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