sys/vfs/nfs/nfs_iod.c

   1 /*
   2  * Copyright (c) 2009 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  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
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34 /*
  35  * NFSIOD operations - now built into the kernel.
  36  */
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/proc.h>
  40 #include <sys/malloc.h>
  41 #include <sys/mount.h>
  42 #include <sys/kernel.h>
  43 #include <sys/mbuf.h>
  44 #include <sys/vnode.h>
  45 #include <sys/fcntl.h>
  46 #include <sys/protosw.h>
  47 #include <sys/resourcevar.h>
  48 #include <sys/socket.h>
  49 #include <sys/socketvar.h>
  50 #include <sys/socketops.h>
  51 #include <sys/syslog.h>
  52 #include <sys/thread.h>
  53 #include <sys/tprintf.h>
  54 #include <sys/sysctl.h>
  55 #include <sys/signalvar.h>
  56 #include <sys/mutex.h>
  57
  58 #include <sys/signal2.h>
  59 #include <sys/thread2.h>
  60 #include <sys/mutex2.h>
  61
  62 #include <netinet/in.h>
  63 #include <netinet/tcp.h>
  64
  65 #include "rpcv2.h"
  66 #include "nfsproto.h"
  67 #include "nfs.h"
  68 #include "xdr_subs.h"
  69 #include "nfsm_subs.h"
  70 #include "nfsmount.h"
  71 #include "nfsnode.h"
  72 #include "nfsrtt.h"
  73
  74 /*
  75  * nfs service connection reader thread
  76  */
  77 void
  78 nfssvc_iod_reader(void *arg)
  79 {
  80         struct nfsmount *nmp = arg;
  81         struct nfsm_info *info;
  82         struct nfsreq *req;
  83         int error;
  84
  85         lwkt_gettoken(&nmp->nm_token);
  86
  87         if (nmp->nm_rxstate == NFSSVC_INIT)
  88                 nmp->nm_rxstate = NFSSVC_PENDING;
  89         for (;;) {
  90                 if (nmp->nm_rxstate == NFSSVC_WAITING) {
  91                         if (TAILQ_FIRST(&nmp->nm_reqq) == NULL &&
  92                             TAILQ_FIRST(&nmp->nm_reqrxq) == NULL) {
  93                                 tsleep(&nmp->nm_rxstate, 0, "nfsidl", 0);
  94                         } else {
  95                                 /*
  96                                  * This can happen during shutdown, we don't
  97                                  * want to hardloop.
  98                                  */
  99                                 error = nfs_reply(nmp, NULL);
 100                                 if (error && error != EWOULDBLOCK) {
 101                                         tsleep(&nmp->nm_rxstate, 0,
 102                                                 "nfsxxx", hz / 10);
 103                                 }
 104                         }
 105                         continue;
 106                 }
 107                 if (nmp->nm_rxstate != NFSSVC_PENDING)
 108                         break;
 109                 nmp->nm_rxstate = NFSSVC_WAITING;
 110
 111                 /*
 112                  * Process requests which have received replies.  Only
 113                  * process the post-reply states.  If we get EINPROGRESS
 114                  * it means the request went back to an auth or retransmit
 115                  * state and we let the iod_writer thread deal with it.
 116                  *
 117                  * Any lock on the request is strictly temporary due to
 118                  * MP races (XXX).
 119                  *
 120                  * If the request completes we run the info->done call
 121                  * to finish up the I/O.
 122                  */
 123                 while ((req = TAILQ_FIRST(&nmp->nm_reqrxq)) != NULL) {
 124                         if (req->r_flags & R_LOCKED) {
 125                                 while (req->r_flags & R_LOCKED) {
 126                                         req->r_flags |= R_WANTED;
 127                                         tsleep(req, 0, "nfstrac", 0);
 128                                 }
 129                                 continue;
 130                         }
 131                         TAILQ_REMOVE(&nmp->nm_reqrxq, req, r_chain);
 132                         info = req->r_info;
 133                         KKASSERT(info);
 134                         info->error = nfs_request(info,
 135                                                   NFSM_STATE_PROCESSREPLY,
 136                                                   NFSM_STATE_DONE);
 137                         if (info->error == EINPROGRESS) {
 138                                 kprintf("rxq: move info %p back to txq\n", info);
 139                                 TAILQ_INSERT_TAIL(&nmp->nm_reqtxq, req, r_chain);
 140                                 nfssvc_iod_writer_wakeup(nmp);
 141                         } else {
 142                                 atomic_subtract_int(&nmp->nm_bioqlen, 1);
 143                                 info->done(info);
 144                         }
 145                 }
 146         }
 147         nmp->nm_rxthread = NULL;
 148         nmp->nm_rxstate = NFSSVC_DONE;
 149
 150         lwkt_reltoken(&nmp->nm_token);
 151         wakeup(&nmp->nm_rxthread);
 152 }
 153
 154 /*
 155  * nfs service connection writer thread
 156  *
 157  * The writer sits on the send side of the client's socket and
 158  * does both the initial processing of BIOs and also transmission
 159  * and retransmission of nfsreq's.
