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. Neither the name of the University nor the names of its contributors
  17  *    may be used to endorse or promote products derived from this software
  18  *    without specific prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  *      @(#)nfs_socket.c        8.5 (Berkeley) 3/30/95
  33  * $FreeBSD: src/sys/nfs/nfs_socket.c,v 1.60.2.6 2003/03/26 01:44:46 alfred Exp $
  34  */
  35
  36 /*
  37  * Socket operations for use by nfs
  38  */
  39
  40 #include <sys/param.h>
  41 #include <sys/systm.h>
  42 #include <sys/proc.h>
  43 #include <sys/malloc.h>
  44 #include <sys/mount.h>
  45 #include <sys/kernel.h>
  46 #include <sys/mbuf.h>
  47 #include <sys/vnode.h>
  48 #include <sys/fcntl.h>
  49 #include <sys/protosw.h>
  50 #include <sys/resourcevar.h>
  51 #include <sys/socket.h>
  52 #include <sys/socketvar.h>
  53 #include <sys/socketops.h>
  54 #include <sys/syslog.h>
  55 #include <sys/thread.h>
  56 #include <sys/tprintf.h>
  57 #include <sys/sysctl.h>
  58 #include <sys/signalvar.h>
  59
  60 #include <sys/signal2.h>
  61 #include <sys/mutex2.h>
  62 #include <sys/socketvar2.h>
  63
  64 #include <netinet/in.h>
  65 #include <netinet/tcp.h>
  66 #include <sys/thread2.h>
  67
  68 #include "rpcv2.h"
  69 #include "nfsproto.h"
  70 #include "nfs.h"
  71 #include "xdr_subs.h"
  72 #include "nfsm_subs.h"
  73 #include "nfsmount.h"
  74 #include "nfsnode.h"
  75 #include "nfsrtt.h"
  76
  77 #define TRUE    1
  78 #define FALSE   0
  79
  80 /*
  81  * RTT calculations are scaled by 256 (8 bits).  A proper fractional
  82  * RTT will still be calculated even with a slow NFS timer.
  83  */
  84 #define NFS_SRTT(r)     (r)->r_nmp->nm_srtt[proct[(r)->r_procnum]]
  85 #define NFS_SDRTT(r)    (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum]]
  86 #define NFS_RTT_SCALE_BITS      8       /* bits */
  87 #define NFS_RTT_SCALE           256     /* value */
  88
  89 /*
  90  * Defines which timer to use for the procnum.
  91  * 0 - default
  92  * 1 - getattr
  93  * 2 - lookup
  94  * 3 - read
  95  * 4 - write
  96  */
  97 static int proct[NFS_NPROCS] = {
  98         0, 1, 0, 2, 1, 3, 3, 4, 0, 0,   /* 00-09        */
  99         0, 0, 0, 0, 0, 0, 3, 3, 0, 0,   /* 10-19        */
 100         0, 5, 0, 0, 0, 0,               /* 20-29        */
 101 };
 102
 103 static int multt[NFS_NPROCS] = {
 104         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 00-09        */
 105         1, 1, 1, 1, 1, 1, 1, 1, 1, 1,   /* 10-19        */
 106         1, 2, 1, 1, 1, 1,               /* 20-29        */
 107 };
 108
 109 static int nfs_backoff[8] = { 2, 3, 5, 8, 13, 21, 34, 55 };
 110 static int nfs_realign_test;
 111 static int nfs_realign_count;
 112 static int nfs_showrtt;
 113 static int nfs_showrexmit;
 114 int nfs_maxasyncbio = NFS_MAXASYNCBIO;
 115
 116 SYSCTL_DECL(_vfs_nfs);
 117
 118 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0,
 119     "Number of times mbufs have been tested for bad alignment");
 120 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0,
 121     "Number of realignments for badly aligned mbuf data");
 122 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrtt, CTLFLAG_RW, &nfs_showrtt, 0,
 123     "Show round trip time output");
 124 SYSCTL_INT(_vfs_nfs, OID_AUTO, showrexmit, CTLFLAG_RW, &nfs_showrexmit, 0,
 125     "Show retransmits info");
 126 SYSCTL_INT(_vfs_nfs, OID_AUTO, maxasyncbio, CTLFLAG_RW, &nfs_maxasyncbio, 0,
 127     "Max number of asynchronous bio's");
 128
 129 static int nfs_request_setup(nfsm_info_t info);
 130 static int nfs_request_auth(struct nfsreq *rep);
 131 static int nfs_request_try(struct nfsreq *rep);
 132 static int nfs_request_waitreply(struct nfsreq *rep);
 133 static int nfs_request_processreply(nfsm_info_t info, int);
 134
 135 int nfsrtton = 0;
 136 struct nfsrtt nfsrtt;
 137 struct callout  nfs_timer_handle;
 138
 139 static int      nfs_msg (struct thread *,char *,char *);
 140 static int      nfs_rcvlock (struct nfsmount *nmp, struct nfsreq *myreq);
 141 static void     nfs_rcvunlock (struct nfsmount *nmp);
 142 static void     nfs_realign (struct mbuf **pm, int hsiz);
 143 static int      nfs_receive (struct nfsmount *nmp, struct nfsreq *rep,
 144                                 struct sockaddr **aname, struct mbuf **mp);
 145 static void     nfs_softterm (struct nfsreq *rep, int islocked);
 146 static void     nfs_hardterm (struct nfsreq *rep, int islocked);
 147 static int      nfs_reconnect (struct nfsmount *nmp, struct nfsreq *rep);
 148 #ifndef NFS_NOSERVER
 149 static int      nfsrv_getstream (struct nfssvc_sock *, int, int *);
 150 static void     nfs_timer_req(struct nfsreq *req);
 151 static void     nfs_checkpkt(struct mbuf *m, int len);
 152
 153 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
 154                                     struct nfssvc_sock *slp,
 155                                     struct thread *td,
 156                                     struct mbuf **mreqp) = {
 157         nfsrv_null,
 158         nfsrv_getattr,
 159         nfsrv_setattr,
 160         nfsrv_lookup,
 161         nfsrv3_access,
 162         nfsrv_readlink,
 163         nfsrv_read,
 164         nfsrv_write,
 165         nfsrv_create,
 166         nfsrv_mkdir,
 167         nfsrv_symlink,
 168         nfsrv_mknod,
 169         nfsrv_remove,
 170         nfsrv_rmdir,
 171         nfsrv_rename,
 172         nfsrv_link,
 173         nfsrv_readdir,
 174         nfsrv_readdirplus,
 175         nfsrv_statfs,
 176         nfsrv_fsinfo,
 177         nfsrv_pathconf,
 178         nfsrv_commit,
 179         nfsrv_noop,
 180         nfsrv_noop,
 181         nfsrv_noop,
 182         nfsrv_noop
 183 };
 184 #endif /* NFS_NOSERVER */
 185
 186 /*
 187  * Initialize sockets and congestion for a new NFS connection.
 188  * We do not free the sockaddr if error.
 189  */
 190 int
 191 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
 192 {
 193         struct socket *so;
 194         int error;
 195         struct sockaddr *saddr;
 196         struct sockaddr_in *sin;
 197         struct thread *td = &thread0; /* only used for socreate and sobind */
 198
 199         nmp->nm_so = so = NULL;
 200         if (nmp->nm_flag & NFSMNT_FORCE)
 201                 return (EINVAL);
 202         saddr = nmp->nm_nam;
 203         error = socreate(saddr->sa_family, &so, nmp->nm_sotype,
 204                 nmp->nm_soproto, td);
 205         if (error)
 206                 goto bad;
 207         nmp->nm_soflags = so->so_proto->pr_flags;
 208
 209         /*
 210          * Some servers require that the client port be a reserved port number.
 211          */
 212         if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
 213                 struct sockopt sopt;
 214                 int ip;
 215                 struct sockaddr_in ssin;
 216
 217                 bzero(&sopt, sizeof sopt);
 218                 ip = IP_PORTRANGE_LOW;
 219                 sopt.sopt_level = IPPROTO_IP;
 220                 sopt.sopt_name = IP_PORTRANGE;
 221                 sopt.sopt_val = (void *)&ip;
 222                 sopt.sopt_valsize = sizeof(ip);
 223                 sopt.sopt_td = NULL;
 224                 error = sosetopt(so, &sopt);
 225                 if (error)
 226                         goto bad;
 227                 bzero(&ssin, sizeof ssin);
 228                 sin = &ssin;
 229                 sin->sin_len = sizeof (struct sockaddr_in);
 230                 sin->sin_family = AF_INET;
 231                 sin->sin_addr.s_addr = INADDR_ANY;
 232                 sin->sin_port = htons(0);
 233                 error = sobind(so, (struct sockaddr *)sin, td);
 234                 if (error)
 235                         goto bad;
 236                 bzero(&sopt, sizeof sopt);
 237                 ip = IP_PORTRANGE_DEFAULT;
 238                 sopt.sopt_level = IPPROTO_IP;
 239                 sopt.sopt_name = IP_PORTRANGE;
 240                 sopt.sopt_val = (void *)&ip;
 241                 sopt.sopt_valsize = sizeof(ip);
 242                 sopt.sopt_td = NULL;
 243                 error = sosetopt(so, &sopt);
 244                 if (error)
 245                         goto bad;
 246         }
 247
 248         /*
 249          * Protocols that do not require connections may be optionally left
 250          * unconnected for servers that reply from a port other than NFS_PORT.
 251          */
 252         if (nmp->nm_flag & NFSMNT_NOCONN) {
 253                 if (nmp->nm_soflags & PR_CONNREQUIRED) {
 254                         error = ENOTCONN;
 255                         goto bad;
 256                 }
 257         } else {
 258                 error = soconnect(so, nmp->nm_nam, td, TRUE);
 259                 if (error)
 260                         goto bad;
 261
 262                 /*
 263                  * Wait for the connection to complete. Cribbed from the
 264                  * connect system call but with the wait timing out so
 265                  * that interruptible mounts don't hang here for a long time.
 266                  */
 267                 crit_enter();
 268                 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
 269                         (void) tsleep((caddr_t)&so->so_timeo, 0,
 270                                 "nfscon", 2 * hz);
 271                         if ((so->so_state & SS_ISCONNECTING) &&
 272                             so->so_error == 0 && rep &&
 273                             (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
 274                                 soclrstate(so, SS_ISCONNECTING);
 275                                 crit_exit();
 276                                 goto bad;
 277                         }
 278                 }
 279                 if (so->so_error) {
 280                         error = so->so_error;
 281                         so->so_error = 0;
 282                         crit_exit();
 283                         goto bad;
 284                 }
 285                 crit_exit();
 286         }
 287         so->so_rcv.ssb_timeo = (5 * hz);
 288         so->so_snd.ssb_timeo = (5 * hz);
 289
 290         /*
 291          * Get buffer reservation size from sysctl, but impose reasonable
 292          * limits.
 293          */
 294         if (nmp->nm_sotype == SOCK_STREAM) {
 295                 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
 296                         struct sockopt sopt;
 297                         int val;
 298
 299                         bzero(&sopt, sizeof sopt);
 300                         sopt.sopt_level = SOL_SOCKET;
 301                         sopt.sopt_name = SO_KEEPALIVE;
 302                         sopt.sopt_val = &val;
 303                         sopt.sopt_valsize = sizeof val;
 304                         val = 1;
 305                         sosetopt(so, &sopt);
 306                 }
 307                 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
 308                         struct sockopt sopt;
 309                         int val;
 310
 311                         bzero(&sopt, sizeof sopt);
 312                         sopt.sopt_level = IPPROTO_TCP;
 313                         sopt.sopt_name = TCP_NODELAY;
 314                         sopt.sopt_val = &val;
 315                         sopt.sopt_valsize = sizeof val;
 316                         val = 1;
 317                         sosetopt(so, &sopt);
 318
 319                         bzero(&sopt, sizeof sopt);
 320                         sopt.sopt_level = IPPROTO_TCP;
 321                         sopt.sopt_name = TCP_FASTKEEP;
 322                         sopt.sopt_val = &val;
 323                         sopt.sopt_valsize = sizeof val;
 324                         val = 1;
 325                         sosetopt(so, &sopt);
 326                 }
 327         }
 328         error = soreserve(so, nfs_soreserve, nfs_soreserve, NULL);
 329         if (error)
 330                 goto bad;
 331         atomic_set_int(&so->so_rcv.ssb_flags, SSB_NOINTR);
 332         atomic_set_int(&so->so_snd.ssb_flags, SSB_NOINTR);
 333
 334         /*
 335          * Clear AUTOSIZE, otherwise the socket buffer could be reduced
 336          * to the point where rpc's cannot be queued using the mbuf
 337          * interface.
 338          */
 339         atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
 340         atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
 341
 342         /* Initialize other non-zero congestion variables */
 343         nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] =
 344                 nmp->nm_srtt[3] = (NFS_TIMEO << NFS_RTT_SCALE_BITS);
 345         nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
 346                 nmp->nm_sdrtt[3] = 0;
 347         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
 348         nmp->nm_timeouts = 0;
 349
 350         /*
 351          * Assign nm_so last.  The moment nm_so is assigned the nfs_timer()
 352          * can mess with the socket.
