }
909                         nmp->nm_timeouts = 0;
910                 }
911                 /*
912                  * If not matched to a request, drop it.
913                  * If it's mine, get out.
914                  */
915                 if (rep == NULL) {
916                         nfsstats.rpcunexpected++;
917                         m_freem(mrep);
918                 } else if (rep == myrep) {
919                         if (rep->r_mrep == NULL)
920                                 panic("nfsreply nil");
921                         return (0);
922                 }
923                 if (myrep->r_flags & R_GETONEREP)
924                         return (0);
925         }
926 }
927
928 /*
929  * nfs_request - goes something like this
930  *      - fill in request struct
931  *      - links it into list
932  *      - calls nfs_send() for first transmit
933  *      - calls nfs_receive() to get reply
934  *      - break down rpc header and return with nfs reply pointed to
935  *        by mrep or error
936  * nb: always frees up mreq mbuf list
937  */
938 int
939 nfs_request(struct vnode *vp, struct mbuf *mrest, int procnum,
940             struct thread *td, struct ucred *cred, struct mbuf **mrp,
941             struct mbuf **mdp, caddr_t *dposp)
942 {
943         struct mbuf *mrep, *m2;
944         struct nfsreq *rep;
945         u_int32_t *tl;
946         int i;
947         struct nfsmount *nmp;
948         struct mbuf *m, *md, *mheadend;
949         char nickv[RPCX_NICKVERF];
950         time_t waituntil;
951         caddr_t dpos, cp2;
952         int t1, error = 0, mrest_len, auth_len, auth_type;
953         int trylater_delay = 15, trylater_cnt = 0, failed_auth = 0;
954         int verf_len, verf_type;
955         u_int32_t xid;
956         char *auth_str, *verf_str;
957         NFSKERBKEY_T key;               /* save session key */
958
959         /* Reject requests while attempting a forced unmount. */
960         if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
961                 m_freem(mrest);
962                 return (ESTALE);
963         }
964         nmp = VFSTONFS(vp->v_mount);
965         MALLOC(rep, struct nfsreq *, sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
966         rep->r_nmp = nmp;
967         rep->r_vp = vp;
968         rep->r_td = td;
969         rep->r_procnum = procnum;
970         rep->r_mreq = NULL;
971         i = 0;
972         m = mrest;
973         while (m) {
974                 i += m->m_len;
975                 m = m->m_next;
976         }
977         mrest_len = i;
978
979         /*
980          * Get the RPC header with authorization.
981          */
982 kerbauth:
983         verf_str = auth_str = NULL;
984         if (nmp->nm_flag & NFSMNT_KERB) {
985                 verf_str = nickv;
986                 verf_len = sizeof (nickv);
987                 auth_type = RPCAUTH_KERB4;
988                 bzero((caddr_t)key, sizeof (key));
989                 if (failed_auth || nfs_getnickauth(nmp, cred, &auth_str,
990                         &auth_len, verf_str, verf_len)) {
991                         error = nfs_getauth(nmp, rep, cred, &auth_str,
992                                 &auth_len, verf_str, &verf_len, key);
993                         if (error) {
994                                 kfree((caddr_t)rep, M_NFSREQ);
995                                 m_freem(mrest);
996                                 return (error);
997                         }
998                 }
999         } else {
1000                 auth_type = RPCAUTH_UNIX;
1001                 if (cred->cr_ngroups < 1)
1002                         panic("nfsreq nogrps");
1003                 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1004                         nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1005                         5 * NFSX_UNSIGNED;
1006         }
1007         m = nfsm_rpchead(cred, nmp->nm_flag, procnum, auth_type, auth_len,
1008              auth_str, verf_len, verf_str, mrest, mrest_len, &mheadend, &xid);
1009         if (auth_str)
1010                 kfree(auth_str, M_TEMP);
1011
1012         /*
1013          * For stream protocols, insert a Sun RPC Record Mark.
1014          */
1015         if (nmp->nm_sotype == SOCK_STREAM) {
1016                 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1017                 if (m == NULL) {
1018                         kfree(rep, M_NFSREQ);
1019                         return (ENOBUFS);
1020                 }
1021                 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1022                          (m->m_pkthdr.len - NFSX_UNSIGNED));
1023         }
1024         rep->r_mreq = m;
1025         rep->r_xid = xid;
1026 tryagain:
1027         if (nmp->nm_flag & NFSMNT_SOFT)
1028                 rep->r_retry = nmp->nm_retry;
1029         else
1030                 rep->r_retry = NFS_MAXREXMIT + 1;       /* past clip limit */
1031         rep->r_rtt = rep->r_rexmit = 0;
1032         if (proct[procnum] > 0)
1033                 rep->r_flags = R_TIMING | R_MASKTIMER;
1034         else
1035                 rep->r_flags = R_MASKTIMER;
1036         rep->r_mrep = NULL;
1037
1038         /*
1039          * Do the client side RPC.
1040          */
1041         nfsstats.rpcrequests++;
1042
1043         /*
1044          * Chain request into list of outstanding requests. Be sure
1045          * to put it LAST so timer finds oldest requests first.  Note
1046          * that R_MASKTIMER is set at the moment to prevent any timer
1047          * action on this request while we are still doing processing on
1048          * it below.  splsoftclock() primarily protects nm_sent.  Note
1049          * that we may block in this code so there is no atomicy guarentee.
1050          */
1051         crit_enter();
1052         TAILQ_INSERT_TAIL(&nfs_reqq, rep, r_chain);
1053
1054         /*
1055          * If backing off another request or avoiding congestion, don't
1056          * send this one now but let timer do it.  If not timing a request,
1057          * do it now. 
1058          *
1059          * Even though the timer will not mess with our request there is
1060          * still the possibility that we will race a reply (which clears
1061          * R_SENT), especially on localhost connections, so be very careful
1062          * when setting R_SENT.  We could set R_SENT prior to calling
1063          * nfs_send() but why bother if the response occurs that quickly?
1064          */
1065         if (nmp->nm_so && (nmp->nm_sotype != SOCK_DGRAM ||
1066             (nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1067             nmp->nm_sent < nmp->nm_cwnd)) {
1068                 if (nmp->nm_soflags & PR_CONNREQUIRED)
1069                         error = nfs_sndlock(rep);
1070                 if (!error) {
1071                         m2 = m_copym(m, 0, M_COPYALL, MB_WAIT);
1072                         error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1073                         if (nmp->nm_soflags & PR_CONNREQUIRED)
1074                                 nfs_sndunlock(rep);
1075                 }
1076                 if (!error && (rep->r_flags & R_MUSTRESEND) == 0 &&
1077                     rep->r_mrep == NULL) {
1078                         KASSERT((rep->r_flags & R_SENT) == 0,
1079                                 ("R_SENT ASSERT %p", rep));
1080                         nmp->nm_sent += NFS_CWNDSCALE;
1081                         rep->r_flags |= R_SENT;
1082                 }
1083         } else {
1084                 rep->r_rtt = -1;
1085         }
1086
1087         /*
1088          * Let the timer do what it will with the request, then
1089          * wait for the reply from our send or the timer's.
