sys/vfs/nfs/nfs_socket.c

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