sys/vfs/nfs/nfs_socket.c

   1 /*
   2  * Copyright (c) 1989, 1991, 1993, 1995
   3  *      The Regents of the University of California.  All rights reserved.
   4  *
   5  * This code is derived from software contributed to Berkeley by
   6  * Rick Macklem at The University of Guelph.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 3. All advertising materials mentioning features or use of this software
  17  *    must display the following acknowledgement:
  18  *      This product includes software developed by the University of
  19  *      California, Berkeley and its contributors.
  20  * 4. Neither the name of the University nor the names of its contributors
  21  *    may be used to endorse or promote products derived from this software
  22  *    without specific prior written permission.
  23  *
  24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  34  * SUCH DAMAGE.
  35  *
  36  *      @(#)nfs_socket.c        8.5 (Berkeley) 3/30/95
  37  * $FreeBSD: src/sys/nfs/nfs_socket.c,v 1.60.2.6 2003/03/26 01:44:46 alfred Exp $
  38  */
  39
  40 /*
  41  * Socket operations for use by nfs
  42  */
  43
  44 #include <sys/param.h>
  45 #include <sys/systm.h>
  46 #include <sys/proc.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mount.h>
  49 #include <sys/kernel.h>
  50 #include <sys/mbuf.h>
  51 #include <sys/vnode.h>
  52 #include <sys/fcntl.h>
  53 #include <sys/protosw.h>
  54 #include <sys/resourcevar.h>
  55 #include <sys/socket.h>
  56 #include <sys/socketvar.h>
  57 #include <sys/socketops.h>
  58 #include <sys/syslog.h>
  59 #include <sys/thread.h>
  60 #include <sys/tprintf.h>
  61 #include <sys/sysctl.h>
  62 #include <sys/signalvar.h>
  63
  64 #include <sys/signal2.h>
  65 #include <sys/mutex2.h>
  66 #include <sys/socketvar2.h>
  67
  68 #include <netinet/in.h>
  69 #include <netinet/tcp.h>
  70 #include <sys/thread2.h>
  71
  72 #include "rpcv2.h"
  73 #include "nfsproto.h"
  74 #include "nfs.h"
  75 #include "xdr_subs.h"
  76 #include "nfsm_subs.h"
  77 #include "nfsmount.h"
  78 #include "nfsnode.h"
  79 #include "nfsrtt.h"
  80
  81 #define TRUE    1
  82 #define FALSE   0
  83
  84 /*
  85  * RTT calculations are scaled by 256 (8 bits).  A proper fractional
  86  * RTT will still be calculated even with a slow NFS timer.
  87  */
  88 #define NFS_SRTT(r)     (r)->r_nmp->nm_srtt[proct[(r)->r_procnum]]
  89 #define NFS_SDRTT(r)    (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum]]
  90 #define NFS_RTT_SCALE_BITS      8       /* bits */
  91 #define NFS_RTT_SCALE           256     /* value */
  92
  93 /*
  94  * Defines which timer to use for the procnum.
  95  * 0 - default
  96  * 1 - getattr
  97  * 2 - lookup
  98  * 3 - read
  99  * 4 - write
 100  */
 101 static int proct[NFS_NPROCS] = {
 102         0, 1, 0, 2, 1, 3, 3, 4, 0, 0,   /* 00-09        */
 103         0, 0, 0, 0, 0, 0, 3, 3, 0, 0,   /* 10-19        */
 104         0, 5, 0, 0, 0, 0,               /* 20-29        */
 105 };
 106
 107 static int multt[NFS_NPROCS] = {
 108         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 00-09        */
 109         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 10-19        */
 110         1, 2, 1, 1, 1, 1,               /* 20-29        */
 111 };
 112
 113 static int nfs_backoff[8] = { 2, 3, 5, 8, 13, 21, 34, 55 };
 114 static int nfs_realign_test;
 115 static int nfs_realign_count;
 116 static int nfs_showrtt;
 117 static int nfs_showrexmit;
 118 int nfs_maxasyncbio = NFS_MAXASYNCBIO;
 119
 120 SYSCTL_DECL(_vfs_nfs);
 121
 122 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0,
 123     "Number of times mbufs have been tested for bad alignment");
 124 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0,
 125     "Number of realignments for badly aligned mbuf data");
 126 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrtt, CTLFLAG_RW, &nfs_showrtt, 0,
 127     "Show round trip time output");
 128 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrexmit, CTLFLAG_RW, &nfs_showrexmit, 0,
 129     "Show retransmits info");
 130 SYSCTL_INT(_vfs_nfs, OID_AUTO, maxasyncbio, CTLFLAG_RW, &nfs_maxasyncbio, 0,
 131     "Max number of asynchronous bio's");
 132
 133 static int nfs_request_setup(nfsm_info_t info);
 134 static int nfs_request_auth(struct nfsreq *rep);
 135 static int nfs_request_try(struct nfsreq *rep);
 136 static int nfs_request_waitreply(struct nfsreq *rep);
 137 static int nfs_request_processreply(nfsm_info_t info, int);
 138
 139 int nfsrtton = 0;
 140 struct nfsrtt nfsrtt;
 141 struct callout  nfs_timer_handle;
 142
 143 static int      nfs_msg (struct thread *,char *,char *);
 144 static int      nfs_rcvlock (struct nfsmount *nmp, struct nfsreq *myreq);
 145 static void     nfs_rcvunlock (struct nfsmount *nmp);
 146 static void     nfs_realign (struct mbuf **pm, int hsiz);
 147 static int      nfs_receive (struct nfsmount *nmp, struct nfsreq *rep,
 148                                 struct sockaddr **aname, struct mbuf **mp);
 149 static void     nfs_softterm (struct nfsreq *rep, int islocked);
 150 static void     nfs_hardterm (struct nfsreq *rep, int islocked);
 151 static int      nfs_reconnect (struct nfsmount *nmp, struct nfsreq *rep);
 152 #ifndef NFS_NOSERVER
 153 static int      nfsrv_getstream (struct nfssvc_sock *, int, int *);
 154 static void     nfs_timer_req(struct nfsreq *req);
 155 static void     nfs_checkpkt(struct mbuf *m, int len);
 156
 157 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
 158                                     struct nfssvc_sock *slp,
 159                                     struct thread *td,
 160                                     struct mbuf **mreqp) = {
 161         nfsrv_null,
 162         nfsrv_getattr,
 163         nfsrv_setattr,
 164         nfsrv_lookup,
 165         nfsrv3_access,
 166         nfsrv_readlink,
 167         nfsrv_read,
 168         nfsrv_write,
 169         nfsrv_create,
 170         nfsrv_mkdir,
 171         nfsrv_symlink,
 172         nfsrv_mknod,
 173         nfsrv_remove,
 174         nfsrv_rmdir,
 175         nfsrv_rename,
 176         nfsrv_link,
 177         nfsrv_readdir,
 178         nfsrv_readdirplus,
 179         nfsrv_statfs,
 180         nfsrv_fsinfo,
 181         nfsrv_pathconf,
 182         nfsrv_commit,
 183         nfsrv_noop,
 184         nfsrv_noop,
 185         nfsrv_noop,
 186         nfsrv_noop
 187 };
 188 #endif /* NFS_NOSERVER */
 189
 190 /*
 191  * Initialize sockets and congestion for a new NFS connection.
 192  * We do not free the sockaddr if error.
 193  */
 194 int
 195 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
 196 {
 197         struct socket *so;
 198         int error;
 199         struct sockaddr *saddr;
 200         struct sockaddr_in *sin;
 201         struct thread *td = &thread0; /* only used for socreate and sobind */
 202
 203         nmp->nm_so = so = NULL;
 204         if (nmp->nm_flag & NFSMNT_FORCE)
 205                 return (EINVAL);
 206         saddr = nmp->nm_nam;
 207         error = socreate(saddr->sa_family, &so, nmp->nm_sotype,
 208                 nmp->nm_soproto, td);
 209         if (error)
 210                 goto bad;
 211         nmp->nm_soflags = so->so_proto->pr_flags;
 212
 213         /*
 214          * Some servers require that the client port be a reserved port number.
 215          */
 216         if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
 217                 struct sockopt sopt;
 218                 int ip;
 219                 struct sockaddr_in ssin;
 220
 221                 bzero(&sopt, sizeof sopt);
 222                 ip = IP_PORTRANGE_LOW;
 223                 sopt.sopt_level = IPPROTO_IP;
 224                 sopt.sopt_name = IP_PORTRANGE;
 225                 sopt.sopt_val = (void *)&ip;
 226                 sopt.sopt_valsize = sizeof(ip);
 227                 sopt.sopt_td = NULL;
 228                 error = sosetopt(so, &sopt);
 229                 if (error)
 230                         goto bad;
 231                 bzero(&ssin, sizeof ssin);
 232                 sin = &ssin;
 233                 sin->sin_len = sizeof (struct sockaddr_in);
 234                 sin->sin_family = AF_INET;
 235                 sin->sin_addr.s_addr = INADDR_ANY;
 236                 sin->sin_port = htons(0);
 237                 error = sobind(so, (struct sockaddr *)sin, td);
 238                 if (error)
 239                         goto bad;
 240                 bzero(&sopt, sizeof sopt);
 241                 ip = IP_PORTRANGE_DEFAULT;
 242                 sopt.sopt_level = IPPROTO_IP;
 243                 sopt.sopt_name = IP_PORTRANGE;
 244                 sopt.sopt_val = (void *)&ip;
 245                 sopt.sopt_valsize = sizeof(ip);
 246                 sopt.sopt_td = NULL;
 247                 error = sosetopt(so, &sopt);
 248                 if (error)
 249                         goto bad;
 250         }
 251
 252         /*
 253          * Protocols that do not require connections may be optionally left
 254          * unconnected for servers that reply from a port other than NFS_PORT.
 255          */
 256         if (nmp->nm_flag & NFSMNT_NOCONN) {
 257                 if (nmp->nm_soflags & PR_CONNREQUIRED) {
 258                         error = ENOTCONN;
 259                         goto bad;
 260                 }
 261         } else {
 262                 error = soconnect(so, nmp->nm_nam, td);
 263                 if (error)
 264                         goto bad;
 265
 266                 /*
 267                  * Wait for the connection to complete. Cribbed from the
 268                  * connect system call but with the wait timing out so
 269                  * that interruptible mounts don't hang here for a long time.
 270                  */
 271                 crit_enter();
 272                 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
 273                         (void) tsleep((caddr_t)&so->so_timeo, 0,
 274                                 "nfscon", 2 * hz);
 275                         if ((so->so_state & SS_ISCONNECTING) &&
 276                             so->so_error == 0 && rep &&
 277                             (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
 278                                 soclrstate(so, SS_ISCONNECTING);
 279                                 crit_exit();
 280                                 goto bad;
 281                         }
 282                 }
 283                 if (so->so_error) {
 284                         error = so->so_error;
 285                         so->so_error = 0;
 286                         crit_exit();
 287                         goto bad;
 288                 }
 289                 crit_exit();
 290         }
 291         so->so_rcv.ssb_timeo = (5 * hz);
 292         so->so_snd.ssb_timeo = (5 * hz);
 293
 294         /*
 295          * Get buffer reservation size from sysctl, but impose reasonable
 296          * limits.
 297          */
 298         if (nmp->nm_sotype == SOCK_STREAM) {
 299                 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
 300                         struct sockopt sopt;
 301                         int val;
 302
 303                         bzero(&sopt, sizeof sopt);
 304                         sopt.sopt_level = SOL_SOCKET;
 305                         sopt.sopt_name = SO_KEEPALIVE;
 306                         sopt.sopt_val = &val;
 307                         sopt.sopt_valsize = sizeof val;
 308                         val = 1;
 309                         sosetopt(so, &sopt);
 310                 }
 311                 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
 312                         struct sockopt sopt;
 313                         int val;
 314
 315                         bzero(&sopt, sizeof sopt);
 316                         sopt.sopt_level = IPPROTO_TCP;
 317                         sopt.sopt_name = TCP_NODELAY;
 318                         sopt.sopt_val = &val;
 319                         sopt.sopt_valsize = sizeof val;
 320                         val = 1;
 321                         sosetopt(so, &sopt);
 322
 323                         bzero(&sopt, sizeof sopt);
 324                         sopt.sopt_level = IPPROTO_TCP;
 325                         sopt.sopt_name = TCP_FASTKEEP;
 326                         sopt.sopt_val = &val;
 327                         sopt.sopt_valsize = sizeof val;
 328                         val = 1;
 329                         sosetopt(so, &sopt);
 330                 }
 331         }
 332         error = soreserve(so, nfs_soreserve, nfs_soreserve, NULL);
 333         if (error)
 334                 goto bad;
 335         atomic_set_int(&so->so_rcv.ssb_flags, SSB_NOINTR);
 336         atomic_set_int(&so->so_snd.ssb_flags, SSB_NOINTR);
 337
 338         /* Initialize other non-zero congestion variables */
 339         nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] =
 340                 nmp->nm_srtt[3] = (NFS_TIMEO << NFS_RTT_SCALE_BITS);
 341         nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
 342                 nmp->nm_sdrtt[3] = 0;
 343         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
 344         nmp->nm_timeouts = 0;
 345
 346         /*
 347          * Assign nm_so last.  The moment nm_so is assigned the nfs_timer()
 348          * can mess with the socket.