1090          */
1091         if (!error || error == EPIPE) {
1092                 rep->r_flags &= ~R_MASKTIMER;
1093                 crit_exit();
1094                 error = nfs_reply(rep);
1095                 crit_enter();
1096         }
1097
1098         /*
1099          * RPC done, unlink the request, but don't rip it out from under
1100          * the callout timer.
1101          */
1102         while (rep->r_flags & R_LOCKED) {
1103                 nfs_timer_raced = 1;
1104                 tsleep(&nfs_timer_raced, 0, "nfstrac", 0);
1105         }
1106         TAILQ_REMOVE(&nfs_reqq, rep, r_chain);
1107
1108         /*
1109          * Decrement the outstanding request count.
1110          */
1111         if (rep->r_flags & R_SENT) {
1112                 rep->r_flags &= ~R_SENT;
1113                 nmp->nm_sent -= NFS_CWNDSCALE;
1114         }
1115         crit_exit();
1116
1117         /*
1118          * If there was a successful reply and a tprintf msg.
1119          * tprintf a response.
1120          */
1121         if (!error && (rep->r_flags & R_TPRINTFMSG))
1122                 nfs_msg(rep->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1123                     "is alive again");
1124         mrep = rep->r_mrep;
1125         md = rep->r_md;
1126         dpos = rep->r_dpos;
1127         if (error) {
1128                 m_freem(rep->r_mreq);
1129                 kfree((caddr_t)rep, M_NFSREQ);
1130                 return (error);
1131         }
1132
1133         /*
1134          * break down the rpc header and check if ok
1135          */
1136         nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1137         if (*tl++ == rpc_msgdenied) {
1138                 if (*tl == rpc_mismatch)
1139                         error = EOPNOTSUPP;
1140                 else if ((nmp->nm_flag & NFSMNT_KERB) && *tl++ == rpc_autherr) {
1141                         if (!failed_auth) {
1142                                 failed_auth++;
1143                                 mheadend->m_next = NULL;
1144                                 m_freem(mrep);
1145                                 m_freem(rep->r_mreq);
1146                                 goto kerbauth;
1147                         } else
1148                                 error = EAUTH;
1149                 } else
1150                         error = EACCES;
1151                 m_freem(mrep);
1152                 m_freem(rep->r_mreq);
1153                 kfree((caddr_t)rep, M_NFSREQ);
1154                 return (error);
1155         }
1156
1157         /*
1158          * Grab any Kerberos verifier, otherwise just throw it away.
1159          */
1160         verf_type = fxdr_unsigned(int, *tl++);
1161         i = fxdr_unsigned(int32_t, *tl);
1162         if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1163                 error = nfs_savenickauth(nmp, cred, i, key, &md, &dpos, mrep);
1164                 if (error)
1165                         goto nfsmout;
1166         } else if (i > 0)
1167                 nfsm_adv(nfsm_rndup(i));
1168         nfsm_dissect(tl, u_int32_t *, NFSX_UNSIGNED);
1169         /* 0 == ok */
1170         if (*tl == 0) {
1171                 nfsm_dissect(tl, u_int32_t *, NFSX_UNSIGNED);
1172                 if (*tl != 0) {
1173                         error = fxdr_unsigned(int, *tl);
1174                         if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1175                                 error == NFSERR_TRYLATER) {
1176                                 m_freem(mrep);
1177                                 error = 0;
1178                                 waituntil = time_second + trylater_delay;
1179                                 while (time_second < waituntil)
1180                                         (void) tsleep((caddr_t)&lbolt,
1181                                                 0, "nqnfstry", 0);
1182                                 trylater_delay *= nfs_backoff[trylater_cnt];
1183                                 if (trylater_cnt < 7)
1184                                         trylater_cnt++;
1185                                 goto tryagain;
1186                         }
1187
1188                         /*
1189                          * If the File Handle was stale, invalidate the
1190                          * lookup cache, just in case.
1191                          *
1192                          * To avoid namecache<->vnode deadlocks we must
1193                          * release the vnode lock if we hold it.
1194                          */
1195                         if (error == ESTALE) {
1196                                 int ltype;
1197
1198                                 ltype = lockstatus(&vp->v_lock, curthread);
1199                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1200                                         lockmgr(&vp->v_lock, LK_RELEASE);
1201                                 cache_inval_vp(vp, CINV_CHILDREN);
1202                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1203                                         lockmgr(&vp->v_lock, ltype);
1204                         }
1205                         if (nmp->nm_flag & NFSMNT_NFSV3) {
1206                                 *mrp = mrep;
1207                                 *mdp = md;
1208                                 *dposp = dpos;
1209                                 error |= NFSERR_RETERR;
1210                         } else
1211                                 m_freem(mrep);
1212                         m_freem(rep->r_mreq);
1213                         kfree((caddr_t)rep, M_NFSREQ);
1214                         return (error);
1215                 }
1216
1217                 *mrp = mrep;
1218                 *mdp = md;
1219                 *dposp = dpos;
1220                 m_freem(rep->r_mreq);
1221                 FREE((caddr_t)rep, M_NFSREQ);
1222                 return (0);
1223         }
1224         m_freem(mrep);
1225         error = EPROTONOSUPPORT;
1226 nfsmout:
1227         m_freem(rep->r_mreq);
1228         kfree((caddr_t)rep, M_NFSREQ);
1229         return (error);
1230 }
1231
1232 #ifndef NFS_NOSERVER
1233 /*
1234  * Generate the rpc reply header
1235  * siz arg. is used to decide if adding a cluster is worthwhile
1236  */
1237 int
1238 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1239             int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1240 {
1241         u_int32_t *tl;
1242         struct mbuf *mreq;
1243         caddr_t bpos;
1244         struct mbuf *mb, *mb2;
1245
1246         siz += RPC_REPLYSIZ;
1247         mb = mreq = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1248         mreq->m_pkthdr.len = 0;
1249         /*
1250          * If this is not a cluster, try and leave leading space
1251          * for the lower level headers.
