sys/vfs/nfs/nfs_socket.c

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