Home | History | Annotate | Line # | Download | only in nfs
nfs_syscalls.c revision 1.132
      1  1.132        ad /*	$NetBSD: nfs_syscalls.c,v 1.132 2008/03/21 21:55:01 ad Exp $	*/
      2   1.11       cgd 
      3    1.1       cgd /*
      4   1.10   mycroft  * Copyright (c) 1989, 1993
      5   1.10   mycroft  *	The Regents of the University of California.  All rights reserved.
      6    1.1       cgd  *
      7    1.1       cgd  * This code is derived from software contributed to Berkeley by
      8    1.1       cgd  * Rick Macklem at The University of Guelph.
      9    1.1       cgd  *
     10    1.1       cgd  * Redistribution and use in source and binary forms, with or without
     11    1.1       cgd  * modification, are permitted provided that the following conditions
     12    1.1       cgd  * are met:
     13    1.1       cgd  * 1. Redistributions of source code must retain the above copyright
     14    1.1       cgd  *    notice, this list of conditions and the following disclaimer.
     15    1.1       cgd  * 2. Redistributions in binary form must reproduce the above copyright
     16    1.1       cgd  *    notice, this list of conditions and the following disclaimer in the
     17    1.1       cgd  *    documentation and/or other materials provided with the distribution.
     18   1.70       agc  * 3. Neither the name of the University nor the names of its contributors
     19    1.1       cgd  *    may be used to endorse or promote products derived from this software
     20    1.1       cgd  *    without specific prior written permission.
     21    1.1       cgd  *
     22    1.1       cgd  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     23    1.1       cgd  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     24    1.1       cgd  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     25    1.1       cgd  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     26    1.1       cgd  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     27    1.1       cgd  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     28    1.1       cgd  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     29    1.1       cgd  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     30    1.1       cgd  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     31    1.1       cgd  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     32    1.1       cgd  * SUCH DAMAGE.
     33    1.1       cgd  *
     34   1.19      fvdl  *	@(#)nfs_syscalls.c	8.5 (Berkeley) 3/30/95
     35    1.1       cgd  */
     36   1.49     lukem 
     37   1.49     lukem #include <sys/cdefs.h>
     38  1.132        ad __KERNEL_RCSID(0, "$NetBSD: nfs_syscalls.c,v 1.132 2008/03/21 21:55:01 ad Exp $");
     39   1.28   thorpej 
     40   1.28   thorpej #include "fs_nfs.h"
     41   1.43     bjh21 #include "opt_nfs.h"
     42   1.30   thorpej #include "opt_nfsserver.h"
     43   1.31  jonathan #include "opt_iso.h"
     44   1.41      fvdl #include "opt_inet.h"
     45   1.39    tsarna #include "opt_compat_netbsd.h"
     46    1.1       cgd 
     47    1.7   mycroft #include <sys/param.h>
     48    1.7   mycroft #include <sys/systm.h>
     49    1.7   mycroft #include <sys/kernel.h>
     50    1.7   mycroft #include <sys/file.h>
     51    1.7   mycroft #include <sys/stat.h>
     52    1.7   mycroft #include <sys/vnode.h>
     53    1.7   mycroft #include <sys/mount.h>
     54    1.7   mycroft #include <sys/proc.h>
     55   1.10   mycroft #include <sys/uio.h>
     56    1.7   mycroft #include <sys/malloc.h>
     57  1.124      yamt #include <sys/kmem.h>
     58    1.7   mycroft #include <sys/buf.h>
     59    1.7   mycroft #include <sys/mbuf.h>
     60    1.7   mycroft #include <sys/socket.h>
     61    1.7   mycroft #include <sys/socketvar.h>
     62   1.39    tsarna #include <sys/signalvar.h>
     63    1.7   mycroft #include <sys/domain.h>
     64    1.7   mycroft #include <sys/protosw.h>
     65   1.10   mycroft #include <sys/namei.h>
     66   1.10   mycroft #include <sys/syslog.h>
     67   1.18  christos #include <sys/filedesc.h>
     68   1.39    tsarna #include <sys/kthread.h>
     69   1.92      elad #include <sys/kauth.h>
     70   1.13       cgd #include <sys/syscallargs.h>
     71   1.13       cgd 
     72    1.7   mycroft #include <netinet/in.h>
     73    1.7   mycroft #include <netinet/tcp.h>
     74   1.10   mycroft #ifdef ISO
     75   1.10   mycroft #include <netiso/iso.h>
     76   1.10   mycroft #endif
     77   1.19      fvdl #include <nfs/xdr_subs.h>
     78   1.10   mycroft #include <nfs/rpcv2.h>
     79   1.19      fvdl #include <nfs/nfsproto.h>
     80    1.7   mycroft #include <nfs/nfs.h>
     81   1.19      fvdl #include <nfs/nfsm_subs.h>
     82    1.7   mycroft #include <nfs/nfsrvcache.h>
     83   1.10   mycroft #include <nfs/nfsmount.h>
     84   1.10   mycroft #include <nfs/nfsnode.h>
     85   1.10   mycroft #include <nfs/nfsrtt.h>
     86   1.18  christos #include <nfs/nfs_var.h>
     87    1.1       cgd 
     88    1.1       cgd /* Global defs. */
     89   1.19      fvdl extern int32_t (*nfsrv3_procs[NFS_NPROCS]) __P((struct nfsrv_descript *,
     90   1.19      fvdl 						struct nfssvc_sock *,
     91   1.84  christos 						struct lwp *, struct mbuf **));
     92   1.19      fvdl extern int nfsrvw_procrastinate;
     93   1.51      matt 
     94   1.47       chs struct nfssvc_sock *nfs_udpsock;
     95   1.47       chs #ifdef ISO
     96   1.47       chs struct nfssvc_sock *nfs_cltpsock;
     97   1.47       chs #endif
     98   1.41      fvdl #ifdef INET6
     99   1.41      fvdl struct nfssvc_sock *nfs_udp6sock;
    100   1.41      fvdl #endif
    101   1.10   mycroft int nuidhash_max = NFS_MAXUIDHASH;
    102   1.18  christos #ifdef NFSSERVER
    103   1.10   mycroft static int nfs_numnfsd = 0;
    104   1.10   mycroft static struct nfsdrt nfsdrt;
    105   1.18  christos #endif
    106   1.51      matt 
    107   1.66      yamt #ifdef NFSSERVER
    108  1.111      yamt kmutex_t nfsd_lock;
    109   1.51      matt struct nfssvc_sockhead nfssvc_sockhead;
    110  1.111      yamt kcondvar_t nfsd_initcv;
    111   1.66      yamt struct nfssvc_sockhead nfssvc_sockpending;
    112   1.51      matt struct nfsdhead nfsd_head;
    113   1.66      yamt struct nfsdidlehead nfsd_idle_head;
    114   1.51      matt 
    115   1.51      matt int nfssvc_sockhead_flag;
    116   1.51      matt int nfsd_head_flag;
    117   1.66      yamt #endif
    118    1.1       cgd 
    119   1.24   thorpej #ifdef NFS
    120  1.124      yamt /*
    121  1.124      yamt  * locking order:
    122  1.124      yamt  *	nfs_iodlist_lock -> nid_lock -> nm_lock
    123  1.124      yamt  */
    124  1.123      yamt kmutex_t nfs_iodlist_lock;
    125  1.124      yamt struct nfs_iodlist nfs_iodlist_idle;
    126  1.124      yamt struct nfs_iodlist nfs_iodlist_all;
    127   1.40    tsarna int nfs_niothreads = -1; /* == "0, and has never been set" */
    128   1.23   thorpej #endif
    129   1.19      fvdl 
    130   1.18  christos #ifdef NFSSERVER
    131   1.65      yamt static struct nfssvc_sock *nfsrv_sockalloc __P((void));
    132   1.77      yamt static void nfsrv_sockfree __P((struct nfssvc_sock *));
    133   1.86      yamt static void nfsd_rt __P((int, struct nfsrv_descript *, int));
    134   1.18  christos #endif
    135   1.19      fvdl 
    136    1.1       cgd /*
    137    1.1       cgd  * NFS server system calls
    138    1.1       cgd  */
    139    1.1       cgd 
    140    1.1       cgd 
    141    1.1       cgd /*
    142   1.19      fvdl  * Nfs server pseudo system call for the nfsd's
    143   1.10   mycroft  * Based on the flag value it either:
    144   1.10   mycroft  * - adds a socket to the selection list
    145   1.10   mycroft  * - remains in the kernel as an nfsd
    146   1.10   mycroft  * - remains in the kernel as an nfsiod
    147    1.1       cgd  */
    148   1.18  christos int
    149  1.128       dsl sys_nfssvc(struct lwp *l, const struct sys_nfssvc_args *uap, register_t *retval)
    150   1.15   thorpej {
    151  1.128       dsl 	/* {
    152   1.13       cgd 		syscallarg(int) flag;
    153  1.107  christos 		syscallarg(void *) argp;
    154  1.128       dsl 	} */
    155   1.23   thorpej 	int error;
    156   1.18  christos #ifdef NFSSERVER
    157  1.132        ad 	file_t *fp;
    158   1.10   mycroft 	struct mbuf *nam;
    159   1.10   mycroft 	struct nfsd_args nfsdarg;
    160   1.10   mycroft 	struct nfsd_srvargs nfsd_srvargs, *nsd = &nfsd_srvargs;
    161   1.10   mycroft 	struct nfsd *nfsd;
    162   1.10   mycroft 	struct nfssvc_sock *slp;
    163   1.18  christos 	struct nfsuid *nuidp;
    164   1.18  christos #endif
    165    1.1       cgd 
    166    1.1       cgd 	/*
    167    1.1       cgd 	 * Must be super user
    168    1.1       cgd 	 */
    169  1.131      elad 	error = kauth_authorize_network(l->l_cred, KAUTH_NETWORK_NFS,
    170  1.131      elad 	    KAUTH_REQ_NETWORK_NFS_SVC, NULL, NULL, NULL);
    171   1.71   thorpej 	if (error)
    172    1.1       cgd 		return (error);
    173   1.83   thorpej 
    174   1.83   thorpej 	/* Initialize NFS server / client shared data. */
    175   1.83   thorpej 	nfs_init();
    176   1.83   thorpej 
    177   1.67      yamt #ifdef NFSSERVER
    178  1.111      yamt 	mutex_enter(&nfsd_lock);
    179   1.12   mycroft 	while (nfssvc_sockhead_flag & SLP_INIT) {
    180  1.111      yamt 		cv_wait(&nfsd_initcv, &nfsd_lock);
    181   1.10   mycroft 	}
    182  1.111      yamt 	mutex_exit(&nfsd_lock);
    183   1.67      yamt #endif
    184   1.13       cgd 	if (SCARG(uap, flag) & NFSSVC_BIOD) {
    185   1.39    tsarna #if defined(NFS) && defined(COMPAT_14)
    186  1.118      yamt 		error = kpause("nfsbiod", true, 0, NULL); /* dummy impl */
    187   1.19      fvdl #else
    188   1.10   mycroft 		error = ENOSYS;
    189   1.19      fvdl #endif
    190   1.13       cgd 	} else if (SCARG(uap, flag) & NFSSVC_MNTD) {
    191   1.10   mycroft 		error = ENOSYS;
    192   1.13       cgd 	} else if (SCARG(uap, flag) & NFSSVC_ADDSOCK) {
    193   1.10   mycroft #ifndef NFSSERVER
    194   1.10   mycroft 		error = ENOSYS;
    195   1.19      fvdl #else
    196  1.107  christos 		error = copyin(SCARG(uap, argp), (void *)&nfsdarg,
    197   1.19      fvdl 		    sizeof(nfsdarg));
    198   1.18  christos 		if (error)
    199   1.10   mycroft 			return (error);
    200   1.35   thorpej 		/* getsock() will use the descriptor for us */
    201  1.132        ad 		error = getsock(nfsdarg.sock, &fp);
    202   1.19      fvdl 		if (error)
    203   1.10   mycroft 			return (error);
    204   1.10   mycroft 		/*
    205   1.10   mycroft 		 * Get the client address for connected sockets.
