]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/netinet/tcp_hostcache.c
zfs: merge openzfs/zfs@21bd76613 (zfs-2.1-release) into stable/13
[FreeBSD/FreeBSD.git] / sys / netinet / tcp_hostcache.c
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 2002 Andre Oppermann, Internet Business Solutions AG
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. The name of the author may not be used to endorse or promote
16  *    products derived from this software without specific prior written
17  *    permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31
32 /*
33  * The tcp_hostcache moves the tcp-specific cached metrics from the routing
34  * table to a dedicated structure indexed by the remote IP address.  It keeps
35  * information on the measured TCP parameters of past TCP sessions to allow
36  * better initial start values to be used with later connections to/from the
37  * same source.  Depending on the network parameters (delay, max MTU,
38  * congestion window) between local and remote sites, this can lead to
39  * significant speed-ups for new TCP connections after the first one.
40  *
41  * Due to the tcp_hostcache, all TCP-specific metrics information in the
42  * routing table have been removed.  The inpcb no longer keeps a pointer to
43  * the routing entry, and protocol-initiated route cloning has been removed
44  * as well.  With these changes, the routing table has gone back to being
45  * more lightwight and only carries information related to packet forwarding.
46  *
47  * tcp_hostcache is designed for multiple concurrent access in SMP
48  * environments and high contention.  All bucket rows have their own lock and
49  * thus multiple lookups and modifies can be done at the same time as long as
50  * they are in different bucket rows.  If a request for insertion of a new
51  * record can't be satisfied, it simply returns an empty structure.  Nobody
52  * and nothing outside of tcp_hostcache.c will ever point directly to any
53  * entry in the tcp_hostcache.  All communication is done in an
54  * object-oriented way and only functions of tcp_hostcache will manipulate
55  * hostcache entries.  Otherwise, we are unable to achieve good behaviour in
56  * concurrent access situations.  Since tcp_hostcache is only caching
57  * information, there are no fatal consequences if we either can't satisfy
58  * any particular request or have to drop/overwrite an existing entry because
59  * of bucket limit memory constrains.
60  */
61
62 /*
63  * Many thanks to jlemon for basic structure of tcp_syncache which is being
64  * followed here.
65  */
66
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
69
70 #include "opt_inet6.h"
71
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/hash.h>
75 #include <sys/jail.h>
76 #include <sys/kernel.h>
77 #include <sys/lock.h>
78 #include <sys/mutex.h>
79 #include <sys/malloc.h>
80 #include <sys/proc.h>
81 #include <sys/sbuf.h>
82 #include <sys/socket.h>
83 #include <sys/socketvar.h>
84 #include <sys/sysctl.h>
85
86 #include <net/if.h>
87 #include <net/if_var.h>
88 #include <net/route.h>
89 #include <net/vnet.h>
90
91 #include <netinet/in.h>
92 #include <netinet/in_systm.h>
93 #include <netinet/ip.h>
94 #include <netinet/in_var.h>
95 #include <netinet/in_pcb.h>
96 #include <netinet/ip_var.h>
97 #ifdef INET6
98 #include <netinet/ip6.h>
99 #include <netinet6/ip6_var.h>
100 #endif
101 #include <netinet/tcp.h>
102 #include <netinet/tcp_var.h>
103 #include <netinet/tcp_hostcache.h>
104 #ifdef INET6
105 #include <netinet6/tcp6_var.h>
106 #endif
107
108 #include <vm/uma.h>
109
110 /* Arbitrary values */
111 #define TCP_HOSTCACHE_HASHSIZE          512
112 #define TCP_HOSTCACHE_BUCKETLIMIT       30
113 #define TCP_HOSTCACHE_EXPIRE            60*60   /* one hour */