1252          */
1253         if ((max_hdr + siz) < MINCLSIZE)
1254                 mreq->m_data += max_hdr;
1255         tl = mtod(mreq, u_int32_t *);
1256         mreq->m_len = 6 * NFSX_UNSIGNED;
1257         bpos = ((caddr_t)tl) + mreq->m_len;
1258         *tl++ = txdr_unsigned(nd->nd_retxid);
1259         *tl++ = rpc_reply;
1260         if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1261                 *tl++ = rpc_msgdenied;
1262                 if (err & NFSERR_AUTHERR) {
1263                         *tl++ = rpc_autherr;
1264                         *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1265                         mreq->m_len -= NFSX_UNSIGNED;
1266                         bpos -= NFSX_UNSIGNED;
1267                 } else {
1268                         *tl++ = rpc_mismatch;
1269                         *tl++ = txdr_unsigned(RPC_VER2);
1270                         *tl = txdr_unsigned(RPC_VER2);
1271                 }
1272         } else {
1273                 *tl++ = rpc_msgaccepted;
1274
1275                 /*
1276                  * For Kerberos authentication, we must send the nickname
1277                  * verifier back, otherwise just RPCAUTH_NULL.
1278                  */
1279                 if (nd->nd_flag & ND_KERBFULL) {
1280                     struct nfsuid *nuidp;
1281                     struct timeval ktvin, ktvout;
1282
1283                     for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1284                         nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1285                         if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1286                             (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1287                              &nuidp->nu_haddr, nd->nd_nam2)))
1288                             break;
1289                     }
1290                     if (nuidp) {
1291                         ktvin.tv_sec =
1292                             txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1293                         ktvin.tv_usec =
1294                             txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1295
1296                         /*
1297                          * Encrypt the timestamp in ecb mode using the
1298                          * session key.
1299                          */
1300 #ifdef NFSKERB
1301                         XXX
1302 #endif
1303
1304                         *tl++ = rpc_auth_kerb;
1305                         *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1306                         *tl = ktvout.tv_sec;
1307                         nfsm_build(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1308                         *tl++ = ktvout.tv_usec;
1309                         *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1310                     } else {
1311                         *tl++ = 0;
1312                         *tl++ = 0;
1313                     }
1314                 } else {
1315                         *tl++ = 0;
1316                         *tl++ = 0;
1317                 }
1318                 switch (err) {
1319                 case EPROGUNAVAIL:
1320                         *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1321                         break;
1322                 case EPROGMISMATCH:
1323                         *tl = txdr_unsigned(RPC_PROGMISMATCH);
1324                         nfsm_build(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
1325                         *tl++ = txdr_unsigned(2);
1326                         *tl = txdr_unsigned(3);
1327                         break;
1328                 case EPROCUNAVAIL:
1329                         *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1330                         break;
1331                 case EBADRPC:
1332                         *tl = txdr_unsigned(RPC_GARBAGE);
1333                         break;
1334                 default:
1335                         *tl = 0;
1336                         if (err != NFSERR_RETVOID) {
1337                                 nfsm_build(tl, u_int32_t *, NFSX_UNSIGNED);
1338                                 if (err)
1339                                     *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1340                                 else
1341                                     *tl = 0;
1342                         }
1343                         break;
1344                 };
1345         }
1346
1347         if (mrq != NULL)
1348             *mrq = mreq;
1349         *mbp = mb;
1350         *bposp = bpos;
1351         if (err != 0 && err != NFSERR_RETVOID)
1352                 nfsstats.srvrpc_errs++;
1353         return (0);
1354 }
1355
1356
1357 #endif /* NFS_NOSERVER */
1358 /*
1359  * Nfs timer routine
1360  * Scan the nfsreq list and retranmit any requests that have timed out
1361  * To avoid retransmission attempts on STREAM sockets (in the future) make
1362  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1363  */
1364 void
1365 nfs_timer(void *arg /* never used */)
1366 {
1367         struct nfsreq *rep;
1368         struct mbuf *m;
1369         struct socket *so;
1370         struct nfsmount *nmp;
1371         int timeo;
1372         int error;
1373 #ifndef NFS_NOSERVER
1374         struct nfssvc_sock *slp;
1375         u_quad_t cur_usec;
1376 #endif /* NFS_NOSERVER */
1377         struct thread *td = &thread0; /* XXX for credentials, will break if sleep */
1378
1379         crit_enter();
1380         TAILQ_FOREACH(rep, &nfs_reqq, r_chain) {
1381                 nmp = rep->r_nmp;
1382                 if (rep->r_mrep || (rep->r_flags & (R_SOFTTERM|R_MASKTIMER)))
1383                         continue;
1384                 rep->r_flags |= R_LOCKED;
1385                 if (nfs_sigintr(nmp, rep, rep->r_td)) {
1386                         nfs_softterm(rep);
1387                         goto skip;
1388                 }
1389                 if (rep->r_rtt >= 0) {
1390                         rep->r_rtt++;
1391                         if (nmp->nm_flag & NFSMNT_DUMBTIMR)
1392                                 timeo = nmp->nm_timeo;
1393                         else
1394                                 timeo = NFS_RTO(nmp, proct[rep->r_procnum]);
1395                         if (nmp->nm_timeouts > 0)
1396                                 timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1397                         if (rep->r_rtt <= timeo)
1398                                 goto skip;
1399                         if (nmp->nm_timeouts < 8)
1400                                 nmp->nm_timeouts++;
1401                 }
1402                 /*
1403                  * Check for server not responding
1404                  */
1405                 if ((rep->r_flags & R_TPRINTFMSG) == 0 &&
1406                      rep->r_rexmit > nmp->nm_deadthresh) {
1407                         nfs_msg(rep->r_td,
1408                             nmp->nm_mountp->mnt_stat.f_mntfromname,
1409                             "not responding");
1410                         rep->r_flags |= R_TPRINTFMSG;
1411                 }
1412                 if (rep->r_rexmit >= rep->r_retry) {    /* too many */
1413                         nfsstats.rpctimeouts++;
1414                         nfs_softterm(rep);
1415                         goto skip;
1416                 }
1417                 if (nmp->nm_sotype != SOCK_DGRAM) {
1418                         if (++rep->r_rexmit > NFS_MAXREXMIT)
1419                                 rep->r_rexmit = NFS_MAXREXMIT;
1420                         goto skip;
1421                 }
1422                 if ((so = nmp->nm_so) == NULL)
1423                         goto skip;
1424
1425                 /*
1426                  * If there is enough space and the window allows..
1427                  *      Resend it
1428                  * Set r_rtt to -1 in case we fail to send it now.
1429                  */
1430                 rep->r_rtt = -1;
1431                 if (ssb_space(&so->so_snd) >= rep->r_mreq->m_pkthdr.len &&
1432                    ((nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1433                     (rep->r_flags & R_SENT) ||
1434                     nmp->nm_sent < nmp->nm_cwnd) &&
1435                    (m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1436                         if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1437                             error = so_pru_send(so, 0, m, NULL, NULL, td);
1438                         else
1439                             error = so_pru_send(so, 0, m, nmp->nm_nam,
1440                                 NULL, td);
1441                         if (error) {
1442                                 if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1443                                         so->so_error = 0;
1444                         } else if (rep->r_mrep == NULL) {
1445                                 /*
1446                                  * Iff first send, start timing
1447                                  * else turn timing off, backoff timer
1448                                  * and divide congestion window by 2.