    206   1.10   mycroft 		 */
    207   1.10   mycroft 		if (nfsdarg.name == NULL || nfsdarg.namelen == 0)
    208   1.10   mycroft 			nam = (struct mbuf *)0;
    209   1.18  christos 		else {
    210   1.18  christos 			error = sockargs(&nam, nfsdarg.name, nfsdarg.namelen,
    211   1.19      fvdl 				MT_SONAME);
    212   1.35   thorpej 			if (error) {
    213  1.132        ad 				fd_putfile(nfsdarg.sock);
    214   1.18  christos 				return (error);
    215   1.35   thorpej 			}
    216   1.18  christos 		}
    217   1.10   mycroft 		error = nfssvc_addsock(fp, nam);
    218  1.132        ad 		fd_putfile(nfsdarg.sock);
    219   1.10   mycroft #endif /* !NFSSERVER */
    220   1.82      jmmv 	} else if (SCARG(uap, flag) & NFSSVC_SETEXPORTSLIST) {
    221   1.82      jmmv #ifndef NFSSERVER
    222   1.82      jmmv 		error = ENOSYS;
    223   1.82      jmmv #else
    224   1.82      jmmv 		struct export_args *args;
    225   1.82      jmmv 		struct mountd_exports_list mel;
    226   1.82      jmmv 
    227   1.82      jmmv 		error = copyin(SCARG(uap, argp), &mel, sizeof(mel));
    228   1.82      jmmv 		if (error != 0)
    229   1.82      jmmv 			return error;
    230   1.82      jmmv 
    231   1.82      jmmv 		args = (struct export_args *)malloc(mel.mel_nexports *
    232   1.82      jmmv 		    sizeof(struct export_args), M_TEMP, M_WAITOK);
    233   1.82      jmmv 		error = copyin(mel.mel_exports, args, mel.mel_nexports *
    234   1.82      jmmv 		    sizeof(struct export_args));
    235   1.82      jmmv 		if (error != 0) {
    236   1.82      jmmv 			free(args, M_TEMP);
    237   1.82      jmmv 			return error;
    238   1.82      jmmv 		}
    239   1.82      jmmv 		mel.mel_exports = args;
    240   1.82      jmmv 
    241   1.84  christos 		error = mountd_set_exports_list(&mel, l);
    242   1.82      jmmv 
    243   1.82      jmmv 		free(args, M_TEMP);
    244   1.82      jmmv #endif /* !NFSSERVER */
    245   1.10   mycroft 	} else {
    246   1.10   mycroft #ifndef NFSSERVER
    247   1.10   mycroft 		error = ENOSYS;
    248   1.19      fvdl #else
    249  1.107  christos 		error = copyin(SCARG(uap, argp), (void *)nsd, sizeof (*nsd));
    250   1.18  christos 		if (error)
    251   1.10   mycroft 			return (error);
    252   1.13       cgd 		if ((SCARG(uap, flag) & NFSSVC_AUTHIN) &&
    253   1.19      fvdl 		    ((nfsd = nsd->nsd_nfsd)) != NULL &&
    254  1.127      yamt 		    (nfsd->nfsd_slp->ns_flags & SLP_VALID)) {
    255   1.19      fvdl 			slp = nfsd->nfsd_slp;
    256   1.10   mycroft 
    257   1.10   mycroft 			/*
    258   1.10   mycroft 			 * First check to see if another nfsd has already
    259   1.10   mycroft 			 * added this credential.
    260   1.10   mycroft 			 */
    261   1.92      elad 			LIST_FOREACH(nuidp, NUIDHASH(slp, nsd->nsd_cr.cr_uid),
    262   1.57      yamt 			    nu_hash) {
    263   1.93      yamt 				if (kauth_cred_geteuid(nuidp->nu_cr) ==
    264   1.93      yamt 				    nsd->nsd_cr.cr_uid &&
    265   1.19      fvdl 				    (!nfsd->nfsd_nd->nd_nam2 ||
    266   1.19      fvdl 				     netaddr_match(NU_NETFAM(nuidp),
    267   1.19      fvdl 				     &nuidp->nu_haddr, nfsd->nfsd_nd->nd_nam2)))
    268   1.10   mycroft 					break;
    269   1.10   mycroft 			}
    270   1.19      fvdl 			if (nuidp) {
    271   1.93      yamt 			    kauth_cred_hold(nuidp->nu_cr);
    272   1.93      yamt 			    nfsd->nfsd_nd->nd_cr = nuidp->nu_cr;
    273   1.19      fvdl 			    nfsd->nfsd_nd->nd_flag |= ND_KERBFULL;
    274   1.19      fvdl 			} else {
    275   1.10   mycroft 			    /*
    276   1.10   mycroft 			     * Nope, so we will.
    277   1.10   mycroft 			     */
    278   1.10   mycroft 			    if (slp->ns_numuids < nuidhash_max) {
    279   1.10   mycroft 				slp->ns_numuids++;
    280  1.130      yamt 				nuidp = kmem_alloc(sizeof(*nuidp), KM_SLEEP);
    281   1.10   mycroft 			    } else
    282   1.10   mycroft 				nuidp = (struct nfsuid *)0;
    283  1.127      yamt 			    if ((slp->ns_flags & SLP_VALID) == 0) {
    284   1.10   mycroft 				if (nuidp)
    285  1.130      yamt 				    kmem_free(nuidp, sizeof(*nuidp));
    286   1.10   mycroft 			    } else {
    287   1.10   mycroft 				if (nuidp == (struct nfsuid *)0) {
    288   1.57      yamt 				    nuidp = TAILQ_FIRST(&slp->ns_uidlruhead);
    289   1.12   mycroft 				    LIST_REMOVE(nuidp, nu_hash);
    290   1.12   mycroft 				    TAILQ_REMOVE(&slp->ns_uidlruhead, nuidp,
    291   1.12   mycroft 					nu_lru);
    292   1.19      fvdl 				    if (nuidp->nu_flag & NU_NAM)
    293   1.19      fvdl 					m_freem(nuidp->nu_nam);
    294   1.10   mycroft 			        }
    295   1.19      fvdl 				nuidp->nu_flag = 0;
    296   1.98     pooka 				kauth_uucred_to_cred(nuidp->nu_cr,
    297   1.98     pooka 				    &nsd->nsd_cr);
    298   1.19      fvdl 				nuidp->nu_timestamp = nsd->nsd_timestamp;
    299   1.95    kardel 				nuidp->nu_expire = time_second + nsd->nsd_ttl;
    300   1.19      fvdl 				/*
    301   1.19      fvdl 				 * and save the session key in nu_key.
    302   1.19      fvdl 				 */
    303   1.32     perry 				memcpy(nuidp->nu_key, nsd->nsd_key,
    304   1.32     perry 				    sizeof(nsd->nsd_key));
    305   1.19      fvdl 				if (nfsd->nfsd_nd->nd_nam2) {
    306   1.19      fvdl 				    struct sockaddr_in *saddr;
    307   1.19      fvdl 
    308   1.19      fvdl 				    saddr = mtod(nfsd->nfsd_nd->nd_nam2,
    309   1.19      fvdl 					 struct sockaddr_in *);
    310   1.19      fvdl 				    switch (saddr->sin_family) {
    311   1.19      fvdl 				    case AF_INET:
    312   1.19      fvdl 					nuidp->nu_flag |= NU_INETADDR;
    313   1.19      fvdl 					nuidp->nu_inetaddr =
    314   1.19      fvdl 					     saddr->sin_addr.s_addr;
    315   1.19      fvdl 					break;
    316   1.19      fvdl 				    case AF_ISO:
    317   1.19      fvdl 				    default:
    318   1.19      fvdl 					nuidp->nu_flag |= NU_NAM;
    319   1.19      fvdl 					nuidp->nu_nam = m_copym(
    320   1.19      fvdl 					    nfsd->nfsd_nd->nd_nam2, 0,
    321   1.19      fvdl 					     M_COPYALL, M_WAIT);
    322   1.19      fvdl 					break;
    323   1.19      fvdl 				    };
    324   1.19      fvdl 				}
    325   1.12   mycroft 				TAILQ_INSERT_TAIL(&slp->ns_uidlruhead, nuidp,
    326   1.19      fvdl 					nu_lru);
    327   1.12   mycroft 				LIST_INSERT_HEAD(NUIDHASH(slp, nsd->nsd_uid),
    328   1.19      fvdl 					nuidp, nu_hash);
    329   1.93      yamt 				kauth_cred_hold(nuidp->nu_cr);
    330   1.93      yamt 				nfsd->nfsd_nd->nd_cr = nuidp->nu_cr;
    331   1.19      fvdl 				nfsd->nfsd_nd->nd_flag |= ND_KERBFULL;
    332   1.10   mycroft 			    }
    333   1.10   mycroft 			}
    334   1.10   mycroft 		}
    335   1.13       cgd 		if ((SCARG(uap, flag) & NFSSVC_AUTHINFAIL) &&
    336   1.13       cgd 		    (nfsd = nsd->nsd_nfsd))
    337   1.19      fvdl 			nfsd->nfsd_flag |= NFSD_AUTHFAIL;
    338   1.53   thorpej 		error = nfssvc_nfsd(nsd, SCARG(uap, argp), l);
    339   1.10   mycroft #endif /* !NFSSERVER */
    340   1.10   mycroft 	}
    341   1.10   mycroft 	if (error == EINTR || error == ERESTART)
    342   1.10   mycroft 		error = 0;
    343   1.10   mycroft 	return (error);
    344   1.10   mycroft }
    345   1.10   mycroft 
    346   1.10   mycroft #ifdef NFSSERVER
    347   1.54   thorpej MALLOC_DEFINE(M_NFSD, "NFS daemon", "Nfs server daemon structure");
    348   1.54   thorpej 
    349   1.65      yamt static struct nfssvc_sock *
    350   1.65      yamt nfsrv_sockalloc()
    351   1.65      yamt {
    352   1.65      yamt 	struct nfssvc_sock *slp;
    353   1.78     perry 
    354  1.130      yamt 	slp = kmem_alloc(sizeof(*slp), KM_SLEEP);
    355   1.65      yamt 	memset(slp, 0, sizeof (struct nfssvc_sock));
    356  1.127      yamt 	/* XXX could be IPL_SOFTNET */
    357  1.127      yamt 	mutex_init(&slp->ns_lock, MUTEX_DRIVER, IPL_VM);
    358  1.127      yamt 	mutex_init(&slp->ns_alock, MUTEX_DRIVER, IPL_VM);
    359  1.111      yamt 	cv_init(&slp->ns_cv, "nfsdsock");
    360   1.65      yamt 	TAILQ_INIT(&slp->ns_uidlruhead);
    361   1.86      yamt 	LIST_INIT(&slp->ns_tq);
    362   1.86      yamt 	SIMPLEQ_INIT(&slp->ns_sendq);
    363  1.111      yamt 	mutex_enter(&nfsd_lock);
    364   1.65      yamt 	TAILQ_INSERT_TAIL(&nfssvc_sockhead, slp, ns_chain);
    365  1.111      yamt 	mutex_exit(&nfsd_lock);
    366   1.65      yamt 
    367   1.65      yamt 	return slp;
    368   1.65      yamt }
    369   1.65      yamt 
    370   1.77      yamt static void
    371   1.77      yamt nfsrv_sockfree(struct nfssvc_sock *slp)
    372   1.77      yamt {
    373   1.77      yamt 
    374   1.77      yamt 	KASSERT(slp->ns_so == NULL);
    375   1.77      yamt 	KASSERT(slp->ns_fp == NULL);
    376  1.127      yamt 	KASSERT((slp->ns_flags & SLP_VALID) == 0);
    377  1.111      yamt 	mutex_destroy(&slp->ns_lock);
    378  1.127      yamt 	mutex_destroy(&slp->ns_alock);
    379  1.111      yamt 	cv_destroy(&slp->ns_cv);
    380  1.130      yamt 	kmem_free(slp, sizeof(*slp));
    381   1.77      yamt }
    382   1.77      yamt 
    383   1.10   mycroft /*
    384   1.10   mycroft  * Adds a socket to the list for servicing by nfsds.