 353          */
 354         nmp->nm_so = so;
 355         return (0);
 356
 357 bad:
 358         if (so) {
 359                 soshutdown(so, SHUT_RDWR);
 360                 soclose(so, FNONBLOCK);
 361         }
 362         return (error);
 363 }
 364
 365 /*
 366  * Reconnect routine:
 367  * Called when a connection is broken on a reliable protocol.
 368  * - clean up the old socket
 369  * - nfs_connect() again
 370  * - set R_NEEDSXMIT for all outstanding requests on mount point
 371  * If this fails the mount point is DEAD!
 372  * nb: Must be called with the nfs_sndlock() set on the mount point.
 373  */
 374 static int
 375 nfs_reconnect(struct nfsmount *nmp, struct nfsreq *rep)
 376 {
 377         struct nfsreq *req;
 378         int error;
 379
 380         nfs_disconnect(nmp);
 381         if (nmp->nm_rxstate >= NFSSVC_STOPPING)
 382                 return (EINTR);
 383         while ((error = nfs_connect(nmp, rep)) != 0) {
 384                 if (error == EINTR || error == ERESTART)
 385                         return (EINTR);
 386                 if (error == EINVAL)
 387                         return (error);
 388                 if (nmp->nm_rxstate >= NFSSVC_STOPPING)
 389                         return (EINTR);
 390                 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
 391         }
 392
 393         /*
 394          * Loop through outstanding request list and fix up all requests
 395          * on old socket.
 396          */
 397         crit_enter();
 398         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
 399                 KKASSERT(req->r_nmp == nmp);
 400                 req->r_flags |= R_NEEDSXMIT;
 401         }
 402         crit_exit();
 403         return (0);
 404 }
 405
 406 /*
 407  * NFS disconnect. Clean up and unlink.
 408  */
 409 void
 410 nfs_disconnect(struct nfsmount *nmp)
 411 {
 412         struct socket *so;
 413
 414         if (nmp->nm_so) {
 415                 so = nmp->nm_so;
 416                 nmp->nm_so = NULL;
 417                 soshutdown(so, SHUT_RDWR);
 418                 soclose(so, FNONBLOCK);
 419         }
 420 }
 421
 422 void
 423 nfs_safedisconnect(struct nfsmount *nmp)
 424 {
 425         int error;
 426
 427         error = nfs_rcvlock(nmp, NULL);
 428         nfs_disconnect(nmp);
 429         if (error == 0)
 430                 nfs_rcvunlock(nmp);
 431 }
 432
 433 /*
 434  * This is the nfs send routine. For connection based socket types, it
 435  * must be called with an nfs_sndlock() on the socket.
 436  * "rep == NULL" indicates that it has been called from a server.
 437  * For the client side:
 438  * - return EINTR if the RPC is terminated, 0 otherwise
 439  * - set R_NEEDSXMIT if the send fails for any reason
 440  * - do any cleanup required by recoverable socket errors (?)
 441  * For the server side:
 442  * - return EINTR or ERESTART if interrupted by a signal
 443  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
 444  * - do any cleanup required by recoverable socket errors (?)
 445  */
 446 int
 447 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
 448          struct nfsreq *rep)
 449 {
 450         struct sockaddr *sendnam;
 451         int error, soflags, flags;
 452
 453         if (rep) {
 454                 if (rep->r_flags & R_SOFTTERM) {
 455                         m_freem(top);
 456                         return (EINTR);
 457                 }
 458                 if ((so = rep->r_nmp->nm_so) == NULL) {
 459                         rep->r_flags |= R_NEEDSXMIT;
 460                         m_freem(top);
 461                         return (0);
 462                 }
 463                 rep->r_flags &= ~R_NEEDSXMIT;
 464                 soflags = rep->r_nmp->nm_soflags;
 465         } else {
 466                 soflags = so->so_proto->pr_flags;
 467         }
 468         if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
 469                 sendnam = NULL;
 470         else
 471                 sendnam = nam;
 472         if (so->so_type == SOCK_SEQPACKET)
 473                 flags = MSG_EOR;
 474         else
 475                 flags = 0;
 476
 477         /*
 478          * calls pru_sosend -> sosend -> so_pru_send -> netrpc
 479          */
 480         error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
 481                               curthread /*XXX*/);
 482
 483         /*
 484          * ENOBUFS for dgram sockets is transient and non fatal.
 485          * No need to log, and no need to break a soft mount.
 486          */
 487         if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
 488                 error = 0;
 489                 /*
 490                  * do backoff retransmit on client
 491                  */
 492                 if (rep) {
 493                         if ((rep->r_nmp->nm_state & NFSSTA_SENDSPACE) == 0) {
 494                                 rep->r_nmp->nm_state |= NFSSTA_SENDSPACE;
 495                                 kprintf("Warning: NFS: Insufficient sendspace "
 496                                         "(%lu),\n"
 497                                         "\t You must increase vfs.nfs.soreserve"
 498                                         "or decrease vfs.nfs.maxasyncbio\n",
 499                                         so->so_snd.ssb_hiwat);
 500                         }
 501                         rep->r_flags |= R_NEEDSXMIT;
 502                 }
 503         }
 504
 505         if (error) {
 506                 if (rep) {
 507                         log(LOG_INFO, "nfs send error %d for server %s\n",error,
 508                             rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
 509                         /*
 510                          * Deal with errors for the client side.
 511                          */
 512                         if (rep->r_flags & R_SOFTTERM)
 513                                 error = EINTR;
 514                         else
 515                                 rep->r_flags |= R_NEEDSXMIT;
 516                 } else {
 517                         log(LOG_INFO, "nfsd send error %d\n", error);
 518                 }
 519
 520                 /*
 521                  * Handle any recoverable (soft) socket errors here. (?)
 522                  */
 523                 if (error != EINTR && error != ERESTART &&
 524                         error != EWOULDBLOCK && error != EPIPE)
 525                         error = 0;
 526         }
 527         return (error);
 528 }
 529
 530 /*
 531  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
 532  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
 533  * Mark and consolidate the data into a new mbuf list.
 534  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
 535  *     small mbufs.
 536  * For SOCK_STREAM we must be very careful to read an entire record once
 537  * we have read any of it, even if the system call has been interrupted.
 538  */
 539 static int
 540 nfs_receive(struct nfsmount *nmp, struct nfsreq *rep,
 541             struct sockaddr **aname, struct mbuf **mp)
 542 {
 543         struct socket *so;
 544         struct sockbuf sio;
 545         struct uio auio;
 546         struct iovec aio;
 547         struct mbuf *m;
 548         struct mbuf *control;
 549         u_int32_t len;
 550         struct sockaddr **getnam;
 551         int error, sotype, rcvflg;
 552         struct thread *td = curthread;  /* XXX */
 553
 554         /*
 555          * Set up arguments for soreceive()
 556          */
 557         *mp = NULL;
 558         *aname = NULL;
 559         sotype = nmp->nm_sotype;
 560
 561         /*
 562          * For reliable protocols, lock against other senders/receivers
 563          * in case a reconnect is necessary.
 564          * For SOCK_STREAM, first get the Record Mark to find out how much
 565          * more there is to get.
 566          * We must lock the socket against other receivers
 567          * until we have an entire rpc request/reply.
 568          */
 569         if (sotype != SOCK_DGRAM) {
 570                 error = nfs_sndlock(nmp, rep);
 571                 if (error)
 572                         return (error);
 573 tryagain:
 574                 /*
 575                  * Check for fatal errors and resending request.
 576                  */
 577                 /*
 578                  * Ugh: If a reconnect attempt just happened, nm_so
 579                  * would have changed. NULL indicates a failed
 580                  * attempt that has essentially shut down this
 581                  * mount point.
 582                  */
 583                 if (rep && (rep->r_mrep || (rep->r_flags & R_SOFTTERM))) {
 584                         nfs_sndunlock(nmp);
 585                         return (EINTR);
 586                 }
 587                 so = nmp->nm_so;
 588                 if (so == NULL) {
 589                         error = nfs_reconnect(nmp, rep);
 590                         if (error) {
 591                                 nfs_sndunlock(nmp);
 592                                 return (error);
 593                         }
 594                         goto tryagain;
 595                 }
 596                 while (rep && (rep->r_flags & R_NEEDSXMIT)) {
 597                         m = m_copym(rep->r_mreq, 0, M_COPYALL, M_WAITOK);
 598                         nfsstats.rpcretries++;
 599                         error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
 600                         if (error) {
 601                                 if (error == EINTR || error == ERESTART ||
 602                                     (error = nfs_reconnect(nmp, rep)) != 0) {
 603                                         nfs_sndunlock(nmp);
 604                                         return (error);
 605                                 }
 606                                 goto tryagain;
 607                         }
 608                 }
 609                 nfs_sndunlock(nmp);
 610                 if (sotype == SOCK_STREAM) {
 611                         /*
 612                          * Get the length marker from the stream
 613                          */
 614                         aio.iov_base = (caddr_t)&len;
 615                         aio.iov_len = sizeof(u_int32_t);
 616                         auio.uio_iov = &aio;
 617                         auio.uio_iovcnt = 1;
 618                         auio.uio_segflg = UIO_SYSSPACE;
 619                         auio.uio_rw = UIO_READ;
 620                         auio.uio_offset = 0;
 621                         auio.uio_resid = sizeof(u_int32_t);
 622                         auio.uio_td = td;
 623                         do {
 624                            rcvflg = MSG_WAITALL;
 625                            error = so_pru_soreceive(so, NULL, &auio, NULL,
 626                                                     NULL, &rcvflg);
 627                            if (error == EWOULDBLOCK && rep) {
 628                                 if (rep->r_flags & R_SOFTTERM)
 629                                         return (EINTR);
 630                            }
 631                         } while (error == EWOULDBLOCK);
 632
 633                         if (error == 0 && auio.uio_resid > 0) {
 634                             /*
 635                              * Only log short packets if not EOF
 636                              */
 637                             if (auio.uio_resid != sizeof(u_int32_t)) {
 638                                 log(LOG_INFO,
 639                                     "short receive (%d/%d) from nfs server %s\n",
 640                                     (int)(sizeof(u_int32_t) - auio.uio_resid),
 641                                     (int)sizeof(u_int32_t),
 642                                     nmp->nm_mountp->mnt_stat.f_mntfromname);
 643                             }
 644                             error = EPIPE;
 645                         }
 646                         if (error)
 647                                 goto errout;
 648                         len = ntohl(len) & ~0x80000000;
 649                         /*
 650                          * This is SERIOUS! We are out of sync with the sender
 651                          * and forcing a disconnect/reconnect is all I can do.
 652                          */
 653                         if (len > NFS_MAXPACKET) {
 654                             log(LOG_ERR, "%s (%d) from nfs server %s\n",
 655                                 "impossible packet length",
 656                                 len,
 657                                 nmp->nm_mountp->mnt_stat.f_mntfromname);
 658                             error = EFBIG;
 659                             goto errout;
 660                         }
 661
 662                         /*
 663                          * Get the rest of the packet as an mbuf chain
 664                          */
 665                         sbinit(&sio, len);
 666                         do {
 667                             rcvflg = MSG_WAITALL;
 668                             error = so_pru_soreceive(so, NULL, NULL, &sio,
 669                                                      NULL, &rcvflg);
 670                         } while (error == EWOULDBLOCK || error == EINTR ||
 671                                  error == ERESTART);
 672                         if (error == 0 && sio.sb_cc != len) {
 673                             if (sio.sb_cc != 0) {
 674                                 log(LOG_INFO,
 675                                     "short receive (%zu/%d) from nfs server %s\n",
 676                                     (size_t)len - auio.uio_resid, len,
 677                                     nmp->nm_mountp->mnt_stat.f_mntfromname);
 678                             }
 679                             error = EPIPE;
 680                         }
 681                         *mp = sio.sb_mb;
 682                 } else {
 683                         /*
 684                          * Non-stream, so get the whole packet by not
 685                          * specifying MSG_WAITALL and by specifying a large
 686                          * length.
 687                          *
 688                          * We have no use for control msg., but must grab them
 689                          * and then throw them away so we know what is going
 690                          * on.
 691                          */
 692                         sbinit(&sio, 100000000);
 693                         do {
 694                             rcvflg = 0;
 695                             error =  so_pru_soreceive(so, NULL, NULL, &sio,
 696                                                       &control, &rcvflg);
 697                             if (control)
 698                                 m_freem(control);
 699                             if (error == EWOULDBLOCK && rep) {
 700                                 if (rep->r_flags & R_SOFTTERM) {
 701                                         m_freem(sio.sb_mb);
 702                                         return (EINTR);
 703                                 }
 704                             }
 705                         } while (error == EWOULDBLOCK ||
 706                                  (error == 0 && sio.sb_mb == NULL && control));
 707                         if ((rcvflg & MSG_EOR) == 0)
 708                                 kprintf("Egad!!\n");
 709                         if (error == 0 && sio.sb_mb == NULL)
 710                                 error = EPIPE;
 711                         len = sio.sb_cc;
 712                         *mp = sio.sb_mb;
 713                 }
 714 errout:
 715                 if (error && error != EINTR && error != ERESTART) {
 716                         m_freem(*mp);
 717                         *mp = NULL;
 718                         if (error != EPIPE) {
 719                                 log(LOG_INFO,
 720                                     "receive error %d from nfs server %s\n",
 721                                     error,
 722                                  nmp->nm_mountp->mnt_stat.f_mntfromname);
 723                         }
 724                         error = nfs_sndlock(nmp, rep);
 725                         if (!error) {
 726                                 error = nfs_reconnect(nmp, rep);
 727                                 if (!error)
 728                                         goto tryagain;
 729                                 else
 730                                         nfs_sndunlock(nmp);
 731                         }
 732                 }
 733         } else {
 734                 if ((so = nmp->nm_so) == NULL)
 735                         return (EACCES);
 736                 if (so->so_state & SS_ISCONNECTED)
 737                         getnam = NULL;
 738                 else
 739                         getnam = aname;
 740                 sbinit(&sio, 100000000);
 741                 do {
 742                         rcvflg = 0;
 743                         error =  so_pru_soreceive(so, getnam, NULL, &sio,
 744                                                   NULL, &rcvflg);
 745                         if (error == EWOULDBLOCK && rep &&
 746                             (rep->r_flags & R_SOFTTERM)) {
 747                                 m_freem(sio.sb_mb);
 748                                 return (EINTR);
 749                         }
 750                 } while (error == EWOULDBLOCK);
 751
 752                 len = sio.sb_cc;
 753                 *mp = sio.sb_mb;
 754
 755                 /*
 756                  * A shutdown may result in no error and no mbuf.
