in_rmx.c revision 9470
1/* 2 * Copyright 1994, 1995 Massachusetts Institute of Technology 3 * 4 * Permission to use, copy, modify, and distribute this software and 5 * its documentation for any purpose and without fee is hereby 6 * granted, provided that both the above copyright notice and this 7 * permission notice appear in all copies, that both the above 8 * copyright notice and this permission notice appear in all 9 * supporting documentation, and that the name of M.I.T. not be used 10 * in advertising or publicity pertaining to distribution of the 11 * software without specific, written prior permission. M.I.T. makes 12 * no representations about the suitability of this software for any 13 * purpose. It is provided "as is" without express or implied 14 * warranty. 15 * 16 * THIS SOFTWARE IS PROVIDED BY M.I.T. ``AS IS''. M.I.T. DISCLAIMS 17 * ALL EXPRESS OR IMPLIED WARRANTIES WITH REGARD TO THIS SOFTWARE, 18 * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF 19 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT 20 * SHALL M.I.T. BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF 23 * USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 24 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 25 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 26 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 * 29 * $Id: in_rmx.c,v 1.14 1995/06/21 19:48:53 wollman Exp $ 30 */ 31 32/* 33 * This code does two things necessary for the enhanced TCP metrics to 34 * function in a useful manner: 35 * 1) It marks all non-host routes as `cloning', thus ensuring that 36 * every actual reference to such a route actually gets turned 37 * into a reference to a host route to the specific destination 38 * requested. 39 * 2) When such routes lose all their references, it arranges for them 40 * to be deleted in some random collection of circumstances, so that 41 * a large quantity of stale routing data is not kept in kernel memory 42 * indefinitely. See in_rtqtimo() below for the exact mechanism. 43 */ 44 45#include <sys/param.h> 46#include <sys/systm.h> 47#include <sys/kernel.h> 48#include <sys/queue.h> 49#include <sys/socket.h> 50#include <sys/socketvar.h> 51#include <sys/mbuf.h> 52#include <sys/syslog.h> 53 54#include <net/if.h> 55#include <net/route.h> 56#include <netinet/in.h> 57#include <netinet/in_systm.h> 58#include <netinet/in_var.h> 59 60#include <netinet/ip.h> 61#include <netinet/ip_var.h> 62 63#include <netinet/tcp.h> 64#include <netinet/tcp_seq.h> 65#include <netinet/tcp_timer.h> 66#include <netinet/tcp_var.h> 67#include <netinet/tcpip.h> 68 69#define RTPRF_OURS RTF_PROTO3 /* set on routes we manage */ 70 71/* 72 * Do what we need to do when inserting a route. 73 */ 74static struct radix_node * 75in_addroute(void *v_arg, void *n_arg, struct radix_node_head *head, 76 struct radix_node *treenodes) 77{ 78 struct rtentry *rt = (struct rtentry *)treenodes; 79 struct sockaddr_in *sin = (struct sockaddr_in *)rt_key(rt); 80 81 /* 82 * For IP, all unicast non-host routes are automatically cloning. 83 */ 84 if(!