    385   1.10   mycroft  */
    386   1.18  christos int
    387   1.10   mycroft nfssvc_addsock(fp, mynam)
    388  1.132        ad 	file_t *fp;
    389   1.10   mycroft 	struct mbuf *mynam;
    390   1.10   mycroft {
    391   1.38  augustss 	struct mbuf *m;
    392   1.38  augustss 	int siz;
    393   1.38  augustss 	struct nfssvc_sock *slp;
    394   1.38  augustss 	struct socket *so;
    395   1.10   mycroft 	struct nfssvc_sock *tslp;
    396   1.10   mycroft 	int error, s;
    397   1.10   mycroft 
    398    1.1       cgd 	so = (struct socket *)fp->f_data;
    399   1.10   mycroft 	tslp = (struct nfssvc_sock *)0;
    400   1.10   mycroft 	/*
    401   1.10   mycroft 	 * Add it to the list, as required.
    402   1.10   mycroft 	 */
    403   1.10   mycroft 	if (so->so_proto->pr_protocol == IPPROTO_UDP) {
    404   1.41      fvdl #ifdef INET6
    405   1.41      fvdl 		if (so->so_proto->pr_domain->dom_family == AF_INET6)
    406   1.41      fvdl 			tslp = nfs_udp6sock;
    407   1.41      fvdl 		else
    408   1.41      fvdl #endif
    409   1.10   mycroft 		tslp = nfs_udpsock;
    410  1.127      yamt 		if (tslp->ns_flags & SLP_VALID) {
    411   1.10   mycroft 			m_freem(mynam);
    412   1.10   mycroft 			return (EPERM);
    413   1.10   mycroft 		}
    414   1.10   mycroft #ifdef ISO
    415   1.10   mycroft 	} else if (so->so_proto->pr_protocol == ISOPROTO_CLTP) {
    416   1.10   mycroft 		tslp = nfs_cltpsock;
    417  1.127      yamt 		if (tslp->ns_flags & SLP_VALID) {
    418   1.10   mycroft 			m_freem(mynam);
    419   1.10   mycroft 			return (EPERM);
    420   1.10   mycroft 		}
    421   1.10   mycroft #endif /* ISO */
    422   1.10   mycroft 	}
    423   1.10   mycroft 	if (so->so_type == SOCK_STREAM)
    424   1.19      fvdl 		siz = NFS_MAXPACKET + sizeof (u_long);
    425   1.10   mycroft 	else
    426    1.1       cgd 		siz = NFS_MAXPACKET;
    427   1.78     perry 	error = soreserve(so, siz, siz);
    428   1.19      fvdl 	if (error) {
    429   1.10   mycroft 		m_freem(mynam);
    430   1.10   mycroft 		return (error);
    431   1.10   mycroft 	}
    432    1.1       cgd 
    433    1.1       cgd 	/*
    434    1.1       cgd 	 * Set protocol specific options { for now TCP only } and
    435    1.1       cgd 	 * reserve some space. For datagram sockets, this can get called
    436    1.1       cgd 	 * repeatedly for the same socket, but that isn't harmful.
    437    1.1       cgd 	 */
    438   1.10   mycroft 	if (so->so_type == SOCK_STREAM) {
    439   1.55      matt 		m = m_get(M_WAIT, MT_SOOPTS);
    440   1.55      matt 		MCLAIM(m, &nfs_mowner);
    441   1.19      fvdl 		*mtod(m, int32_t *) = 1;
    442   1.19      fvdl 		m->m_len = sizeof(int32_t);
    443    1.1       cgd 		sosetopt(so, SOL_SOCKET, SO_KEEPALIVE, m);
    444    1.1       cgd 	}
    445   1.41      fvdl 	if ((so->so_proto->pr_domain->dom_family == AF_INET
    446   1.41      fvdl #ifdef INET6
    447   1.41      fvdl 	    || so->so_proto->pr_domain->dom_family == AF_INET6
    448   1.41      fvdl #endif
    449   1.41      fvdl 	    ) &&
    450   1.10   mycroft 	    so->so_proto->pr_protocol == IPPROTO_TCP) {
    451   1.55      matt 		m = m_get(M_WAIT, MT_SOOPTS);
    452   1.55      matt 		MCLAIM(m, &nfs_mowner);
    453   1.19      fvdl 		*mtod(m, int32_t *) = 1;
    454   1.19      fvdl 		m->m_len = sizeof(int32_t);
    455    1.1       cgd 		sosetopt(so, IPPROTO_TCP, TCP_NODELAY, m);
    456    1.1       cgd 	}
    457    1.1       cgd 	so->so_rcv.sb_flags &= ~SB_NOINTR;
    458    1.1       cgd 	so->so_rcv.sb_timeo = 0;
    459    1.1       cgd 	so->so_snd.sb_flags &= ~SB_NOINTR;
    460    1.1       cgd 	so->so_snd.sb_timeo = 0;
    461  1.111      yamt 	if (tslp) {
    462   1.10   mycroft 		slp = tslp;
    463  1.111      yamt 	} else {
    464   1.65      yamt 		slp = nfsrv_sockalloc();
    465   1.10   mycroft 	}
    466   1.10   mycroft 	slp->ns_so = so;
    467   1.10   mycroft 	slp->ns_nam = mynam;
    468  1.127      yamt 	mutex_enter(&fp->f_lock);
    469   1.10   mycroft 	fp->f_count++;
    470  1.127      yamt 	mutex_exit(&fp->f_lock);
    471   1.10   mycroft 	slp->ns_fp = fp;
    472  1.127      yamt 	slp->ns_flags = SLP_VALID;
    473  1.127      yamt 	slp->ns_aflags = SLP_A_NEEDQ;
    474  1.127      yamt 	slp->ns_gflags = 0;
    475  1.127      yamt 	slp->ns_sflags = 0;
    476  1.129        ad 	KERNEL_LOCK(1, curlwp);
    477   1.14   mycroft 	s = splsoftnet();
    478  1.107  christos 	so->so_upcallarg = (void *)slp;
    479  1.127      yamt 	so->so_upcall = nfsrv_soupcall;
    480   1.29      matt 	so->so_rcv.sb_flags |= SB_UPCALL;
    481  1.127      yamt 	splx(s);
    482  1.129        ad 	KERNEL_UNLOCK_ONE(curlwp);
    483   1.10   mycroft 	nfsrv_wakenfsd(slp);
    484   1.10   mycroft 	return (0);
    485   1.10   mycroft }
    486    1.1       cgd 
    487   1.10   mycroft /*
    488   1.10   mycroft  * Called by nfssvc() for nfsds. Just loops around servicing rpc requests
    489   1.10   mycroft  * until it is killed by a signal.
    490   1.10   mycroft  */
    491   1.18  christos int
    492   1.53   thorpej nfssvc_nfsd(nsd, argp, l)
    493   1.10   mycroft 	struct nfsd_srvargs *nsd;
    494  1.107  christos 	void *argp;
    495   1.53   thorpej 	struct lwp *l;
    496   1.10   mycroft {
    497   1.95    kardel 	struct timeval tv;
    498   1.38  augustss 	struct mbuf *m;
    499   1.38  augustss 	struct nfssvc_sock *slp;
    500   1.19      fvdl 	struct nfsd *nfsd = nsd->nsd_nfsd;
    501   1.19      fvdl 	struct nfsrv_descript *nd = NULL;
    502   1.19      fvdl 	struct mbuf *mreq;
    503   1.19      fvdl 	u_quad_t cur_usec;
    504  1.111      yamt 	int error = 0, cacherep, siz, sotype, writes_todo;
    505   1.53   thorpej 	struct proc *p = l->l_proc;
    506  1.111      yamt 	int s;
    507  1.127      yamt 	bool doreinit;
    508   1.19      fvdl 
    509   1.19      fvdl #ifndef nolint
    510   1.19      fvdl 	cacherep = RC_DOIT;
    511   1.19      fvdl 	writes_todo = 0;
    512   1.19      fvdl #endif
    513  1.117        ad 	uvm_lwp_hold(l);
    514  1.111      yamt 	if (nfsd == NULL) {
    515  1.130      yamt 		nsd->nsd_nfsd = nfsd = kmem_alloc(sizeof(*nfsd), KM_SLEEP);
    516  1.111      yamt 		memset(nfsd, 0, sizeof (struct nfsd));
    517  1.111      yamt 		cv_init(&nfsd->nfsd_cv, "nfsd");
    518   1.19      fvdl 		nfsd->nfsd_procp = p;
    519  1.111      yamt 		mutex_enter(&nfsd_lock);
    520  1.127      yamt 		while ((nfssvc_sockhead_flag & SLP_INIT) != 0) {
    521  1.127      yamt 			KASSERT(nfs_numnfsd == 0);
    522  1.127      yamt 			cv_wait(&nfsd_initcv, &nfsd_lock);
    523  1.127      yamt 		}
    524   1.19      fvdl 		TAILQ_INSERT_TAIL(&nfsd_head, nfsd, nfsd_chain);
    525   1.10   mycroft 		nfs_numnfsd++;
    526  1.111      yamt 		mutex_exit(&nfsd_lock);
    527   1.10   mycroft 	}
    528    1.1       cgd 	/*
    529   1.10   mycroft 	 * Loop getting rpc requests until SIGKILL.
