freebsd-dev/sys/netinet/in_rmx.c
andre 9ed61330fb Introduce tcp_hostcache and remove the tcp specific metrics from
the routing table.  Move all usage and references in the tcp stack
from the routing table metrics to the tcp hostcache.

It caches measured parameters of past tcp sessions to provide better
initial start values for following connections from or to the same
source or destination.  Depending on the network parameters to/from
the remote host this can lead to significant speedups for new tcp
connections after the first one because they inherit and shortcut
the learning curve.

tcp_hostcache is designed for multiple concurrent access in SMP
environments with high contention and is hash indexed by remote
ip address.

It removes significant locking requirements from the tcp stack with
regard to the routing table.

Reviewed by:	sam (mentor), bms
Reviewed by:	-net, -current, core@kame.net (IPv6 parts)
Approved by:	re (scottl)
2003-11-20 20:07:39 +00:00

397 lines
11 KiB
C

/*
* Copyright 1994, 1995 Massachusetts Institute of Technology
*
* Permission to use, copy, modify, and distribute this software and
* its documentation for any purpose and without fee is hereby
* granted, provided that both the above copyright notice and this
* permission notice appear in all copies, that both the above
* copyright notice and this permission notice appear in all
* supporting documentation, and that the name of M.I.T. not be used
* in advertising or publicity pertaining to distribution of the
* software without specific, written prior permission. M.I.T. makes
* no representations about the suitability of this software for any
* purpose. It is provided "as is" without express or implied
* warranty.
*
* THIS SOFTWARE IS PROVIDED BY M.I.T. ``AS IS''. M.I.T. DISCLAIMS
* ALL EXPRESS OR IMPLIED WARRANTIES WITH REGARD TO THIS SOFTWARE,
* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT
* SHALL M.I.T. BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF
* USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
* $FreeBSD$
*/
/*
* This code does two things necessary for the enhanced TCP metrics to
* function in a useful manner:
* 1) It marks all non-host routes as `cloning', thus ensuring that
* every actual reference to such a route actually gets turned
* into a reference to a host route to the specific destination
* requested.
* 2) When such routes lose all their references, it arranges for them
* to be deleted in some random collection of circumstances, so that
* a large quantity of stale routing data is not kept in kernel memory
* indefinitely. See in_rtqtimo() below for the exact mechanism.
*/
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/sysctl.h>
#include <sys/socket.h>
#include <sys/mbuf.h>
#include <sys/syslog.h>
#include <sys/callout.h>
#include <net/if.h>
#include <net/route.h>
#include <netinet/in.h>
#include <netinet/in_var.h>
#include <netinet/ip_var.h>
extern int in_inithead(void **head, int off);
#define RTPRF_OURS RTF_PROTO3 /* set on routes we manage */
/*
* Do what we need to do when inserting a route.
*/
static struct radix_node *
in_addroute(void *v_arg, void *n_arg, struct radix_node_head *head,
struct radix_node *treenodes)
{
struct rtentry *rt = (struct rtentry *)treenodes;
struct sockaddr_in *sin = (struct sockaddr_in *)rt_key(rt);
struct radix_node *ret;
/*
* A little bit of help for both IP output and input:
* For host routes, we make sure that RTF_BROADCAST
* is set for anything that looks like a broadcast address.
* This way, we can avoid an expensive call to in_broadcast()
* in ip_output() most of the time (because the route passed
* to ip_output() is almost always a host route).
*
* We also do the same for local addresses, with the thought
* that this might one day be used to speed up ip_input().
*
* We also mark routes to multicast addresses as such, because