 757                  * Convert to EPIPE.
 758                  */
 759                 if (*mp == NULL && error == 0)
 760                         error = EPIPE;
 761         }
 762         if (error) {
 763                 m_freem(*mp);
 764                 *mp = NULL;
 765         }
 766
 767         /*
 768          * Search for any mbufs that are not a multiple of 4 bytes long
 769          * or with m_data not longword aligned.
 770          * These could cause pointer alignment problems, so copy them to
 771          * well aligned mbufs.
 772          */
 773         nfs_realign(mp, 5 * NFSX_UNSIGNED);
 774         return (error);
 775 }
 776
 777 /*
 778  * Implement receipt of reply on a socket.
 779  *
 780  * We must search through the list of received datagrams matching them
 781  * with outstanding requests using the xid, until ours is found.
 782  *
 783  * If myrep is NULL we process packets on the socket until
 784  * interrupted or until nm_reqrxq is non-empty.
 785  */
 786 /* ARGSUSED */
 787 int
 788 nfs_reply(struct nfsmount *nmp, struct nfsreq *myrep)
 789 {
 790         struct nfsreq *rep;
 791         struct sockaddr *nam;
 792         u_int32_t rxid;
 793         u_int32_t *tl;
 794         int error;
 795         struct nfsm_info info;
 796
 797         /*
 798          * Loop around until we get our own reply
 799          */
 800         for (;;) {
 801                 /*
 802                  * Lock against other receivers so that I don't get stuck in
 803                  * sbwait() after someone else has received my reply for me.
 804                  * Also necessary for connection based protocols to avoid
 805                  * race conditions during a reconnect.
 806                  *
 807                  * If nfs_rcvlock() returns EALREADY, that means that
 808                  * the reply has already been recieved by another
 809                  * process and we can return immediately.  In this
 810                  * case, the lock is not taken to avoid races with
 811                  * other processes.
 812                  */
 813                 info.mrep = NULL;
 814
 815                 error = nfs_rcvlock(nmp, myrep);
 816                 if (error == EALREADY)
 817                         return (0);
 818                 if (error)
 819                         return (error);
 820
 821                 /*
 822                  * If myrep is NULL we are the receiver helper thread.
 823                  * Stop waiting for incoming replies if there are
 824                  * messages sitting on reqrxq that we need to process,
 825                  * or if a shutdown request is pending.
 826                  */
 827                 if (myrep == NULL && (TAILQ_FIRST(&nmp->nm_reqrxq) ||
 828                     nmp->nm_rxstate > NFSSVC_PENDING)) {
 829                         nfs_rcvunlock(nmp);
 830                         return(EWOULDBLOCK);
 831                 }
 832
 833                 /*
 834                  * Get the next Rpc reply off the socket
 835                  *
 836                  * We cannot release the receive lock until we've
 837                  * filled in rep->r_mrep, otherwise a waiting
 838                  * thread may deadlock in soreceive with no incoming
 839                  * packets expected.
 840                  */
 841                 error = nfs_receive(nmp, myrep, &nam, &info.mrep);
 842                 if (error) {
 843                         /*
 844                          * Ignore routing errors on connectionless protocols??
 845                          */
 846                         nfs_rcvunlock(nmp);
 847                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
 848                                 if (nmp->nm_so == NULL)
 849                                         return (error);
 850                                 nmp->nm_so->so_error = 0;
 851                                 continue;
 852                         }
 853                         return (error);
 854                 }
 855                 if (nam)
 856                         kfree(nam, M_SONAME);
 857
 858                 /*
 859                  * Get the xid and check that it is an rpc reply
 860                  */
 861                 info.md = info.mrep;
 862                 info.dpos = mtod(info.md, caddr_t);
 863                 NULLOUT(tl = nfsm_dissect(&info, 2*NFSX_UNSIGNED));
 864                 rxid = *tl++;
 865                 if (*tl != rpc_reply) {
 866                         nfsstats.rpcinvalid++;
 867                         m_freem(info.mrep);
 868                         info.mrep = NULL;
 869 nfsmout:
 870                         nfs_rcvunlock(nmp);
 871                         continue;
 872                 }
 873
 874                 /*
 875                  * Loop through the request list to match up the reply
 876                  * Iff no match, just drop the datagram.  On match, set
 877                  * r_mrep atomically to prevent the timer from messing
 878                  * around with the request after we have exited the critical
 879                  * section.
 880                  */
 881                 crit_enter();
 882                 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
 883                         if (rep->r_mrep == NULL && rxid == rep->r_xid)
 884                                 break;
 885                 }
 886
 887                 /*
 888                  * Fill in the rest of the reply if we found a match.
 889                  *
 890                  * Deal with duplicate responses if there was no match.
 891                  */
 892                 if (rep) {
 893                         rep->r_md = info.md;
 894                         rep->r_dpos = info.dpos;
 895                         if (nfsrtton) {
 896                                 struct rttl *rt;
 897
 898                                 rt = &nfsrtt.rttl[nfsrtt.pos];
 899                                 rt->proc = rep->r_procnum;
 900                                 rt->rto = 0;
 901                                 rt->sent = 0;
 902                                 rt->cwnd = nmp->nm_maxasync_scaled;
 903                                 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
 904                                 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
 905                                 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
 906                                 getmicrotime(&rt->tstamp);
 907                                 if (rep->r_flags & R_TIMING)
 908                                         rt->rtt = rep->r_rtt;
 909                                 else
 910                                         rt->rtt = 1000000;
 911                                 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
 912                         }
 913
 914                         /*
 915                          * New congestion control is based only on async
 916                          * requests.
 917                          */
 918                         if (nmp->nm_maxasync_scaled < NFS_MAXASYNC_SCALED)
 919                                 ++nmp->nm_maxasync_scaled;
 920                         if (rep->r_flags & R_SENT) {
 921                                 rep->r_flags &= ~R_SENT;
 922                         }
 923                         /*
 924                          * Update rtt using a gain of 0.125 on the mean
 925                          * and a gain of 0.25 on the deviation.
 926                          *
 927                          * NOTE SRTT/SDRTT are only good if R_TIMING is set.
 928                          */
 929                         if ((rep->r_flags & R_TIMING) && rep->r_rexmit == 0) {
 930                                 /*
 931                                  * Since the timer resolution of
 932                                  * NFS_HZ is so course, it can often
 933                                  * result in r_rtt == 0. Since
 934                                  * r_rtt == N means that the actual
 935                                  * rtt is between N+dt and N+2-dt ticks,
 936                                  * add 1.
 937                                  */
 938                                 int n;
 939                                 int d;
 940
 941 #define NFSRSB  NFS_RTT_SCALE_BITS
 942                                 n = ((NFS_SRTT(rep) * 7) +
 943                                      (rep->r_rtt << NFSRSB)) >> 3;
 944                                 d = n - NFS_SRTT(rep);
 945                                 NFS_SRTT(rep) = n;
 946
 947                                 /*
 948                                  * Don't let the jitter calculation decay
 949                                  * too quickly, but we want a fast rampup.
 950                                  */
 951                                 if (d < 0)
 952                                         d = -d;
 953                                 d <<= NFSRSB;
 954                                 if (d < NFS_SDRTT(rep))
 955                                         n = ((NFS_SDRTT(rep) * 15) + d) >> 4;
 956                                 else
 957                                         n = ((NFS_SDRTT(rep) * 3) + d) >> 2;
 958                                 NFS_SDRTT(rep) = n;
 959 #undef NFSRSB
 960                         }
 961                         nmp->nm_timeouts = 0;
 962                         rep->r_mrep = info.mrep;
 963                         nfs_hardterm(rep, 0);
 964                 } else {
 965                         /*
 966                          * Extract vers, prog, nfsver, procnum.  A duplicate
 967                          * response means we didn't wait long enough so
 968                          * we increase the SRTT to avoid future spurious
 969                          * timeouts.
 970                          */
 971                         u_int procnum = nmp->nm_lastreprocnum;
 972                         int n;
 973
 974                         if (procnum < NFS_NPROCS && proct[procnum]) {
 975                                 if (nfs_showrexmit)
 976                                         kprintf("D");
 977                                 n = nmp->nm_srtt[proct[procnum]];
 978                                 n += NFS_ASYSCALE * NFS_HZ;
 979                                 if (n < NFS_ASYSCALE * NFS_HZ * 10)
 980                                         n = NFS_ASYSCALE * NFS_HZ * 10;
 981                                 nmp->nm_srtt[proct[procnum]] = n;
 982                         }
 983                 }
 984                 nfs_rcvunlock(nmp);
 985                 crit_exit();
 986
 987                 /*
 988                  * If not matched to a request, drop it.
 989                  * If it's mine, get out.
 990                  */
 991                 if (rep == NULL) {
 992                         nfsstats.rpcunexpected++;
 993                         m_freem(info.mrep);
 994                         info.mrep = NULL;
 995                 } else if (rep == myrep) {
 996                         if (rep->r_mrep == NULL)
 997                                 panic("nfsreply nil");
 998                         return (0);
 999                 }
1000         }
1001 }
1002
1003 /*
1004  * Run the request state machine until the target state is reached
1005  * or a fatal error occurs.  The target state is not run.  Specifying
1006  * a target of NFSM_STATE_DONE runs the state machine until the rpc
1007  * is complete.
1008  *
1009  * EINPROGRESS is returned for all states other then the DONE state,
1010  * indicating that the rpc is still in progress.
1011  */
1012 int
1013 nfs_request(struct nfsm_info *info, nfsm_state_t bstate, nfsm_state_t estate)
1014 {
1015         struct nfsreq *req;
1016
1017         while (info->state >= bstate && info->state < estate) {
1018                 switch(info->state) {
1019                 case NFSM_STATE_SETUP:
1020                         /*
1021                          * Setup the nfsreq.  Any error which occurs during
1022                          * this state is fatal.
1023                          */
1024                         info->error = nfs_request_setup(info);
1025                         if (info->error) {
1026                                 info->state = NFSM_STATE_DONE;
1027                                 return (info->error);
1028                         } else {
1029                                 req = info->req;
1030                                 req->r_mrp = &info->mrep;
1031                                 req->r_mdp = &info->md;
1032                                 req->r_dposp = &info->dpos;
1033                                 info->state = NFSM_STATE_AUTH;
1034                         }
1035                         break;
1036                 case NFSM_STATE_AUTH:
1037                         /*
1038                          * Authenticate the nfsreq.  Any error which occurs
1039                          * during this state is fatal.
1040                          */
1041                         info->error = nfs_request_auth(info->req);
1042                         if (info->error) {
1043                                 info->state = NFSM_STATE_DONE;
1044                                 return (info->error);
1045                         } else {
1046                                 info->state = NFSM_STATE_TRY;
1047                         }
1048                         break;
1049                 case NFSM_STATE_TRY:
1050                         /*
1051                          * Transmit or retransmit attempt.  An error in this
1052                          * state is ignored and we always move on to the
1053                          * next state.
1054                          *
1055                          * This can trivially race the receiver if the
1056                          * request is asynchronous.  nfs_request_try()
1057                          * will thus set the state for us and we
1058                          * must also return immediately if we are
1059                          * running an async state machine, because
1060                          * info can become invalid due to races after
1061                          * try() returns.
1062                          */
1063                         if (info->req->r_flags & R_ASYNC) {
1064                                 nfs_request_try(info->req);
1065                                 if (estate == NFSM_STATE_WAITREPLY)
1066                                         return (EINPROGRESS);
1067                         } else {
1068                                 nfs_request_try(info->req);
1069                                 info->state = NFSM_STATE_WAITREPLY;
1070                         }
1071                         break;
1072                 case NFSM_STATE_WAITREPLY:
1073                         /*
1074                          * Wait for a reply or timeout and move on to the
1075                          * next state.  The error returned by this state
1076                          * is passed to the processing code in the next
1077                          * state.
1078                          */
1079                         info->error = nfs_request_waitreply(info->req);
1080                         info->state = NFSM_STATE_PROCESSREPLY;
1081                         break;
1082                 case NFSM_STATE_PROCESSREPLY:
1083                         /*
1084                          * Process the reply or timeout.  Errors which occur
1085                          * in this state may cause the state machine to
1086                          * go back to an earlier state, and are fatal
1087                          * otherwise.
