kernel - Attempt to fix lost rpc issue with NFS timeout/retry
[dragonfly.git] / sys / vfs / nfs / nfs_socket.c
1 /*
2  * Copyright (c) 1989, 1991, 1993, 1995
3  *      The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Rick Macklem at The University of Guelph.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *      This product includes software developed by the University of
19  *      California, Berkeley and its contributors.
20  * 4. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *      @(#)nfs_socket.c        8.5 (Berkeley) 3/30/95
37  * $FreeBSD: src/sys/nfs/nfs_socket.c,v 1.60.2.6 2003/03/26 01:44:46 alfred Exp $
38  * $DragonFly: src/sys/vfs/nfs/nfs_socket.c,v 1.45 2007/05/18 17:05:13 dillon Exp $
39  */
40
41 /*
42  * Socket operations for use by nfs
43  */
44
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/proc.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/kernel.h>
51 #include <sys/mbuf.h>
52 #include <sys/vnode.h>
53 #include <sys/fcntl.h>
54 #include <sys/protosw.h>
55 #include <sys/resourcevar.h>
56 #include <sys/socket.h>
57 #include <sys/socketvar.h>
58 #include <sys/socketops.h>
59 #include <sys/syslog.h>
60 #include <sys/thread.h>
61 #include <sys/tprintf.h>
62 #include <sys/sysctl.h>
63 #include <sys/signalvar.h>
64
65 #include <sys/signal2.h>
66 #include <sys/mutex2.h>
67 #include <sys/socketvar2.h>
68
69 #include <netinet/in.h>
70 #include <netinet/tcp.h>
71 #include <sys/thread2.h>
72
73 #include "rpcv2.h"
74 #include "nfsproto.h"
75 #include "nfs.h"
76 #include "xdr_subs.h"
77 #include "nfsm_subs.h"
78 #include "nfsmount.h"
79 #include "nfsnode.h"
80 #include "nfsrtt.h"
81
82 #define TRUE    1
83 #define FALSE   0
84
85 /*
86  * RTT calculations are scaled by 256 (8 bits).  A proper fractional
87  * RTT will still be calculated even with a slow NFS timer.
88  */
89 #define NFS_SRTT(r)     (r)->r_nmp->nm_srtt[proct[(r)->r_procnum]]
90 #define NFS_SDRTT(r)    (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum]]
91 #define NFS_RTT_SCALE_BITS      8       /* bits */
92 #define NFS_RTT_SCALE           256     /* value */
93
94 /*
95  * Defines which timer to use for the procnum.
96  * 0 - default
97  * 1 - getattr
98  * 2 - lookup
99  * 3 - read
100  * 4 - write
101  */
102 static int proct[NFS_NPROCS] = {
103         0, 1, 0, 2, 1, 3, 3, 4, 0, 0,   /* 00-09        */
104         0, 0, 0, 0, 0, 0, 3, 3, 0, 0,   /* 10-19        */
105         0, 5, 0, 0, 0, 0,               /* 20-29        */
106 };
107
108 static int multt[NFS_NPROCS] = {
109         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 00-09        */
110         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 10-19        */
111         1, 2, 1, 1, 1, 1,               /* 20-29        */
112 };
113
114 static int nfs_backoff[8] = { 2, 3, 5, 8, 13, 21, 34, 55 };
115 static int nfs_realign_test;
116 static int nfs_realign_count;
117 static int nfs_showrtt;
118 static int nfs_showrexmit;
119 int nfs_maxasyncbio = NFS_MAXASYNCBIO;
120
121 SYSCTL_DECL(_vfs_nfs);
122
123 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0, "");
124 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0, "");
125 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrtt, CTLFLAG_RW, &nfs_showrtt, 0, "");
126 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrexmit, CTLFLAG_RW, &nfs_showrexmit, 0, "");
127 SYSCTL_INT(_vfs_nfs, OID_AUTO, maxasyncbio, CTLFLAG_RW, &nfs_maxasyncbio, 0, "");
128
129 static int nfs_request_setup(nfsm_info_t info);
130 static int nfs_request_auth(struct nfsreq *rep);
131 static int nfs_request_try(struct nfsreq *rep);
132 static int nfs_request_waitreply(struct nfsreq *rep);
133 static int nfs_request_processreply(nfsm_info_t info, int);
134
135 int nfsrtton = 0;
136 struct nfsrtt nfsrtt;
137 struct callout  nfs_timer_handle;
138
139 static int      nfs_msg (struct thread *,char *,char *);
140 static int      nfs_rcvlock (struct nfsmount *nmp, struct nfsreq *myreq);
141 static void     nfs_rcvunlock (struct nfsmount *nmp);
142 static void     nfs_realign (struct mbuf **pm, int hsiz);
143 static int      nfs_receive (struct nfsmount *nmp, struct nfsreq *rep,
144                                 struct sockaddr **aname, struct mbuf **mp);
145 static void     nfs_softterm (struct nfsreq *rep, int islocked);
146 static void     nfs_hardterm (struct nfsreq *rep, int islocked);
147 static int      nfs_reconnect (struct nfsmount *nmp, struct nfsreq *rep);
148 #ifndef NFS_NOSERVER 
149 static int      nfsrv_getstream (struct nfssvc_sock *, int, int *);
150 static void     nfs_timer_req(struct nfsreq *req);
151 static void     nfs_checkpkt(struct mbuf *m, int len);
152
153 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
154                                     struct nfssvc_sock *slp,
155                                     struct thread *td,
156                                     struct mbuf **mreqp) = {
157         nfsrv_null,
158         nfsrv_getattr,
159         nfsrv_setattr,
160         nfsrv_lookup,
161         nfsrv3_access,
162         nfsrv_readlink,
163         nfsrv_read,
164         nfsrv_write,
165         nfsrv_create,
166         nfsrv_mkdir,
167         nfsrv_symlink,
168         nfsrv_mknod,
169         nfsrv_remove,
170         nfsrv_rmdir,
171         nfsrv_rename,
172         nfsrv_link,
173         nfsrv_readdir,
174         nfsrv_readdirplus,
175         nfsrv_statfs,
176         nfsrv_fsinfo,
177         nfsrv_pathconf,
178         nfsrv_commit,
179         nfsrv_noop,
180         nfsrv_noop,
181         nfsrv_noop,
182         nfsrv_noop
183 };
184 #endif /* NFS_NOSERVER */
185
186 /*
187  * Initialize sockets and congestion for a new NFS connection.
188  * We do not free the sockaddr if error.
189  */
190 int
191 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
192 {
193         struct socket *so;
194         int error;
195         struct sockaddr *saddr;
196         struct sockaddr_in *sin;
197         struct thread *td = &thread0; /* only used for socreate and sobind */
198
199         nmp->nm_so = so = NULL;
200         if (nmp->nm_flag & NFSMNT_FORCE)
201                 return (EINVAL);
202         saddr = nmp->nm_nam;
203         error = socreate(saddr->sa_family, &so, nmp->nm_sotype,
204                 nmp->nm_soproto, td);
205         if (error)
206                 goto bad;
207         nmp->nm_soflags = so->so_proto->pr_flags;
208
209         /*
210          * Some servers require that the client port be a reserved port number.
211          */
212         if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
213                 struct sockopt sopt;
214                 int ip;
215                 struct sockaddr_in ssin;
216
217                 bzero(&sopt, sizeof sopt);
218                 ip = IP_PORTRANGE_LOW;
219                 sopt.sopt_level = IPPROTO_IP;
220                 sopt.sopt_name = IP_PORTRANGE;
221                 sopt.sopt_val = (void *)&ip;
222                 sopt.sopt_valsize = sizeof(ip);
223                 sopt.sopt_td = NULL;
224                 error = sosetopt(so, &sopt);
225                 if (error)
226                         goto bad;
227                 bzero(&ssin, sizeof ssin);
228                 sin = &ssin;
229                 sin->sin_len = sizeof (struct sockaddr_in);
230                 sin->sin_family = AF_INET;
231                 sin->sin_addr.s_addr = INADDR_ANY;
232                 sin->sin_port = htons(0);
233                 error = sobind(so, (struct sockaddr *)sin, td);
234                 if (error)
235                         goto bad;
236                 bzero(&sopt, sizeof sopt);
237                 ip = IP_PORTRANGE_DEFAULT;
238                 sopt.sopt_level = IPPROTO_IP;
239                 sopt.sopt_name = IP_PORTRANGE;
240                 sopt.sopt_val = (void *)&ip;
241                 sopt.sopt_valsize = sizeof(ip);
242                 sopt.sopt_td = NULL;
243                 error = sosetopt(so, &sopt);
244                 if (error)
245                         goto bad;
246         }
247
248         /*
249          * Protocols that do not require connections may be optionally left
250          * unconnected for servers that reply from a port other than NFS_PORT.
251          */
252         if (nmp->nm_flag & NFSMNT_NOCONN) {
253                 if (nmp->nm_soflags & PR_CONNREQUIRED) {
254                         error = ENOTCONN;
255                         goto bad;
256                 }
257         } else {
258                 error = soconnect(so, nmp->nm_nam, td);
259                 if (error)
260                         goto bad;
261
262                 /*
263                  * Wait for the connection to complete. Cribbed from the
264                  * connect system call but with the wait timing out so
265                  * that interruptible mounts don't hang here for a long time.
266                  */
267                 crit_enter();
268                 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
269                         (void) tsleep((caddr_t)&so->so_timeo, 0,
270                                 "nfscon", 2 * hz);
271                         if ((so->so_state & SS_ISCONNECTING) &&
272                             so->so_error == 0 && rep &&
273                             (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
274                                 soclrstate(so, SS_ISCONNECTING);
275                                 crit_exit();
276                                 goto bad;
277                         }
278                 }
279                 if (so->so_error) {
280                         error = so->so_error;
281                         so->so_error = 0;
282                         crit_exit();
283                         goto bad;
284                 }
285                 crit_exit();
286         }
287         so->so_rcv.ssb_timeo = (5 * hz);
288         so->so_snd.ssb_timeo = (5 * hz);
289
290         /*
291          * Get buffer reservation size from sysctl, but impose reasonable
292          * limits.
293          */
294         if (nmp->nm_sotype == SOCK_STREAM) {
295                 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
296                         struct sockopt sopt;
297                         int val;
298
299                         bzero(&sopt, sizeof sopt);
300                         sopt.sopt_level = SOL_SOCKET;
301                         sopt.sopt_name = SO_KEEPALIVE;
302                         sopt.sopt_val = &val;
303                         sopt.sopt_valsize = sizeof val;
304                         val = 1;
305                         sosetopt(so, &sopt);
306                 }
307                 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
308                         struct sockopt sopt;
309                         int val;
310
311                         bzero(&sopt, sizeof sopt);
312                         sopt.sopt_level = IPPROTO_TCP;
313                         sopt.sopt_name = TCP_NODELAY;
314                         sopt.sopt_val = &val;
315                         sopt.sopt_valsize = sizeof val;
316                         val = 1;
317                         sosetopt(so, &sopt);
318
319                         bzero(&sopt, sizeof sopt);
320                         sopt.sopt_level = IPPROTO_TCP;
321                         sopt.sopt_name = TCP_FASTKEEP;
322                         sopt.sopt_val = &val;
323                         sopt.sopt_valsize = sizeof val;
324                         val = 1;
325                         sosetopt(so, &sopt);
326                 }
327         }
328         error = soreserve(so, nfs_soreserve, nfs_soreserve, NULL);
329         if (error)
330                 goto bad;
331         atomic_set_int(&so->so_rcv.ssb_flags, SSB_NOINTR);
332         atomic_set_int(&so->so_snd.ssb_flags, SSB_NOINTR);
333
334         /* Initialize other non-zero congestion variables */
335         nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] = 
336                 nmp->nm_srtt[3] = (NFS_TIMEO << NFS_RTT_SCALE_BITS);
337         nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
338                 nmp->nm_sdrtt[3] = 0;
339         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
340         nmp->nm_timeouts = 0;
341
342         /*
343          * Assign nm_so last.  The moment nm_so is assigned the nfs_timer()
344          * can mess with the socket.
345          */
346         nmp->nm_so = so;
347         return (0);
348
349 bad:
350         if (so) {
351                 soshutdown(so, SHUT_RDWR);
352                 soclose(so, FNONBLOCK);
353         }
354         return (error);
355 }
356
357 /*
358  * Reconnect routine:
359  * Called when a connection is broken on a reliable protocol.
360  * - clean up the old socket
361  * - nfs_connect() again
362  * - set R_NEEDSXMIT for all outstanding requests on mount point
363  * If this fails the mount point is DEAD!