1449                                  *
1450                                  * It is possible for the so_pru_send() to
1451                                  * block and for us to race a reply so we
1452                                  * only do this if the reply field has not 
1453                                  * been filled in.  R_LOCKED will prevent
1454                                  * the request from being ripped out from under
1455                                  * us entirely.
1456                                  */
1457                                 if (rep->r_flags & R_SENT) {
1458                                         rep->r_flags &= ~R_TIMING;
1459                                         if (++rep->r_rexmit > NFS_MAXREXMIT)
1460                                                 rep->r_rexmit = NFS_MAXREXMIT;
1461                                         nmp->nm_cwnd >>= 1;
1462                                         if (nmp->nm_cwnd < NFS_CWNDSCALE)
1463                                                 nmp->nm_cwnd = NFS_CWNDSCALE;
1464                                         nfsstats.rpcretries++;
1465                                 } else {
1466                                         rep->r_flags |= R_SENT;
1467                                         nmp->nm_sent += NFS_CWNDSCALE;
1468                                 }
1469                                 rep->r_rtt = 0;
1470                         }
1471                 }
1472 skip:
1473                 rep->r_flags &= ~R_LOCKED;
1474         }
1475 #ifndef NFS_NOSERVER
1476
1477         /*
1478          * Scan the write gathering queues for writes that need to be
1479          * completed now.
1480          */
1481         cur_usec = nfs_curusec();
1482         TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1483             if (slp->ns_tq.lh_first && slp->ns_tq.lh_first->nd_time<=cur_usec)
1484                 nfsrv_wakenfsd(slp, 1);
1485         }
1486 #endif /* NFS_NOSERVER */
1487
1488         /*
1489          * Due to possible blocking, a client operation may be waiting for
1490          * us to finish processing this request so it can remove it.
1491          */
1492         if (nfs_timer_raced) {
1493                 nfs_timer_raced = 0;
1494                 wakeup(&nfs_timer_raced);
1495         }
1496         crit_exit();
1497         callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer, NULL);
1498 }
1499
1500 /*
1501  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1502  * wait for all requests to complete. This is used by forced unmounts
1503  * to terminate any outstanding RPCs.
1504  */
1505 int
1506 nfs_nmcancelreqs(struct nfsmount *nmp)
1507 {
1508         struct nfsreq *req;
1509         int i;
1510
1511         crit_enter();
1512         TAILQ_FOREACH(req, &nfs_reqq, r_chain) {
1513                 if (nmp != req->r_nmp || req->r_mrep != NULL ||
1514                     (req->r_flags & R_SOFTTERM)) {
1515                         continue;
1516                 }
1517                 nfs_softterm(req);
1518         }
1519         crit_exit();
1520
1521         for (i = 0; i < 30; i++) {
1522                 crit_enter();
1523                 TAILQ_FOREACH(req, &nfs_reqq, r_chain) {
1524                         if (nmp == req->r_nmp)
1525                                 break;
1526                 }
1527                 crit_exit();
1528                 if (req == NULL)
1529                         return (0);
1530                 tsleep(&lbolt, 0, "nfscancel", 0);
1531         }
1532         return (EBUSY);
1533 }
1534
1535 /*
1536  * Flag a request as being about to terminate (due to NFSMNT_INT/NFSMNT_SOFT).
1537  * The nm_send count is decremented now to avoid deadlocks when the process in
1538  * soreceive() hasn't yet managed to send its own request.
1539  *
1540  * This routine must be called at splsoftclock() to protect r_flags and
1541  * nm_sent.
1542  */
1543
1544 static void
1545 nfs_softterm(struct nfsreq *rep)
1546 {
1547         rep->r_flags |= R_SOFTTERM;
1548
1549         if (rep->r_flags & R_SENT) {
1550                 rep->r_nmp->nm_sent -= NFS_CWNDSCALE;
1551                 rep->r_flags &= ~R_SENT;
1552         }
1553 }
1554
1555 /*
1556  * Test for a termination condition pending on the process.
1557  * This is used for NFSMNT_INT mounts.
1558  */
1559 int
1560 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
1561 {
1562         sigset_t tmpset;
1563         struct proc *p;
1564         struct lwp *lp;
1565
1566         if (rep && (rep->r_flags & R_SOFTTERM))
1567                 return (EINTR);
1568         /* Terminate all requests while attempting a forced unmount. */
1569         if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
1570                 return (EINTR);
1571         if (!(nmp->nm_flag & NFSMNT_INT))
1572                 return (0);
1573         /* td might be NULL YYY */
1574         if (td == NULL || (p = td->td_proc) == NULL)
1575                 return (0);
1576
1577         lp = td->td_lwp;
1578         tmpset = lwp_sigpend(lp);
1579         SIGSETNAND(tmpset, lp->lwp_sigmask);
1580         SIGSETNAND(tmpset, p->p_sigignore);
1581         if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
1582                 return (EINTR);
1583
1584         return (0);
1585 }
1586
1587 /*
1588  * Lock a socket against others.
1589  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
1590  * and also to avoid race conditions between the processes with nfs requests
1591  * in progress when a reconnect is necessary.
1592  */
1593 int
1594 nfs_sndlock(struct nfsreq *rep)
1595 {
1596         int *statep = &rep->r_nmp->nm_state;
1597         struct thread *td;
1598         int slptimeo;
1599         int slpflag;
1600         int error;
1601
1602         slpflag = 0;
1603         slptimeo = 0;
1604         td = rep->r_td;
1605         if (rep->r_nmp->nm_flag & NFSMNT_INT)
1606                 slpflag = PCATCH;
1607
1608         error = 0;
1609         crit_enter();
1610         while (*statep & NFSSTA_SNDLOCK) {
1611                 *statep |= NFSSTA_WANTSND;
1612                 if (nfs_sigintr(rep->r_nmp, rep, td)) {
1613                         error = EINTR;
1614                         break;
1615                 }
1616                 tsleep((caddr_t)statep, slpflag, "nfsndlck", slptimeo);
1617                 if (slpflag == PCATCH) {
1618                         slpflag = 0;
1619                         slptimeo = 2 * hz;
1620                 }
1621         }
1622         /* Always fail if our request has been cancelled. */
1623         if ((rep->r_flags & R_SOFTTERM))
1624                 error = EINTR;
1625         if (error == 0)
1626                 *statep |= NFSSTA_SNDLOCK;
1627         crit_exit();
1628         return (error);
1629 }
1630
1631 /*
1632  * Unlock the stream socket for others.