1088                          */
1089                         info->error = nfs_request_processreply(info,
1090                                                                info->error);
1091                         switch(info->error) {
1092                         case ENEEDAUTH:
1093                                 info->state = NFSM_STATE_AUTH;
1094                                 break;
1095                         case EAGAIN:
1096                                 info->state = NFSM_STATE_TRY;
1097                                 break;
1098                         default:
1099                                 /*
1100                                  * Operation complete, with or without an
1101                                  * error.  We are done.
1102                                  */
1103                                 info->req = NULL;
1104                                 info->state = NFSM_STATE_DONE;
1105                                 return (info->error);
1106                         }
1107                         break;
1108                 case NFSM_STATE_DONE:
1109                         /*
1110                          * Shouldn't be reached
1111                          */
1112                         return (info->error);
1113                         /* NOT REACHED */
1114                 }
1115         }
1116
1117         /*
1118          * If we are done return the error code (if any).
1119          * Otherwise return EINPROGRESS.
1120          */
1121         if (info->state == NFSM_STATE_DONE)
1122                 return (info->error);
1123         return (EINPROGRESS);
1124 }
1125
1126 /*
1127  * nfs_request - goes something like this
1128  *      - fill in request struct
1129  *      - links it into list
1130  *      - calls nfs_send() for first transmit
1131  *      - calls nfs_receive() to get reply
1132  *      - break down rpc header and return with nfs reply pointed to
1133  *        by mrep or error
1134  * nb: always frees up mreq mbuf list
1135  */
1136 static int
1137 nfs_request_setup(nfsm_info_t info)
1138 {
1139         struct nfsreq *req;
1140         struct nfsmount *nmp;
1141         struct mbuf *m;
1142         int i;
1143
1144         /*
1145          * Reject requests while attempting a forced unmount.
1146          */
1147         if (info->vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
1148                 m_freem(info->mreq);
1149                 info->mreq = NULL;
1150                 return (EIO);
1151         }
1152         nmp = VFSTONFS(info->vp->v_mount);
1153         req = kmalloc(sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
1154         req->r_nmp = nmp;
1155         req->r_vp = info->vp;
1156         req->r_td = info->td;
1157         req->r_procnum = info->procnum;
1158         req->r_mreq = NULL;
1159         req->r_cred = info->cred;
1160
1161         i = 0;
1162         m = info->mreq;
1163         while (m) {
1164                 i += m->m_len;
1165                 m = m->m_next;
1166         }
1167         req->r_mrest = info->mreq;
1168         req->r_mrest_len = i;
1169
1170         /*
1171          * The presence of a non-NULL r_info in req indicates
1172          * async completion via our helper threads.  See the receiver
1173          * code.
1174          */
1175         if (info->bio) {
1176                 req->r_info = info;
1177                 req->r_flags = R_ASYNC;
1178         } else {
1179                 req->r_info = NULL;
1180                 req->r_flags = 0;
1181         }
1182         info->req = req;
1183         return(0);
1184 }
1185
1186 static int
1187 nfs_request_auth(struct nfsreq *rep)
1188 {
1189         struct nfsmount *nmp = rep->r_nmp;
1190         struct mbuf *m;
1191         char nickv[RPCX_NICKVERF];
1192         int error = 0, auth_len, auth_type;
1193         int verf_len;
1194         u_int32_t xid;
1195         char *auth_str, *verf_str;
1196         struct ucred *cred;
1197
1198         cred = rep->r_cred;
1199         rep->r_failed_auth = 0;
1200
1201         /*
1202          * Get the RPC header with authorization.
1203          */
1204         verf_str = auth_str = NULL;
1205         if (nmp->nm_flag & NFSMNT_KERB) {
1206                 verf_str = nickv;
1207                 verf_len = sizeof (nickv);
1208                 auth_type = RPCAUTH_KERB4;
1209                 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1210                 if (rep->r_failed_auth ||
1211                     nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1212                                     verf_str, verf_len)) {
1213                         error = nfs_getauth(nmp, rep, cred, &auth_str,
1214                                 &auth_len, verf_str, &verf_len, rep->r_key);
1215                         if (error) {
1216                                 m_freem(rep->r_mrest);
1217                                 rep->r_mrest = NULL;
1218                                 kfree((caddr_t)rep, M_NFSREQ);
1219                                 return (error);
1220                         }
1221                 }
1222         } else {
1223                 auth_type = RPCAUTH_UNIX;
1224                 if (cred->cr_ngroups < 1)
1225                         panic("nfsreq nogrps");
1226                 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1227                         nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1228                         5 * NFSX_UNSIGNED;
1229         }
1230         if (rep->r_mrest)
1231                 nfs_checkpkt(rep->r_mrest, rep->r_mrest_len);
1232         m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1233                         auth_len, auth_str, verf_len, verf_str,
1234                         rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1235         rep->r_mrest = NULL;
1236         if (auth_str)
1237                 kfree(auth_str, M_TEMP);
1238
1239         /*
1240          * For stream protocols, insert a Sun RPC Record Mark.
1241          */
1242         if (nmp->nm_sotype == SOCK_STREAM) {
1243                 M_PREPEND(m, NFSX_UNSIGNED, M_WAITOK);
1244                 if (m == NULL) {
1245                         kfree(rep, M_NFSREQ);
1246                         return (ENOBUFS);
1247                 }
1248                 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1249                          (m->m_pkthdr.len - NFSX_UNSIGNED));
1250         }
1251
1252         nfs_checkpkt(m, m->m_pkthdr.len);
1253
1254         rep->r_mreq = m;
1255         rep->r_xid = xid;
1256         return (0);
1257 }
1258
1259 static int
1260 nfs_request_try(struct nfsreq *rep)
1261 {
1262         struct nfsmount *nmp = rep->r_nmp;
1263         struct mbuf *m2;
1264         int error;
1265
1266         /*
1267          * Request is not on any queue, only the owner has access to it
1268          * so it should not be locked by anyone atm.
1269          *
1270          * Interlock to prevent races.  While locked the only remote
1271          * action possible is for r_mrep to be set (once we enqueue it).
1272          */
1273         if (rep->r_flags == 0xdeadc0de) {
1274                 print_backtrace(-1);
1275                 panic("flags nbad");
1276         }
1277         KKASSERT((rep->r_flags & (R_LOCKED | R_ONREQQ)) == 0);
1278         if (nmp->nm_flag & NFSMNT_SOFT)
1279                 rep->r_retry = nmp->nm_retry;
1280         else
1281                 rep->r_retry = NFS_MAXREXMIT + 1;       /* past clip limit */
1282         rep->r_rtt = rep->r_rexmit = 0;
1283         if (proct[rep->r_procnum] > 0)
1284                 rep->r_flags |= R_TIMING | R_LOCKED;
1285         else
1286                 rep->r_flags |= R_LOCKED;
1287         rep->r_mrep = NULL;
1288
1289         nfsstats.rpcrequests++;
1290
1291         if (nmp->nm_flag & NFSMNT_FORCE) {
1292                 rep->r_flags |= R_SOFTTERM;
1293                 rep->r_flags &= ~R_LOCKED;
1294                 if (rep->r_info)
1295                         rep->r_info->error = EINTR;
1296                 return (0);
1297         }
1298         rep->r_flags |= R_NEEDSXMIT;    /* in case send lock races us */
1299
1300         /*
1301          * Do the client side RPC.
1302          *
1303          * Chain request into list of outstanding requests. Be sure
1304          * to put it LAST so timer finds oldest requests first.  Note
1305          * that our control of R_LOCKED prevents the request from
1306          * getting ripped out from under us or transmitted by the
1307          * timer code.
1308          *
1309          * For requests with info structures we must atomically set the
1310          * info's state because the structure could become invalid upon
1311          * return due to races (i.e., if async)
1312          */
1313         crit_enter();
1314         mtx_link_init(&rep->r_link);
1315         KKASSERT((rep->r_flags & R_ONREQQ) == 0);
1316         TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1317         rep->r_flags |= R_ONREQQ;
1318         ++nmp->nm_reqqlen;
1319         if (rep->r_flags & R_ASYNC)
1320                 rep->r_info->state = NFSM_STATE_WAITREPLY;
1321         crit_exit();
1322
1323         error = 0;
1324
1325         /*
1326          * Send if we can.  Congestion control is not handled here any more
1327          * becausing trying to defer the initial send based on the nfs_timer
1328          * requires having a very fast nfs_timer, which is silly.
1329          */
1330         if (nmp->nm_so) {
1331                 if (nmp->nm_soflags & PR_CONNREQUIRED)
1332                         error = nfs_sndlock(nmp, rep);
1333                 if (error == 0 && (rep->r_flags & R_NEEDSXMIT)) {
1334                         m2 = m_copym(rep->r_mreq, 0, M_COPYALL, M_WAITOK);
1335                         error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1336                         rep->r_flags &= ~R_NEEDSXMIT;
1337                         if ((rep->r_flags & R_SENT) == 0) {
1338                                 rep->r_flags |= R_SENT;
1339                         }
1340                         if (nmp->nm_soflags & PR_CONNREQUIRED)
1341                                 nfs_sndunlock(nmp);
1342                 }
1343         } else {
1344                 rep->r_rtt = -1;
1345         }
1346         if (error == EPIPE)
1347                 error = 0;
1348
1349         /*
1350          * Release the lock.  The only remote action that may have occurred
1351          * would have been the setting of rep->r_mrep.  If this occured
1352          * and the request was async we have to move it to the reader
1353          * thread's queue for action.
1354          *
1355          * For async requests also make sure the reader is woken up so
1356          * it gets on the socket to read responses.
1357          */
1358         crit_enter();
1359         if (rep->r_flags & R_ASYNC) {
1360                 if (rep->r_mrep)
1361                         nfs_hardterm(rep, 1);
1362                 rep->r_flags &= ~R_LOCKED;
1363                 nfssvc_iod_reader_wakeup(nmp);
1364         } else {
1365                 rep->r_flags &= ~R_LOCKED;
1366         }
1367         if (rep->r_flags & R_WANTED) {
1368                 rep->r_flags &= ~R_WANTED;
1369                 wakeup(rep);
1370         }
1371         crit_exit();
1372         return (error);
1373 }
1374
1375 /*
1376  * This code is only called for synchronous requests.  Completed synchronous
1377  * requests are left on reqq and we remove them before moving on to the
1378  * processing state.
1379  */
1380 static int
1381 nfs_request_waitreply(struct nfsreq *rep)
1382 {
1383         struct nfsmount *nmp = rep->r_nmp;
1384         int error;
1385
1386         KKASSERT((rep->r_flags & R_ASYNC) == 0);
1387
1388         /*
1389          * Wait until the request is finished.
1390          */
1391         error = nfs_reply(nmp, rep);
1392
1393         /*
1394          * RPC done, unlink the request, but don't rip it out from under
1395          * the callout timer.
1396          *
1397          * Once unlinked no other receiver or the timer will have
1398          * visibility, so we do not have to set R_LOCKED.
1399          */
1400         crit_enter();
1401         while (rep->r_flags & R_LOCKED) {
1402                 rep->r_flags |= R_WANTED;
1403                 tsleep(rep, 0, "nfstrac", 0);
1404         }
1405         KKASSERT(rep->r_flags & R_ONREQQ);
1406         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1407         rep->r_flags &= ~R_ONREQQ;
1408         --nmp->nm_reqqlen;
1409         if (TAILQ_FIRST(&nmp->nm_bioq) &&
1410             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
1411                 nfssvc_iod_writer_wakeup(nmp);
1412         }
1413         crit_exit();
1414
1415         /*
1416          * Decrement the outstanding request count.
1417          */
1418         if (rep->r_flags & R_SENT) {
1419                 rep->r_flags &= ~R_SENT;
1420         }
1421         return (error);
1422 }
1423
1424 /*
1425  * Process reply with error returned from nfs_requet_waitreply().
1426  *
1427  * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1428  * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1429  */
1430 static int
1431 nfs_request_processreply(nfsm_info_t info, int error)
1432 {
1433         struct nfsreq *req = info->req;
1434         struct nfsmount *nmp = req->r_nmp;
1435         u_int32_t *tl;
1436         int verf_type;
1437         int i;
1438
1439         /*
1440          * If there was a successful reply and a tprintf msg.
1441          * tprintf a response.
1442          */
1443         if (error == 0 && (req->r_flags & R_TPRINTFMSG)) {
1444                 nfs_msg(req->r_td,
1445                         nmp->nm_mountp->mnt_stat.f_mntfromname,
1446                         "is alive again");
1447         }
1448
1449         /*
1450          * Assign response and handle any pre-process error.  Response
1451          * fields can be NULL if an error is already pending.
