1 /* 2 * Copyright (c) 1989 The Regents of the University of California. 3 * 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 * %sccs.include.redist.c% 9 * 10 * @(#)nfs_srvcache.c 7.16 (Berkeley) 07/22/92 11 */ 12 13 /* 14 * Reference: Chet Juszczak, "Improving the Performance and Correctness 15 * of an NFS Server", in Proc. Winter 1989 USENIX Conference, 16 * pages 53-63. San Diego, February 1989. 17 */ 18 #include <sys/param.h> 19 #include <sys/vnode.h> 20 #include <sys/mount.h> 21 #include <sys/kernel.h> 22 #include <sys/systm.h> 23 #include <sys/proc.h> 24 #include <sys/mbuf.h> 25 #include <sys/malloc.h> 26 #include <sys/socket.h> 27 #include <sys/socketvar.h> 28 #include <netinet/in.h> 29 #ifdef ISO 30 #include <netiso/iso.h> 31 #endif 32 #include <nfs/nfsm_subs.h> 33 #include <nfs/rpcv2.h> 34 #include <nfs/nfsv2.h> 35 #include <nfs/nfs.h> 36 #include <nfs/nfsrvcache.h> 37 #include <nfs/nqnfs.h> 38 39 long numnfsrvcache, desirednfsrvcache = NFSRVCACHESIZ; 40 41 #define NFSRCHASH(xid) (((xid) + ((xid) >> 16)) & rheadhash) 42 static struct nfsrvcache *nfsrvlruhead, **nfsrvlrutail = &nfsrvlruhead; 43 static struct nfsrvcache **rheadhtbl; 44 static u_long rheadhash; 45 46 #define TRUE 1 47 #define FALSE 0 48 49 #define NETFAMILY(rp) \ 50 (((rp)->rc_flag & RC_INETADDR) ? AF_INET : AF_ISO) 51 52 /* 53 * Static array that defines which nfs rpc's are nonidempotent 54 */ 55 int nonidempotent[NFS_NPROCS] = { 56 FALSE, 57 FALSE, 58 TRUE, 59 FALSE, 60 FALSE, 61 FALSE, 62 FALSE, 63 FALSE, 64 TRUE, 65 TRUE, 66 TRUE, 67 TRUE, 68 TRUE, 69 TRUE, 70 TRUE, 71 TRUE, 72 FALSE, 73 FALSE, 74 FALSE, 75 FALSE, 76 FALSE, 77 FALSE, 78 }; 79 80 /* True iff the rpc reply is an nfs status ONLY! */ 81 static int repliesstatus[NFS_NPROCS] = { 82 FALSE, 83 FALSE, 84 FALSE, 85 FALSE, 86 FALSE, 87 FALSE, 88 FALSE, 89 FALSE, 90 FALSE, 91 FALSE, 92 TRUE, 93 TRUE, 94 TRUE, 95 TRUE, 96 FALSE, 97 TRUE, 98 FALSE, 99 FALSE, 100 FALSE, 101 FALSE, 102 FALSE, 103 FALSE, 104 }; 105 106 /* 107 * Initialize the server request cache list 108 */ 109 nfsrv_initcache() 110 { 111 112 rheadhtbl = hashinit(desirednfsrvcache, M_NFSD, &rheadhash); 113 } 114 115 /* 116 * Look for the request in the cache 117 * If found then 118 * return action and optionally reply 119 * else 120 * insert it in the cache 121 * 122 * The rules are as follows: 123 * - if in progress, return DROP request 124 * - if completed within DELAY of the current time, return DROP it 125 * - if completed a longer time ago return REPLY if the reply was cached or 126 * return DOIT 127 * Update/add new request at end of lru list 128 */ 129 nfsrv_getcache(nam, nd, repp) 130 struct mbuf *nam; 131 register struct nfsd *nd; 132 struct mbuf **repp; 133 { 134 register struct nfsrvcache *rp, *rq, **rpp; 135 struct mbuf *mb; 136 struct sockaddr_in *saddr; 137 caddr_t bpos; 138 int ret; 139 140 if (nd->nd_nqlflag != NQL_NOVAL) 141 return (RC_DOIT); 142 rpp = &rheadhtbl[NFSRCHASH(nd->nd_retxid)]; 143 loop: 144 for (rp = *rpp; rp; rp = rp->rc_forw) { 145 if (nd->nd_retxid == rp->rc_xid && nd->nd_procnum == rp->rc_proc && 146 netaddr_match(NETFAMILY(rp), &rp->rc_haddr, (union nethostaddr *)0, nam)) { 147 if ((rp->rc_flag & RC_LOCKED) != 0) { 148 rp->rc_flag |= RC_WANTED; 149 (void) tsleep((caddr_t)rp, PZERO-1, "nfsrc", 0); 150 goto loop; 151 } 152 rp->rc_flag |= RC_LOCKED; 153 /* If not at end of LRU chain, move it there */ 154 if (rp->rc_next) { 155 /* remove from LRU chain */ 156 *rp->rc_prev = rp->rc_next; 157 rp->rc_next->rc_prev = rp->rc_prev; 158 /* and replace at end of it */ 159 rp->rc_next = NULL; 160 rp->rc_prev = nfsrvlrutail; 161 *nfsrvlrutail = rp; 162 nfsrvlrutail = &rp->rc_next; 163 } 164 if (rp->rc_state == RC_UNUSED) 165 panic("nfsrv cache"); 166 if (rp->rc_state == RC_INPROG || 167 (time.tv_sec - rp->rc_timestamp) < RC_DELAY) { 168 nfsstats.srvcache_inproghits++; 169 ret = RC_DROPIT; 170 } else if (rp->rc_flag & RC_REPSTATUS) { 171 nfsstats.srvcache_idemdonehits++; 172 nfs_rephead(0, nd, rp->rc_status, 173 0, (u_quad_t *)0, repp, &mb, &bpos); 174 rp->rc_timestamp = time.tv_sec; 175 ret = RC_REPLY; 176 } else if (rp->rc_flag & RC_REPMBUF) { 177 nfsstats.srvcache_idemdonehits++; 178 *repp = m_copym(rp->rc_reply, 0, M_COPYALL, 179 M_WAIT); 180 rp->rc_timestamp = time.tv_sec; 181 ret = RC_REPLY; 182 } else { 183 nfsstats.srvcache_nonidemdonehits++; 184 rp->rc_state = RC_INPROG; 185 ret = RC_DOIT; 186 } 187 rp->rc_flag &= ~RC_LOCKED; 188 if (rp->rc_flag & RC_WANTED) { 189 rp->rc_flag &= ~RC_WANTED; 190 wakeup((caddr_t)rp); 191 } 192 return (ret); 193 } 194 } 195 nfsstats.srvcache_misses++; 196 if (numnfsrvcache < desirednfsrvcache) { 197 rp = (struct nfsrvcache *)malloc((u_long)sizeof *rp, 198 M_NFSD, M_WAITOK); 199 bzero((char *)rp, sizeof *rp); 200 numnfsrvcache++; 201 rp->rc_flag = RC_LOCKED; 202 } else { 203 rp = nfsrvlruhead; 204 while ((rp->rc_flag & RC_LOCKED) != 0) { 205 rp->rc_flag |= RC_WANTED; 206 (void) tsleep((caddr_t)rp, PZERO-1, "nfsrc", 0); 207 rp = nfsrvlruhead; 208 } 209 rp->rc_flag |= RC_LOCKED; 210 /* remove from hash chain */ 211 if (rq = rp->rc_forw) 212 rq->rc_back = rp->rc_back; 213 *rp->rc_back = rq; 214 /* remove from LRU chain */ 215 *rp->rc_prev = rp->rc_next; 216 rp->rc_next->rc_prev = rp->rc_prev; 217 if (rp->rc_flag & RC_REPMBUF) 218 m_freem(rp->rc_reply); 219 if (rp->rc_flag & RC_NAM) 220 MFREE(rp->rc_nam, mb); 221 rp->rc_flag &= (RC_LOCKED | RC_WANTED); 222 } 