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
  57 #include <sys/signal2.h>
  58 #include <sys/thread2.h>
  59 #include <sys/mutex2.h>
  60
  61 #include <netinet/in.h>
  62 #include <netinet/tcp.h>
  63
  64 #include "rpcv2.h"
  65 #include "nfsproto.h"
  66 #include "nfs.h"
  67 #include "xdr_subs.h"
  68 #include "nfsm_subs.h"
  69 #include "nfsmount.h"
  70 #include "nfsnode.h"
  71 #include "nfsrtt.h"
  72
  73 /*
  74  * nfs service connection reader thread
  75  */
  76 void
  77 nfssvc_iod_reader(void *arg)
  78 {
  79         struct nfsmount *nmp = arg;
  80         struct nfsm_info *info;
  81         struct nfsreq *req;
  82         int error;
  83
  84         lwkt_gettoken(&nmp->nm_token);
  85
  86         if (nmp->nm_rxstate == NFSSVC_INIT)
  87                 nmp->nm_rxstate = NFSSVC_PENDING;
  88         for (;;) {
  89                 if (nmp->nm_rxstate == NFSSVC_WAITING) {
  90                         if (TAILQ_FIRST(&nmp->nm_reqq) == NULL &&
  91                             TAILQ_FIRST(&nmp->nm_reqrxq) == NULL) {
  92                                 tsleep(&nmp->nm_rxstate, 0, "nfsidl", 0);
  93                         } else {
  94                                 /*
  95                                  * This can happen during shutdown, we don't
  96                                  * want to hardloop.
  97                                  */
  98                                 error = nfs_reply(nmp, NULL);
  99                                 if (error && error != EWOULDBLOCK) {
 100                                         tsleep(&nmp->nm_rxstate, 0,
 101                                                 "nfsxxx", hz / 10);
 102                                 }
 103                         }
 104                         continue;
 105                 }
 106                 if (nmp->nm_rxstate != NFSSVC_PENDING)
 107                         break;
 108                 nmp->nm_rxstate = NFSSVC_WAITING;
 109
 110                 /*
 111                  * Process requests which have received replies.  Only
 112                  * process the post-reply states.  If we get EINPROGRESS
 113                  * it means the request went back to an auth or retransmit
 114                  * state and we let the iod_writer thread deal with it.
 115                  *
 116                  * Any lock on the request is strictly temporary due to
 117                  * MP races (XXX).
 118                  *
 119                  * If the request completes we run the info->done call
 120                  * to finish up the I/O.
 121                  */
 122                 while ((req = TAILQ_FIRST(&nmp->nm_reqrxq)) != NULL) {
 123                         if (req->r_flags & R_LOCKED) {
 124                                 while (req->r_flags & R_LOCKED) {
 125                                         req->r_flags |= R_WANTED;
 126                                         tsleep(req, 0, "nfstrac", 0);
 127                                 }
 128                                 continue;
 129                         }
 130                         TAILQ_REMOVE(&nmp->nm_reqrxq, req, r_chain);
 131                         info = req->r_info;
 132                         KKASSERT(info);
 133                         info->error = nfs_request(info,
 134                                                   NFSM_STATE_PROCESSREPLY,
 135                                                   NFSM_STATE_DONE);
 136                         if (info->error == EINPROGRESS) {
 137                                 kprintf("rxq: move info %p back to txq\n", info);
 138                                 TAILQ_INSERT_TAIL(&nmp->nm_reqtxq, req, r_chain);
 139                                 nfssvc_iod_writer_wakeup(nmp);
 140                         } else {
 141                                 atomic_subtract_int(&nmp->nm_bioqlen, 1);
 142                                 info->done(info);
 143                         }
 144                 }
 145         }
 146         nmp->nm_rxthread = NULL;
 147         nmp->nm_rxstate = NFSSVC_DONE;
 148
 149         lwkt_reltoken(&nmp->nm_token);
 150         wakeup(&nmp->nm_rxthread);
 151 }
 152
 153 /*
 154  * nfs service connection writer thread
 155  *
 156  * The writer sits on the send side of the client's socket and
 157  * does both the initial processing of BIOs and also transmission
 158  * and retransmission of nfsreq's.
