sys/vfs/nfs/nfs_srvcache.c

   1 /*
   2  * Copyright (c) 1989, 1993
   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_srvcache.c      8.3 (Berkeley) 3/30/95
  37  * $FreeBSD: src/sys/nfs/nfs_srvcache.c,v 1.21 2000/02/13 03:32:06 peter Exp $
  38  * $DragonFly: src/sys/vfs/nfs/nfs_srvcache.c,v 1.13 2008/01/05 14:02:41 swildner Exp $
  39  */
  40
  41 /*
  42  * Reference: Chet Juszczak, "Improving the Performance and Correctness
  43  *              of an NFS Server", in Proc. Winter 1989 USENIX Conference,
  44  *              pages 53-63. San Diego, February 1989.
  45  */
  46 #include <sys/param.h>
  47 #include <sys/malloc.h>
  48 #include <sys/mount.h>
  49 #include <sys/systm.h>
  50 #include <sys/mbuf.h>
  51 #include <sys/socket.h>
  52 #include <sys/socketvar.h>      /* for dup_sockaddr */
  53
  54 #include <netinet/in.h>
  55 #include "rpcv2.h"
  56 #include "nfsproto.h"
  57 #include "nfs.h"
  58 #include "nfsrvcache.h"
  59
  60 #ifndef NFS_NOSERVER
  61 static long numnfsrvcache;
  62 static long desirednfsrvcache = NFSRVCACHESIZ;
  63
  64 #define NFSRCHASH(xid) \
  65         (&nfsrvhashtbl[((xid) + ((xid) >> 24)) & nfsrvhash])
  66 static LIST_HEAD(nfsrvhash, nfsrvcache) *nfsrvhashtbl;
  67 static TAILQ_HEAD(nfsrvlru, nfsrvcache) nfsrvlruhead;
  68 static u_long nfsrvhash;
  69
  70 #define TRUE    1
  71 #define FALSE   0
  72
  73 #define NETFAMILY(rp) \
  74                 (((rp)->rc_flag & RC_INETADDR) ? AF_INET : AF_ISO)
  75
  76 struct lwkt_token srvcache_token = LWKT_TOKEN_MP_INITIALIZER(srvcache_token);
  77
  78 /*
  79  * Static array that defines which nfs rpc's are nonidempotent
  80  */
  81 static int nonidempotent[NFS_NPROCS] = {
  82         FALSE,
  83         FALSE,
  84         TRUE,
  85         FALSE,
  86         FALSE,
  87         FALSE,
  88         FALSE,
  89         TRUE,
  90         TRUE,
  91         TRUE,
  92         TRUE,
  93         TRUE,
  94         TRUE,
  95         TRUE,
  96         TRUE,
  97         TRUE,
  98         FALSE,
  99         FALSE,
 100         FALSE,
 101         FALSE,
 102         FALSE,
 103         FALSE,
 104         FALSE,
 105         FALSE,
 106         FALSE,
 107         FALSE,
 108 };
 109
 110 /* True iff the rpc reply is an nfs status ONLY! */
 111 static int nfsv2_repstat[NFS_NPROCS] = {
 112         FALSE,
 113         FALSE,
 114         FALSE,
 115         FALSE,
 116         FALSE,
 117         FALSE,
 118         FALSE,
 119         FALSE,
 120         FALSE,
 121         FALSE,
 122         TRUE,
 123         TRUE,
 124         TRUE,
 125         TRUE,
 126         FALSE,
 127         TRUE,
 128         FALSE,
 129         FALSE,
 130 };
 131
 132 /*
 133  * Initialize the server request cache list
 134  */
 135 void
 136 nfsrv_initcache(void)
 137 {
 138         nfsrvhashtbl = hashinit(desirednfsrvcache, M_NFSD, &nfsrvhash);
 139         TAILQ_INIT(&nfsrvlruhead);
 140 }
 141
 142 /*
 143  * Look for the request in the cache
 144  * If found then
 145  *    return action and optionally reply
 146  * else
 147  *    insert it in the cache
 148  *
 149  * The rules are as follows:
 150  * - if in progress, return DROP request
 151  * - if completed within DELAY of the current time, return DROP it
 152  * - if completed a longer time ago return REPLY if the reply was cached or
 153  *   return DOIT
 154  * Update/add new request at end of lru list
 155  */
 156 int
 157 nfsrv_getcache(struct nfsrv_descript *nd, struct nfssvc_sock *slp,
 158                struct mbuf **repp)
 159 {
 160         struct nfsrvcache *rp;
 161         struct mbuf *mb;
 162         struct sockaddr_in *saddr;
 163         caddr_t bpos;
 164         int ret;
 165
 166         /*
 167          * Don't cache recent requests for reliable transport protocols.
 168          * (Maybe we should for the case of a reconnect, but..)