364  * nb: Must be called with the nfs_sndlock() set on the mount point.
365  */
366 static int
367 nfs_reconnect(struct nfsmount *nmp, struct nfsreq *rep)
368 {
369         struct nfsreq *req;
370         int error;
371
372         nfs_disconnect(nmp);
373         if (nmp->nm_rxstate >= NFSSVC_STOPPING)
374                 return (EINTR);
375         while ((error = nfs_connect(nmp, rep)) != 0) {
376                 if (error == EINTR || error == ERESTART)
377                         return (EINTR);
378                 if (error == EINVAL)
379                         return (error);
380                 if (nmp->nm_rxstate >= NFSSVC_STOPPING)
381                         return (EINTR);
382                 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
383         }
384
385         /*
386          * Loop through outstanding request list and fix up all requests
387          * on old socket.
388          */
389         crit_enter();
390         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
391                 KKASSERT(req->r_nmp == nmp);
392                 req->r_flags |= R_NEEDSXMIT;
393         }
394         crit_exit();
395         return (0);
396 }
397
398 /*
399  * NFS disconnect. Clean up and unlink.
400  */
401 void
402 nfs_disconnect(struct nfsmount *nmp)
403 {
404         struct socket *so;
405
406         if (nmp->nm_so) {
407                 so = nmp->nm_so;
408                 nmp->nm_so = NULL;
409                 soshutdown(so, SHUT_RDWR);
410                 soclose(so, FNONBLOCK);
411         }
412 }
413
414 void
415 nfs_safedisconnect(struct nfsmount *nmp)
416 {
417         nfs_rcvlock(nmp, NULL);
418         nfs_disconnect(nmp);
419         nfs_rcvunlock(nmp);
420 }
421
422 /*
423  * This is the nfs send routine. For connection based socket types, it
424  * must be called with an nfs_sndlock() on the socket.
425  * "rep == NULL" indicates that it has been called from a server.
426  * For the client side:
427  * - return EINTR if the RPC is terminated, 0 otherwise
428  * - set R_NEEDSXMIT if the send fails for any reason
429  * - do any cleanup required by recoverable socket errors (?)
430  * For the server side:
431  * - return EINTR or ERESTART if interrupted by a signal
432  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
433  * - do any cleanup required by recoverable socket errors (?)
434  */
435 int
436 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
437          struct nfsreq *rep)
438 {
439         struct sockaddr *sendnam;
440         int error, soflags, flags;
441
442         if (rep) {
443                 if (rep->r_flags & R_SOFTTERM) {
444                         m_freem(top);
445                         return (EINTR);
446                 }
447                 if ((so = rep->r_nmp->nm_so) == NULL) {
448                         rep->r_flags |= R_NEEDSXMIT;
449                         m_freem(top);
450                         return (0);
451                 }
452                 rep->r_flags &= ~R_NEEDSXMIT;
453                 soflags = rep->r_nmp->nm_soflags;
454         } else {
455                 soflags = so->so_proto->pr_flags;
456         }
457         if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
458                 sendnam = NULL;
459         else
460                 sendnam = nam;
461         if (so->so_type == SOCK_SEQPACKET)
462                 flags = MSG_EOR;
463         else
464                 flags = 0;
465
466         /*
467          * calls pru_sosend -> sosend -> so_pru_send -> netrpc
468          */
469         error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
470                               curthread /*XXX*/);
471
472         /*
473          * ENOBUFS for dgram sockets is transient and non fatal.
474          * No need to log, and no need to break a soft mount.
475          */
476         if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
477                 error = 0;
478                 /*
479                  * do backoff retransmit on client
480                  */
481                 if (rep) {
482                         if ((rep->r_nmp->nm_state & NFSSTA_SENDSPACE) == 0) {
483                                 rep->r_nmp->nm_state |= NFSSTA_SENDSPACE;
484                                 kprintf("Warning: NFS: Insufficient sendspace "
485                                         "(%lu),\n"
486                                         "\t You must increase vfs.nfs.soreserve"
487                                         "or decrease vfs.nfs.maxasyncbio\n",
488                                         so->so_snd.ssb_hiwat);
489                         }
490                         rep->r_flags |= R_NEEDSXMIT;
491                 }
492         }
493
494         if (error) {
495                 if (rep) {
496                         log(LOG_INFO, "nfs send error %d for server %s\n",error,
497                             rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
498                         /*
499                          * Deal with errors for the client side.
500                          */
501                         if (rep->r_flags & R_SOFTTERM)
502                                 error = EINTR;
503                         else
504                                 rep->r_flags |= R_NEEDSXMIT;
505                 } else {
506                         log(LOG_INFO, "nfsd send error %d\n", error);
507                 }
508
509                 /*
510                  * Handle any recoverable (soft) socket errors here. (?)
511                  */
512                 if (error != EINTR && error != ERESTART &&
513                         error != EWOULDBLOCK && error != EPIPE)
514                         error = 0;
515         }
516         return (error);
517 }
518
519 /*
520  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
521  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
522  * Mark and consolidate the data into a new mbuf list.
523  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
524  *     small mbufs.
525  * For SOCK_STREAM we must be very careful to read an entire record once
526  * we have read any of it, even if the system call has been interrupted.
527  */
528 static int
529 nfs_receive(struct nfsmount *nmp, struct nfsreq *rep,
530             struct sockaddr **aname, struct mbuf **mp)
531 {
532         struct socket *so;
533         struct sockbuf sio;
534         struct uio auio;
535         struct iovec aio;
536         struct mbuf *m;
537         struct mbuf *control;
538         u_int32_t len;
539         struct sockaddr **getnam;
540         int error, sotype, rcvflg;
541         struct thread *td = curthread;  /* XXX */
542
543         /*
544          * Set up arguments for soreceive()
545          */
546         *mp = NULL;
547         *aname = NULL;
548         sotype = nmp->nm_sotype;
549
550         /*
551          * For reliable protocols, lock against other senders/receivers
552          * in case a reconnect is necessary.
553          * For SOCK_STREAM, first get the Record Mark to find out how much
554          * more there is to get.
555          * We must lock the socket against other receivers
556          * until we have an entire rpc request/reply.
557          */
558         if (sotype != SOCK_DGRAM) {
559                 error = nfs_sndlock(nmp, rep);
560                 if (error)
561                         return (error);
562 tryagain:
563                 /*
564                  * Check for fatal errors and resending request.
565                  */
566                 /*
567                  * Ugh: If a reconnect attempt just happened, nm_so
568                  * would have changed. NULL indicates a failed
569                  * attempt that has essentially shut down this
570                  * mount point.
571                  */
572                 if (rep && (rep->r_mrep || (rep->r_flags & R_SOFTTERM))) {
573                         nfs_sndunlock(nmp);
574                         return (EINTR);
575                 }
576                 so = nmp->nm_so;
577                 if (so == NULL) {
578                         error = nfs_reconnect(nmp, rep);
579                         if (error) {
580                                 nfs_sndunlock(nmp);
581                                 return (error);
582                         }
583                         goto tryagain;
584                 }
585                 while (rep && (rep->r_flags & R_NEEDSXMIT)) {
586                         m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
587                         nfsstats.rpcretries++;
588                         error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
589                         if (error) {
590                                 if (error == EINTR || error == ERESTART ||
591                                     (error = nfs_reconnect(nmp, rep)) != 0) {
592                                         nfs_sndunlock(nmp);
593                                         return (error);
594                                 }
595                                 goto tryagain;
596                         }
597                 }
598                 nfs_sndunlock(nmp);
599                 if (sotype == SOCK_STREAM) {
600                         /*
601                          * Get the length marker from the stream
602                          */
603                         aio.iov_base = (caddr_t)&len;
604                         aio.iov_len = sizeof(u_int32_t);
605                         auio.uio_iov = &aio;
606                         auio.uio_iovcnt = 1;
607                         auio.uio_segflg = UIO_SYSSPACE;
608                         auio.uio_rw = UIO_READ;
609                         auio.uio_offset = 0;
610                         auio.uio_resid = sizeof(u_int32_t);
611                         auio.uio_td = td;
612                         do {
613                            rcvflg = MSG_WAITALL;
614                            error = so_pru_soreceive(so, NULL, &auio, NULL,
615                                                     NULL, &rcvflg);
616                            if (error == EWOULDBLOCK && rep) {
617                                 if (rep->r_flags & R_SOFTTERM)
618                                         return (EINTR);
619                            }
620                         } while (error == EWOULDBLOCK);
621
622                         if (error == 0 && auio.uio_resid > 0) {
623                             /*
624                              * Only log short packets if not EOF
625                              */
626                             if (auio.uio_resid != sizeof(u_int32_t))
627                             log(LOG_INFO,
628                                  "short receive (%d/%d) from nfs server %s\n",
629                                  (int)(sizeof(u_int32_t) - auio.uio_resid),
630                                  (int)sizeof(u_int32_t),
631                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
632                             error = EPIPE;
633                         }
634                         if (error)
635                                 goto errout;
636                         len = ntohl(len) & ~0x80000000;
637                         /*
638                          * This is SERIOUS! We are out of sync with the sender
639                          * and forcing a disconnect/reconnect is all I can do.
640                          */
641                         if (len > NFS_MAXPACKET) {
642                             log(LOG_ERR, "%s (%d) from nfs server %s\n",
643                                 "impossible packet length",
644                                 len,
645                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
646                             error = EFBIG;
647                             goto errout;
648                         }
649
650                         /*
651                          * Get the rest of the packet as an mbuf chain
652                          */
653                         sbinit(&sio, len);
654                         do {
655                             rcvflg = MSG_WAITALL;
656                             error = so_pru_soreceive(so, NULL, NULL, &sio,
657                                                      NULL, &rcvflg);
658                         } while (error == EWOULDBLOCK || error == EINTR ||
659                                  error == ERESTART);
660                         if (error == 0 && sio.sb_cc != len) {
661                             if (sio.sb_cc != 0)
662                             log(LOG_INFO,
663                                 "short receive (%zu/%d) from nfs server %s\n",
664                                 (size_t)len - auio.uio_resid, len,
665                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
666                             error = EPIPE;
667                         }
668                         *mp = sio.sb_mb;
669                 } else {
670                         /*
671                          * Non-stream, so get the whole packet by not
672                          * specifying MSG_WAITALL and by specifying a large
673                          * length.
674                          *
675                          * We have no use for control msg., but must grab them
676                          * and then throw them away so we know what is going
677                          * on.
678                          */
679                         sbinit(&sio, 100000000);
680                         do {
681                             rcvflg = 0;
682                             error =  so_pru_soreceive(so, NULL, NULL, &sio,
683                                                       &control, &rcvflg);
684                             if (control)
685                                 m_freem(control);
686                             if (error == EWOULDBLOCK && rep) {
687                                 if (rep->r_flags & R_SOFTTERM) {
688                                         m_freem(sio.sb_mb);
689                                         return (EINTR);
690                                 }
691                             }
692                         } while (error == EWOULDBLOCK ||
693                                  (error == 0 && sio.sb_mb == NULL && control));
694                         if ((rcvflg & MSG_EOR) == 0)
695                                 kprintf("Egad!!\n");
696                         if (error == 0 && sio.sb_mb == NULL)
697                                 error = EPIPE;
698                         len = sio.sb_cc;
699                         *mp = sio.sb_mb;
700                 }
701 errout:
702                 if (error && error != EINTR && error != ERESTART) {
703                         m_freem(*mp);
704                         *mp = NULL;
705                         if (error != EPIPE) {
706                                 log(LOG_INFO,
707                                     "receive error %d from nfs server %s\n",
708                                     error,
709                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
710                         }
711                         error = nfs_sndlock(nmp, rep);
712                         if (!error) {
713                                 error = nfs_reconnect(nmp, rep);
714                                 if (!error)
715                                         goto tryagain;
716                                 else
717                                         nfs_sndunlock(nmp);
718                         }
719                 }
720         } else {
721                 if ((so = nmp->nm_so) == NULL)
722                         return (EACCES);
723                 if (so->so_state & SS_ISCONNECTED)
724                         getnam = NULL;
725                 else
726                         getnam = aname;
727                 sbinit(&sio, 100000000);
728                 do {
729                         rcvflg = 0;
730                         error =  so_pru_soreceive(so, getnam, NULL, &sio,
731                                                   NULL, &rcvflg);
732                         if (error == EWOULDBLOCK && rep &&
733                             (rep->r_flags & R_SOFTTERM)) {
734                                 m_freem(sio.sb_mb);
735                                 return (EINTR);
736                         }
737                 } while (error == EWOULDBLOCK);
738
739                 len = sio.sb_cc;
740                 *mp = sio.sb_mb;
741
742                 /*
743                  * A shutdown may result in no error and no mbuf.