114 #define TCP_HOSTCACHE_PRUNE             5*60    /* every 5 minutes */
115
116 VNET_DEFINE_STATIC(struct tcp_hostcache, tcp_hostcache);
117 #define V_tcp_hostcache         VNET(tcp_hostcache)
118
119 VNET_DEFINE_STATIC(struct callout, tcp_hc_callout);
120 #define V_tcp_hc_callout        VNET(tcp_hc_callout)
121
122 static struct hc_metrics *tcp_hc_lookup(struct in_conninfo *);
123 static struct hc_metrics *tcp_hc_insert(struct in_conninfo *);
124 static int sysctl_tcp_hc_list(SYSCTL_HANDLER_ARGS);
125 static int sysctl_tcp_hc_histo(SYSCTL_HANDLER_ARGS);
126 static int sysctl_tcp_hc_purgenow(SYSCTL_HANDLER_ARGS);
127 static void tcp_hc_purge_internal(int);
128 static void tcp_hc_purge(void *);
129
130 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hostcache,
131     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
132     "TCP Host cache");
133
134 VNET_DEFINE(int, tcp_use_hostcache) = 1;
135 #define V_tcp_use_hostcache  VNET(tcp_use_hostcache)
136 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
137     &VNET_NAME(tcp_use_hostcache), 0,
138     "Enable the TCP hostcache");
139
140 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
141     &VNET_NAME(tcp_hostcache.cache_limit), 0,
142     "Overall entry limit for hostcache");
143
144 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
145     &VNET_NAME(tcp_hostcache.hashsize), 0,
146     "Size of TCP hostcache hashtable");
147
148 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, bucketlimit,
149     CTLFLAG_VNET | CTLFLAG_RDTUN, &VNET_NAME(tcp_hostcache.bucket_limit), 0,
150     "Per-bucket hash limit for hostcache");
151
152 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, count, CTLFLAG_VNET | CTLFLAG_RD,
153      &VNET_NAME(tcp_hostcache.cache_count), 0,
154     "Current number of entries in hostcache");
155
156 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, expire, CTLFLAG_VNET | CTLFLAG_RW,
157     &VNET_NAME(tcp_hostcache.expire), 0,
158     "Expire time of TCP hostcache entries");
159
160 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, prune, CTLFLAG_VNET | CTLFLAG_RW,
161     &VNET_NAME(tcp_hostcache.prune), 0,
162     "Time between purge runs");
163
164 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, purge, CTLFLAG_VNET | CTLFLAG_RW,
165     &VNET_NAME(tcp_hostcache.purgeall), 0,
166     "Expire all entries on next purge run");
167
168 SYSCTL_PROC(_net_inet_tcp_hostcache, OID_AUTO, list,
169     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_MPSAFE,
170     0, 0, sysctl_tcp_hc_list, "A",
171     "List of all hostcache entries");
172
173 SYSCTL_PROC(_net_inet_tcp_hostcache, OID_AUTO, histo,
174     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_MPSAFE,
175     0, 0, sysctl_tcp_hc_histo, "A",
176     "Print a histogram of hostcache hashbucket utilization");
177
178 SYSCTL_PROC(_net_inet_tcp_hostcache, OID_AUTO, purgenow,
179     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
180     NULL, 0, sysctl_tcp_hc_purgenow, "I",
181     "Immediately purge all entries");
182
183 static MALLOC_DEFINE(M_HOSTCACHE, "hostcache", "TCP hostcache");
184
185 /* Use jenkins_hash32(), as in other parts of the tcp stack */
186 #define HOSTCACHE_HASH(ip) \
187         (jenkins_hash32((uint32_t *)(ip), 1, V_tcp_hostcache.hashsalt) & \
188          V_tcp_hostcache.hashmask)
189
190 #define HOSTCACHE_HASH6(ip6)                            \
191         (jenkins_hash32((uint32_t *)&((ip6)->s6_addr32[0]), 4, \
192          V_tcp_hostcache.hashsalt) & \
193          V_tcp_hostcache.hashmask)
194
195 #define THC_LOCK(lp)            mtx_lock(lp)
196 #define THC_UNLOCK(lp)          mtx_unlock(lp)
197
198 void
199 tcp_hc_init(void)
200 {
201         u_int cache_limit;
202         int i;
203
204         /*
205          * Initialize hostcache structures.
