2 * Copyright (c) 2002 Andre Oppermann, Internet Business Solutions AG
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. The name of the author may not be used to endorse or promote
14 * products derived from this software without specific prior written
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * The tcp_hostcache moves the tcp specific cached metrics from the routing
34 * table into a dedicated structure indexed by the remote IP address. It
35 * keeps information on the measured tcp parameters of past tcp sessions
36 * to have better initial start values for following connections from the
37 * same source. Depending on the network parameters (delay, bandwidth, max
38 * MTU, congestion window) between local and remote site this can lead to
39 * significant speedups for new tcp connections after the first one.
41 * Due to this new tcp_hostcache all tcp specific metrics information in
42 * the routing table has been removed. The INPCB no longer keeps a pointer
43 * to the routing entry and protocol initiated route cloning has been
44 * removed as well. With these changes the routing table has gone back
45 * to being more lightwight and only carries information related to packet
48 * Tcp_hostcache is designed for multiple concurrent access in SMP
49 * environments and high contention. All bucket rows have their own
50 * lock and thus multiple lookups and modifies can be done at the same
51 * time as long as they are in different bucket rows. If a request for
52 * insertion of a new record can't be satisfied it simply returns an
53 * empty structure. Nobody and nothing shall ever point directly to
54 * any entry in tcp_hostcache. All communication is done in an object
55 * oriented way and only funtions of tcp_hostcache will manipulate hostcache
56 * entries. Otherwise we are unable to achieve good behaviour in concurrent
57 * access situations. Since tcp_hostcache is only caching information there
58 * are no fatal consequences if we either can't satisfy any particular request
59 * or have to drop/overwrite an existing entry because of bucket limit
64 * Many thanks to jlemon for basic structure of tcp_syncache which is being
68 #include "opt_inet6.h"
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/kernel.h>
74 #include <sys/mutex.h>
75 #include <sys/malloc.h>
76 #include <sys/socket.h>
77 #include <sys/socketvar.h>
78 #include <sys/sysctl.h>
82 #include <netinet/in.h>
83 #include <netinet/in_systm.h>
84 #include <netinet/ip.h>
85 #include <netinet/in_var.h>
86 #include <netinet/in_pcb.h>
87 #include <netinet/ip_var.h>
89 #include <netinet/ip6.h>
90 #include <netinet6/ip6_var.h>
92 #include <netinet/tcp.h>
93 #include <netinet/tcp_var.h>
95 #include <netinet6/tcp6_var.h>
101 TAILQ_HEAD(hc_qhead, hc_metrics);
104 struct hc_qhead hch_bucket;
111 TAILQ_ENTRY(hc_metrics) rmx_q;
112 struct hc_head *rmx_head; /* head of bucket tail queue */
113 struct in_addr ip4; /* IP address */
114 struct in6_addr ip6; /* IP6 address */
115 /* endpoint specific values for tcp */
116 u_long rmx_mtu; /* MTU for this path */
117 u_long rmx_ssthresh; /* outbound gateway buffer limit */
118 u_long rmx_rtt; /* estimated round trip time */
119 u_long rmx_rttvar; /* estimated rtt variance */
120 u_long rmx_bandwidth; /* estimated bandwidth */
121 u_long rmx_cwnd; /* congestion window */
122 u_long rmx_sendpipe; /* outbound delay-bandwidth product */
123 u_long rmx_recvpipe; /* inbound delay-bandwidth product */
124 /* tcp hostcache internal data */
125 int rmx_expire; /* lifetime for object */
126 u_long rmx_hits; /* number of hits */
127 u_long rmx_updates; /* number of updates */
130 /* Arbitrary values */
131 #define TCP_HOSTCACHE_HASHSIZE 512
132 #define TCP_HOSTCACHE_BUCKETLIMIT 30
133 #define TCP_HOSTCACHE_EXPIRE 60*60 /* one hour */
134 #define TCP_HOSTCACHE_PRUNE 5*60 /* every 5 minutes */
136 struct tcp_hostcache {
137 struct hc_head *hashbase;