744                  * Convert to EPIPE.
745                  */
746                 if (*mp == NULL && error == 0)
747                         error = EPIPE;
748         }
749         if (error) {
750                 m_freem(*mp);
751                 *mp = NULL;
752         }
753
754         /*
755          * Search for any mbufs that are not a multiple of 4 bytes long
756          * or with m_data not longword aligned.
757          * These could cause pointer alignment problems, so copy them to
758          * well aligned mbufs.
759          */
760         nfs_realign(mp, 5 * NFSX_UNSIGNED);
761         return (error);
762 }
763
764 /*
765  * Implement receipt of reply on a socket.
766  *
767  * We must search through the list of received datagrams matching them
768  * with outstanding requests using the xid, until ours is found.
769  *
770  * If myrep is NULL we process packets on the socket until
771  * interrupted or until nm_reqrxq is non-empty.
772  */
773 /* ARGSUSED */
774 int
775 nfs_reply(struct nfsmount *nmp, struct nfsreq *myrep)
776 {
777         struct nfsreq *rep;
778         struct sockaddr *nam;
779         u_int32_t rxid;
780         u_int32_t *tl;
781         int error;
782         struct nfsm_info info;
783
784         /*
785          * Loop around until we get our own reply
786          */
787         for (;;) {
788                 /*
789                  * Lock against other receivers so that I don't get stuck in
790                  * sbwait() after someone else has received my reply for me.
791                  * Also necessary for connection based protocols to avoid
792                  * race conditions during a reconnect.
793                  *
794                  * If nfs_rcvlock() returns EALREADY, that means that
795                  * the reply has already been recieved by another
796                  * process and we can return immediately.  In this
797                  * case, the lock is not taken to avoid races with
798                  * other processes.
799                  */
800                 info.mrep = NULL;
801
802                 error = nfs_rcvlock(nmp, myrep);
803                 if (error == EALREADY)
804                         return (0);
805                 if (error)
806                         return (error);
807
808                 /*
809                  * If myrep is NULL we are the receiver helper thread.
810                  * Stop waiting for incoming replies if there are
811                  * messages sitting on reqrxq that we need to process,
812                  * or if a shutdown request is pending.
813                  */
814                 if (myrep == NULL && (TAILQ_FIRST(&nmp->nm_reqrxq) ||
815                     nmp->nm_rxstate > NFSSVC_PENDING)) {
816                         nfs_rcvunlock(nmp);
817                         return(EWOULDBLOCK);
818                 }
819
820                 /*
821                  * Get the next Rpc reply off the socket
822                  *
823                  * We cannot release the receive lock until we've
824                  * filled in rep->r_mrep, otherwise a waiting
825                  * thread may deadlock in soreceive with no incoming
826                  * packets expected.
827                  */
828                 error = nfs_receive(nmp, myrep, &nam, &info.mrep);
829                 if (error) {
830                         /*
831                          * Ignore routing errors on connectionless protocols??
832                          */
833                         nfs_rcvunlock(nmp);
834                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
835                                 if (nmp->nm_so == NULL)
836                                         return (error);
837                                 nmp->nm_so->so_error = 0;
838                                 continue;
839                         }
840                         return (error);
841                 }
842                 if (nam)
843                         FREE(nam, M_SONAME);
844
845                 /*
846                  * Get the xid and check that it is an rpc reply
847                  */
848                 info.md = info.mrep;
849                 info.dpos = mtod(info.md, caddr_t);
850                 NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
851                 rxid = *tl++;
852                 if (*tl != rpc_reply) {
853                         nfsstats.rpcinvalid++;
854                         m_freem(info.mrep);
855                         info.mrep = NULL;
856 nfsmout:
857                         nfs_rcvunlock(nmp);
858                         continue;
859                 }
860
861                 /*
862                  * Loop through the request list to match up the reply
863                  * Iff no match, just drop the datagram.  On match, set
864                  * r_mrep atomically to prevent the timer from messing
865                  * around with the request after we have exited the critical
866                  * section.
867                  */
868                 crit_enter();
869                 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
870                         if (rep->r_mrep == NULL && rxid == rep->r_xid)
871                                 break;
872                 }
873
874                 /*
875                  * Fill in the rest of the reply if we found a match.
876                  *
877                  * Deal with duplicate responses if there was no match.
878                  */
879                 if (rep) {
880                         rep->r_md = info.md;
881                         rep->r_dpos = info.dpos;
882                         if (nfsrtton) {
883                                 struct rttl *rt;
884
885                                 rt = &nfsrtt.rttl[nfsrtt.pos];
886                                 rt->proc = rep->r_procnum;
887                                 rt->rto = 0;
888                                 rt->sent = 0;
889                                 rt->cwnd = nmp->nm_maxasync_scaled;
890                                 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
891                                 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
892                                 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
893                                 getmicrotime(&rt->tstamp);
894                                 if (rep->r_flags & R_TIMING)
895                                         rt->rtt = rep->r_rtt;
896                                 else
897                                         rt->rtt = 1000000;
898                                 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
899                         }
900
901                         /*
902                          * New congestion control is based only on async
903                          * requests.
904                          */
905                         if (nmp->nm_maxasync_scaled < NFS_MAXASYNC_SCALED)
906                                 ++nmp->nm_maxasync_scaled;
907                         if (rep->r_flags & R_SENT) {
908                                 rep->r_flags &= ~R_SENT;
909                         }
910                         /*
911                          * Update rtt using a gain of 0.125 on the mean
912                          * and a gain of 0.25 on the deviation.
913                          *
914                          * NOTE SRTT/SDRTT are only good if R_TIMING is set.
915                          */
916                         if ((rep->r_flags & R_TIMING) && rep->r_rexmit == 0) {
917                                 /*
918                                  * Since the timer resolution of
919                                  * NFS_HZ is so course, it can often
920                                  * result in r_rtt == 0. Since
921                                  * r_rtt == N means that the actual
922                                  * rtt is between N+dt and N+2-dt ticks,
923                                  * add 1.
924                                  */
925                                 int n;
926                                 int d;
927
928 #define NFSRSB  NFS_RTT_SCALE_BITS
929                                 n = ((NFS_SRTT(rep) * 7) +
930                                      (rep->r_rtt << NFSRSB)) >> 3;
931                                 d = n - NFS_SRTT(rep);
932                                 NFS_SRTT(rep) = n;
933
934                                 /*
935                                  * Don't let the jitter calculation decay
936                                  * too quickly, but we want a fast rampup.
937                                  */
938                                 if (d < 0)
939                                         d = -d;
940                                 d <<= NFSRSB;
941                                 if (d < NFS_SDRTT(rep))
942                                         n = ((NFS_SDRTT(rep) * 15) + d) >> 4;
943                                 else
944                                         n = ((NFS_SDRTT(rep) * 3) + d) >> 2;
945                                 NFS_SDRTT(rep) = n;
946 #undef NFSRSB
947                         }
948                         nmp->nm_timeouts = 0;
949                         rep->r_mrep = info.mrep;
950                         nfs_hardterm(rep, 0);
951                 } else {
952                         /*
953                          * Extract vers, prog, nfsver, procnum.  A duplicate
954                          * response means we didn't wait long enough so
955                          * we increase the SRTT to avoid future spurious
956                          * timeouts.
957                          */
958                         u_int procnum = nmp->nm_lastreprocnum;
959                         int n;
960
961                         if (procnum < NFS_NPROCS && proct[procnum]) {
962                                 if (nfs_showrexmit)
963                                         kprintf("D");
964                                 n = nmp->nm_srtt[proct[procnum]];
965                                 n += NFS_ASYSCALE * NFS_HZ;
966                                 if (n < NFS_ASYSCALE * NFS_HZ * 10)
967                                         n = NFS_ASYSCALE * NFS_HZ * 10;
968                                 nmp->nm_srtt[proct[procnum]] = n;
969                         }
970                 }
971                 nfs_rcvunlock(nmp);
972                 crit_exit();
973
974                 /*
975                  * If not matched to a request, drop it.
976                  * If it's mine, get out.
977                  */
978                 if (rep == NULL) {
979                         nfsstats.rpcunexpected++;
980                         m_freem(info.mrep);
981                         info.mrep = NULL;
982                 } else if (rep == myrep) {
983                         if (rep->r_mrep == NULL)
984                                 panic("nfsreply nil");
985                         return (0);
986                 }
987         }
988 }
989
990 /*
991  * Run the request state machine until the target state is reached
992  * or a fatal error occurs.  The target state is not run.  Specifying
993  * a target of NFSM_STATE_DONE runs the state machine until the rpc
994  * is complete.
995  *
996  * EINPROGRESS is returned for all states other then the DONE state,
997  * indicating that the rpc is still in progress.
998  */
999 int
1000 nfs_request(struct nfsm_info *info, nfsm_state_t bstate, nfsm_state_t estate)
1001 {
1002         struct nfsreq *req;
1003
1004         while (info->state >= bstate && info->state < estate) {
1005                 switch(info->state) {
1006                 case NFSM_STATE_SETUP:
1007                         /*
1008                          * Setup the nfsreq.  Any error which occurs during
1009                          * this state is fatal.
1010                          */
1011                         info->error = nfs_request_setup(info);
1012                         if (info->error) {
1013                                 info->state = NFSM_STATE_DONE;
1014                                 return (info->error);
1015                         } else {
1016                                 req = info->req;
1017                                 req->r_mrp = &info->mrep;
1018                                 req->r_mdp = &info->md;
1019                                 req->r_dposp = &info->dpos;
1020                                 info->state = NFSM_STATE_AUTH;
1021                         }
1022                         break;
1023                 case NFSM_STATE_AUTH:
1024                         /*
1025                          * Authenticate the nfsreq.  Any error which occurs
1026                          * during this state is fatal.
1027                          */
1028                         info->error = nfs_request_auth(info->req);
1029                         if (info->error) {
1030                                 info->state = NFSM_STATE_DONE;
1031                                 return (info->error);
1032                         } else {
1033                                 info->state = NFSM_STATE_TRY;
1034                         }
1035                         break;
1036                 case NFSM_STATE_TRY:
1037                         /*
1038                          * Transmit or retransmit attempt.  An error in this
1039                          * state is ignored and we always move on to the
1040                          * next state.
1041                          *
1042                          * This can trivially race the receiver if the
1043                          * request is asynchronous.  nfs_request_try()
1044                          * will thus set the state for us and we
1045                          * must also return immediately if we are
1046                          * running an async state machine, because
1047                          * info can become invalid due to races after
1048                          * try() returns.
1049                          */
1050                         if (info->req->r_flags & R_ASYNC) {
1051                                 nfs_request_try(info->req);
1052                                 if (estate == NFSM_STATE_WAITREPLY)
1053                                         return (EINPROGRESS);
1054                         } else {
1055                                 nfs_request_try(info->req);
1056                                 info->state = NFSM_STATE_WAITREPLY;
1057                         }
1058                         break;
1059                 case NFSM_STATE_WAITREPLY:
1060                         /*
1061                          * Wait for a reply or timeout and move on to the
1062                          * next state.  The error returned by this state
1063                          * is passed to the processing code in the next
1064                          * state.
1065                          */
1066                         info->error = nfs_request_waitreply(info->req);
1067                         info->state = NFSM_STATE_PROCESSREPLY;
1068                         break;
1069                 case NFSM_STATE_PROCESSREPLY:
1070                         /*
1071                          * Process the reply or timeout.  Errors which occur
1072                          * in this state may cause the state machine to
1073                          * go back to an earlier state, and are fatal
1074                          * otherwise.
1075                          */
1076                         info->error = nfs_request_processreply(info,
1077                                                                info->error);
1078                         switch(info->error) {
1079                         case ENEEDAUTH:
1080                                 info->state = NFSM_STATE_AUTH;
1081                                 break;
1082                         case EAGAIN:
1083                                 info->state = NFSM_STATE_TRY;
1084                                 break;
1085                         default:
1086                                 /*
1087                                  * Operation complete, with or without an
1088                                  * error.  We are done.
1089                                  */
1090                                 info->req = NULL;
1091                                 info->state = NFSM_STATE_DONE;
1092                                 return (info->error);
1093                         }
1094                         break;
1095                 case NFSM_STATE_DONE:
1096                         /*
1097                          * Shouldn't be reached
1098                          */
1099                         return (info->error);
1100                         /* NOT REACHED */
1101                 }
1102         }
1103
1104         /*
1105          * If we are done return the error code (if any).
1106          * Otherwise return EINPROGRESS.
