in_rmx.c revision 46381
14074Swollman/* 26399Swollman * Copyright 1994, 1995 Massachusetts Institute of Technology 34074Swollman * 46399Swollman * Permission to use, copy, modify, and distribute this software and 56399Swollman * its documentation for any purpose and without fee is hereby 66399Swollman * granted, provided that both the above copyright notice and this 76399Swollman * permission notice appear in all copies, that both the above 86399Swollman * copyright notice and this permission notice appear in all 96399Swollman * supporting documentation, and that the name of M.I.T. not be used 106399Swollman * in advertising or publicity pertaining to distribution of the 116399Swollman * software without specific, written prior permission. M.I.T. makes 126399Swollman * no representations about the suitability of this software for any 136399Swollman * purpose. It is provided "as is" without express or implied 146399Swollman * warranty. 158876Srgrimes * 166399Swollman * THIS SOFTWARE IS PROVIDED BY M.I.T. ``AS IS''. M.I.T. DISCLAIMS 176399Swollman * ALL EXPRESS OR IMPLIED WARRANTIES WITH REGARD TO THIS SOFTWARE, 186399Swollman * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF 196399Swollman * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT 206399Swollman * SHALL M.I.T. BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 216399Swollman * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 226399Swollman * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF 236399Swollman * USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 246399Swollman * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 256399Swollman * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 266399Swollman * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 276399Swollman * SUCH DAMAGE. 284074Swollman * 2946381Sbillf * $Id: in_rmx.c,v 1.35 1998/08/05 16:59:20 bde Exp $ 304074Swollman */ 314074Swollman 324074Swollman/* 334074Swollman * This code does two things necessary for the enhanced TCP metrics to 344074Swollman * function in a useful manner: 354074Swollman * 1) It marks all non-host routes as `cloning', thus ensuring that 364074Swollman * every actual reference to such a route actually gets turned 374074Swollman * into a reference to a host route to the specific destination 384074Swollman * requested. 394074Swollman * 2) When such routes lose all their references, it arranges for them 404074Swollman * to be deleted in some random collection of circumstances, so that 414074Swollman * a large quantity of stale routing data is not kept in kernel memory 424074Swollman * indefinitely. See in_rtqtimo() below for the exact mechanism. 434074Swollman */ 444074Swollman 454074Swollman#include <sys/param.h> 464074Swollman#include <sys/systm.h> 474074Swollman#include <sys/kernel.h> 4812172Sphk#include <sys/sysctl.h> 494074Swollman#include <sys/socket.h> 504074Swollman#include <sys/mbuf.h> 514893Swollman#include <sys/syslog.h> 524074Swollman 534074Swollman#include <net/if.h> 544074Swollman#include <net/route.h> 554074Swollman#include <netinet/in.h> 564074Swollman#include <netinet/in_var.h> 574074Swollman 5812579Sbdeextern int in_inithead __P((void **head, int off)); 5912579Sbde 605101Swollman#define RTPRF_OURS RTF_PROTO3 /* set on routes we manage */ 614074Swollman 624074Swollman/* 634074Swollman * Do what we need to do when inserting a route. 