]> git.proxmox.com Git - mirror_iproute2.git/blame - Modules/tcp_diag.c
projects / mirror_iproute2.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
(Logical change 1.3)
[mirror_iproute2.git] / Modules / tcp_diag.c
CommitLineData
aba5acdf
SH		 1/*
2 * tcp_diag.c Module for monitoring TCP sockets.
3 *
4 * Version: $
5 *
6 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
7 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public License
10 * as published by the Free Software Foundation; either version
11 * 2 of the License, or (at your option) any later version.
12 */
13
14#include <linux/config.h>
15#include <linux/module.h>
16#include <linux/types.h>
17#include <linux/fcntl.h>
18#include <linux/random.h>
19#include <linux/cache.h>
20#include <linux/init.h>
21
22#include <net/icmp.h>
23#include <net/tcp.h>
24#include <net/ipv6.h>
25#include <net/inet_common.h>
26
27#include <linux/inet.h>
28#include <linux/stddef.h>
29
30#include "tcp_diag.h"
31
32static struct sock *tcpnl;
33
34
35#define TCPDIAG_PUT(skb, attrtype, attrlen) \
36({ int rtalen = RTA_LENGTH(attrlen); \
37 struct rtattr *rta; \
38 if (skb_tailroom(skb) < RTA_ALIGN(rtalen)) goto nlmsg_failure; \
39 rta = (void*)__skb_put(skb, RTA_ALIGN(rtalen)); \
40 rta->rta_type = attrtype; \
41 rta->rta_len = rtalen; \
42 RTA_DATA(rta); })
43
44static int tcpdiag_fill(struct sk_buff *skb, struct sock *sk,
45 int ext, u32 pid, u32 seq)
46{
47 struct tcp_opt *tp = &sk->tp_pinfo.af_tcp;
48 struct tcpdiagmsg *r;
49 struct nlmsghdr *nlh;
50 struct tcp_info *info = NULL;
51 struct tcpdiag_meminfo *minfo = NULL;
52 unsigned char *b = skb->tail;
53
54 nlh = NLMSG_PUT(skb, pid, seq, TCPDIAG_GETSOCK, sizeof(*r));
55 r = NLMSG_DATA(nlh);
56 if (sk->state != TCP_TIME_WAIT) {
57 if (ext & (1<<(TCPDIAG_MEMINFO-1)))
58 minfo = TCPDIAG_PUT(skb, TCPDIAG_MEMINFO, sizeof(*minfo));
59 if (ext & (1<<(TCPDIAG_INFO-1)))
60 info = TCPDIAG_PUT(skb, TCPDIAG_INFO, sizeof(*info));
61 }
62 r->tcpdiag_family = sk->family;
63 r->tcpdiag_state = sk->state;
64 r->tcpdiag_timer = 0;
65 r->tcpdiag_retrans = 0;
66
67 r->id.tcpdiag_sport = sk->sport;
68 r->id.tcpdiag_dport = sk->dport;
69 r->id.tcpdiag_src[0] = sk->rcv_saddr;
70 r->id.tcpdiag_dst[0] = sk->daddr;
71 r->id.tcpdiag_if = sk->bound_dev_if;
72 *((struct sock **)&r->id.tcpdiag_cookie) = sk;
73
74 if (r->tcpdiag_state == TCP_TIME_WAIT) {
75 struct tcp_tw_bucket *tw = (struct tcp_tw_bucket*)sk;
76 long tmo = tw->ttd - jiffies;
77 if (tmo < 0)
78 tmo = 0;
79
80 r->tcpdiag_state = tw->substate;
81 r->tcpdiag_timer = 3;
82 r->tcpdiag_expires = (tmo*1000+HZ-1)/HZ;
83 r->tcpdiag_rqueue = 0;
84 r->tcpdiag_wqueue = 0;
85 r->tcpdiag_uid = 0;
86 r->tcpdiag_inode = 0;
87#ifdef CONFIG_IPV6
88 if (r->tcpdiag_family == AF_INET6) {
89 memcpy(r->id.tcpdiag_src, &tw->v6_rcv_saddr, 16);
90 memcpy(r->id.tcpdiag_dst, &tw->v6_daddr, 16);
91 }
92#endif
93 nlh->nlmsg_len = skb->tail - b;