1633  */
1634 void
1635 nfs_sndunlock(struct nfsreq *rep)
1636 {
1637         int *statep = &rep->r_nmp->nm_state;
1638
1639         if ((*statep & NFSSTA_SNDLOCK) == 0)
1640                 panic("nfs sndunlock");
1641         crit_enter();
1642         *statep &= ~NFSSTA_SNDLOCK;
1643         if (*statep & NFSSTA_WANTSND) {
1644                 *statep &= ~NFSSTA_WANTSND;
1645                 wakeup((caddr_t)statep);
1646         }
1647         crit_exit();
1648 }
1649
1650 static int
1651 nfs_rcvlock(struct nfsreq *rep)
1652 {
1653         int *statep = &rep->r_nmp->nm_state;
1654         int slpflag;
1655         int slptimeo;
1656         int error;
1657
1658         /*
1659          * Unconditionally check for completion in case another nfsiod
1660          * get the packet while the caller was blocked, before the caller
1661          * called us.  Packet reception is handled by mainline code which
1662          * is protected by the BGL at the moment.
1663          *
1664          * We do not strictly need the second check just before the
1665          * tsleep(), but it's good defensive programming.
1666          */
1667         if (rep->r_mrep != NULL)
1668                 return (EALREADY);
1669
1670         if (rep->r_nmp->nm_flag & NFSMNT_INT)
1671                 slpflag = PCATCH;
1672         else
1673                 slpflag = 0;
1674         slptimeo = 0;
1675         error = 0;
1676         crit_enter();
1677         while (*statep & NFSSTA_RCVLOCK) {
1678                 if (nfs_sigintr(rep->r_nmp, rep, rep->r_td)) {
1679                         error = EINTR;
1680                         break;
1681                 }
1682                 if (rep->r_mrep != NULL) {
1683                         error = EALREADY;
1684                         break;
1685                 }
1686                 *statep |= NFSSTA_WANTRCV;
1687                 tsleep((caddr_t)statep, slpflag, "nfsrcvlk", slptimeo);
1688                 /*
1689                  * If our reply was recieved while we were sleeping,
1690                  * then just return without taking the lock to avoid a
1691                  * situation where a single iod could 'capture' the
1692                  * recieve lock.
1693                  */
1694                 if (rep->r_mrep != NULL) {
1695                         error = EALREADY;
1696                         break;
1697                 }
1698                 if (slpflag == PCATCH) {
1699                         slpflag = 0;
1700                         slptimeo = 2 * hz;
1701                 }
1702         }
1703         if (error == 0) {
1704                 *statep |= NFSSTA_RCVLOCK;
1705                 rep->r_nmp->nm_rcvlock_td = curthread;  /* DEBUGGING */
1706         }
1707         crit_exit();
1708         return (error);
1709 }
1710
1711 /*
1712  * Unlock the stream socket for others.
1713  */
1714 static void
1715 nfs_rcvunlock(struct nfsreq *rep)
1716 {
1717         int *statep = &rep->r_nmp->nm_state;
1718
1719         if ((*statep & NFSSTA_RCVLOCK) == 0)
1720                 panic("nfs rcvunlock");
1721         crit_enter();
1722         rep->r_nmp->nm_rcvlock_td = (void *)-1; /* DEBUGGING */
1723         *statep &= ~NFSSTA_RCVLOCK;
1724         if (*statep & NFSSTA_WANTRCV) {
1725                 *statep &= ~NFSSTA_WANTRCV;
1726                 wakeup((caddr_t)statep);
1727         }
1728         crit_exit();
1729 }
1730
1731 /*
1732  *      nfs_realign:
1733  *
1734  *      Check for badly aligned mbuf data and realign by copying the unaligned
1735  *      portion of the data into a new mbuf chain and freeing the portions
1736  *      of the old chain that were replaced.
1737  *
1738  *      We cannot simply realign the data within the existing mbuf chain
1739  *      because the underlying buffers may contain other rpc commands and
1740  *      we cannot afford to overwrite them.
1741  *
1742  *      We would prefer to avoid this situation entirely.  The situation does
1743  *      not occur with NFS/UDP and is supposed to only occassionally occur
1744  *      with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
1745  */
1746 static void
1747 nfs_realign(struct mbuf **pm, int hsiz)
1748 {
1749         struct mbuf *m;
1750         struct mbuf *n = NULL;
1751         int off = 0;
1752
1753         ++nfs_realign_test;
1754
1755         while ((m = *pm) != NULL) {
1756                 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3)) {
1757                         n = m_getl(m->m_len, MB_WAIT, MT_DATA, 0, NULL);
1758                         n->m_len = 0;
1759                         break;
1760                 }
1761                 pm = &m->m_next;
1762         }
1763
1764         /*
1765          * If n is non-NULL, loop on m copying data, then replace the
1766          * portion of the chain that had to be realigned.
1767          */
1768         if (n != NULL) {
1769                 ++nfs_realign_count;
1770                 while (m) {
1771                         m_copyback(n, off, m->m_len, mtod(m, caddr_t));
1772                         off += m->m_len;
1773                         m = m->m_next;
1774                 }
1775                 m_freem(*pm);
1776                 *pm = n;
1777         }
1778 }
1779
1780 #ifndef NFS_NOSERVER
1781
1782 /*
1783  * Parse an RPC request
1784  * - verify it
1785  * - fill in the cred struct.
1786  */
1787 int
1788 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
1789 {
1790         int len, i;
1791         u_int32_t *tl;
1792         int32_t t1;
1793         struct uio uio;
1794         struct iovec iov;
1795         caddr_t dpos, cp2, cp;
1796         u_int32_t nfsvers, auth_type;
1797         uid_t nickuid;
1798         int error = 0, ticklen;
1799         struct mbuf *mrep, *md;
1800         struct nfsuid *nuidp;
1801         struct timeval tvin, tvout;
1802 #if 0                           /* until encrypted keys are implemented */
1803         NFSKERBKEYSCHED_T keys; /* stores key schedule */
1804 #endif
1805
1806         mrep = nd->nd_mrep;
1807         md = nd->nd_md;
1808         dpos = nd->nd_dpos;
1809         if (has_header) {
1810                 nfsm_dissect(tl, u_int32_t *, 10 * NFSX_UNSIGNED);
1811                 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
1812                 if (*tl++ != rpc_call) {
1813                         m_freem(mrep);
1814                         return (EBADRPC);
1815                 }
1816         } else
1817                 nfsm_dissect(tl, u_int32_t *, 8 * NFSX_UNSIGNED);
1818         nd->nd_repstat = 0;
1819         nd->nd_flag = 0;
1820         if (*tl++ != rpc_vers) {
1821                 nd->nd_repstat = ERPCMISMATCH;
1822                 nd->nd_procnum = NFSPROC_NOOP;
1823                 return (0);
1824         }
1825         if (*tl != nfs_prog) {
1826                 nd->nd_repstat = EPROGUNAVAIL;
1827                 nd->nd_procnum = NFSPROC_NOOP;
1828                 return (0);
1829         }
1830         tl++;
1831         nfsvers = fxdr_unsigned(u_int32_t, *tl++);
1832         if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
1833                 nd->nd_repstat = EPROGMISMATCH;
1834                 nd->nd_procnum = NFSPROC_NOOP;
1835                 return (0);
1836         }
1837         if (nfsvers == NFS_VER3)
1838                 nd->nd_flag = ND_NFSV3;
1839         nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
1840         if (nd->nd_procnum == NFSPROC_NULL)
1841                 return (0);
1842         if (nd->nd_procnum >= NFS_NPROCS ||
1843                 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
1844                 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
1845                 nd->nd_repstat = EPROCUNAVAIL;
1846                 nd->nd_procnum = NFSPROC_NOOP;
1847                 return (0);
1848         }
1849         if ((nd->nd_flag & ND_NFSV3) == 0)
1850                 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
1851         auth_type = *tl++;
1852         len = fxdr_unsigned(int, *tl++);
1853         if (len < 0 || len > RPCAUTH_MAXSIZ) {
1854                 m_freem(mrep);
1855                 return (EBADRPC);
1856         }
1857
1858         nd->nd_flag &= ~ND_KERBAUTH;
1859         /*
1860          * Handle auth_unix or auth_kerb.