1107          */
1108         if (info->state == NFSM_STATE_DONE)
1109                 return (info->error);
1110         return (EINPROGRESS);
1111 }
1112
1113 /*
1114  * nfs_request - goes something like this
1115  *      - fill in request struct
1116  *      - links it into list
1117  *      - calls nfs_send() for first transmit
1118  *      - calls nfs_receive() to get reply
1119  *      - break down rpc header and return with nfs reply pointed to
1120  *        by mrep or error
1121  * nb: always frees up mreq mbuf list
1122  */
1123 static int
1124 nfs_request_setup(nfsm_info_t info)
1125 {
1126         struct nfsreq *req;
1127         struct nfsmount *nmp;
1128         struct mbuf *m;
1129         int i;
1130
1131         /*
1132          * Reject requests while attempting a forced unmount.
1133          */
1134         if (info->vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
1135                 m_freem(info->mreq);
1136                 info->mreq = NULL;
1137                 return (ESTALE);
1138         }
1139         nmp = VFSTONFS(info->vp->v_mount);
1140         req = kmalloc(sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
1141         req->r_nmp = nmp;
1142         req->r_vp = info->vp;
1143         req->r_td = info->td;
1144         req->r_procnum = info->procnum;
1145         req->r_mreq = NULL;
1146         req->r_cred = info->cred;
1147
1148         i = 0;
1149         m = info->mreq;
1150         while (m) {
1151                 i += m->m_len;
1152                 m = m->m_next;
1153         }
1154         req->r_mrest = info->mreq;
1155         req->r_mrest_len = i;
1156
1157         /*
1158          * The presence of a non-NULL r_info in req indicates
1159          * async completion via our helper threads.  See the receiver
1160          * code.
1161          */
1162         if (info->bio) {
1163                 req->r_info = info;
1164                 req->r_flags = R_ASYNC;
1165         } else {
1166                 req->r_info = NULL;
1167                 req->r_flags = 0;
1168         }
1169         info->req = req;
1170         return(0);
1171 }
1172
1173 static int
1174 nfs_request_auth(struct nfsreq *rep)
1175 {
1176         struct nfsmount *nmp = rep->r_nmp;
1177         struct mbuf *m;
1178         char nickv[RPCX_NICKVERF];
1179         int error = 0, auth_len, auth_type;
1180         int verf_len;
1181         u_int32_t xid;
1182         char *auth_str, *verf_str;
1183         struct ucred *cred;
1184
1185         cred = rep->r_cred;
1186         rep->r_failed_auth = 0;
1187
1188         /*
1189          * Get the RPC header with authorization.
1190          */
1191         verf_str = auth_str = NULL;
1192         if (nmp->nm_flag & NFSMNT_KERB) {
1193                 verf_str = nickv;
1194                 verf_len = sizeof (nickv);
1195                 auth_type = RPCAUTH_KERB4;
1196                 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1197                 if (rep->r_failed_auth ||
1198                     nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1199                                     verf_str, verf_len)) {
1200                         error = nfs_getauth(nmp, rep, cred, &auth_str,
1201                                 &auth_len, verf_str, &verf_len, rep->r_key);
1202                         if (error) {
1203                                 m_freem(rep->r_mrest);
1204                                 rep->r_mrest = NULL;
1205                                 kfree((caddr_t)rep, M_NFSREQ);
1206                                 return (error);
1207                         }
1208                 }
1209         } else {
1210                 auth_type = RPCAUTH_UNIX;
1211                 if (cred->cr_ngroups < 1)
1212                         panic("nfsreq nogrps");
1213                 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1214                         nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1215                         5 * NFSX_UNSIGNED;
1216         }
1217         if (rep->r_mrest)
1218                 nfs_checkpkt(rep->r_mrest, rep->r_mrest_len);
1219         m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1220                         auth_len, auth_str, verf_len, verf_str,
1221                         rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1222         rep->r_mrest = NULL;
1223         if (auth_str)
1224                 kfree(auth_str, M_TEMP);
1225
1226         /*
1227          * For stream protocols, insert a Sun RPC Record Mark.
1228          */
1229         if (nmp->nm_sotype == SOCK_STREAM) {
1230                 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1231                 if (m == NULL) {
1232                         kfree(rep, M_NFSREQ);
1233                         return (ENOBUFS);
1234                 }
1235                 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1236                          (m->m_pkthdr.len - NFSX_UNSIGNED));
1237         }
1238
1239         nfs_checkpkt(m, m->m_pkthdr.len);
1240
1241         rep->r_mreq = m;
1242         rep->r_xid = xid;
1243         return (0);
1244 }
1245
1246 static int
1247 nfs_request_try(struct nfsreq *rep)
1248 {
1249         struct nfsmount *nmp = rep->r_nmp;
1250         struct mbuf *m2;
1251         int error;
1252
1253         /*
1254          * Request is not on any queue, only the owner has access to it
1255          * so it should not be locked by anyone atm.
1256          *
1257          * Interlock to prevent races.  While locked the only remote
1258          * action possible is for r_mrep to be set (once we enqueue it).
1259          */
1260         if (rep->r_flags == 0xdeadc0de) {
1261                 print_backtrace(-1);
1262                 panic("flags nbad\n");
1263         }
1264         KKASSERT((rep->r_flags & (R_LOCKED | R_ONREQQ)) == 0);
1265         if (nmp->nm_flag & NFSMNT_SOFT)
1266                 rep->r_retry = nmp->nm_retry;
1267         else
1268                 rep->r_retry = NFS_MAXREXMIT + 1;       /* past clip limit */
1269         rep->r_rtt = rep->r_rexmit = 0;
1270         if (proct[rep->r_procnum] > 0)
1271                 rep->r_flags |= R_TIMING | R_LOCKED;
1272         else
1273                 rep->r_flags |= R_LOCKED;
1274         rep->r_mrep = NULL;
1275
1276         nfsstats.rpcrequests++;
1277
1278         if (nmp->nm_flag & NFSMNT_FORCE) {
1279                 rep->r_flags |= R_SOFTTERM;
1280                 rep->r_flags &= ~R_LOCKED;
1281                 return (0);
1282         }
1283         rep->r_flags |= R_NEEDSXMIT;    /* in case send lock races us */
1284
1285         /*
1286          * Do the client side RPC.
1287          *
1288          * Chain request into list of outstanding requests. Be sure
1289          * to put it LAST so timer finds oldest requests first.  Note
1290          * that our control of R_LOCKED prevents the request from
1291          * getting ripped out from under us or transmitted by the
1292          * timer code.
1293          *
1294          * For requests with info structures we must atomically set the
1295          * info's state because the structure could become invalid upon
1296          * return due to races (i.e., if async)
1297          */
1298         crit_enter();
1299         mtx_link_init(&rep->r_link);
1300         KKASSERT((rep->r_flags & R_ONREQQ) == 0);
1301         TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1302         rep->r_flags |= R_ONREQQ;
1303         ++nmp->nm_reqqlen;
1304         if (rep->r_flags & R_ASYNC)
1305                 rep->r_info->state = NFSM_STATE_WAITREPLY;
1306         crit_exit();
1307
1308         error = 0;
1309
1310         /*
1311          * Send if we can.  Congestion control is not handled here any more
1312          * becausing trying to defer the initial send based on the nfs_timer
1313          * requires having a very fast nfs_timer, which is silly.
1314          */
1315         if (nmp->nm_so) {
1316                 if (nmp->nm_soflags & PR_CONNREQUIRED)
1317                         error = nfs_sndlock(nmp, rep);
1318                 if (error == 0 && (rep->r_flags & R_NEEDSXMIT)) {
1319                         m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1320                         error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1321                         rep->r_flags &= ~R_NEEDSXMIT;
1322                         if ((rep->r_flags & R_SENT) == 0) {
1323                                 rep->r_flags |= R_SENT;
1324                         }
1325                         if (nmp->nm_soflags & PR_CONNREQUIRED)
1326                                 nfs_sndunlock(nmp);
1327                 }
1328         } else {
1329                 rep->r_rtt = -1;
1330         }
1331         if (error == EPIPE)
1332                 error = 0;
1333
1334         /*
1335          * Release the lock.  The only remote action that may have occurred
1336          * would have been the setting of rep->r_mrep.  If this occured
1337          * and the request was async we have to move it to the reader
1338          * thread's queue for action.
1339          *
1340          * For async requests also make sure the reader is woken up so
1341          * it gets on the socket to read responses.
1342          */
1343         crit_enter();
1344         if (rep->r_flags & R_ASYNC) {
1345                 if (rep->r_mrep)
1346                         nfs_hardterm(rep, 1);
1347                 rep->r_flags &= ~R_LOCKED;
1348                 nfssvc_iod_reader_wakeup(nmp);
1349         } else {
1350                 rep->r_flags &= ~R_LOCKED;
1351         }
1352         if (rep->r_flags & R_WANTED) {
1353                 rep->r_flags &= ~R_WANTED;
1354                 wakeup(rep);
1355         }
1356         crit_exit();
1357         return (error);
1358 }
1359
1360 /*
1361  * This code is only called for synchronous requests.  Completed synchronous
1362  * requests are left on reqq and we remove them before moving on to the
1363  * processing state.
1364  */
1365 static int
1366 nfs_request_waitreply(struct nfsreq *rep)
1367 {
1368         struct nfsmount *nmp = rep->r_nmp;
1369         int error;
1370
1371         KKASSERT((rep->r_flags & R_ASYNC) == 0);
1372
1373         /*
1374          * Wait until the request is finished.
1375          */
1376         error = nfs_reply(nmp, rep);
1377
1378         /*
1379          * RPC done, unlink the request, but don't rip it out from under
1380          * the callout timer.
1381          *
1382          * Once unlinked no other receiver or the timer will have
1383          * visibility, so we do not have to set R_LOCKED.
1384          */
1385         crit_enter();
1386         while (rep->r_flags & R_LOCKED) {
1387                 rep->r_flags |= R_WANTED;
1388                 tsleep(rep, 0, "nfstrac", 0);
1389         }
1390         KKASSERT(rep->r_flags & R_ONREQQ);
1391         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1392         rep->r_flags &= ~R_ONREQQ;
1393         --nmp->nm_reqqlen;
1394         if (TAILQ_FIRST(&nmp->nm_bioq) &&
1395             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1396                 nfssvc_iod_writer_wakeup(nmp);
1397         }
1398         crit_exit();
1399
1400         /*
1401          * Decrement the outstanding request count.
1402          */
1403         if (rep->r_flags & R_SENT) {
1404                 rep->r_flags &= ~R_SENT;
1405         }
1406         return (error);
1407 }
1408
1409 /*
1410  * Process reply with error returned from nfs_requet_waitreply().
1411  *
1412  * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1413  * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1414  */
1415 static int
1416 nfs_request_processreply(nfsm_info_t info, int error)
1417 {
1418         struct nfsreq *req = info->req;
1419         struct nfsmount *nmp = req->r_nmp;
1420         u_int32_t *tl;
1421         int verf_type;
1422         int i;
1423
1424         /*
1425          * If there was a successful reply and a tprintf msg.
1426          * tprintf a response.
1427          */
1428         if (error == 0 && (req->r_flags & R_TPRINTFMSG)) {
1429                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1430                     "is alive again");
1431         }
1432         info->mrep = req->r_mrep;
1433         info->md = req->r_md;
1434         info->dpos = req->r_dpos;
1435         if (error) {
1436                 m_freem(req->r_mreq);
1437                 req->r_mreq = NULL;
1438                 kfree(req, M_NFSREQ);
1439                 info->req = NULL;
1440                 return (error);
1441         }
1442
1443         /*
1444          * break down the rpc header and check if ok
1445          */
1446         NULLOUT(tl = nfsm_dissect(info, 3 * NFSX_UNSIGNED));
1447         if (*tl++ == rpc_msgdenied) {
1448                 if (*tl == rpc_mismatch) {
1449                         error = EOPNOTSUPP;
1450                 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1451                            *tl++ == rpc_autherr) {
1452                         if (req->r_failed_auth == 0) {
1453                                 req->r_failed_auth++;
1454                                 req->r_mheadend->m_next = NULL;
1455                                 m_freem(info->mrep);
1456                                 info->mrep = NULL;
1457                                 m_freem(req->r_mreq);
1458                                 req->r_mreq = NULL;
1459                                 return (ENEEDAUTH);
1460                         } else {
1461                                 error = EAUTH;
1462                         }
1463                 } else {
1464                         error = EACCES;
1465                 }
1466                 m_freem(info->mrep);
1467                 info->mrep = NULL;
1468                 m_freem(req->r_mreq);
1469                 req->r_mreq = NULL;
1470                 kfree(req, M_NFSREQ);
1471                 info->req = NULL;
1472                 return (error);
1473         }
1474
1475         /*
1476          * Grab any Kerberos verifier, otherwise just throw it away.
