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  * Estimate rto for an nfs rpc sent via. an unreliable datagram.
  87  * Use the mean and mean deviation of rtt for the appropriate type of rpc
  88  * for the frequent rpcs and a default for the others.
  89  * The justification for doing "other" this way is that these rpcs
  90  * happen so infrequently that timer est. would probably be stale.
  91  * Also, since many of these rpcs are
  92  * non-idempotent, a conservative timeout is desired.
  93  * getattr, lookup - A+2D
  94  * read, write     - A+4D
  95  * other           - nm_timeo
  96  */
  97 #define NFS_RTO(n, t) \
  98         ((t) == 0 ? (n)->nm_timeo : \
  99          ((t) < 3 ? \
 100           (((((n)->nm_srtt[t-1] + 3) >> 2) + (n)->nm_sdrtt[t-1] + 1) >> 1) : \
 101           ((((n)->nm_srtt[t-1] + 7) >> 3) + (n)->nm_sdrtt[t-1] + 1)))
 102 #define NFS_SRTT(r)     (r)->r_nmp->nm_srtt[proct[(r)->r_procnum] - 1]
 103 #define NFS_SDRTT(r)    (r)->r_nmp->nm_sdrtt[proct[(r)->r_procnum] - 1]
 104
 105 /*
 106  * Defines which timer to use for the procnum.
 107  * 0 - default
 108  * 1 - getattr
 109  * 2 - lookup
 110  * 3 - read
 111  * 4 - write
 112  */
 113 static int proct[NFS_NPROCS] = {
 114         0, 1, 0, 2, 1, 3, 3, 4, 0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 0, 0, 0, 0, 0,
 115         0, 0, 0,
 116 };
 117
 118 static int nfs_realign_test;
 119 static int nfs_realign_count;
 120 static int nfs_bufpackets = 4;
 121 static int nfs_timer_raced;
 122
 123 SYSCTL_DECL(_vfs_nfs);
 124
 125 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_test, CTLFLAG_RW, &nfs_realign_test, 0, "");
 126 SYSCTL_INT(_vfs_nfs, OID_AUTO, realign_count, CTLFLAG_RW, &nfs_realign_count, 0, "");
 127 SYSCTL_INT(_vfs_nfs, OID_AUTO, bufpackets, CTLFLAG_RW, &nfs_bufpackets, 0, "");
 128
 129
 130 /*
 131  * There is a congestion window for outstanding rpcs maintained per mount
 132  * point. The cwnd size is adjusted in roughly the way that:
 133  * Van Jacobson, Congestion avoidance and Control, In "Proceedings of
 134  * SIGCOMM '88". ACM, August 1988.
 135  * describes for TCP. The cwnd size is chopped in half on a retransmit timeout
 136  * and incremented by 1/cwnd when each rpc reply is received and a full cwnd
 137  * of rpcs is in progress.
 138  * (The sent count and cwnd are scaled for integer arith.)
 139  * Variants of "slow start" were tried and were found to be too much of a
 140  * performance hit (ave. rtt 3 times larger),
 141  * I suspect due to the large rtt that nfs rpcs have.
 142  */
 143 #define NFS_CWNDSCALE   256
 144 #define NFS_MAXCWND     (NFS_CWNDSCALE * 32)
 145 static int nfs_backoff[8] = { 2, 4, 8, 16, 32, 64, 128, 256, };
 146 int nfsrtton = 0;
 147 struct nfsrtt nfsrtt;
 148 struct callout  nfs_timer_handle;
 149
 150 static int      nfs_msg (struct thread *,char *,char *);
 151 static int      nfs_rcvlock (struct nfsreq *);
 152 static void     nfs_rcvunlock (struct nfsreq *);
 153 static void     nfs_realign (struct mbuf **pm, int hsiz);
 154 static int      nfs_receive (struct nfsreq *rep, struct sockaddr **aname,
 155                                  struct mbuf **mp);
 156 static void     nfs_softterm (struct nfsreq *rep);
 157 static int      nfs_reconnect (struct nfsreq *rep);
 158 #ifndef NFS_NOSERVER
 159 static int      nfsrv_getstream (struct nfssvc_sock *, int, int *);
 160 static void     nfs_timer_req(struct nfsreq *req);
 161
 162 int (*nfsrv3_procs[NFS_NPROCS]) (struct nfsrv_descript *nd,
 163                                     struct nfssvc_sock *slp,
 164                                     struct thread *td,
 165                                     struct mbuf **mreqp) = {
 166         nfsrv_null,
 167         nfsrv_getattr,
 168         nfsrv_setattr,
 169         nfsrv_lookup,
 170         nfsrv3_access,
 171         nfsrv_readlink,
 172         nfsrv_read,
 173         nfsrv_write,
 174         nfsrv_create,
 175         nfsrv_mkdir,
 176         nfsrv_symlink,
 177         nfsrv_mknod,
 178         nfsrv_remove,
 179         nfsrv_rmdir,
 180         nfsrv_rename,
 181         nfsrv_link,
 182         nfsrv_readdir,
 183         nfsrv_readdirplus,
 184         nfsrv_statfs,
 185         nfsrv_fsinfo,
 186         nfsrv_pathconf,
 187         nfsrv_commit,
 188         nfsrv_noop,
 189         nfsrv_noop,
 190         nfsrv_noop,
 191         nfsrv_noop
 192 };
 193 #endif /* NFS_NOSERVER */
 194
 195 /*
 196  * Initialize sockets and congestion for a new NFS connection.
 197  * We do not free the sockaddr if error.
 198  */
 199 int
 200 nfs_connect(struct nfsmount *nmp, struct nfsreq *rep)
 201 {
 202         struct socket *so;
 203         int error, rcvreserve, sndreserve;
 204         int pktscale;
 205         struct sockaddr *saddr;
 206         struct sockaddr_in *sin;
 207         struct thread *td = &thread0; /* only used for socreate and sobind */
 208
 209         nmp->nm_so = NULL;
 210         saddr = nmp->nm_nam;
 211         error = socreate(saddr->sa_family, &nmp->nm_so, nmp->nm_sotype,
 212                 nmp->nm_soproto, td);
 213         if (error)
 214                 goto bad;
 215         so = nmp->nm_so;
 216         nmp->nm_soflags = so->so_proto->pr_flags;
 217
 218         /*
 219          * Some servers require that the client port be a reserved port number.
 220          */
 221         if (saddr->sa_family == AF_INET && (nmp->nm_flag & NFSMNT_RESVPORT)) {
 222                 struct sockopt sopt;
 223                 int ip;
 224                 struct sockaddr_in ssin;
 225
 226                 bzero(&sopt, sizeof sopt);
 227                 ip = IP_PORTRANGE_LOW;
 228                 sopt.sopt_level = IPPROTO_IP;
 229                 sopt.sopt_name = IP_PORTRANGE;
 230                 sopt.sopt_val = (void *)&ip;
 231                 sopt.sopt_valsize = sizeof(ip);
 232                 sopt.sopt_td = NULL;
 233                 error = sosetopt(so, &sopt);
 234                 if (error)
 235                         goto bad;
 236                 bzero(&ssin, sizeof ssin);
 237                 sin = &ssin;
 238                 sin->sin_len = sizeof (struct sockaddr_in);
 239                 sin->sin_family = AF_INET;
 240                 sin->sin_addr.s_addr = INADDR_ANY;
 241                 sin->sin_port = htons(0);
 242                 error = sobind(so, (struct sockaddr *)sin, td);
 243                 if (error)
 244                         goto bad;
 245                 bzero(&sopt, sizeof sopt);
 246                 ip = IP_PORTRANGE_DEFAULT;
 247                 sopt.sopt_level = IPPROTO_IP;
 248                 sopt.sopt_name = IP_PORTRANGE;
 249                 sopt.sopt_val = (void *)&ip;
 250                 sopt.sopt_valsize = sizeof(ip);
 251                 sopt.sopt_td = NULL;
 252                 error = sosetopt(so, &sopt);
 253                 if (error)
 254                         goto bad;
 255         }
 256
 257         /*
 258          * Protocols that do not require connections may be optionally left
 259          * unconnected for servers that reply from a port other than NFS_PORT.
 260          */
 261         if (nmp->nm_flag & NFSMNT_NOCONN) {
 262                 if (nmp->nm_soflags & PR_CONNREQUIRED) {
 263                         error = ENOTCONN;
 264                         goto bad;
 265                 }
 266         } else {
 267                 error = soconnect(so, nmp->nm_nam, td);
 268                 if (error)
 269                         goto bad;
 270
 271                 /*
 272                  * Wait for the connection to complete. Cribbed from the
 273                  * connect system call but with the wait timing out so
 274                  * that interruptible mounts don't hang here for a long time.
 275                  */
 276                 crit_enter();
 277                 while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) {
 278                         (void) tsleep((caddr_t)&so->so_timeo, 0,
 279                                 "nfscon", 2 * hz);
 280                         if ((so->so_state & SS_ISCONNECTING) &&
 281                             so->so_error == 0 && rep &&
 282                             (error = nfs_sigintr(nmp, rep, rep->r_td)) != 0){
 283                                 so->so_state &= ~SS_ISCONNECTING;
 284                                 crit_exit();
 285                                 goto bad;
 286                         }
 287                 }
 288                 if (so->so_error) {
 289                         error = so->so_error;
 290                         so->so_error = 0;
 291                         crit_exit();
 292                         goto bad;
 293                 }
 294                 crit_exit();
 295         }
 296         so->so_rcv.ssb_timeo = (5 * hz);
 297         so->so_snd.ssb_timeo = (5 * hz);
 298
 299         /*
 300          * Get buffer reservation size from sysctl, but impose reasonable
 301          * limits.
 302          */
 303         pktscale = nfs_bufpackets;
 304         if (pktscale < 2)
 305                 pktscale = 2;
 306         if (pktscale > 64)
 307                 pktscale = 64;
 308
 309         if (nmp->nm_sotype == SOCK_DGRAM) {
 310                 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
 311                 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
 312                     NFS_MAXPKTHDR) * pktscale;
 313         } else if (nmp->nm_sotype == SOCK_SEQPACKET) {
 314                 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR) * pktscale;
 315                 rcvreserve = (max(nmp->nm_rsize, nmp->nm_readdirsize) +
 316                     NFS_MAXPKTHDR) * pktscale;
 317         } else {
 318                 if (nmp->nm_sotype != SOCK_STREAM)
 319                         panic("nfscon sotype");
 320                 if (so->so_proto->pr_flags & PR_CONNREQUIRED) {
 321                         struct sockopt sopt;
 322                         int val;
 323
 324                         bzero(&sopt, sizeof sopt);
 325                         sopt.sopt_level = SOL_SOCKET;
 326                         sopt.sopt_name = SO_KEEPALIVE;
 327                         sopt.sopt_val = &val;
 328                         sopt.sopt_valsize = sizeof val;
 329                         val = 1;
 330                         sosetopt(so, &sopt);
 331                 }
 332                 if (so->so_proto->pr_protocol == IPPROTO_TCP) {
 333                         struct sockopt sopt;
 334                         int val;
 335
 336                         bzero(&sopt, sizeof sopt);
 337                         sopt.sopt_level = IPPROTO_TCP;
 338                         sopt.sopt_name = TCP_NODELAY;
 339                         sopt.sopt_val = &val;
 340                         sopt.sopt_valsize = sizeof val;
 341                         val = 1;
 342                         sosetopt(so, &sopt);
 343                 }
 344                 sndreserve = (nmp->nm_wsize + NFS_MAXPKTHDR +
 345                     sizeof (u_int32_t)) * pktscale;
 346                 rcvreserve = (nmp->nm_rsize + NFS_MAXPKTHDR +
 347                     sizeof (u_int32_t)) * pktscale;
 348         }
 349         error = soreserve(so, sndreserve, rcvreserve,
 350                           &td->td_proc->p_rlimit[RLIMIT_SBSIZE]);
 351         if (error)
 352                 goto bad;
 353         so->so_rcv.ssb_flags |= SSB_NOINTR;
 354         so->so_snd.ssb_flags |= SSB_NOINTR;
 355
 356         /* Initialize other non-zero congestion variables */
 357         nmp->nm_srtt[0] = nmp->nm_srtt[1] = nmp->nm_srtt[2] =
 358                 nmp->nm_srtt[3] = (NFS_TIMEO << 3);
 359         nmp->nm_sdrtt[0] = nmp->nm_sdrtt[1] = nmp->nm_sdrtt[2] =
 360                 nmp->nm_sdrtt[3] = 0;
 361         nmp->nm_cwnd = NFS_MAXCWND / 2;     /* Initial send window */
 362         nmp->nm_sent = 0;
 363         nmp->nm_timeouts = 0;
 364         return (0);
 365
 366 bad:
 367         nfs_disconnect(nmp);
 368         return (error);
 369 }
 370
 371 /*
 372  * Reconnect routine:
 373  * Called when a connection is broken on a reliable protocol.