1861          */
1862         if (auth_type == rpc_auth_unix) {
1863                 len = fxdr_unsigned(int, *++tl);
1864                 if (len < 0 || len > NFS_MAXNAMLEN) {
1865                         m_freem(mrep);
1866                         return (EBADRPC);
1867                 }
1868                 nfsm_adv(nfsm_rndup(len));
1869                 nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1870                 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
1871                 nd->nd_cr.cr_ref = 1;
1872                 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
1873                 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
1874                 len = fxdr_unsigned(int, *tl);
1875                 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
1876                         m_freem(mrep);
1877                         return (EBADRPC);
1878                 }
1879                 nfsm_dissect(tl, u_int32_t *, (len + 2) * NFSX_UNSIGNED);
1880                 for (i = 1; i <= len; i++)
1881                     if (i < NGROUPS)
1882                         nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
1883                     else
1884                         tl++;
1885                 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
1886                 if (nd->nd_cr.cr_ngroups > 1)
1887                     nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
1888                 len = fxdr_unsigned(int, *++tl);
1889                 if (len < 0 || len > RPCAUTH_MAXSIZ) {
1890                         m_freem(mrep);
1891                         return (EBADRPC);
1892                 }
1893                 if (len > 0)
1894                         nfsm_adv(nfsm_rndup(len));
1895         } else if (auth_type == rpc_auth_kerb) {
1896                 switch (fxdr_unsigned(int, *tl++)) {
1897                 case RPCAKN_FULLNAME:
1898                         ticklen = fxdr_unsigned(int, *tl);
1899                         *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
1900                         uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
1901                         nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
1902                         if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
1903                                 m_freem(mrep);
1904                                 return (EBADRPC);
1905                         }
1906                         uio.uio_offset = 0;
1907                         uio.uio_iov = &iov;
1908                         uio.uio_iovcnt = 1;
1909                         uio.uio_segflg = UIO_SYSSPACE;
1910                         iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
1911                         iov.iov_len = RPCAUTH_MAXSIZ - 4;
1912                         nfsm_mtouio(&uio, uio.uio_resid);
1913                         nfsm_dissect(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
1914                         if (*tl++ != rpc_auth_kerb ||
1915                                 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
1916                                 kprintf("Bad kerb verifier\n");
1917                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
1918                                 nd->nd_procnum = NFSPROC_NOOP;
1919                                 return (0);
1920                         }
1921                         nfsm_dissect(cp, caddr_t, 4 * NFSX_UNSIGNED);
1922                         tl = (u_int32_t *)cp;
1923                         if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
1924                                 kprintf("Not fullname kerb verifier\n");
1925                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
1926                                 nd->nd_procnum = NFSPROC_NOOP;
1927                                 return (0);
1928                         }
1929                         cp += NFSX_UNSIGNED;
1930                         bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
1931                         nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
1932                         nd->nd_flag |= ND_KERBFULL;
1933                         nfsd->nfsd_flag |= NFSD_NEEDAUTH;
1934                         break;
1935                 case RPCAKN_NICKNAME:
1936                         if (len != 2 * NFSX_UNSIGNED) {
1937                                 kprintf("Kerb nickname short\n");
1938                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
1939                                 nd->nd_procnum = NFSPROC_NOOP;
1940                                 return (0);
1941                         }
1942                         nickuid = fxdr_unsigned(uid_t, *tl);
1943                         nfsm_dissect(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
1944                         if (*tl++ != rpc_auth_kerb ||
1945                                 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
1946                                 kprintf("Kerb nick verifier bad\n");
1947                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
1948                                 nd->nd_procnum = NFSPROC_NOOP;
1949                                 return (0);
1950                         }
1951                         nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1952                         tvin.tv_sec = *tl++;
1953                         tvin.tv_usec = *tl;
1954
1955                         for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
1956                             nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1957                                 if (nuidp->nu_cr.cr_uid == nickuid &&
1958                                     (!nd->nd_nam2 ||
1959                                      netaddr_match(NU_NETFAM(nuidp),
1960                                       &nuidp->nu_haddr, nd->nd_nam2)))
1961                                         break;
1962                         }
1963                         if (!nuidp) {
1964                                 nd->nd_repstat =
1965                                         (NFSERR_AUTHERR|AUTH_REJECTCRED);
1966                                 nd->nd_procnum = NFSPROC_NOOP;
1967                                 return (0);
1968                         }
1969
1970                         /*
1971                          * Now, decrypt the timestamp using the session key
1972                          * and validate it.
1973                          */
1974 #ifdef NFSKERB
1975                         XXX
1976 #endif
1977
1978                         tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
1979                         tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
1980                         if (nuidp->nu_expire < time_second ||
1981                             nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
1982                             (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
1983                              nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
1984                                 nuidp->nu_expire = 0;
1985                                 nd->nd_repstat =
1986                                     (NFSERR_AUTHERR|AUTH_REJECTVERF);
1987                                 nd->nd_procnum = NFSPROC_NOOP;
1988                                 return (0);
1989                         }
1990                         nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
1991                         nd->nd_flag |= ND_KERBNICK;
1992                 };
1993         } else {
1994                 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
1995                 nd->nd_procnum = NFSPROC_NOOP;
1996                 return (0);
1997         }
1998
1999         nd->nd_md = md;
2000         nd->nd_dpos = dpos;
2001         return (0);
2002 nfsmout:
2003         return (error);
2004 }
2005
2006 #endif
2007
2008 /*
2009  * Send a message to the originating process's terminal.  The thread and/or
2010  * process may be NULL.  YYY the thread should not be NULL but there may
2011  * still be some uio_td's that are still being passed as NULL through to
2012  * nfsm_request().