206          */
207         atomic_store_int(&V_tcp_hostcache.cache_count, 0);
208         V_tcp_hostcache.hashsize = TCP_HOSTCACHE_HASHSIZE;
209         V_tcp_hostcache.bucket_limit = TCP_HOSTCACHE_BUCKETLIMIT;
210         V_tcp_hostcache.expire = TCP_HOSTCACHE_EXPIRE;
211         V_tcp_hostcache.prune = TCP_HOSTCACHE_PRUNE;
212         V_tcp_hostcache.hashsalt = arc4random();
213
214         TUNABLE_INT_FETCH("net.inet.tcp.hostcache.hashsize",
215             &V_tcp_hostcache.hashsize);
216         if (!powerof2(V_tcp_hostcache.hashsize)) {
217                 printf("WARNING: hostcache hash size is not a power of 2.\n");
218                 V_tcp_hostcache.hashsize = TCP_HOSTCACHE_HASHSIZE; /* default */
219         }
220         V_tcp_hostcache.hashmask = V_tcp_hostcache.hashsize - 1;
221
222         TUNABLE_INT_FETCH("net.inet.tcp.hostcache.bucketlimit",
223             &V_tcp_hostcache.bucket_limit);
224
225         cache_limit = V_tcp_hostcache.hashsize * V_tcp_hostcache.bucket_limit;
226         V_tcp_hostcache.cache_limit = cache_limit;
227         TUNABLE_INT_FETCH("net.inet.tcp.hostcache.cachelimit",
228             &V_tcp_hostcache.cache_limit);
229         if (V_tcp_hostcache.cache_limit > cache_limit)
230                 V_tcp_hostcache.cache_limit = cache_limit;
231
232         /*
233          * Allocate the hash table.
234          */
235         V_tcp_hostcache.hashbase = (struct hc_head *)
236             malloc(V_tcp_hostcache.hashsize * sizeof(struct hc_head),
237                    M_HOSTCACHE, M_WAITOK | M_ZERO);
238
239         /*
240          * Initialize the hash buckets.
241          */
242         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
243                 TAILQ_INIT(&V_tcp_hostcache.hashbase[i].hch_bucket);
244                 V_tcp_hostcache.hashbase[i].hch_length = 0;
245                 mtx_init(&V_tcp_hostcache.hashbase[i].hch_mtx, "tcp_hc_entry",
246                           NULL, MTX_DEF);
247         }
248
249         /*
250          * Allocate the hostcache entries.
251          */
252         V_tcp_hostcache.zone =
253             uma_zcreate("hostcache", sizeof(struct hc_metrics),
254             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
255         uma_zone_set_max(V_tcp_hostcache.zone, V_tcp_hostcache.cache_limit);
256
257         /*
258          * Set up periodic cache cleanup.
259          */
260         callout_init(&V_tcp_hc_callout, 1);
261         callout_reset(&V_tcp_hc_callout, V_tcp_hostcache.prune * hz,
262             tcp_hc_purge, curvnet);
263 }
264
265 #ifdef VIMAGE
266 void
267 tcp_hc_destroy(void)
268 {
269         int i;
270
271         callout_drain(&V_tcp_hc_callout);
272
273         /* Purge all hc entries. */
274         tcp_hc_purge_internal(1);
275
276         /* Free the uma zone and the allocated hash table. */
277         uma_zdestroy(V_tcp_hostcache.zone);
278
279         for (i = 0; i < V_tcp_hostcache.hashsize; i++)
280                 mtx_destroy(&V_tcp_hostcache.hashbase[i].hch_mtx);
281         free(V_tcp_hostcache.hashbase, M_HOSTCACHE);
282 }
283 #endif
284
285 /*
286  * Internal function: look up an entry in the hostcache or return NULL.
287  *
288  * If an entry has been returned, the caller becomes responsible for
289  * unlocking the bucket row after he is done reading/modifying the entry.
290  */
291 static struct hc_metrics *
292 tcp_hc_lookup(struct in_conninfo *inc)
293 {
294         int hash;
295         struct hc_head *hc_head;
296         struct hc_metrics *hc_entry;
297
298         if (!V_tcp_use_hostcache)
299                 return NULL;
300
301         KASSERT(inc != NULL, ("tcp_hc_lookup with NULL in_conninfo pointer"));
302
303         /*
304          * Hash the foreign ip address.