 349          */
 350         nmp->nm_so = so;
 351         return (0);
 352
 353 bad:
 354         if (so) {
 355                 soshutdown(so, SHUT_RDWR);
 356                 soclose(so, FNONBLOCK);
 357         }
 358         return (error);
 359 }
 360
 361 /*
 362  * Reconnect routine:
 363  * Called when a connection is broken on a reliable protocol.
 364  * - clean up the old socket
 365  * - nfs_connect() again
 366  * - set R_NEEDSXMIT for all outstanding requests on mount point
 367  * If this fails the mount point is DEAD!
 368  * nb: Must be called with the nfs_sndlock() set on the mount point.
 369  */
 370 static int
 371 nfs_reconnect(struct nfsmount *nmp, struct nfsreq *rep)
 372 {
 373         struct nfsreq *req;
 374         int error;
 375
 376         nfs_disconnect(nmp);
 377         if (nmp->nm_rxstate >= NFSSVC_STOPPING)
 378                 return (EINTR);
 379         while ((error = nfs_connect(nmp, rep)) != 0) {
 380                 if (error == EINTR || error == ERESTART)
 381                         return (EINTR);
 382                 if (error == EINVAL)
 383                         return (error);
 384                 if (nmp->nm_rxstate >= NFSSVC_STOPPING)
 385                         return (EINTR);
 386                 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
 387         }
 388
 389         /*
 390          * Loop through outstanding request list and fix up all requests
 391          * on old socket.
 392          */
 393         crit_enter();
 394         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
 395                 KKASSERT(req->r_nmp == nmp);
 396                 req->r_flags |= R_NEEDSXMIT;
 397         }
 398         crit_exit();
 399         return (0);
 400 }
 401
 402 /*
 403  * NFS disconnect. Clean up and unlink.
 404  */
 405 void
 406 nfs_disconnect(struct nfsmount *nmp)
 407 {
 408         struct socket *so;
 409
 410         if (nmp->nm_so) {
 411                 so = nmp->nm_so;
 412                 nmp->nm_so = NULL;
 413                 soshutdown(so, SHUT_RDWR);
 414                 soclose(so, FNONBLOCK);
 415         }
 416 }
 417
 418 void
 419 nfs_safedisconnect(struct nfsmount *nmp)
 420 {
 421         nfs_rcvlock(nmp, NULL);
 422         nfs_disconnect(nmp);
 423         nfs_rcvunlock(nmp);
 424 }
 425
 426 /*
 427  * This is the nfs send routine. For connection based socket types, it
 428  * must be called with an nfs_sndlock() on the socket.
 429  * "rep == NULL" indicates that it has been called from a server.
 430  * For the client side:
 431  * - return EINTR if the RPC is terminated, 0 otherwise
 432  * - set R_NEEDSXMIT if the send fails for any reason
 433  * - do any cleanup required by recoverable socket errors (?)
 434  * For the server side:
 435  * - return EINTR or ERESTART if interrupted by a signal
 436  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
 437  * - do any cleanup required by recoverable socket errors (?)
 438  */
 439 int
 440 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
 441          struct nfsreq *rep)
 442 {
 443         struct sockaddr *sendnam;
 444         int error, soflags, flags;
 445
 446         if (rep) {
 447                 if (rep->r_flags & R_SOFTTERM) {
 448                         m_freem(top);
 449                         return (EINTR);
 450                 }
 451                 if ((so = rep->r_nmp->nm_so) == NULL) {
 452                         rep->r_flags |= R_NEEDSXMIT;
 453                         m_freem(top);
 454                         return (0);
 455                 }
 456                 rep->r_flags &= ~R_NEEDSXMIT;
 457                 soflags = rep->r_nmp->nm_soflags;
 458         } else {
 459                 soflags = so->so_proto->pr_flags;
 460         }
 461         if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
 462                 sendnam = NULL;
 463         else
 464                 sendnam = nam;
 465         if (so->so_type == SOCK_SEQPACKET)
 466                 flags = MSG_EOR;
 467         else
 468                 flags = 0;
 469
 470         /*
 471          * calls pru_sosend -> sosend -> so_pru_send -> netrpc
 472          */
 473         error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
 474                               curthread /*XXX*/);
 475
 476         /*
 477          * ENOBUFS for dgram sockets is transient and non fatal.
 478          * No need to log, and no need to break a soft mount.
 479          */
 480         if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
 481                 error = 0;
 482                 /*
 483                  * do backoff retransmit on client
 484                  */
 485                 if (rep) {
 486                         if ((rep->r_nmp->nm_state & NFSSTA_SENDSPACE) == 0) {
 487                                 rep->r_nmp->nm_state |= NFSSTA_SENDSPACE;
 488                                 kprintf("Warning: NFS: Insufficient sendspace "
 489                                         "(%lu),\n"
 490                                         "\t You must increase vfs.nfs.soreserve"
 491                                         "or decrease vfs.nfs.maxasyncbio\n",
 492                                         so->so_snd.ssb_hiwat);
 493                         }
 494                         rep->r_flags |= R_NEEDSXMIT;
 495                 }
 496         }
 497
 498         if (error) {
 499                 if (rep) {
 500                         log(LOG_INFO, "nfs send error %d for server %s\n",error,
 501                             rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
 502                         /*
 503                          * Deal with errors for the client side.
 504                          */
 505                         if (rep->r_flags & R_SOFTTERM)
 506                                 error = EINTR;
 507                         else
 508                                 rep->r_flags |= R_NEEDSXMIT;
 509                 } else {
 510                         log(LOG_INFO, "nfsd send error %d\n", error);
 511                 }
 512
 513                 /*
 514                  * Handle any recoverable (soft) socket errors here. (?)
 515                  */
 516                 if (error != EINTR && error != ERESTART &&
 517                         error != EWOULDBLOCK && error != EPIPE)
 518                         error = 0;
 519         }
 520         return (error);
 521 }
 522
 523 /*
 524  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
 525  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
 526  * Mark and consolidate the data into a new mbuf list.
 527  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
 528  *     small mbufs.
 529  * For SOCK_STREAM we must be very careful to read an entire record once
 530  * we have read any of it, even if the system call has been interrupted.
 531  */
 532 static int
 533 nfs_receive(struct nfsmount *nmp, struct nfsreq *rep,
 534             struct sockaddr **aname, struct mbuf **mp)
 535 {
 536         struct socket *so;
 537         struct sockbuf sio;
 538         struct uio auio;
 539         struct iovec aio;
 540         struct mbuf *m;
 541         struct mbuf *control;
 542         u_int32_t len;
 543         struct sockaddr **getnam;
 544         int error, sotype, rcvflg;
 545         struct thread *td = curthread;  /* XXX */
 546
 547         /*
 548          * Set up arguments for soreceive()
 549          */
 550         *mp = NULL;
 551         *aname = NULL;
 552         sotype = nmp->nm_sotype;
 553
 554         /*
 555          * For reliable protocols, lock against other senders/receivers
 556          * in case a reconnect is necessary.
 557          * For SOCK_STREAM, first get the Record Mark to find out how much
 558          * more there is to get.
 559          * We must lock the socket against other receivers
 560          * until we have an entire rpc request/reply.
 561          */
 562         if (sotype != SOCK_DGRAM) {
 563                 error = nfs_sndlock(nmp, rep);
 564                 if (error)
 565                         return (error);
 566 tryagain:
 567                 /*
 568                  * Check for fatal errors and resending request.
 569                  */
 570                 /*
 571                  * Ugh: If a reconnect attempt just happened, nm_so
 572                  * would have changed. NULL indicates a failed
 573                  * attempt that has essentially shut down this
 574                  * mount point.
 575                  */
 576                 if (rep && (rep->r_mrep || (rep->r_flags & R_SOFTTERM))) {
 577                         nfs_sndunlock(nmp);
 578                         return (EINTR);
 579                 }
 580                 so = nmp->nm_so;
 581                 if (so == NULL) {
 582                         error = nfs_reconnect(nmp, rep);
 583                         if (error) {
 584                                 nfs_sndunlock(nmp);
 585                                 return (error);
 586                         }
 587                         goto tryagain;
 588                 }
 589                 while (rep && (rep->r_flags & R_NEEDSXMIT)) {
 590                         m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
 591                         nfsstats.rpcretries++;
 592                         error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
 593                         if (error) {
 594                                 if (error == EINTR || error == ERESTART ||
 595                                     (error = nfs_reconnect(nmp, rep)) != 0) {
 596                                         nfs_sndunlock(nmp);
 597                                         return (error);
 598                                 }
 599                                 goto tryagain;
 600                         }
 601                 }
 602                 nfs_sndunlock(nmp);
 603                 if (sotype == SOCK_STREAM) {
 604                         /*
 605                          * Get the length marker from the stream
 606                          */
 607                         aio.iov_base = (caddr_t)&len;
 608                         aio.iov_len = sizeof(u_int32_t);
 609                         auio.uio_iov = &aio;
 610                         auio.uio_iovcnt = 1;
 611                         auio.uio_segflg = UIO_SYSSPACE;
 612                         auio.uio_rw = UIO_READ;
 613                         auio.uio_offset = 0;
 614                         auio.uio_resid = sizeof(u_int32_t);
 615                         auio.uio_td = td;
 616                         do {
 617                            rcvflg = MSG_WAITALL;
 618                            error = so_pru_soreceive(so, NULL, &auio, NULL,
 619                                                     NULL, &rcvflg);
 620                            if (error == EWOULDBLOCK && rep) {
 621                                 if (rep->r_flags & R_SOFTTERM)
 622                                         return (EINTR);
 623                            }
 624                         } while (error == EWOULDBLOCK);
 625
 626                         if (error == 0 && auio.uio_resid > 0) {
 627                             /*
 628                              * Only log short packets if not EOF
 629                              */
 630                             if (auio.uio_resid != sizeof(u_int32_t)) {
 631                                 log(LOG_INFO,
 632                                     "short receive (%d/%d) from nfs server %s\n",
 633                                     (int)(sizeof(u_int32_t) - auio.uio_resid),
 634                                     (int)sizeof(u_int32_t),
 635                                     nmp->nm_mountp->mnt_stat.f_mntfromname);
 636                             }
 637                             error = EPIPE;
 638                         }
 639                         if (error)
 640                                 goto errout;
 641                         len = ntohl(len) & ~0x80000000;
 642                         /*
 643                          * This is SERIOUS! We are out of sync with the sender
 644                          * and forcing a disconnect/reconnect is all I can do.
 645                          */
 646                         if (len > NFS_MAXPACKET) {
 647                             log(LOG_ERR, "%s (%d) from nfs server %s\n",
 648                                 "impossible packet length",
 649                                 len,
 650                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
 651                             error = EFBIG;
 652                             goto errout;
 653                         }
 654
 655                         /*
 656                          * Get the rest of the packet as an mbuf chain
 657                          */
 658                         sbinit(&sio, len);
 659                         do {
 660                             rcvflg = MSG_WAITALL;
 661                             error = so_pru_soreceive(so, NULL, NULL, &sio,
 662                                                      NULL, &rcvflg);
 663                         } while (error == EWOULDBLOCK || error == EINTR ||
 664                                  error == ERESTART);
 665                         if (error == 0 && sio.sb_cc != len) {
 666                             if (sio.sb_cc != 0) {
 667                                 log(LOG_INFO,
 668                                     "short receive (%zu/%d) from nfs server %s\n",
 669                                     (size_t)len - auio.uio_resid, len,
 670                                     nmp->nm_mountp->mnt_stat.f_mntfromname);
 671                             }
 672                             error = EPIPE;
 673                         }
 674                         *mp = sio.sb_mb;
 675                 } else {
 676                         /*
 677                          * Non-stream, so get the whole packet by not
 678                          * specifying MSG_WAITALL and by specifying a large
 679                          * length.
 680                          *
 681                          * We have no use for control msg., but must grab them
 682                          * and then throw them away so we know what is going
 683                          * on.