    530    1.1       cgd 	 */
    531    1.1       cgd 	for (;;) {
    532  1.127      yamt 		bool dummy;
    533  1.127      yamt 
    534  1.113      yamt 		if ((curcpu()->ci_schedstate.spc_flags & SPCF_SHOULDYIELD)
    535  1.113      yamt 		    != 0) {
    536  1.113      yamt 			preempt();
    537  1.113      yamt 		}
    538   1.85      yamt 		if (nfsd->nfsd_slp == NULL) {
    539  1.111      yamt 			mutex_enter(&nfsd_lock);
    540   1.85      yamt 			while (nfsd->nfsd_slp == NULL &&
    541   1.12   mycroft 			    (nfsd_head_flag & NFSD_CHECKSLP) == 0) {
    542   1.66      yamt 				SLIST_INSERT_HEAD(&nfsd_idle_head, nfsd,
    543   1.66      yamt 				    nfsd_idle);
    544  1.111      yamt 				error = cv_wait_sig(&nfsd->nfsd_cv, &nfsd_lock);
    545   1.66      yamt 				if (error) {
    546   1.66      yamt 					slp = nfsd->nfsd_slp;
    547   1.66      yamt 					nfsd->nfsd_slp = NULL;
    548   1.66      yamt 					if (!slp)
    549   1.66      yamt 						SLIST_REMOVE(&nfsd_idle_head,
    550   1.66      yamt 						    nfsd, nfsd, nfsd_idle);
    551  1.111      yamt 					mutex_exit(&nfsd_lock);
    552   1.66      yamt 					if (slp) {
    553   1.66      yamt 						nfsrv_wakenfsd(slp);
    554   1.66      yamt 						nfsrv_slpderef(slp);
    555   1.66      yamt 					}
    556   1.10   mycroft 					goto done;
    557   1.66      yamt 				}
    558   1.10   mycroft 			}
    559   1.85      yamt 			if (nfsd->nfsd_slp == NULL &&
    560   1.12   mycroft 			    (nfsd_head_flag & NFSD_CHECKSLP) != 0) {
    561   1.66      yamt 				slp = TAILQ_FIRST(&nfssvc_sockpending);
    562   1.66      yamt 				if (slp) {
    563  1.127      yamt 					KASSERT((slp->ns_gflags & SLP_G_DOREC)
    564  1.121      yamt 					    != 0);
    565   1.66      yamt 					TAILQ_REMOVE(&nfssvc_sockpending, slp,
    566   1.66      yamt 					    ns_pending);
    567  1.127      yamt 					slp->ns_gflags &= ~SLP_G_DOREC;
    568   1.66      yamt 					slp->ns_sref++;
    569   1.66      yamt 					nfsd->nfsd_slp = slp;
    570   1.66      yamt 				} else
    571   1.12   mycroft 					nfsd_head_flag &= ~NFSD_CHECKSLP;
    572   1.10   mycroft 			}
    573  1.116      yamt 			KASSERT(nfsd->nfsd_slp == NULL ||
    574  1.116      yamt 			    nfsd->nfsd_slp->ns_sref > 0);
    575  1.111      yamt 			mutex_exit(&nfsd_lock);
    576   1.85      yamt 			if ((slp = nfsd->nfsd_slp) == NULL)
    577   1.10   mycroft 				continue;
    578  1.127      yamt 			if (slp->ns_flags & SLP_VALID) {
    579  1.127      yamt 				bool more;
    580  1.127      yamt 
    581  1.127      yamt 				if (nfsdsock_testbits(slp, SLP_A_NEEDQ)) {
    582  1.127      yamt 					nfsrv_rcv(slp);
    583   1.10   mycroft 				}
    584  1.127      yamt 				if (nfsdsock_testbits(slp, SLP_A_DISCONN)) {
    585  1.119      yamt 					nfsrv_zapsock(slp);
    586  1.119      yamt 				}
    587  1.127      yamt 				error = nfsrv_dorec(slp, nfsd, &nd, &more);
    588   1.95    kardel 				getmicrotime(&tv);
    589   1.95    kardel 				cur_usec = (u_quad_t)tv.tv_sec * 1000000 +
    590   1.95    kardel 					(u_quad_t)tv.tv_usec;
    591  1.127      yamt 				writes_todo = 0;
    592  1.127      yamt 				if (error) {
    593  1.127      yamt 					struct nfsrv_descript *nd2;
    594  1.127      yamt 
    595  1.127      yamt 					mutex_enter(&nfsd_lock);
    596  1.127      yamt 					nd2 = LIST_FIRST(&slp->ns_tq);
    597  1.127      yamt 					if (nd2 != NULL &&
    598  1.127      yamt 					    nd2->nd_time <= cur_usec) {
    599  1.127      yamt 						error = 0;
    600  1.127      yamt 						cacherep = RC_DOIT;
    601  1.127      yamt 						writes_todo = 1;
    602  1.127      yamt 					}
    603  1.127      yamt 					mutex_exit(&nfsd_lock);
    604  1.127      yamt 				}
    605  1.127      yamt 				if (error == 0 && more) {
    606   1.85      yamt 					nfsrv_wakenfsd(slp);
    607   1.85      yamt 				}
    608    1.1       cgd 			}
    609   1.10   mycroft 		} else {
    610   1.10   mycroft 			error = 0;
    611   1.19      fvdl 			slp = nfsd->nfsd_slp;
    612   1.10   mycroft 		}
    613  1.116      yamt 		KASSERT(slp != NULL);
    614  1.116      yamt 		KASSERT(nfsd->nfsd_slp == slp);
    615  1.127      yamt 		if (error || (slp->ns_flags & SLP_VALID) == 0) {
    616   1.19      fvdl 			if (nd) {
    617   1.93      yamt 				nfsdreq_free(nd);
    618   1.19      fvdl 				nd = NULL;
    619   1.19      fvdl 			}
    620   1.85      yamt 			nfsd->nfsd_slp = NULL;
    621   1.10   mycroft 			nfsrv_slpderef(slp);
    622    1.1       cgd 			continue;
    623    1.1       cgd 		}
    624   1.86      yamt 		sotype = slp->ns_so->so_type;
    625   1.19      fvdl 		if (nd) {
    626   1.95    kardel 			getmicrotime(&nd->nd_starttime);
    627   1.63      yamt 			if (nd->nd_nam2)
    628   1.63      yamt 				nd->nd_nam = nd->nd_nam2;
    629   1.63      yamt 			else
    630   1.63      yamt 				nd->nd_nam = slp->ns_nam;
    631   1.63      yamt 
    632   1.63      yamt 			/*
    633   1.63      yamt 			 * Check to see if authorization is needed.
    634   1.63      yamt 			 */
    635   1.63      yamt 			if (nfsd->nfsd_flag & NFSD_NEEDAUTH) {
    636   1.63      yamt 				nfsd->nfsd_flag &= ~NFSD_NEEDAUTH;
    637   1.63      yamt 				nsd->nsd_haddr = mtod(nd->nd_nam,
    638   1.63      yamt 				    struct sockaddr_in *)->sin_addr.s_addr;
    639   1.63      yamt 				nsd->nsd_authlen = nfsd->nfsd_authlen;
    640   1.63      yamt 				nsd->nsd_verflen = nfsd->nfsd_verflen;
    641   1.63      yamt 				if (!copyout(nfsd->nfsd_authstr,
    642   1.63      yamt 				    nsd->nsd_authstr, nfsd->nfsd_authlen) &&
    643   1.63      yamt 				    !copyout(nfsd->nfsd_verfstr,
    644   1.63      yamt 				    nsd->nsd_verfstr, nfsd->nfsd_verflen) &&
    645   1.63      yamt 				    !copyout(nsd, argp, sizeof (*nsd))) {
    646  1.117        ad 					uvm_lwp_rele(l);
    647   1.63      yamt 					return (ENEEDAUTH);
    648   1.63      yamt 				}
    649   1.63      yamt 				cacherep = RC_DROPIT;
    650   1.10   mycroft 			} else
    651   1.63      yamt 				cacherep = nfsrv_getcache(nd, slp, &mreq);
    652   1.63      yamt 
    653  1.100      yamt 			if (nfsd->nfsd_flag & NFSD_AUTHFAIL) {
    654   1.63      yamt 				nfsd->nfsd_flag &= ~NFSD_AUTHFAIL;
    655   1.10   mycroft 				nd->nd_procnum = NFSPROC_NOOP;
    656   1.63      yamt 				nd->nd_repstat =
    657   1.63      yamt 				    (NFSERR_AUTHERR | AUTH_TOOWEAK);
    658   1.10   mycroft 				cacherep = RC_DOIT;
    659   1.63      yamt 			}
    660   1.10   mycroft 		}
    661   1.10   mycroft 
    662   1.19      fvdl 		/*
    663   1.19      fvdl 		 * Loop to get all the write rpc relies that have been
    664   1.19      fvdl 		 * gathered together.
    665   1.19      fvdl 		 */
    666   1.19      fvdl 		do {
    667   1.34  sommerfe #ifdef DIAGNOSTIC
    668   1.63      yamt 			int lockcount;
    669   1.34  sommerfe #endif
    670   1.63      yamt 			switch (cacherep) {
    671   1.63      yamt 			case RC_DOIT:
    672   1.34  sommerfe #ifdef DIAGNOSTIC
    673   1.63      yamt 				/*
    674   1.63      yamt 				 * NFS server procs should neither release
    675   1.63      yamt 				 * locks already held, nor leave things
    676   1.63      yamt 				 * locked.  Catch this sooner, rather than
    677   1.63      yamt 				 * later (when we try to relock something we
    678   1.63      yamt 				 * already have locked).  Careful inspection
    679   1.63      yamt 				 * of the failing routine usually turns up the
    680   1.63      yamt 				 * lock leak.. once we know what it is..
    681   1.63      yamt 				 */
    682   1.63      yamt 				lockcount = l->l_locks;
    683   1.34  sommerfe #endif
    684   1.80      onoe 				mreq = NULL;
    685   1.88      yamt 				netexport_rdlock();
    686   1.90  christos 				if (writes_todo || nd == NULL ||
    687   1.90  christos 				     (!(nd->nd_flag & ND_NFSV3) &&
    688   1.63      yamt 				     nd->nd_procnum == NFSPROC_WRITE &&
    689  1.100      yamt 				     nfsrvw_procrastinate > 0))
    690   1.63      yamt 					error = nfsrv_writegather(&nd, slp,
    691   1.84  christos 					    l, &mreq);
    692   1.63      yamt 				else
    693   1.63      yamt 					error =
    694   1.63      yamt 					    (*(nfsrv3_procs[nd->nd_procnum]))
    695   1.84  christos 					    (nd, slp, l, &mreq);
    696   1.88      yamt 				netexport_rdunlock();
    697   1.34  sommerfe #ifdef DIAGNOSTIC
    698   1.63      yamt 				if (l->l_locks != lockcount) {
    699   1.63      yamt 					/*
    700   1.63      yamt 					 * If you see this panic, audit
    701   1.63      yamt 					 * nfsrv3_procs[nd->nd_procnum] for
    702   1.63      yamt 					 * vnode locking errors (usually, it's
    703   1.63      yamt 					 * due to forgetting to vput()
    704   1.63      yamt 					 * something).