1452          */
1453         info->mrep = req->r_mrep;
1454         info->md = req->r_md;
1455         info->dpos = req->r_dpos;
1456
1457         if (error) {
1458                 m_freem(req->r_mreq);
1459                 req->r_mreq = NULL;
1460                 kfree(req, M_NFSREQ);
1461                 info->req = NULL;
1462                 return (error);
1463         }
1464
1465         /*
1466          * break down the rpc header and check if ok
1467          */
1468         NULLOUT(tl = nfsm_dissect(info, 3 * NFSX_UNSIGNED));
1469         if (*tl++ == rpc_msgdenied) {
1470                 if (*tl == rpc_mismatch) {
1471                         error = EOPNOTSUPP;
1472                 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1473                            *tl++ == rpc_autherr) {
1474                         if (req->r_failed_auth == 0) {
1475                                 req->r_failed_auth++;
1476                                 req->r_mheadend->m_next = NULL;
1477                                 m_freem(info->mrep);
1478                                 info->mrep = NULL;
1479                                 m_freem(req->r_mreq);
1480                                 req->r_mreq = NULL;
1481                                 return (ENEEDAUTH);
1482                         } else {
1483                                 error = EAUTH;
1484                         }
1485                 } else {
1486                         error = EACCES;
1487                 }
1488                 m_freem(info->mrep);
1489                 info->mrep = NULL;
1490                 m_freem(req->r_mreq);
1491                 req->r_mreq = NULL;
1492                 kfree(req, M_NFSREQ);
1493                 info->req = NULL;
1494                 return (error);
1495         }
1496
1497         /*
1498          * Grab any Kerberos verifier, otherwise just throw it away.
1499          */
1500         verf_type = fxdr_unsigned(int, *tl++);
1501         i = fxdr_unsigned(int32_t, *tl);
1502         if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1503                 error = nfs_savenickauth(nmp, req->r_cred, i, req->r_key,
1504                                          &info->md, &info->dpos, info->mrep);
1505                 if (error)
1506                         goto nfsmout;
1507         } else if (i > 0) {
1508                 ERROROUT(nfsm_adv(info, nfsm_rndup(i)));
1509         }
1510         NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1511         /* 0 == ok */
1512         if (*tl == 0) {
1513                 NULLOUT(tl = nfsm_dissect(info, NFSX_UNSIGNED));
1514                 if (*tl != 0) {
1515                         error = fxdr_unsigned(int, *tl);
1516
1517                         /*
1518                          * Does anyone even implement this?  Just impose
1519                          * a 1-second delay.
1520                          */
1521                         if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1522                                 error == NFSERR_TRYLATER) {
1523                                 m_freem(info->mrep);
1524                                 info->mrep = NULL;
1525                                 error = 0;
1526
1527                                 tsleep((caddr_t)&lbolt, 0, "nqnfstry", 0);
1528                                 return (EAGAIN);        /* goto tryagain */
1529                         }
1530
1531 #if 0
1532                         /*
1533                          * XXX We can't do this here any more because the
1534                          *     caller may be holding a shared lock on the
1535                          *     namecache entry.
1536                          *
1537                          * If the File Handle was stale, invalidate the
1538                          * lookup cache, just in case.
1539                          *
1540                          * To avoid namecache<->vnode deadlocks we must
1541                          * release the vnode lock if we hold it.
1542                          */
1543                         if (error == ESTALE) {
1544                                 struct vnode *vp = req->r_vp;
1545                                 int ltype;
1546
1547                                 ltype = lockstatus(&vp->v_lock, curthread);
1548                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1549                                         lockmgr(&vp->v_lock, LK_RELEASE);
1550                                 cache_inval_vp(vp, CINV_CHILDREN);
1551                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1552                                         lockmgr(&vp->v_lock, ltype);
1553                         }
1554 #endif
1555                         if (nmp->nm_flag & NFSMNT_NFSV3) {
1556                                 KKASSERT(*req->r_mrp == info->mrep);
1557                                 KKASSERT(*req->r_mdp == info->md);
1558                                 KKASSERT(*req->r_dposp == info->dpos);
1559                                 error |= NFSERR_RETERR;
1560                         } else {
1561                                 m_freem(info->mrep);
1562                                 info->mrep = NULL;
1563                         }
1564                         m_freem(req->r_mreq);
1565                         req->r_mreq = NULL;
1566                         kfree(req, M_NFSREQ);
1567                         info->req = NULL;
1568                         return (error);
1569                 }
1570
1571                 KKASSERT(*req->r_mrp == info->mrep);
1572                 KKASSERT(*req->r_mdp == info->md);
1573                 KKASSERT(*req->r_dposp == info->dpos);
1574                 m_freem(req->r_mreq);
1575                 req->r_mreq = NULL;
1576                 kfree(req, M_NFSREQ);
1577                 return (0);
1578         }
1579         m_freem(info->mrep);
1580         info->mrep = NULL;
1581         error = EPROTONOSUPPORT;
1582 nfsmout:
1583         m_freem(req->r_mreq);
1584         req->r_mreq = NULL;
1585         kfree(req, M_NFSREQ);
1586         info->req = NULL;
1587         return (error);
1588 }
1589
1590 #ifndef NFS_NOSERVER
1591 /*
1592  * Generate the rpc reply header
1593  * siz arg. is used to decide if adding a cluster is worthwhile
1594  */
1595 int
1596 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1597             int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1598 {
1599         u_int32_t *tl;
1600         struct nfsm_info info;
1601
1602         siz += RPC_REPLYSIZ;
1603         info.mb = m_getl(max_hdr + siz, M_WAITOK, MT_DATA, M_PKTHDR, NULL);
1604         info.mreq = info.mb;
1605         info.mreq->m_pkthdr.len = 0;
1606         /*
1607          * If this is not a cluster, try and leave leading space
1608          * for the lower level headers.
1609          */
1610         if ((max_hdr + siz) < MINCLSIZE)
1611                 info.mreq->m_data += max_hdr;
1612         tl = mtod(info.mreq, u_int32_t *);
1613         info.mreq->m_len = 6 * NFSX_UNSIGNED;
1614         info.bpos = ((caddr_t)tl) + info.mreq->m_len;
1615         *tl++ = txdr_unsigned(nd->nd_retxid);
1616         *tl++ = rpc_reply;
1617         if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1618                 *tl++ = rpc_msgdenied;
1619                 if (err & NFSERR_AUTHERR) {
1620                         *tl++ = rpc_autherr;
1621                         *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1622                         info.mreq->m_len -= NFSX_UNSIGNED;
1623                         info.bpos -= NFSX_UNSIGNED;
1624                 } else {
1625                         *tl++ = rpc_mismatch;
1626                         *tl++ = txdr_unsigned(RPC_VER2);
1627                         *tl = txdr_unsigned(RPC_VER2);
1628                 }
1629         } else {
1630                 *tl++ = rpc_msgaccepted;
1631
1632                 /*
1633                  * For Kerberos authentication, we must send the nickname
1634                  * verifier back, otherwise just RPCAUTH_NULL.
1635                  */
1636                 if (nd->nd_flag & ND_KERBFULL) {
1637                     struct nfsuid *nuidp;
1638                     struct timeval ktvout;
1639
1640                     for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1641                         nuidp != NULL; nuidp = nuidp->nu_hash.le_next) {
1642                         if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1643                             (!nd->nd_nam2 || netaddr_match(AF_INET,
1644                              &nuidp->nu_haddr, nd->nd_nam2)))
1645                             break;
1646                     }
1647                     if (nuidp) {
1648                         /*
1649                          * Encrypt the timestamp in ecb mode using the
1650                          * session key.
1651                          */
1652 #ifdef NFSKERB
1653                         XXX
1654 #else
1655                         ktvout.tv_sec = 0;
1656                         ktvout.tv_usec = 0;
1657 #endif
1658
1659                         *tl++ = rpc_auth_kerb;
1660                         *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1661                         *tl = ktvout.tv_sec;
1662                         tl = nfsm_build(&info, 3 * NFSX_UNSIGNED);
1663                         *tl++ = ktvout.tv_usec;
1664                         *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1665                     } else {
1666                         *tl++ = 0;
1667                         *tl++ = 0;
1668                     }
1669                 } else {
1670                         *tl++ = 0;
1671                         *tl++ = 0;
1672                 }
1673                 switch (err) {
1674                 case EPROGUNAVAIL:
1675                         *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1676                         break;
1677                 case EPROGMISMATCH:
1678                         *tl = txdr_unsigned(RPC_PROGMISMATCH);
1679                         tl = nfsm_build(&info, 2 * NFSX_UNSIGNED);
1680                         *tl++ = txdr_unsigned(2);
1681                         *tl = txdr_unsigned(3);
1682                         break;
1683                 case EPROCUNAVAIL:
1684                         *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1685                         break;
1686                 case EBADRPC:
1687                         *tl = txdr_unsigned(RPC_GARBAGE);
1688                         break;
1689                 default:
1690                         *tl = 0;
1691                         if (err != NFSERR_RETVOID) {
1692                                 tl = nfsm_build(&info, NFSX_UNSIGNED);
1693                                 if (err)
1694                                     *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1695                                 else
1696                                     *tl = 0;
1697                         }
1698                         break;
1699                 }
1700         }
1701
1702         if (mrq != NULL)
1703             *mrq = info.mreq;
1704         *mbp = info.mb;
1705         *bposp = info.bpos;
1706         if (err != 0 && err != NFSERR_RETVOID)
1707                 nfsstats.srvrpc_errs++;
1708         return (0);
1709 }
1710
1711
1712 #endif /* NFS_NOSERVER */
1713
1714 /*
1715  * Nfs timer routine.
1716  *
1717  * Scan the nfsreq list and retranmit any requests that have timed out
1718  * To avoid retransmission attempts on STREAM sockets (in the future) make
1719  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1720  *
1721  * Requests with attached responses, terminated requests, and
1722  * locked requests are ignored.  Locked requests will be picked up
1723  * in a later timer call.
1724  */
1725 void
1726 nfs_timer_callout(void *arg /* never used */)
1727 {
1728         struct nfsmount *nmp;
1729         struct nfsreq *req;
1730 #ifndef NFS_NOSERVER
1731         struct nfssvc_sock *slp;
1732         u_quad_t cur_usec;
1733 #endif /* NFS_NOSERVER */
1734
1735         lwkt_gettoken(&nfs_token);
1736         TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1737                 lwkt_gettoken(&nmp->nm_token);
1738                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1739                         KKASSERT(nmp == req->r_nmp);
1740                         if (req->r_mrep)
1741                                 continue;
1742                         if (req->r_flags & (R_SOFTTERM | R_LOCKED))
1743                                 continue;
1744
1745                         /*
1746                          * Handle timeout/retry.  Be sure to process r_mrep
1747                          * for async requests that completed while we had
1748                          * the request locked or they will hang in the reqq
1749                          * forever.
1750                          */
1751                         req->r_flags |= R_LOCKED;
1752                         if (nfs_sigintr(nmp, req, req->r_td)) {
1753                                 nfs_softterm(req, 1);
1754                                 req->r_flags &= ~R_LOCKED;
1755                         } else {
1756                                 nfs_timer_req(req);
1757                                 if (req->r_flags & R_ASYNC) {
1758                                         if (req->r_mrep)
1759                                                 nfs_hardterm(req, 1);
1760                                         req->r_flags &= ~R_LOCKED;
1761                                         nfssvc_iod_reader_wakeup(nmp);
1762                                 } else {
1763                                         req->r_flags &= ~R_LOCKED;
1764                                 }
1765                         }
1766                         if (req->r_flags & R_WANTED) {
1767                                 req->r_flags &= ~R_WANTED;
1768                                 wakeup(req);
1769                         }
1770                 }
1771                 lwkt_reltoken(&nmp->nm_token);
1772         }
1773 #ifndef NFS_NOSERVER
1774
1775         /*
1776          * Scan the write gathering queues for writes that need to be
1777          * completed now.
1778          */
1779         cur_usec = nfs_curusec();
1780
1781         TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1782                 /* XXX race against removal */
1783                 if (lwkt_trytoken(&slp->ns_token)) {
1784                         if (slp->ns_tq.lh_first &&
1785                             (slp->ns_tq.lh_first->nd_time <= cur_usec)) {
1786                                 nfsrv_wakenfsd(slp, 1);
1787                         }
1788                         lwkt_reltoken(&slp->ns_token);
1789                 }
1790         }
1791 #endif /* NFS_NOSERVER */
1792
1793         callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer_callout, NULL);
1794         lwkt_reltoken(&nfs_token);
1795 }
1796
1797 static
1798 void
1799 nfs_timer_req(struct nfsreq *req)
1800 {
1801         struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1802         struct nfsmount *nmp = req->r_nmp;
1803         struct mbuf *m;
1804         struct socket *so;
1805         int timeo;
1806         int error;
1807
1808         /*
1809          * rtt ticks and timeout calculation.  Return if the timeout
1810          * has not been reached yet, unless the packet is flagged
1811          * for an immediate send.
1812          *
1813          * The mean rtt doesn't help when we get random I/Os, we have
1814          * to multiply by fairly large numbers.
1815          */
1816         if (req->r_rtt >= 0) {
1817                 /*
1818                  * Calculate the timeout to test against.
1819                  */
1820                 req->r_rtt++;
1821                 if (nmp->nm_flag & NFSMNT_DUMBTIMR) {
1822                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1823                 } else if (req->r_flags & R_TIMING) {
1824                         timeo = NFS_SRTT(req) + NFS_SDRTT(req);
1825                 } else {
1826                         timeo = nmp->nm_timeo << NFS_RTT_SCALE_BITS;
1827                 }
1828                 timeo *= multt[req->r_procnum];
1829                 /* timeo is still scaled by SCALE_BITS */
1830
1831 #define NFSFS   (NFS_RTT_SCALE * NFS_HZ)
1832                 if (req->r_flags & R_TIMING) {
1833                         static long last_time;
1834                         if (nfs_showrtt && last_time != time_uptime) {
1835                                 kprintf("rpccmd %d NFS SRTT %d SDRTT %d "
1836                                         "timeo %d.%03d\n",
1837                                         proct[req->r_procnum],
1838                                         NFS_SRTT(req), NFS_SDRTT(req),
1839                                         timeo / NFSFS,
1840                                         timeo % NFSFS * 1000 /  NFSFS);
1841                                 last_time = time_uptime;
1842                         }
1843                 }
1844 #undef NFSFS
1845
1846                 /*
1847                  * deal with nfs_timer jitter.