 684                          */
 685                         sbinit(&sio, 100000000);
 686                         do {
 687                             rcvflg = 0;
 688                             error =  so_pru_soreceive(so, NULL, NULL, &sio,
 689                                                       &control, &rcvflg);
 690                             if (control)
 691                                 m_freem(control);
 692                             if (error == EWOULDBLOCK && rep) {
 693                                 if (rep->r_flags & R_SOFTTERM) {
 694                                         m_freem(sio.sb_mb);
 695                                         return (EINTR);
 696                                 }
 697                             }
 698                         } while (error == EWOULDBLOCK ||
 699                                  (error == 0 && sio.sb_mb == NULL && control));
 700                         if ((rcvflg & MSG_EOR) == 0)
 701                                 kprintf("Egad!!\n");
 702                         if (error == 0 && sio.sb_mb == NULL)
 703                                 error = EPIPE;
 704                         len = sio.sb_cc;
 705                         *mp = sio.sb_mb;
 706                 }
 707 errout:
 708                 if (error && error != EINTR && error != ERESTART) {
 709                         m_freem(*mp);
 710                         *mp = NULL;
 711                         if (error != EPIPE) {
 712                                 log(LOG_INFO,
 713                                     "receive error %d from nfs server %s\n",
 714                                     error,
 715                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
 716                         }
 717                         error = nfs_sndlock(nmp, rep);
 718                         if (!error) {
 719                                 error = nfs_reconnect(nmp, rep);
 720                                 if (!error)
 721                                         goto tryagain;
 722                                 else
 723                                         nfs_sndunlock(nmp);
 724                         }
 725                 }
 726         } else {
 727                 if ((so = nmp->nm_so) == NULL)
 728                         return (EACCES);
 729                 if (so->so_state & SS_ISCONNECTED)
 730                         getnam = NULL;
 731                 else
 732                         getnam = aname;
 733                 sbinit(&sio, 100000000);
 734                 do {
 735                         rcvflg = 0;
 736                         error =  so_pru_soreceive(so, getnam, NULL, &sio,
 737                                                   NULL, &rcvflg);
 738                         if (error == EWOULDBLOCK && rep &&
 739                             (rep->r_flags & R_SOFTTERM)) {
 740                                 m_freem(sio.sb_mb);
 741                                 return (EINTR);
 742                         }
 743                 } while (error == EWOULDBLOCK);
 744
 745                 len = sio.sb_cc;
 746                 *mp = sio.sb_mb;
 747
 748                 /*
 749                  * A shutdown may result in no error and no mbuf.
 750                  * Convert to EPIPE.
 751                  */
 752                 if (*mp == NULL && error == 0)
 753                         error = EPIPE;
 754         }
 755         if (error) {
 756                 m_freem(*mp);
 757                 *mp = NULL;
 758         }
 759
 760         /*
 761          * Search for any mbufs that are not a multiple of 4 bytes long
 762          * or with m_data not longword aligned.
 763          * These could cause pointer alignment problems, so copy them to
 764          * well aligned mbufs.
 765          */
 766         nfs_realign(mp, 5 * NFSX_UNSIGNED);
 767         return (error);
 768 }
 769
 770 /*
 771  * Implement receipt of reply on a socket.
 772  *
 773  * We must search through the list of received datagrams matching them
 774  * with outstanding requests using the xid, until ours is found.
 775  *
 776  * If myrep is NULL we process packets on the socket until
 777  * interrupted or until nm_reqrxq is non-empty.
 778  */
 779 /* ARGSUSED */
 780 int
 781 nfs_reply(struct nfsmount *nmp, struct nfsreq *myrep)
 782 {
 783         struct nfsreq *rep;
 784         struct sockaddr *nam;
 785         u_int32_t rxid;
 786         u_int32_t *tl;
 787         int error;
 788         struct nfsm_info info;
 789
 790         /*
 791          * Loop around until we get our own reply
 792          */
 793         for (;;) {
 794                 /*
 795                  * Lock against other receivers so that I don't get stuck in
 796                  * sbwait() after someone else has received my reply for me.
 797                  * Also necessary for connection based protocols to avoid
 798                  * race conditions during a reconnect.
 799                  *
 800                  * If nfs_rcvlock() returns EALREADY, that means that
 801                  * the reply has already been recieved by another
 802                  * process and we can return immediately.  In this
 803                  * case, the lock is not taken to avoid races with
 804                  * other processes.
 805                  */
 806                 info.mrep = NULL;
 807
 808                 error = nfs_rcvlock(nmp, myrep);
 809                 if (error == EALREADY)
 810                         return (0);
 811                 if (error)
 812                         return (error);
 813
 814                 /*
 815                  * If myrep is NULL we are the receiver helper thread.
 816                  * Stop waiting for incoming replies if there are
 817                  * messages sitting on reqrxq that we need to process,
 818                  * or if a shutdown request is pending.
 819                  */
 820                 if (myrep == NULL && (TAILQ_FIRST(&nmp->nm_reqrxq) ||
 821                     nmp->nm_rxstate > NFSSVC_PENDING)) {
 822                         nfs_rcvunlock(nmp);
 823                         return(EWOULDBLOCK);
 824                 }
 825
 826                 /*
 827                  * Get the next Rpc reply off the socket
 828                  *
 829                  * We cannot release the receive lock until we've
 830                  * filled in rep->r_mrep, otherwise a waiting
 831                  * thread may deadlock in soreceive with no incoming
 832                  * packets expected.
 833                  */
 834                 error = nfs_receive(nmp, myrep, &nam, &info.mrep);
 835                 if (error) {
 836                         /*
 837                          * Ignore routing errors on connectionless protocols??
 838                          */
 839                         nfs_rcvunlock(nmp);
 840                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
 841                                 if (nmp->nm_so == NULL)
 842                                         return (error);
 843                                 nmp->nm_so->so_error = 0;
 844                                 continue;
 845                         }
 846                         return (error);
 847                 }
 848                 if (nam)
 849                         kfree(nam, M_SONAME);
 850
 851                 /*
 852                  * Get the xid and check that it is an rpc reply
 853                  */
 854                 info.md = info.mrep;
 855                 info.dpos = mtod(info.md, caddr_t);
 856                 NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
 857                 rxid = *tl++;
 858                 if (*tl != rpc_reply) {
 859                         nfsstats.rpcinvalid++;
 860                         m_freem(info.mrep);
 861                         info.mrep = NULL;
 862 nfsmout:
 863                         nfs_rcvunlock(nmp);
 864                         continue;
 865                 }
 866
 867                 /*
 868                  * Loop through the request list to match up the reply
 869                  * Iff no match, just drop the datagram.  On match, set
 870                  * r_mrep atomically to prevent the timer from messing
 871                  * around with the request after we have exited the critical
 872                  * section.
 873                  */
 874                 crit_enter();
 875                 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
 876                         if (rep->r_mrep == NULL && rxid == rep->r_xid)
 877                                 break;
 878                 }
 879
 880                 /*
 881                  * Fill in the rest of the reply if we found a match.
 882                  *
 883                  * Deal with duplicate responses if there was no match.
 884                  */
 885                 if (rep) {
 886                         rep->r_md = info.md;
 887                         rep->r_dpos = info.dpos;
 888                         if (nfsrtton) {
 889                                 struct rttl *rt;
 890
 891                                 rt = &nfsrtt.rttl[nfsrtt.pos];
 892                                 rt->proc = rep->r_procnum;
 893                                 rt->rto = 0;
 894                                 rt->sent = 0;
 895                                 rt->cwnd = nmp->nm_maxasync_scaled;
 896                                 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
 897                                 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
 898                                 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
 899                                 getmicrotime(&rt->tstamp);
 900                                 if (rep->r_flags & R_TIMING)
 901                                         rt->rtt = rep->r_rtt;
 902                                 else
 903                                         rt->rtt = 1000000;
 904                                 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
 905                         }
 906
 907                         /*
 908                          * New congestion control is based only on async
 909                          * requests.
 910                          */
 911                         if (nmp->nm_maxasync_scaled < NFS_MAXASYNC_SCALED)
 912                                 ++nmp->nm_maxasync_scaled;
 913                         if (rep->r_flags & R_SENT) {
 914                                 rep->r_flags &= ~R_SENT;
 915                         }
 916                         /*
 917                          * Update rtt using a gain of 0.125 on the mean
 918                          * and a gain of 0.25 on the deviation.
 919                          *
 920                          * NOTE SRTT/SDRTT are only good if R_TIMING is set.
 921                          */
 922                         if ((rep->r_flags & R_TIMING) && rep->r_rexmit == 0) {
 923                                 /*
 924                                  * Since the timer resolution of
 925                                  * NFS_HZ is so course, it can often
 926                                  * result in r_rtt == 0. Since
 927                                  * r_rtt == N means that the actual
 928                                  * rtt is between N+dt and N+2-dt ticks,
 929                                  * add 1.
 930                                  */
 931                                 int n;
 932                                 int d;
 933
 934 #define NFSRSB  NFS_RTT_SCALE_BITS
 935                                 n = ((NFS_SRTT(rep) * 7) +
 936                                      (rep->r_rtt << NFSRSB)) >> 3;
 937                                 d = n - NFS_SRTT(rep);
 938                                 NFS_SRTT(rep) = n;
 939
 940                                 /*
 941                                  * Don't let the jitter calculation decay
 942                                  * too quickly, but we want a fast rampup.
 943                                  */
 944                                 if (d < 0)
 945                                         d = -d;
 946                                 d <<= NFSRSB;
 947                                 if (d < NFS_SDRTT(rep))
 948                                         n = ((NFS_SDRTT(rep) * 15) + d) >> 4;
 949                                 else
 950                                         n = ((NFS_SDRTT(rep) * 3) + d) >> 2;
 951                                 NFS_SDRTT(rep) = n;
 952 #undef NFSRSB
 953                         }
 954                         nmp->nm_timeouts = 0;
 955                         rep->r_mrep = info.mrep;
 956                         nfs_hardterm(rep, 0);
 957                 } else {
 958                         /*
 959                          * Extract vers, prog, nfsver, procnum.  A duplicate
 960                          * response means we didn't wait long enough so
 961                          * we increase the SRTT to avoid future spurious
 962                          * timeouts.
 963                          */
 964                         u_int procnum = nmp->nm_lastreprocnum;
 965                         int n;
 966
 967                         if (procnum < NFS_NPROCS && proct[procnum]) {
 968                                 if (nfs_showrexmit)
 969                                         kprintf("D");
 970                                 n = nmp->nm_srtt[proct[procnum]];
 971                                 n += NFS_ASYSCALE * NFS_HZ;
 972                                 if (n < NFS_ASYSCALE * NFS_HZ * 10)
 973                                         n = NFS_ASYSCALE * NFS_HZ * 10;
 974                                 nmp->nm_srtt[proct[procnum]] = n;
 975                         }
 976                 }
 977                 nfs_rcvunlock(nmp);
 978                 crit_exit();
 979
 980                 /*
 981                  * If not matched to a request, drop it.
 982                  * If it's mine, get out.
 983                  */
 984                 if (rep == NULL) {
 985                         nfsstats.rpcunexpected++;
 986                         m_freem(info.mrep);
 987                         info.mrep = NULL;
 988                 } else if (rep == myrep) {
 989                         if (rep->r_mrep == NULL)
 990                                 panic("nfsreply nil");
 991                         return (0);
 992                 }
 993         }
 994 }
 995
 996 /*
 997  * Run the request state machine until the target state is reached
 998  * or a fatal error occurs.  The target state is not run.  Specifying
 999  * a target of NFSM_STATE_DONE runs the state machine until the rpc
1000  * is complete.
1001  *
1002  * EINPROGRESS is returned for all states other then the DONE state,
1003  * indicating that the rpc is still in progress.
1004  */
1005 int
1006 nfs_request(struct nfsm_info *info, nfsm_state_t bstate, nfsm_state_t estate)
1007 {
1008         struct nfsreq *req;
1009
1010         while (info->state >= bstate && info->state < estate) {
1011                 switch(info->state) {
1012                 case NFSM_STATE_SETUP:
1013                         /*
1014                          * Setup the nfsreq.  Any error which occurs during
1015                          * this state is fatal.
1016                          */
1017                         info->error = nfs_request_setup(info);
1018                         if (info->error) {
1019                                 info->state = NFSM_STATE_DONE;
1020                                 return (info->error);
1021                         } else {
1022                                 req = info->req;
1023                                 req->r_mrp = &info->mrep;
1024                                 req->r_mdp = &info->md;
1025                                 req->r_dposp = &info->dpos;
1026                                 info->state = NFSM_STATE_AUTH;
1027                         }
1028                         break;
1029                 case NFSM_STATE_AUTH:
1030                         /*
1031                          * Authenticate the nfsreq.  Any error which occurs
1032                          * during this state is fatal.
1033                          */
1034                         info->error = nfs_request_auth(info->req);
1035                         if (info->error) {
1036                                 info->state = NFSM_STATE_DONE;
1037                                 return (info->error);
1038                         } else {
1039                                 info->state = NFSM_STATE_TRY;
1040                         }
1041                         break;
1042                 case NFSM_STATE_TRY:
1043                         /*
1044                          * Transmit or retransmit attempt.  An error in this
1045                          * state is ignored and we always move on to the
1046                          * next state.
1047                          *
1048                          * This can trivially race the receiver if the
1049                          * request is asynchronous.  nfs_request_try()
1050                          * will thus set the state for us and we
1051                          * must also return immediately if we are
1052                          * running an async state machine, because
1053                          * info can become invalid due to races after
1054                          * try() returns.