94 return skb->len;
95 }
96
97#ifdef CONFIG_IPV6
98 if (r->tcpdiag_family == AF_INET6) {
99 memcpy(r->id.tcpdiag_src, &sk->net_pinfo.af_inet6.rcv_saddr, 16);
100 memcpy(r->id.tcpdiag_dst, &sk->net_pinfo.af_inet6.daddr, 16);
101 }
102#endif
103
104#define EXPIRES_IN_MS(tmo) ((tmo-jiffies)*1000+HZ-1)/HZ
105
106 if (tp->pending == TCP_TIME_RETRANS) {
107 r->tcpdiag_timer = 1;
108 r->tcpdiag_retrans = tp->retransmits;
109 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
110 } else if (tp->pending == TCP_TIME_PROBE0) {
111 r->tcpdiag_timer = 4;
112 r->tcpdiag_retrans = tp->probes_out;
113 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
114 } else if (timer_pending(&sk->timer)) {
115 r->tcpdiag_timer = 2;
116 r->tcpdiag_retrans = tp->probes_out;
117 r->tcpdiag_expires = EXPIRES_IN_MS(sk->timer.expires);
118 } else {
119 r->tcpdiag_timer = 0;
120 r->tcpdiag_expires = 0;
121 }
122#undef EXPIRES_IN_MS
123
124 r->tcpdiag_rqueue = tp->rcv_nxt - tp->copied_seq;
125 r->tcpdiag_wqueue = tp->write_seq - tp->snd_una;
126 r->tcpdiag_uid = sock_i_uid(sk);
127 r->tcpdiag_inode = sock_i_ino(sk);
128
129 if (minfo) {
130 minfo->tcpdiag_rmem = atomic_read(&sk->rmem_alloc);
131 minfo->tcpdiag_wmem = sk->wmem_queued;
132 minfo->tcpdiag_fmem = sk->forward_alloc;
133 minfo->tcpdiag_tmem = atomic_read(&sk->wmem_alloc);
134 }
135
136 if (info) {
137 u32 now = tcp_time_stamp;
138
139 info->tcpi_state = sk->state;
140 info->tcpi_ca_state = tp->ca_state;
141 info->tcpi_retransmits = tp->retransmits;
142 info->tcpi_probes = tp->probes_out;
143 info->tcpi_backoff = tp->backoff;
144 info->tcpi_options = 0;
145 if (tp->tstamp_ok)
146 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
147 if (tp->sack_ok)
148 info->tcpi_options |= TCPI_OPT_SACK;
149 if (tp->wscale_ok) {
150 info->tcpi_options |= TCPI_OPT_WSCALE;
151 info->tcpi_snd_wscale = tp->snd_wscale;
152 info->tcpi_rcv_wscale = tp->rcv_wscale;
153 } else {
154 info->tcpi_snd_wscale = 0;
155 info->tcpi_rcv_wscale = 0;
156 }
157#ifdef CONFIG_INET_ECN
158 if (tp->ecn_flags&TCP_ECN_OK)
159 info->tcpi_options |= TCPI_OPT_ECN;
160#endif
161
162 info->tcpi_rto = (1000000*tp->rto)/HZ;
163 info->tcpi_ato = (1000000*tp->ack.ato)/HZ;
164 info->tcpi_snd_mss = tp->mss_cache;
165 info->tcpi_rcv_mss = tp->ack.rcv_mss;
166
167 info->tcpi_unacked = tp->packets_out;
168 info->tcpi_sacked = tp->sacked_out;
169 info->tcpi_lost = tp->lost_out;
170 info->tcpi_retrans = tp->retrans_out;
171 info->tcpi_fackets = tp->fackets_out;
172
173 info->tcpi_last_data_sent = ((now - tp->lsndtime)*1000)/HZ;
174 info->tcpi_last_ack_sent = 0;
175 info->tcpi_last_data_recv = ((now - tp->ack.lrcvtime)*1000)/HZ;
176 info->tcpi_last_ack_recv = ((now - tp->rcv_tstamp)*1000)/HZ;
177
178 info->tcpi_pmtu = tp->pmtu_cookie;
179 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
180 info->tcpi_rtt = ((1000000*tp->srtt)/HZ)>>3;
181 info->tcpi_rttvar = ((1000000*tp->mdev)/HZ)>>2;