1848                  */
1849                 timeo = (timeo >> NFS_RTT_SCALE_BITS) + 1;
1850                 if (timeo < 2)
1851                         timeo = 2;
1852
1853                 if (nmp->nm_timeouts > 0)
1854                         timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1855                 if (timeo > NFS_MAXTIMEO)
1856                         timeo = NFS_MAXTIMEO;
1857                 if (req->r_rtt <= timeo) {
1858                         if ((req->r_flags & R_NEEDSXMIT) == 0)
1859                                 return;
1860                 } else if (nmp->nm_timeouts < 8) {
1861                         nmp->nm_timeouts++;
1862                 }
1863         }
1864
1865         /*
1866          * Check for server not responding
1867          */
1868         if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1869              req->r_rexmit > nmp->nm_deadthresh) {
1870                 nfs_msg(req->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1871                         "not responding");
1872                 req->r_flags |= R_TPRINTFMSG;
1873         }
1874         if (req->r_rexmit >= req->r_retry) {    /* too many */
1875                 nfsstats.rpctimeouts++;
1876                 nfs_softterm(req, 1);
1877                 return;
1878         }
1879
1880         /*
1881          * Generally disable retransmission on reliable sockets,
1882          * unless the request is flagged for immediate send.
1883          */
1884         if (nmp->nm_sotype != SOCK_DGRAM) {
1885                 if (++req->r_rexmit > NFS_MAXREXMIT)
1886                         req->r_rexmit = NFS_MAXREXMIT;
1887                 if ((req->r_flags & R_NEEDSXMIT) == 0)
1888                         return;
1889         }
1890
1891         /*
1892          * Stop here if we do not have a socket!
1893          */
1894         if ((so = nmp->nm_so) == NULL)
1895                 return;
1896
1897         /*
1898          * If there is enough space and the window allows.. resend it.
1899          *
1900          * r_rtt is left intact in case we get an answer after the
1901          * retry that was a reply to the original packet.
1902          *
1903          * NOTE: so_pru_send()
1904          */
1905         if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1906             (req->r_flags & (R_SENT | R_NEEDSXMIT)) &&
1907            (m = m_copym(req->r_mreq, 0, M_COPYALL, M_NOWAIT))){
1908                 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1909                     error = so_pru_send(so, 0, m, NULL, NULL, td);
1910                 else
1911                     error = so_pru_send(so, 0, m, nmp->nm_nam, NULL, td);
1912                 if (error) {
1913                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1914                                 so->so_error = 0;
1915                         req->r_flags |= R_NEEDSXMIT;
1916                 } else if (req->r_mrep == NULL) {
1917                         /*
1918                          * Iff first send, start timing
1919                          * else turn timing off, backoff timer
1920                          * and divide congestion window by 2.
1921                          *
1922                          * It is possible for the so_pru_send() to
1923                          * block and for us to race a reply so we
1924                          * only do this if the reply field has not
1925                          * been filled in.  R_LOCKED will prevent
1926                          * the request from being ripped out from under
1927                          * us entirely.
1928                          *
1929                          * Record the last resent procnum to aid us
1930                          * in duplicate detection on receive.
1931                          */
1932                         if ((req->r_flags & R_NEEDSXMIT) == 0) {
1933                                 if (nfs_showrexmit)
1934                                         kprintf("X");
1935                                 if (++req->r_rexmit > NFS_MAXREXMIT)
1936                                         req->r_rexmit = NFS_MAXREXMIT;
1937                                 nmp->nm_maxasync_scaled >>= 1;
1938                                 if (nmp->nm_maxasync_scaled < NFS_MINASYNC_SCALED)
1939                                         nmp->nm_maxasync_scaled = NFS_MINASYNC_SCALED;
1940                                 nfsstats.rpcretries++;
1941                                 nmp->nm_lastreprocnum = req->r_procnum;
1942                         } else {
1943                                 req->r_flags |= R_SENT;
1944                                 req->r_flags &= ~R_NEEDSXMIT;
1945                         }
1946                 }
1947         }
1948 }
1949
1950 /*
1951  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1952  * wait for all requests to complete. This is used by forced unmounts
1953  * to terminate any outstanding RPCs.
1954  *
1955  * Locked requests cannot be canceled but will be marked for
1956  * soft-termination.
1957  */
1958 int
1959 nfs_nmcancelreqs(struct nfsmount *nmp)
1960 {
1961         struct nfsreq *req;
1962         int i;
1963
1964         crit_enter();
1965         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1966                 if (req->r_mrep != NULL || (req->r_flags & R_SOFTTERM))
1967                         continue;
1968                 nfs_softterm(req, 0);
1969         }
1970         /* XXX  the other two queues as well */
1971         crit_exit();
1972
1973         for (i = 0; i < 30; i++) {
1974                 crit_enter();
1975                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1976                         if (nmp == req->r_nmp)
1977                                 break;
1978                 }
1979                 crit_exit();
1980                 if (req == NULL)
1981                         return (0);
1982                 tsleep(&lbolt, 0, "nfscancel", 0);
1983         }
1984         return (EBUSY);
1985 }
1986
1987 /*
1988  * Soft-terminate a request, effectively marking it as failed.
1989  *
1990  * Must be called from within a critical section.
1991  */
1992 static void
1993 nfs_softterm(struct nfsreq *rep, int islocked)
1994 {
1995         rep->r_flags |= R_SOFTTERM;
1996         if (rep->r_info)
1997                 rep->r_info->error = EINTR;
1998         nfs_hardterm(rep, islocked);
1999 }
2000
2001 /*
2002  * Hard-terminate a request, typically after getting a response.
2003  *
2004  * The state machine can still decide to re-issue it later if necessary.
2005  *
2006  * Must be called from within a critical section.
2007  */
2008 static void
2009 nfs_hardterm(struct nfsreq *rep, int islocked)
2010 {
2011         struct nfsmount *nmp = rep->r_nmp;
2012
2013         /*
2014          * The nm_send count is decremented now to avoid deadlocks
2015          * when the process in soreceive() hasn't yet managed to send
2016          * its own request.
2017          */
2018         if (rep->r_flags & R_SENT) {
2019                 rep->r_flags &= ~R_SENT;
2020         }
2021
2022         /*
2023          * If we locked the request or nobody else has locked the request,
2024          * and the request is async, we can move it to the reader thread's
2025          * queue now and fix up the state.
2026          *
2027          * If we locked the request or nobody else has locked the request,
2028          * we can wake up anyone blocked waiting for a response on the
2029          * request.
2030          */
2031         if (islocked || (rep->r_flags & R_LOCKED) == 0) {
2032                 if ((rep->r_flags & (R_ONREQQ | R_ASYNC)) ==
2033                     (R_ONREQQ | R_ASYNC)) {
2034                         rep->r_flags &= ~R_ONREQQ;
2035                         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
2036                         --nmp->nm_reqqlen;
2037                         TAILQ_INSERT_TAIL(&nmp->nm_reqrxq, rep, r_chain);
2038                         KKASSERT(rep->r_info->state == NFSM_STATE_TRY ||
2039                                  rep->r_info->state == NFSM_STATE_WAITREPLY);
2040
2041                         /*
2042                          * When setting the state to PROCESSREPLY we must
2043                          * roll-up any error not related to the contents of
2044                          * the reply (i.e. if there is no contents).
2045                          */
2046                         rep->r_info->state = NFSM_STATE_PROCESSREPLY;
2047                         nfssvc_iod_reader_wakeup(nmp);
2048                         if (TAILQ_FIRST(&nmp->nm_bioq) &&
2049                             nmp->nm_reqqlen <= nfs_maxasyncbio * 2 / 3) {
2050                                 nfssvc_iod_writer_wakeup(nmp);
2051                         }
2052                 }
2053                 mtx_abort_link(&nmp->nm_rxlock, &rep->r_link);
2054         }
2055 }
2056
2057 /*
2058  * Test for a termination condition pending on the process.
2059  * This is used for NFSMNT_INT mounts.
2060  */
2061 int
2062 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
2063 {
2064         sigset_t tmpset;
2065         struct proc *p;
2066         struct lwp *lp;
2067
2068         if (rep && (rep->r_flags & R_SOFTTERM))
2069                 return (EINTR);
2070         /* Terminate all requests while attempting a forced unmount. */
2071         if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
2072                 return (EINTR);
2073         if (!(nmp->nm_flag & NFSMNT_INT))
2074                 return (0);
2075         /* td might be NULL YYY */
2076         if (td == NULL || (p = td->td_proc) == NULL)
2077                 return (0);
2078
2079         lp = td->td_lwp;
2080         tmpset = lwp_sigpend(lp);
2081         SIGSETNAND(tmpset, lp->lwp_sigmask);
2082         SIGSETNAND(tmpset, p->p_sigignore);
2083         if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
2084                 return (EINTR);
2085
2086         return (0);
2087 }
2088
2089 /*
2090  * Lock a socket against others.
2091  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
2092  * and also to avoid race conditions between the processes with nfs requests
2093  * in progress when a reconnect is necessary.
2094  */
2095 int
2096 nfs_sndlock(struct nfsmount *nmp, struct nfsreq *rep)
2097 {
2098         mtx_t *mtx = &nmp->nm_txlock;
2099         struct thread *td;
2100         int slptimeo;
2101         int slpflag;
2102         int error;
2103
2104         slpflag = 0;
2105         slptimeo = 0;
2106         td = rep ? rep->r_td : NULL;
2107         if (nmp->nm_flag & NFSMNT_INT)
2108                 slpflag = PCATCH;
2109
2110         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2111                 if (nfs_sigintr(nmp, rep, td)) {
2112                         error = EINTR;
2113                         break;
2114                 }
2115                 error = mtx_lock_ex(mtx, slpflag, slptimeo);
2116                 if (error == 0)
2117                         break;
2118                 if (slpflag == PCATCH) {
2119                         slpflag = 0;
2120                         slptimeo = 2 * hz;
2121                 }
2122         }
2123         /* Always fail if our request has been cancelled. */
2124         if (rep && (rep->r_flags & R_SOFTTERM)) {
2125                 if (error == 0)
2126                         mtx_unlock(mtx);
2127                 error = EINTR;
2128         }
2129         return (error);
2130 }
2131
2132 /*
2133  * Unlock the stream socket for others.
2134  */
2135 void
2136 nfs_sndunlock(struct nfsmount *nmp)
2137 {
2138         mtx_unlock(&nmp->nm_txlock);
2139 }
2140
2141 /*
2142  * Lock the receiver side of the socket.
2143  *
2144  * rep may be NULL.
2145  */
2146 static int
2147 nfs_rcvlock(struct nfsmount *nmp, struct nfsreq *rep)
2148 {
2149         mtx_t *mtx = &nmp->nm_rxlock;
2150         int slpflag;
2151         int slptimeo;
2152         int error;
2153
2154         /*
2155          * Unconditionally check for completion in case another nfsiod
2156          * get the packet while the caller was blocked, before the caller
2157          * called us.  Packet reception is handled by mainline code which
2158          * is protected by the BGL at the moment.
2159          *
2160          * We do not strictly need the second check just before the
2161          * tsleep(), but it's good defensive programming.
2162          */
2163         if (rep && rep->r_mrep != NULL)
2164                 return (EALREADY);
2165
2166         if (nmp->nm_flag & NFSMNT_INT)
2167                 slpflag = PCATCH;
2168         else
2169                 slpflag = 0;
2170         slptimeo = 0;
2171
2172         while ((error = mtx_lock_ex_try(mtx)) != 0) {
2173                 if (nfs_sigintr(nmp, rep, (rep ? rep->r_td : NULL))) {
2174                         error = EINTR;
2175                         break;
2176                 }
2177                 if (rep && rep->r_mrep != NULL) {
2178                         error = EALREADY;
2179                         break;
2180                 }
2181
2182                 /*
2183                  * NOTE: can return ENOLCK, but in that case rep->r_mrep
2184                  *       will already be set.
2185                  */
2186                 if (rep) {
2187                         error = mtx_lock_ex_link(mtx, &rep->r_link,
2188                                                  slpflag, slptimeo);
2189                 } else {
2190                         error = mtx_lock_ex(mtx, slpflag, slptimeo);
2191                 }
2192                 if (error == 0)
2193                         break;
2194
2195                 /*
2196                  * If our reply was recieved while we were sleeping,
2197                  * then just return without taking the lock to avoid a
2198                  * situation where a single iod could 'capture' the
2199                  * recieve lock.