1477          */
1478         verf_type = fxdr_unsigned(int, *tl++);
1479         i = fxdr_unsigned(int32_t, *tl);
1480         if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1481                 error = nfs_savenickauth(nmp, req->r_cred, i, req->r_key,
1482                                          &info->md, &info->dpos, info->mrep);
1483                 if (error)
1484                         goto nfsmout;
1485         } else if (i > 0) {
1486                 ERROROUT(nfsm_adv(info, nfsm_rndup(i)));
1487         }
1488         NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1489         /* 0 == ok */
1490         if (*tl == 0) {
1491                 NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1492                 if (*tl != 0) {
1493                         error = fxdr_unsigned(int, *tl);
1494
1495                         /*
1496                          * Does anyone even implement this?  Just impose
1497                          * a 1-second delay.
1498                          */
1499                         if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1500                                 error == NFSERR_TRYLATER) {
1501                                 m_freem(info->mrep);
1502                                 info->mrep = NULL;
1503                                 error = 0;
1504
1505                                 tsleep((caddr_t)&lbolt, 0, "nqnfstry", 0);
1506                                 return (EAGAIN);        /* goto tryagain */
1507                         }
1508
1509                         /*
1510                          * If the File Handle was stale, invalidate the
1511                          * lookup cache, just in case.
1512                          *
1513                          * To avoid namecache<->vnode deadlocks we must
1514                          * release the vnode lock if we hold it.
1515                          */
1516                         if (error == ESTALE) {
1517                                 struct vnode *vp = req->r_vp;
1518                                 int ltype;
1519
1520                                 ltype = lockstatus(&vp->v_lock, curthread);
1521                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1522                                         lockmgr(&vp->v_lock, LK_RELEASE);
1523                                 cache_inval_vp(vp, CINV_CHILDREN);
1524                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1525                                         lockmgr(&vp->v_lock, ltype);
1526                         }
1527                         if (nmp->nm_flag & NFSMNT_NFSV3) {
1528                                 KKASSERT(*req->r_mrp == info->mrep);
1529                                 KKASSERT(*req->r_mdp == info->md);
1530                                 KKASSERT(*req->r_dposp == info->dpos);
1531                                 error |= NFSERR_RETERR;
1532                         } else {
1533                                 m_freem(info->mrep);
1534                                 info->mrep = NULL;
1535                         }
1536                         m_freem(req->r_mreq);
1537                         req->r_mreq = NULL;
1538                         kfree(req, M_NFSREQ);
1539                         info->req = NULL;
1540                         return (error);
1541                 }
1542
1543                 KKASSERT(*req->r_mrp == info->mrep);
1544                 KKASSERT(*req->r_mdp == info->md);
1545                 KKASSERT(*req->r_dposp == info->dpos);
1546                 m_freem(req->r_mreq);
1547                 req->r_mreq = NULL;
1548                 FREE(req, M_NFSREQ);
1549                 return (0);
1550         }
1551         m_freem(info->mrep);
1552         info->mrep = NULL;
1553         error = EPROTONOSUPPORT;
1554 nfsmout:
1555         m_freem(req->r_mreq);
1556         req->r_mreq = NULL;
1557         kfree(req, M_NFSREQ);
1558         info->req = NULL;
1559         return (error);
1560 }
1561
1562 #ifndef NFS_NOSERVER
1563 /*
1564  * Generate the rpc reply header
1565  * siz arg. is used to decide if adding a cluster is worthwhile
1566  */
1567 int
1568 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1569             int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1570 {
1571         u_int32_t *tl;
1572         struct nfsm_info info;
1573
1574         siz += RPC_REPLYSIZ;
1575         info.mb = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1576         info.mreq = info.mb;
1577         info.mreq->m_pkthdr.len = 0;
1578         /*
1579          * If this is not a cluster, try and leave leading space
1580          * for the lower level headers.
1581          */
1582         if ((max_hdr + siz) < MINCLSIZE)
1583                 info.mreq->m_data += max_hdr;
1584         tl = mtod(info.mreq, u_int32_t *);
1585         info.mreq->m_len = 6 * NFSX_UNSIGNED;
1586         info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1587         *tl++ = txdr_unsigned(nd->nd_retxid);
1588         *tl++ = rpc_reply;
1589         if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1590                 *tl++ = rpc_msgdenied;
1591                 if (err & NFSERR_AUTHERR) {
1592                         *tl++ = rpc_autherr;
1593                         *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1594                         info.mreq->m_len -= NFSX_UNSIGNED;
1595                         info.bpos -= NFSX_UNSIGNED;
1596                 } else {
1597                         *tl++ = rpc_mismatch;
1598                         *tl++ = txdr_unsigned(RPC_VER2);
1599                         *tl = txdr_unsigned(RPC_VER2);
1600                 }
1601         } else {
1602                 *tl++ = rpc_msgaccepted;
1603
1604                 /*
1605                  * For Kerberos authentication, we must send the nickname
1606                  * verifier back, otherwise just RPCAUTH_NULL.
1607                  */
1608                 if (nd->nd_flag & ND_KERBFULL) {
1609                     struct nfsuid *nuidp;
1610                     struct timeval ktvin, ktvout;
1611
1612                     for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1613                         nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1614                         if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1615                             (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1616                              &nuidp->nu_haddr, nd->nd_nam2)))
1617                             break;
1618                     }
1619                     if (nuidp) {
1620                         ktvin.tv_sec =
1621                             txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1622                         ktvin.tv_usec =
1623                             txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1624
1625                         /*
1626                          * Encrypt the timestamp in ecb mode using the
1627                          * session key.
1628                          */
1629 #ifdef NFSKERB
1630                         XXX
1631 #else
1632                         ktvout.tv_sec = 0;
1633                         ktvout.tv_usec = 0;
1634 #endif
1635
1636                         *tl++ = rpc_auth_kerb;
1637                         *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1638                         *tl = ktvout.tv_sec;
1639                         tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1640                         *tl++ = ktvout.tv_usec;
1641                         *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1642                     } else {
1643                         *tl++ = 0;
1644                         *tl++ = 0;
1645                     }
1646                 } else {
1647                         *tl++ = 0;
1648                         *tl++ = 0;
1649                 }
1650                 switch (err) {
1651                 case EPROGUNAVAIL:
1652                         *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1653                         break;
1654                 case EPROGMISMATCH:
1655                         *tl = txdr_unsigned(RPC_PROGMISMATCH);
1656                         tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1657                         *tl++ = txdr_unsigned(2);
1658                         *tl = txdr_unsigned(3);
1659                         break;
1660                 case EPROCUNAVAIL:
1661                         *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1662                         break;
1663                 case EBADRPC:
1664                         *tl = txdr_unsigned(RPC_GARBAGE);
1665                         break;
1666                 default:
1667                         *tl = 0;
1668                         if (err != NFSERR_RETVOID) {
1669                                 tl = nfsm_build(&info, NFSX_UNSIGNED);
1670                                 if (err)
1671                                     *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1672                                 else
1673                                     *tl = 0;
1674                         }
1675                         break;
1676                 };
1677         }
1678
1679         if (mrq != NULL)
1680             *mrq = info.mreq;
1681         *mbp = info.mb;
1682         *bposp = info.bpos;
1683         if (err != 0 && err != NFSERR_RETVOID)
1684                 nfsstats.srvrpc_errs++;
1685         return (0);
1686 }
1687
1688
1689 #endif /* NFS_NOSERVER */
1690
1691 /*
1692  * Nfs timer routine.
1693  *
1694  * Scan the nfsreq list and retranmit any requests that have timed out
1695  * To avoid retransmission attempts on STREAM sockets (in the future) make
1696  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1697  *
1698  * Requests with attached responses, terminated requests, and
1699  * locked requests are ignored.  Locked requests will be picked up
1700  * in a later timer call.
1701  */
1702 void
1703 nfs_timer_callout(void *arg /* never used */)
1704 {
1705         struct nfsmount *nmp;
1706         struct nfsreq *req;
1707 #ifndef NFS_NOSERVER
1708         struct nfssvc_sock *slp;
1709         u_quad_t cur_usec;
1710 #endif /* NFS_NOSERVER */
1711
1712         lwkt_gettoken(&nfs_token);
1713         TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1714                 lwkt_gettoken(&nmp->nm_token);
1715                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1716                         KKASSERT(nmp == req->r_nmp);
1717                         if (req->r_mrep)
1718                                 continue;
1719                         if (req->r_flags & (R_SOFTTERM | R_LOCKED))
1720                                 continue;
1721
1722                         /*
1723                          * Handle timeout/retry.  Be sure to process r_mrep
1724                          * for async requests that completed while we had
1725                          * the request locked or they will hang in the reqq
1726                          * forever.
1727                          */
1728                         req->r_flags |= R_LOCKED;
1729                         if (nfs_sigintr(nmp, req, req->r_td)) {
1730                                 nfs_softterm(req, 1);
1731                                 req->r_flags &= ~R_LOCKED;
1732                         } else {
1733                                 nfs_timer_req(req);
1734                                 if (req->r_flags & R_ASYNC) {
1735                                         if (req->r_mrep)
1736                                                 nfs_hardterm(req, 1);
1737                                         req->r_flags &= ~R_LOCKED;
1738                                         nfssvc_iod_reader_wakeup(nmp);
1739                                 } else {
1740                                         req->r_flags &= ~R_LOCKED;
1741                                 }
1742                         }
1743                         if (req->r_flags & R_WANTED) {
1744                                 req->r_flags &= ~R_WANTED;
1745                                 wakeup(req);
1746                         }
1747                 }
1748                 lwkt_reltoken(&nmp->nm_token);
1749         }
1750 #ifndef NFS_NOSERVER
1751
1752         /*
1753          * Scan the write gathering queues for writes that need to be
1754          * completed now.
1755          */
1756         cur_usec = nfs_curusec();
1757
1758         TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1759                 /* XXX race against removal */
1760                 if (lwkt_trytoken(&slp->ns_token)) {
1761                         if (slp->ns_tq.lh_first &&
1762                             (slp->ns_tq.lh_first->nd_time <= cur_usec)) {
1763                                 nfsrv_wakenfsd(slp, 1);
1764                         }
1765                         lwkt_reltoken(&slp->ns_token);
1766                 }
1767         }
1768 #endif /* NFS_NOSERVER */
1769
1770         callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer_callout, NULL);
1771         lwkt_reltoken(&nfs_token);
1772 }
1773
1774 static
1775 void
1776 nfs_timer_req(struct nfsreq *req)
1777 {
1778         struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1779         struct nfsmount *nmp = req->r_nmp;
1780         struct mbuf *m;
1781         struct socket *so;
1782         int timeo;
1783         int error;
1784
1785         /*
1786          * rtt ticks and timeout calculation.  Return if the timeout
1787          * has not been reached yet, unless the packet is flagged
1788          * for an immediate send.
1789          *
1790          * The mean rtt doesn't help when we get random I/Os, we have
1791          * to multiply by fairly large numbers.
1792          */
1793         if (req->r_rtt >= 0) {
1794                 /*
1795                  * Calculate the timeout to test against.
1796                  */
1797                 req->r_rtt++;
1798                 if (nmp->nm_flag & NFSMNT_DUMBTIMR) {
1799                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1800                 } else if (req->r_flags & R_TIMING) {
1801                         timeo = NFS_SRTT(req) + NFS_SDRTT(req);
1802                 } else {
1803                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1804                 }
1805                 timeo *= multt[req->r_procnum];
1806                 /* timeo is still scaled by SCALE_BITS */
1807
1808 #define NFSFS   (NFS_RTT_SCALE * NFS_HZ)
1809                 if (req->r_flags & R_TIMING) {
1810                         static long last_time;
1811                         if (nfs_showrtt && last_time != time_second) {
1812                                 kprintf("rpccmd %d NFS SRTT %d SDRTT %d "
1813                                         "timeo %d.%03d\n",
1814                                         proct[req->r_procnum],
1815                                         NFS_SRTT(req), NFS_SDRTT(req),
1816                                         timeo / NFSFS,
1817                                         timeo % NFSFS * 1000 /  NFSFS);
1818                                 last_time = time_second;
1819                         }
1820                 }
1821 #undef NFSFS
1822
1823                 /*
1824                  * deal with nfs_timer jitter.