644074Swollman */ 654074Swollmanstatic struct radix_node * 664074Swollmanin_addroute(void *v_arg, void *n_arg, struct radix_node_head *head, 674074Swollman struct radix_node *treenodes) 684074Swollman{ 694074Swollman struct rtentry *rt = (struct rtentry *)treenodes; 709470Swollman struct sockaddr_in *sin = (struct sockaddr_in *)rt_key(rt); 7113581Sfenner struct radix_node *ret; 724074Swollman 734074Swollman /* 745792Swollman * For IP, all unicast non-host routes are automatically cloning. 754074Swollman */ 7615652Swollman if(IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) 7715652Swollman rt->rt_flags |= RTF_MULTICAST; 7815652Swollman 7915652Swollman if(!(rt->rt_flags & (RTF_HOST | RTF_CLONING | RTF_MULTICAST))) { 8015652Swollman rt->rt_flags |= RTF_PRCLONING; 8115652Swollman } 8215652Swollman 8315652Swollman /* 8415652Swollman * A little bit of help for both IP output and input: 8515652Swollman * For host routes, we make sure that RTF_BROADCAST 8615652Swollman * is set for anything that looks like a broadcast address. 8715652Swollman * This way, we can avoid an expensive call to in_broadcast() 8815652Swollman * in ip_output() most of the time (because the route passed 8915652Swollman * to ip_output() is almost always a host route). 9015652Swollman * 9115652Swollman * We also do the same for local addresses, with the thought 9215652Swollman * that this might one day be used to speed up ip_input(). 9315652Swollman * 9415652Swollman * We also mark routes to multicast addresses as such, because 9515652Swollman * it's easy to do and might be useful (but this is much more 9615652Swollman * dubious since it's so easy to inspect the address). (This 9715652Swollman * is done above.) 9815652Swollman */ 9915652Swollman if (rt->rt_flags & RTF_HOST) { 10015652Swollman if (in_broadcast(sin->sin_addr, rt->rt_ifp)) { 10115652Swollman rt->rt_flags |= RTF_BROADCAST; 10215652Swollman } else { 10315652Swollman#define satosin(sa) ((struct sockaddr_in *)sa) 10415652Swollman if (satosin(rt->rt_ifa->ifa_addr)->sin_addr.s_addr 10515652Swollman == sin->sin_addr.s_addr) 10615652Swollman rt->rt_flags |= RTF_LOCAL; 10715652Swollman#undef satosin 1085792Swollman } 1095792Swollman } 1104074Swollman 1119470Swollman if (!rt->rt_rmx.rmx_mtu && !(rt->rt_rmx.rmx_locks & RTV_MTU) 1129470Swollman && rt->rt_ifp) 11310881Swollman rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu; 1149470Swollman 11513581Sfenner ret = rn_addroute(v_arg, n_arg, head, treenodes); 11613581Sfenner if (ret == NULL && rt->rt_flags & RTF_HOST) { 11713581Sfenner struct rtentry *rt2; 11813581Sfenner /* 11913581Sfenner * We are trying to add a host route, but can't. 12013581Sfenner * Find out if it is because of an 12113581Sfenner * ARP entry and delete it if so. 12213581Sfenner */ 12313581Sfenner rt2 = rtalloc1((struct sockaddr *)sin, 0, 12413581Sfenner RTF_CLONING | RTF_PRCLONING); 12513581Sfenner if (rt2) { 12613581Sfenner if (rt2->rt_flags & RTF_LLINFO && 12713581Sfenner rt2->rt_flags & RTF_HOST && 12813581Sfenner rt2->rt_gateway && 12913581Sfenner rt2->rt_gateway->sa_family == AF_LINK) { 13013581Sfenner rtrequest(RTM_DELETE, 13113581Sfenner (struct sockaddr *)rt_key(rt2), 13213581Sfenner rt2->rt_gateway, 13313581Sfenner rt_mask(rt2), rt2->rt_flags, 0); 13413581Sfenner ret = rn_addroute(v_arg, n_arg, head, 13513581Sfenner treenodes); 13613581Sfenner } 13713581Sfenner RTFREE(rt2); 13813581Sfenner } 13913581Sfenner } 14013581Sfenner return ret; 1414074Swollman} 1424074Swollman 1434074Swollman/* 1444105Swollman * This code is the inverse of in_clsroute: on first reference, if we 1454105Swollman * were managing the route, stop doing so and set the expiration timer 1464105Swollman * back off again. 