 374  * - clean up the old socket
 375  * - nfs_connect() again
 376  * - set R_MUSTRESEND for all outstanding requests on mount point
 377  * If this fails the mount point is DEAD!
 378  * nb: Must be called with the nfs_sndlock() set on the mount point.
 379  */
 380 static int
 381 nfs_reconnect(struct nfsreq *rep)
 382 {
 383         struct nfsreq *req;
 384         struct nfsmount *nmp = rep->r_nmp;
 385         int error;
 386
 387         nfs_disconnect(nmp);
 388         while ((error = nfs_connect(nmp, rep)) != 0) {
 389                 if (error == EINTR || error == ERESTART)
 390                         return (EINTR);
 391                 (void) tsleep((caddr_t)&lbolt, 0, "nfscon", 0);
 392         }
 393
 394         /*
 395          * Loop through outstanding request list and fix up all requests
 396          * on old socket.
 397          */
 398         crit_enter();
 399         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
 400                 KKASSERT(req->r_nmp == nmp);
 401                 req->r_flags |= R_MUSTRESEND;
 402         }
 403         crit_exit();
 404         return (0);
 405 }
 406
 407 /*
 408  * NFS disconnect. Clean up and unlink.
 409  */
 410 void
 411 nfs_disconnect(struct nfsmount *nmp)
 412 {
 413         struct socket *so;
 414
 415         if (nmp->nm_so) {
 416                 so = nmp->nm_so;
 417                 nmp->nm_so = NULL;
 418                 soshutdown(so, SHUT_RDWR);
 419                 soclose(so, FNONBLOCK);
 420         }
 421 }
 422
 423 void
 424 nfs_safedisconnect(struct nfsmount *nmp)
 425 {
 426         struct nfsreq dummyreq;
 427
 428         bzero(&dummyreq, sizeof(dummyreq));
 429         dummyreq.r_nmp = nmp;
 430         dummyreq.r_td = NULL;
 431         mtx_link_init(&dummyreq.r_link);
 432         nfs_rcvlock(&dummyreq);
 433         nfs_disconnect(nmp);
 434         nfs_rcvunlock(&dummyreq);
 435 }
 436
 437 /*
 438  * This is the nfs send routine. For connection based socket types, it
 439  * must be called with an nfs_sndlock() on the socket.
 440  * "rep == NULL" indicates that it has been called from a server.
 441  * For the client side:
 442  * - return EINTR if the RPC is terminated, 0 otherwise
 443  * - set R_MUSTRESEND if the send fails for any reason
 444  * - do any cleanup required by recoverable socket errors (?)
 445  * For the server side:
 446  * - return EINTR or ERESTART if interrupted by a signal
 447  * - return EPIPE if a connection is lost for connection based sockets (TCP...)
 448  * - do any cleanup required by recoverable socket errors (?)
 449  */
 450 int
 451 nfs_send(struct socket *so, struct sockaddr *nam, struct mbuf *top,
 452          struct nfsreq *rep)
 453 {
 454         struct sockaddr *sendnam;
 455         int error, soflags, flags;
 456
 457         if (rep) {
 458                 if (rep->r_flags & R_SOFTTERM) {
 459                         m_freem(top);
 460                         return (EINTR);
 461                 }
 462                 if ((so = rep->r_nmp->nm_so) == NULL) {
 463                         rep->r_flags |= R_MUSTRESEND;
 464                         m_freem(top);
 465                         return (0);
 466                 }
 467                 rep->r_flags &= ~R_MUSTRESEND;
 468                 soflags = rep->r_nmp->nm_soflags;
 469         } else
 470                 soflags = so->so_proto->pr_flags;
 471         if ((soflags & PR_CONNREQUIRED) || (so->so_state & SS_ISCONNECTED))
 472                 sendnam = NULL;
 473         else
 474                 sendnam = nam;
 475         if (so->so_type == SOCK_SEQPACKET)
 476                 flags = MSG_EOR;
 477         else
 478                 flags = 0;
 479
 480         error = so_pru_sosend(so, sendnam, NULL, top, NULL, flags,
 481             curthread /*XXX*/);
 482         /*
 483          * ENOBUFS for dgram sockets is transient and non fatal.
 484          * No need to log, and no need to break a soft mount.
 485          */
 486         if (error == ENOBUFS && so->so_type == SOCK_DGRAM) {
 487                 error = 0;
 488                 if (rep)                /* do backoff retransmit on client */
 489                         rep->r_flags |= R_MUSTRESEND;
 490         }
 491
 492         if (error) {
 493                 if (rep) {
 494                         log(LOG_INFO, "nfs send error %d for server %s\n",error,
 495                             rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
 496                         /*
 497                          * Deal with errors for the client side.
 498                          */
 499                         if (rep->r_flags & R_SOFTTERM)
 500                                 error = EINTR;
 501                         else
 502                                 rep->r_flags |= R_MUSTRESEND;
 503                 } else
 504                         log(LOG_INFO, "nfsd send error %d\n", error);
 505
 506                 /*
 507                  * Handle any recoverable (soft) socket errors here. (?)
 508                  */
 509                 if (error != EINTR && error != ERESTART &&
 510                         error != EWOULDBLOCK && error != EPIPE)
 511                         error = 0;
 512         }
 513         return (error);
 514 }
 515
 516 /*
 517  * Receive a Sun RPC Request/Reply. For SOCK_DGRAM, the work is all
 518  * done by soreceive(), but for SOCK_STREAM we must deal with the Record
 519  * Mark and consolidate the data into a new mbuf list.
 520  * nb: Sometimes TCP passes the data up to soreceive() in long lists of
 521  *     small mbufs.
 522  * For SOCK_STREAM we must be very careful to read an entire record once
 523  * we have read any of it, even if the system call has been interrupted.
 524  */
 525 static int
 526 nfs_receive(struct nfsreq *rep, struct sockaddr **aname, struct mbuf **mp)
 527 {
 528         struct socket *so;
 529         struct sockbuf sio;
 530         struct uio auio;
 531         struct iovec aio;
 532         struct mbuf *m;
 533         struct mbuf *control;
 534         u_int32_t len;
 535         struct sockaddr **getnam;
 536         int error, sotype, rcvflg;
 537         struct thread *td = curthread;  /* XXX */
 538
 539         /*
 540          * Set up arguments for soreceive()
 541          */
 542         *mp = NULL;
 543         *aname = NULL;
 544         sotype = rep->r_nmp->nm_sotype;
 545
 546         /*
 547          * For reliable protocols, lock against other senders/receivers
 548          * in case a reconnect is necessary.
 549          * For SOCK_STREAM, first get the Record Mark to find out how much
 550          * more there is to get.
 551          * We must lock the socket against other receivers
 552          * until we have an entire rpc request/reply.
 553          */
 554         if (sotype != SOCK_DGRAM) {
 555                 error = nfs_sndlock(rep);
 556                 if (error)
 557                         return (error);
 558 tryagain:
 559                 /*
 560                  * Check for fatal errors and resending request.
 561                  */
 562                 /*
 563                  * Ugh: If a reconnect attempt just happened, nm_so
 564                  * would have changed. NULL indicates a failed
 565                  * attempt that has essentially shut down this
 566                  * mount point.
 567                  */
 568                 if (rep->r_mrep || (rep->r_flags & R_SOFTTERM)) {
 569                         nfs_sndunlock(rep);
 570                         return (EINTR);
 571                 }
 572                 so = rep->r_nmp->nm_so;
 573                 if (!so) {
 574                         error = nfs_reconnect(rep);
 575                         if (error) {
 576                                 nfs_sndunlock(rep);
 577                                 return (error);
 578                         }
 579                         goto tryagain;
 580                 }
 581                 while (rep->r_flags & R_MUSTRESEND) {
 582                         m = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
 583                         nfsstats.rpcretries++;
 584                         error = nfs_send(so, rep->r_nmp->nm_nam, m, rep);
 585                         if (error) {
 586                                 if (error == EINTR || error == ERESTART ||
 587                                     (error = nfs_reconnect(rep)) != 0) {
 588                                         nfs_sndunlock(rep);
 589                                         return (error);
 590                                 }
 591                                 goto tryagain;
 592                         }
 593                 }
 594                 nfs_sndunlock(rep);
 595                 if (sotype == SOCK_STREAM) {
 596                         /*
 597                          * Get the length marker from the stream
 598                          */
 599                         aio.iov_base = (caddr_t)&len;
 600                         aio.iov_len = sizeof(u_int32_t);
 601                         auio.uio_iov = &aio;
 602                         auio.uio_iovcnt = 1;
 603                         auio.uio_segflg = UIO_SYSSPACE;
 604                         auio.uio_rw = UIO_READ;
 605                         auio.uio_offset = 0;
 606                         auio.uio_resid = sizeof(u_int32_t);
 607                         auio.uio_td = td;
 608                         do {
 609                            rcvflg = MSG_WAITALL;
 610                            error = so_pru_soreceive(so, NULL, &auio, NULL,
 611                                                     NULL, &rcvflg);
 612                            if (error == EWOULDBLOCK && rep) {
 613                                 if (rep->r_flags & R_SOFTTERM)
 614                                         return (EINTR);
 615                            }
 616                         } while (error == EWOULDBLOCK);
 617
 618                         if (error == 0 && auio.uio_resid > 0) {
 619                             /*
 620                              * Only log short packets if not EOF
 621                              */
 622                             if (auio.uio_resid != sizeof(u_int32_t))
 623                             log(LOG_INFO,
 624                                  "short receive (%d/%d) from nfs server %s\n",
 625                                  (int)(sizeof(u_int32_t) - auio.uio_resid),
 626                                  (int)sizeof(u_int32_t),
 627                                  rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
 628                             error = EPIPE;
 629                         }
 630                         if (error)
 631                                 goto errout;
 632                         len = ntohl(len) & ~0x80000000;
 633                         /*
 634                          * This is SERIOUS! We are out of sync with the sender
 635                          * and forcing a disconnect/reconnect is all I can do.
 636                          */
 637                         if (len > NFS_MAXPACKET) {
 638                             log(LOG_ERR, "%s (%d) from nfs server %s\n",
 639                                 "impossible packet length",
 640                                 len,
 641                                 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
 642                             error = EFBIG;
 643                             goto errout;
 644                         }
 645
 646                         /*
 647                          * Get the rest of the packet as an mbuf chain
 648                          */
 649                         sbinit(&sio, len);
 650                         do {
 651                             rcvflg = MSG_WAITALL;
 652                             error = so_pru_soreceive(so, NULL, NULL, &sio,
 653                                                      NULL, &rcvflg);
 654                         } while (error == EWOULDBLOCK || error == EINTR ||
 655                                  error == ERESTART);
 656                         if (error == 0 && sio.sb_cc != len) {
 657                             if (sio.sb_cc != 0)
 658                             log(LOG_INFO,
 659                                 "short receive (%d/%d) from nfs server %s\n",
 660                                 len - auio.uio_resid, len,
 661                                 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
 662                             error = EPIPE;
 663                         }
 664                         *mp = sio.sb_mb;
 665                 } else {
 666                         /*
 667                          * Non-stream, so get the whole packet by not
 668                          * specifying MSG_WAITALL and by specifying a large
 669                          * length.