2200                  */
2201                 if (rep && rep->r_mrep != NULL) {
2202                         error = EALREADY;
2203                         break;
2204                 }
2205                 if (slpflag == PCATCH) {
2206                         slpflag = 0;
2207                         slptimeo = 2 * hz;
2208                 }
2209         }
2210         if (error == 0) {
2211                 if (rep && rep->r_mrep != NULL) {
2212                         error = EALREADY;
2213                         mtx_unlock(mtx);
2214                 }
2215         }
2216         return (error);
2217 }
2218
2219 /*
2220  * Unlock the stream socket for others.
2221  */
2222 static void
2223 nfs_rcvunlock(struct nfsmount *nmp)
2224 {
2225         mtx_unlock(&nmp->nm_rxlock);
2226 }
2227
2228 /*
2229  * nfs_realign:
2230  *
2231  * Check for badly aligned mbuf data and realign by copying the unaligned
2232  * portion of the data into a new mbuf chain and freeing the portions
2233  * of the old chain that were replaced.
2234  *
2235  * We cannot simply realign the data within the existing mbuf chain
2236  * because the underlying buffers may contain other rpc commands and
2237  * we cannot afford to overwrite them.
2238  *
2239  * We would prefer to avoid this situation entirely.  The situation does
2240  * not occur with NFS/UDP and is supposed to only occassionally occur
2241  * with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
2242  *
2243  * NOTE!  M_NOWAIT cannot be used here.  The mbufs must be acquired
2244  *        because the rpc request OR reply cannot be thrown away.  TCP NFS
2245  *        mounts do not retry their RPCs unless the TCP connection itself
2246  *        is dropped so throwing away a RPC will basically cause the NFS
2247  *        operation to lockup indefinitely.
2248  */
2249 static void
2250 nfs_realign(struct mbuf **pm, int hsiz)
2251 {
2252         struct mbuf *m;
2253         struct mbuf *n = NULL;
2254
2255         /*
2256          * Check for misalignemnt
2257          */
2258         ++nfs_realign_test;
2259         while ((m = *pm) != NULL) {
2260                 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3))
2261                         break;
2262                 pm = &m->m_next;
2263         }
2264
2265         /*
2266          * If misalignment found make a completely new copy.
2267          */
2268         if (m) {
2269                 ++nfs_realign_count;
2270                 n = m_dup_data(m, M_WAITOK);
2271                 m_freem(*pm);
2272                 *pm = n;
2273         }
2274 }
2275
2276 #ifndef NFS_NOSERVER
2277
2278 /*
2279  * Parse an RPC request
2280  * - verify it
2281  * - fill in the cred struct.
2282  */
2283 int
2284 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
2285 {
2286         int len, i;
2287         u_int32_t *tl;
2288         struct uio uio;
2289         struct iovec iov;
2290         caddr_t cp;
2291         u_int32_t nfsvers, auth_type;
2292         uid_t nickuid;
2293         int error = 0, ticklen;
2294         struct nfsuid *nuidp;
2295         struct timeval tvin, tvout;
2296         struct nfsm_info info;
2297 #if 0                           /* until encrypted keys are implemented */
2298         NFSKERBKEYSCHED_T keys; /* stores key schedule */
2299 #endif
2300
2301         info.mrep = nd->nd_mrep;
2302         info.md = nd->nd_md;
2303         info.dpos = nd->nd_dpos;
2304
2305         if (has_header) {
2306                 NULLOUT(tl = nfsm_dissect(&info, 10 * NFSX_UNSIGNED));
2307                 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
2308                 if (*tl++ != rpc_call) {
2309                         m_freem(info.mrep);
2310                         return (EBADRPC);
2311                 }
2312         } else {
2313                 NULLOUT(tl = nfsm_dissect(&info, 8 * NFSX_UNSIGNED));
2314         }
2315         nd->nd_repstat = 0;
2316         nd->nd_flag = 0;
2317         if (*tl++ != rpc_vers) {
2318                 nd->nd_repstat = ERPCMISMATCH;
2319                 nd->nd_procnum = NFSPROC_NOOP;
2320                 return (0);
2321         }
2322         if (*tl != nfs_prog) {
2323                 nd->nd_repstat = EPROGUNAVAIL;
2324                 nd->nd_procnum = NFSPROC_NOOP;
2325                 return (0);
2326         }
2327         tl++;
2328         nfsvers = fxdr_unsigned(u_int32_t, *tl++);
2329         if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
2330                 nd->nd_repstat = EPROGMISMATCH;
2331                 nd->nd_procnum = NFSPROC_NOOP;
2332                 return (0);
2333         }
2334         if (nfsvers == NFS_VER3)
2335                 nd->nd_flag = ND_NFSV3;
2336         nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
2337         if (nd->nd_procnum == NFSPROC_NULL)
2338                 return (0);
2339         if (nd->nd_procnum >= NFS_NPROCS ||
2340                 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
2341                 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
2342                 nd->nd_repstat = EPROCUNAVAIL;
2343                 nd->nd_procnum = NFSPROC_NOOP;
2344                 return (0);
2345         }
2346         if ((nd->nd_flag & ND_NFSV3) == 0)
2347                 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
2348         auth_type = *tl++;
2349         len = fxdr_unsigned(int, *tl++);
2350         if (len < 0 || len > RPCAUTH_MAXSIZ) {
2351                 m_freem(info.mrep);
2352                 return (EBADRPC);
2353         }
2354
2355         nd->nd_flag &= ~ND_KERBAUTH;
2356         /*
2357          * Handle auth_unix or auth_kerb.
2358          */
2359         if (auth_type == rpc_auth_unix) {
2360                 len = fxdr_unsigned(int, *++tl);
2361                 if (len < 0 || len > NFS_MAXNAMLEN) {
2362                         m_freem(info.mrep);
2363                         return (EBADRPC);
2364                 }
2365                 ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2366                 NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2367                 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
2368                 nd->nd_cr.cr_ref = 1;
2369                 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
2370                 nd->nd_cr.cr_ruid = nd->nd_cr.cr_svuid = nd->nd_cr.cr_uid;
2371                 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
2372                 nd->nd_cr.cr_rgid = nd->nd_cr.cr_svgid = nd->nd_cr.cr_gid;
2373                 len = fxdr_unsigned(int, *tl);
2374                 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
2375                         m_freem(info.mrep);
2376                         return (EBADRPC);
2377                 }
2378                 NULLOUT(tl = nfsm_dissect(&info, (len + 2) * NFSX_UNSIGNED));
2379                 for (i = 1; i <= len; i++)
2380                     if (i < NGROUPS)
2381                         nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
2382                     else
2383                         tl++;
2384                 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
2385                 if (nd->nd_cr.cr_ngroups > 1)
2386                     nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
2387                 len = fxdr_unsigned(int, *++tl);
2388                 if (len < 0 || len > RPCAUTH_MAXSIZ) {
2389                         m_freem(info.mrep);
2390                         return (EBADRPC);
2391                 }
2392                 if (len > 0) {
2393                         ERROROUT(nfsm_adv(&info, nfsm_rndup(len)));
2394                 }
2395         } else if (auth_type == rpc_auth_kerb) {
2396                 switch (fxdr_unsigned(int, *tl++)) {
2397                 case RPCAKN_FULLNAME:
2398                         ticklen = fxdr_unsigned(int, *tl);
2399                         *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
2400                         uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
2401                         nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
2402                         if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
2403                                 m_freem(info.mrep);
2404                                 return (EBADRPC);
2405                         }
2406                         uio.uio_offset = 0;
2407                         uio.uio_iov = &iov;
2408                         uio.uio_iovcnt = 1;
2409                         uio.uio_segflg = UIO_SYSSPACE;
2410                         iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2411                         iov.iov_len = RPCAUTH_MAXSIZ - 4;
2412                         ERROROUT(nfsm_mtouio(&info, &uio, uio.uio_resid));
2413                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2414                         if (*tl++ != rpc_auth_kerb ||
2415                                 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2416                                 kprintf("Bad kerb verifier\n");
2417                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2418                                 nd->nd_procnum = NFSPROC_NOOP;
2419                                 return (0);
2420                         }
2421                         NULLOUT(cp = nfsm_dissect(&info, 4 * NFSX_UNSIGNED));
2422                         tl = (u_int32_t *)cp;
2423                         if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2424                                 kprintf("Not fullname kerb verifier\n");
2425                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2426                                 nd->nd_procnum = NFSPROC_NOOP;
2427                                 return (0);
2428                         }
2429                         cp += NFSX_UNSIGNED;
2430                         bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2431                         nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2432                         nd->nd_flag |= ND_KERBFULL;
2433                         nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2434                         break;
2435                 case RPCAKN_NICKNAME:
2436                         if (len != 2 * NFSX_UNSIGNED) {
2437                                 kprintf("Kerb nickname short\n");
2438                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2439                                 nd->nd_procnum = NFSPROC_NOOP;
2440                                 return (0);
2441                         }
2442                         nickuid = fxdr_unsigned(uid_t, *tl);
2443                         NULLOUT(tl = nfsm_dissect(&info, 2 * NFSX_UNSIGNED));
2444                         if (*tl++ != rpc_auth_kerb ||
2445                                 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2446                                 kprintf("Kerb nick verifier bad\n");
2447                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2448                                 nd->nd_procnum = NFSPROC_NOOP;
2449                                 return (0);
2450                         }
2451                         NULLOUT(tl = nfsm_dissect(&info, 3 * NFSX_UNSIGNED));
2452                         tvin.tv_sec = *tl++;
2453                         tvin.tv_usec = *tl;
2454
2455                         for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2456                             nuidp != NULL; nuidp = nuidp->nu_hash.le_next) {
2457                                 if (nuidp->nu_cr.cr_uid == nickuid &&
2458                                     (!nd->nd_nam2 ||
2459                                      netaddr_match(AF_INET,
2460                                       &nuidp->nu_haddr, nd->nd_nam2)))
2461                                         break;
2462                         }
2463                         if (!nuidp) {
2464                                 nd->nd_repstat =
2465                                         (NFSERR_AUTHERR|AUTH_REJECTCRED);
2466                                 nd->nd_procnum = NFSPROC_NOOP;
2467                                 return (0);
2468                         }
2469
2470                         /*
2471                          * Now, decrypt the timestamp using the session key
2472                          * and validate it.
2473                          */
2474 #ifdef NFSKERB
2475                         XXX
2476 #else
2477                         tvout.tv_sec = 0;
2478                         tvout.tv_usec = 0;
2479 #endif
2480
2481                         tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2482                         tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2483                         if (nuidp->nu_expire != time_uptime ||
2484                             nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2485                             (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2486                              nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2487                                 nuidp->nu_expire = 0;
2488                                 nd->nd_repstat =
2489                                     (NFSERR_AUTHERR|AUTH_REJECTVERF);
2490                                 nd->nd_procnum = NFSPROC_NOOP;
2491                                 return (0);
2492                         }
2493                         nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2494                         nd->nd_flag |= ND_KERBNICK;
2495                         break;
2496                 }
2497         } else {
2498                 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2499                 nd->nd_procnum = NFSPROC_NOOP;
2500                 return (0);
2501         }
2502
2503         nd->nd_md = info.md;
2504         nd->nd_dpos = info.dpos;
2505         return (0);
2506 nfsmout:
2507         return (error);
2508 }
2509
2510 #endif
2511
2512 /*
2513  * Send a message to the originating process's terminal.  The thread and/or
2514  * process may be NULL.  YYY the thread should not be NULL but there may
2515  * still be some uio_td's that are still being passed as NULL through to
2516  * nfsm_request().
2517  */
2518 static int
2519 nfs_msg(struct thread *td, char *server, char *msg)
2520 {
2521         tpr_t tpr;
2522
2523         if (td && td->td_proc)
2524                 tpr = tprintf_open(td->td_proc);
2525         else
2526                 tpr = NULL;
2527         tprintf(tpr, "nfs server %s: %s\n", server, msg);
2528         tprintf_close(tpr);
2529         return (0);
2530 }
2531
2532 #ifndef NFS_NOSERVER
2533
2534 /*
2535  * Socket upcall routine for nfsd sockets.  This runs in the protocol
2536  * thread and passes waitflag == M_NOWAIT.
2537  */
2538 void
2539 nfsrv_rcv_upcall(struct socket *so, void *arg, int waitflag)
2540 {
2541         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2542
2543         if (slp->ns_needq_upcall == 0) {
2544                 slp->ns_needq_upcall = 1;       /* ok to race */
2545                 lwkt_gettoken(&nfs_token);
2546                 nfsrv_wakenfsd(slp, 1);
2547                 lwkt_reltoken(&nfs_token);
2548         }
2549 #if 0
2550         lwkt_gettoken(&slp->ns_token);
2551         slp->ns_flag |= SLP_NEEDQ;
2552         nfsrv_rcv(so, arg, waitflag);
2553         lwkt_reltoken(&slp->ns_token);
2554 #endif
2555 }
2556
2557 /*
2558  * Process new data on a receive socket.  Essentially do as much as we can
2559  * non-blocking, else punt and it will be called with M_WAITOK from an nfsd.
2560  *
2561  * slp->ns_token is held on call
2562  */
2563 void
2564 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2565 {
2566         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2567         struct mbuf *m;
2568         struct sockaddr *nam;
2569         struct sockbuf sio;
2570         int flags, error;
2571         int nparallel_wakeup = 0;
2572
2573         ASSERT_LWKT_TOKEN_HELD(&slp->ns_token);
2574
2575         if ((slp->ns_flag & SLP_VALID) == 0)
2576                 return;
2577
2578         /*
2579          * Do not allow an infinite number of completed RPC records to build
2580          * up before we stop reading data from the socket.  Otherwise we could
2581          * end up holding onto an unreasonable number of mbufs for requests
2582          * waiting for service.