305          */
306         if (inc->inc_flags & INC_ISIPV6)
307                 hash = HOSTCACHE_HASH6(&inc->inc6_faddr);
308         else
309                 hash = HOSTCACHE_HASH(&inc->inc_faddr);
310
311         hc_head = &V_tcp_hostcache.hashbase[hash];
312
313         /*
314          * Acquire lock for this bucket row; we release the lock if we don't
315          * find an entry, otherwise the caller has to unlock after he is
316          * done.
317          */
318         THC_LOCK(&hc_head->hch_mtx);
319
320         /*
321          * Iterate through entries in bucket row looking for a match.
322          */
323         TAILQ_FOREACH(hc_entry, &hc_head->hch_bucket, rmx_q) {
324                 if (inc->inc_flags & INC_ISIPV6) {
325                         /* XXX: check ip6_zoneid */
326                         if (memcmp(&inc->inc6_faddr, &hc_entry->ip6,
327                             sizeof(inc->inc6_faddr)) == 0)
328                                 return hc_entry;
329                 } else {
330                         if (memcmp(&inc->inc_faddr, &hc_entry->ip4,
331                             sizeof(inc->inc_faddr)) == 0)
332                                 return hc_entry;
333                 }
334         }
335
336         /*
337          * We were unsuccessful and didn't find anything.
338          */
339         THC_UNLOCK(&hc_head->hch_mtx);
340         return NULL;
341 }
342
343 /*
344  * Internal function: insert an entry into the hostcache or return NULL if
345  * unable to allocate a new one.
346  *
347  * If an entry has been returned, the caller becomes responsible for
348  * unlocking the bucket row after he is done reading/modifying the entry.
349  */
350 static struct hc_metrics *
351 tcp_hc_insert(struct in_conninfo *inc)
352 {
353         int hash;
354         struct hc_head *hc_head;
355         struct hc_metrics *hc_entry;
356
357         if (!V_tcp_use_hostcache)
358                 return NULL;
359
360         KASSERT(inc != NULL, ("tcp_hc_insert with NULL in_conninfo pointer"));
361
362         /*
363          * Hash the foreign ip address.
364          */
365         if (inc->inc_flags & INC_ISIPV6)
366                 hash = HOSTCACHE_HASH6(&inc->inc6_faddr);
367         else
368                 hash = HOSTCACHE_HASH(&inc->inc_faddr);
369
370         hc_head = &V_tcp_hostcache.hashbase[hash];
371
372         /*
373          * Acquire lock for this bucket row; we release the lock if we don't
374          * find an entry, otherwise the caller has to unlock after he is
375          * done.
376          */
377         THC_LOCK(&hc_head->hch_mtx);
378
379         /*
380          * If the bucket limit is reached, reuse the least-used element.
381          */
382         if (hc_head->hch_length >= V_tcp_hostcache.bucket_limit ||
383             atomic_load_int(&V_tcp_hostcache.cache_count) >= V_tcp_hostcache.cache_limit) {
384                 hc_entry = TAILQ_LAST(&hc_head->hch_bucket, hc_qhead);
385                 /*
386                  * At first we were dropping the last element, just to
387                  * reacquire it in the next two lines again, which isn't very
388                  * efficient.  Instead just reuse the least used element.
389                  * We may drop something that is still "in-use" but we can be
390                  * "lossy".
391                  * Just give up if this bucket row is empty and we don't have
392                  * anything to replace.
393                  */
394                 if (hc_entry == NULL) {
395                         THC_UNLOCK(&hc_head->hch_mtx);
396                         return NULL;
397                 }
398                 TAILQ_REMOVE(&hc_head->hch_bucket, hc_entry, rmx_q);
399                 KASSERT(V_tcp_hostcache.hashbase[hash].hch_length > 0 &&
400                         V_tcp_hostcache.hashbase[hash].hch_length <=
401                         V_tcp_hostcache.bucket_limit,
402                         ("tcp_hostcache: bucket length range violated at %u: %u",
403                         hash, V_tcp_hostcache.hashbase[hash].hch_length));
404                 V_tcp_hostcache.hashbase[hash].hch_length--;
405                 atomic_subtract_int(&V_tcp_hostcache.cache_count, 1);
406                 TCPSTAT_INC(tcps_hc_bucketoverflow);
407 #if 0
408                 uma_zfree(V_tcp_hostcache.zone, hc_entry);
409 #endif
410         } else {
411                 /*
412                  * Allocate a new entry, or balk if not possible.