(rt->rt_flags & (RTF_HOST | RTF_CLONING))) { 85 if(!IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) { 86 rt->rt_flags |= RTF_PRCLONING; 87 } 88 } 89 90 /* 91 * We also specify a send and receive pipe size for every 92 * route added, to help TCP a bit. TCP doesn't actually 93 * want a true pipe size, which would be prohibitive in memory 94 * costs and is hard to compute anyway; it simply uses these 95 * values to size its buffers. So, we fill them in with the 96 * same values that TCP would have used anyway, and allow the 97 * installing program or the link layer to override these values 98 * as it sees fit. This will hopefully allow TCP more 99 * opportunities to save its ssthresh value. 100 */ 101 if (!rt->rt_rmx.rmx_sendpipe && !(rt->rt_rmx.rmx_locks & RTV_SPIPE)) 102 rt->rt_rmx.rmx_sendpipe = tcp_sendspace; 103 104 if (!rt->rt_rmx.rmx_recvpipe && !(rt->rt_rmx.rmx_locks & RTV_RPIPE)) 105 rt->rt_rmx.rmx_recvpipe = tcp_recvspace; 106 107 /* 108 * Finally, set an MTU, again duplicating logic in TCP. 109 * The in_localaddr() business will go away when we have 110 * proper PMTU discovery. 111 */ 112 if (!rt->rt_rmx.rmx_mtu && !(rt->rt_rmx.rmx_locks & RTV_MTU) 113 && rt->rt_ifp) 114 rt->rt_rmx.rmx_mtu = (in_localaddr(sin->sin_addr) 115 ? rt->rt_ifp->if_mtu 116 : tcp_mssdflt + sizeof(struct tcpiphdr)); 117 118 return rn_addroute(v_arg, n_arg, head, treenodes); 119} 120 121/* 122 * This code is the inverse of in_clsroute: on first reference, if we 123 * were managing the route, stop doing so and set the expiration timer 124 * back off again. 125 */ 126static struct radix_node * 127in_matroute(void *v_arg, struct radix_node_head *head) 128{ 129 struct radix_node *rn = rn_match(v_arg, head); 130 struct rtentry *rt = (struct rtentry *)rn; 131 132 if(rt && rt->rt_refcnt == 0) { /* this is first reference */ 133 if(rt->rt_flags & RTPRF_OURS) { 134 rt->rt_flags &= ~RTPRF_OURS; 135 rt->rt_rmx.rmx_expire = 0; 136 } 137 } 138 return rn; 139} 140 141/* MIB variables: net.inet.ip.{rtexpire,rtmaxcache,rtminexpire}. */ 142int rtq_reallyold = 60*60; /* one hour is ``really old'' */ 143int rtq_toomany = 128; /* 128 cached routes is ``too many'' */ 144int rtq_minreallyold = 10; /* never automatically crank down to less */ 145 146/* 147 * On last reference drop, mark the route as belong to us so that it can be 148 * timed out. 149 */ 150static void 151in_clsroute(struct radix_node *rn, struct radix_node_head *head) 152{ 153 struct rtentry *rt = (struct rtentry *)rn; 154 155 if(!(rt->rt_flags & RTF_UP)) 156 return; /* prophylactic measures */ 157 158 if((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST) 159 return; 160 161 if((rt->rt_flags & (RTF_WASCLONED | RTPRF_OURS)) 162 != RTF_WASCLONED) 163 return; 164 165 /* 166 * As requested by David Greenman: 167 * If rtq_reallyold is 0, just delete the route without 168 * waiting for a timeout cycle to kill it. 169 */ 170 if(rtq_reallyold != 0) { 171 rt->rt_flags |= RTPRF_OURS; 172 rt->rt_rmx.rmx_expire = time.tv_sec + rtq_reallyold; 173 } else { 174 rtrequest(RTM_DELETE, 175 (struct sockaddr *)rt_key(rt), 176 rt->rt_gateway, rt_mask(rt), 177 rt->rt_flags, 0); 178 } 179} 180 181struct rtqk_arg { 182 struct radix_node_head *rnh; 183 int draining; 184 int killed; 185 int found; 186 int updating; 187 time_t nextstop; 188}; 189 190/* 191 * Get rid of old routes. When draining, this deletes everything, even when 192 * the timeout is not expired yet. When updating, this makes sure that 193 * nothing has a timeout longer than the current value of rtq_reallyold. 