2583          *
2584          * This should give pretty good feedback to the TCP layer and
2585          * prevents a memory crunch for other protocols.
2586          *
2587          * Note that the same service socket can be dispatched to several
2588          * nfs servers simultaniously.  The tcp protocol callback calls us
2589          * with M_NOWAIT.  nfsd calls us with M_WAITOK (typically).
2590          */
2591         if (NFSRV_RECLIMIT(slp))
2592                 return;
2593
2594         /*
2595          * Handle protocol specifics to parse an RPC request.  We always
2596          * pull from the socket using non-blocking I/O.
2597          */
2598         if (so->so_type == SOCK_STREAM) {
2599                 /*
2600                  * The data has to be read in an orderly fashion from a TCP
2601                  * stream, unlike a UDP socket.  It is possible for soreceive
2602                  * and/or nfsrv_getstream() to block, so make sure only one
2603                  * entity is messing around with the TCP stream at any given
2604                  * moment.  The receive sockbuf's lock in soreceive is not
2605                  * sufficient.
2606                  */
2607                 if (slp->ns_flag & SLP_GETSTREAM)
2608                         return;
2609                 slp->ns_flag |= SLP_GETSTREAM;
2610
2611                 /*
2612                  * Do soreceive().  Pull out as much data as possible without
2613                  * blocking.
2614                  */
2615                 sbinit(&sio, 1000000000);
2616                 flags = MSG_DONTWAIT;
2617                 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2618                 if (error || sio.sb_mb == NULL) {
2619                         if (error != EWOULDBLOCK)
2620                                 slp->ns_flag |= SLP_DISCONN;
2621                         slp->ns_flag &= ~(SLP_GETSTREAM | SLP_NEEDQ);
2622                         goto done;
2623                 }
2624                 m = sio.sb_mb;
2625                 if (slp->ns_rawend) {
2626                         slp->ns_rawend->m_next = m;
2627                         slp->ns_cc += sio.sb_cc;
2628                 } else {
2629                         slp->ns_raw = m;
2630                         slp->ns_cc = sio.sb_cc;
2631                 }
2632                 while (m->m_next)
2633                         m = m->m_next;
2634                 slp->ns_rawend = m;
2635
2636                 /*
2637                  * Now try and parse as many record(s) as we can out of the
2638                  * raw stream data.  This will set SLP_DOREC.
2639                  */
2640                 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2641                 if (error && error != EWOULDBLOCK)
2642                         slp->ns_flag |= SLP_DISCONN;
2643                 slp->ns_flag &= ~SLP_GETSTREAM;
2644         } else {
2645                 /*
2646                  * For UDP soreceive typically pulls just one packet, loop
2647                  * to get the whole batch.
2648                  */
2649                 do {
2650                         sbinit(&sio, 1000000000);
2651                         flags = MSG_DONTWAIT;
2652                         error = so_pru_soreceive(so, &nam, NULL, &sio,
2653                                                  NULL, &flags);
2654                         if (sio.sb_mb) {
2655                                 struct nfsrv_rec *rec;
2656                                 int mf = (waitflag & M_NOWAIT) ?
2657                                             M_NOWAIT : M_WAITOK;
2658                                 rec = kmalloc(sizeof(struct nfsrv_rec),
2659                                              M_NFSRVDESC, mf);
2660                                 if (!rec) {
2661                                         if (nam)
2662                                                 kfree(nam, M_SONAME);
2663                                         m_freem(sio.sb_mb);
2664                                         continue;
2665                                 }
2666                                 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2667                                 rec->nr_address = nam;
2668                                 rec->nr_packet = sio.sb_mb;
2669                                 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2670                                 ++slp->ns_numrec;
2671                                 slp->ns_flag |= SLP_DOREC;
2672                                 ++nparallel_wakeup;
2673                         } else {
2674                                 slp->ns_flag &= ~SLP_NEEDQ;
2675                         }
2676                         if (error) {
2677                                 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2678                                     && error != EWOULDBLOCK) {
2679                                         slp->ns_flag |= SLP_DISCONN;
2680                                         break;
2681                                 }
2682                         }
2683                         if (NFSRV_RECLIMIT(slp))
2684                                 break;
2685                 } while (sio.sb_mb);
2686         }
2687
2688         /*
2689          * If we were upcalled from the tcp protocol layer and we have
2690          * fully parsed records ready to go, or there is new data pending,
2691          * or something went wrong, try to wake up a nfsd thread to deal
2692          * with it.
2693          */
2694 done:
2695         /* XXX this code is currently not executed (nfsrv_rcv_upcall) */
2696         if (waitflag == M_NOWAIT && (slp->ns_flag & SLP_ACTION_MASK)) {
2697                 lwkt_gettoken(&nfs_token);
2698                 nfsrv_wakenfsd(slp, nparallel_wakeup);
2699                 lwkt_reltoken(&nfs_token);
2700         }
2701 }
2702
2703 /*
2704  * Try and extract an RPC request from the mbuf data list received on a
2705  * stream socket. The "waitflag" argument indicates whether or not it
2706  * can sleep.
2707  */
2708 static int
2709 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2710 {
2711         struct mbuf *m, **mpp;
2712         char *cp1, *cp2;
2713         int len;
2714         struct mbuf *om, *m2, *recm;
2715         u_int32_t recmark;
2716
2717         for (;;) {
2718             if (slp->ns_reclen == 0) {
2719                 if (slp->ns_cc < NFSX_UNSIGNED)
2720                         return (0);
2721                 m = slp->ns_raw;
2722                 if (m->m_len >= NFSX_UNSIGNED) {
2723                         bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2724                         m->m_data += NFSX_UNSIGNED;
2725                         m->m_len -= NFSX_UNSIGNED;
2726                 } else {
2727                         cp1 = (caddr_t)&recmark;
2728                         cp2 = mtod(m, caddr_t);
2729                         while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2730                                 while (m->m_len == 0) {
2731                                         m = m->m_next;
2732                                         cp2 = mtod(m, caddr_t);
2733                                 }
2734                                 *cp1++ = *cp2++;
2735                                 m->m_data++;
2736                                 m->m_len--;
2737                         }
2738                 }
2739                 slp->ns_cc -= NFSX_UNSIGNED;
2740                 recmark = ntohl(recmark);
2741                 slp->ns_reclen = recmark & ~0x80000000;
2742                 if (recmark & 0x80000000)
2743                         slp->ns_flag |= SLP_LASTFRAG;
2744                 else
2745                         slp->ns_flag &= ~SLP_LASTFRAG;
2746                 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2747                         log(LOG_ERR, "%s (%d) from nfs client\n",
2748                             "impossible packet length",
2749                             slp->ns_reclen);
2750                         return (EPERM);
2751                 }
2752             }
2753
2754             /*
2755              * Now get the record part.
2756              *
2757              * Note that slp->ns_reclen may be 0.  Linux sometimes
2758              * generates 0-length RPCs
2759              */
2760             recm = NULL;
2761             if (slp->ns_cc == slp->ns_reclen) {
2762                 recm = slp->ns_raw;
2763                 slp->ns_raw = slp->ns_rawend = NULL;
2764                 slp->ns_cc = slp->ns_reclen = 0;
2765             } else if (slp->ns_cc > slp->ns_reclen) {
2766                 len = 0;
2767                 m = slp->ns_raw;
2768                 om = NULL;
2769
2770                 while (len < slp->ns_reclen) {
2771                         if ((len + m->m_len) > slp->ns_reclen) {
2772                                 m2 = m_copym(m, 0, slp->ns_reclen - len,
2773                                         waitflag);
2774                                 if (m2) {
2775                                         if (om) {
2776                                                 om->m_next = m2;
2777                                                 recm = slp->ns_raw;
2778                                         } else
2779                                                 recm = m2;
2780                                         m->m_data += slp->ns_reclen - len;
2781                                         m->m_len -= slp->ns_reclen - len;
2782                                         len = slp->ns_reclen;
2783                                 } else {
2784                                         return (EWOULDBLOCK);
2785                                 }
2786                         } else if ((len + m->m_len) == slp->ns_reclen) {
2787                                 om = m;
2788                                 len += m->m_len;
2789                                 m = m->m_next;
2790                                 recm = slp->ns_raw;
2791                                 om->m_next = NULL;
2792                         } else {
2793                                 om = m;
2794                                 len += m->m_len;
2795                                 m = m->m_next;
2796                         }
2797                 }
2798                 slp->ns_raw = m;
2799                 slp->ns_cc -= len;
2800                 slp->ns_reclen = 0;
2801             } else {
2802                 return (0);
2803             }
2804
2805             /*
2806              * Accumulate the fragments into a record.
2807              */
2808             mpp = &slp->ns_frag;
2809             while (*mpp)
2810                 mpp = &((*mpp)->m_next);
2811             *mpp = recm;
2812             if (slp->ns_flag & SLP_LASTFRAG) {
2813                 struct nfsrv_rec *rec;
2814                 int mf = (waitflag & M_NOWAIT) ? M_NOWAIT : M_WAITOK;
2815                 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2816                 if (!rec) {
2817                     m_freem(slp->ns_frag);
2818                 } else {
2819                     nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2820                     rec->nr_address = NULL;
2821                     rec->nr_packet = slp->ns_frag;
2822                     STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2823                     ++slp->ns_numrec;
2824                     slp->ns_flag |= SLP_DOREC;
2825                     ++*countp;
2826                 }
2827                 slp->ns_frag = NULL;
2828             }
2829         }
2830 }
2831
2832 #ifdef INVARIANTS
2833
2834 /*
2835  * Sanity check our mbuf chain.
2836  */
2837 static void
2838 nfs_checkpkt(struct mbuf *m, int len)
2839 {
2840         int xlen = 0;
2841         while (m) {
2842                 xlen += m->m_len;
2843                 m = m->m_next;
2844         }
2845         if (xlen != len) {
2846                 panic("nfs_checkpkt: len mismatch %d/%d mbuf %p",
2847                         xlen, len, m);
2848         }
2849 }
2850
2851 #else
2852
2853 static void
2854 nfs_checkpkt(struct mbuf *m __unused, int len __unused)
2855 {
2856 }
2857
2858 #endif
2859
2860 /*
2861  * Parse an RPC header.
2862  *
2863  * If the socket is invalid or no records are pending we return ENOBUFS.
2864  * The caller must deal with NEEDQ races.
2865  */
2866 int
2867 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2868             struct nfsrv_descript **ndp)
2869 {
2870         struct nfsrv_rec *rec;
2871         struct mbuf *m;
2872         struct sockaddr *nam;
2873         struct nfsrv_descript *nd;
2874         int error;
2875
2876         *ndp = NULL;
2877         if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2878                 return (ENOBUFS);
2879         rec = STAILQ_FIRST(&slp->ns_rec);
2880         STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2881         KKASSERT(slp->ns_numrec > 0);
2882         if (--slp->ns_numrec == 0)
2883                 slp->ns_flag &= ~SLP_DOREC;
2884         nam = rec->nr_address;
2885         m = rec->nr_packet;
2886         kfree(rec, M_NFSRVDESC);
2887         nd = kmalloc(sizeof(struct nfsrv_descript), M_NFSRVDESC, M_WAITOK);
2888         nd->nd_md = nd->nd_mrep = m;
2889         nd->nd_nam2 = nam;
2890         nd->nd_dpos = mtod(m, caddr_t);
2891         error = nfs_getreq(nd, nfsd, TRUE);
2892         if (error) {
2893                 if (nam) {
2894                         kfree(nam, M_SONAME);
2895                 }
2896                 kfree((caddr_t)nd, M_NFSRVDESC);
2897                 return (error);
2898         }
2899         *ndp = nd;
2900         nfsd->nfsd_nd = nd;
2901         return (0);
2902 }
2903
2904 /*
2905  * Try to assign service sockets to nfsd threads based on the number
2906  * of new rpc requests that have been queued on the service socket.
2907  *
2908  * If no nfsd's are available or additonal requests are pending, set the
2909  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2910  * the work in the nfssvc_sock list when it is finished processing its
2911  * current work.  This flag is only cleared when an nfsd can not find
2912  * any new work to perform.
2913  */
2914 void
2915 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2916 {
2917         struct nfsd *nd;
2918
2919         if ((slp->ns_flag & SLP_VALID) == 0)
2920                 return;
2921         if (nparallel <= 1)
2922                 nparallel = 1;
2923         TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2924                 if (nd->nfsd_flag & NFSD_WAITING) {
2925                         nd->nfsd_flag &= ~NFSD_WAITING;
2926                         if (nd->nfsd_slp)
2927                                 panic("nfsd wakeup");
2928                         nfsrv_slpref(slp);
2929                         nd->nfsd_slp = slp;
2930                         wakeup((caddr_t)nd);
2931                         if (--nparallel == 0)
2932                                 break;
2933                 }
2934         }
2935
2936         /*
2937          * If we couldn't assign slp then the NFSDs are all busy and
2938          * we set a flag indicating that there is pending work.
2939          */
2940         if (nparallel)
2941                 nfsd_head_flag |= NFSD_CHECKSLP;
2942 }
2943 #endif /* NFS_NOSERVER */