413                  */
414                 hc_entry = uma_zalloc(V_tcp_hostcache.zone, M_NOWAIT);
415                 if (hc_entry == NULL) {
416                         THC_UNLOCK(&hc_head->hch_mtx);
417                         return NULL;
418                 }
419         }
420
421         /*
422          * Initialize basic information of hostcache entry.
423          */
424         bzero(hc_entry, sizeof(*hc_entry));
425         if (inc->inc_flags & INC_ISIPV6) {
426                 hc_entry->ip6 = inc->inc6_faddr;
427                 hc_entry->ip6_zoneid = inc->inc6_zoneid;
428         } else
429                 hc_entry->ip4 = inc->inc_faddr;
430         hc_entry->rmx_head = hc_head;
431         hc_entry->rmx_expire = V_tcp_hostcache.expire;
432
433         /*
434          * Put it upfront.
435          */
436         TAILQ_INSERT_HEAD(&hc_head->hch_bucket, hc_entry, rmx_q);
437         V_tcp_hostcache.hashbase[hash].hch_length++;
438         KASSERT(V_tcp_hostcache.hashbase[hash].hch_length <
439                 V_tcp_hostcache.bucket_limit,
440                 ("tcp_hostcache: bucket length too high at %u: %u",
441                 hash, V_tcp_hostcache.hashbase[hash].hch_length));
442         atomic_add_int(&V_tcp_hostcache.cache_count, 1);
443         TCPSTAT_INC(tcps_hc_added);
444
445         return hc_entry;
446 }
447
448 /*
449  * External function: look up an entry in the hostcache and fill out the
450  * supplied TCP metrics structure.  Fills in NULL when no entry was found or
451  * a value is not set.
452  */
453 void
454 tcp_hc_get(struct in_conninfo *inc, struct hc_metrics_lite *hc_metrics_lite)
455 {
456         struct hc_metrics *hc_entry;
457
458         if (!V_tcp_use_hostcache) {
459                 bzero(hc_metrics_lite, sizeof(*hc_metrics_lite));
460                 return;
461         }
462
463         /*
464          * Find the right bucket.
465          */
466         hc_entry = tcp_hc_lookup(inc);
467
468         /*
469          * If we don't have an existing object.
470          */
471         if (hc_entry == NULL) {
472                 bzero(hc_metrics_lite, sizeof(*hc_metrics_lite));
473                 return;
474         }
475         hc_entry->rmx_hits++;
476         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
477
478         hc_metrics_lite->rmx_mtu = hc_entry->rmx_mtu;
479         hc_metrics_lite->rmx_ssthresh = hc_entry->rmx_ssthresh;
480         hc_metrics_lite->rmx_rtt = hc_entry->rmx_rtt;
481         hc_metrics_lite->rmx_rttvar = hc_entry->rmx_rttvar;
482         hc_metrics_lite->rmx_cwnd = hc_entry->rmx_cwnd;
483         hc_metrics_lite->rmx_sendpipe = hc_entry->rmx_sendpipe;
484         hc_metrics_lite->rmx_recvpipe = hc_entry->rmx_recvpipe;
485
486         /*
487          * Unlock bucket row.
488          */
489         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
490 }
491
492 /*
493  * External function: look up an entry in the hostcache and return the
494  * discovered path MTU.  Returns 0 if no entry is found or value is not
495  * set.
496  */
497 uint32_t
498 tcp_hc_getmtu(struct in_conninfo *inc)
499 {
500         struct hc_metrics *hc_entry;
501         uint32_t mtu;
502
503         if (!V_tcp_use_hostcache)
504                 return 0;
505
506         hc_entry = tcp_hc_lookup(inc);
507         if (hc_entry == NULL) {
508                 return 0;
509         }
510         hc_entry->rmx_hits++;
511         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
512
513         mtu = hc_entry->rmx_mtu;
514         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
515         return mtu;
516 }
517
518 /*
519  * External function: update the MTU value of an entry in the hostcache.