* it's easy to do and might be useful (but this is much more
* dubious since it's so easy to inspect the address).
*/
if (rt->rt_flags & RTF_HOST) {
if (in_broadcast(sin->sin_addr, rt->rt_ifp)) {
rt->rt_flags |= RTF_BROADCAST;
} else if (satosin(rt->rt_ifa->ifa_addr)->sin_addr.s_addr ==
sin->sin_addr.s_addr) {
rt->rt_flags |= RTF_LOCAL;
}
}
if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr)))
rt->rt_flags |= RTF_MULTICAST;
if (!rt->rt_rmx.rmx_mtu && rt->rt_ifp)
rt->rt_rmx.rmx_mtu = rt->rt_ifp->if_mtu;
ret = rn_addroute(v_arg, n_arg, head, treenodes);
if (ret == NULL && rt->rt_flags & RTF_HOST) {
struct rtentry *rt2;
/*
* We are trying to add a host route, but can't.
* Find out if it is because of an
* ARP entry and delete it if so.
*/
rt2 = rtalloc1((struct sockaddr *)sin, 0, RTF_CLONING);
if (rt2) {
if (rt2->rt_flags & RTF_LLINFO &&
rt2->rt_flags & RTF_HOST &&
rt2->rt_gateway &&
rt2->rt_gateway->sa_family == AF_LINK) {
rtexpunge(rt2);
RTFREE_LOCKED(rt2);
ret = rn_addroute(v_arg, n_arg, head,
treenodes);
} else
RTFREE_LOCKED(rt2);
}
}
return ret;
}
/*
* This code is the inverse of in_clsroute: on first reference, if we
* were managing the route, stop doing so and set the expiration timer
* back off again.
*/
static struct radix_node *
in_matroute(void *v_arg, struct radix_node_head *head)
{
struct radix_node *rn = rn_match(v_arg, head);
struct rtentry *rt = (struct rtentry *)rn;
/*XXX locking? */
if (rt && rt->rt_refcnt == 0) { /* this is first reference */
if (rt->rt_flags & RTPRF_OURS) {
rt->rt_flags &= ~RTPRF_OURS;
rt->rt_rmx.rmx_expire = 0;
}
}
return rn;
}
static int rtq_reallyold = 60*60; /* one hour is "really old" */
SYSCTL_INT(_net_inet_ip, IPCTL_RTEXPIRE, rtexpire, CTLFLAG_RW,
&rtq_reallyold, 0, "Default expiration time on dynamically learned routes");
static int rtq_minreallyold = 10; /* never automatically crank down to less */
SYSCTL_INT(_net_inet_ip, IPCTL_RTMINEXPIRE, rtminexpire, CTLFLAG_RW,
&rtq_minreallyold, 0,
"Minimum time to attempt to hold onto dynamically learned routes");
static int rtq_toomany = 128; /* 128 cached routes is "too many" */
SYSCTL_INT(_net_inet_ip, IPCTL_RTMAXCACHE, rtmaxcache, CTLFLAG_RW,
&rtq_toomany, 0, "Upper limit on dynamically learned routes");
/*
* On last reference drop, mark the route as belong to us so that it can be
* timed out.
*/
static void
in_clsroute(struct radix_node *rn, struct radix_node_head *head)
{
struct rtentry *rt = (struct rtentry *)rn;
RT_LOCK_ASSERT(rt);
if (!(rt->rt_flags & RTF_UP))
return; /* prophylactic measures */
if ((rt->rt_flags & (RTF_LLINFO | RTF_HOST)) != RTF_HOST)
return;
if ((rt->rt_flags & (RTF_WASCLONED | RTPRF_OURS)) != RTF_WASCLONED)
return;
/*
* If rtq_reallyold is 0, just delete the route without
* waiting for a timeout cycle to kill it.
*/
if (rtq_reallyold != 0) {
rt->rt_flags |= RTPRF_OURS;
rt->rt_rmx.rmx_expire = time_second + rtq_reallyold;
} else {
rtexpunge(rt);
}
}
struct rtqk_arg {
struct radix_node_head *rnh;
int draining;
int killed;
int found;
int updating;
time_t nextstop;
};
/*
* Get rid of old routes. When draining, this deletes everything, even when
* the timeout is not expired yet. When updating, this makes sure that
* nothing has a timeout longer than the current value of rtq_reallyold.
*/
static int
in_rtqkill(struct radix_node *rn, void *rock)
{
struct rtqk_arg *ap = rock;
struct rtentry *rt = (struct rtentry *)rn;
int err;
if (rt->rt_flags & RTPRF_OURS) {
ap->found++;
if (ap->draining || rt->rt_rmx.rmx_expire <= time_second) {
if (rt->rt_refcnt > 0)
panic("rtqkill route really not free");
err = rtrequest(RTM_DELETE,
(struct sockaddr *)rt_key(rt),
rt->rt_gateway, rt_mask(rt),
rt->rt_flags, 0);
if (err) {
log(LOG_WARNING, "in_rtqkill: error %d\n", err);
} else {
ap->killed++;
}
} else {
if (ap->updating &&
(rt->rt_rmx.rmx_expire - time_second >
rtq_reallyold)) {
rt->rt_rmx.rmx_expire =