    705   1.63      yamt 					 */
    706   1.46      matt #ifdef DEBUG
    707   1.63      yamt 					extern void printlockedvnodes(void);
    708   1.63      yamt 					printlockedvnodes();
    709   1.46      matt #endif
    710   1.63      yamt 					printf("nfsd: locking botch in op %d"
    711   1.63      yamt 					    " (before %d, after %d)\n",
    712   1.63      yamt 					    nd ? nd->nd_procnum : -1,
    713   1.63      yamt 					    lockcount, l->l_locks);
    714   1.63      yamt 				}
    715   1.34  sommerfe #endif
    716   1.80      onoe 				if (mreq == NULL) {
    717   1.80      onoe 					if (nd != NULL) {
    718   1.80      onoe 						if (nd->nd_nam2)
    719   1.80      onoe 							m_free(nd->nd_nam2);
    720   1.80      onoe 						if (nd->nd_mrep)
    721   1.80      onoe 							m_freem(nd->nd_mrep);
    722   1.80      onoe 					}
    723   1.63      yamt 					break;
    724   1.80      onoe 				}
    725   1.63      yamt 				if (error) {
    726  1.100      yamt 					nfsstats.srv_errs++;
    727  1.106   thorpej 					nfsrv_updatecache(nd, false, mreq);
    728   1.63      yamt 					if (nd->nd_nam2)
    729   1.63      yamt 						m_freem(nd->nd_nam2);
    730   1.63      yamt 					break;
    731   1.63      yamt 				}
    732   1.63      yamt 				nfsstats.srvrpccnt[nd->nd_procnum]++;
    733  1.106   thorpej 				nfsrv_updatecache(nd, true, mreq);
    734   1.63      yamt 				nd->nd_mrep = (struct mbuf *)0;
    735   1.63      yamt 			case RC_REPLY:
    736   1.63      yamt 				m = mreq;
    737   1.63      yamt 				siz = 0;
    738   1.63      yamt 				while (m) {
    739   1.63      yamt 					siz += m->m_len;
    740   1.63      yamt 					m = m->m_next;
    741   1.63      yamt 				}
    742   1.63      yamt 				if (siz <= 0 || siz > NFS_MAXPACKET) {
    743   1.63      yamt 					printf("mbuf siz=%d\n",siz);
    744   1.63      yamt 					panic("Bad nfs svc reply");
    745   1.63      yamt 				}
    746   1.63      yamt 				m = mreq;
    747   1.63      yamt 				m->m_pkthdr.len = siz;
    748   1.63      yamt 				m->m_pkthdr.rcvif = (struct ifnet *)0;
    749   1.63      yamt 				/*
    750   1.63      yamt 				 * For stream protocols, prepend a Sun RPC
    751   1.63      yamt 				 * Record Mark.
    752   1.63      yamt 				 */
    753   1.63      yamt 				if (sotype == SOCK_STREAM) {
    754   1.63      yamt 					M_PREPEND(m, NFSX_UNSIGNED, M_WAIT);
    755   1.63      yamt 					*mtod(m, u_int32_t *) =
    756   1.63      yamt 					    htonl(0x80000000 | siz);
    757   1.63      yamt 				}
    758   1.86      yamt 				nd->nd_mreq = m;
    759   1.86      yamt 				if (nfsrtton) {
    760   1.86      yamt 					nfsd_rt(slp->ns_so->so_type, nd,
    761   1.86      yamt 					    cacherep);
    762   1.63      yamt 				}
    763   1.86      yamt 				error = nfsdsock_sendreply(slp, nd);
    764   1.86      yamt 				nd = NULL;
    765   1.63      yamt 				if (error == EPIPE)
    766   1.63      yamt 					nfsrv_zapsock(slp);
    767   1.63      yamt 				if (error == EINTR || error == ERESTART) {
    768  1.114      yamt 					nfsd->nfsd_slp = NULL;
    769   1.63      yamt 					nfsrv_slpderef(slp);
    770   1.63      yamt 					goto done;
    771   1.63      yamt 				}
    772   1.19      fvdl 				break;
    773   1.63      yamt 			case RC_DROPIT:
    774   1.63      yamt 				if (nfsrtton)
    775   1.63      yamt 					nfsd_rt(sotype, nd, cacherep);
    776   1.63      yamt 				m_freem(nd->nd_mrep);
    777   1.63      yamt 				m_freem(nd->nd_nam2);
    778    1.1       cgd 				break;
    779    1.1       cgd 			}
    780   1.63      yamt 			if (nd) {
    781   1.93      yamt 				nfsdreq_free(nd);
    782   1.63      yamt 				nd = NULL;
    783   1.63      yamt 			}
    784   1.63      yamt 
    785    1.1       cgd 			/*
    786   1.63      yamt 			 * Check to see if there are outstanding writes that
    787   1.63      yamt 			 * need to be serviced.
    788    1.1       cgd 			 */
    789   1.95    kardel 			getmicrotime(&tv);
    790   1.95    kardel 			cur_usec = (u_quad_t)tv.tv_sec * 1000000 +
    791   1.95    kardel 			    (u_quad_t)tv.tv_usec;
    792   1.63      yamt 			s = splsoftclock();
    793   1.63      yamt 			if (LIST_FIRST(&slp->ns_tq) &&
    794   1.63      yamt 			    LIST_FIRST(&slp->ns_tq)->nd_time <= cur_usec) {
    795   1.63      yamt 				cacherep = RC_DOIT;
    796   1.63      yamt 				writes_todo = 1;
    797   1.63      yamt 			} else
    798   1.63      yamt 				writes_todo = 0;
    799   1.63      yamt 			splx(s);
    800   1.19      fvdl 		} while (writes_todo);
    801  1.127      yamt 		if (nfsrv_dorec(slp, nfsd, &nd, &dummy)) {
    802   1.19      fvdl 			nfsd->nfsd_slp = NULL;
    803   1.10   mycroft 			nfsrv_slpderef(slp);
    804   1.10   mycroft 		}
    805    1.1       cgd 	}
    806   1.10   mycroft done:
    807  1.111      yamt 	mutex_enter(&nfsd_lock);
    808   1.19      fvdl 	TAILQ_REMOVE(&nfsd_head, nfsd, nfsd_chain);
    809  1.127      yamt 	doreinit = --nfs_numnfsd == 0;
    810  1.127      yamt 	if (doreinit)
    811  1.127      yamt 		nfssvc_sockhead_flag |= SLP_INIT;
    812  1.111      yamt 	mutex_exit(&nfsd_lock);
    813  1.111      yamt 	cv_destroy(&nfsd->nfsd_cv);
    814  1.130      yamt 	kmem_free(nfsd, sizeof(*nfsd));
    815  1.111      yamt 	nsd->nsd_nfsd = NULL;
    816  1.127      yamt 	if (doreinit)
    817  1.106   thorpej 		nfsrv_init(true);	/* Reinitialize everything */
    818  1.117        ad 	uvm_lwp_rele(l);
    819    1.1       cgd 	return (error);
    820    1.1       cgd }
    821    1.1       cgd 
    822   1.10   mycroft /*
    823   1.10   mycroft  * Shut down a socket associated with an nfssvc_sock structure.
    824   1.10   mycroft  * Should be called with the send lock set, if required.
    825   1.10   mycroft  * The trick here is to increment the sref at the start, so that the nfsds
    826   1.10   mycroft  * will stop using it and clear ns_flag at the end so that it will not be
    827   1.10   mycroft  * reassigned during cleanup.
    828   1.73      yamt  *
    829   1.73      yamt  * called at splsoftnet.
    830   1.10   mycroft  */
    831   1.18  christos void
    832   1.10   mycroft nfsrv_zapsock(slp)
    833   1.38  augustss 	struct nfssvc_sock *slp;
    834   1.10   mycroft {
    835   1.38  augustss 	struct nfsuid *nuidp, *nnuidp;
    836  1.127      yamt 	struct nfsrv_descript *nwp;
    837   1.10   mycroft 	struct socket *so;
    838  1.104      yamt 	struct mbuf *m;
    839   1.19      fvdl 	int s;
    840   1.10   mycroft 
    841   1.85      yamt 	if (nfsdsock_drain(slp)) {
    842   1.74      yamt 		return;
    843   1.74      yamt 	}
    844  1.111      yamt 	mutex_enter(&nfsd_lock);
    845  1.127      yamt 	if (slp->ns_gflags & SLP_G_DOREC) {
    846   1.73      yamt 		TAILQ_REMOVE(&nfssvc_sockpending, slp, ns_pending);
    847  1.127      yamt 		slp->ns_gflags &= ~SLP_G_DOREC;
    848   1.73      yamt 	}
    849  1.111      yamt 	mutex_exit(&nfsd_lock);
    850   1.77      yamt 
    851   1.77      yamt 	so = slp->ns_so;
    852   1.77      yamt 	KASSERT(so != NULL);
    853  1.129        ad 	KERNEL_LOCK(1, curlwp);
    854  1.127      yamt 	s = splsoftnet();
    855   1.77      yamt 	so->so_upcall = NULL;
    856   1.77      yamt 	so->so_upcallarg = NULL;
    857   1.77      yamt 	so->so_rcv.sb_flags &= ~SB_UPCALL;
    858  1.127      yamt 	splx(s);
    859   1.77      yamt 	soshutdown(so, SHUT_RDWR);
    860  1.129        ad 	KERNEL_UNLOCK_ONE(curlwp);
    861   1.77      yamt 
    862   1.77      yamt 	if (slp->ns_nam)
    863   1.77      yamt 		m_free(slp->ns_nam);
    864   1.77      yamt 	m_freem(slp->ns_raw);
    865  1.104      yamt 	m = slp->ns_rec;
    866  1.104      yamt 	while (m != NULL) {
    867  1.104      yamt 		struct mbuf *n;
    868  1.104      yamt 
    869  1.104      yamt 		n = m->m_nextpkt;
    870  1.104      yamt 		m_freem(m);
    871  1.104      yamt 		m = n;
    872  1.104      yamt 	}
    873   1.77      yamt 	for (nuidp = TAILQ_FIRST(&slp->ns_uidlruhead); nuidp != 0;
    874   1.77      yamt 	    nuidp = nnuidp) {
    875   1.77      yamt 		nnuidp = TAILQ_NEXT(nuidp, nu_lru);
    876   1.77      yamt 		LIST_REMOVE(nuidp, nu_hash);
    877   1.77      yamt 		TAILQ_REMOVE(&slp->ns_uidlruhead, nuidp, nu_lru);
    878   1.77      yamt 		if (nuidp->nu_flag & NU_NAM)
    879   1.77      yamt 			m_freem(nuidp->nu_nam);
    880  1.130      yamt 		kmem_free(nuidp, sizeof(*nuidp));
    881   1.77      yamt 	}
    882  1.127      yamt 	mutex_enter(&nfsd_lock);
    883  1.127      yamt 	while ((nwp = LIST_FIRST(&slp->ns_tq)) != NULL) {
    884   1.77      yamt 		LIST_REMOVE(nwp, nd_tq);
    885  1.127      yamt 		mutex_exit(&nfsd_lock);
    886   1.93      yamt 		nfsdreq_free(nwp);
    887  1.127      yamt 		mutex_enter(&nfsd_lock);
    888   1.10   mycroft 	}
    889  1.127      yamt 	mutex_exit(&nfsd_lock);
    890   1.10   mycroft }
    891   1.10   mycroft 
    892   1.10   mycroft /*
    893   1.10   mycroft  * Derefence a server socket structure. If it has no more references and
    894   1.10   mycroft  * is no longer valid, you can throw it away.