520  * Creates a new entry if none was found.
521  */
522 void
523 tcp_hc_updatemtu(struct in_conninfo *inc, uint32_t mtu)
524 {
525         struct hc_metrics *hc_entry;
526
527         if (!V_tcp_use_hostcache)
528                 return;
529
530         /*
531          * Find the right bucket.
532          */
533         hc_entry = tcp_hc_lookup(inc);
534
535         /*
536          * If we don't have an existing object, try to insert a new one.
537          */
538         if (hc_entry == NULL) {
539                 hc_entry = tcp_hc_insert(inc);
540                 if (hc_entry == NULL)
541                         return;
542         }
543         hc_entry->rmx_updates++;
544         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
545
546         hc_entry->rmx_mtu = mtu;
547
548         /*
549          * Put it upfront so we find it faster next time.
550          */
551         TAILQ_REMOVE(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
552         TAILQ_INSERT_HEAD(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
553
554         /*
555          * Unlock bucket row.
556          */
557         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
558 }
559
560 /*
561  * External function: update the TCP metrics of an entry in the hostcache.
562  * Creates a new entry if none was found.
563  */
564 void
565 tcp_hc_update(struct in_conninfo *inc, struct hc_metrics_lite *hcml)
566 {
567         struct hc_metrics *hc_entry;
568
569         if (!V_tcp_use_hostcache)
570                 return;
571
572         hc_entry = tcp_hc_lookup(inc);
573         if (hc_entry == NULL) {
574                 hc_entry = tcp_hc_insert(inc);
575                 if (hc_entry == NULL)
576                         return;
577         }
578         hc_entry->rmx_updates++;
579         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
580
581         if (hcml->rmx_rtt != 0) {
582                 if (hc_entry->rmx_rtt == 0)
583                         hc_entry->rmx_rtt = hcml->rmx_rtt;
584                 else
585                         hc_entry->rmx_rtt = ((uint64_t)hc_entry->rmx_rtt +
586                             (uint64_t)hcml->rmx_rtt) / 2;
587                 TCPSTAT_INC(tcps_cachedrtt);
588         }
589         if (hcml->rmx_rttvar != 0) {
590                 if (hc_entry->rmx_rttvar == 0)
591                         hc_entry->rmx_rttvar = hcml->rmx_rttvar;
592                 else
593                         hc_entry->rmx_rttvar = ((uint64_t)hc_entry->rmx_rttvar +
594                             (uint64_t)hcml->rmx_rttvar) / 2;
595                 TCPSTAT_INC(tcps_cachedrttvar);
596         }
597         if (hcml->rmx_ssthresh != 0) {
598                 if (hc_entry->rmx_ssthresh == 0)
599                         hc_entry->rmx_ssthresh = hcml->rmx_ssthresh;
600                 else
601                         hc_entry->rmx_ssthresh =
602                             (hc_entry->rmx_ssthresh + hcml->rmx_ssthresh) / 2;
603                 TCPSTAT_INC(tcps_cachedssthresh);
604         }
605         if (hcml->rmx_cwnd != 0) {
606                 if (hc_entry->rmx_cwnd == 0)
607                         hc_entry->rmx_cwnd = hcml->rmx_cwnd;
608                 else
609                         hc_entry->rmx_cwnd = ((uint64_t)hc_entry->rmx_cwnd +
610                             (uint64_t)hcml->rmx_cwnd) / 2;
611                 /* TCPSTAT_INC(tcps_cachedcwnd); */
612         }
613         if (hcml->rmx_sendpipe != 0) {
614                 if (hc_entry->rmx_sendpipe == 0)
615                         hc_entry->rmx_sendpipe = hcml->rmx_sendpipe;
616                 else
617                         hc_entry->rmx_sendpipe =
618                             ((uint64_t)hc_entry->rmx_sendpipe +
619                             (uint64_t)hcml->rmx_sendpipe) /2;
620                 /* TCPSTAT_INC(tcps_cachedsendpipe); */
621         }
622         if (hcml->rmx_recvpipe != 0) {
623                 if (hc_entry->rmx_recvpipe == 0)
624                         hc_entry->rmx_recvpipe = hcml->rmx_recvpipe;
625                 else
626                         hc_entry->rmx_recvpipe =
627                             ((uint64_t)hc_entry->rmx_recvpipe +
628                             (uint64_t)hcml->rmx_recvpipe) /2;
629                 /* TCPSTAT_INC(tcps_cachedrecvpipe); */
630         }
631
632         TAILQ_REMOVE(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
633         TAILQ_INSERT_HEAD(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
634         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
635 }
636
637 /*
638  * Sysctl function: prints the list and values of all hostcache entries in
639  * unsorted order.