time_second + rtq_reallyold;
}
ap->nextstop = lmin(ap->nextstop,
rt->rt_rmx.rmx_expire);
}
}
return 0;
}
#define RTQ_TIMEOUT 60*10 /* run no less than once every ten minutes */
static int rtq_timeout = RTQ_TIMEOUT;
static struct callout rtq_timer;
static void
in_rtqtimo(void *rock)
{
struct radix_node_head *rnh = rock;
struct rtqk_arg arg;
struct timeval atv;
static time_t last_adjusted_timeout = 0;
arg.found = arg.killed = 0;
arg.rnh = rnh;
arg.nextstop = time_second + rtq_timeout;
arg.draining = arg.updating = 0;
RADIX_NODE_HEAD_LOCK(rnh);
rnh->rnh_walktree(rnh, in_rtqkill, &arg);
RADIX_NODE_HEAD_UNLOCK(rnh);
/*
* Attempt to be somewhat dynamic about this:
* If there are ``too many'' routes sitting around taking up space,
* then crank down the timeout, and see if we can't make some more
* go away. However, we make sure that we will never adjust more
* than once in rtq_timeout seconds, to keep from cranking down too
* hard.
*/
if ((arg.found - arg.killed > rtq_toomany) &&
(time_second - last_adjusted_timeout >= rtq_timeout) &&
rtq_reallyold > rtq_minreallyold) {
rtq_reallyold = 2 * rtq_reallyold / 3;
if (rtq_reallyold < rtq_minreallyold) {
rtq_reallyold = rtq_minreallyold;
}
last_adjusted_timeout = time_second;
#ifdef DIAGNOSTIC
log(LOG_DEBUG, "in_rtqtimo: adjusted rtq_reallyold to %d\n",
rtq_reallyold);
#endif
arg.found = arg.killed = 0;
arg.updating = 1;
RADIX_NODE_HEAD_LOCK(rnh);
rnh->rnh_walktree(rnh, in_rtqkill, &arg);
RADIX_NODE_HEAD_UNLOCK(rnh);
}
atv.tv_usec = 0;
atv.tv_sec = arg.nextstop - time_second;
callout_reset(&rtq_timer, tvtohz(&atv), in_rtqtimo, rock);
}
void
in_rtqdrain(void)
{
struct radix_node_head *rnh = rt_tables[AF_INET];
struct rtqk_arg arg;
arg.found = arg.killed = 0;
arg.rnh = rnh;
arg.nextstop = 0;
arg.draining = 1;
arg.updating = 0;
RADIX_NODE_HEAD_LOCK(rnh);
rnh->rnh_walktree(rnh, in_rtqkill, &arg);
RADIX_NODE_HEAD_UNLOCK(rnh);
}
/*
* Initialize our routing tree.
*/
int
in_inithead(void **head, int off)
{
struct radix_node_head *rnh;
if (!rn_inithead(head, off))
return 0;
if (head != (void **)&rt_tables[AF_INET]) /* BOGUS! */
return 1; /* only do this for the real routing table */
rnh = *head;
rnh->rnh_addaddr = in_addroute;
rnh->rnh_matchaddr = in_matroute;
rnh->rnh_close = in_clsroute;
callout_init(&rtq_timer, CALLOUT_MPSAFE);
in_rtqtimo(rnh); /* kick off timeout first time */
return 1;
}
/*
* This zaps old routes when the interface goes down or interface
* address is deleted. In the latter case, it deletes static routes
* that point to this address. If we don't do this, we may end up
* using the old address in the future. The ones we always want to
* get rid of are things like ARP entries, since the user might down
* the interface, walk over to a completely different network, and
* plug back in.
*/
struct in_ifadown_arg {
struct radix_node_head *rnh;
struct ifaddr *ifa;
int del;
};
static int
in_ifadownkill(struct radix_node *rn, void *xap)
{
struct in_ifadown_arg *ap = xap;
struct rtentry *rt = (struct rtentry *)rn;
RT_LOCK(rt);
if (rt->rt_ifa == ap->ifa &&
(ap->del || !(rt->rt_flags & RTF_STATIC))) {
/*
* We need to disable the automatic prune that happens
* in this case in rtrequest() because it will blow
* away the pointers that rn_walktree() needs in order
* continue our descent. We will end up deleting all
* the routes that rtrequest() would have in any case,
* so that behavior is not needed there.
*/
rt->rt_flags &= ~RTF_CLONING;
rtexpunge(rt);
}
RT_UNLOCK(rt);
return 0;
}
int
in_ifadown(struct ifaddr *ifa, int delete)
{
struct in_ifadown_arg arg;
struct radix_node_head *rnh;
if (ifa->ifa_addr->sa_family != AF_INET)
return 1;
arg.rnh = rnh = rt_tables[AF_INET];
arg.ifa = ifa;
arg.del = delete;
RADIX_NODE_HEAD_LOCK(rnh);
rnh->rnh_walktree(rnh, in_ifadownkill, &arg);
RADIX_NODE_HEAD_UNLOCK(rnh);
ifa->ifa_flags &= ~IFA_ROUTE; /* XXXlocking? */
return 0;
}