 670                          *
 671                          * We have no use for control msg., but must grab them
 672                          * and then throw them away so we know what is going
 673                          * on.
 674                          */
 675                         sbinit(&sio, 100000000);
 676                         do {
 677                             rcvflg = 0;
 678                             error =  so_pru_soreceive(so, NULL, NULL, &sio,
 679                                                       &control, &rcvflg);
 680                             if (control)
 681                                 m_freem(control);
 682                             if (error == EWOULDBLOCK && rep) {
 683                                 if (rep->r_flags & R_SOFTTERM) {
 684                                         m_freem(sio.sb_mb);
 685                                         return (EINTR);
 686                                 }
 687                             }
 688                         } while (error == EWOULDBLOCK ||
 689                                  (error == 0 && sio.sb_mb == NULL && control));
 690                         if ((rcvflg & MSG_EOR) == 0)
 691                                 kprintf("Egad!!\n");
 692                         if (error == 0 && sio.sb_mb == NULL)
 693                                 error = EPIPE;
 694                         len = sio.sb_cc;
 695                         *mp = sio.sb_mb;
 696                 }
 697 errout:
 698                 if (error && error != EINTR && error != ERESTART) {
 699                         m_freem(*mp);
 700                         *mp = NULL;
 701                         if (error != EPIPE) {
 702                                 log(LOG_INFO,
 703                                     "receive error %d from nfs server %s\n",
 704                                     error,
 705                                  rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
 706                         }
 707                         error = nfs_sndlock(rep);
 708                         if (!error) {
 709                                 error = nfs_reconnect(rep);
 710                                 if (!error)
 711                                         goto tryagain;
 712                                 else
 713                                         nfs_sndunlock(rep);
 714                         }
 715                 }
 716         } else {
 717                 if ((so = rep->r_nmp->nm_so) == NULL)
 718                         return (EACCES);
 719                 if (so->so_state & SS_ISCONNECTED)
 720                         getnam = NULL;
 721                 else
 722                         getnam = aname;
 723                 sbinit(&sio, 100000000);
 724                 do {
 725                         rcvflg = 0;
 726                         error =  so_pru_soreceive(so, getnam, NULL, &sio,
 727                                                   NULL, &rcvflg);
 728                         if (error == EWOULDBLOCK &&
 729                             (rep->r_flags & R_SOFTTERM)) {
 730                                 m_freem(sio.sb_mb);
 731                                 return (EINTR);
 732                         }
 733                 } while (error == EWOULDBLOCK);
 734                 len = sio.sb_cc;
 735                 *mp = sio.sb_mb;
 736         }
 737         if (error) {
 738                 m_freem(*mp);
 739                 *mp = NULL;
 740         }
 741         /*
 742          * Search for any mbufs that are not a multiple of 4 bytes long
 743          * or with m_data not longword aligned.
 744          * These could cause pointer alignment problems, so copy them to
 745          * well aligned mbufs.
 746          */
 747         nfs_realign(mp, 5 * NFSX_UNSIGNED);
 748         return (error);
 749 }
 750
 751 /*
 752  * Implement receipt of reply on a socket.
 753  * We must search through the list of received datagrams matching them
 754  * with outstanding requests using the xid, until ours is found.
 755  */
 756 /* ARGSUSED */
 757 int
 758 nfs_reply(struct nfsreq *myrep)
 759 {
 760         struct nfsreq *rep;
 761         struct nfsmount *nmp = myrep->r_nmp;
 762         int32_t t1;
 763         struct mbuf *mrep, *md;
 764         struct sockaddr *nam;
 765         u_int32_t rxid, *tl;
 766         caddr_t dpos, cp2;
 767         int error;
 768
 769         /*
 770          * Loop around until we get our own reply
 771          */
 772         for (;;) {
 773                 /*
 774                  * Lock against other receivers so that I don't get stuck in
 775                  * sbwait() after someone else has received my reply for me.
 776                  * Also necessary for connection based protocols to avoid
 777                  * race conditions during a reconnect.
 778                  *
 779                  * If nfs_rcvlock() returns EALREADY, that means that
 780                  * the reply has already been recieved by another
 781                  * process and we can return immediately.  In this
 782                  * case, the lock is not taken to avoid races with
 783                  * other processes.
 784                  */
 785                 error = nfs_rcvlock(myrep);
 786                 if (error == EALREADY)
 787                         return (0);
 788                 if (error)
 789                         return (error);
 790                 /*
 791                  * Get the next Rpc reply off the socket
 792                  */
 793                 error = nfs_receive(myrep, &nam, &mrep);
 794                 nfs_rcvunlock(myrep);
 795                 if (error) {
 796                         /*
 797                          * Ignore routing errors on connectionless protocols??
 798                          */
 799                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
 800                                 nmp->nm_so->so_error = 0;
 801                                 if (myrep->r_flags & R_GETONEREP)
 802                                         return (0);
 803                                 continue;
 804                         }
 805                         return (error);
 806                 }
 807                 if (nam)
 808                         FREE(nam, M_SONAME);
 809
 810                 /*
 811                  * Get the xid and check that it is an rpc reply
 812                  */
 813                 md = mrep;
 814                 dpos = mtod(md, caddr_t);
 815                 nfsm_dissect(tl, u_int32_t *, 2*NFSX_UNSIGNED);
 816                 rxid = *tl++;
 817                 if (*tl != rpc_reply) {
 818                         nfsstats.rpcinvalid++;
 819                         m_freem(mrep);
 820 nfsmout:
 821                         if (myrep->r_flags & R_GETONEREP)
 822                                 return (0);
 823                         continue;
 824                 }
 825
 826                 /*
 827                  * Loop through the request list to match up the reply
 828                  * Iff no match, just drop the datagram.  On match, set
 829                  * r_mrep atomically to prevent the timer from messing
 830                  * around with the request after we have exited the critical
 831                  * section.
 832                  */
 833                 crit_enter();
 834                 TAILQ_FOREACH(rep, &nmp->nm_reqq, r_chain) {
 835                         if (rep->r_mrep == NULL && rxid == rep->r_xid)
 836                                 break;
 837                 }
 838                 crit_exit();
 839
 840                 /*
 841                  * Fill in the rest of the reply if we found a match.
 842                  */
 843                 if (rep) {
 844                         rep->r_md = md;
 845                         rep->r_dpos = dpos;
 846                         if (nfsrtton) {
 847                                 struct rttl *rt;
 848
 849                                 rt = &nfsrtt.rttl[nfsrtt.pos];
 850                                 rt->proc = rep->r_procnum;
 851                                 rt->rto = NFS_RTO(nmp, proct[rep->r_procnum]);
 852                                 rt->sent = nmp->nm_sent;
 853                                 rt->cwnd = nmp->nm_cwnd;
 854                                 rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
 855                                 rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
 856                                 rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
 857                                 getmicrotime(&rt->tstamp);
 858                                 if (rep->r_flags & R_TIMING)
 859                                         rt->rtt = rep->r_rtt;
 860                                 else
 861                                         rt->rtt = 1000000;
 862                                 nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
 863                         }
 864                         /*
 865                          * Update congestion window.
 866                          * Do the additive increase of
 867                          * one rpc/rtt.
 868                          */
 869                         if (nmp->nm_cwnd <= nmp->nm_sent) {
 870                                 nmp->nm_cwnd +=
 871                                    (NFS_CWNDSCALE * NFS_CWNDSCALE +
 872                                    (nmp->nm_cwnd >> 1)) / nmp->nm_cwnd;
 873                                 if (nmp->nm_cwnd > NFS_MAXCWND)
 874                                         nmp->nm_cwnd = NFS_MAXCWND;
 875                         }
 876                         crit_enter();   /* nfs_timer interlock for nm_sent */
 877                         if (rep->r_flags & R_SENT) {
 878                                 rep->r_flags &= ~R_SENT;
 879                                 nmp->nm_sent -= NFS_CWNDSCALE;
 880                         }
 881                         crit_exit();
 882                         /*
 883                          * Update rtt using a gain of 0.125 on the mean
 884                          * and a gain of 0.25 on the deviation.
 885                          */
 886                         if (rep->r_flags & R_TIMING) {
 887                                 /*
 888                                  * Since the timer resolution of
 889                                  * NFS_HZ is so course, it can often
 890                                  * result in r_rtt == 0. Since
 891                                  * r_rtt == N means that the actual
 892                                  * rtt is between N+dt and N+2-dt ticks,
 893                                  * add 1.
 894                                  */
 895                                 t1 = rep->r_rtt + 1;
 896                                 t1 -= (NFS_SRTT(rep) >> 3);
 897                                 NFS_SRTT(rep) += t1;
 898                                 if (t1 < 0)
 899                                         t1 = -t1;
 900                                 t1 -= (NFS_SDRTT(rep) >> 2);
 901                                 NFS_SDRTT(rep) += t1;
 902                         }
 903                         nmp->nm_timeouts = 0;
 904                         rep->r_mrep = mrep;
 905                         mtx_abort_ex_link(&rep->r_nmp->nm_rxlock, &rep->r_link);
 906                 }
 907                 /*
 908                  * If not matched to a request, drop it.
 909                  * If it's mine, get out.
 910                  */
 911                 if (rep == NULL) {
 912                         nfsstats.rpcunexpected++;
 913                         m_freem(mrep);
 914                 } else if (rep == myrep) {
 915                         if (rep->r_mrep == NULL)
 916                                 panic("nfsreply nil");
 917                         return (0);
 918                 }
 919                 if (myrep->r_flags & R_GETONEREP)
 920                         return (0);
 921         }
 922 }
 923
 924 int
 925 nfs_request(struct vnode *vp, struct mbuf *mrest, int procnum,
 926             struct thread *td, struct ucred *cred, struct mbuf **mrp,
 927             struct mbuf **mdp, caddr_t *dposp)
 928 {
 929         struct nfsreq *rep = NULL;
 930         int error;
 931
 932         error = nfs_request_setup(vp, mrest, procnum, td, cred, &rep);
 933         if (error)
 934                 return (error);
 935         rep->r_mrp = mrp;
 936         rep->r_mdp = mdp;
 937         rep->r_dposp = dposp;
 938 needauth:
 939         error = nfs_request_auth(rep);
 940         if (error)
 941                 return (error);
 942 tryagain:
 943         error = nfs_request_try(rep);           /* error ignored */
 944         error = nfs_request_waitreply(rep);     /* pass to process */
 945         error = nfs_request_processreply(rep, error);
 946         if (error == ENEEDAUTH)
 947                 goto needauth;
 948         if (error == EAGAIN)
 949                 goto tryagain;
 950         return (error);
 951 }
 952
 953 /*
 954  * nfs_request - goes something like this
 955  *      - fill in request struct
 956  *      - links it into list
 957  *      - calls nfs_send() for first transmit
 958  *      - calls nfs_receive() to get reply
 959  *      - break down rpc header and return with nfs reply pointed to
 960  *        by mrep or error
 961  * nb: always frees up mreq mbuf list
 962  */
 963 int
 964 nfs_request_setup(struct vnode *vp, struct mbuf *mrest, int procnum,
 965             struct thread *td, struct ucred *cred,
 966             struct nfsreq **repp)
 967 {
 968         struct nfsreq *rep;
 969         struct nfsmount *nmp;
 970         struct mbuf *m;
 971         int i;
 972
 973         /* Reject requests while attempting a forced unmount. */
 974         if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) {
 975                 m_freem(mrest);
 976                 return (ESTALE);
 977         }
 978         nmp = VFSTONFS(vp->v_mount);
 979         MALLOC(rep, struct nfsreq *, sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
 980         rep->r_nmp = nmp;
 981         rep->r_vp = vp;
 982         rep->r_td = td;
 983         rep->r_procnum = procnum;
 984         rep->r_mreq = NULL;
 985         i = 0;
 986         m = mrest;
 987         while (m) {
 988                 i += m->m_len;
 989                 m = m->m_next;
 990         }
 991         rep->r_mrest = mrest;
 992         rep->r_mrest_len = i;
 993         rep->r_cred = cred;
 994         *repp = rep;
 995         return(0);
 996 }
 997
 998 int
 999 nfs_request_auth(struct nfsreq *rep)
1000 {
1001         struct nfsmount *nmp = rep->r_nmp;
1002         struct mbuf *m;
1003         char nickv[RPCX_NICKVERF];
1004         int error = 0, auth_len, auth_type;
1005         int verf_len;
1006         u_int32_t xid;
1007         char *auth_str, *verf_str;
1008         struct ucred *cred;
1009
1010         cred = rep->r_cred;
1011         rep->r_failed_auth = 0;
1012
1013         /*
1014          * Get the RPC header with authorization.