640  */
641 static int
642 sysctl_tcp_hc_list(SYSCTL_HANDLER_ARGS)
643 {
644         const int linesize = 128;
645         struct sbuf sb;
646         int i, error, len;
647         struct hc_metrics *hc_entry;
648         char ip4buf[INET_ADDRSTRLEN];
649 #ifdef INET6
650         char ip6buf[INET6_ADDRSTRLEN];
651 #endif
652
653         if (jailed_without_vnet(curthread->td_ucred) != 0)
654                 return (EPERM);
655
656         /* Optimize Buffer length query by sbin/sysctl */
657         if (req->oldptr == NULL) {
658                 len = (atomic_load_int(&V_tcp_hostcache.cache_count) + 1) *
659                         linesize;
660                 return (SYSCTL_OUT(req, NULL, len));
661         }
662
663         error = sysctl_wire_old_buffer(req, 0);
664         if (error != 0) {
665                 return(error);
666         }
667
668         /* Use a buffer sized for one full bucket */
669         sbuf_new_for_sysctl(&sb, NULL, V_tcp_hostcache.bucket_limit *
670                 linesize, req);
671
672         sbuf_printf(&sb,
673                 "\nIP address        MTU  SSTRESH      RTT   RTTVAR "
674                 "    CWND SENDPIPE RECVPIPE HITS  UPD  EXP\n");
675         sbuf_drain(&sb);
676
677 #define msec(u) (((u) + 500) / 1000)
678         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
679                 THC_LOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
680                 TAILQ_FOREACH(hc_entry, &V_tcp_hostcache.hashbase[i].hch_bucket,
681                               rmx_q) {
682                         sbuf_printf(&sb,
683                             "%-15s %5u %8u %6lums %6lums %8u %8u %8u %4lu "
684                             "%4lu %4i\n",
685                             hc_entry->ip4.s_addr ?
686                                 inet_ntoa_r(hc_entry->ip4, ip4buf) :
687 #ifdef INET6
688                                 ip6_sprintf(ip6buf, &hc_entry->ip6),
689 #else
690                                 "IPv6?",
691 #endif
692                             hc_entry->rmx_mtu,
693                             hc_entry->rmx_ssthresh,
694                             msec((u_long)hc_entry->rmx_rtt *
695                                 (RTM_RTTUNIT / (hz * TCP_RTT_SCALE))),
696                             msec((u_long)hc_entry->rmx_rttvar *
697                                 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE))),
698                             hc_entry->rmx_cwnd,
699                             hc_entry->rmx_sendpipe,
700                             hc_entry->rmx_recvpipe,
701                             hc_entry->rmx_hits,
702                             hc_entry->rmx_updates,
703                             hc_entry->rmx_expire);
704                 }
705                 THC_UNLOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
706                 sbuf_drain(&sb);
707         }
708 #undef msec
709         error = sbuf_finish(&sb);
710         sbuf_delete(&sb);
711         return(error);
712 }
713
714 /*
715  * Sysctl function: prints a histogram of the hostcache hashbucket
716  * utilization.