 169          */
 170         if (!nd->nd_nam2)
 171                 return (RC_DOIT);
 172
 173         lwkt_gettoken(&srvcache_token);
 174 loop:
 175         for (rp = NFSRCHASH(nd->nd_retxid)->lh_first; rp != 0;
 176             rp = rp->rc_hash.le_next) {
 177             if (nd->nd_retxid == rp->rc_xid && nd->nd_procnum == rp->rc_proc &&
 178                 netaddr_match(NETFAMILY(rp), &rp->rc_haddr, nd->nd_nam)) {
 179                         NFS_DPF(RC, ("H%03x", rp->rc_xid & 0xfff));
 180                         if ((rp->rc_flag & RC_LOCKED) != 0) {
 181                                 rp->rc_flag |= RC_WANTED;
 182                                 tsleep((caddr_t)rp, 0, "nfsrc", 0);
 183                                 goto loop;
 184                         }
 185                         rp->rc_flag |= RC_LOCKED;
 186                         /* If not at end of LRU chain, move it there */
 187                         if (TAILQ_NEXT(rp, rc_lru) != NULL) {
 188                                 TAILQ_REMOVE(&nfsrvlruhead, rp, rc_lru);
 189                                 TAILQ_INSERT_TAIL(&nfsrvlruhead, rp, rc_lru);
 190                         }
 191                         if (rp->rc_state == RC_UNUSED)
 192                                 panic("nfsrv cache");
 193                         if (rp->rc_state == RC_INPROG) {
 194                                 nfsstats.srvcache_inproghits++;
 195                                 ret = RC_DROPIT;
 196                         } else if (rp->rc_flag & RC_REPSTATUS) {
 197                                 nfsstats.srvcache_nonidemdonehits++;
 198                                 nfs_rephead(0, nd, slp, rp->rc_status,
 199                                             repp, &mb, &bpos);
 200                                 ret = RC_REPLY;
 201                         } else if (rp->rc_flag & RC_REPMBUF) {
 202                                 nfsstats.srvcache_nonidemdonehits++;
 203                                 *repp = m_copym(rp->rc_reply, 0, M_COPYALL,
 204                                                 MB_WAIT);
 205                                 ret = RC_REPLY;
 206                         } else {
 207                                 nfsstats.srvcache_idemdonehits++;
 208                                 rp->rc_state = RC_INPROG;
 209                                 ret = RC_DOIT;
 210                         }
 211                         rp->rc_flag &= ~RC_LOCKED;
 212                         if (rp->rc_flag & RC_WANTED) {
 213                                 rp->rc_flag &= ~RC_WANTED;
 214                                 wakeup((caddr_t)rp);
 215                         }
 216                         lwkt_reltoken(&srvcache_token);
 217                         return (ret);
 218                 }
 219         }
 220
 221         nfsstats.srvcache_misses++;
 222         NFS_DPF(RC, ("M%03x", nd->nd_retxid & 0xfff));
 223         if (numnfsrvcache < desirednfsrvcache) {
 224                 rp = kmalloc((u_long)sizeof *rp, M_NFSD, M_WAITOK | M_ZERO);
 225                 numnfsrvcache++;
 226                 rp->rc_flag = RC_LOCKED;
 227         } else {
 228                 rp = TAILQ_FIRST(&nfsrvlruhead);
 229                 while ((rp->rc_flag & RC_LOCKED) != 0) {
 230                         rp->rc_flag |= RC_WANTED;
 231                         tsleep((caddr_t)rp, 0, "nfsrc", 0);
 232                         rp = TAILQ_FIRST(&nfsrvlruhead);
 233                 }
 234                 rp->rc_flag |= RC_LOCKED;
 235                 LIST_REMOVE(rp, rc_hash);
 236                 TAILQ_REMOVE(&nfsrvlruhead, rp, rc_lru);
 237                 if (rp->rc_flag & RC_REPMBUF) {
 238                         m_freem(rp->rc_reply);
 239                         rp->rc_reply = NULL;
 240                         rp->rc_flag &= ~RC_REPMBUF;
 241                 }
 242                 if (rp->rc_flag & RC_NAM) {
 243                         FREE(rp->rc_nam, M_SONAME);
 244                         rp->rc_nam = NULL;
 245                         rp->rc_flag &= ~RC_NAM;
 246                 }
 247         }
 248         TAILQ_INSERT_TAIL(&nfsrvlruhead, rp, rc_lru);
 249
 250         rp->rc_state = RC_INPROG;
 251         rp->rc_xid = nd->nd_retxid;
 252         saddr = (struct sockaddr_in *)nd->nd_nam;
 253         switch (saddr->sin_family) {
 254         case AF_INET:
 255                 rp->rc_flag |= RC_INETADDR;
 256                 rp->rc_inetaddr = saddr->sin_addr.s_addr;
 257                 break;
 258         case AF_ISO:
 259         default:
 260                 rp->rc_flag |= RC_NAM;
 261                 rp->rc_nam = dup_sockaddr(nd->nd_nam);
 262                 break;
 263         };
 264         rp->rc_proc = nd->nd_procnum;
 265         LIST_INSERT_HEAD(NFSRCHASH(nd->nd_retxid), rp, rc_hash);
 266         rp->rc_flag &= ~RC_LOCKED;
 267         if (rp->rc_flag & RC_WANTED) {
 268                 rp->rc_flag &= ~RC_WANTED;
 269                 wakeup((caddr_t)rp);