1015          */
1016         verf_str = auth_str = NULL;
1017         if (nmp->nm_flag & NFSMNT_KERB) {
1018                 verf_str = nickv;
1019                 verf_len = sizeof (nickv);
1020                 auth_type = RPCAUTH_KERB4;
1021                 bzero((caddr_t)rep->r_key, sizeof(rep->r_key));
1022                 if (rep->r_failed_auth ||
1023                     nfs_getnickauth(nmp, cred, &auth_str, &auth_len,
1024                                     verf_str, verf_len)) {
1025                         error = nfs_getauth(nmp, rep, cred, &auth_str,
1026                                 &auth_len, verf_str, &verf_len, rep->r_key);
1027                         if (error) {
1028                                 m_freem(rep->r_mrest);
1029                                 rep->r_mrest = NULL;
1030                                 kfree((caddr_t)rep, M_NFSREQ);
1031                                 return (error);
1032                         }
1033                 }
1034         } else {
1035                 auth_type = RPCAUTH_UNIX;
1036                 if (cred->cr_ngroups < 1)
1037                         panic("nfsreq nogrps");
1038                 auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
1039                         nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
1040                         5 * NFSX_UNSIGNED;
1041         }
1042         m = nfsm_rpchead(cred, nmp->nm_flag, rep->r_procnum, auth_type,
1043                         auth_len, auth_str, verf_len, verf_str,
1044                         rep->r_mrest, rep->r_mrest_len, &rep->r_mheadend, &xid);
1045         rep->r_mrest = NULL;
1046         if (auth_str)
1047                 kfree(auth_str, M_TEMP);
1048
1049         /*
1050          * For stream protocols, insert a Sun RPC Record Mark.
1051          */
1052         if (nmp->nm_sotype == SOCK_STREAM) {
1053                 M_PREPEND(m, NFSX_UNSIGNED, MB_WAIT);
1054                 if (m == NULL) {
1055                         kfree(rep, M_NFSREQ);
1056                         return (ENOBUFS);
1057                 }
1058                 *mtod(m, u_int32_t *) = htonl(0x80000000 |
1059                          (m->m_pkthdr.len - NFSX_UNSIGNED));
1060         }
1061         rep->r_mreq = m;
1062         rep->r_xid = xid;
1063         return (0);
1064 }
1065
1066 int
1067 nfs_request_try(struct nfsreq *rep)
1068 {
1069         struct nfsmount *nmp = rep->r_nmp;
1070         struct mbuf *m2;
1071         int error;
1072
1073         if (nmp->nm_flag & NFSMNT_SOFT)
1074                 rep->r_retry = nmp->nm_retry;
1075         else
1076                 rep->r_retry = NFS_MAXREXMIT + 1;       /* past clip limit */
1077         rep->r_rtt = rep->r_rexmit = 0;
1078         if (proct[rep->r_procnum] > 0)
1079                 rep->r_flags = R_TIMING | R_MASKTIMER;
1080         else
1081                 rep->r_flags = R_MASKTIMER;
1082         rep->r_mrep = NULL;
1083
1084         /*
1085          * Do the client side RPC.
1086          */
1087         nfsstats.rpcrequests++;
1088
1089         /*
1090          * Chain request into list of outstanding requests. Be sure
1091          * to put it LAST so timer finds oldest requests first.  Note
1092          * that R_MASKTIMER is set at the moment to prevent any timer
1093          * action on this request while we are still doing processing on
1094          * it below.  splsoftclock() primarily protects nm_sent.  Note
1095          * that we may block in this code so there is no atomicy guarentee.
1096          */
1097         crit_enter();
1098         TAILQ_INSERT_TAIL(&nmp->nm_reqq, rep, r_chain);
1099         mtx_link_init(&rep->r_link);
1100
1101         error = 0;
1102
1103         /*
1104          * If backing off another request or avoiding congestion, don't
1105          * send this one now but let timer do it.  If not timing a request,
1106          * do it now.
1107          *
1108          * Even though the timer will not mess with our request there is
1109          * still the possibility that we will race a reply (which clears
1110          * R_SENT), especially on localhost connections, so be very careful
1111          * when setting R_SENT.  We could set R_SENT prior to calling
1112          * nfs_send() but why bother if the response occurs that quickly?
1113          */
1114         if (nmp->nm_so && (nmp->nm_sotype != SOCK_DGRAM ||
1115             (nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1116             nmp->nm_sent < nmp->nm_cwnd)) {
1117                 if (nmp->nm_soflags & PR_CONNREQUIRED)
1118                         error = nfs_sndlock(rep);
1119                 if (!error) {
1120                         m2 = m_copym(rep->r_mreq, 0, M_COPYALL, MB_WAIT);
1121                         error = nfs_send(nmp->nm_so, nmp->nm_nam, m2, rep);
1122                         if (nmp->nm_soflags & PR_CONNREQUIRED)
1123                                 nfs_sndunlock(rep);
1124                 }
1125                 if (!error && (rep->r_flags & R_MUSTRESEND) == 0 &&
1126                     rep->r_mrep == NULL) {
1127                         KASSERT((rep->r_flags & R_SENT) == 0,
1128                                 ("R_SENT ASSERT %p", rep));
1129                         nmp->nm_sent += NFS_CWNDSCALE;
1130                         rep->r_flags |= R_SENT;
1131                 }
1132         } else {
1133                 rep->r_rtt = -1;
1134         }
1135         if (error == EPIPE)
1136                 error = 0;
1137         /*
1138          * Let the timer do what it will with the request, then
1139          * wait for the reply from our send or the timer's.
1140          */
1141         if (error == 0)
1142                 rep->r_flags &= ~R_MASKTIMER;
1143         crit_exit();
1144         return (error);
1145 }
1146
1147 int
1148 nfs_request_waitreply(struct nfsreq *rep)
1149 {
1150         struct nfsmount *nmp = rep->r_nmp;
1151         int error;
1152
1153
1154         error = nfs_reply(rep);
1155         crit_enter();
1156
1157         /*
1158          * RPC done, unlink the request, but don't rip it out from under
1159          * the callout timer.
1160          */
1161         while (rep->r_flags & R_LOCKED) {
1162                 nfs_timer_raced = 1;
1163                 tsleep(&nfs_timer_raced, 0, "nfstrac", 0);
1164         }
1165         TAILQ_REMOVE(&nmp->nm_reqq, rep, r_chain);
1166
1167         /*
1168          * Decrement the outstanding request count.
1169          */
1170         if (rep->r_flags & R_SENT) {
1171                 rep->r_flags &= ~R_SENT;
1172                 nmp->nm_sent -= NFS_CWNDSCALE;
1173         }
1174         crit_exit();
1175
1176         return (error);
1177 }
1178
1179 /*
1180  * Process reply with error returned from nfs_requet_waitreply().
1181  *
1182  * Returns EAGAIN if it wants us to loop up to nfs_request_try() again.
1183  * Returns ENEEDAUTH if it wants us to loop up to nfs_request_auth() again.
1184  */
1185 int
1186 nfs_request_processreply(struct nfsreq *rep, int error)
1187 {
1188         struct nfsmount *nmp = rep->r_nmp;
1189         time_t waituntil;
1190         caddr_t dpos, cp2;
1191         struct mbuf *mrep;
1192         struct mbuf *md;
1193         u_int32_t *tl;
1194         int trylater_delay = 15, trylater_cnt = 0;
1195         int verf_type;
1196         int t1;
1197         int i;
1198
1199         /*
1200          * If there was a successful reply and a tprintf msg.
1201          * tprintf a response.
1202          */
1203         if (!error && (rep->r_flags & R_TPRINTFMSG))
1204                 nfs_msg(rep->r_td, nmp->nm_mountp->mnt_stat.f_mntfromname,
1205                     "is alive again");
1206         mrep = rep->r_mrep;
1207         md = rep->r_md;
1208         dpos = rep->r_dpos;
1209         if (error) {
1210                 m_freem(rep->r_mreq);
1211                 kfree((caddr_t)rep, M_NFSREQ);
1212                 return (error);
1213         }
1214
1215         /*
1216          * break down the rpc header and check if ok
1217          */
1218         nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1219         if (*tl++ == rpc_msgdenied) {
1220                 if (*tl == rpc_mismatch) {
1221                         error = EOPNOTSUPP;
1222                 } else if ((nmp->nm_flag & NFSMNT_KERB) &&
1223                            *tl++ == rpc_autherr) {
1224                         if (!rep->r_failed_auth) {
1225                                 rep->r_failed_auth++;
1226                                 rep->r_mheadend->m_next = NULL;
1227                                 m_freem(mrep);
1228                                 m_freem(rep->r_mreq);
1229                                 return (ENEEDAUTH);
1230                         } else {
1231                                 error = EAUTH;
1232                         }
1233                 } else {
1234                         error = EACCES;
1235                 }
1236                 m_freem(mrep);
1237                 m_freem(rep->r_mreq);
1238                 kfree((caddr_t)rep, M_NFSREQ);
1239                 return (error);
1240         }
1241
1242         /*
1243          * Grab any Kerberos verifier, otherwise just throw it away.
1244          */
1245         verf_type = fxdr_unsigned(int, *tl++);
1246         i = fxdr_unsigned(int32_t, *tl);
1247         if ((nmp->nm_flag & NFSMNT_KERB) && verf_type == RPCAUTH_KERB4) {
1248                 error = nfs_savenickauth(nmp, rep->r_cred, i,
1249                                          rep->r_key, &md, &dpos, mrep);
1250                 if (error)
1251                         goto nfsmout;
1252         } else if (i > 0)
1253                 nfsm_adv(nfsm_rndup(i));
1254         nfsm_dissect(tl, u_int32_t *, NFSX_UNSIGNED);
1255         /* 0 == ok */
1256         if (*tl == 0) {
1257                 nfsm_dissect(tl, u_int32_t *, NFSX_UNSIGNED);
1258                 if (*tl != 0) {
1259                         error = fxdr_unsigned(int, *tl);
1260                         if ((nmp->nm_flag & NFSMNT_NFSV3) &&
1261                                 error == NFSERR_TRYLATER) {
1262                                 m_freem(mrep);
1263                                 error = 0;
1264                                 waituntil = time_second + trylater_delay;
1265                                 while (time_second < waituntil)
1266                                         (void) tsleep((caddr_t)&lbolt,
1267                                                 0, "nqnfstry", 0);
1268                                 trylater_delay *= nfs_backoff[trylater_cnt];
1269                                 if (trylater_cnt < 7)
1270                                         trylater_cnt++;
1271                                 rep->r_flags &= ~R_MASKTIMER;
1272                                 return (EAGAIN);        /* goto tryagain */
1273                         }
1274
1275                         /*
1276                          * If the File Handle was stale, invalidate the
1277                          * lookup cache, just in case.
1278                          *
1279                          * To avoid namecache<->vnode deadlocks we must
1280                          * release the vnode lock if we hold it.