1055                          */
1056                         if (info->req->r_flags & R_ASYNC) {
1057                                 nfs_request_try(info->req);
1058                                 if (estate == NFSM_STATE_WAITREPLY)
1059                                         return (EINPROGRESS);
1060                         } else {
1061                                 nfs_request_try(info->req);
1062                                 info->state = NFSM_STATE_WAITREPLY;
1063                         }
1064                         break;
1065                 case NFSM_STATE_WAITREPLY:
1066                         /*
1067                          * Wait for a reply or timeout and move on to the
1068                          * next state.  The error returned by this state
1069                          * is passed to the processing code in the next
1070                          * state.
1071                          */
1072                         info->error = nfs_request_waitreply(info->req);
1073                         info->state = NFSM_STATE_PROCESSREPLY;
1074                         break;
1075                 case NFSM_STATE_PROCESSREPLY:
1076                         /*
1077                          * Process the reply or timeout.  Errors which occur
1078                          * in this state may cause the state machine to
1079                          * go back to an earlier state, and are fatal
1080                          * otherwise.
1081                          */
1082                         info->error = nfs_request_processreply(info,
1083                                                                info->error);
1084                         switch(info->error) {
1085                         case ENEEDAUTH:
1086                                 info->state = NFSM_STATE_AUTH;
1087                                 break;
1088                         case EAGAIN:
1089                                 info->state = NFSM_STATE_TRY;
1090                                 break;
1091                         default:
1092                                 /*
1093                                  * Operation complete, with or without an
1094                                  * error.  We are done.
1095                                  */
1096                                 info->req = NULL;
1097                                 info->state = NFSM_STATE_DONE;
1098                                 return (info->error);
1099                         }
1100                         break;
1101                 case NFSM_STATE_DONE:
1102                         /*
1103                          * Shouldn't be reached
1104                          */
1105                         return (info->error);
1106                         /* NOT REACHED */
1107                 }
1108         }
1109
1110         /*
1111          * If we are done return the error code (if any).
1112          * Otherwise return EINPROGRESS.
1113          */
1114         if (info->state == NFSM_STATE_DONE)
1115                 return (info->error);
1116         return (EINPROGRESS);
1117 }
1118
1119 /*
1120  * nfs_request - goes something like this
1121  *      - fill in request struct
1122  *      - links it into list
1123  *      - calls nfs_send() for first transmit
1124  *      - calls nfs_receive() to get reply
1125  *      - break down rpc header and return with nfs reply pointed to
1126  *        by mrep or error
1127  * nb: always frees up mreq mbuf list
1128  */
1129 static int
1130 nfs_request_setup(nfsm_info_t info)
1131 {
1132         struct nfsreq *req;
1133         struct nfsmount *nmp;
1134         struct mbuf *m;
1135         int i;
1136
1137         /*
1138          * Reject requests while attempting a forced unmount.
1139          */
1140         if (info->vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
1141                 m_freem(info->mreq);
1142                 info->mreq = NULL;
1143                 return (EIO);
1144         }
1145         nmp = VFSTONFS(info->vp->v_mount);
1146         req = kmalloc(sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
1147         req->r_nmp = nmp;
1148         req->r_vp = info->vp;
1149         req->r_td = info->td;
1150         req->r_procnum = info->procnum;
1151         req->r_mreq = NULL;
1152         req->r_cred = info->cred;
1153
1154         i = 0;
1155         m = info->mreq;
1156         while (m) {
1157                 i += m->m_len;
1158                 m = m->m_next;
1159         }
1160         req->r_mrest = info->mreq;
1161         req->r_mrest_len = i;
1162
1163         /*
1164          * The presence of a non-NULL r_info in req indicates
1165          * async completion via our helper threads.  See the receiver
1166          * code.
1167          */
1168         if (info->bio) {
1169                 req->r_info = info;
1170                 req->r_flags = R_ASYNC;
1171         } else {
1172                 req->r_info = NULL;
1173                 req->r_flags = 0;
1174         }
1175         info->req = req;
1176         return(0);
1177 }
1178
1179 static int
1180 nfs_request_auth(struct nfsreq *rep)
1181 {
1182         struct nfsmount *nmp = rep->r_nmp;
1183         struct mbuf *m;
1184         char nickv[RPCX_NICKVERF];
1185         int error = 0, auth_len, auth_type;
1186         int verf_len;
1187         u_int32_t xid;
1188         char *auth_str, *verf_str;
1189         struct ucred *cred;
1190
1191         cred = rep->r_cred;
1192         rep->r_failed_auth = 0;
1193
1194         /*
1195          * Get the RPC header with authorization.
1196          */
1197         verf_str = auth_str = NULL;
1198         if (nmp->nm_flag & NFSMNT_KERB) {
1199                 verf_str = nickv;
1200                 verf_len = sizeof (nickv);
1201                 auth_type = RPCAUTH_KERB4;
1202                 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1203                 if (rep->r_failed_auth ||
1204                     nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1205                                     verf_str, verf_len)) {
1206                         error = nfs_getauth(nmp, rep, cred, &auth_str,
1207                                 &auth_len, verf_str, &verf_len, rep->r_key);
1208                         if (error) {
1209                                 m_freem(rep->r_mrest);
1210                                 rep->r_mrest = NULL;
1211                                 kfree((caddr_t)rep, M_NFSREQ);
1212                                 return (error);
1213                         }
1214                 }
1215         } else {
1216                 auth_type = RPCAUTH_UNIX;
1217                 if (cred->cr_ngroups < 1)
1218                         panic("nfsreq nogrps");
1219                 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1220                         nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1221                         5 * NFSX_UNSIGNED;
1222         }
1223         if (rep->r_mrest)
1224                 nfs_checkpkt(rep->r_mrest, rep->r_mrest_len);
1225         m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1226                         auth_len, auth_str, verf_len, verf_str,
1227                         rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1228         rep->r_mrest = NULL;
1229         if (auth_str)
1230                 kfree(auth_str, M_TEMP);
1231
1232         /*
1233          * For stream protocols, insert a Sun RPC Record Mark.
1234          */
1235         if (nmp->nm_sotype == SOCK_STREAM) {
1236                 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1237                 if (m == NULL) {
1238                         kfree(rep, M_NFSREQ);
1239                         return (ENOBUFS);
1240                 }
1241                 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1242                          (m->m_pkthdr.len - NFSX_UNSIGNED));
1243         }
1244
1245         nfs_checkpkt(m, m->m_pkthdr.len);
1246
1247         rep->r_mreq = m;
1248         rep->r_xid = xid;
1249         return (0);
1250 }
1251
1252 static int
1253 nfs_request_try(struct nfsreq *rep)
1254 {
1255         struct nfsmount *nmp = rep->r_nmp;
1256         struct mbuf *m2;
1257         int error;
1258
1259         /*
1260          * Request is not on any queue, only the owner has access to it
1261          * so it should not be locked by anyone atm.
1262          *
1263          * Interlock to prevent races.  While locked the only remote
1264          * action possible is for r_mrep to be set (once we enqueue it).
1265          */
1266         if (rep->r_flags == 0xdeadc0de) {
1267                 print_backtrace(-1);
1268                 panic("flags nbad");
1269         }
1270         KKASSERT((rep->r_flags & (R_LOCKED | R_ONREQQ)) == 0);
1271         if (nmp->nm_flag & NFSMNT_SOFT)
1272                 rep->r_retry = nmp->nm_retry;
1273         else
1274                 rep->r_retry = NFS_MAXREXMIT + 1;       /* past clip limit */
1275         rep->r_rtt = rep->r_rexmit = 0;
1276         if (proct[rep->r_procnum] > 0)
1277                 rep->r_flags |= R_TIMING | R_LOCKED;
1278         else
1279                 rep->r_flags |= R_LOCKED;
1280         rep->r_mrep = NULL;
1281
1282         nfsstats.rpcrequests++;
1283
1284         if (nmp->nm_flag & NFSMNT_FORCE) {
1285                 rep->r_flags |= R_SOFTTERM;
1286                 rep->r_flags &= ~R_LOCKED;
1287                 return (0);
1288         }
1289         rep->r_flags |= R_NEEDSXMIT;    /* in case send lock races us */
1290
1291         /*
1292          * Do the client side RPC.
1293          *
1294          * Chain request into list of outstanding requests. Be sure
1295          * to put it LAST so timer finds oldest requests first.  Note
1296          * that our control of R_LOCKED prevents the request from
1297          * getting ripped out from under us or transmitted by the
1298          * timer code.
1299          *
1300          * For requests with info structures we must atomically set the
1301          * info's state because the structure could become invalid upon
1302          * return due to races (i.e., if async)
1303          */
1304         crit_enter();
1305         mtx_link_init(&rep->r_link);
1306         KKASSERT((rep->r_flags & R_ONREQQ) == 0);
1307         TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1308         rep->r_flags |= R_ONREQQ;
1309         ++nmp->nm_reqqlen;
1310         if (rep->r_flags & R_ASYNC)
1311                 rep->r_info->state = NFSM_STATE_WAITREPLY;
1312         crit_exit();
1313
1314         error = 0;
1315
1316         /*
1317          * Send if we can.  Congestion control is not handled here any more
1318          * becausing trying to defer the initial send based on the nfs_timer
1319          * requires having a very fast nfs_timer, which is silly.
1320          */
1321         if (nmp->nm_so) {
1322                 if (nmp->nm_soflags & PR_CONNREQUIRED)
1323                         error = nfs_sndlock(nmp, rep);
1324                 if (error == 0 && (rep->r_flags & R_NEEDSXMIT)) {
1325                         m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1326                         error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1327                         rep->r_flags &= ~R_NEEDSXMIT;
1328                         if ((rep->r_flags & R_SENT) == 0) {
1329                                 rep->r_flags |= R_SENT;
1330                         }
1331                         if (nmp->nm_soflags & PR_CONNREQUIRED)
1332                                 nfs_sndunlock(nmp);
1333                 }
1334         } else {
1335                 rep->r_rtt = -1;
1336         }
1337         if (error == EPIPE)
1338                 error = 0;
1339
1340         /*
1341          * Release the lock.  The only remote action that may have occurred
1342          * would have been the setting of rep->r_mrep.  If this occured
1343          * and the request was async we have to move it to the reader
1344          * thread's queue for action.
1345          *
1346          * For async requests also make sure the reader is woken up so
1347          * it gets on the socket to read responses.
1348          */
1349         crit_enter();
1350         if (rep->r_flags & R_ASYNC) {
1351                 if (rep->r_mrep)
1352                         nfs_hardterm(rep, 1);
1353                 rep->r_flags &= ~R_LOCKED;
1354                 nfssvc_iod_reader_wakeup(nmp);
1355         } else {
1356                 rep->r_flags &= ~R_LOCKED;
1357         }
1358         if (rep->r_flags & R_WANTED) {
1359                 rep->r_flags &= ~R_WANTED;
1360                 wakeup(rep);
1361         }
1362         crit_exit();
1363         return (error);
1364 }
1365
1366 /*
1367  * This code is only called for synchronous requests.  Completed synchronous
1368  * requests are left on reqq and we remove them before moving on to the
1369  * processing state.
1370  */
1371 static int
1372 nfs_request_waitreply(struct nfsreq *rep)
1373 {
1374         struct nfsmount *nmp = rep->r_nmp;
1375         int error;
1376
1377         KKASSERT((rep->r_flags & R_ASYNC) == 0);
1378
1379         /*
1380          * Wait until the request is finished.
1381          */
1382         error = nfs_reply(nmp, rep);
1383
1384         /*
1385          * RPC done, unlink the request, but don't rip it out from under
1386          * the callout timer.
1387          *
1388          * Once unlinked no other receiver or the timer will have
1389          * visibility, so we do not have to set R_LOCKED.
1390          */
1391         crit_enter();
1392         while (rep->r_flags & R_LOCKED) {
1393                 rep->r_flags |= R_WANTED;
1394                 tsleep(rep, 0, "nfstrac", 0);
1395         }
1396         KKASSERT(rep->r_flags & R_ONREQQ);
1397         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1398         rep->r_flags &= ~R_ONREQQ;
1399         --nmp->nm_reqqlen;
1400         if (TAILQ_FIRST(&nmp->nm_bioq) &&
1401             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1402                 nfssvc_iod_writer_wakeup(nmp);
1403         }
1404         crit_exit();
1405
1406         /*
1407          * Decrement the outstanding request count.
1408          */
1409         if (rep->r_flags & R_SENT) {
1410                 rep->r_flags &= ~R_SENT;
1411         }
1412         return (error);
1413 }
1414
1415 /*
1416  * Process reply with error returned from nfs_requet_waitreply().
1417  *
1418  * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1419  * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1420  */
1421 static int
1422 nfs_request_processreply(nfsm_info_t info, int error)
1423 {
1424         struct nfsreq *req = info->req;
1425         struct nfsmount *nmp = req->r_nmp;
1426         u_int32_t *tl;
1427         int verf_type;
1428         int i;
1429
1430         /*
1431          * If there was a successful reply and a tprintf msg.
1432          * tprintf a response.
