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