1281                          */
1282                         if (error == ESTALE) {
1283                                 struct vnode *vp = rep->r_vp;
1284                                 int ltype;
1285
1286                                 ltype = lockstatus(&vp->v_lock, curthread);
1287                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1288                                         lockmgr(&vp->v_lock, LK_RELEASE);
1289                                 cache_inval_vp(vp, CINV_CHILDREN);
1290                                 if (ltype == LK_EXCLUSIVE || ltype == LK_SHARED)
1291                                         lockmgr(&vp->v_lock, ltype);
1292                         }
1293                         if (nmp->nm_flag & NFSMNT_NFSV3) {
1294                                 *rep->r_mrp = mrep;
1295                                 *rep->r_mdp = md;
1296                                 *rep->r_dposp = dpos;
1297                                 error |= NFSERR_RETERR;
1298                         } else
1299                                 m_freem(mrep);
1300                         m_freem(rep->r_mreq);
1301                         kfree((caddr_t)rep, M_NFSREQ);
1302                         return (error);
1303                 }
1304
1305                 *rep->r_mrp = mrep;
1306                 *rep->r_mdp = md;
1307                 *rep->r_dposp = dpos;
1308                 m_freem(rep->r_mreq);
1309                 FREE((caddr_t)rep, M_NFSREQ);
1310                 return (0);
1311         }
1312         m_freem(mrep);
1313         error = EPROTONOSUPPORT;
1314 nfsmout:
1315         m_freem(rep->r_mreq);
1316         kfree((caddr_t)rep, M_NFSREQ);
1317         return (error);
1318 }
1319
1320 #ifndef NFS_NOSERVER
1321 /*
1322  * Generate the rpc reply header
1323  * siz arg. is used to decide if adding a cluster is worthwhile
1324  */
1325 int
1326 nfs_rephead(int siz, struct nfsrv_descript *nd, struct nfssvc_sock *slp,
1327             int err, struct mbuf **mrq, struct mbuf **mbp, caddr_t *bposp)
1328 {
1329         u_int32_t *tl;
1330         struct mbuf *mreq;
1331         caddr_t bpos;
1332         struct mbuf *mb, *mb2;
1333
1334         siz += RPC_REPLYSIZ;
1335         mb = mreq = m_getl(max_hdr + siz, MB_WAIT, MT_DATA, M_PKTHDR, NULL);
1336         mreq->m_pkthdr.len = 0;
1337         /*
1338          * If this is not a cluster, try and leave leading space
1339          * for the lower level headers.
1340          */
1341         if ((max_hdr + siz) < MINCLSIZE)
1342                 mreq->m_data += max_hdr;
1343         tl = mtod(mreq, u_int32_t *);
1344         mreq->m_len = 6 * NFSX_UNSIGNED;
1345         bpos = ((caddr_t)tl) + mreq->m_len;
1346         *tl++ = txdr_unsigned(nd->nd_retxid);
1347         *tl++ = rpc_reply;
1348         if (err == ERPCMISMATCH || (err & NFSERR_AUTHERR)) {
1349                 *tl++ = rpc_msgdenied;
1350                 if (err & NFSERR_AUTHERR) {
1351                         *tl++ = rpc_autherr;
1352                         *tl = txdr_unsigned(err & ~NFSERR_AUTHERR);
1353                         mreq->m_len -= NFSX_UNSIGNED;
1354                         bpos -= NFSX_UNSIGNED;
1355                 } else {
1356                         *tl++ = rpc_mismatch;
1357                         *tl++ = txdr_unsigned(RPC_VER2);
1358                         *tl = txdr_unsigned(RPC_VER2);
1359                 }
1360         } else {
1361                 *tl++ = rpc_msgaccepted;
1362
1363                 /*
1364                  * For Kerberos authentication, we must send the nickname
1365                  * verifier back, otherwise just RPCAUTH_NULL.
1366                  */
1367                 if (nd->nd_flag & ND_KERBFULL) {
1368                     struct nfsuid *nuidp;
1369                     struct timeval ktvin, ktvout;
1370
1371                     for (nuidp = NUIDHASH(slp, nd->nd_cr.cr_uid)->lh_first;
1372                         nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
1373                         if (nuidp->nu_cr.cr_uid == nd->nd_cr.cr_uid &&
1374                             (!nd->nd_nam2 || netaddr_match(NU_NETFAM(nuidp),
1375                              &nuidp->nu_haddr, nd->nd_nam2)))
1376                             break;
1377                     }
1378                     if (nuidp) {
1379                         ktvin.tv_sec =
1380                             txdr_unsigned(nuidp->nu_timestamp.tv_sec - 1);
1381                         ktvin.tv_usec =
1382                             txdr_unsigned(nuidp->nu_timestamp.tv_usec);
1383
1384                         /*
1385                          * Encrypt the timestamp in ecb mode using the
1386                          * session key.
1387                          */
1388 #ifdef NFSKERB
1389                         XXX
1390 #endif
1391
1392                         *tl++ = rpc_auth_kerb;
1393                         *tl++ = txdr_unsigned(3 * NFSX_UNSIGNED);
1394                         *tl = ktvout.tv_sec;
1395                         nfsm_build(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1396                         *tl++ = ktvout.tv_usec;
1397                         *tl++ = txdr_unsigned(nuidp->nu_cr.cr_uid);
1398                     } else {
1399                         *tl++ = 0;
1400                         *tl++ = 0;
1401                     }
1402                 } else {
1403                         *tl++ = 0;
1404                         *tl++ = 0;
1405                 }
1406                 switch (err) {
1407                 case EPROGUNAVAIL:
1408                         *tl = txdr_unsigned(RPC_PROGUNAVAIL);
1409                         break;
1410                 case EPROGMISMATCH:
1411                         *tl = txdr_unsigned(RPC_PROGMISMATCH);
1412                         nfsm_build(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
1413                         *tl++ = txdr_unsigned(2);
1414                         *tl = txdr_unsigned(3);
1415                         break;
1416                 case EPROCUNAVAIL:
1417                         *tl = txdr_unsigned(RPC_PROCUNAVAIL);
1418                         break;
1419                 case EBADRPC:
1420                         *tl = txdr_unsigned(RPC_GARBAGE);
1421                         break;
1422                 default:
1423                         *tl = 0;
1424                         if (err != NFSERR_RETVOID) {
1425                                 nfsm_build(tl, u_int32_t *, NFSX_UNSIGNED);
1426                                 if (err)
1427                                     *tl = txdr_unsigned(nfsrv_errmap(nd, err));
1428                                 else
1429                                     *tl = 0;
1430                         }
1431                         break;
1432                 };
1433         }
1434
1435         if (mrq != NULL)
1436             *mrq = mreq;
1437         *mbp = mb;
1438         *bposp = bpos;
1439         if (err != 0 && err != NFSERR_RETVOID)
1440                 nfsstats.srvrpc_errs++;
1441         return (0);
1442 }
1443
1444
1445 #endif /* NFS_NOSERVER */
1446 /*
1447  * Nfs timer routine
1448  * Scan the nfsreq list and retranmit any requests that have timed out
1449  * To avoid retransmission attempts on STREAM sockets (in the future) make
1450  * sure to set the r_retry field to 0 (implies nm_retry == 0).
1451  */
1452 void
1453 nfs_timer(void *arg /* never used */)
1454 {
1455         struct nfsmount *nmp;
1456         struct nfsreq *req;
1457 #ifndef NFS_NOSERVER
1458         struct nfssvc_sock *slp;
1459         u_quad_t cur_usec;
1460 #endif /* NFS_NOSERVER */
1461
1462         crit_enter();
1463         TAILQ_FOREACH(nmp, &nfs_mountq, nm_entry) {
1464                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1465                         KKASSERT(nmp == req->r_nmp);
1466                         if (req->r_mrep ||
1467                             (req->r_flags & (R_SOFTTERM|R_MASKTIMER))) {
1468                                 continue;
1469                         }
1470                         req->r_flags |= R_LOCKED;
1471                         if (nfs_sigintr(nmp, req, req->r_td)) {
1472                                 nfs_softterm(req);
1473                         } else {
1474                                 nfs_timer_req(req);
1475                         }
1476                         req->r_flags &= ~R_LOCKED;
1477                 }
1478         }
1479 #ifndef NFS_NOSERVER
1480
1481         /*
1482          * Scan the write gathering queues for writes that need to be
1483          * completed now.
1484          */
1485         cur_usec = nfs_curusec();
1486         TAILQ_FOREACH(slp, &nfssvc_sockhead, ns_chain) {
1487             if (slp->ns_tq.lh_first && slp->ns_tq.lh_first->nd_time<=cur_usec)
1488                 nfsrv_wakenfsd(slp, 1);
1489         }
1490 #endif /* NFS_NOSERVER */
1491
1492         /*
1493          * Due to possible blocking, a client operation may be waiting for
1494          * us to finish processing this request so it can remove it.
1495          */
1496         if (nfs_timer_raced) {
1497                 nfs_timer_raced = 0;
1498                 wakeup(&nfs_timer_raced);
1499         }
1500         crit_exit();
1501         callout_reset(&nfs_timer_handle, nfs_ticks, nfs_timer, NULL);
1502 }
1503
1504 static
1505 void
1506 nfs_timer_req(struct nfsreq *req)
1507 {
1508         struct thread *td = &thread0; /* XXX for creds, will break if sleep */
1509         struct nfsmount *nmp = req->r_nmp;
1510         struct mbuf *m;
1511         struct socket *so;
1512         int timeo;
1513         int error;
1514
1515         if (req->r_rtt >= 0) {
1516                 req->r_rtt++;
1517                 if (nmp->nm_flag & NFSMNT_DUMBTIMR)
1518                         timeo = nmp->nm_timeo;
1519                 else
1520                         timeo = NFS_RTO(nmp, proct[req->r_procnum]);
1521                 if (nmp->nm_timeouts > 0)
1522                         timeo *= nfs_backoff[nmp->nm_timeouts - 1];
1523                 if (req->r_rtt <= timeo)
1524                         return;
1525                 if (nmp->nm_timeouts < 8)
1526                         nmp->nm_timeouts++;
1527         }
1528         /*
1529          * Check for server not responding
1530          */
1531         if ((req->r_flags & R_TPRINTFMSG) == 0 &&
1532              req->r_rexmit > nmp->nm_deadthresh) {
1533                 nfs_msg(req->r_td,
1534                     nmp->nm_mountp->mnt_stat.f_mntfromname,
1535                     "not responding");
1536                 req->r_flags |= R_TPRINTFMSG;
1537         }
1538         if (req->r_rexmit >= req->r_retry) {    /* too many */
1539                 nfsstats.rpctimeouts++;
1540                 nfs_softterm(req);
1541                 return;
1542         }
1543         if (nmp->nm_sotype != SOCK_DGRAM) {
1544                 if (++req->r_rexmit > NFS_MAXREXMIT)
1545                         req->r_rexmit = NFS_MAXREXMIT;
1546                 return;
1547         }
1548         if ((so = nmp->nm_so) == NULL)
1549                 return;
1550
1551         /*
1552          * If there is enough space and the window allows..
1553          *      Resend it
1554          * Set r_rtt to -1 in case we fail to send it now.
1555          */
1556         req->r_rtt = -1;
1557         if (ssb_space(&so->so_snd) >= req->r_mreq->m_pkthdr.len &&
1558            ((nmp->nm_flag & NFSMNT_DUMBTIMR) ||
1559             (req->r_flags & R_SENT) ||
1560             nmp->nm_sent < nmp->nm_cwnd) &&
1561            (m = m_copym(req->r_mreq, 0, M_COPYALL, MB_DONTWAIT))){
1562                 if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
1563                     error = so_pru_send(so, 0, m, NULL, NULL, td);
1564                 else
1565                     error = so_pru_send(so, 0, m, nmp->nm_nam,
1566                         NULL, td);
1567                 if (error) {
1568                         if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
1569                                 so->so_error = 0;
1570                 } else if (req->r_mrep == NULL) {
1571                         /*
1572                          * Iff first send, start timing
1573                          * else turn timing off, backoff timer
1574                          * and divide congestion window by 2.
1575                          *
1576                          * It is possible for the so_pru_send() to
1577                          * block and for us to race a reply so we
1578                          * only do this if the reply field has not
1579                          * been filled in.  R_LOCKED will prevent
1580                          * the request from being ripped out from under
1581                          * us entirely.