1433          */
1434         if (error == 0 && (req->r_flags & R_TPRINTFMSG)) {
1435                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1436                     "is alive again");
1437         }
1438         info->mrep = req->r_mrep;
1439         info->md = req->r_md;
1440         info->dpos = req->r_dpos;
1441         if (error) {
1442                 m_freem(req->r_mreq);
1443                 req->r_mreq = NULL;
1444                 kfree(req, M_NFSREQ);
1445                 info->req = NULL;
1446                 return (error);
1447         }
1448
1449         /*
1450          * break down the rpc header and check if ok
1451          */
1452         NULLOUT(tl = nfsm_dissect(info, 3 * NFSX_UNSIGNED));
1453         if (*tl++ == rpc_msgdenied) {
1454                 if (*tl == rpc_mismatch) {
1455                         error = EOPNOTSUPP;
1456                 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1457                            *tl++ == rpc_autherr) {
1458                         if (req->r_failed_auth == 0) {
1459                                 req->r_failed_auth++;
1460                                 req->r_mheadend->m_next = NULL;
1461                                 m_freem(info->mrep);
1462                                 info->mrep = NULL;
1463                                 m_freem(req->r_mreq);
1464                                 req->r_mreq = NULL;
1465                                 return (ENEEDAUTH);
1466                         } else {
1467                                 error = EAUTH;
1468                         }
1469                 } else {
1470                         error = EACCES;
1471                 }
1472                 m_freem(info->mrep);
1473                 info->mrep = NULL;
1474                 m_freem(req->r_mreq);
1475                 req->r_mreq = NULL;
1476                 kfree(req, M_NFSREQ);
1477                 info->req = NULL;
1478                 return (error);
1479         }
1480
1481         /*
1482          * Grab any Kerberos verifier, otherwise just throw it away.
1483          */
1484         verf_type = fxdr_unsigned(int, *tl++);
1485         i = fxdr_unsigned(int32_t, *tl);
1486         if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1487                 error = nfs_savenickauth(nmp, req->r_cred, i, req->r_key,
1488                                          &info->md, &info->dpos, info->mrep);
1489                 if (error)
1490                         goto nfsmout;
1491         } else if (i > 0) {
1492                 ERROROUT(nfsm_adv(info, nfsm_rndup(i)));
1493         }
1494         NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1495         /* 0 == ok */
1496         if (*tl == 0) {
1497                 NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1498                 if (*tl != 0) {
1499                         error = fxdr_unsigned(int, *tl);
1500
1501                         /*
1502                          * Does anyone even implement this?  Just impose
1503                          * a 1-second delay.
1504                          */
1505                         if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1506                                 error == NFSERR_TRYLATER) {
1507                                 m_freem(info->mrep);
1508                                 info->mrep = NULL;
1509                                 error = 0;
1510
1511                                 tsleep((caddr_t)&lbolt, 0, "nqnfstry", 0);
1512                                 return (EAGAIN);        /* goto tryagain */
1513                         }
1514
1515                         /*
1516                          * If the File Handle was stale, invalidate the
1517                          * lookup cache, just in case.
1518                          *
1519                          * To avoid namecache<->vnode deadlocks we must
1520                          * release the vnode lock if we hold it.
1521                          */
1522                         if (error == ESTALE) {
1523                                 struct vnode *vp = req->r_vp;
1524                                 int ltype;
1525
1526                                 ltype = lockstatus(&vp->v_lock, curthread);
1527                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1528                                         lockmgr(&vp->v_lock, LK_RELEASE);
1529                                 cache_inval_vp(vp, CINV_CHILDREN);
1530                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1531                                         lockmgr(&vp->v_lock, ltype);
1532                         }
1533                         if (nmp->nm_flag & NFSMNT_NFSV3) {
1534                                 KKASSERT(*req->r_mrp == info->mrep);
1535                                 KKASSERT(*req->r_mdp == info->md);
1536                                 KKASSERT(*req->r_dposp == info->dpos);
1537                                 error |= NFSERR_RETERR;
1538                         } else {
1539                                 m_freem(info->mrep);
1540                                 info->mrep = NULL;
1541                         }
1542                         m_freem(req->r_mreq);
1543                         req->r_mreq = NULL;
1544                         kfree(req, M_NFSREQ);
1545                         info->req = NULL;
1546                         return (error);
1547                 }
1548
1549                 KKASSERT(*req->r_mrp == info->mrep);
1550                 KKASSERT(*req->r_mdp == info->md);
1551                 KKASSERT(*req->r_dposp == info->dpos);
1552                 m_freem(req->r_mreq);
1553                 req->r_mreq = NULL;
1554                 kfree(req, M_NFSREQ);
1555                 return (0);
1556         }
1557         m_freem(info->mrep);
1558         info->mrep = NULL;
1559         error = EPROTONOSUPPORT;
1560 nfsmout:
1561         m_freem(req->r_mreq);
1562         req->r_mreq = NULL;
1563         kfree(req, M_NFSREQ);
1564         info->req = NULL;
1565         return (error);
1566 }
1567
1568 #ifndef NFS_NOSERVER
1569 /*
1570  * Generate the rpc reply header
1571  * siz arg. is used to decide if adding a cluster is worthwhile
1572  */
1573 int
1574 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1575             int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1576 {
1577         u_int32_t *tl;
1578         struct nfsm_info info;
1579
1580         siz += RPC_REPLYSIZ;
1581         info.mb = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1582         info.mreq = info.mb;
1583         info.mreq->m_pkthdr.len = 0;
1584         /*
1585          * If this is not a cluster, try and leave leading space
1586          * for the lower level headers.
1587          */
1588         if ((max_hdr + siz) < MINCLSIZE)
1589                 info.mreq->m_data += max_hdr;
1590         tl = mtod(info.mreq, u_int32_t *);
1591         info.mreq->m_len = 6 * NFSX_UNSIGNED;
1592         info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1593         *tl++ = txdr_unsigned(nd->nd_retxid);
1594         *tl++ = rpc_reply;
1595         if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1596                 *tl++ = rpc_msgdenied;
1597                 if (err & NFSERR_AUTHERR) {
1598                         *tl++ = rpc_autherr;
1599                         *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1600                         info.mreq->m_len -= NFSX_UNSIGNED;
1601                         info.bpos -= NFSX_UNSIGNED;
1602                 } else {
1603                         *tl++ = rpc_mismatch;
1604                         *tl++ = txdr_unsigned(RPC_VER2);
1605                         *tl = txdr_unsigned(RPC_VER2);
1606                 }
1607         } else {
1608                 *tl++ = rpc_msgaccepted;
1609
1610                 /*
1611                  * For Kerberos authentication, we must send the nickname
1612                  * verifier back, otherwise just RPCAUTH_NULL.
1613                  */
1614                 if (nd->nd_flag & ND_KERBFULL) {
1615                     struct nfsuid *nuidp;
1616                     struct timeval ktvout;
1617
1618                     for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1619                         nuidp != NULL; nuidp = nuidp->nu_hash.le_next) {
1620                         if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1621                             (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1622                              &nuidp->nu_haddr, nd->nd_nam2)))
1623                             break;
1624                     }
1625                     if (nuidp) {
1626                         /*
1627                          * Encrypt the timestamp in ecb mode using the
1628                          * session key.
1629                          */
1630 #ifdef NFSKERB
1631                         XXX
1632 #else
1633                         ktvout.tv_sec = 0;
1634                         ktvout.tv_usec = 0;
1635 #endif
1636
1637                         *tl++ = rpc_auth_kerb;
1638                         *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1639                         *tl = ktvout.tv_sec;
1640                         tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1641                         *tl++ = ktvout.tv_usec;
1642                         *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1643                     } else {
1644                         *tl++ = 0;
1645                         *tl++ = 0;
1646                     }
1647                 } else {
1648                         *tl++ = 0;
1649                         *tl++ = 0;
1650                 }
1651                 switch (err) {
1652                 case EPROGUNAVAIL:
1653                         *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1654                         break;
1655                 case EPROGMISMATCH:
1656                         *tl = txdr_unsigned(RPC_PROGMISMATCH);
1657                         tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1658                         *tl++ = txdr_unsigned(2);
1659                         *tl = txdr_unsigned(3);
1660                         break;
1661                 case EPROCUNAVAIL:
1662                         *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1663                         break;
1664                 case EBADRPC:
1665                         *tl = txdr_unsigned(RPC_GARBAGE);
1666                         break;
1667                 default:
1668                         *tl = 0;
1669                         if (err != NFSERR_RETVOID) {
1670                                 tl = nfsm_build(&info, NFSX_UNSIGNED);
1671                                 if (err)
1672                                     *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1673                                 else
1674                                     *tl = 0;
1675                         }
1676                         break;
1677                 };
1678         }
1679
1680         if (mrq != NULL)
1681             *mrq = info.mreq;
1682         *mbp = info.mb;
1683         *bposp = info.bpos;
1684         if (err != 0 && err != NFSERR_RETVOID)
1685                 nfsstats.srvrpc_errs++;
1686         return (0);
1687 }
1688
1689
1690 #endif /* NFS_NOSERVER */
1691
1692 /*
1693  * Nfs timer routine.
1694  *
1695  * Scan the nfsreq list and retranmit any requests that have timed out
1696  * To avoid retransmission attempts on STREAM sockets (in the future) make
1697  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1698  *
1699  * Requests with attached responses, terminated requests, and
1700  * locked requests are ignored.  Locked requests will be picked up
1701  * in a later timer call.
1702  */
1703 void
1704 nfs_timer_callout(void *arg /* never used */)
1705 {
1706         struct nfsmount *nmp;
1707         struct nfsreq *req;
1708 #ifndef NFS_NOSERVER
1709         struct nfssvc_sock *slp;
1710         u_quad_t cur_usec;
1711 #endif /* NFS_NOSERVER */
1712
1713         lwkt_gettoken(&nfs_token);
1714         TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1715                 lwkt_gettoken(&nmp->nm_token);
1716                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1717                         KKASSERT(nmp == req->r_nmp);
1718                         if (req->r_mrep)
1719                                 continue;
1720                         if (req->r_flags & (R_SOFTTERM | R_LOCKED))
1721                                 continue;
1722
1723                         /*
1724                          * Handle timeout/retry.  Be sure to process r_mrep
1725                          * for async requests that completed while we had
1726                          * the request locked or they will hang in the reqq
1727                          * forever.
1728                          */
1729                         req->r_flags |= R_LOCKED;
1730                         if (nfs_sigintr(nmp, req, req->r_td)) {
1731                                 nfs_softterm(req, 1);
1732                                 req->r_flags &= ~R_LOCKED;
1733                         } else {
1734                                 nfs_timer_req(req);
1735                                 if (req->r_flags & R_ASYNC) {
1736                                         if (req->r_mrep)
1737                                                 nfs_hardterm(req, 1);
1738                                         req->r_flags &= ~R_LOCKED;
1739                                         nfssvc_iod_reader_wakeup(nmp);
1740                                 } else {
1741                                         req->r_flags &= ~R_LOCKED;
1742                                 }
1743                         }
1744                         if (req->r_flags & R_WANTED) {
1745                                 req->r_flags &= ~R_WANTED;
1746                                 wakeup(req);
1747                         }
1748                 }
1749                 lwkt_reltoken(&nmp->nm_token);
1750         }
1751 #ifndef NFS_NOSERVER
1752
1753         /*
1754          * Scan the write gathering queues for writes that need to be
1755          * completed now.
1756          */
1757         cur_usec = nfs_curusec();
1758
1759         TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1760                 /* XXX race against removal */
1761                 if (lwkt_trytoken(&slp->ns_token)) {
1762                         if (slp->ns_tq.lh_first &&
1763                             (slp->ns_tq.lh_first->nd_time <= cur_usec)) {
1764                                 nfsrv_wakenfsd(slp, 1);
1765                         }
1766                         lwkt_reltoken(&slp->ns_token);
1767                 }
1768         }
1769 #endif /* NFS_NOSERVER */
1770
1771         callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer_callout, NULL);
1772         lwkt_reltoken(&nfs_token);
1773 }
1774
1775 static
1776 void
1777 nfs_timer_req(struct nfsreq *req)
1778 {
1779         struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1780         struct nfsmount *nmp = req->r_nmp;
1781         struct mbuf *m;
1782         struct socket *so;
1783         int timeo;
1784         int error;
1785
1786         /*
1787          * rtt ticks and timeout calculation.  Return if the timeout
1788          * has not been reached yet, unless the packet is flagged
1789          * for an immediate send.
1790          *
1791          * The mean rtt doesn't help when we get random I/Os, we have
1792          * to multiply by fairly large numbers.
1793          */
1794         if (req->r_rtt >= 0) {
1795                 /*
1796                  * Calculate the timeout to test against.