1474074Swollman */ 1484074Swollmanstatic struct radix_node * 1494074Swollmanin_matroute(void *v_arg, struct radix_node_head *head) 1504074Swollman{ 1514074Swollman struct radix_node *rn = rn_match(v_arg, head); 1524074Swollman struct rtentry *rt = (struct rtentry *)rn; 1534074Swollman 1544074Swollman if(rt && rt->rt_refcnt == 0) { /* this is first reference */ 1555101Swollman if(rt->rt_flags & RTPRF_OURS) { 1565101Swollman rt->rt_flags &= ~RTPRF_OURS; 1574105Swollman rt->rt_rmx.rmx_expire = 0; 1584074Swollman } 1594074Swollman } 1604074Swollman return rn; 1614074Swollman} 1624074Swollman 16312296Sphkstatic int rtq_reallyold = 60*60; 16412172Sphk /* one hour is ``really old'' */ 16546381SbillfSYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire, CTLFLAG_RW, 16646381Sbillf &rtq_reallyold , 0, 16746381Sbillf "Default expiration time on dynamically learned routes"); 16812172Sphk 16912296Sphkstatic int rtq_minreallyold = 10; 17012172Sphk /* never automatically crank down to less */ 17146381SbillfSYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire, CTLFLAG_RW, 17246381Sbillf &rtq_minreallyold , 0, 17346381Sbillf "Minimum time to attempt to hold onto dynamically learned routes"); 17412172Sphk 17512296Sphkstatic int rtq_toomany = 128; 17612172Sphk /* 128 cached routes is ``too many'' */ 17746381SbillfSYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache, CTLFLAG_RW, 17846381Sbillf &rtq_toomany , 0, "Upper limit on dynamically learned routes"); 1794105Swollman 1804074Swollman/* 1815792Swollman * On last reference drop, mark the route as belong to us so that it can be 1824074Swollman * timed out. 1834074Swollman */ 1844074Swollmanstatic void 1854074Swollmanin_clsroute(struct radix_node *rn, struct radix_node_head *head) 1864074Swollman{ 1874074Swollman struct rtentry *rt = (struct rtentry *)rn; 1888876Srgrimes 1895792Swollman if(!(rt->rt_flags & RTF_UP)) 1905792Swollman return; /* prophylactic measures */ 1915792Swollman 1924105Swollman if((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST) 1934074Swollman return; 1944074Swollman 1958876Srgrimes if((rt->rt_flags & (RTF_WASCLONED | RTPRF_OURS)) 1965101Swollman != RTF_WASCLONED) 1974105Swollman return; 1984074Swollman 1996399Swollman /* 2006399Swollman * As requested by David Greenman: 2016399Swollman * If rtq_reallyold is 0, just delete the route without 2026399Swollman * waiting for a timeout cycle to kill it. 2036399Swollman */ 2046399Swollman if(rtq_reallyold != 0) { 2056399Swollman rt->rt_flags |= RTPRF_OURS; 20634961Sphk rt->rt_rmx.rmx_expire = time_second + rtq_reallyold; 2076399Swollman } else { 2086399Swollman rtrequest(RTM_DELETE, 2096399Swollman (struct sockaddr *)rt_key(rt), 2106399Swollman rt->rt_gateway, rt_mask(rt), 2116399Swollman rt->rt_flags, 0); 2126399Swollman } 2134074Swollman} 2144074Swollman 2154893Swollmanstruct rtqk_arg { 2164893Swollman struct radix_node_head *rnh; 2175101Swollman int draining; 2184893Swollman int killed; 2194893Swollman int found; 2206400Swollman int updating; 2214893Swollman time_t nextstop; 2224893Swollman}; 2234893Swollman 2244074Swollman/* 2255101Swollman * Get rid of old routes. When draining, this deletes everything, even when 2266400Swollman * the timeout is not expired yet. When updating, this makes sure that 2276400Swollman * nothing has a timeout longer than the current value of rtq_reallyold. 