194 */ 195static int 196in_rtqkill(struct radix_node *rn, void *rock) 197{ 198 struct rtqk_arg *ap = rock; 199 struct radix_node_head *rnh = ap->rnh; 200 struct rtentry *rt = (struct rtentry *)rn; 201 int err; 202 203 if(rt->rt_flags & RTPRF_OURS) { 204 ap->found++; 205 206 if(ap->draining || rt->rt_rmx.rmx_expire <= time.tv_sec) { 207 if(rt->rt_refcnt > 0) 208 panic("rtqkill route really not free"); 209 210 err = rtrequest(RTM_DELETE, 211 (struct sockaddr *)rt_key(rt), 212 rt->rt_gateway, rt_mask(rt), 213 rt->rt_flags, 0); 214 if(err) { 215 log(LOG_WARNING, "in_rtqkill: error %d\n", err); 216 } else { 217 ap->killed++; 218 } 219 } else { 220 if(ap->updating 221 && (rt->rt_rmx.rmx_expire - time.tv_sec 222 > rtq_reallyold)) { 223 rt->rt_rmx.rmx_expire = time.tv_sec 224 + rtq_reallyold; 225 } 226 ap->nextstop = lmin(ap->nextstop, 227 rt->rt_rmx.rmx_expire); 228 } 229 } 230 231 return 0; 232} 233 234#define RTQ_TIMEOUT 60*10 /* run no less than once every ten minutes */ 235int rtq_timeout = RTQ_TIMEOUT; 236 237static void 238in_rtqtimo(void *rock) 239{ 240 struct radix_node_head *rnh = rock; 241 struct rtqk_arg arg; 242 struct timeval atv; 243 static time_t last_adjusted_timeout = 0; 244 int s; 245 246 arg.found = arg.killed = 0; 247 arg.rnh = rnh; 248 arg.nextstop = time.tv_sec + rtq_timeout; 249 arg.draining = arg.updating = 0; 250 s = splnet(); 251 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 252 splx(s); 253 254 /* 255 * Attempt to be somewhat dynamic about this: 256 * If there are ``too many'' routes sitting around taking up space, 257 * then crank down the timeout, and see if we can't make some more 258 * go away. However, we make sure that we will never adjust more 259 * than once in rtq_timeout seconds, to keep from cranking down too 260 * hard. 261 */ 262 if((arg.found - arg.killed > rtq_toomany) 263 && (time.tv_sec - last_adjusted_timeout >= rtq_timeout) 264 && rtq_reallyold > rtq_minreallyold) { 265 rtq_reallyold = 2*rtq_reallyold / 3; 266 if(rtq_reallyold < rtq_minreallyold) { 267 rtq_reallyold = rtq_minreallyold; 268 } 269 270 last_adjusted_timeout = time.tv_sec; 271 log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n", 272 rtq_reallyold); 273 arg.found = arg.killed = 0; 274 arg.updating = 1; 275 s = splnet(); 276 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 277 splx(s); 278 } 279 280 atv.tv_usec = 0; 281 atv.tv_sec = arg.nextstop; 282 timeout(in_rtqtimo, rock, hzto(&atv)); 283} 284 285void 286in_rtqdrain(void) 287{ 288 struct radix_node_head *rnh = rt_tables[AF_INET]; 289 struct rtqk_arg arg; 290 int s; 291 arg.found = arg.killed = 0; 292 arg.rnh = rnh; 293 arg.nextstop = 0; 294 arg.draining = 1; 295 arg.updating = 0; 296 s = splnet(); 297 rnh->rnh_walktree(rnh, in_rtqkill, &arg); 298 splx(s); 299} 300 301/* 302 * Initialize our routing tree. 303 */ 304int 305in_inithead(void **head, int off) 306{ 307 struct radix_node_head *rnh; 308 309 if(!rn_inithead(head, off)) 310 return 0; 311 312 if(head != (void **)&rt_tables[AF_INET]) /* BOGUS! */ 313 return 1; /* only do this for the real routing table */ 314 315 rnh = *head; 316 rnh->rnh_addaddr = in_addroute; 317 rnh->rnh_matchaddr = in_matroute; 318 rnh->rnh_close = in_clsroute; 319 in_rtqtimo(rnh); /* kick off timeout first time */ 320 return 1; 321} 322 323