1797                  */
1798                 req->r_rtt++;
1799                 if (nmp->nm_flag & NFSMNT_DUMBTIMR) {
1800                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1801                 } else if (req->r_flags & R_TIMING) {
1802                         timeo = NFS_SRTT(req) + NFS_SDRTT(req);
1803                 } else {
1804                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1805                 }
1806                 timeo *= multt[req->r_procnum];
1807                 /* timeo is still scaled by SCALE_BITS */
1808
1809 #define NFSFS   (NFS_RTT_SCALE * NFS_HZ)
1810                 if (req->r_flags & R_TIMING) {
1811                         static long last_time;
1812                         if (nfs_showrtt && last_time != time_second) {
1813                                 kprintf("rpccmd %d NFS SRTT %d SDRTT %d "
1814                                         "timeo %d.%03d\n",
1815                                         proct[req->r_procnum],
1816                                         NFS_SRTT(req), NFS_SDRTT(req),
1817                                         timeo / NFSFS,
1818                                         timeo % NFSFS * 1000 /  NFSFS);
1819                                 last_time = time_second;
1820                         }
1821                 }
1822 #undef NFSFS
1823
1824                 /*
1825                  * deal with nfs_timer jitter.
1826                  */
1827                 timeo = (timeo >> NFS_RTT_SCALE_BITS) + 1;
1828                 if (timeo < 2)
1829                         timeo = 2;
1830
1831                 if (nmp->nm_timeouts > 0)
1832                         timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1833                 if (timeo > NFS_MAXTIMEO)
1834                         timeo = NFS_MAXTIMEO;
1835                 if (req->r_rtt <= timeo) {
1836                         if ((req->r_flags & R_NEEDSXMIT) == 0)
1837                                 return;
1838                 } else if (nmp->nm_timeouts < 8) {
1839                         nmp->nm_timeouts++;
1840                 }
1841         }
1842
1843         /*
1844          * Check for server not responding
1845          */
1846         if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1847              req->r_rexmit > nmp->nm_deadthresh) {
1848                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1849                         "not responding");
1850                 req->r_flags |= R_TPRINTFMSG;
1851         }
1852         if (req->r_rexmit >= req->r_retry) {    /* too many */
1853                 nfsstats.rpctimeouts++;
1854                 nfs_softterm(req, 1);
1855                 return;
1856         }
1857
1858         /*
1859          * Generally disable retransmission on reliable sockets,
1860          * unless the request is flagged for immediate send.
1861          */
1862         if (nmp->nm_sotype != SOCK_DGRAM) {
1863                 if (++req->r_rexmit > NFS_MAXREXMIT)
1864                         req->r_rexmit = NFS_MAXREXMIT;
1865                 if ((req->r_flags & R_NEEDSXMIT) == 0)
1866                         return;
1867         }
1868
1869         /*
1870          * Stop here if we do not have a socket!
1871          */
1872         if ((so = nmp->nm_so) == NULL)
1873                 return;
1874
1875         /*
1876          * If there is enough space and the window allows.. resend it.
1877          *
1878          * r_rtt is left intact in case we get an answer after the
1879          * retry that was a reply to the original packet.
1880          *
1881          * NOTE: so_pru_send()
1882          */
1883         if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1884             (req->r_flags & (R_SENT | R_NEEDSXMIT)) &&
1885            (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1886                 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1887                     error = so_pru_send(so, 0, m, NULL, NULL, td);
1888                 else
1889                     error = so_pru_send(so, 0, m, nmp->nm_nam, NULL, td);
1890                 if (error) {
1891                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1892                                 so->so_error = 0;
1893                         req->r_flags |= R_NEEDSXMIT;
1894                 } else if (req->r_mrep == NULL) {
1895                         /*
1896                          * Iff first send, start timing
1897                          * else turn timing off, backoff timer
1898                          * and divide congestion window by 2.
1899                          *
1900                          * It is possible for the so_pru_send() to
1901                          * block and for us to race a reply so we
1902                          * only do this if the reply field has not
1903                          * been filled in.  R_LOCKED will prevent
1904                          * the request from being ripped out from under
1905                          * us entirely.
1906                          *
1907                          * Record the last resent procnum to aid us
1908                          * in duplicate detection on receive.
1909                          */
1910                         if ((req->r_flags & R_NEEDSXMIT) == 0) {
1911                                 if (nfs_showrexmit)
1912                                         kprintf("X");
1913                                 if (++req->r_rexmit > NFS_MAXREXMIT)
1914                                         req->r_rexmit = NFS_MAXREXMIT;
1915                                 nmp->nm_maxasync_scaled >>= 1;
1916                                 if (nmp->nm_maxasync_scaled < NFS_MINASYNC_SCALED)
1917                                         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
1918                                 nfsstats.rpcretries++;
1919                                 nmp->nm_lastreprocnum = req->r_procnum;
1920                         } else {
1921                                 req->r_flags |= R_SENT;
1922                                 req->r_flags &= ~R_NEEDSXMIT;
1923                         }
1924                 }
1925         }
1926 }
1927
1928 /*
1929  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1930  * wait for all requests to complete. This is used by forced unmounts
1931  * to terminate any outstanding RPCs.
1932  *
1933  * Locked requests cannot be canceled but will be marked for
1934  * soft-termination.
1935  */
1936 int
1937 nfs_nmcancelreqs(struct nfsmount *nmp)
1938 {
1939         struct nfsreq *req;
1940         int i;
1941
1942         crit_enter();
1943         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1944                 if (req->r_mrep != NULL || (req->r_flags & R_SOFTTERM))
1945                         continue;
1946                 nfs_softterm(req, 0);
1947         }
1948         /* XXX  the other two queues as well */
1949         crit_exit();
1950
1951         for (i = 0; i < 30; i++) {
1952                 crit_enter();
1953                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1954                         if (nmp == req->r_nmp)
1955                                 break;
1956                 }
1957                 crit_exit();
1958                 if (req == NULL)
1959                         return (0);
1960                 tsleep(&lbolt, 0, "nfscancel", 0);
1961         }
1962         return (EBUSY);
1963 }
1964
1965 /*
1966  * Soft-terminate a request, effectively marking it as failed.
1967  *
1968  * Must be called from within a critical section.
1969  */
1970 static void
1971 nfs_softterm(struct nfsreq *rep, int islocked)
1972 {
1973         rep->r_flags |= R_SOFTTERM;
1974         nfs_hardterm(rep, islocked);
1975 }
1976
1977 /*
1978  * Hard-terminate a request, typically after getting a response.
1979  *
1980  * The state machine can still decide to re-issue it later if necessary.
1981  *
1982  * Must be called from within a critical section.
1983  */
1984 static void
1985 nfs_hardterm(struct nfsreq *rep, int islocked)
1986 {
1987         struct nfsmount *nmp = rep->r_nmp;
1988
1989         /*
1990          * The nm_send count is decremented now to avoid deadlocks
1991          * when the process in soreceive() hasn't yet managed to send
1992          * its own request.
1993          */
1994         if (rep->r_flags & R_SENT) {
1995                 rep->r_flags &= ~R_SENT;
1996         }
1997
1998         /*
1999          * If we locked the request or nobody else has locked the request,
2000          * and the request is async, we can move it to the reader thread's
2001          * queue now and fix up the state.
2002          *
2003          * If we locked the request or nobody else has locked the request,
2004          * we can wake up anyone blocked waiting for a response on the
2005          * request.
2006          */
2007         if (islocked || (rep->r_flags & R_LOCKED) == 0) {
2008                 if ((rep->r_flags & (R_ONREQQ | R_ASYNC)) ==
2009                     (R_ONREQQ | R_ASYNC)) {
2010                         rep->r_flags &= ~R_ONREQQ;
2011                         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
2012                         --nmp->nm_reqqlen;
2013                         TAILQ_INSERT_TAIL(&nmp->nm_reqrxq, rep, r_chain);
2014                         KKASSERT(rep->r_info->state == NFSM_STATE_TRY ||
2015                                  rep->r_info->state == NFSM_STATE_WAITREPLY);
2016                         rep->r_info->state = NFSM_STATE_PROCESSREPLY;
2017                         nfssvc_iod_reader_wakeup(nmp);
2018                         if (TAILQ_FIRST(&nmp->nm_bioq) &&
2019                             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
2020                                 nfssvc_iod_writer_wakeup(nmp);
2021                         }
2022                 }
2023                 mtx_abort_ex_link(&nmp->nm_rxlock, &rep->r_link);
2024         }
2025 }
2026
2027 /*
2028  * Test for a termination condition pending on the process.
2029  * This is used for NFSMNT_INT mounts.
2030  */
2031 int
2032 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
2033 {
2034         sigset_t tmpset;
2035         struct proc *p;
2036         struct lwp *lp;
2037
2038         if (rep && (rep->r_flags & R_SOFTTERM))
2039                 return (EINTR);
2040         /* Terminate all requests while attempting a forced unmount. */
2041         if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
2042                 return (EINTR);
2043         if (!(nmp->nm_flag & NFSMNT_INT))
2044                 return (0);
2045         /* td might be NULL YYY */
2046         if (td == NULL || (p = td->td_proc) == NULL)
2047                 return (0);
2048
2049         lp = td->td_lwp;
2050         tmpset = lwp_sigpend(lp);
2051         SIGSETNAND(tmpset, lp->lwp_sigmask);
2052         SIGSETNAND(tmpset, p->p_sigignore);
2053         if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
2054                 return (EINTR);
2055
2056         return (0);
2057 }
2058
2059 /*
2060  * Lock a socket against others.
2061  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
2062  * and also to avoid race conditions between the processes with nfs requests
2063  * in progress when a reconnect is necessary.
2064  */
2065 int
2066 nfs_sndlock(struct nfsmount *nmp, struct nfsreq *rep)
2067 {
2068         mtx_t mtx = &nmp->nm_txlock;
2069         struct thread *td;
2070         int slptimeo;
2071         int slpflag;
2072         int error;
2073
2074         slpflag = 0;
2075         slptimeo = 0;
2076         td = rep ? rep->r_td : NULL;
2077         if (nmp->nm_flag & NFSMNT_INT)
2078                 slpflag = PCATCH;
2079
2080         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2081                 if (nfs_sigintr(nmp, rep, td)) {
2082                         error = EINTR;
2083                         break;
2084                 }
2085                 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
2086                 if (error == 0)
2087                         break;
2088                 if (slpflag == PCATCH) {
2089                         slpflag = 0;
2090                         slptimeo = 2 * hz;
2091                 }
2092         }
2093         /* Always fail if our request has been cancelled. */
2094         if (rep && (rep->r_flags & R_SOFTTERM)) {
2095                 if (error == 0)
2096                         mtx_unlock(mtx);
2097                 error = EINTR;
2098         }
2099         return (error);
2100 }
2101
2102 /*
2103  * Unlock the stream socket for others.
2104  */
2105 void
2106 nfs_sndunlock(struct nfsmount *nmp)
2107 {
2108         mtx_unlock(&nmp->nm_txlock);
2109 }
2110
2111 /*
2112  * Lock the receiver side of the socket.
2113  *
2114  * rep may be NULL.
2115  */
2116 static int
2117 nfs_rcvlock(struct nfsmount *nmp, struct nfsreq *rep)
2118 {
2119         mtx_t mtx = &nmp->nm_rxlock;
2120         int slpflag;
2121         int slptimeo;
2122         int error;
2123
2124         /*
2125          * Unconditionally check for completion in case another nfsiod
2126          * get the packet while the caller was blocked, before the caller
2127          * called us.  Packet reception is handled by mainline code which
2128          * is protected by the BGL at the moment.
2129          *
2130          * We do not strictly need the second check just before the
2131          * tsleep(), but it's good defensive programming.
2132          */
2133         if (rep && rep->r_mrep != NULL)
2134                 return (EALREADY);
2135
2136         if (nmp->nm_flag & NFSMNT_INT)
2137                 slpflag = PCATCH;
2138         else
2139                 slpflag = 0;
2140         slptimeo = 0;
2141
2142         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2143                 if (nfs_sigintr(nmp, rep, (rep ? rep->r_td : NULL))) {
2144                         error = EINTR;
2145                         break;
2146                 }
2147                 if (rep && rep->r_mrep != NULL) {
2148                         error = EALREADY;
2149                         break;
2150                 }
2151
2152                 /*
2153                  * NOTE: can return ENOLCK, but in that case rep->r_mrep
2154                  *       will already be set.
2155                  */
2156                 if (rep) {
2157                         error = mtx_lock_ex_link(mtx, &rep->r_link,
2158                                                  "nfsrcvlk",
2159                                                  slpflag, slptimeo);
2160                 } else {
2161                         error = mtx_lock_ex(mtx, "nfsrcvlk", slpflag, slptimeo);
2162                 }
2163                 if (error == 0)
2164                         break;
2165
2166                 /*
2167                  * If our reply was recieved while we were sleeping,
2168                  * then just return without taking the lock to avoid a
2169                  * situation where a single iod could 'capture' the
2170                  * recieve lock.