1582                          */
1583                         if (req->r_flags & R_SENT) {
1584                                 req->r_flags &= ~R_TIMING;
1585                                 if (++req->r_rexmit > NFS_MAXREXMIT)
1586                                         req->r_rexmit = NFS_MAXREXMIT;
1587                                 nmp->nm_cwnd >>= 1;
1588                                 if (nmp->nm_cwnd < NFS_CWNDSCALE)
1589                                         nmp->nm_cwnd = NFS_CWNDSCALE;
1590                                 nfsstats.rpcretries++;
1591                         } else {
1592                                 req->r_flags |= R_SENT;
1593                                 nmp->nm_sent += NFS_CWNDSCALE;
1594                         }
1595                         req->r_rtt = 0;
1596                 }
1597         }
1598 }
1599
1600 /*
1601  * Mark all of an nfs mount's outstanding requests with R_SOFTTERM and
1602  * wait for all requests to complete. This is used by forced unmounts
1603  * to terminate any outstanding RPCs.
1604  */
1605 int
1606 nfs_nmcancelreqs(struct nfsmount *nmp)
1607 {
1608         struct nfsreq *req;
1609         int i;
1610
1611         crit_enter();
1612         TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1613                 if (nmp != req->r_nmp || req->r_mrep != NULL ||
1614                     (req->r_flags & R_SOFTTERM)) {
1615                         continue;
1616                 }
1617                 nfs_softterm(req);
1618         }
1619         crit_exit();
1620
1621         for (i = 0; i < 30; i++) {
1622                 crit_enter();
1623                 TAILQ_FOREACH(req, &nmp->nm_reqq, r_chain) {
1624                         if (nmp == req->r_nmp)
1625                                 break;
1626                 }
1627                 crit_exit();
1628                 if (req == NULL)
1629                         return (0);
1630                 tsleep(&lbolt, 0, "nfscancel", 0);
1631         }
1632         return (EBUSY);
1633 }
1634
1635 /*
1636  * Flag a request as being about to terminate (due to NFSMNT_INT/NFSMNT_SOFT).
1637  * The nm_send count is decremented now to avoid deadlocks when the process in
1638  * soreceive() hasn't yet managed to send its own request.
1639  *
1640  * This routine must be called at splsoftclock() to protect r_flags and
1641  * nm_sent.
1642  */
1643
1644 static void
1645 nfs_softterm(struct nfsreq *rep)
1646 {
1647         rep->r_flags |= R_SOFTTERM;
1648
1649         if (rep->r_flags & R_SENT) {
1650                 rep->r_nmp->nm_sent -= NFS_CWNDSCALE;
1651                 rep->r_flags &= ~R_SENT;
1652         }
1653 }
1654
1655 /*
1656  * Test for a termination condition pending on the process.
1657  * This is used for NFSMNT_INT mounts.
1658  */
1659 int
1660 nfs_sigintr(struct nfsmount *nmp, struct nfsreq *rep, struct thread *td)
1661 {
1662         sigset_t tmpset;
1663         struct proc *p;
1664         struct lwp *lp;
1665
1666         if (rep && (rep->r_flags & R_SOFTTERM))
1667                 return (EINTR);
1668         /* Terminate all requests while attempting a forced unmount. */
1669         if (nmp->nm_mountp->mnt_kern_flag & MNTK_UNMOUNTF)
1670                 return (EINTR);
1671         if (!(nmp->nm_flag & NFSMNT_INT))
1672                 return (0);
1673         /* td might be NULL YYY */
1674         if (td == NULL || (p = td->td_proc) == NULL)
1675                 return (0);
1676
1677         lp = td->td_lwp;
1678         tmpset = lwp_sigpend(lp);
1679         SIGSETNAND(tmpset, lp->lwp_sigmask);
1680         SIGSETNAND(tmpset, p->p_sigignore);
1681         if (SIGNOTEMPTY(tmpset) && NFSINT_SIGMASK(tmpset))
1682                 return (EINTR);
1683
1684         return (0);
1685 }
1686
1687 /*
1688  * Lock a socket against others.
1689  * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
1690  * and also to avoid race conditions between the processes with nfs requests
1691  * in progress when a reconnect is necessary.
1692  */
1693 int
1694 nfs_sndlock(struct nfsreq *rep)
1695 {
1696         mtx_t mtx = &rep->r_nmp->nm_txlock;
1697         struct thread *td;
1698         int slptimeo;
1699         int slpflag;
1700         int error;
1701
1702         slpflag = 0;
1703         slptimeo = 0;
1704         td = rep->r_td;
1705         if (rep->r_nmp->nm_flag & NFSMNT_INT)
1706                 slpflag = PCATCH;
1707
1708         while ((error = mtx_lock_ex_try(mtx)) != 0) {
1709                 if (nfs_sigintr(rep->r_nmp, rep, td)) {
1710                         error = EINTR;
1711                         break;
1712                 }
1713                 error = mtx_lock_ex(mtx, "nfsndlck", slpflag, slptimeo);
1714                 if (error == 0)
1715                         break;
1716                 if (slpflag == PCATCH) {
1717                         slpflag = 0;
1718                         slptimeo = 2 * hz;
1719                 }
1720         }
1721         /* Always fail if our request has been cancelled. */
1722         if (rep->r_flags & R_SOFTTERM) {
1723                 if (error == 0)
1724                         mtx_unlock(mtx);
1725                 error = EINTR;
1726         }
1727         return (error);
1728 }
1729
1730 /*
1731  * Unlock the stream socket for others.
1732  */
1733 void
1734 nfs_sndunlock(struct nfsreq *rep)
1735 {
1736         mtx_t mtx = &rep->r_nmp->nm_txlock;
1737
1738         mtx_unlock(mtx);
1739 }
1740
1741 static int
1742 nfs_rcvlock(struct nfsreq *rep)
1743 {
1744         mtx_t mtx = &rep->r_nmp->nm_rxlock;
1745         int slpflag;
1746         int slptimeo;
1747         int error;
1748
1749         /*
1750          * Unconditionally check for completion in case another nfsiod
1751          * get the packet while the caller was blocked, before the caller
1752          * called us.  Packet reception is handled by mainline code which
1753          * is protected by the BGL at the moment.
1754          *
1755          * We do not strictly need the second check just before the
1756          * tsleep(), but it's good defensive programming.
1757          */
1758         if (rep->r_mrep != NULL)
1759                 return (EALREADY);
1760
1761         if (rep->r_nmp->nm_flag & NFSMNT_INT)
1762                 slpflag = PCATCH;
1763         else
1764                 slpflag = 0;
1765         slptimeo = 0;
1766
1767         while ((error = mtx_lock_ex_try(mtx)) != 0) {
1768                 if (nfs_sigintr(rep->r_nmp, rep, rep->r_td)) {
1769                         error = EINTR;
1770                         break;
1771                 }
1772                 if (rep->r_mrep != NULL) {
1773                         error = EALREADY;
1774                         break;
1775                 }
1776
1777                 /*
1778                  * NOTE: can return ENOLCK, but in that case rep->r_mrep
1779                  *       will already be set.
1780                  */
1781                 error = mtx_lock_ex_link(mtx, &rep->r_link, "nfsrcvlk",
1782                                          slpflag, slptimeo);
1783                 if (error == 0)
1784                         break;
1785
1786                 /*
1787                  * If our reply was recieved while we were sleeping,
1788                  * then just return without taking the lock to avoid a
1789                  * situation where a single iod could 'capture' the
1790                  * recieve lock.
1791                  */
1792                 if (rep->r_mrep != NULL) {
1793                         error = EALREADY;
1794                         break;
1795                 }
1796                 if (slpflag == PCATCH) {
1797                         slpflag = 0;
1798                         slptimeo = 2 * hz;
1799                 }
1800         }
1801         if (error == 0) {
1802                 if (rep->r_mrep != NULL) {
1803                         error = EALREADY;
1804                         mtx_unlock(mtx);
1805                 }
1806         }
1807         return (error);
1808 }
1809
1810 /*
1811  * Unlock the stream socket for others.
1812  */
1813 static void
1814 nfs_rcvunlock(struct nfsreq *rep)
1815 {
1816         mtx_t mtx = &rep->r_nmp->nm_rxlock;
1817
1818         mtx_unlock(mtx);
1819 }
1820
1821 /*
1822  *      nfs_realign:
1823  *
1824  *      Check for badly aligned mbuf data and realign by copying the unaligned
1825  *      portion of the data into a new mbuf chain and freeing the portions
1826  *      of the old chain that were replaced.
1827  *
1828  *      We cannot simply realign the data within the existing mbuf chain
1829  *      because the underlying buffers may contain other rpc commands and
1830  *      we cannot afford to overwrite them.
1831  *
1832  *      We would prefer to avoid this situation entirely.  The situation does
1833  *      not occur with NFS/UDP and is supposed to only occassionally occur
1834  *      with TCP.  Use vfs.nfs.realign_count and realign_test to check this.
1835  */
1836 static void
1837 nfs_realign(struct mbuf **pm, int hsiz)
1838 {
1839         struct mbuf *m;
1840         struct mbuf *n = NULL;
1841         int off = 0;
1842
1843         ++nfs_realign_test;
1844
1845         while ((m = *pm) != NULL) {
1846                 if ((m->m_len & 0x3) || (mtod(m, intptr_t) & 0x3)) {
1847                         n = m_getl(m->m_len, MB_WAIT, MT_DATA, 0, NULL);
1848                         n->m_len = 0;
1849                         break;
1850                 }
1851                 pm = &m->m_next;
1852         }
1853
1854         /*
1855          * If n is non-NULL, loop on m copying data, then replace the
1856          * portion of the chain that had to be realigned.
1857          */
1858         if (n != NULL) {
1859                 ++nfs_realign_count;
1860                 while (m) {
1861                         m_copyback(n, off, m->m_len, mtod(m, caddr_t));
1862                         off += m->m_len;
1863                         m = m->m_next;
1864                 }
1865                 m_freem(*pm);
1866                 *pm = n;
1867         }
1868 }
1869
1870 #ifndef NFS_NOSERVER
1871
1872 /*
1873  * Parse an RPC request
1874  * - verify it
1875  * - fill in the cred struct.
1876  */
1877 int
1878 nfs_getreq(struct nfsrv_descript *nd, struct nfsd *nfsd, int has_header)
1879 {
1880         int len, i;
1881         u_int32_t *tl;
1882         int32_t t1;
1883         struct uio uio;
1884         struct iovec iov;
1885         caddr_t dpos, cp2, cp;
1886         u_int32_t nfsvers, auth_type;
1887         uid_t nickuid;
1888         int error = 0, ticklen;
1889         struct mbuf *mrep, *md;
1890         struct nfsuid *nuidp;
1891         struct timeval tvin, tvout;
1892 #if 0                           /* until encrypted keys are implemented */
1893         NFSKERBKEYSCHED_T keys; /* stores key schedule */
1894 #endif
1895
1896         mrep = nd->nd_mrep;
1897         md = nd->nd_md;
1898         dpos = nd->nd_dpos;
1899         if (has_header) {
1900                 nfsm_dissect(tl, u_int32_t *, 10 * NFSX_UNSIGNED);
1901                 nd->nd_retxid = fxdr_unsigned(u_int32_t, *tl++);
1902                 if (*tl++ != rpc_call) {
1903                         m_freem(mrep);
1904                         return (EBADRPC);
1905                 }
1906         } else
1907                 nfsm_dissect(tl, u_int32_t *, 8 * NFSX_UNSIGNED);
1908         nd->nd_repstat = 0;
1909         nd->nd_flag = 0;
1910         if (*tl++ != rpc_vers) {
1911                 nd->nd_repstat = ERPCMISMATCH;
1912                 nd->nd_procnum = NFSPROC_NOOP;
1913                 return (0);
1914         }
1915         if (*tl != nfs_prog) {
1916                 nd->nd_repstat = EPROGUNAVAIL;
1917                 nd->nd_procnum = NFSPROC_NOOP;
1918                 return (0);
1919         }
1920         tl++;
1921         nfsvers = fxdr_unsigned(u_int32_t, *tl++);
1922         if (nfsvers < NFS_VER2 || nfsvers > NFS_VER3) {
1923                 nd->nd_repstat = EPROGMISMATCH;
1924                 nd->nd_procnum = NFSPROC_NOOP;
1925                 return (0);
1926         }
1927         if (nfsvers == NFS_VER3)
1928                 nd->nd_flag = ND_NFSV3;
1929         nd->nd_procnum = fxdr_unsigned(u_int32_t, *tl++);
1930         if (nd->nd_procnum == NFSPROC_NULL)
1931                 return (0);
1932         if (nd->nd_procnum >= NFS_NPROCS ||
1933                 (nd->nd_procnum >= NQNFSPROC_GETLEASE) ||
1934                 (!nd->nd_flag && nd->nd_procnum > NFSV2PROC_STATFS)) {
1935                 nd->nd_repstat = EPROCUNAVAIL;
1936                 nd->nd_procnum = NFSPROC_NOOP;
1937                 return (0);
1938         }
1939         if ((nd->nd_flag & ND_NFSV3) == 0)
1940                 nd->nd_procnum = nfsv3_procid[nd->nd_procnum];
1941         auth_type = *tl++;
1942         len = fxdr_unsigned(int, *tl++);
1943         if (len < 0 || len > RPCAUTH_MAXSIZ) {
1944                 m_freem(mrep);
1945                 return (EBADRPC);
1946         }
1947
1948         nd->nd_flag &= ~ND_KERBAUTH;
1949         /*
1950          * Handle auth_unix or auth_kerb.