2284074Swollman */ 2294893Swollmanstatic int 2304893Swollmanin_rtqkill(struct radix_node *rn, void *rock) 2314074Swollman{ 2324893Swollman struct rtqk_arg *ap = rock; 2334893Swollman struct rtentry *rt = (struct rtentry *)rn; 2344893Swollman int err; 2354893Swollman 2365101Swollman if(rt->rt_flags & RTPRF_OURS) { 2374893Swollman ap->found++; 2384893Swollman 23934961Sphk if(ap->draining || rt->rt_rmx.rmx_expire <= time_second) { 2404893Swollman if(rt->rt_refcnt > 0) 2417170Sdg panic("rtqkill route really not free"); 2424893Swollman 2434893Swollman err = rtrequest(RTM_DELETE, 2444893Swollman (struct sockaddr *)rt_key(rt), 2454893Swollman rt->rt_gateway, rt_mask(rt), 2464893Swollman rt->rt_flags, 0); 2474893Swollman if(err) { 2486568Sdg log(LOG_WARNING, "in_rtqkill: error %d\n", err); 2494893Swollman } else { 2504893Swollman ap->killed++; 2514893Swollman } 2524893Swollman } else { 2538876Srgrimes if(ap->updating 25434961Sphk && (rt->rt_rmx.rmx_expire - time_second 2556400Swollman > rtq_reallyold)) { 25634961Sphk rt->rt_rmx.rmx_expire = time_second 2576400Swollman + rtq_reallyold; 2586400Swollman } 2594893Swollman ap->nextstop = lmin(ap->nextstop, 2604893Swollman rt->rt_rmx.rmx_expire); 2614893Swollman } 2624893Swollman } 2634893Swollman 2644893Swollman return 0; 2654074Swollman} 2664074Swollman 2676399Swollman#define RTQ_TIMEOUT 60*10 /* run no less than once every ten minutes */ 26812296Sphkstatic int rtq_timeout = RTQ_TIMEOUT; 2695792Swollman 2704074Swollmanstatic void 2714074Swollmanin_rtqtimo(void *rock) 2724074Swollman{ 2734893Swollman struct radix_node_head *rnh = rock; 2744893Swollman struct rtqk_arg arg; 2754893Swollman struct timeval atv; 2766399Swollman static time_t last_adjusted_timeout = 0; 2775180Swollman int s; 2784074Swollman 2794893Swollman arg.found = arg.killed = 0; 2804893Swollman arg.rnh = rnh; 28134961Sphk arg.nextstop = time_second + rtq_timeout; 2826400Swollman arg.draining = arg.updating = 0; 2835180Swollman s = splnet(); 2844893Swollman rnh->rnh_walktree(rnh, in_rtqkill, &arg); 2855180Swollman splx(s); 2866399Swollman 2876399Swollman /* 2886399Swollman * Attempt to be somewhat dynamic about this: 2896399Swollman * If there are ``too many'' routes sitting around taking up space, 2906399Swollman * then crank down the timeout, and see if we can't make some more 2916399Swollman * go away. However, we make sure that we will never adjust more 2926399Swollman * than once in rtq_timeout seconds, to keep from cranking down too 2936399Swollman * hard. 2946399Swollman */ 2956399Swollman if((arg.found - arg.killed > rtq_toomany) 29634961Sphk && (time_second - last_adjusted_timeout >= rtq_timeout) 2976399Swollman && rtq_reallyold > rtq_minreallyold) { 2986399Swollman rtq_reallyold = 2*rtq_reallyold / 3; 2996399Swollman if(rtq_reallyold < rtq_minreallyold) { 3006399Swollman rtq_reallyold = rtq_minreallyold; 3016399Swollman } 3026399Swollman 30334961Sphk last_adjusted_timeout = time_second; 30416542Snate#ifdef DIAGNOSTIC 3056568Sdg log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n", 3066399Swollman rtq_reallyold); 30716542Snate#endif 3086399Swollman arg.found = arg.killed = 0; 3096400Swollman arg.updating = 1; 3106399Swollman s = splnet(); 3116399Swollman rnh->rnh_walktree(rnh, in_rtqkill, &arg); 3126399Swollman splx(s); 3136399Swollman } 3146399Swollman 3154893Swollman atv.tv_usec = 0; 31638128Sbde atv.tv_sec = arg.nextstop - time_second; 31738128Sbde timeout(in_rtqtimo, rock, tvtohz(&atv)); 3184074Swollman} 3194074Swollman 32012933Swollmanvoid 3214893Swollmanin_rtqdrain(void) 3224893Swollman{ 3235101Swollman struct radix_node_head *rnh = rt_tables[AF_INET]; 3245101Swollman struct rtqk_arg arg; 3255180Swollman int s; 3265101Swollman arg.found = arg.killed = 0; 3275101Swollman arg.rnh = rnh; 3285101Swollman arg.nextstop = 0; 3295101Swollman arg.draining = 1; 3306400Swollman arg.updating = 0; 3315180Swollman s = splnet(); 3325101Swollman rnh->rnh_walktree(rnh, in_rtqkill, &arg); 3335101Swollman splx(s); 3344893Swollman} 3354893Swollman 3364074Swollman/* 3374074Swollman * Initialize our routing tree. 3384074Swollman */ 3394074Swollmanint 3404074Swollmanin_inithead(void **head, int off) 3414074Swollman{ 3424074Swollman struct radix_node_head *rnh; 3434074Swollman 3444074Swollman if(!rn_inithead(head, off)) 3454074Swollman return 0; 3464074Swollman 3474896Swollman if(head != (void **)&rt_tables[AF_INET]) /* BOGUS! */ 3484896Swollman return 1; /* only do this for the real routing table */ 3494896Swollman 3504074Swollman rnh = *head; 3514074Swollman rnh->rnh_addaddr = in_addroute; 3524074Swollman rnh->rnh_matchaddr = in_matroute; 3534074Swollman rnh->rnh_close = in_clsroute; 3544105Swollman in_rtqtimo(rnh); /* kick off timeout first time */ 3554074Swollman return 1; 3564074Swollman} 3574074Swollman 35822672Swollman 35922672Swollman/* 36022672Swollman * This zaps old routes when the interface goes down. 36122672Swollman * Currently it doesn't delete static routes; there are 36222672Swollman * arguments one could make for both behaviors. For the moment, 36322672Swollman * we will adopt the Principle of Least Surprise and leave them 36422672Swollman * alone (with the knowledge that this will not be enough for some 36522672Swollman * people). The ones we really want to get rid of are things like ARP 36622672Swollman * entries, since the user might down the interface, walk over to a completely 36722672Swollman * different network, and plug back in. 36822672Swollman */ 36922672Swollmanstruct in_ifadown_arg { 37022672Swollman struct radix_node_head *rnh; 37122672Swollman struct ifaddr *ifa; 37222672Swollman}; 37322672Swollman 37422672Swollmanstatic int 37522672Swollmanin_ifadownkill(struct radix_node *rn, void *xap) 37622672Swollman{ 37722672Swollman struct in_ifadown_arg *ap = xap; 37822672Swollman struct rtentry *rt = (struct rtentry *)rn; 37922672Swollman int err; 38022672Swollman 38122672Swollman if (rt->rt_ifa == ap->ifa && !(rt->rt_flags & RTF_STATIC)) { 38234914Speter /* 38334914Speter * We need to disable the automatic prune that happens 38434914Speter * in this case in rtrequest() because it will blow 38534914Speter * away the pointers that rn_walktree() needs in order 38634914Speter * continue our descent. We will end up deleting all 38734914Speter * the routes that rtrequest() would have in any case, 38834914Speter * so that behavior is not needed there. 38934914Speter */ 39034914Speter rt->rt_flags &= ~RTF_PRCLONING; 39122672Swollman err = rtrequest(RTM_DELETE, (struct sockaddr *)rt_key(rt), 39222672Swollman rt->rt_gateway, rt_mask(rt), rt->rt_flags, 0); 39322672Swollman if (err) { 39422672Swollman log(LOG_WARNING, "in_ifadownkill: error %d\n", err); 39522672Swollman } 39622672Swollman } 39722672Swollman return 0; 39822672Swollman} 39922672Swollman 40022672Swollmanint 40122672Swollmanin_ifadown(struct ifaddr *ifa) 40222672Swollman{ 40322672Swollman struct in_ifadown_arg arg; 40422672Swollman struct radix_node_head *rnh; 40522672Swollman 40622672Swollman if (ifa->ifa_addr->sa_family != AF_INET) 40722672Swollman return 1; 40822672Swollman 40922672Swollman arg.rnh = rnh = rt_tables[AF_INET]; 41022672Swollman arg.ifa = ifa; 41122672Swollman rnh->rnh_walktree(rnh, in_ifadownkill, &arg); 41222672Swollman ifa->ifa_flags &= ~IFA_ROUTE; 41322672Swollman return 0; 41422672Swollman} 415