2171                  */
2172                 if (rep && rep->r_mrep != NULL) {
2173                         error = EALREADY;
2174                         break;
2175                 }
2176                 if (slpflag == PCATCH) {
2177                         slpflag = 0;
2178                         slptimeo = 2 * hz;
2179                 }
2180         }
2181         if (error == 0) {
2182                 if (rep && rep->r_mrep != NULL) {
2183                         error = EALREADY;
2184                         mtx_unlock(mtx);
2185                 }
2186         }
2187         return (error);
2188 }
2189
2190 /*
2191  * Unlock the stream socket for others.
2192  */
2193 static void
2194 nfs_rcvunlock(struct nfsmount *nmp)
2195 {
2196         mtx_unlock(&nmp->nm_rxlock);
2197 }
2198
2199 /*
2200  * nfs_realign:
2201  *
2202  * Check for badly aligned mbuf data and realign by copying the unaligned
2203  * portion of the data into a new mbuf chain and freeing the portions
2204  * of the old chain that were replaced.
2205  *
2206  * We cannot simply realign the data within the existing mbuf chain
2207  * because the underlying buffers may contain other rpc commands and
2208  * we cannot afford to overwrite them.
2209  *
2210  * We would prefer to avoid this situation entirely.  The situation does
2211  * not occur with NFS/UDP and is supposed to only occassionally occur
2212  * with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
2213  *
2214  * NOTE!  MB_DONTWAIT cannot be used here.  The mbufs must be acquired
2215  *        because the rpc request OR reply cannot be thrown away.  TCP NFS
2216  *        mounts do not retry their RPCs unless the TCP connection itself
2217  *        is dropped so throwing away a RPC will basically cause the NFS
2218  *        operation to lockup indefinitely.
2219  */
2220 static void
2221 nfs_realign(struct mbuf **pm, int hsiz)
2222 {
2223         struct mbuf *m;
2224         struct mbuf *n = NULL;
2225
2226         /*
2227          * Check for misalignemnt
2228          */
2229         ++nfs_realign_test;
2230         while ((m = *pm) != NULL) {
2231                 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3))
2232                         break;
2233                 pm = &m->m_next;
2234         }
2235
2236         /*
2237          * If misalignment found make a completely new copy.
2238          */
2239         if (m) {
2240                 ++nfs_realign_count;
2241                 n = m_dup_data(m, MB_WAIT);
2242                 m_freem(*pm);
2243                 *pm = n;
2244         }
2245 }
2246
2247 #ifndef NFS_NOSERVER
2248
2249 /*
2250  * Parse an RPC request
2251  * - verify it
2252  * - fill in the cred struct.
2253  */
2254 int
2255 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
2256 {
2257         int len, i;
2258         u_int32_t *tl;
2259         struct uio uio;
2260         struct iovec iov;
2261         caddr_t cp;
2262         u_int32_t nfsvers, auth_type;
2263         uid_t nickuid;
2264         int error = 0, ticklen;
2265         struct nfsuid *nuidp;
2266         struct timeval tvin, tvout;
2267         struct nfsm_info info;
2268 #if 0                           /* until encrypted keys are implemented */
2269         NFSKERBKEYSCHED_T keys; /* stores key schedule */
2270 #endif
2271
2272         info.mrep = nd->nd_mrep;
2273         info.md = nd->nd_md;
2274         info.dpos = nd->nd_dpos;
2275
2276         if (has_header) {
2277                 NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
2278                 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
2279                 if (*tl++ != rpc_call) {
2280                         m_freem(info.mrep);
2281                         return (EBADRPC);
2282                 }
2283         } else {
2284                 NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
2285         }
2286         nd->nd_repstat = 0;
2287         nd->nd_flag = 0;
2288         if (*tl++ != rpc_vers) {
2289                 nd->nd_repstat = ERPCMISMATCH;
2290                 nd->nd_procnum = NFSPROC_NOOP;
2291                 return (0);
2292         }
2293         if (*tl != nfs_prog) {
2294                 nd->nd_repstat = EPROGUNAVAIL;
2295                 nd->nd_procnum = NFSPROC_NOOP;
2296                 return (0);
2297         }
2298         tl++;
2299         nfsvers = fxdr_unsigned(u_int32_t, *tl++);
2300         if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
2301                 nd->nd_repstat = EPROGMISMATCH;
2302                 nd->nd_procnum = NFSPROC_NOOP;
2303                 return (0);
2304         }
2305         if (nfsvers == NFS_VER3)
2306                 nd->nd_flag = ND_NFSV3;
2307         nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
2308         if (nd->nd_procnum == NFSPROC_NULL)
2309                 return (0);
2310         if (nd->nd_procnum >= NFS_NPROCS ||
2311                 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
2312                 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
2313                 nd->nd_repstat = EPROCUNAVAIL;
2314                 nd->nd_procnum = NFSPROC_NOOP;
2315                 return (0);
2316         }
2317         if ((nd->nd_flag & ND_NFSV3) == 0)
2318                 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
2319         auth_type = *tl++;
2320         len = fxdr_unsigned(int, *tl++);
2321         if (len < 0 || len > RPCAUTH_MAXSIZ) {
2322                 m_freem(info.mrep);
2323                 return (EBADRPC);
2324         }
2325
2326         nd->nd_flag &= ~ND_KERBAUTH;
2327         /*
2328          * Handle auth_unix or auth_kerb.
2329          */
2330         if (auth_type == rpc_auth_unix) {
2331                 len = fxdr_unsigned(int, *++tl);
2332                 if (len < 0 || len > NFS_MAXNAMLEN) {
2333                         m_freem(info.mrep);
2334                         return (EBADRPC);
2335                 }
2336                 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2337                 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2338                 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
2339                 nd->nd_cr.cr_ref = 1;
2340                 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
2341                 nd->nd_cr.cr_ruid = nd->nd_cr.cr_svuid = nd->nd_cr.cr_uid;
2342                 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
2343                 nd->nd_cr.cr_rgid = nd->nd_cr.cr_svgid = nd->nd_cr.cr_gid;
2344                 len = fxdr_unsigned(int, *tl);
2345                 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
2346                         m_freem(info.mrep);
2347                         return (EBADRPC);
2348                 }
2349                 NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
2350                 for (i = 1; i <= len; i++)
2351                     if (i < NGROUPS)
2352                         nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
2353                     else
2354                         tl++;
2355                 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
2356                 if (nd->nd_cr.cr_ngroups > 1)
2357                     nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
2358                 len = fxdr_unsigned(int, *++tl);
2359                 if (len < 0 || len > RPCAUTH_MAXSIZ) {
2360                         m_freem(info.mrep);
2361                         return (EBADRPC);
2362                 }
2363                 if (len > 0) {
2364                         ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2365                 }
2366         } else if (auth_type == rpc_auth_kerb) {
2367                 switch (fxdr_unsigned(int, *tl++)) {
2368                 case RPCAKN_FULLNAME:
2369                         ticklen = fxdr_unsigned(int, *tl);
2370                         *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2371                         uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2372                         nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2373                         if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2374                                 m_freem(info.mrep);
2375                                 return (EBADRPC);
2376                         }
2377                         uio.uio_offset = 0;
2378                         uio.uio_iov = &iov;
2379                         uio.uio_iovcnt = 1;
2380                         uio.uio_segflg = UIO_SYSSPACE;
2381                         iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2382                         iov.iov_len = RPCAUTH_MAXSIZ - 4;
2383                         ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2384                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2385                         if (*tl++ != rpc_auth_kerb ||
2386                                 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2387                                 kprintf("Bad kerb verifier\n");
2388                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2389                                 nd->nd_procnum = NFSPROC_NOOP;
2390                                 return (0);
2391                         }
2392                         NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2393                         tl = (u_int32_t *)cp;
2394                         if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2395                                 kprintf("Not fullname kerb verifier\n");
2396                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2397                                 nd->nd_procnum = NFSPROC_NOOP;
2398                                 return (0);
2399                         }
2400                         cp += NFSX_UNSIGNED;
2401                         bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2402                         nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2403                         nd->nd_flag |= ND_KERBFULL;
2404                         nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2405                         break;
2406                 case RPCAKN_NICKNAME:
2407                         if (len != 2 * NFSX_UNSIGNED) {
2408                                 kprintf("Kerb nickname short\n");
2409                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2410                                 nd->nd_procnum = NFSPROC_NOOP;
2411                                 return (0);
2412                         }
2413                         nickuid = fxdr_unsigned(uid_t, *tl);
2414                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2415                         if (*tl++ != rpc_auth_kerb ||
2416                                 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2417                                 kprintf("Kerb nick verifier bad\n");
2418                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2419                                 nd->nd_procnum = NFSPROC_NOOP;
2420                                 return (0);
2421                         }
2422                         NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2423                         tvin.tv_sec = *tl++;
2424                         tvin.tv_usec = *tl;
2425
2426                         for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2427                             nuidp != NULL; nuidp = nuidp->nu_hash.le_next) {
2428                                 if (nuidp->nu_cr.cr_uid == nickuid &&
2429                                     (!nd->nd_nam2 ||
2430                                      netaddr_match(NU_NETFAM(nuidp),
2431                                       &nuidp->nu_haddr, nd->nd_nam2)))
2432                                         break;
2433                         }
2434                         if (!nuidp) {
2435                                 nd->nd_repstat =
2436                                         (NFSERR_AUTHERR|AUTH_REJECTCRED);
2437                                 nd->nd_procnum = NFSPROC_NOOP;
2438                                 return (0);
2439                         }
2440
2441                         /*
2442                          * Now, decrypt the timestamp using the session key
2443                          * and validate it.
2444                          */
2445 #ifdef NFSKERB
2446                         XXX
2447 #else
2448                         tvout.tv_sec = 0;
2449                         tvout.tv_usec = 0;
2450 #endif
2451
2452                         tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2453                         tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2454                         if (nuidp->nu_expire < time_second ||
2455                             nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2456                             (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2457                              nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2458                                 nuidp->nu_expire = 0;
2459                                 nd->nd_repstat =
2460                                     (NFSERR_AUTHERR|AUTH_REJECTVERF);
2461                                 nd->nd_procnum = NFSPROC_NOOP;
2462                                 return (0);
2463                         }
2464                         nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2465                         nd->nd_flag |= ND_KERBNICK;
2466                         break;
2467                 }
2468         } else {
2469                 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2470                 nd->nd_procnum = NFSPROC_NOOP;
2471                 return (0);
2472         }
2473
2474         nd->nd_md = info.md;
2475         nd->nd_dpos = info.dpos;
2476         return (0);
2477 nfsmout:
2478         return (error);
2479 }
2480
2481 #endif
2482
2483 /*
2484  * Send a message to the originating process's terminal.  The thread and/or
2485  * process may be NULL.  YYY the thread should not be NULL but there may
2486  * still be some uio_td's that are still being passed as NULL through to
2487  * nfsm_request().
2488  */
2489 static int
2490 nfs_msg(struct thread *td, char *server, char *msg)
2491 {
2492         tpr_t tpr;
2493
2494         if (td && td->td_proc)
2495                 tpr = tprintf_open(td->td_proc);
2496         else
2497                 tpr = NULL;
2498         tprintf(tpr, "nfs server %s: %s\n", server, msg);
2499         tprintf_close(tpr);
2500         return (0);
2501 }
2502
2503 #ifndef NFS_NOSERVER
2504
2505 /*
2506  * Socket upcall routine for nfsd sockets.  This runs in the protocol
2507  * thread and passes waitflag == MB_DONTWAIT.
2508  */
2509 void
2510 nfsrv_rcv_upcall(struct socket *so, void *arg, int waitflag)
2511 {
2512         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2513
2514         if (slp->ns_needq_upcall == 0) {
2515                 slp->ns_needq_upcall = 1;       /* ok to race */
2516                 lwkt_gettoken(&nfs_token);
2517                 nfsrv_wakenfsd(slp, 1);
2518                 lwkt_reltoken(&nfs_token);
2519         }
2520 #if 0
2521         lwkt_gettoken(&slp->ns_token);
2522         slp->ns_flag |= SLP_NEEDQ;
2523         nfsrv_rcv(so, arg, waitflag);
2524         lwkt_reltoken(&slp->ns_token);
2525 #endif
2526 }
2527
2528 /*
2529  * Process new data on a receive socket.  Essentially do as much as we can
2530  * non-blocking, else punt and it will be called with MB_WAIT from an nfsd.
2531  *
2532  * slp->ns_token is held on call
2533  */
2534 void
2535 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2536 {
2537         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2538         struct mbuf *m;
2539         struct sockaddr *nam;
2540         struct sockbuf sio;
2541         int flags, error;
2542         int nparallel_wakeup = 0;
2543
2544         ASSERT_LWKT_TOKEN_HELD(&slp->ns_token);
2545
2546         if ((slp->ns_flag & SLP_VALID) == 0)
2547                 return;
2548
2549         /*
2550          * Do not allow an infinite number of completed RPC records to build
2551          * up before we stop reading data from the socket.  Otherwise we could
2552          * end up holding onto an unreasonable number of mbufs for requests
2553          * waiting for service.
2554          *
2555          * This should give pretty good feedback to the TCP layer and
2556          * prevents a memory crunch for other protocols.