223 /* place at end of LRU list */ 224 rp->rc_next = NULL; 225 rp->rc_prev = nfsrvlrutail; 226 *nfsrvlrutail = rp; 227 nfsrvlrutail = &rp->rc_next; 228 rp->rc_state = RC_INPROG; 229 rp->rc_xid = nd->nd_retxid; 230 saddr = mtod(nam, struct sockaddr_in *); 231 switch (saddr->sin_family) { 232 case AF_INET: 233 rp->rc_flag |= RC_INETADDR; 234 rp->rc_inetaddr = saddr->sin_addr.s_addr; 235 break; 236 case AF_ISO: 237 default: 238 rp->rc_flag |= RC_NAM; 239 rp->rc_nam = m_copym(nam, 0, M_COPYALL, M_WAIT); 240 break; 241 }; 242 rp->rc_proc = nd->nd_procnum; 243 /* insert into hash chain */ 244 if (rq = *rpp) 245 rq->rc_back = &rp->rc_forw; 246 rp->rc_next = rq; 247 rp->rc_back = rpp; 248 *rpp = rp; 249 rp->rc_flag &= ~RC_LOCKED; 250 if (rp->rc_flag & RC_WANTED) { 251 rp->rc_flag &= ~RC_WANTED; 252 wakeup((caddr_t)rp); 253 } 254 return (RC_DOIT); 255 } 256 257 /* 258 * Update a request cache entry after the rpc has been done 259 */ 260 void 261 nfsrv_updatecache(nam, nd, repvalid, repmbuf) 262 struct mbuf *nam; 263 register struct nfsd *nd; 264 int repvalid; 265 struct mbuf *repmbuf; 266 { 267 register struct nfsrvcache *rp; 268 269 if (nd->nd_nqlflag != NQL_NOVAL) 270 return; 271 loop: 272 for (rp = rheadhtbl[NFSRCHASH(nd->nd_retxid)]; rp; rp = rp->rc_forw) { 273 if (nd->nd_retxid == rp->rc_xid && nd->nd_procnum == rp->rc_proc && 274 netaddr_match(NETFAMILY(rp), &rp->rc_haddr, (union nethostaddr *)0, nam)) { 275 if ((rp->rc_flag & RC_LOCKED) != 0) { 276 rp->rc_flag |= RC_WANTED; 277 (void) tsleep((caddr_t)rp, PZERO-1, "nfsrc", 0); 278 goto loop; 279 } 280 rp->rc_flag |= RC_LOCKED; 281 rp->rc_state = RC_DONE; 282 /* 283 * If we have a valid reply update status and save 284 * the reply for non-idempotent rpc's. 285 * Otherwise invalidate entry by setting the timestamp 286 * to nil. 287 */ 288 if (repvalid) { 289 rp->rc_timestamp = time.tv_sec; 290 if (nonidempotent[nd->nd_procnum]) { 291 if (repliesstatus[nd->nd_procnum]) { 292 rp->rc_status = nd->nd_repstat; 293 rp->rc_flag |= RC_REPSTATUS; 294 } else { 295 rp->rc_reply = m_copym(repmbuf, 296 0, M_COPYALL, M_WAIT); 297 rp->rc_flag |= RC_REPMBUF; 298 } 299 } 300 } else { 301 rp->rc_timestamp = 0; 302 } 303 rp->rc_flag &= ~RC_LOCKED; 304 if (rp->rc_flag & RC_WANTED) { 305 rp->rc_flag &= ~RC_WANTED; 306 wakeup((caddr_t)rp); 307 } 308 return; 309 } 310 } 311 } 312 313 /* 314 * Clean out the cache. Called when the last nfsd terminates. 315 */ 316 void 317 nfsrv_cleancache() 318 { 319 register struct nfsrvcache *rp, *nextrp; 320 321 for (rp = nfsrvlruhead; rp; rp = nextrp) { 322 nextrp = rp->rc_next; 323 free(rp, M_NFSD); 324 } 325 bzero((char *)rheadhtbl, (rheadhash + 1) * sizeof(void *)); 326 numnfsrvcache = 0; 327 } 328