2013  */
2014 static int
2015 nfs_msg(struct thread *td, char *server, char *msg)
2016 {
2017         tpr_t tpr;
2018
2019         if (td && td->td_proc)
2020                 tpr = tprintf_open(td->td_proc);
2021         else
2022                 tpr = NULL;
2023         tprintf(tpr, "nfs server %s: %s\n", server, msg);
2024         tprintf_close(tpr);
2025         return (0);
2026 }
2027
2028 #ifndef NFS_NOSERVER
2029 /*
2030  * Socket upcall routine for the nfsd sockets.
2031  * The caddr_t arg is a pointer to the "struct nfssvc_sock".
2032  * Essentially do as much as possible non-blocking, else punt and it will
2033  * be called with MB_WAIT from an nfsd.
2034  */
2035 void
2036 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2037 {
2038         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2039         struct mbuf *m;
2040         struct sockaddr *nam;
2041         struct sockbuf sio;
2042         int flags, error;
2043         int nparallel_wakeup = 0;
2044
2045         if ((slp->ns_flag & SLP_VALID) == 0)
2046                 return;
2047
2048         /*
2049          * Do not allow an infinite number of completed RPC records to build 
2050          * up before we stop reading data from the socket.  Otherwise we could
2051          * end up holding onto an unreasonable number of mbufs for requests
2052          * waiting for service.
2053          *
2054          * This should give pretty good feedback to the TCP
2055          * layer and prevents a memory crunch for other protocols.
2056          *
2057          * Note that the same service socket can be dispatched to several
2058          * nfs servers simultaniously.
2059          *
2060          * the tcp protocol callback calls us with MB_DONTWAIT.  
2061          * nfsd calls us with MB_WAIT (typically).
2062          */
2063         if (waitflag == MB_DONTWAIT && slp->ns_numrec >= nfsd_waiting / 2 + 1) {
2064                 slp->ns_flag |= SLP_NEEDQ;
2065                 goto dorecs;
2066         }
2067
2068         /*
2069          * Handle protocol specifics to parse an RPC request.  We always
2070          * pull from the socket using non-blocking I/O.
2071          */
2072         if (so->so_type == SOCK_STREAM) {
2073                 /*
2074                  * The data has to be read in an orderly fashion from a TCP
2075                  * stream, unlike a UDP socket.  It is possible for soreceive
2076                  * and/or nfsrv_getstream() to block, so make sure only one
2077                  * entity is messing around with the TCP stream at any given
2078                  * moment.  The receive sockbuf's lock in soreceive is not
2079                  * sufficient.
2080                  *
2081                  * Note that this procedure can be called from any number of
2082                  * NFS severs *OR* can be upcalled directly from a TCP
2083                  * protocol thread.
2084                  */
2085                 if (slp->ns_flag & SLP_GETSTREAM) {
2086                         slp->ns_flag |= SLP_NEEDQ;
2087                         goto dorecs;
2088                 }
2089                 slp->ns_flag |= SLP_GETSTREAM;
2090
2091                 /*
2092                  * Do soreceive().  Pull out as much data as possible without
2093                  * blocking.
2094                  */
2095                 sbinit(&sio, 1000000000);
2096                 flags = MSG_DONTWAIT;
2097                 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2098                 if (error || sio.sb_mb == NULL) {
2099                         if (error == EWOULDBLOCK)
2100                                 slp->ns_flag |= SLP_NEEDQ;
2101                         else
2102                                 slp->ns_flag |= SLP_DISCONN;
2103                         slp->ns_flag &= ~SLP_GETSTREAM;
2104                         goto dorecs;
2105                 }
2106                 m = sio.sb_mb;
2107                 if (slp->ns_rawend) {
2108                         slp->ns_rawend->m_next = m;
2109                         slp->ns_cc += sio.sb_cc;
2110                 } else {
2111                         slp->ns_raw = m;
2112                         slp->ns_cc = sio.sb_cc;
2113                 }
2114                 while (m->m_next)
2115                         m = m->m_next;
2116                 slp->ns_rawend = m;
2117
2118                 /*
2119                  * Now try and parse as many record(s) as we can out of the
2120                  * raw stream data.
2121                  */
2122                 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2123                 if (error) {
2124                         if (error == EPERM)
2125                                 slp->ns_flag |= SLP_DISCONN;
2126                         else
2127                                 slp->ns_flag |= SLP_NEEDQ;
2128                 }
2129                 slp->ns_flag &= ~SLP_GETSTREAM;
2130         } else {
2131                 /*
2132                  * For UDP soreceive typically pulls just one packet, loop
2133                  * to get the whole batch.
2134                  */
2135                 do {
2136                         sbinit(&sio, 1000000000);
2137                         flags = MSG_DONTWAIT;
2138                         error = so_pru_soreceive(so, &nam, NULL, &sio,
2139                                                  NULL, &flags);
2140                         if (sio.sb_mb) {
2141                                 struct nfsrv_rec *rec;
2142                                 int mf = (waitflag & MB_DONTWAIT) ?
2143                                             M_NOWAIT : M_WAITOK;
2144                                 rec = kmalloc(sizeof(struct nfsrv_rec),
2145                                              M_NFSRVDESC, mf);
2146                                 if (!rec) {
2147                                         if (nam)
2148                                                 FREE(nam, M_SONAME);
2149                                         m_freem(sio.sb_mb);
2150                                         continue;
2151                                 }
2152                                 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2153                                 rec->nr_address = nam;
2154                                 rec->nr_packet = sio.sb_mb;
2155                                 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2156                                 ++slp->ns_numrec;
2157                                 ++nparallel_wakeup;
2158                         }
2159                         if (error) {
2160                                 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2161                                         && error != EWOULDBLOCK) {
2162                                         slp->ns_flag |= SLP_DISCONN;
2163                                         goto dorecs;
2164                                 }
2165                         }
2166                 } while (sio.sb_mb);
2167         }
2168
2169         /*
2170          * If we were upcalled from the tcp protocol layer and we have
2171          * fully parsed records ready to go, or there is new data pending,
2172          * or something went wrong, try to wake up an nfsd thread to deal
2173          * with it.
2174          */
2175 dorecs:
2176         if (waitflag == MB_DONTWAIT && (slp->ns_numrec > 0
2177              || (slp->ns_flag & (SLP_NEEDQ | SLP_DISCONN)))) {
2178                 nfsrv_wakenfsd(slp, nparallel_wakeup);
2179         }
2180 }
2181
2182 /*
2183  * Try and extract an RPC request from the mbuf data list received on a
2184  * stream socket. The "waitflag" argument indicates whether or not it
2185  * can sleep.