1951          */
1952         if (auth_type == rpc_auth_unix) {
1953                 len = fxdr_unsigned(int, *++tl);
1954                 if (len < 0 || len > NFS_MAXNAMLEN) {
1955                         m_freem(mrep);
1956                         return (EBADRPC);
1957                 }
1958                 nfsm_adv(nfsm_rndup(len));
1959                 nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
1960                 bzero((caddr_t)&nd->nd_cr, sizeof (struct ucred));
1961                 nd->nd_cr.cr_ref = 1;
1962                 nd->nd_cr.cr_uid = fxdr_unsigned(uid_t, *tl++);
1963                 nd->nd_cr.cr_gid = fxdr_unsigned(gid_t, *tl++);
1964                 len = fxdr_unsigned(int, *tl);
1965                 if (len < 0 || len > RPCAUTH_UNIXGIDS) {
1966                         m_freem(mrep);
1967                         return (EBADRPC);
1968                 }
1969                 nfsm_dissect(tl, u_int32_t *, (len + 2) * NFSX_UNSIGNED);
1970                 for (i = 1; i <= len; i++)
1971                     if (i < NGROUPS)
1972                         nd->nd_cr.cr_groups[i] = fxdr_unsigned(gid_t, *tl++);
1973                     else
1974                         tl++;
1975                 nd->nd_cr.cr_ngroups = (len >= NGROUPS) ? NGROUPS : (len + 1);
1976                 if (nd->nd_cr.cr_ngroups > 1)
1977                     nfsrvw_sort(nd->nd_cr.cr_groups, nd->nd_cr.cr_ngroups);
1978                 len = fxdr_unsigned(int, *++tl);
1979                 if (len < 0 || len > RPCAUTH_MAXSIZ) {
1980                         m_freem(mrep);
1981                         return (EBADRPC);
1982                 }
1983                 if (len > 0)
1984                         nfsm_adv(nfsm_rndup(len));
1985         } else if (auth_type == rpc_auth_kerb) {
1986                 switch (fxdr_unsigned(int, *tl++)) {
1987                 case RPCAKN_FULLNAME:
1988                         ticklen = fxdr_unsigned(int, *tl);
1989                         *((u_int32_t *)nfsd->nfsd_authstr) = *tl;
1990                         uio.uio_resid = nfsm_rndup(ticklen) + NFSX_UNSIGNED;
1991                         nfsd->nfsd_authlen = uio.uio_resid + NFSX_UNSIGNED;
1992                         if (uio.uio_resid > (len - 2 * NFSX_UNSIGNED)) {
1993                                 m_freem(mrep);
1994                                 return (EBADRPC);
1995                         }
1996                         uio.uio_offset = 0;
1997                         uio.uio_iov = &iov;
1998                         uio.uio_iovcnt = 1;
1999                         uio.uio_segflg = UIO_SYSSPACE;
2000                         iov.iov_base = (caddr_t)&nfsd->nfsd_authstr[4];
2001                         iov.iov_len = RPCAUTH_MAXSIZ - 4;
2002                         nfsm_mtouio(&uio, uio.uio_resid);
2003                         nfsm_dissect(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
2004                         if (*tl++ != rpc_auth_kerb ||
2005                                 fxdr_unsigned(int, *tl) != 4 * NFSX_UNSIGNED) {
2006                                 kprintf("Bad kerb verifier\n");
2007                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2008                                 nd->nd_procnum = NFSPROC_NOOP;
2009                                 return (0);
2010                         }
2011                         nfsm_dissect(cp, caddr_t, 4 * NFSX_UNSIGNED);
2012                         tl = (u_int32_t *)cp;
2013                         if (fxdr_unsigned(int, *tl) != RPCAKN_FULLNAME) {
2014                                 kprintf("Not fullname kerb verifier\n");
2015                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2016                                 nd->nd_procnum = NFSPROC_NOOP;
2017                                 return (0);
2018                         }
2019                         cp += NFSX_UNSIGNED;
2020                         bcopy(cp, nfsd->nfsd_verfstr, 3 * NFSX_UNSIGNED);
2021                         nfsd->nfsd_verflen = 3 * NFSX_UNSIGNED;
2022                         nd->nd_flag |= ND_KERBFULL;
2023                         nfsd->nfsd_flag |= NFSD_NEEDAUTH;
2024                         break;
2025                 case RPCAKN_NICKNAME:
2026                         if (len != 2 * NFSX_UNSIGNED) {
2027                                 kprintf("Kerb nickname short\n");
2028                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADCRED);
2029                                 nd->nd_procnum = NFSPROC_NOOP;
2030                                 return (0);
2031                         }
2032                         nickuid = fxdr_unsigned(uid_t, *tl);
2033                         nfsm_dissect(tl, u_int32_t *, 2 * NFSX_UNSIGNED);
2034                         if (*tl++ != rpc_auth_kerb ||
2035                                 fxdr_unsigned(int, *tl) != 3 * NFSX_UNSIGNED) {
2036                                 kprintf("Kerb nick verifier bad\n");
2037                                 nd->nd_repstat = (NFSERR_AUTHERR|AUTH_BADVERF);
2038                                 nd->nd_procnum = NFSPROC_NOOP;
2039                                 return (0);
2040                         }
2041                         nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
2042                         tvin.tv_sec = *tl++;
2043                         tvin.tv_usec = *tl;
2044
2045                         for (nuidp = NUIDHASH(nfsd->nfsd_slp,nickuid)->lh_first;
2046                             nuidp != 0; nuidp = nuidp->nu_hash.le_next) {
2047                                 if (nuidp->nu_cr.cr_uid == nickuid &&
2048                                     (!nd->nd_nam2 ||
2049                                      netaddr_match(NU_NETFAM(nuidp),
2050                                       &nuidp->nu_haddr, nd->nd_nam2)))
2051                                         break;
2052                         }
2053                         if (!nuidp) {
2054                                 nd->nd_repstat =
2055                                         (NFSERR_AUTHERR|AUTH_REJECTCRED);
2056                                 nd->nd_procnum = NFSPROC_NOOP;
2057                                 return (0);
2058                         }
2059
2060                         /*
2061                          * Now, decrypt the timestamp using the session key
2062                          * and validate it.
2063                          */
2064 #ifdef NFSKERB
2065                         XXX
2066 #endif
2067
2068                         tvout.tv_sec = fxdr_unsigned(long, tvout.tv_sec);
2069                         tvout.tv_usec = fxdr_unsigned(long, tvout.tv_usec);
2070                         if (nuidp->nu_expire < time_second ||
2071                             nuidp->nu_timestamp.tv_sec > tvout.tv_sec ||
2072                             (nuidp->nu_timestamp.tv_sec == tvout.tv_sec &&
2073                              nuidp->nu_timestamp.tv_usec > tvout.tv_usec)) {
2074                                 nuidp->nu_expire = 0;
2075                                 nd->nd_repstat =
2076                                     (NFSERR_AUTHERR|AUTH_REJECTVERF);
2077                                 nd->nd_procnum = NFSPROC_NOOP;
2078                                 return (0);
2079                         }
2080                         nfsrv_setcred(&nuidp->nu_cr, &nd->nd_cr);
2081                         nd->nd_flag |= ND_KERBNICK;
2082                 };
2083         } else {
2084                 nd->nd_repstat = (NFSERR_AUTHERR | AUTH_REJECTCRED);
2085                 nd->nd_procnum = NFSPROC_NOOP;
2086                 return (0);
2087         }
2088
2089         nd->nd_md = md;
2090         nd->nd_dpos = dpos;
2091         return (0);
2092 nfsmout:
2093         return (error);
2094 }
2095
2096 #endif
2097
2098 /*
2099  * Send a message to the originating process's terminal.  The thread and/or
2100  * process may be NULL.  YYY the thread should not be NULL but there may
2101  * still be some uio_td's that are still being passed as NULL through to
2102  * nfsm_request().
2103  */
2104 static int
2105 nfs_msg(struct thread *td, char *server, char *msg)
2106 {
2107         tpr_t tpr;
2108
2109         if (td && td->td_proc)
2110                 tpr = tprintf_open(td->td_proc);
2111         else
2112                 tpr = NULL;
2113         tprintf(tpr, "nfs server %s: %s\n", server, msg);
2114         tprintf_close(tpr);
2115         return (0);
2116 }
2117
2118 #ifndef NFS_NOSERVER
2119 /*
2120  * Socket upcall routine for the nfsd sockets.
2121  * The caddr_t arg is a pointer to the "struct nfssvc_sock".
2122  * Essentially do as much as possible non-blocking, else punt and it will
2123  * be called with MB_WAIT from an nfsd.
2124  */
2125 void
2126 nfsrv_rcv(struct socket *so, void *arg, int waitflag)
2127 {
2128         struct nfssvc_sock *slp = (struct nfssvc_sock *)arg;
2129         struct mbuf *m;
2130         struct sockaddr *nam;
2131         struct sockbuf sio;
2132         int flags, error;
2133         int nparallel_wakeup = 0;
2134
2135         if ((slp->ns_flag & SLP_VALID) == 0)
2136                 return;
2137
2138         /*
2139          * Do not allow an infinite number of completed RPC records to build
2140          * up before we stop reading data from the socket.  Otherwise we could
2141          * end up holding onto an unreasonable number of mbufs for requests
2142          * waiting for service.
2143          *
2144          * This should give pretty good feedback to the TCP
2145          * layer and prevents a memory crunch for other protocols.
2146          *
2147          * Note that the same service socket can be dispatched to several
2148          * nfs servers simultaniously.
2149          *
2150          * the tcp protocol callback calls us with MB_DONTWAIT.
2151          * nfsd calls us with MB_WAIT (typically).
2152          */
2153         if (waitflag == MB_DONTWAIT && slp->ns_numrec >= nfsd_waiting / 2 + 1) {
2154                 slp->ns_flag |= SLP_NEEDQ;
2155                 goto dorecs;
2156         }
2157
2158         /*
2159          * Handle protocol specifics to parse an RPC request.  We always
2160          * pull from the socket using non-blocking I/O.
2161          */
2162         if (so->so_type == SOCK_STREAM) {
2163                 /*
2164                  * The data has to be read in an orderly fashion from a TCP
2165                  * stream, unlike a UDP socket.  It is possible for soreceive
2166                  * and/or nfsrv_getstream() to block, so make sure only one
2167                  * entity is messing around with the TCP stream at any given
2168                  * moment.  The receive sockbuf's lock in soreceive is not
2169                  * sufficient.
2170                  *
2171                  * Note that this procedure can be called from any number of
2172                  * NFS severs *OR* can be upcalled directly from a TCP
2173                  * protocol thread.
2174                  */
2175                 if (slp->ns_flag & SLP_GETSTREAM) {
2176                         slp->ns_flag |= SLP_NEEDQ;
2177                         goto dorecs;
2178                 }
2179                 slp->ns_flag |= SLP_GETSTREAM;
2180
2181                 /*
2182                  * Do soreceive().  Pull out as much data as possible without
2183                  * blocking.