2557          *
2558          * Note that the same service socket can be dispatched to several
2559          * nfs servers simultaniously.  The tcp protocol callback calls us
2560          * with MB_DONTWAIT.  nfsd calls us with MB_WAIT (typically).
2561          */
2562         if (NFSRV_RECLIMIT(slp))
2563                 return;
2564
2565         /*
2566          * Handle protocol specifics to parse an RPC request.  We always
2567          * pull from the socket using non-blocking I/O.
2568          */
2569         if (so->so_type == SOCK_STREAM) {
2570                 /*
2571                  * The data has to be read in an orderly fashion from a TCP
2572                  * stream, unlike a UDP socket.  It is possible for soreceive
2573                  * and/or nfsrv_getstream() to block, so make sure only one
2574                  * entity is messing around with the TCP stream at any given
2575                  * moment.  The receive sockbuf's lock in soreceive is not
2576                  * sufficient.
2577                  */
2578                 if (slp->ns_flag & SLP_GETSTREAM)
2579                         return;
2580                 slp->ns_flag |= SLP_GETSTREAM;
2581
2582                 /*
2583                  * Do soreceive().  Pull out as much data as possible without
2584                  * blocking.
2585                  */
2586                 sbinit(&sio, 1000000000);
2587                 flags = MSG_DONTWAIT;
2588                 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2589                 if (error || sio.sb_mb == NULL) {
2590                         if (error != EWOULDBLOCK)
2591                                 slp->ns_flag |= SLP_DISCONN;
2592                         slp->ns_flag &= ~(SLP_GETSTREAM | SLP_NEEDQ);
2593                         goto done;
2594                 }
2595                 m = sio.sb_mb;
2596                 if (slp->ns_rawend) {
2597                         slp->ns_rawend->m_next = m;
2598                         slp->ns_cc += sio.sb_cc;
2599                 } else {
2600                         slp->ns_raw = m;
2601                         slp->ns_cc = sio.sb_cc;
2602                 }
2603                 while (m->m_next)
2604                         m = m->m_next;
2605                 slp->ns_rawend = m;
2606
2607                 /*
2608                  * Now try and parse as many record(s) as we can out of the
2609                  * raw stream data.  This will set SLP_DOREC.
2610                  */
2611                 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2612                 if (error && error != EWOULDBLOCK)
2613                         slp->ns_flag |= SLP_DISCONN;
2614                 slp->ns_flag &= ~SLP_GETSTREAM;
2615         } else {
2616                 /*
2617                  * For UDP soreceive typically pulls just one packet, loop
2618                  * to get the whole batch.
2619                  */
2620                 do {
2621                         sbinit(&sio, 1000000000);
2622                         flags = MSG_DONTWAIT;
2623                         error = so_pru_soreceive(so, &nam, NULL, &sio,
2624                                                  NULL, &flags);
2625                         if (sio.sb_mb) {
2626                                 struct nfsrv_rec *rec;
2627                                 int mf = (waitflag & MB_DONTWAIT) ?
2628                                             M_NOWAIT : M_WAITOK;
2629                                 rec = kmalloc(sizeof(struct nfsrv_rec),
2630                                              M_NFSRVDESC, mf);
2631                                 if (!rec) {
2632                                         if (nam)
2633                                                 kfree(nam, M_SONAME);
2634                                         m_freem(sio.sb_mb);
2635                                         continue;
2636                                 }
2637                                 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2638                                 rec->nr_address = nam;
2639                                 rec->nr_packet = sio.sb_mb;
2640                                 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2641                                 ++slp->ns_numrec;
2642                                 slp->ns_flag |= SLP_DOREC;
2643                                 ++nparallel_wakeup;
2644                         } else {
2645                                 slp->ns_flag &= ~SLP_NEEDQ;
2646                         }
2647                         if (error) {
2648                                 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2649                                     && error != EWOULDBLOCK) {
2650                                         slp->ns_flag |= SLP_DISCONN;
2651                                         break;
2652                                 }
2653                         }
2654                         if (NFSRV_RECLIMIT(slp))
2655                                 break;
2656                 } while (sio.sb_mb);
2657         }
2658
2659         /*
2660          * If we were upcalled from the tcp protocol layer and we have
2661          * fully parsed records ready to go, or there is new data pending,
2662          * or something went wrong, try to wake up a nfsd thread to deal
2663          * with it.
2664          */
2665 done:
2666         /* XXX this code is currently not executed (nfsrv_rcv_upcall) */
2667         if (waitflag == MB_DONTWAIT && (slp->ns_flag & SLP_ACTION_MASK)) {
2668                 lwkt_gettoken(&nfs_token);
2669                 nfsrv_wakenfsd(slp, nparallel_wakeup);
2670                 lwkt_reltoken(&nfs_token);
2671         }
2672 }
2673
2674 /*
2675  * Try and extract an RPC request from the mbuf data list received on a
2676  * stream socket. The "waitflag" argument indicates whether or not it
2677  * can sleep.
2678  */
2679 static int
2680 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2681 {
2682         struct mbuf *m, **mpp;
2683         char *cp1, *cp2;
2684         int len;
2685         struct mbuf *om, *m2, *recm;
2686         u_int32_t recmark;
2687
2688         for (;;) {
2689             if (slp->ns_reclen == 0) {
2690                 if (slp->ns_cc < NFSX_UNSIGNED)
2691                         return (0);
2692                 m = slp->ns_raw;
2693                 if (m->m_len >= NFSX_UNSIGNED) {
2694                         bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2695                         m->m_data += NFSX_UNSIGNED;
2696                         m->m_len -= NFSX_UNSIGNED;
2697                 } else {
2698                         cp1 = (caddr_t)&recmark;
2699                         cp2 = mtod(m, caddr_t);
2700                         while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2701                                 while (m->m_len == 0) {
2702                                         m = m->m_next;
2703                                         cp2 = mtod(m, caddr_t);
2704                                 }
2705                                 *cp1++ = *cp2++;
2706                                 m->m_data++;
2707                                 m->m_len--;
2708                         }
2709                 }
2710                 slp->ns_cc -= NFSX_UNSIGNED;
2711                 recmark = ntohl(recmark);
2712                 slp->ns_reclen = recmark & ~0x80000000;
2713                 if (recmark & 0x80000000)
2714                         slp->ns_flag |= SLP_LASTFRAG;
2715                 else
2716                         slp->ns_flag &= ~SLP_LASTFRAG;
2717                 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2718                         log(LOG_ERR, "%s (%d) from nfs client\n",
2719                             "impossible packet length",
2720                             slp->ns_reclen);
2721                         return (EPERM);
2722                 }
2723             }
2724
2725             /*
2726              * Now get the record part.
2727              *
2728              * Note that slp->ns_reclen may be 0.  Linux sometimes
2729              * generates 0-length RPCs
2730              */
2731             recm = NULL;
2732             if (slp->ns_cc == slp->ns_reclen) {
2733                 recm = slp->ns_raw;
2734                 slp->ns_raw = slp->ns_rawend = NULL;
2735                 slp->ns_cc = slp->ns_reclen = 0;
2736             } else if (slp->ns_cc > slp->ns_reclen) {
2737                 len = 0;
2738                 m = slp->ns_raw;
2739                 om = NULL;
2740
2741                 while (len < slp->ns_reclen) {
2742                         if ((len + m->m_len) > slp->ns_reclen) {
2743                                 m2 = m_copym(m, 0, slp->ns_reclen - len,
2744                                         waitflag);
2745                                 if (m2) {
2746                                         if (om) {
2747                                                 om->m_next = m2;
2748                                                 recm = slp->ns_raw;
2749                                         } else
2750                                                 recm = m2;
2751                                         m->m_data += slp->ns_reclen - len;
2752                                         m->m_len -= slp->ns_reclen - len;
2753                                         len = slp->ns_reclen;
2754                                 } else {
2755                                         return (EWOULDBLOCK);
2756                                 }
2757                         } else if ((len + m->m_len) == slp->ns_reclen) {
2758                                 om = m;
2759                                 len += m->m_len;
2760                                 m = m->m_next;
2761                                 recm = slp->ns_raw;
2762                                 om->m_next = NULL;
2763                         } else {
2764                                 om = m;
2765                                 len += m->m_len;
2766                                 m = m->m_next;
2767                         }
2768                 }
2769                 slp->ns_raw = m;
2770                 slp->ns_cc -= len;
2771                 slp->ns_reclen = 0;
2772             } else {
2773                 return (0);
2774             }
2775
2776             /*
2777              * Accumulate the fragments into a record.
2778              */
2779             mpp = &slp->ns_frag;
2780             while (*mpp)
2781                 mpp = &((*mpp)->m_next);
2782             *mpp = recm;
2783             if (slp->ns_flag & SLP_LASTFRAG) {
2784                 struct nfsrv_rec *rec;
2785                 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2786                 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2787                 if (!rec) {
2788                     m_freem(slp->ns_frag);
2789                 } else {
2790                     nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2791                     rec->nr_address = NULL;
2792                     rec->nr_packet = slp->ns_frag;
2793                     STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2794                     ++slp->ns_numrec;
2795                     slp->ns_flag |= SLP_DOREC;
2796                     ++*countp;
2797                 }
2798                 slp->ns_frag = NULL;
2799             }
2800         }
2801 }
2802
2803 #ifdef INVARIANTS
2804
2805 /*
2806  * Sanity check our mbuf chain.
2807  */
2808 static void
2809 nfs_checkpkt(struct mbuf *m, int len)
2810 {
2811         int xlen = 0;
2812         while (m) {
2813                 xlen += m->m_len;
2814                 m = m->m_next;
2815         }
2816         if (xlen != len) {
2817                 panic("nfs_checkpkt: len mismatch %d/%d mbuf %p",
2818                         xlen, len, m);
2819         }
2820 }
2821
2822 #else
2823
2824 static void
2825 nfs_checkpkt(struct mbuf *m __unused, int len __unused)
2826 {
2827 }
2828
2829 #endif
2830
2831 /*
2832  * Parse an RPC header.
2833  *
2834  * If the socket is invalid or no records are pending we return ENOBUFS.
2835  * The caller must deal with NEEDQ races.
2836  */
2837 int
2838 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2839             struct nfsrv_descript **ndp)
2840 {
2841         struct nfsrv_rec *rec;
2842         struct mbuf *m;
2843         struct sockaddr *nam;
2844         struct nfsrv_descript *nd;
2845         int error;
2846
2847         *ndp = NULL;
2848         if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2849                 return (ENOBUFS);
2850         rec = STAILQ_FIRST(&slp->ns_rec);
2851         STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2852         KKASSERT(slp->ns_numrec > 0);
2853         if (--slp->ns_numrec == 0)
2854                 slp->ns_flag &= ~SLP_DOREC;
2855         nam = rec->nr_address;
2856         m = rec->nr_packet;
2857         kfree(rec, M_NFSRVDESC);
2858         nd = kmalloc(sizeof(struct nfsrv_descript), M_NFSRVDESC, M_WAITOK);
2859         nd->nd_md = nd->nd_mrep = m;
2860         nd->nd_nam2 = nam;
2861         nd->nd_dpos = mtod(m, caddr_t);
2862         error = nfs_getreq(nd, nfsd, TRUE);
2863         if (error) {
2864                 if (nam) {
2865                         kfree(nam, M_SONAME);
2866                 }
2867                 kfree((caddr_t)nd, M_NFSRVDESC);
2868                 return (error);
2869         }
2870         *ndp = nd;
2871         nfsd->nfsd_nd = nd;
2872         return (0);
2873 }
2874
2875 /*
2876  * Try to assign service sockets to nfsd threads based on the number
2877  * of new rpc requests that have been queued on the service socket.
2878  *
2879  * If no nfsd's are available or additonal requests are pending, set the
2880  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2881  * the work in the nfssvc_sock list when it is finished processing its
2882  * current work.  This flag is only cleared when an nfsd can not find
2883  * any new work to perform.
2884  */
2885 void
2886 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2887 {
2888         struct nfsd *nd;
2889
2890         if ((slp->ns_flag & SLP_VALID) == 0)
2891                 return;
2892         if (nparallel <= 1)
2893                 nparallel = 1;
2894         TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2895                 if (nd->nfsd_flag & NFSD_WAITING) {
2896                         nd->nfsd_flag &= ~NFSD_WAITING;
2897                         if (nd->nfsd_slp)
2898                                 panic("nfsd wakeup");
2899                         nfsrv_slpref(slp);
2900                         nd->nfsd_slp = slp;
2901                         wakeup((caddr_t)nd);
2902                         if (--nparallel == 0)
2903                                 break;
2904                 }
2905         }
2906
2907         /*
2908          * If we couldn't assign slp then the NFSDs are all busy and
2909          * we set a flag indicating that there is pending work.
2910          */
2911         if (nparallel)
2912                 nfsd_head_flag |= NFSD_CHECKSLP;
2913 }
2914 #endif /* NFS_NOSERVER */