 160  *
 161  * The writer processes both new BIOs from nm_bioq and retransmit
 162  * or state machine jumpbacks from nm_reqtxq
 163  */
 164 void
 165 nfssvc_iod_writer(void *arg)
 166 {
 167         struct nfsmount *nmp = arg;
 168         struct bio *bio;
 169         struct nfsreq *req;
 170         struct vnode *vp;
 171         nfsm_info_t info;
 172
 173         lwkt_gettoken(&nmp->nm_token);
 174
 175         if (nmp->nm_txstate == NFSSVC_INIT)
 176                 nmp->nm_txstate = NFSSVC_PENDING;
 177
 178         for (;;) {
 179                 if (nmp->nm_txstate == NFSSVC_WAITING) {
 180                         tsleep(&nmp->nm_txstate, 0, "nfsidl", 0);
 181                         continue;
 182                 }
 183                 if (nmp->nm_txstate != NFSSVC_PENDING)
 184                         break;
 185                 nmp->nm_txstate = NFSSVC_WAITING;
 186
 187                 /*
 188                  * Eep, we could blow out the mbuf allocator if we just
 189                  * did everything the kernel wanted us to do.
 190                  */
 191                 while ((bio = TAILQ_FIRST(&nmp->nm_bioq)) != NULL) {
 192                         if (nmp->nm_reqqlen > nfs_maxasyncbio)
 193                                 break;
 194                         TAILQ_REMOVE(&nmp->nm_bioq, bio, bio_act);
 195                         vp = bio->bio_driver_info;
 196                         nfs_startio(vp, bio, NULL);
 197                 }
 198
 199                 /*
 200                  * Process reauths & retransmits.  If we get an EINPROGRESS
 201                  * it means the state transitioned to WAITREPLY or later.
 202                  * Otherwise the request completed (probably with an error
 203                  * since we didn't get to a replied state).
 204                  */
 205                 while ((req = TAILQ_FIRST(&nmp->nm_reqtxq)) != NULL) {
 206                         TAILQ_REMOVE(&nmp->nm_reqtxq, req, r_chain);
 207                         info = req->r_info;
 208                         KKASSERT(info);
 209                         info->error = nfs_request(info,
 210                                                   NFSM_STATE_AUTH,
 211                                                   NFSM_STATE_WAITREPLY);
 212                         if (info->error == EINPROGRESS) {
 213                                 ;
 214                         } else {
 215                                 atomic_subtract_int(&nmp->nm_bioqlen, 1);
 216                                 info->done(info);
 217                         }
 218                 }
 219         }
 220         nmp->nm_txthread = NULL;
 221         nmp->nm_txstate = NFSSVC_DONE;
 222         lwkt_reltoken(&nmp->nm_token);
 223         wakeup(&nmp->nm_txthread);
 224 }
 225
 226 void
 227 nfssvc_iod_stop1(struct nfsmount *nmp)
 228 {
 229         nmp->nm_txstate = NFSSVC_STOPPING;
 230         nmp->nm_rxstate = NFSSVC_STOPPING;
 231 }
 232
 233 void
 234 nfssvc_iod_stop2(struct nfsmount *nmp)
 235 {
 236         wakeup(&nmp->nm_txstate);
 237         while (nmp->nm_txthread)
 238                 tsleep(&nmp->nm_txthread, 0, "nfssttx", hz*2);
 239         wakeup(&nmp->nm_rxstate);
 240         while (nmp->nm_rxthread)
 241                 tsleep(&nmp->nm_rxthread, 0, "nfsstrx", hz*2);
 242 }
 243
 244 void
 245 nfssvc_iod_writer_wakeup(struct nfsmount *nmp)
 246 {
 247         if (nmp->nm_txstate == NFSSVC_WAITING) {
 248                 nmp->nm_txstate = NFSSVC_PENDING;
 249                 wakeup(&nmp->nm_txstate);
 250         }
 251 }
 252
 253 void
 254 nfssvc_iod_reader_wakeup(struct nfsmount *nmp)
 255 {
 256         if (nmp->nm_rxstate == NFSSVC_WAITING) {
 257                 nmp->nm_rxstate = NFSSVC_PENDING;
 258                 wakeup(&nmp->nm_rxstate);
 259         }
 260 }