147 static struct tcp_hostcache tcp_hostcache;
149 static struct callout tcp_hc_callout;
151 static struct hc_metrics *tcp_hc_lookup(struct in_conninfo *);
152 static struct hc_metrics *tcp_hc_insert(struct in_conninfo *);
153 static int sysctl_tcp_hc_list(SYSCTL_HANDLER_ARGS);
154 static void tcp_hc_purge(void *);
156 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hostcache, CTLFLAG_RW, 0, "TCP Host cache");
158 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, cachelimit, CTLFLAG_RDTUN,
159 &tcp_hostcache.cache_limit, 0, "Overall entry limit for hostcache");
161 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, hashsize, CTLFLAG_RDTUN,
162 &tcp_hostcache.hashsize, 0, "Size of TCP hostcache hashtable");
164 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, bucketlimit, CTLFLAG_RDTUN,
165 &tcp_hostcache.bucket_limit, 0, "Per-bucket hash limit for hostcache");
167 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, count, CTLFLAG_RD,
168 &tcp_hostcache.cache_count, 0, "Current number of entries in hostcache");
170 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, expire, CTLFLAG_RW,
171 &tcp_hostcache.expire, 0, "Expire time of TCP hostcache entries");
173 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, purge, CTLFLAG_RW,
174 &tcp_hostcache.purgeall, 0, "Expire all entires on next purge run");
176 SYSCTL_PROC(_net_inet_tcp_hostcache, OID_AUTO, list,
177 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_SKIP, 0, 0,
178 sysctl_tcp_hc_list, "A", "List of all hostcache entries");
181 static MALLOC_DEFINE(M_HOSTCACHE, "hostcache", "TCP hostcache");
183 #define HOSTCACHE_HASH(ip) \
184 (((ip)->s_addr ^ ((ip)->s_addr >> 7) ^ ((ip)->s_addr >> 17)) & \
185 tcp_hostcache.hashmask)
187 /* XXX: What is the recommended hash to get good entropy for IPv6 addresses? */
188 #define HOSTCACHE_HASH6(ip6) \
189 (((ip6)->s6_addr32[0] ^ \
190 (ip6)->s6_addr32[1] ^ \
191 (ip6)->s6_addr32[2] ^ \
192 (ip6)->s6_addr32[3]) & \
193 tcp_hostcache.hashmask)
195 #define THC_LOCK(lp) mtx_lock(lp)
196 #define THC_UNLOCK(lp) mtx_unlock(lp)
204 * Initialize hostcache structures
206 tcp_hostcache.cache_count = 0;
207 tcp_hostcache.hashsize = TCP_HOSTCACHE_HASHSIZE;
208 tcp_hostcache.bucket_limit = TCP_HOSTCACHE_BUCKETLIMIT;
209 tcp_hostcache.cache_limit =
210 tcp_hostcache.hashsize * tcp_hostcache.bucket_limit;
211 tcp_hostcache.expire = TCP_HOSTCACHE_EXPIRE;
213 TUNABLE_INT_FETCH("net.inet.tcp.hostcache.hashsize",
214 &tcp_hostcache.hashsize);
215 TUNABLE_INT_FETCH("net.inet.tcp.hostcache.cachelimit",
216 &tcp_hostcache.cache_limit);
217 TUNABLE_INT_FETCH("net.inet.tcp.hostcache.bucketlimit",
218 &tcp_hostcache.bucket_limit);
219 if (!powerof2(tcp_hostcache.hashsize)) {
220 printf("WARNING: hostcache hash size is not a power of 2.\n");
221 tcp_hostcache.hashsize = 512; /* safe default */
223 tcp_hostcache.hashmask = tcp_hostcache.hashsize - 1;
226 * Allocate the hash table
228 tcp_hostcache.hashbase = (struct hc_head *)
229 malloc(tcp_hostcache.hashsize * sizeof(struct hc_head),
230 M_HOSTCACHE, M_WAITOK | M_ZERO);
233 * Initialize the hash buckets
235 for (i = 0; i < tcp_hostcache.hashsize; i++) {
236 TAILQ_INIT(&tcp_hostcache.hashbase[i].hch_bucket);
237 tcp_hostcache.hashbase[i].hch_length = 0;
238 mtx_init(&tcp_hostcache.hashbase[i].hch_mtx, "tcp_hc_entry",
243 * Allocate the hostcache entries.
245 tcp_hostcache.zone = uma_zcreate("hostcache", sizeof(struct hc_metrics),
246 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
247 uma_zone_set_max(tcp_hostcache.zone, tcp_hostcache.cache_limit);
250 * Set up periodic cache cleanup.
252 callout_init(&tcp_hc_callout, CALLOUT_MPSAFE);
253 callout_reset(&tcp_hc_callout, TCP_HOSTCACHE_PRUNE * hz, tcp_hc_purge, 0);
257 * Internal function: lookup an entry in the hostcache or return NULL.
259 * If an entry has been returned, the caller becomes responsible for
260 * unlocking the bucket row after he is done reading/modifying the entry.