1825                  */
1826                 timeo = (timeo >> NFS_RTT_SCALE_BITS) + 1;
1827                 if (timeo < 2)
1828                         timeo = 2;
1829
1830                 if (nmp->nm_timeouts > 0)
1831                         timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1832                 if (timeo > NFS_MAXTIMEO)
1833                         timeo = NFS_MAXTIMEO;
1834                 if (req->r_rtt <= timeo) {
1835                         if ((req->r_flags & R_NEEDSXMIT) == 0)
1836                                 return;
1837                 } else if (nmp->nm_timeouts < 8) {
1838                         nmp->nm_timeouts++;
1839                 }
1840         }
1841
1842         /*
1843          * Check for server not responding
1844          */
1845         if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1846              req->r_rexmit > nmp->nm_deadthresh) {
1847                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1848                         "not responding");
1849                 req->r_flags |= R_TPRINTFMSG;
1850         }
1851         if (req->r_rexmit >= req->r_retry) {    /* too many */
1852                 nfsstats.rpctimeouts++;
1853                 nfs_softterm(req, 1);
1854                 return;
1855         }
1856
1857         /*
1858          * Generally disable retransmission on reliable sockets,
1859          * unless the request is flagged for immediate send.
1860          */
1861         if (nmp->nm_sotype != SOCK_DGRAM) {
1862                 if (++req->r_rexmit > NFS_MAXREXMIT)
1863                         req->r_rexmit = NFS_MAXREXMIT;
1864                 if ((req->r_flags & R_NEEDSXMIT) == 0)
1865                         return;
1866         }
1867
1868         /*
1869          * Stop here if we do not have a socket!
1870          */
1871         if ((so = nmp->nm_so) == NULL)
1872                 return;
1873
1874         /*
1875          * If there is enough space and the window allows.. resend it.
1876          *
1877          * r_rtt is left intact in case we get an answer after the
1878          * retry that was a reply to the original packet.
1879          *
1880          * NOTE: so_pru_send()
1881          */
1882         if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1883             (req->r_flags & (R_SENT | R_NEEDSXMIT)) &&
1884            (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1885                 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1886                     error = so_pru_send(so, 0, m, NULL, NULL, td);
1887                 else
1888                     error = so_pru_send(so, 0, m, nmp->nm_nam, NULL, td);
1889                 if (error) {
1890                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1891                                 so->so_error = 0;
1892                         req->r_flags |= R_NEEDSXMIT;
1893                 } else if (req->r_mrep == NULL) {
1894                         /*
1895                          * Iff first send, start timing
1896                          * else turn timing off, backoff timer
1897                          * and divide congestion window by 2.
1898                          *
1899                          * It is possible for the so_pru_send() to
1900                          * block and for us to race a reply so we
1901                          * only do this if the reply field has not
1902                          * been filled in.  R_LOCKED will prevent
1903                          * the request from being ripped out from under
1904                          * us entirely.
1905                          *
1906                          * Record the last resent procnum to aid us
1907                          * in duplicate detection on receive.
1908                          */
1909                         if ((req->r_flags & R_NEEDSXMIT) == 0) {
1910                                 if (nfs_showrexmit)
1911                                         kprintf("X");
1912                                 if (++req->r_rexmit > NFS_MAXREXMIT)
1913                                         req->r_rexmit = NFS_MAXREXMIT;
1914                                 nmp->nm_maxasync_scaled >>= 1;
1915                                 if (nmp->nm_maxasync_scaled < NFS_MINASYNC_SCALED)
1916                                         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
1917                                 nfsstats.rpcretries++;
1918                                 nmp->nm_lastreprocnum = req->r_procnum;
1919                         } else {
1920                                 req->r_flags |= R_SENT;
1921                                 req->r_flags &= ~R_NEEDSXMIT;
1922                         }
1923                 }
1924         }
1925 }
1926
1927 /*
1928  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1929  * wait for all requests to complete. This is used by forced unmounts
1930  * to terminate any outstanding RPCs.
1931  *
1932  * Locked requests cannot be canceled but will be marked for
1933  * soft-termination.
1934  */
1935 int
1936 nfs_nmcancelreqs(struct nfsmount *nmp)
1937 {
1938         struct nfsreq *req;
1939         int i;
1940
1941         crit_enter();
1942         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1943                 if (req->r_mrep != NULL || (req->r_flags & R_SOFTTERM))
1944                         continue;
1945                 nfs_softterm(req, 0);
1946         }
1947         /* XXX  the other two queues as well */
1948         crit_exit();
1949
1950         for (i = 0; i < 30; i++) {
1951                 crit_enter();
1952                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1953                         if (nmp == req->r_nmp)
1954                                 break;
1955                 }
1956                 crit_exit();
1957                 if (req == NULL)
1958                         return (0);
1959                 tsleep(&lbolt, 0, "nfscancel", 0);
1960         }
1961         return (EBUSY);
1962 }
1963
1964 /*
1965  * Soft-terminate a request, effectively marking it as failed.
1966  *
1967  * Must be called from within a critical section.
1968  */
1969 static void
1970 nfs_softterm(struct nfsreq *rep, int islocked)
1971 {
1972         rep->r_flags |= R_SOFTTERM;
1973         nfs_hardterm(rep, islocked);
1974 }
1975
1976 /*
1977  * Hard-terminate a request, typically after getting a response.
1978  *
1979  * The state machine can still decide to re-issue it later if necessary.
1980  *
1981  * Must be called from within a critical section.
1982  */
1983 static void
1984 nfs_hardterm(struct nfsreq *rep, int islocked)
1985 {
1986         struct nfsmount *nmp = rep->r_nmp;
1987
1988         /*
1989          * The nm_send count is decremented now to avoid deadlocks
1990          * when the process in soreceive() hasn't yet managed to send
1991          * its own request.
1992          */
1993         if (rep->r_flags & R_SENT) {
1994                 rep->r_flags &= ~R_SENT;
1995         }
1996
1997         /*
1998          * If we locked the request or nobody else has locked the request,
1999          * and the request is async, we can move it to the reader thread's
2000          * queue now and fix up the state.
2001          *
2002          * If we locked the request or nobody else has locked the request,
2003          * we can wake up anyone blocked waiting for a response on the
2004          * request.
2005          */
2006         if (islocked || (rep->r_flags & R_LOCKED) == 0) {
2007                 if ((rep->r_flags & (R_ONREQQ | R_ASYNC)) ==
2008                     (R_ONREQQ | R_ASYNC)) {
2009                         rep->r_flags &= ~R_ONREQQ;
2010                         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
2011                         --nmp->nm_reqqlen;
2012                         TAILQ_INSERT_TAIL(&nmp->nm_reqrxq, rep, r_chain);
2013                         KKASSERT(rep->r_info->state == NFSM_STATE_TRY ||
2014                                  rep->r_info->state == NFSM_STATE_WAITREPLY);
2015                         rep->r_info->state = NFSM_STATE_PROCESSREPLY;
2016                         nfssvc_iod_reader_wakeup(nmp);
2017                         if (TAILQ_FIRST(&nmp->nm_bioq) &&
2018                             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
2019                                 nfssvc_iod_writer_wakeup(nmp);
2020                         }
2021                 }
2022                 mtx_abort_ex_link(&nmp->nm_rxlock, &rep->r_link);
2023         }
2024 }
2025
2026 /*
2027  * Test for a termination condition pending on the process.
2028  * This is used for NFSMNT_INT mounts.
2029  */
2030 int
2031 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
2032 {
2033         sigset_t tmpset;
2034         struct proc *p;
2035         struct lwp *lp;
2036
2037         if (rep && (rep->r_flags & R_SOFTTERM))
2038                 return (EINTR);
2039         /* Terminate all requests while attempting a forced unmount. */
2040         if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
2041                 return (EINTR);
2042         if (!(nmp->nm_flag & NFSMNT_INT))
2043                 return (0);
2044         /* td might be NULL YYY */
2045         if (td == NULL || (p = td->td_proc) == NULL)
2046                 return (0);
2047
2048         lp = td->td_lwp;
2049         tmpset = lwp_sigpend(lp);
2050         SIGSETNAND(tmpset, lp->lwp_sigmask);
2051         SIGSETNAND(tmpset, p->p_sigignore);
2052         if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
2053                 return (EINTR);
2054
2055         return (0);
2056 }
2057
2058 /*
2059  * Lock a socket against others.
2060  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
2061  * and also to avoid race conditions between the processes with nfs requests
2062  * in progress when a reconnect is necessary.
2063  */
2064 int
2065 nfs_sndlock(struct nfsmount *nmp, struct nfsreq *rep)
2066 {
2067         mtx_t mtx = &nmp->nm_txlock;
2068         struct thread *td;
2069         int slptimeo;
2070         int slpflag;
2071         int error;
2072
2073         slpflag = 0;
2074         slptimeo = 0;
2075         td = rep ? rep->r_td : NULL;
2076         if (nmp->nm_flag & NFSMNT_INT)
2077                 slpflag = PCATCH;
2078
2079         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2080                 if (nfs_sigintr(nmp, rep, td)) {
2081                         error = EINTR;
2082                         break;
2083                 }
2084                 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
2085                 if (error == 0)
2086                         break;
2087                 if (slpflag == PCATCH) {
2088                         slpflag = 0;
2089                         slptimeo = 2 * hz;
2090                 }
2091         }
2092         /* Always fail if our request has been cancelled. */
2093         if (rep && (rep->r_flags & R_SOFTTERM)) {
2094                 if (error == 0)
2095                         mtx_unlock(mtx);
2096                 error = EINTR;
2097         }
2098         return (error);
2099 }
2100
2101 /*
2102  * Unlock the stream socket for others.
2103  */
2104 void
2105 nfs_sndunlock(struct nfsmount *nmp)
2106 {
2107         mtx_unlock(&nmp->nm_txlock);
2108 }
2109
2110 /*
2111  * Lock the receiver side of the socket.
2112  *
2113  * rep may be NULL.
2114  */
2115 static int
2116 nfs_rcvlock(struct nfsmount *nmp, struct nfsreq *rep)
2117 {
2118         mtx_t mtx = &nmp->nm_rxlock;
2119         int slpflag;
2120         int slptimeo;
2121         int error;
2122
2123         /*
2124          * Unconditionally check for completion in case another nfsiod
2125          * get the packet while the caller was blocked, before the caller
2126          * called us.  Packet reception is handled by mainline code which
2127          * is protected by the BGL at the moment.
2128          *
2129          * We do not strictly need the second check just before the
2130          * tsleep(), but it's good defensive programming.
2131          */
2132         if (rep && rep->r_mrep != NULL)
2133                 return (EALREADY);
2134
2135         if (nmp->nm_flag & NFSMNT_INT)
2136                 slpflag = PCATCH;
2137         else
2138                 slpflag = 0;
2139         slptimeo = 0;
2140
2141         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2142                 if (nfs_sigintr(nmp, rep, (rep ? rep->r_td : NULL))) {
2143                         error = EINTR;
2144                         break;
2145                 }
2146                 if (rep && rep->r_mrep != NULL) {
2147                         error = EALREADY;
2148                         break;
2149                 }
2150
2151                 /*
2152                  * NOTE: can return ENOLCK, but in that case rep->r_mrep
2153                  *       will already be set.
2154                  */
2155                 if (rep) {
2156                         error = mtx_lock_ex_link(mtx, &rep->r_link,
2157                                                  "nfsrcvlk",
2158                                                  slpflag, slptimeo);
2159                 } else {
2160                         error = mtx_lock_ex(mtx, "nfsrcvlk", slpflag, slptimeo);
2161                 }
2162                 if (error == 0)
2163                         break;
2164
2165                 /*
2166                  * If our reply was recieved while we were sleeping,
2167                  * then just return without taking the lock to avoid a
2168                  * situation where a single iod could 'capture' the
2169                  * recieve lock.
2170                  */
2171                 if (rep && rep->r_mrep != NULL) {
2172                         error = EALREADY;
2173                         break;
2174                 }
2175                 if (slpflag == PCATCH) {
2176                         slpflag = 0;
2177                         slptimeo = 2 * hz;
2178                 }
2179         }
2180         if (error == 0) {
2181                 if (rep && rep->r_mrep != NULL) {
2182                         error = EALREADY;
2183                         mtx_unlock(mtx);
2184                 }
2185         }
2186         return (error);
2187 }
2188
2189 /*
2190  * Unlock the stream socket for others.
2191  */
2192 static void
2193 nfs_rcvunlock(struct nfsmount *nmp)
2194 {
2195         mtx_unlock(&nmp->nm_rxlock);
2196 }
2197
2198 /*
2199  * nfs_realign:
2200  *
2201  * Check for badly aligned mbuf data and realign by copying the unaligned
2202  * portion of the data into a new mbuf chain and freeing the portions
2203  * of the old chain that were replaced.