    895   1.10   mycroft  */
    896   1.10   mycroft void
    897   1.10   mycroft nfsrv_slpderef(slp)
    898   1.38  augustss 	struct nfssvc_sock *slp;
    899   1.10   mycroft {
    900  1.116      yamt 	uint32_t ref;
    901   1.66      yamt 
    902  1.116      yamt 	mutex_enter(&nfsd_lock);
    903  1.116      yamt 	KASSERT(slp->ns_sref > 0);
    904  1.116      yamt 	ref = --slp->ns_sref;
    905  1.116      yamt 	mutex_exit(&nfsd_lock);
    906  1.127      yamt 	if (ref == 0 && (slp->ns_flags & SLP_VALID) == 0) {
    907  1.132        ad 		file_t *fp;
    908  1.111      yamt 
    909  1.111      yamt 		mutex_enter(&nfsd_lock);
    910  1.127      yamt 		KASSERT((slp->ns_gflags & SLP_G_DOREC) == 0);
    911   1.12   mycroft 		TAILQ_REMOVE(&nfssvc_sockhead, slp, ns_chain);
    912  1.111      yamt 		mutex_exit(&nfsd_lock);
    913   1.85      yamt 
    914   1.85      yamt 		fp = slp->ns_fp;
    915   1.85      yamt 		if (fp != NULL) {
    916   1.85      yamt 			slp->ns_fp = NULL;
    917   1.85      yamt 			KASSERT(fp != NULL);
    918   1.85      yamt 			KASSERT(fp->f_data == slp->ns_so);
    919  1.132        ad 			KASSERT(fp->f_count > 0);
    920  1.132        ad 			closef(fp);
    921   1.85      yamt 			slp->ns_so = NULL;
    922   1.85      yamt 		}
    923   1.85      yamt 
    924   1.77      yamt 		nfsrv_sockfree(slp);
    925   1.10   mycroft 	}
    926   1.10   mycroft }
    927   1.10   mycroft 
    928   1.10   mycroft /*
    929   1.10   mycroft  * Initialize the data structures for the server.
    930   1.10   mycroft  * Handshake with any new nfsds starting up to avoid any chance of
    931   1.10   mycroft  * corruption.
    932   1.10   mycroft  */
    933   1.10   mycroft void
    934   1.10   mycroft nfsrv_init(terminating)
    935   1.10   mycroft 	int terminating;
    936   1.10   mycroft {
    937   1.66      yamt 	struct nfssvc_sock *slp;
    938   1.10   mycroft 
    939  1.111      yamt 	if (!terminating) {
    940  1.127      yamt 		/* XXX could be IPL_SOFTNET */
    941  1.127      yamt 		mutex_init(&nfsd_lock, MUTEX_DRIVER, IPL_VM);
    942  1.111      yamt 		cv_init(&nfsd_initcv, "nfsdinit");
    943  1.111      yamt 	}
    944  1.111      yamt 
    945  1.111      yamt 	mutex_enter(&nfsd_lock);
    946  1.127      yamt 	if (!terminating && (nfssvc_sockhead_flag & SLP_INIT) != 0)
    947   1.10   mycroft 		panic("nfsd init");
    948   1.12   mycroft 	nfssvc_sockhead_flag |= SLP_INIT;
    949   1.66      yamt 
    950   1.10   mycroft 	if (terminating) {
    951  1.114      yamt 		KASSERT(SLIST_EMPTY(&nfsd_idle_head));
    952  1.114      yamt 		KASSERT(TAILQ_EMPTY(&nfsd_head));
    953   1.66      yamt 		while ((slp = TAILQ_FIRST(&nfssvc_sockhead)) != NULL) {
    954  1.111      yamt 			mutex_exit(&nfsd_lock);
    955   1.85      yamt 			KASSERT(slp->ns_sref == 0);
    956   1.85      yamt 			slp->ns_sref++;
    957   1.85      yamt 			nfsrv_zapsock(slp);
    958   1.85      yamt 			nfsrv_slpderef(slp);
    959  1.111      yamt 			mutex_enter(&nfsd_lock);
    960   1.10   mycroft 		}
    961  1.114      yamt 		KASSERT(TAILQ_EMPTY(&nfssvc_sockpending));
    962  1.111      yamt 		mutex_exit(&nfsd_lock);
    963   1.10   mycroft 		nfsrv_cleancache();	/* And clear out server cache */
    964   1.66      yamt 	} else {
    965  1.111      yamt 		mutex_exit(&nfsd_lock);
    966   1.26      fvdl 		nfs_pub.np_valid = 0;
    967   1.66      yamt 	}
    968   1.12   mycroft 
    969   1.12   mycroft 	TAILQ_INIT(&nfssvc_sockhead);
    970   1.66      yamt 	TAILQ_INIT(&nfssvc_sockpending);
    971   1.12   mycroft 
    972   1.12   mycroft 	TAILQ_INIT(&nfsd_head);
    973   1.66      yamt 	SLIST_INIT(&nfsd_idle_head);
    974   1.12   mycroft 	nfsd_head_flag &= ~NFSD_CHECKSLP;
    975   1.12   mycroft 
    976   1.65      yamt 	nfs_udpsock = nfsrv_sockalloc();
    977   1.41      fvdl 
    978   1.41      fvdl #ifdef INET6
    979   1.65      yamt 	nfs_udp6sock = nfsrv_sockalloc();
    980   1.41      fvdl #endif
    981   1.12   mycroft 
    982   1.47       chs #ifdef ISO
    983   1.65      yamt 	nfs_cltpsock = nfsrv_sockalloc();
    984   1.47       chs #endif
    985   1.66      yamt 
    986  1.111      yamt 	mutex_enter(&nfsd_lock);
    987  1.111      yamt 	nfssvc_sockhead_flag &= ~SLP_INIT;
    988  1.111      yamt 	cv_broadcast(&nfsd_initcv);
    989  1.111      yamt 	mutex_exit(&nfsd_lock);
    990   1.10   mycroft }
    991   1.10   mycroft 
    992   1.10   mycroft /*
    993   1.10   mycroft  * Add entries to the server monitor log.
    994   1.10   mycroft  */
    995   1.10   mycroft static void
    996   1.19      fvdl nfsd_rt(sotype, nd, cacherep)
    997   1.10   mycroft 	int sotype;
    998   1.38  augustss 	struct nfsrv_descript *nd;
    999   1.10   mycroft 	int cacherep;
   1000   1.10   mycroft {
   1001   1.95    kardel 	struct timeval tv;
   1002   1.38  augustss 	struct drt *rt;
   1003   1.10   mycroft 
   1004   1.10   mycroft 	rt = &nfsdrt.drt[nfsdrt.pos];
   1005   1.10   mycroft 	if (cacherep == RC_DOIT)
   1006   1.10   mycroft 		rt->flag = 0;
   1007   1.10   mycroft 	else if (cacherep == RC_REPLY)
   1008   1.10   mycroft 		rt->flag = DRT_CACHEREPLY;
   1009   1.10   mycroft 	else
   1010   1.10   mycroft 		rt->flag = DRT_CACHEDROP;
   1011   1.10   mycroft 	if (sotype == SOCK_STREAM)
   1012   1.10   mycroft 		rt->flag |= DRT_TCP;
   1013  1.102      yamt 	if (nd->nd_flag & ND_NFSV3)
   1014   1.19      fvdl 		rt->flag |= DRT_NFSV3;
   1015   1.10   mycroft 	rt->proc = nd->nd_procnum;
   1016   1.19      fvdl 	if (mtod(nd->nd_nam, struct sockaddr *)->sa_family == AF_INET)
   1017   1.19      fvdl 	    rt->ipadr = mtod(nd->nd_nam, struct sockaddr_in *)->sin_addr.s_addr;
   1018   1.10   mycroft 	else
   1019   1.19      fvdl 	    rt->ipadr = INADDR_ANY;
   1020   1.95    kardel 	getmicrotime(&tv);
   1021   1.95    kardel 	rt->resptime = ((tv.tv_sec - nd->nd_starttime.tv_sec) * 1000000) +
   1022   1.95    kardel 		(tv.tv_usec - nd->nd_starttime.tv_usec);
   1023   1.95    kardel 	rt->tstamp = tv;
   1024   1.10   mycroft 	nfsdrt.pos = (nfsdrt.pos + 1) % NFSRTTLOGSIZ;
   1025   1.10   mycroft }
   1026    1.3     glass #endif /* NFSSERVER */
   1027    1.3     glass 
   1028   1.24   thorpej #ifdef NFS
   1029   1.22   thorpej 
   1030   1.22   thorpej int nfs_defect = 0;
   1031    1.1       cgd /*
   1032   1.39    tsarna  * Asynchronous I/O threads for client nfs.
   1033   1.10   mycroft  * They do read-ahead and write-behind operations on the block I/O cache.
   1034   1.10   mycroft  * Never returns unless it fails or gets killed.
   1035    1.1       cgd  */
   1036   1.39    tsarna 
   1037  1.118      yamt static void
   1038  1.118      yamt nfssvc_iod(void *arg)
   1039    1.1       cgd {
   1040   1.38  augustss 	struct buf *bp;
   1041   1.58      yamt 	struct nfs_iod *myiod;
   1042   1.22   thorpej 	struct nfsmount *nmp;
   1043    1.1       cgd 
   1044  1.124      yamt 	myiod = kmem_alloc(sizeof(*myiod), KM_SLEEP);
   1045  1.124      yamt 	mutex_init(&myiod->nid_lock, MUTEX_DEFAULT, IPL_NONE);
   1046  1.124      yamt 	cv_init(&myiod->nid_cv, "nfsiod");
   1047  1.124      yamt 	myiod->nid_exiting = false;
   1048  1.124      yamt 	myiod->nid_mount = NULL;
   1049  1.123      yamt 	mutex_enter(&nfs_iodlist_lock);
   1050  1.124      yamt 	LIST_INSERT_HEAD(&nfs_iodlist_all, myiod, nid_all);
   1051  1.123      yamt 	mutex_exit(&nfs_iodlist_lock);
   1052  1.118      yamt 
   1053    1.1       cgd 	for (;;) {
   1054  1.124      yamt 		mutex_enter(&nfs_iodlist_lock);
   1055  1.124      yamt 		LIST_INSERT_HEAD(&nfs_iodlist_idle, myiod, nid_idle);
   1056  1.124      yamt 		mutex_exit(&nfs_iodlist_lock);
   1057  1.124      yamt 
   1058  1.109      yamt 		mutex_enter(&myiod->nid_lock);
   1059  1.106   thorpej 		while (/*CONSTCOND*/ true) {
   1060   1.60      yamt 			nmp = myiod->nid_mount;
   1061   1.60      yamt 			if (nmp) {
   1062  1.124      yamt 				myiod->nid_mount = NULL;
   1063  1.124      yamt 				break;
   1064  1.124      yamt 			}
   1065  1.124      yamt 			if (__predict_false(myiod->nid_exiting)) {
   1066  1.124      yamt 				/*
   1067  1.124      yamt 				 * drop nid_lock to preserve locking order.