262 static struct hc_metrics *
263 tcp_hc_lookup(struct in_conninfo *inc)
266 struct hc_head *hc_head;
267 struct hc_metrics *hc_entry;
269 KASSERT(inc != NULL, ("tcp_hc_lookup with NULL in_conninfo pointer"));
272 * Hash the foreign ip address.
275 hash = HOSTCACHE_HASH6(&inc->inc6_faddr);
277 hash = HOSTCACHE_HASH(&inc->inc_faddr);
279 hc_head = &tcp_hostcache.hashbase[hash];
282 * aquire lock for this bucket row
283 * we release the lock if we don't find an entry,
284 * otherwise the caller has to unlock after he is done
286 THC_LOCK(&hc_head->hch_mtx);
289 * circle through entries in bucket row looking for a match
291 TAILQ_FOREACH(hc_entry, &hc_head->hch_bucket, rmx_q) {
292 if (inc->inc_isipv6) {
293 if (memcmp(&inc->inc6_faddr, &hc_entry->ip6,
294 sizeof(inc->inc6_faddr)) == 0)
297 if (memcmp(&inc->inc_faddr, &hc_entry->ip4,
298 sizeof(inc->inc_faddr)) == 0)
304 * We were unsuccessful and didn't find anything
306 THC_UNLOCK(&hc_head->hch_mtx);
311 * Internal function: insert an entry into the hostcache or return NULL
312 * if unable to allocate a new one.
314 * If an entry has been returned, the caller becomes responsible for
315 * unlocking the bucket row after he is done reading/modifying the entry.
317 static struct hc_metrics *
318 tcp_hc_insert(struct in_conninfo *inc)
321 struct hc_head *hc_head;
322 struct hc_metrics *hc_entry;
324 KASSERT(inc != NULL, ("tcp_hc_insert with NULL in_conninfo pointer"));
327 * Hash the foreign ip address
330 hash = HOSTCACHE_HASH6(&inc->inc6_faddr);
332 hash = HOSTCACHE_HASH(&inc->inc_faddr);
334 hc_head = &tcp_hostcache.hashbase[hash];
337 * aquire lock for this bucket row
338 * we release the lock if we don't find an entry,
339 * otherwise the caller has to unlock after he is done
341 THC_LOCK(&hc_head->hch_mtx);
344 * If the bucket limit is reached reuse the least used element
346 if (hc_head->hch_length >= tcp_hostcache.bucket_limit ||
347 tcp_hostcache.cache_count >= tcp_hostcache.cache_limit) {
348 hc_entry = TAILQ_LAST(&hc_head->hch_bucket, hc_qhead);
350 * At first we were dropping the last element, just to
351 * reaquire it in the next two lines again which ain't
352 * very efficient. Instead just reuse the least used element.
353 * Maybe we drop something that is still "in-use" but we can
356 TAILQ_REMOVE(&hc_head->hch_bucket, hc_entry, rmx_q);
357 tcp_hostcache.hashbase[hash].hch_length--;
358 tcp_hostcache.cache_count--;
359 tcpstat.tcps_hc_bucketoverflow++;
361 uma_zfree(tcp_hostcache.zone, hc_entry);
365 * Allocate a new entry, or balk if not possible
367 hc_entry = uma_zalloc(tcp_hostcache.zone, M_NOWAIT);
368 if (hc_entry == NULL) {
369 THC_UNLOCK(&hc_head->hch_mtx);
375 * Initialize basic information of hostcache entry
377 bzero(hc_entry, sizeof(*hc_entry));
379 bcopy(&inc->inc6_faddr, &hc_entry->ip6, sizeof(hc_entry->ip6));
381 hc_entry->ip4 = inc->inc_faddr;
382 hc_entry->rmx_head = hc_head;
383 hc_entry->rmx_expire = tcp_hostcache.expire;
388 TAILQ_INSERT_HEAD(&hc_head->hch_bucket, hc_entry, rmx_q);
389 tcp_hostcache.hashbase[hash].hch_length++;
390 tcp_hostcache.cache_count++;
391 tcpstat.tcps_hc_added++;
397 * External function: lookup an entry in the hostcache and fill out the
398 * supplied tcp metrics structure. Fills in null when no entry was found
399 * or a value is not set.