 159  *
 160  * The writer processes both new BIOs from nm_bioq and retransmit
 161  * or state machine jumpbacks from nm_reqtxq
 162  */
 163 void
 164 nfssvc_iod_writer(void *arg)
 165 {
 166         struct nfsmount *nmp = arg;
 167         struct bio *bio;
 168         struct nfsreq *req;
 169         struct vnode *vp;
 170         nfsm_info_t info;
 171
 172         lwkt_gettoken(&nmp->nm_token);
 173
 174         if (nmp->nm_txstate == NFSSVC_INIT)
 175                 nmp->nm_txstate = NFSSVC_PENDING;
 176
 177         for (;;) {
 178                 if (nmp->nm_txstate == NFSSVC_WAITING) {
 179                         tsleep(&nmp->nm_txstate, 0, "nfsidl", 0);
 180                         continue;
 181                 }
 182                 if (nmp->nm_txstate != NFSSVC_PENDING)
 183                         break;
 184                 nmp->nm_txstate = NFSSVC_WAITING;
 185
 186                 /*
 187                  * Eep, we could blow out the mbuf allocator if we just
 188                  * did everything the kernel wanted us to do.
 189                  */
 190                 while ((bio = TAILQ_FIRST(&nmp->nm_bioq)) != NULL) {
 191                         if (nmp->nm_reqqlen > nfs_maxasyncbio)
 192                                 break;
 193                         TAILQ_REMOVE(&nmp->nm_bioq, bio, bio_act);
 194                         vp = bio->bio_driver_info;
 195                         nfs_startio(vp, bio, NULL);
 196                 }
 197
 198                 /*
 199                  * Process reauths & retransmits.  If we get an EINPROGRESS
 200                  * it means the state transitioned to WAITREPLY or later.
 201                  * Otherwise the request completed (probably with an error
 202                  * since we didn't get to a replied state).
 203                  */
 204                 while ((req = TAILQ_FIRST(&nmp->nm_reqtxq)) != NULL) {
 205                         TAILQ_REMOVE(&nmp->nm_reqtxq, req, r_chain);
 206                         info = req->r_info;
 207                         KKASSERT(info);
 208                         info->error = nfs_request(info,
 209                                                   NFSM_STATE_AUTH,
 210                                                   NFSM_STATE_WAITREPLY);
 211                         if (info->error == EINPROGRESS) {
 212                                 ;
 213                         } else {
 214                                 atomic_subtract_int(&nmp->nm_bioqlen, 1);
 215                                 info->done(info);
 216                         }
 217                 }
 218         }
 219         nmp->nm_txthread = NULL;
 220         nmp->nm_txstate = NFSSVC_DONE;
 221         lwkt_reltoken(&nmp->nm_token);
 222         wakeup(&nmp->nm_txthread);
 223 }
 224
 225 void
 226 nfssvc_iod_stop1(struct nfsmount *nmp)
 227 {
 228         nmp->nm_txstate = NFSSVC_STOPPING;
 229         nmp->nm_rxstate = NFSSVC_STOPPING;
 230 }
 231
 232 void
 233 nfssvc_iod_stop2(struct nfsmount *nmp)
 234 {
 235         wakeup(&nmp->nm_txstate);
 236         while (nmp->nm_txthread)
 237                 tsleep(&nmp->nm_txthread, 0, "nfssttx", hz*2);
 238         wakeup(&nmp->nm_rxstate);
 239         while (nmp->nm_rxthread)
 240                 tsleep(&nmp->nm_rxthread, 0, "nfsstrx", hz*2);
 241 }
 242
 243 void
 244 nfssvc_iod_writer_wakeup(struct nfsmount *nmp)
 245 {
 246         if (nmp->nm_txstate == NFSSVC_WAITING) {
 247                 nmp->nm_txstate = NFSSVC_PENDING;
 248                 wakeup(&nmp->nm_txstate);
 249         }
 250 }
 251
 252 void
 253 nfssvc_iod_reader_wakeup(struct nfsmount *nmp)
 254 {
 255         if (nmp->nm_rxstate == NFSSVC_WAITING) {
 256                 nmp->nm_rxstate = NFSSVC_PENDING;
 257                 wakeup(&nmp->nm_rxstate);
 258         }
 259 }