   1068  1.124      yamt 				 */
   1069  1.124      yamt 				mutex_exit(&myiod->nid_lock);
   1070  1.124      yamt 				mutex_enter(&nfs_iodlist_lock);
   1071  1.124      yamt 				mutex_enter(&myiod->nid_lock);
   1072  1.124      yamt 				/*
   1073  1.124      yamt 				 * recheck nid_mount because nfs_asyncio can
   1074  1.124      yamt 				 * pick us in the meantime as we are still on
   1075  1.124      yamt 				 * nfs_iodlist_lock.
   1076  1.124      yamt 				 */
   1077  1.124      yamt 				if (myiod->nid_mount != NULL) {
   1078  1.124      yamt 					mutex_exit(&nfs_iodlist_lock);
   1079  1.124      yamt 					continue;
   1080   1.60      yamt 				}
   1081  1.124      yamt 				LIST_REMOVE(myiod, nid_idle);
   1082  1.124      yamt 				mutex_exit(&nfs_iodlist_lock);
   1083   1.60      yamt 				goto quit;
   1084  1.110      yamt 			}
   1085  1.118      yamt 			cv_wait(&myiod->nid_cv, &myiod->nid_lock);
   1086   1.59      yamt 		}
   1087  1.124      yamt 		mutex_exit(&myiod->nid_lock);
   1088   1.60      yamt 
   1089  1.124      yamt 		mutex_enter(&nmp->nm_lock);
   1090   1.60      yamt 		while ((bp = TAILQ_FIRST(&nmp->nm_bufq)) != NULL) {
   1091   1.59      yamt 			/* Take one off the front of the list */
   1092   1.59      yamt 			TAILQ_REMOVE(&nmp->nm_bufq, bp, b_freelist);
   1093   1.59      yamt 			nmp->nm_bufqlen--;
   1094  1.123      yamt 			if (nmp->nm_bufqlen < 2 * nmp->nm_bufqiods) {
   1095  1.109      yamt 				cv_broadcast(&nmp->nm_aiocv);
   1096   1.59      yamt 			}
   1097  1.109      yamt 			mutex_exit(&nmp->nm_lock);
   1098  1.124      yamt 			KERNEL_LOCK(1, curlwp);
   1099   1.79  christos 			(void)nfs_doio(bp);
   1100  1.124      yamt 			KERNEL_UNLOCK_LAST(curlwp);
   1101  1.109      yamt 			mutex_enter(&nmp->nm_lock);
   1102   1.59      yamt 			/*
   1103   1.81    rpaulo 			 * If there are more than one iod on this mount,
   1104   1.81    rpaulo 			 * then defect so that the iods can be shared out
   1105   1.81    rpaulo 			 * fairly between the mounts
   1106   1.59      yamt 			 */
   1107   1.59      yamt 			if (nfs_defect && nmp->nm_bufqiods > 1) {
   1108   1.59      yamt 				break;
   1109   1.59      yamt 			}
   1110   1.59      yamt 		}
   1111  1.124      yamt 		KASSERT(nmp->nm_bufqiods > 0);
   1112  1.124      yamt 		nmp->nm_bufqiods--;
   1113  1.109      yamt 		mutex_exit(&nmp->nm_lock);
   1114    1.1       cgd 	}
   1115   1.60      yamt quit:
   1116  1.124      yamt 	KASSERT(myiod->nid_mount == NULL);
   1117  1.124      yamt 	mutex_exit(&myiod->nid_lock);
   1118  1.124      yamt 
   1119  1.124      yamt 	cv_destroy(&myiod->nid_cv);
   1120  1.124      yamt 	mutex_destroy(&myiod->nid_lock);
   1121  1.124      yamt 	kmem_free(myiod, sizeof(*myiod));
   1122   1.39    tsarna 
   1123  1.118      yamt 	kthread_exit(0);
   1124   1.60      yamt }
   1125   1.60      yamt 
   1126   1.60      yamt void
   1127   1.60      yamt nfs_iodinit()
   1128   1.60      yamt {
   1129   1.60      yamt 
   1130  1.123      yamt 	mutex_init(&nfs_iodlist_lock, MUTEX_DEFAULT, IPL_NONE);
   1131  1.124      yamt 	LIST_INIT(&nfs_iodlist_all);
   1132  1.124      yamt 	LIST_INIT(&nfs_iodlist_idle);
   1133    1.1       cgd }
   1134    1.3     glass 
   1135  1.124      yamt int
   1136  1.124      yamt nfs_set_niothreads(int newval)
   1137   1.39    tsarna {
   1138  1.124      yamt 	struct nfs_iod *nid;
   1139  1.124      yamt 	int error = 0;
   1140  1.124      yamt 
   1141  1.124      yamt #if defined(MULTIPROCESSOR)
   1142  1.124      yamt         int hold_count;
   1143  1.124      yamt #endif /* defined(MULTIPROCESSOR) */
   1144  1.124      yamt 
   1145  1.124      yamt 	KERNEL_UNLOCK_ALL(curlwp, &hold_count);
   1146   1.78     perry 
   1147  1.123      yamt 	mutex_enter(&nfs_iodlist_lock);
   1148  1.124      yamt 	/* clamp to sane range */
   1149  1.124      yamt 	nfs_niothreads = max(0, min(newval, NFS_MAXASYNCDAEMON));
   1150   1.39    tsarna 
   1151  1.124      yamt 	while (nfs_numasync != nfs_niothreads && error == 0) {
   1152  1.124      yamt 		while (nfs_numasync < nfs_niothreads) {
   1153   1.39    tsarna 
   1154  1.124      yamt 			/*
   1155  1.124      yamt 			 * kthread_create can wait for pagedaemon and
   1156  1.124      yamt 			 * pagedaemon can wait for nfsiod which needs to acquire
   1157  1.124      yamt 			 * nfs_iodlist_lock.
   1158  1.124      yamt 			 */
   1159  1.124      yamt 
   1160  1.124      yamt 			mutex_exit(&nfs_iodlist_lock);
   1161  1.124      yamt 			error = kthread_create(PRI_NONE, KTHREAD_MPSAFE, NULL,
   1162  1.123      yamt 			    nfssvc_iod, NULL, NULL, "nfsio");
   1163  1.124      yamt 			mutex_enter(&nfs_iodlist_lock);
   1164  1.124      yamt 			if (error) {
   1165  1.124      yamt 				/* give up */
   1166  1.124      yamt 				nfs_niothreads = nfs_numasync;
   1167  1.124      yamt 				break;
   1168  1.124      yamt 			}
   1169  1.124      yamt 			nfs_numasync++;
   1170   1.39    tsarna 		}
   1171  1.124      yamt 		while (nfs_numasync > nfs_niothreads) {
   1172  1.124      yamt 			nid = LIST_FIRST(&nfs_iodlist_all);
   1173  1.124      yamt 			if (nid == NULL) {
   1174  1.124      yamt 				/* iod has not started yet. */
   1175  1.124      yamt 				kpause("nfsiorm", false, hz, &nfs_iodlist_lock);
   1176  1.124      yamt 				continue;
   1177   1.39    tsarna 			}
   1178  1.124      yamt 			LIST_REMOVE(nid, nid_all);
   1179  1.124      yamt 			mutex_enter(&nid->nid_lock);
   1180  1.124      yamt 			KASSERT(!nid->nid_exiting);
   1181  1.124      yamt 			nid->nid_exiting = true;
   1182  1.124      yamt 			cv_signal(&nid->nid_cv);
   1183  1.124      yamt 			mutex_exit(&nid->nid_lock);
   1184  1.124      yamt 			nfs_numasync--;
   1185  1.118      yamt 		}
   1186   1.39    tsarna 	}
   1187  1.123      yamt 	mutex_exit(&nfs_iodlist_lock);
   1188  1.124      yamt 
   1189  1.124      yamt 	KERNEL_LOCK(hold_count, curlwp);
   1190  1.124      yamt 	return error;
   1191   1.39    tsarna }
   1192   1.19      fvdl 
   1193   1.10   mycroft /*
   1194   1.10   mycroft  * Get an authorization string for the uid by having the mount_nfs sitting
   1195   1.10   mycroft  * on this mount point porpous out of the kernel and do it.
   1196   1.10   mycroft  */
   1197   1.18  christos int
   1198   1.19      fvdl nfs_getauth(nmp, rep, cred, auth_str, auth_len, verf_str, verf_len, key)
   1199   1.38  augustss 	struct nfsmount *nmp;
   1200   1.10   mycroft 	struct nfsreq *rep;
   1201   1.92      elad 	kauth_cred_t cred;
   1202   1.10   mycroft 	char **auth_str;
   1203   1.10   mycroft 	int *auth_len;
   1204   1.19      fvdl 	char *verf_str;
   1205   1.19      fvdl 	int *verf_len;
   1206   1.19      fvdl 	NFSKERBKEY_T key;		/* return session key */
   1207   1.10   mycroft {
   1208   1.10   mycroft 	int error = 0;
   1209   1.10   mycroft 
   1210   1.27      fvdl 	while ((nmp->nm_iflag & NFSMNT_WAITAUTH) == 0) {
   1211   1.27      fvdl 		nmp->nm_iflag |= NFSMNT_WANTAUTH;
   1212  1.107  christos 		(void) tsleep((void *)&nmp->nm_authtype, PSOCK,
   1213   1.10   mycroft 			"nfsauth1", 2 * hz);
   1214   1.84  christos 		error = nfs_sigintr(nmp, rep, rep->r_lwp);
   1215   1.18  christos 		if (error) {
   1216   1.27      fvdl 			nmp->nm_iflag &= ~NFSMNT_WANTAUTH;
   1217   1.10   mycroft 			return (error);
   1218   1.10   mycroft 		}
   1219   1.10   mycroft 	}
   1220   1.27      fvdl 	nmp->nm_iflag &= ~(NFSMNT_WAITAUTH | NFSMNT_WANTAUTH);
   1221   1.10   mycroft 	nmp->nm_authstr = *auth_str = (char *)malloc(RPCAUTH_MAXSIZ, M_TEMP, M_WAITOK);
   1222   1.19      fvdl 	nmp->nm_authlen = RPCAUTH_MAXSIZ;
   1223   1.19      fvdl 	nmp->nm_verfstr = verf_str;
   1224   1.19      fvdl 	nmp->nm_verflen = *verf_len;
   1225   1.92      elad 	nmp->nm_authuid = kauth_cred_geteuid(cred);
   1226  1.107  christos 	wakeup((void *)&nmp->nm_authstr);
   1227    1.3     glass 
   1228   1.10   mycroft 	/*
   1229   1.10   mycroft 	 * And wait for mount_nfs to do its stuff.