402 tcp_hc_get(struct in_conninfo *inc, struct hc_metrics_lite *hc_metrics_lite)
404 struct hc_metrics *hc_entry;
407 * Find the right bucket
409 hc_entry = tcp_hc_lookup(inc);
412 * If we don't have an existing object
414 if (hc_entry == NULL) {
415 bzero(hc_metrics_lite, sizeof(*hc_metrics_lite));
418 hc_entry->rmx_hits++;
419 hc_entry->rmx_expire = tcp_hostcache.expire; /* start over again */
421 hc_metrics_lite->rmx_mtu = hc_entry->rmx_mtu;
422 hc_metrics_lite->rmx_ssthresh = hc_entry->rmx_ssthresh;
423 hc_metrics_lite->rmx_rtt = hc_entry->rmx_rtt;
424 hc_metrics_lite->rmx_rttvar = hc_entry->rmx_rttvar;
425 hc_metrics_lite->rmx_bandwidth = hc_entry->rmx_bandwidth;
426 hc_metrics_lite->rmx_cwnd = hc_entry->rmx_cwnd;
427 hc_metrics_lite->rmx_sendpipe = hc_entry->rmx_sendpipe;
428 hc_metrics_lite->rmx_recvpipe = hc_entry->rmx_recvpipe;
433 THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
437 * External function: lookup an entry in the hostcache and return the
438 * discovered path mtu. Returns null if no entry is found or value is not
442 tcp_hc_getmtu(struct in_conninfo *inc)
444 struct hc_metrics *hc_entry;
447 hc_entry = tcp_hc_lookup(inc);
448 if (hc_entry == NULL) {
451 hc_entry->rmx_hits++;
452 hc_entry->rmx_expire = tcp_hostcache.expire; /* start over again */
454 mtu = hc_entry->rmx_mtu;
455 THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
460 * External function: update the mtu value of an entry in the hostcache.
461 * Creates a new entry if none was found.
464 tcp_hc_updatemtu(struct in_conninfo *inc, u_long mtu)
466 struct hc_metrics *hc_entry;
469 * Find the right bucket
471 hc_entry = tcp_hc_lookup(inc);
474 * If we don't have an existing object try to insert a new one
476 if (hc_entry == NULL) {
477 hc_entry = tcp_hc_insert(inc);
478 if (hc_entry == NULL)
481 hc_entry->rmx_updates++;
482 hc_entry->rmx_expire = tcp_hostcache.expire; /* start over again */
484 hc_entry->rmx_mtu = mtu;
487 * put it upfront so we find it faster next time
489 TAILQ_REMOVE(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
490 TAILQ_INSERT_HEAD(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
495 THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
499 * External function: update the tcp metrics of an entry in the hostcache.
500 * Creates a new entry if none was found.
503 tcp_hc_update(struct in_conninfo *inc, struct hc_metrics_lite *hcml)
505 struct hc_metrics *hc_entry;
507 hc_entry = tcp_hc_lookup(inc);
508 if (hc_entry == NULL) {
509 hc_entry = tcp_hc_insert(inc);
510 if (hc_entry == NULL)
513 hc_entry->rmx_updates++;
514 hc_entry->rmx_expire = tcp_hostcache.expire; /* start over again */
516 if (hcml->rmx_rtt != 0) {
517 if (hc_entry->rmx_rtt == 0)
518 hc_entry->rmx_rtt = hcml->rmx_rtt;
521 (hc_entry->rmx_rtt + hcml->rmx_rtt) / 2;
522 tcpstat.tcps_cachedrtt++;
524 if (hcml->rmx_rttvar != 0) {
525 if (hc_entry->rmx_rttvar == 0)
526 hc_entry->rmx_rttvar = hcml->rmx_rttvar;
528 hc_entry->rmx_rttvar =
529 (hc_entry->rmx_rttvar + hcml->rmx_rttvar) / 2;
530 tcpstat.tcps_cachedrttvar++;
532 if (hcml->rmx_ssthresh != 0) {
533 if (hc_entry->rmx_ssthresh == 0)
534 hc_entry->rmx_ssthresh = hcml->rmx_ssthresh;
536 hc_entry->rmx_ssthresh =
537 (hc_entry->rmx_ssthresh + hcml->rmx_ssthresh) / 2;