2204  *
2205  * We cannot simply realign the data within the existing mbuf chain
2206  * because the underlying buffers may contain other rpc commands and
2207  * we cannot afford to overwrite them.
2208  *
2209  * We would prefer to avoid this situation entirely.  The situation does
2210  * not occur with NFS/UDP and is supposed to only occassionally occur
2211  * with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
2212  *
2213  * NOTE!  MB_DONTWAIT cannot be used here.  The mbufs must be acquired
2214  *        because the rpc request OR reply cannot be thrown away.  TCP NFS
2215  *        mounts do not retry their RPCs unless the TCP connection itself
2216  *        is dropped so throwing away a RPC will basically cause the NFS
2217  *        operation to lockup indefinitely.
2218  */
2219 static void
2220 nfs_realign(struct mbuf **pm, int hsiz)
2221 {
2222         struct mbuf *m;
2223         struct mbuf *n = NULL;
2224
2225         /*
2226          * Check for misalignemnt
2227          */
2228         ++nfs_realign_test;
2229         while ((m = *pm) != NULL) {
2230                 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3))
2231                         break;
2232                 pm = &m->m_next;
2233         }
2234
2235         /*
2236          * If misalignment found make a completely new copy.
2237          */
2238         if (m) {
2239                 ++nfs_realign_count;
2240                 n = m_dup_data(m, MB_WAIT);
2241                 m_freem(*pm);
2242                 *pm = n;
2243         }
2244 }
2245
2246 #ifndef NFS_NOSERVER
2247
2248 /*
2249  * Parse an RPC request
2250  * - verify it
2251  * - fill in the cred struct.
2252  */
2253 int
2254 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
2255 {
2256         int len, i;
2257         u_int32_t *tl;
2258         struct uio uio;
2259         struct iovec iov;
2260         caddr_t cp;
2261         u_int32_t nfsvers, auth_type;
2262         uid_t nickuid;
2263         int error = 0, ticklen;
2264         struct nfsuid *nuidp;
2265         struct timeval tvin, tvout;
2266         struct nfsm_info info;
2267 #if 0                           /* until encrypted keys are implemented */
2268         NFSKERBKEYSCHED_T keys; /* stores key schedule */
2269 #endif
2270
2271         info.mrep = nd->nd_mrep;
2272         info.md = nd->nd_md;
2273         info.dpos = nd->nd_dpos;
2274
2275         if (has_header) {
2276                 NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
2277                 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
2278                 if (*tl++ != rpc_call) {
2279                         m_freem(info.mrep);
2280                         return (EBADRPC);
2281                 }
2282         } else {
2283                 NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
2284         }
2285         nd->nd_repstat = 0;
2286         nd->nd_flag = 0;
2287         if (*tl++ != rpc_vers) {
2288                 nd->nd_repstat = ERPCMISMATCH;
2289                 nd->nd_procnum = NFSPROC_NOOP;
2290                 return (0);
2291         }
2292         if (*tl != nfs_prog) {
2293                 nd->nd_repstat = EPROGUNAVAIL;
2294                 nd->nd_procnum = NFSPROC_NOOP;
2295                 return (0);
2296         }
2297         tl++;
2298         nfsvers = fxdr_unsigned(u_int32_t, *tl++);
2299         if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
2300                 nd->nd_repstat = EPROGMISMATCH;
2301                 nd->nd_procnum = NFSPROC_NOOP;
2302                 return (0);
2303         }
2304         if (nfsvers == NFS_VER3)
2305                 nd->nd_flag = ND_NFSV3;
2306         nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
2307         if (nd->nd_procnum == NFSPROC_NULL)
2308                 return (0);
2309         if (nd->nd_procnum >= NFS_NPROCS ||
2310                 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
2311                 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
2312                 nd->nd_repstat = EPROCUNAVAIL;
2313                 nd->nd_procnum = NFSPROC_NOOP;
2314                 return (0);
2315         }
2316         if ((nd->nd_flag & ND_NFSV3) == 0)
2317                 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
2318         auth_type = *tl++;
2319         len = fxdr_unsigned(int, *tl++);
2320         if (len < 0 || len > RPCAUTH_MAXSIZ) {
2321                 m_freem(info.mrep);
2322                 return (EBADRPC);
2323         }
2324
2325         nd->nd_flag &= ~ND_KERBAUTH;
2326         /*
2327          * Handle auth_unix or auth_kerb.
2328          */
2329         if (auth_type == rpc_auth_unix) {
2330                 len = fxdr_unsigned(int, *++tl);
2331                 if (len < 0 || len > NFS_MAXNAMLEN) {
2332                         m_freem(info.mrep);
2333                         return (EBADRPC);
2334                 }
2335                 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2336                 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2337                 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
2338                 nd->nd_cr.cr_ref = 1;
2339                 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
2340                 nd->nd_cr.cr_ruid = nd->nd_cr.cr_svuid = nd->nd_cr.cr_uid;
2341                 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
2342                 nd->nd_cr.cr_rgid = nd->nd_cr.cr_svgid = nd->nd_cr.cr_gid;
2343                 len = fxdr_unsigned(int, *tl);
2344                 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
2345                         m_freem(info.mrep);
2346                         return (EBADRPC);
2347                 }
2348                 NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
2349                 for (i = 1; i <= len; i++)
2350                     if (i < NGROUPS)
2351                         nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
2352                     else
2353                         tl++;
2354                 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
2355                 if (nd->nd_cr.cr_ngroups > 1)
2356                     nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
2357                 len = fxdr_unsigned(int, *++tl);
2358                 if (len < 0 || len > RPCAUTH_MAXSIZ) {
2359                         m_freem(info.mrep);
2360                         return (EBADRPC);
2361                 }
2362                 if (len > 0) {
2363                         ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2364                 }
2365         } else if (auth_type == rpc_auth_kerb) {
2366                 switch (fxdr_unsigned(int, *tl++)) {
2367                 case RPCAKN_FULLNAME:
2368                         ticklen = fxdr_unsigned(int, *tl);
2369                         *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2370                         uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2371                         nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2372                         if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2373                                 m_freem(info.mrep);
2374                                 return (EBADRPC);
2375                         }
2376                         uio.uio_offset = 0;
2377                         uio.uio_iov = &iov;
2378                         uio.uio_iovcnt = 1;
2379                         uio.uio_segflg = UIO_SYSSPACE;
2380                         iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2381                         iov.iov_len = RPCAUTH_MAXSIZ - 4;
2382                         ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2383                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2384                         if (*tl++ != rpc_auth_kerb ||
2385                                 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2386                                 kprintf("Bad kerb verifier\n");
2387                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2388                                 nd->nd_procnum = NFSPROC_NOOP;
2389                                 return (0);
2390                         }
2391                         NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2392                         tl = (u_int32_t *)cp;
2393                         if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2394                                 kprintf("Not fullname kerb verifier\n");
2395                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2396                                 nd->nd_procnum = NFSPROC_NOOP;
2397                                 return (0);
2398                         }
2399                         cp += NFSX_UNSIGNED;
2400                         bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2401                         nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2402                         nd->nd_flag |= ND_KERBFULL;
2403                         nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2404                         break;
2405                 case RPCAKN_NICKNAME:
2406                         if (len != 2 * NFSX_UNSIGNED) {
2407                                 kprintf("Kerb nickname short\n");
2408                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2409                                 nd->nd_procnum = NFSPROC_NOOP;
2410                                 return (0);
2411                         }
2412                         nickuid = fxdr_unsigned(uid_t, *tl);
2413                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2414                         if (*tl++ != rpc_auth_kerb ||
2415                                 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2416                                 kprintf("Kerb nick verifier bad\n");
2417                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2418                                 nd->nd_procnum = NFSPROC_NOOP;
2419                                 return (0);
2420                         }
2421                         NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2422                         tvin.tv_sec = *tl++;
2423                         tvin.tv_usec = *tl;
2424
2425                         for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2426                             nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2427                                 if (nuidp->nu_cr.cr_uid == nickuid &&
2428                                     (!nd->nd_nam2 ||
2429                                      netaddr_match(NU_NETFAM(nuidp),
2430                                       &nuidp->nu_haddr, nd->nd_nam2)))
2431                                         break;
2432                         }
2433                         if (!nuidp) {
2434                                 nd->nd_repstat =
2435                                         (NFSERR_AUTHERR|AUTH_REJECTCRED);
2436                                 nd->nd_procnum = NFSPROC_NOOP;
2437                                 return (0);
2438                         }
2439
2440                         /*
2441                          * Now, decrypt the timestamp using the session key
2442                          * and validate it.
2443                          */
2444 #ifdef NFSKERB
2445                         XXX
2446 #else
2447                         tvout.tv_sec = 0;
2448                         tvout.tv_usec = 0;
2449 #endif
2450
2451                         tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2452                         tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2453                         if (nuidp->nu_expire < time_second ||
2454                             nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2455                             (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2456                              nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2457                                 nuidp->nu_expire = 0;
2458                                 nd->nd_repstat =
2459                                     (NFSERR_AUTHERR|AUTH_REJECTVERF);
2460                                 nd->nd_procnum = NFSPROC_NOOP;
2461                                 return (0);
2462                         }
2463                         nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2464                         nd->nd_flag |= ND_KERBNICK;
2465                 };
2466         } else {
2467                 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2468                 nd->nd_procnum = NFSPROC_NOOP;
2469                 return (0);
2470         }
2471
2472         nd->nd_md = info.md;
2473         nd->nd_dpos = info.dpos;
2474         return (0);
2475 nfsmout:
2476         return (error);
2477 }
2478
2479 #endif
2480
2481 /*
2482  * Send a message to the originating process's terminal.  The thread and/or
2483  * process may be NULL.  YYY the thread should not be NULL but there may
2484  * still be some uio_td's that are still being passed as NULL through to
2485  * nfsm_request().
2486  */
2487 static int
2488 nfs_msg(struct thread *td, char *server, char *msg)
2489 {
2490         tpr_t tpr;
2491
2492         if (td && td->td_proc)
2493                 tpr = tprintf_open(td->td_proc);
2494         else
2495                 tpr = NULL;
2496         tprintf(tpr, "nfs server %s: %s\n", server, msg);
2497         tprintf_close(tpr);
2498         return (0);
2499 }
2500
2501 #ifndef NFS_NOSERVER
2502
2503 /*
2504  * Socket upcall routine for nfsd sockets.  This runs in the protocol
2505  * thread and passes waitflag == MB_DONTWAIT.
2506  */
2507 void
2508 nfsrv_rcv_upcall(struct socket *so, void *arg, int waitflag)
2509 {
2510         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2511
2512         if (slp->ns_needq_upcall == 0) {
2513                 slp->ns_needq_upcall = 1;       /* ok to race */
2514                 lwkt_gettoken(&nfs_token);
2515                 nfsrv_wakenfsd(slp, 1);
2516                 lwkt_reltoken(&nfs_token);
2517         }
2518 #if 0
2519         lwkt_gettoken(&slp->ns_token);
2520         slp->ns_flag |= SLP_NEEDQ;
2521         nfsrv_rcv(so, arg, waitflag);
2522         lwkt_reltoken(&slp->ns_token);
2523 #endif
2524 }
2525
2526 /*
2527  * Process new data on a receive socket.  Essentially do as much as we can
2528  * non-blocking, else punt and it will be called with MB_WAIT from an nfsd.
2529  *
2530  * slp->ns_token is held on call
2531  */
2532 void
2533 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2534 {
2535         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2536         struct mbuf *m;
2537         struct sockaddr *nam;
2538         struct sockbuf sio;
2539         int flags, error;
2540         int nparallel_wakeup = 0;
2541
2542         ASSERT_LWKT_TOKEN_HELD(&slp->ns_token);
2543
2544         if ((slp->ns_flag & SLP_VALID) == 0)
2545                 return;
2546
2547         /*
2548          * Do not allow an infinite number of completed RPC records to build 
2549          * up before we stop reading data from the socket.  Otherwise we could
2550          * end up holding onto an unreasonable number of mbufs for requests
2551          * waiting for service.
2552          *
2553          * This should give pretty good feedback to the TCP layer and
2554          * prevents a memory crunch for other protocols.
2555          *
2556          * Note that the same service socket can be dispatched to several
2557          * nfs servers simultaniously.  The tcp protocol callback calls us
2558          * with MB_DONTWAIT.  nfsd calls us with MB_WAIT (typically).