   1230   1.10   mycroft 	 */
   1231   1.27      fvdl 	while ((nmp->nm_iflag & NFSMNT_HASAUTH) == 0 && error == 0) {
   1232  1.107  christos 		(void) tsleep((void *)&nmp->nm_authlen, PSOCK,
   1233   1.10   mycroft 			"nfsauth2", 2 * hz);
   1234   1.84  christos 		error = nfs_sigintr(nmp, rep, rep->r_lwp);
   1235   1.10   mycroft 	}
   1236   1.27      fvdl 	if (nmp->nm_iflag & NFSMNT_AUTHERR) {
   1237   1.27      fvdl 		nmp->nm_iflag &= ~NFSMNT_AUTHERR;
   1238   1.10   mycroft 		error = EAUTH;
   1239   1.10   mycroft 	}
   1240   1.10   mycroft 	if (error)
   1241  1.107  christos 		free((void *)*auth_str, M_TEMP);
   1242   1.10   mycroft 	else {
   1243   1.10   mycroft 		*auth_len = nmp->nm_authlen;
   1244   1.19      fvdl 		*verf_len = nmp->nm_verflen;
   1245   1.52       chs 		memcpy(key, nmp->nm_key, sizeof (NFSKERBKEY_T));
   1246   1.10   mycroft 	}
   1247   1.27      fvdl 	nmp->nm_iflag &= ~NFSMNT_HASAUTH;
   1248   1.27      fvdl 	nmp->nm_iflag |= NFSMNT_WAITAUTH;
   1249   1.27      fvdl 	if (nmp->nm_iflag & NFSMNT_WANTAUTH) {
   1250   1.27      fvdl 		nmp->nm_iflag &= ~NFSMNT_WANTAUTH;
   1251  1.107  christos 		wakeup((void *)&nmp->nm_authtype);
   1252   1.10   mycroft 	}
   1253   1.19      fvdl 	return (error);
   1254   1.19      fvdl }
   1255   1.19      fvdl 
   1256   1.19      fvdl /*
   1257   1.19      fvdl  * Get a nickname authenticator and verifier.
   1258   1.19      fvdl  */
   1259   1.19      fvdl int
   1260   1.99      yamt nfs_getnickauth(struct nfsmount *nmp, kauth_cred_t cred, char **auth_str,
   1261   1.99      yamt     int *auth_len, char *verf_str, int verf_len)
   1262   1.19      fvdl {
   1263   1.95    kardel 	struct timeval ktvin, ktvout, tv;
   1264   1.38  augustss 	struct nfsuid *nuidp;
   1265   1.38  augustss 	u_int32_t *nickp, *verfp;
   1266   1.19      fvdl 
   1267   1.91       mrg 	memset(&ktvout, 0, sizeof ktvout);	/* XXX gcc */
   1268   1.91       mrg 
   1269   1.19      fvdl #ifdef DIAGNOSTIC
   1270   1.19      fvdl 	if (verf_len < (4 * NFSX_UNSIGNED))
   1271   1.19      fvdl 		panic("nfs_getnickauth verf too small");
   1272   1.19      fvdl #endif
   1273   1.92      elad 	LIST_FOREACH(nuidp, NMUIDHASH(nmp, kauth_cred_geteuid(cred)), nu_hash) {
   1274   1.92      elad 		if (kauth_cred_geteuid(nuidp->nu_cr) == kauth_cred_geteuid(cred))
   1275   1.19      fvdl 			break;
   1276   1.19      fvdl 	}
   1277   1.95    kardel 	if (!nuidp || nuidp->nu_expire < time_second)
   1278   1.19      fvdl 		return (EACCES);
   1279   1.19      fvdl 
   1280   1.19      fvdl 	/*
   1281   1.19      fvdl 	 * Move to the end of the lru list (end of lru == most recently used).
   1282   1.19      fvdl 	 */
   1283   1.19      fvdl 	TAILQ_REMOVE(&nmp->nm_uidlruhead, nuidp, nu_lru);
   1284   1.19      fvdl 	TAILQ_INSERT_TAIL(&nmp->nm_uidlruhead, nuidp, nu_lru);
   1285   1.19      fvdl 
   1286   1.19      fvdl 	nickp = (u_int32_t *)malloc(2 * NFSX_UNSIGNED, M_TEMP, M_WAITOK);
   1287   1.19      fvdl 	*nickp++ = txdr_unsigned(RPCAKN_NICKNAME);
   1288   1.19      fvdl 	*nickp = txdr_unsigned(nuidp->nu_nickname);
   1289   1.19      fvdl 	*auth_str = (char *)nickp;
   1290   1.19      fvdl 	*auth_len = 2 * NFSX_UNSIGNED;
   1291   1.19      fvdl 
   1292   1.19      fvdl 	/*
   1293   1.19      fvdl 	 * Now we must encrypt the verifier and package it up.
   1294   1.19      fvdl 	 */
   1295   1.19      fvdl 	verfp = (u_int32_t *)verf_str;
   1296   1.19      fvdl 	*verfp++ = txdr_unsigned(RPCAKN_NICKNAME);
   1297   1.95    kardel 	getmicrotime(&tv);
   1298   1.95    kardel 	if (tv.tv_sec > nuidp->nu_timestamp.tv_sec ||
   1299   1.95    kardel 	    (tv.tv_sec == nuidp->nu_timestamp.tv_sec &&
   1300   1.95    kardel 	     tv.tv_usec > nuidp->nu_timestamp.tv_usec))
   1301   1.95    kardel 		nuidp->nu_timestamp = tv;
   1302   1.19      fvdl 	else
   1303   1.19      fvdl 		nuidp->nu_timestamp.tv_usec++;
   1304   1.19      fvdl 	ktvin.tv_sec = txdr_unsigned(nuidp->nu_timestamp.tv_sec);
   1305   1.19      fvdl 	ktvin.tv_usec = txdr_unsigned(nuidp->nu_timestamp.tv_usec);
   1306   1.19      fvdl 
   1307   1.19      fvdl 	/*
   1308   1.19      fvdl 	 * Now encrypt the timestamp verifier in ecb mode using the session
   1309   1.19      fvdl 	 * key.
   1310   1.19      fvdl 	 */
   1311   1.19      fvdl #ifdef NFSKERB
   1312   1.19      fvdl 	XXX
   1313   1.19      fvdl #endif
   1314   1.19      fvdl 
   1315   1.19      fvdl 	*verfp++ = ktvout.tv_sec;
   1316   1.19      fvdl 	*verfp++ = ktvout.tv_usec;
   1317   1.19      fvdl 	*verfp = 0;
   1318   1.19      fvdl 	return (0);
   1319   1.19      fvdl }
   1320   1.19      fvdl 
   1321   1.19      fvdl /*
   1322   1.19      fvdl  * Save the current nickname in a hash list entry on the mount point.
   1323   1.19      fvdl  */
   1324   1.19      fvdl int
   1325   1.19      fvdl nfs_savenickauth(nmp, cred, len, key, mdp, dposp, mrep)
   1326   1.38  augustss 	struct nfsmount *nmp;
   1327   1.92      elad 	kauth_cred_t cred;
   1328   1.19      fvdl 	int len;
   1329   1.19      fvdl 	NFSKERBKEY_T key;
   1330   1.19      fvdl 	struct mbuf **mdp;
   1331   1.19      fvdl 	char **dposp;
   1332   1.19      fvdl 	struct mbuf *mrep;
   1333   1.19      fvdl {
   1334   1.38  augustss 	struct nfsuid *nuidp;
   1335   1.38  augustss 	u_int32_t *tl;
   1336   1.38  augustss 	int32_t t1;
   1337   1.19      fvdl 	struct mbuf *md = *mdp;
   1338   1.19      fvdl 	struct timeval ktvin, ktvout;
   1339   1.19      fvdl 	u_int32_t nick;
   1340   1.19      fvdl 	char *dpos = *dposp, *cp2;
   1341   1.19      fvdl 	int deltasec, error = 0;
   1342   1.19      fvdl 
   1343   1.91       mrg 	memset(&ktvout, 0, sizeof ktvout);	 /* XXX gcc */
   1344   1.91       mrg 
   1345   1.19      fvdl 	if (len == (3 * NFSX_UNSIGNED)) {
   1346   1.19      fvdl 		nfsm_dissect(tl, u_int32_t *, 3 * NFSX_UNSIGNED);
   1347   1.19      fvdl 		ktvin.tv_sec = *tl++;
   1348   1.19      fvdl 		ktvin.tv_usec = *tl++;
   1349   1.19      fvdl 		nick = fxdr_unsigned(u_int32_t, *tl);
   1350   1.19      fvdl 
   1351   1.19      fvdl 		/*
   1352   1.19      fvdl 		 * Decrypt the timestamp in ecb mode.
   1353   1.19      fvdl 		 */
   1354   1.19      fvdl #ifdef NFSKERB
   1355   1.19      fvdl 		XXX
   1356   1.19      fvdl #endif
   1357   1.19      fvdl 		ktvout.tv_sec = fxdr_unsigned(long, ktvout.tv_sec);
   1358   1.19      fvdl 		ktvout.tv_usec = fxdr_unsigned(long, ktvout.tv_usec);
   1359   1.95    kardel 		deltasec = time_second - ktvout.tv_sec;
   1360   1.19      fvdl 		if (deltasec < 0)
   1361   1.19      fvdl 			deltasec = -deltasec;
   1362   1.19      fvdl 		/*
   1363   1.19      fvdl 		 * If ok, add it to the hash list for the mount point.
   1364   1.19      fvdl 		 */
   1365   1.19      fvdl 		if (deltasec <= NFS_KERBCLOCKSKEW) {
   1366   1.19      fvdl 			if (nmp->nm_numuids < nuidhash_max) {
   1367   1.19      fvdl 				nmp->nm_numuids++;
   1368  1.130      yamt 				nuidp = kmem_alloc(sizeof(*nuidp), KM_SLEEP);
   1369   1.19      fvdl 			} else {
   1370   1.57      yamt 				nuidp = TAILQ_FIRST(&nmp->nm_uidlruhead);
   1371   1.19      fvdl 				LIST_REMOVE(nuidp, nu_hash);
   1372   1.19      fvdl 				TAILQ_REMOVE(&nmp->nm_uidlruhead, nuidp,
   1373   1.19      fvdl 					nu_lru);
   1374   1.19      fvdl 			}
   1375   1.19      fvdl 			nuidp->nu_flag = 0;
   1376   1.92      elad 			kauth_cred_seteuid(nuidp->nu_cr, kauth_cred_geteuid(cred));
   1377   1.95    kardel 			nuidp->nu_expire = time_second + NFS_KERBTTL;
   1378   1.19      fvdl 			nuidp->nu_timestamp = ktvout;
   1379   1.19      fvdl 			nuidp->nu_nickname = nick;
   1380   1.52       chs 			memcpy(nuidp->nu_key, key, sizeof (NFSKERBKEY_T));
   1381   1.19      fvdl 			TAILQ_INSERT_TAIL(&nmp->nm_uidlruhead, nuidp,
   1382   1.19      fvdl 				nu_lru);
   1383   1.92      elad 			LIST_INSERT_HEAD(NMUIDHASH(nmp, kauth_cred_geteuid(cred)),
   1384   1.19      fvdl 				nuidp, nu_hash);
   1385   1.19      fvdl 		}
   1386   1.19      fvdl 	} else
   1387   1.19      fvdl 		nfsm_adv(nfsm_rndup(len));
   1388   1.19      fvdl nfsmout:
   1389   1.19      fvdl 	*mdp = md;
   1390   1.19      fvdl 	*dposp = dpos;
   1391   1.10   mycroft 	return (error);
   1392   1.10   mycroft }
   1393   1.24   thorpej #endif /* NFS */
   1394