538 tcpstat.tcps_cachedssthresh++;
540 if (hcml->rmx_bandwidth != 0) {
541 if (hc_entry->rmx_bandwidth == 0)
542 hc_entry->rmx_bandwidth = hcml->rmx_bandwidth;
544 hc_entry->rmx_bandwidth =
545 (hc_entry->rmx_bandwidth + hcml->rmx_bandwidth) / 2;
546 /* tcpstat.tcps_cachedbandwidth++; */
548 if (hcml->rmx_cwnd != 0) {
549 if (hc_entry->rmx_cwnd == 0)
550 hc_entry->rmx_cwnd = hcml->rmx_cwnd;
553 (hc_entry->rmx_cwnd + hcml->rmx_cwnd) / 2;
554 /* tcpstat.tcps_cachedcwnd++; */
556 if (hcml->rmx_sendpipe != 0) {
557 if (hc_entry->rmx_sendpipe == 0)
558 hc_entry->rmx_sendpipe = hcml->rmx_sendpipe;
560 hc_entry->rmx_sendpipe =
561 (hc_entry->rmx_sendpipe + hcml->rmx_sendpipe) /2;
562 /* tcpstat.tcps_cachedsendpipe++; */
564 if (hcml->rmx_recvpipe != 0) {
565 if (hc_entry->rmx_recvpipe == 0)
566 hc_entry->rmx_recvpipe = hcml->rmx_recvpipe;
568 hc_entry->rmx_recvpipe =
569 (hc_entry->rmx_recvpipe + hcml->rmx_recvpipe) /2;
570 /* tcpstat.tcps_cachedrecvpipe++; */
573 TAILQ_REMOVE(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
574 TAILQ_INSERT_HEAD(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
575 THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
579 * Sysctl function: prints the list and values of all hostcache entries in
583 sysctl_tcp_hc_list(SYSCTL_HANDLER_ARGS)
589 struct hc_metrics *hc_entry;
591 char ip6buf[INET6_ADDRSTRLEN];
594 bufsize = linesize * (tcp_hostcache.cache_count + 1);
596 p = buf = (char *)malloc(bufsize, M_TEMP, M_WAITOK|M_ZERO);
598 len = snprintf(p, linesize,
599 "\nIP address MTU SSTRESH RTT RTTVAR BANDWIDTH "
600 " CWND SENDPIPE RECVPIPE HITS UPD EXP\n");
603 #define msec(u) (((u) + 500) / 1000)
604 for (i = 0; i < tcp_hostcache.hashsize; i++) {
605 THC_LOCK(&tcp_hostcache.hashbase[i].hch_mtx);
606 TAILQ_FOREACH(hc_entry, &tcp_hostcache.hashbase[i].hch_bucket,
608 len = snprintf(p, linesize,
609 "%-15s %5lu %8lu %6lums %6lums %9lu %8lu %8lu %8lu "
611 hc_entry->ip4.s_addr ? inet_ntoa(hc_entry->ip4) :
613 ip6_sprintf(ip6buf, &hc_entry->ip6),
618 hc_entry->rmx_ssthresh,
619 msec(hc_entry->rmx_rtt *
620 (RTM_RTTUNIT / (hz * TCP_RTT_SCALE))),
621 msec(hc_entry->rmx_rttvar *
622 (RTM_RTTUNIT / (hz * TCP_RTT_SCALE))),
623 hc_entry->rmx_bandwidth * 8,
625 hc_entry->rmx_sendpipe,
626 hc_entry->rmx_recvpipe,
628 hc_entry->rmx_updates,
629 hc_entry->rmx_expire);
632 THC_UNLOCK(&tcp_hostcache.hashbase[i].hch_mtx);
635 error = SYSCTL_OUT(req, buf, p - buf);
641 * Expire and purge (old|all) entries in the tcp_hostcache. Runs periodically
645 tcp_hc_purge(void *arg)
647 struct hc_metrics *hc_entry, *hc_next;
648 int all = (intptr_t)arg;
651 if (tcp_hostcache.purgeall) {
653 tcp_hostcache.purgeall = 0;
656 for (i = 0; i < tcp_hostcache.hashsize; i++) {
657 THC_LOCK(&tcp_hostcache.hashbase[i].hch_mtx);
658 TAILQ_FOREACH_SAFE(hc_entry, &tcp_hostcache.hashbase[i].hch_bucket,
660 if (all || hc_entry->rmx_expire <= 0) {
661 TAILQ_REMOVE(&tcp_hostcache.hashbase[i].hch_bucket,
663 uma_zfree(tcp_hostcache.zone, hc_entry);
664 tcp_hostcache.hashbase[i].hch_length--;
665 tcp_hostcache.cache_count--;
667 hc_entry->rmx_expire -= TCP_HOSTCACHE_PRUNE;
669 THC_UNLOCK(&tcp_hostcache.hashbase[i].hch_mtx);
671 callout_reset(&tcp_hc_callout, TCP_HOSTCACHE_PRUNE * hz, tcp_hc_purge, 0);