2559          */
2560         if (NFSRV_RECLIMIT(slp))
2561                 return;
2562
2563         /*
2564          * Handle protocol specifics to parse an RPC request.  We always
2565          * pull from the socket using non-blocking I/O.
2566          */
2567         if (so->so_type == SOCK_STREAM) {
2568                 /*
2569                  * The data has to be read in an orderly fashion from a TCP
2570                  * stream, unlike a UDP socket.  It is possible for soreceive
2571                  * and/or nfsrv_getstream() to block, so make sure only one
2572                  * entity is messing around with the TCP stream at any given
2573                  * moment.  The receive sockbuf's lock in soreceive is not
2574                  * sufficient.
2575                  */
2576                 if (slp->ns_flag & SLP_GETSTREAM)
2577                         return;
2578                 slp->ns_flag |= SLP_GETSTREAM;
2579
2580                 /*
2581                  * Do soreceive().  Pull out as much data as possible without
2582                  * blocking.
2583                  */
2584                 sbinit(&sio, 1000000000);
2585                 flags = MSG_DONTWAIT;
2586                 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2587                 if (error || sio.sb_mb == NULL) {
2588                         if (error != EWOULDBLOCK)
2589                                 slp->ns_flag |= SLP_DISCONN;
2590                         slp->ns_flag &= ~(SLP_GETSTREAM | SLP_NEEDQ);
2591                         goto done;
2592                 }
2593                 m = sio.sb_mb;
2594                 if (slp->ns_rawend) {
2595                         slp->ns_rawend->m_next = m;
2596                         slp->ns_cc += sio.sb_cc;
2597                 } else {
2598                         slp->ns_raw = m;
2599                         slp->ns_cc = sio.sb_cc;
2600                 }
2601                 while (m->m_next)
2602                         m = m->m_next;
2603                 slp->ns_rawend = m;
2604
2605                 /*
2606                  * Now try and parse as many record(s) as we can out of the
2607                  * raw stream data.  This will set SLP_DOREC.
2608                  */
2609                 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2610                 if (error && error != EWOULDBLOCK)
2611                         slp->ns_flag |= SLP_DISCONN;
2612                 slp->ns_flag &= ~SLP_GETSTREAM;
2613         } else {
2614                 /*
2615                  * For UDP soreceive typically pulls just one packet, loop
2616                  * to get the whole batch.
2617                  */
2618                 do {
2619                         sbinit(&sio, 1000000000);
2620                         flags = MSG_DONTWAIT;
2621                         error = so_pru_soreceive(so, &nam, NULL, &sio,
2622                                                  NULL, &flags);
2623                         if (sio.sb_mb) {
2624                                 struct nfsrv_rec *rec;
2625                                 int mf = (waitflag & MB_DONTWAIT) ?
2626                                             M_NOWAIT : M_WAITOK;
2627                                 rec = kmalloc(sizeof(struct nfsrv_rec),
2628                                              M_NFSRVDESC, mf);
2629                                 if (!rec) {
2630                                         if (nam)
2631                                                 FREE(nam, M_SONAME);
2632                                         m_freem(sio.sb_mb);
2633                                         continue;
2634                                 }
2635                                 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2636                                 rec->nr_address = nam;
2637                                 rec->nr_packet = sio.sb_mb;
2638                                 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2639                                 ++slp->ns_numrec;
2640                                 slp->ns_flag |= SLP_DOREC;
2641                                 ++nparallel_wakeup;
2642                         } else {
2643                                 slp->ns_flag &= ~SLP_NEEDQ;
2644                         }
2645                         if (error) {
2646                                 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2647                                     && error != EWOULDBLOCK) {
2648                                         slp->ns_flag |= SLP_DISCONN;
2649                                         break;
2650                                 }
2651                         }
2652                         if (NFSRV_RECLIMIT(slp))
2653                                 break;
2654                 } while (sio.sb_mb);
2655         }
2656
2657         /*
2658          * If we were upcalled from the tcp protocol layer and we have
2659          * fully parsed records ready to go, or there is new data pending,
2660          * or something went wrong, try to wake up a nfsd thread to deal
2661          * with it.
2662          */
2663 done:
2664         /* XXX this code is currently not executed (nfsrv_rcv_upcall) */
2665         if (waitflag == MB_DONTWAIT && (slp->ns_flag & SLP_ACTION_MASK)) {
2666                 lwkt_gettoken(&nfs_token);
2667                 nfsrv_wakenfsd(slp, nparallel_wakeup);
2668                 lwkt_reltoken(&nfs_token);
2669         }
2670 }
2671
2672 /*
2673  * Try and extract an RPC request from the mbuf data list received on a
2674  * stream socket. The "waitflag" argument indicates whether or not it
2675  * can sleep.
2676  */
2677 static int
2678 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2679 {
2680         struct mbuf *m, **mpp;
2681         char *cp1, *cp2;
2682         int len;
2683         struct mbuf *om, *m2, *recm;
2684         u_int32_t recmark;
2685
2686         for (;;) {
2687             if (slp->ns_reclen == 0) {
2688                 if (slp->ns_cc < NFSX_UNSIGNED)
2689                         return (0);
2690                 m = slp->ns_raw;
2691                 if (m->m_len >= NFSX_UNSIGNED) {
2692                         bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2693                         m->m_data += NFSX_UNSIGNED;
2694                         m->m_len -= NFSX_UNSIGNED;
2695                 } else {
2696                         cp1 = (caddr_t)&recmark;
2697                         cp2 = mtod(m, caddr_t);
2698                         while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2699                                 while (m->m_len == 0) {
2700                                         m = m->m_next;
2701                                         cp2 = mtod(m, caddr_t);
2702                                 }
2703                                 *cp1++ = *cp2++;
2704                                 m->m_data++;
2705                                 m->m_len--;
2706                         }
2707                 }
2708                 slp->ns_cc -= NFSX_UNSIGNED;
2709                 recmark = ntohl(recmark);
2710                 slp->ns_reclen = recmark & ~0x80000000;
2711                 if (recmark & 0x80000000)
2712                         slp->ns_flag |= SLP_LASTFRAG;
2713                 else
2714                         slp->ns_flag &= ~SLP_LASTFRAG;
2715                 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2716                         log(LOG_ERR, "%s (%d) from nfs client\n",
2717                             "impossible packet length",
2718                             slp->ns_reclen);
2719                         return (EPERM);
2720                 }
2721             }
2722
2723             /*
2724              * Now get the record part.
2725              *
2726              * Note that slp->ns_reclen may be 0.  Linux sometimes
2727              * generates 0-length RPCs
2728              */
2729             recm = NULL;
2730             if (slp->ns_cc == slp->ns_reclen) {
2731                 recm = slp->ns_raw;
2732                 slp->ns_raw = slp->ns_rawend = NULL;
2733                 slp->ns_cc = slp->ns_reclen = 0;
2734             } else if (slp->ns_cc > slp->ns_reclen) {
2735                 len = 0;
2736                 m = slp->ns_raw;
2737                 om = NULL;
2738
2739                 while (len < slp->ns_reclen) {
2740                         if ((len + m->m_len) > slp->ns_reclen) {
2741                                 m2 = m_copym(m, 0, slp->ns_reclen - len,
2742                                         waitflag);
2743                                 if (m2) {
2744                                         if (om) {
2745                                                 om->m_next = m2;
2746                                                 recm = slp->ns_raw;
2747                                         } else
2748                                                 recm = m2;
2749                                         m->m_data += slp->ns_reclen - len;
2750                                         m->m_len -= slp->ns_reclen - len;
2751                                         len = slp->ns_reclen;
2752                                 } else {
2753                                         return (EWOULDBLOCK);
2754                                 }
2755                         } else if ((len + m->m_len) == slp->ns_reclen) {
2756                                 om = m;
2757                                 len += m->m_len;
2758                                 m = m->m_next;
2759                                 recm = slp->ns_raw;
2760                                 om->m_next = NULL;
2761                         } else {
2762                                 om = m;
2763                                 len += m->m_len;
2764                                 m = m->m_next;
2765                         }
2766                 }
2767                 slp->ns_raw = m;
2768                 slp->ns_cc -= len;
2769                 slp->ns_reclen = 0;
2770             } else {
2771                 return (0);
2772             }
2773
2774             /*
2775              * Accumulate the fragments into a record.
2776              */
2777             mpp = &slp->ns_frag;
2778             while (*mpp)
2779                 mpp = &((*mpp)->m_next);
2780             *mpp = recm;
2781             if (slp->ns_flag & SLP_LASTFRAG) {
2782                 struct nfsrv_rec *rec;
2783                 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2784                 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2785                 if (!rec) {
2786                     m_freem(slp->ns_frag);
2787                 } else {
2788                     nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2789                     rec->nr_address = NULL;
2790                     rec->nr_packet = slp->ns_frag;
2791                     STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2792                     ++slp->ns_numrec;
2793                     slp->ns_flag |= SLP_DOREC;
2794                     ++*countp;
2795                 }
2796                 slp->ns_frag = NULL;
2797             }
2798         }
2799 }
2800
2801 #ifdef INVARIANTS
2802
2803 /*
2804  * Sanity check our mbuf chain.
2805  */
2806 static void
2807 nfs_checkpkt(struct mbuf *m, int len)
2808 {
2809         int xlen = 0;
2810         while (m) {
2811                 xlen += m->m_len;
2812                 m = m->m_next;
2813         }
2814         if (xlen != len) {
2815                 panic("nfs_checkpkt: len mismatch %d/%d mbuf %p\n",
2816                         xlen, len, m);
2817         }
2818 }
2819
2820 #else
2821
2822 static void
2823 nfs_checkpkt(struct mbuf *m __unused, int len __unused)
2824 {
2825 }
2826
2827 #endif
2828
2829 /*
2830  * Parse an RPC header.
2831  *
2832  * If the socket is invalid or no records are pending we return ENOBUFS.
2833  * The caller must deal with NEEDQ races.
2834  */
2835 int
2836 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2837             struct nfsrv_descript **ndp)
2838 {
2839         struct nfsrv_rec *rec;
2840         struct mbuf *m;
2841         struct sockaddr *nam;
2842         struct nfsrv_descript *nd;
2843         int error;
2844
2845         *ndp = NULL;
2846         if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2847                 return (ENOBUFS);
2848         rec = STAILQ_FIRST(&slp->ns_rec);
2849         STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2850         KKASSERT(slp->ns_numrec > 0);
2851         if (--slp->ns_numrec == 0)
2852                 slp->ns_flag &= ~SLP_DOREC;
2853         nam = rec->nr_address;
2854         m = rec->nr_packet;
2855         kfree(rec, M_NFSRVDESC);
2856         MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2857                 M_NFSRVDESC, M_WAITOK);
2858         nd->nd_md = nd->nd_mrep = m;
2859         nd->nd_nam2 = nam;
2860         nd->nd_dpos = mtod(m, caddr_t);
2861         error = nfs_getreq(nd, nfsd, TRUE);
2862         if (error) {
2863                 if (nam) {
2864                         FREE(nam, M_SONAME);
2865                 }
2866                 kfree((caddr_t)nd, M_NFSRVDESC);
2867                 return (error);
2868         }
2869         *ndp = nd;
2870         nfsd->nfsd_nd = nd;
2871         return (0);
2872 }
2873
2874 /*
2875  * Try to assign service sockets to nfsd threads based on the number
2876  * of new rpc requests that have been queued on the service socket.
2877  *
2878  * If no nfsd's are available or additonal requests are pending, set the
2879  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2880  * the work in the nfssvc_sock list when it is finished processing its
2881  * current work.  This flag is only cleared when an nfsd can not find
2882  * any new work to perform.
2883  */
2884 void
2885 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2886 {
2887         struct nfsd *nd;
2888
2889         if ((slp->ns_flag & SLP_VALID) == 0)
2890                 return;
2891         if (nparallel <= 1)
2892                 nparallel = 1;
2893         TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2894                 if (nd->nfsd_flag & NFSD_WAITING) {
2895                         nd->nfsd_flag &= ~NFSD_WAITING;
2896                         if (nd->nfsd_slp)
2897                                 panic("nfsd wakeup");
2898                         nfsrv_slpref(slp);
2899                         nd->nfsd_slp = slp;
2900                         wakeup((caddr_t)nd);
2901                         if (--nparallel == 0)
2902                                 break;
2903                 }
2904         }
2905
2906         /*
2907          * If we couldn't assign slp then the NFSDs are all busy and
2908          * we set a flag indicating that there is pending work.
2909          */
2910         if (nparallel)
2911                 nfsd_head_flag |= NFSD_CHECKSLP;
2912 }
2913 #endif /* NFS_NOSERVER */