 270         }
 271         lwkt_reltoken(&srvcache_token);
 272
 273         return (RC_DOIT);
 274 }
 275
 276 /*
 277  * Update a request cache entry after the rpc has been done
 278  */
 279 void
 280 nfsrv_updatecache(struct nfsrv_descript *nd, int repvalid, struct mbuf *repmbuf)
 281 {
 282         struct nfsrvcache *rp;
 283
 284         if (!nd->nd_nam2)
 285                 return;
 286
 287         lwkt_gettoken(&srvcache_token);
 288 loop:
 289         for (rp = NFSRCHASH(nd->nd_retxid)->lh_first; rp != 0;
 290             rp = rp->rc_hash.le_next) {
 291             if (nd->nd_retxid == rp->rc_xid && nd->nd_procnum == rp->rc_proc &&
 292                 netaddr_match(NETFAMILY(rp), &rp->rc_haddr, nd->nd_nam)) {
 293                         NFS_DPF(RC, ("U%03x", rp->rc_xid & 0xfff));
 294                         if ((rp->rc_flag & RC_LOCKED) != 0) {
 295                                 rp->rc_flag |= RC_WANTED;
 296                                 tsleep((caddr_t)rp, 0, "nfsrc", 0);
 297                                 goto loop;
 298                         }
 299                         rp->rc_flag |= RC_LOCKED;
 300                         if (rp->rc_state == RC_DONE) {
 301                                 /*
 302                                  * This can occur if the cache is too small.
 303                                  * Retransmits of the same request aren't
 304                                  * dropped so we may see the operation
 305                                  * complete more then once.
 306                                  */
 307                                 if (rp->rc_flag & RC_REPMBUF) {
 308                                         m_freem(rp->rc_reply);
 309                                         rp->rc_reply = NULL;
 310                                         rp->rc_flag &= ~RC_REPMBUF;
 311                                 }
 312                         }
 313                         rp->rc_state = RC_DONE;
 314
 315                         /*
 316                          * If we have a valid reply update status and save
 317                          * the reply for non-idempotent rpc's.
 318                          */
 319                         if (repvalid && nonidempotent[nd->nd_procnum]) {
 320                                 if ((nd->nd_flag & ND_NFSV3) == 0 &&
 321                                   nfsv2_repstat[nfsv2_procid[nd->nd_procnum]]) {
 322                                         rp->rc_status = nd->nd_repstat;
 323                                         rp->rc_flag |= RC_REPSTATUS;
 324                                 } else {
 325                                         if (rp->rc_flag & RC_REPMBUF) {
 326                                                 m_freem(rp->rc_reply);
 327                                                 rp->rc_reply = NULL;
 328                                                 rp->rc_flag &= ~RC_REPMBUF;
 329                                         }
 330                                         rp->rc_reply = m_copym(repmbuf, 0,
 331                                                         M_COPYALL, MB_WAIT);
 332                                         rp->rc_flag |= RC_REPMBUF;
 333                                 }
 334                         }
 335                         rp->rc_flag &= ~RC_LOCKED;
 336                         if (rp->rc_flag & RC_WANTED) {
 337                                 rp->rc_flag &= ~RC_WANTED;
 338                                 wakeup((caddr_t)rp);
 339                         }
 340                         break;
 341                 }
 342         }
 343         lwkt_reltoken(&srvcache_token);
 344         NFS_DPF(RC, ("L%03x", nd->nd_retxid & 0xfff));
 345 }
 346
 347 /*
 348  * Clean out the cache. Called when the last nfsd terminates.
 349  */
 350 void
 351 nfsrv_cleancache(void)
 352 {
 353         struct nfsrvcache *rp;
 354
 355         lwkt_gettoken(&srvcache_token);
 356         while ((rp = TAILQ_FIRST(&nfsrvlruhead)) != NULL) {
 357                 if (rp->rc_flag & RC_LOCKED) {
 358                         rp->rc_flag |= RC_WANTED;
 359                         tsleep((caddr_t)rp, 0, "nfsrc", 0);
 360                         continue;
 361                 }
 362                 LIST_REMOVE(rp, rc_hash);
 363                 TAILQ_REMOVE(&nfsrvlruhead, rp, rc_lru);
 364                 if (rp->rc_flag & RC_REPMBUF) {
 365                         m_freem(rp->rc_reply);
 366                         rp->rc_reply = NULL;
 367                         rp->rc_flag &= ~RC_REPMBUF;
 368                 }
 369                 if (rp->rc_flag & RC_NAM) {
 370                         kfree(rp->rc_nam, M_SONAME);
 371                         rp->rc_nam = NULL;
 372                         rp->rc_flag &= ~RC_NAM;
 373                 }
 374                 kfree(rp, M_NFSD);
 375         }
 376         numnfsrvcache = 0;
 377         lwkt_reltoken(&srvcache_token);
 378 }
 379
 380 #endif /* NFS_NOSERVER */