717  */
718 static int
719 sysctl_tcp_hc_histo(SYSCTL_HANDLER_ARGS)
720 {
721         const int linesize = 50;
722         struct sbuf sb;
723         int i, error;
724         int *histo;
725         u_int hch_length;
726
727         if (jailed_without_vnet(curthread->td_ucred) != 0)
728                 return (EPERM);
729
730         histo = (int *)malloc(sizeof(int) * (V_tcp_hostcache.bucket_limit + 1),
731                         M_TEMP, M_NOWAIT|M_ZERO);
732         if (histo == NULL)
733                 return(ENOMEM);
734
735         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
736                 hch_length = V_tcp_hostcache.hashbase[i].hch_length;
737                 KASSERT(hch_length <= V_tcp_hostcache.bucket_limit,
738                         ("tcp_hostcache: bucket limit exceeded at %u: %u",
739                         i, hch_length));
740                 histo[hch_length]++;
741         }
742
743         /* Use a buffer for 16 lines */
744         sbuf_new_for_sysctl(&sb, NULL, 16 * linesize, req);
745
746         sbuf_printf(&sb, "\nLength\tCount\n");
747         for (i = 0; i <= V_tcp_hostcache.bucket_limit; i++) {
748                 sbuf_printf(&sb, "%u\t%u\n", i, histo[i]);
749         }
750         error = sbuf_finish(&sb);
751         sbuf_delete(&sb);
752         free(histo, M_TEMP);
753         return(error);
754 }
755
756 /*
757  * Caller has to make sure the curvnet is set properly.
758  */
759 static void
760 tcp_hc_purge_internal(int all)
761 {
762         struct hc_metrics *hc_entry, *hc_next;
763         int i;
764
765         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
766                 THC_LOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
767                 TAILQ_FOREACH_SAFE(hc_entry,
768                     &V_tcp_hostcache.hashbase[i].hch_bucket, rmx_q, hc_next) {
769                         KASSERT(V_tcp_hostcache.hashbase[i].hch_length > 0 &&
770                                 V_tcp_hostcache.hashbase[i].hch_length <=
771                                 V_tcp_hostcache.bucket_limit,
772                                 ("tcp_hostcache: bucket length out of range at %u: %u",
773                                 i, V_tcp_hostcache.hashbase[i].hch_length));
774                         if (all || hc_entry->rmx_expire <= 0) {
775                                 TAILQ_REMOVE(&V_tcp_hostcache.hashbase[i].hch_bucket,
776                                               hc_entry, rmx_q);
777                                 uma_zfree(V_tcp_hostcache.zone, hc_entry);
778                                 V_tcp_hostcache.hashbase[i].hch_length--;
779                                 atomic_subtract_int(&V_tcp_hostcache.cache_count, 1);
780                         } else
781                                 hc_entry->rmx_expire -= V_tcp_hostcache.prune;
782                 }
783                 THC_UNLOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
784         }
785 }
786
787 /*
788  * Expire and purge (old|all) entries in the tcp_hostcache.  Runs
789  * periodically from the callout.
790  */
791 static void
792 tcp_hc_purge(void *arg)
793 {
794         CURVNET_SET((struct vnet *) arg);
795         int all = 0;
796
797         if (V_tcp_hostcache.purgeall) {
798                 if (V_tcp_hostcache.purgeall == 2)
799                         V_tcp_hostcache.hashsalt = arc4random();
800                 all = 1;
801                 V_tcp_hostcache.purgeall = 0;
802         }
803
804         tcp_hc_purge_internal(all);
805
806         callout_reset(&V_tcp_hc_callout, V_tcp_hostcache.prune * hz,
807             tcp_hc_purge, arg);
808         CURVNET_RESTORE();
809 }
810
811 /*
812  * Expire and purge all entries in hostcache immediately.
813  */
814 static int
815 sysctl_tcp_hc_purgenow(SYSCTL_HANDLER_ARGS)
816 {
817         int error, val;
818
819         val = 0;
820         error = sysctl_handle_int(oidp, &val, 0, req);
821         if (error || !req->newptr)
822                 return (error);
823
824         if (val == 2)
825                 V_tcp_hostcache.hashsalt = arc4random();
826         tcp_hc_purge_internal(1);
827
828         callout_reset(&V_tcp_hc_callout, V_tcp_hostcache.prune * hz,
829             tcp_hc_purge, curvnet);
830
831         return (0);
832 }