2186  */
2187 static int
2188 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2189 {
2190         struct mbuf *m, **mpp;
2191         char *cp1, *cp2;
2192         int len;
2193         struct mbuf *om, *m2, *recm;
2194         u_int32_t recmark;
2195
2196         for (;;) {
2197             if (slp->ns_reclen == 0) {
2198                 if (slp->ns_cc < NFSX_UNSIGNED)
2199                         return (0);
2200                 m = slp->ns_raw;
2201                 if (m->m_len >= NFSX_UNSIGNED) {
2202                         bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2203                         m->m_data += NFSX_UNSIGNED;
2204                         m->m_len -= NFSX_UNSIGNED;
2205                 } else {
2206                         cp1 = (caddr_t)&recmark;
2207                         cp2 = mtod(m, caddr_t);
2208                         while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2209                                 while (m->m_len == 0) {
2210                                         m = m->m_next;
2211                                         cp2 = mtod(m, caddr_t);
2212                                 }
2213                                 *cp1++ = *cp2++;
2214                                 m->m_data++;
2215                                 m->m_len--;
2216                         }
2217                 }
2218                 slp->ns_cc -= NFSX_UNSIGNED;
2219                 recmark = ntohl(recmark);
2220                 slp->ns_reclen = recmark & ~0x80000000;
2221                 if (recmark & 0x80000000)
2222                         slp->ns_flag |= SLP_LASTFRAG;
2223                 else
2224                         slp->ns_flag &= ~SLP_LASTFRAG;
2225                 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2226                         log(LOG_ERR, "%s (%d) from nfs client\n",
2227                             "impossible packet length",
2228                             slp->ns_reclen);
2229                         return (EPERM);
2230                 }
2231             }
2232
2233             /*
2234              * Now get the record part.
2235              *
2236              * Note that slp->ns_reclen may be 0.  Linux sometimes
2237              * generates 0-length RPCs
2238              */
2239             recm = NULL;
2240             if (slp->ns_cc == slp->ns_reclen) {
2241                 recm = slp->ns_raw;
2242                 slp->ns_raw = slp->ns_rawend = NULL;
2243                 slp->ns_cc = slp->ns_reclen = 0;
2244             } else if (slp->ns_cc > slp->ns_reclen) {
2245                 len = 0;
2246                 m = slp->ns_raw;
2247                 om = NULL;
2248
2249                 while (len < slp->ns_reclen) {
2250                         if ((len + m->m_len) > slp->ns_reclen) {
2251                                 m2 = m_copym(m, 0, slp->ns_reclen - len,
2252                                         waitflag);
2253                                 if (m2) {
2254                                         if (om) {
2255                                                 om->m_next = m2;
2256                                                 recm = slp->ns_raw;
2257                                         } else
2258                                                 recm = m2;
2259                                         m->m_data += slp->ns_reclen - len;
2260                                         m->m_len -= slp->ns_reclen - len;
2261                                         len = slp->ns_reclen;
2262                                 } else {
2263                                         return (EWOULDBLOCK);
2264                                 }
2265                         } else if ((len + m->m_len) == slp->ns_reclen) {
2266                                 om = m;
2267                                 len += m->m_len;
2268                                 m = m->m_next;
2269                                 recm = slp->ns_raw;
2270                                 om->m_next = NULL;
2271                         } else {
2272                                 om = m;
2273                                 len += m->m_len;
2274                                 m = m->m_next;
2275                         }
2276                 }
2277                 slp->ns_raw = m;
2278                 slp->ns_cc -= len;
2279                 slp->ns_reclen = 0;
2280             } else {
2281                 return (0);
2282             }
2283
2284             /*
2285              * Accumulate the fragments into a record.
2286              */
2287             mpp = &slp->ns_frag;
2288             while (*mpp)
2289                 mpp = &((*mpp)->m_next);
2290             *mpp = recm;
2291             if (slp->ns_flag & SLP_LASTFRAG) {
2292                 struct nfsrv_rec *rec;
2293                 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2294                 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2295                 if (!rec) {
2296                     m_freem(slp->ns_frag);
2297                 } else {
2298                     nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2299                     rec->nr_address = NULL;
2300                     rec->nr_packet = slp->ns_frag;
2301                     STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2302                     ++slp->ns_numrec;
2303                     ++*countp;
2304                 }
2305                 slp->ns_frag = NULL;
2306             }
2307         }
2308 }
2309
2310 /*
2311  * Parse an RPC header.
2312  */
2313 int
2314 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2315             struct nfsrv_descript **ndp)
2316 {
2317         struct nfsrv_rec *rec;
2318         struct mbuf *m;
2319         struct sockaddr *nam;
2320         struct nfsrv_descript *nd;
2321         int error;
2322
2323         *ndp = NULL;
2324         if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2325                 return (ENOBUFS);
2326         rec = STAILQ_FIRST(&slp->ns_rec);
2327         STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2328         KKASSERT(slp->ns_numrec > 0);
2329         --slp->ns_numrec;
2330         nam = rec->nr_address;
2331         m = rec->nr_packet;
2332         kfree(rec, M_NFSRVDESC);
2333         MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2334                 M_NFSRVDESC, M_WAITOK);
2335         nd->nd_md = nd->nd_mrep = m;
2336         nd->nd_nam2 = nam;
2337         nd->nd_dpos = mtod(m, caddr_t);
2338         error = nfs_getreq(nd, nfsd, TRUE);
2339         if (error) {
2340                 if (nam) {
2341                         FREE(nam, M_SONAME);
2342                 }
2343                 kfree((caddr_t)nd, M_NFSRVDESC);
2344                 return (error);
2345         }
2346         *ndp = nd;
2347         nfsd->nfsd_nd = nd;
2348         return (0);
2349 }
2350
2351 /*
2352  * Try to assign service sockets to nfsd threads based on the number
2353  * of new rpc requests that have been queued on the service socket.
2354  *
2355  * If no nfsd's are available or additonal requests are pending, set the
2356  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2357  * the work in the nfssvc_sock list when it is finished processing its
2358  * current work.  This flag is only cleared when an nfsd can not find
2359  * any new work to perform.
2360  */
2361 void
2362 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2363 {
2364         struct nfsd *nd;
2365
2366         if ((slp->ns_flag & SLP_VALID) == 0)
2367                 return;
2368         if (nparallel <= 1)
2369                 nparallel = 1;
2370         TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2371                 if (nd->nfsd_flag & NFSD_WAITING) {
2372                         nd->nfsd_flag &= ~NFSD_WAITING;
2373                         if (nd->nfsd_slp)
2374                                 panic("nfsd wakeup");
2375                         slp->ns_sref++;
2376                         nd->nfsd_slp = slp;
2377                         wakeup((caddr_t)nd);
2378                         if (--nparallel == 0)
2379                                 break;
2380                 }
2381         }
2382         if (nparallel) {
2383                 slp->ns_flag |= SLP_DOREC;
2384                 nfsd_head_flag |= NFSD_CHECKSLP;
2385         }
2386 }
2387 #endif /* NFS_NOSERVER */