2184                  */
2185                 sbinit(&sio, 1000000000);
2186                 flags = MSG_DONTWAIT;
2187                 error = so_pru_soreceive(so, &nam, NULL, &sio, NULL, &flags);
2188                 if (error || sio.sb_mb == NULL) {
2189                         if (error == EWOULDBLOCK)
2190                                 slp->ns_flag |= SLP_NEEDQ;
2191                         else
2192                                 slp->ns_flag |= SLP_DISCONN;
2193                         slp->ns_flag &= ~SLP_GETSTREAM;
2194                         goto dorecs;
2195                 }
2196                 m = sio.sb_mb;
2197                 if (slp->ns_rawend) {
2198                         slp->ns_rawend->m_next = m;
2199                         slp->ns_cc += sio.sb_cc;
2200                 } else {
2201                         slp->ns_raw = m;
2202                         slp->ns_cc = sio.sb_cc;
2203                 }
2204                 while (m->m_next)
2205                         m = m->m_next;
2206                 slp->ns_rawend = m;
2207
2208                 /*
2209                  * Now try and parse as many record(s) as we can out of the
2210                  * raw stream data.
2211                  */
2212                 error = nfsrv_getstream(slp, waitflag, &nparallel_wakeup);
2213                 if (error) {
2214                         if (error == EPERM)
2215                                 slp->ns_flag |= SLP_DISCONN;
2216                         else
2217                                 slp->ns_flag |= SLP_NEEDQ;
2218                 }
2219                 slp->ns_flag &= ~SLP_GETSTREAM;
2220         } else {
2221                 /*
2222                  * For UDP soreceive typically pulls just one packet, loop
2223                  * to get the whole batch.
2224                  */
2225                 do {
2226                         sbinit(&sio, 1000000000);
2227                         flags = MSG_DONTWAIT;
2228                         error = so_pru_soreceive(so, &nam, NULL, &sio,
2229                                                  NULL, &flags);
2230                         if (sio.sb_mb) {
2231                                 struct nfsrv_rec *rec;
2232                                 int mf = (waitflag & MB_DONTWAIT) ?
2233                                             M_NOWAIT : M_WAITOK;
2234                                 rec = kmalloc(sizeof(struct nfsrv_rec),
2235                                              M_NFSRVDESC, mf);
2236                                 if (!rec) {
2237                                         if (nam)
2238                                                 FREE(nam, M_SONAME);
2239                                         m_freem(sio.sb_mb);
2240                                         continue;
2241                                 }
2242                                 nfs_realign(&sio.sb_mb, 10 * NFSX_UNSIGNED);
2243                                 rec->nr_address = nam;
2244                                 rec->nr_packet = sio.sb_mb;
2245                                 STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2246                                 ++slp->ns_numrec;
2247                                 ++nparallel_wakeup;
2248                         }
2249                         if (error) {
2250                                 if ((so->so_proto->pr_flags & PR_CONNREQUIRED)
2251                                         && error != EWOULDBLOCK) {
2252                                         slp->ns_flag |= SLP_DISCONN;
2253                                         goto dorecs;
2254                                 }
2255                         }
2256                 } while (sio.sb_mb);
2257         }
2258
2259         /*
2260          * If we were upcalled from the tcp protocol layer and we have
2261          * fully parsed records ready to go, or there is new data pending,
2262          * or something went wrong, try to wake up an nfsd thread to deal
2263          * with it.
2264          */
2265 dorecs:
2266         if (waitflag == MB_DONTWAIT && (slp->ns_numrec > 0
2267              || (slp->ns_flag & (SLP_NEEDQ | SLP_DISCONN)))) {
2268                 nfsrv_wakenfsd(slp, nparallel_wakeup);
2269         }
2270 }
2271
2272 /*
2273  * Try and extract an RPC request from the mbuf data list received on a
2274  * stream socket. The "waitflag" argument indicates whether or not it
2275  * can sleep.
2276  */
2277 static int
2278 nfsrv_getstream(struct nfssvc_sock *slp, int waitflag, int *countp)
2279 {
2280         struct mbuf *m, **mpp;
2281         char *cp1, *cp2;
2282         int len;
2283         struct mbuf *om, *m2, *recm;
2284         u_int32_t recmark;
2285
2286         for (;;) {
2287             if (slp->ns_reclen == 0) {
2288                 if (slp->ns_cc < NFSX_UNSIGNED)
2289                         return (0);
2290                 m = slp->ns_raw;
2291                 if (m->m_len >= NFSX_UNSIGNED) {
2292                         bcopy(mtod(m, caddr_t), (caddr_t)&recmark, NFSX_UNSIGNED);
2293                         m->m_data += NFSX_UNSIGNED;
2294                         m->m_len -= NFSX_UNSIGNED;
2295                 } else {
2296                         cp1 = (caddr_t)&recmark;
2297                         cp2 = mtod(m, caddr_t);
2298                         while (cp1 < ((caddr_t)&recmark) + NFSX_UNSIGNED) {
2299                                 while (m->m_len == 0) {
2300                                         m = m->m_next;
2301                                         cp2 = mtod(m, caddr_t);
2302                                 }
2303                                 *cp1++ = *cp2++;
2304                                 m->m_data++;
2305                                 m->m_len--;
2306                         }
2307                 }
2308                 slp->ns_cc -= NFSX_UNSIGNED;
2309                 recmark = ntohl(recmark);
2310                 slp->ns_reclen = recmark & ~0x80000000;
2311                 if (recmark & 0x80000000)
2312                         slp->ns_flag |= SLP_LASTFRAG;
2313                 else
2314                         slp->ns_flag &= ~SLP_LASTFRAG;
2315                 if (slp->ns_reclen > NFS_MAXPACKET || slp->ns_reclen <= 0) {
2316                         log(LOG_ERR, "%s (%d) from nfs client\n",
2317                             "impossible packet length",
2318                             slp->ns_reclen);
2319                         return (EPERM);
2320                 }
2321             }
2322
2323             /*
2324              * Now get the record part.
2325              *
2326              * Note that slp->ns_reclen may be 0.  Linux sometimes
2327              * generates 0-length RPCs
2328              */
2329             recm = NULL;
2330             if (slp->ns_cc == slp->ns_reclen) {
2331                 recm = slp->ns_raw;
2332                 slp->ns_raw = slp->ns_rawend = NULL;
2333                 slp->ns_cc = slp->ns_reclen = 0;
2334             } else if (slp->ns_cc > slp->ns_reclen) {
2335                 len = 0;
2336                 m = slp->ns_raw;
2337                 om = NULL;
2338
2339                 while (len < slp->ns_reclen) {
2340                         if ((len + m->m_len) > slp->ns_reclen) {
2341                                 m2 = m_copym(m, 0, slp->ns_reclen - len,
2342                                         waitflag);
2343                                 if (m2) {
2344                                         if (om) {
2345                                                 om->m_next = m2;
2346                                                 recm = slp->ns_raw;
2347                                         } else
2348                                                 recm = m2;
2349                                         m->m_data += slp->ns_reclen - len;
2350                                         m->m_len -= slp->ns_reclen - len;
2351                                         len = slp->ns_reclen;
2352                                 } else {
2353                                         return (EWOULDBLOCK);
2354                                 }
2355                         } else if ((len + m->m_len) == slp->ns_reclen) {
2356                                 om = m;
2357                                 len += m->m_len;
2358                                 m = m->m_next;
2359                                 recm = slp->ns_raw;
2360                                 om->m_next = NULL;
2361                         } else {
2362                                 om = m;
2363                                 len += m->m_len;
2364                                 m = m->m_next;
2365                         }
2366                 }
2367                 slp->ns_raw = m;
2368                 slp->ns_cc -= len;
2369                 slp->ns_reclen = 0;
2370             } else {
2371                 return (0);
2372             }
2373
2374             /*
2375              * Accumulate the fragments into a record.
2376              */
2377             mpp = &slp->ns_frag;
2378             while (*mpp)
2379                 mpp = &((*mpp)->m_next);
2380             *mpp = recm;
2381             if (slp->ns_flag & SLP_LASTFRAG) {
2382                 struct nfsrv_rec *rec;
2383                 int mf = (waitflag & MB_DONTWAIT) ? M_NOWAIT : M_WAITOK;
2384                 rec = kmalloc(sizeof(struct nfsrv_rec), M_NFSRVDESC, mf);
2385                 if (!rec) {
2386                     m_freem(slp->ns_frag);
2387                 } else {
2388                     nfs_realign(&slp->ns_frag, 10 * NFSX_UNSIGNED);
2389                     rec->nr_address = NULL;
2390                     rec->nr_packet = slp->ns_frag;
2391                     STAILQ_INSERT_TAIL(&slp->ns_rec, rec, nr_link);
2392                     ++slp->ns_numrec;
2393                     ++*countp;
2394                 }
2395                 slp->ns_frag = NULL;
2396             }
2397         }
2398 }
2399
2400 /*
2401  * Parse an RPC header.
2402  */
2403 int
2404 nfsrv_dorec(struct nfssvc_sock *slp, struct nfsd *nfsd,
2405             struct nfsrv_descript **ndp)
2406 {
2407         struct nfsrv_rec *rec;
2408         struct mbuf *m;
2409         struct sockaddr *nam;
2410         struct nfsrv_descript *nd;
2411         int error;
2412
2413         *ndp = NULL;
2414         if ((slp->ns_flag & SLP_VALID) == 0 || !STAILQ_FIRST(&slp->ns_rec))
2415                 return (ENOBUFS);
2416         rec = STAILQ_FIRST(&slp->ns_rec);
2417         STAILQ_REMOVE_HEAD(&slp->ns_rec, nr_link);
2418         KKASSERT(slp->ns_numrec > 0);
2419         --slp->ns_numrec;
2420         nam = rec->nr_address;
2421         m = rec->nr_packet;
2422         kfree(rec, M_NFSRVDESC);
2423         MALLOC(nd, struct nfsrv_descript *, sizeof (struct nfsrv_descript),
2424                 M_NFSRVDESC, M_WAITOK);
2425         nd->nd_md = nd->nd_mrep = m;
2426         nd->nd_nam2 = nam;
2427         nd->nd_dpos = mtod(m, caddr_t);
2428         error = nfs_getreq(nd, nfsd, TRUE);
2429         if (error) {
2430                 if (nam) {
2431                         FREE(nam, M_SONAME);
2432                 }
2433                 kfree((caddr_t)nd, M_NFSRVDESC);
2434                 return (error);
2435         }
2436         *ndp = nd;
2437         nfsd->nfsd_nd = nd;
2438         return (0);
2439 }
2440
2441 /*
2442  * Try to assign service sockets to nfsd threads based on the number
2443  * of new rpc requests that have been queued on the service socket.
2444  *
2445  * If no nfsd's are available or additonal requests are pending, set the
2446  * NFSD_CHECKSLP flag so that one of the running nfsds will go look for
2447  * the work in the nfssvc_sock list when it is finished processing its
2448  * current work.  This flag is only cleared when an nfsd can not find
2449  * any new work to perform.
2450  */
2451 void
2452 nfsrv_wakenfsd(struct nfssvc_sock *slp, int nparallel)
2453 {
2454         struct nfsd *nd;
2455
2456         if ((slp->ns_flag & SLP_VALID) == 0)
2457                 return;
2458         if (nparallel <= 1)
2459                 nparallel = 1;
2460         TAILQ_FOREACH(nd, &nfsd_head, nfsd_chain) {
2461                 if (nd->nfsd_flag & NFSD_WAITING) {
2462                         nd->nfsd_flag &= ~NFSD_WAITING;
2463                         if (nd->nfsd_slp)
2464                                 panic("nfsd wakeup");
2465                         slp->ns_sref++;
2466                         nd->nfsd_slp = slp;
2467                         wakeup((caddr_t)nd);
2468                         if (--nparallel == 0)
2469                                 break;
2470                 }
2471         }
2472         if (nparallel) {
2473                 slp->ns_flag |= SLP_DOREC;
2474                 nfsd_head_flag |= NFSD_CHECKSLP;
2475         }
2476 }
2477 #endif /* NFS_NOSERVER */