182 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
183 info->tcpi_snd_cwnd = tp->snd_cwnd;
184 info->tcpi_advmss = tp->advmss;
185 info->tcpi_reordering = tp->reordering;
186 }
187
188 nlh->nlmsg_len = skb->tail - b;
189 return skb->len;
190
191nlmsg_failure:
192 skb_trim(skb, b - skb->data);
193 return -1;
194}
195
196extern struct sock *tcp_v4_lookup(u32 saddr, u16 sport, u32 daddr, u16 dport, int dif);
197#ifdef CONFIG_IPV6
198extern struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
199 struct in6_addr *daddr, u16 dport,
200 int dif);
201#endif
202
203static int tcpdiag_get_exact(struct sk_buff *in_skb, struct nlmsghdr *nlh)
204{
205 int err;
206 struct sock *sk;
207 struct tcpdiagreq *req = NLMSG_DATA(nlh);
208 struct sk_buff *rep;
209
210 if (req->tcpdiag_family == AF_INET) {
211 sk = tcp_v4_lookup(req->id.tcpdiag_dst[0], req->id.tcpdiag_dport,
212 req->id.tcpdiag_src[0], req->id.tcpdiag_sport,
213 req->id.tcpdiag_if);
214 }
215#ifdef CONFIG_IPV6
216 else if (req->tcpdiag_family == AF_INET6) {
217 sk = tcp_v6_lookup((struct in6_addr*)req->id.tcpdiag_dst, req->id.tcpdiag_dport,
218 (struct in6_addr*)req->id.tcpdiag_src, req->id.tcpdiag_sport,
219 req->id.tcpdiag_if);
220 }
221#endif
222 else {
223 return -EINVAL;
224 }
225
226 if (sk == NULL)
227 return -ENOENT;
228
229 err = -ESTALE;
230 if ((req->id.tcpdiag_cookie[0] != TCPDIAG_NOCOOKIE ||
231 req->id.tcpdiag_cookie[1] != TCPDIAG_NOCOOKIE) &&
232 sk != *((struct sock **)&req->id.tcpdiag_cookie[0]))
233 goto out;
234
235 err = -ENOMEM;
236 rep = alloc_skb(NLMSG_SPACE(sizeof(struct tcpdiagmsg)+
237 sizeof(struct tcpdiag_meminfo)+
238 sizeof(struct tcp_info)+64), GFP_KERNEL);
239 if (!rep)
240 goto out;
241
242 if (tcpdiag_fill(rep, sk, req->tcpdiag_ext,
243 NETLINK_CB(in_skb).pid,
244 nlh->nlmsg_seq) <= 0)
245 BUG();
246
247 err = netlink_unicast(tcpnl, rep, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
248 if (err > 0)
249 err = 0;
250
251out:
252 if (sk) {
253 if (sk->state == TCP_TIME_WAIT)
254 tcp_tw_put((struct tcp_tw_bucket*)sk);
255 else
256 sock_put(sk);
257 }
258 return err;
259}
260
261int bitstring_match(u32 *a1, u32 *a2, int bits)
262{
263 int words = bits >> 5;
264
265 bits &= 0x1f;
266
267 if (words) {
268 if (memcmp(a1, a2, words << 2))
269 return 0;
270 }
271 if (bits) {
272 __u32 w1, w2;
273 __u32 mask;
274
275 w1 = a1[words];
276 w2 = a2[words];
277
278 mask = htonl((0xffffffff) << (32 - bits));
279
280 if ((w1 ^ w2) & mask)
281 return 0;
282 }
283
284 return 1;
285}
286
287
288int tcpdiag_bc_run(char *bc, int len, struct sock *sk)
289{
290 while (len > 0) {
291 int yes = 1;
292 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
293
294 switch (op->code) {
295 case TCPDIAG_BC_NOP:
296 break;
297 case TCPDIAG_BC_JMP:
298 yes = 0;
299 break;
300 case TCPDIAG_BC_S_GE:
301 yes = (sk->num >= op[1].no);
302 break;
303 case TCPDIAG_BC_S_LE:
304 yes = (sk->num <= op[1].no);
305 break;
306 case TCPDIAG_BC_D_GE:
307 yes = (ntohs(sk->dport) >= op[1].no);
308 break;
309 case TCPDIAG_BC_D_LE:
310 yes = (ntohs(sk->dport) <= op[1].no);
311 break;
312 case TCPDIAG_BC_AUTO:
313 yes = !(sk->userlocks&SOCK_BINDPORT_LOCK);
314 break;
315 case TCPDIAG_BC_S_COND:
316 case TCPDIAG_BC_D_COND:
317 {
318 struct tcpdiag_hostcond *cond = (struct tcpdiag_hostcond*)(op+1);
319 u32 *addr;
320
321 if (cond->port != -1 &&
322 cond->port != (op->code == TCPDIAG_BC_S_COND ? sk->num : ntohs(sk->dport))) {
323 yes = 0;
324 break;
325 }
326
327 if (cond->prefix_len == 0)
328 break;
329
330 if (sk->family == AF_INET6) {
331 if (op->code == TCPDIAG_BC_S_COND)
332 addr = (u32*)&sk->net_pinfo.af_inet6.rcv_saddr;
333 else
334 addr = (u32*)&sk->net_pinfo.af_inet6.daddr;
335 } else {
336 if (op->code == TCPDIAG_BC_S_COND)
337 addr = &sk->rcv_saddr;
338 else
339 addr = &sk->daddr;
340 }
341
342 if (bitstring_match(addr, cond->addr, cond->prefix_len))
343 break;
344 if (sk->family == AF_INET6 && cond->family == AF_INET) {
345 if (addr[0] == 0 && addr[1] == 0 &&
346 addr[2] == __constant_htonl(0xffff) &&
347 bitstring_match(addr+3, cond->addr, cond->prefix_len))
348 break;
349 }
350 yes = 0;
351 break;
352 }
353 }
354
355 if (yes) {
356 len -= op->yes;
357 bc += op->yes;
358 } else {
359 len -= op->no;
360 bc += op->no;
361 }
362 }
363 return (len == 0);
364}
365
366int valid_cc(char *bc, int len, int cc)
367{
368 while (len >= 0) {
369 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
370
371 if (cc > len)
372 return 0;
373 if (cc == len)
374 return 1;
375 if (op->yes < 4)
376 return 0;
377 len -= op->yes;
378 bc += op->yes;
379 }
380 return 0;
381}
382
383int tcpdiag_bc_audit(char *bytecode, int bytecode_len)
384{
385 char *bc = bytecode;
386 int len = bytecode_len;
387
388 while (len > 0) {
389 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
390
391//printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
392 switch (op->code) {
393 case TCPDIAG_BC_AUTO:
394 case TCPDIAG_BC_S_COND:
395 case TCPDIAG_BC_D_COND:
396 case TCPDIAG_BC_S_GE:
397 case TCPDIAG_BC_S_LE:
398 case TCPDIAG_BC_D_GE:
399 case TCPDIAG_BC_D_LE:
400 if (op->yes < 4 || op->yes > len+4)
401 return -EINVAL;
402 case TCPDIAG_BC_JMP:
403 if (op->no < 4 || op->no > len+4)
404 return -EINVAL;
405 if (op->no < len &&
406 !valid_cc(bytecode, bytecode_len, len-op->no))
407 return -EINVAL;
408 break;
409 case TCPDIAG_BC_NOP:
410 if (op->yes < 4 || op->yes > len+4)
411 return -EINVAL;
412 break;
413 default:
414 return -EINVAL;
415 }
416 bc += op->yes;
417 len -= op->yes;
418 }
419 return len == 0 ? 0 : -EINVAL;
420}
421
422
423int tcpdiag_dump(struct sk_buff *skb, struct netlink_callback *cb)
424{
425 int i, num;
426 int s_i, s_num;
427 struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
428 struct rtattr *bc = NULL;
429
430 if (cb->nlh->nlmsg_len > 4+NLMSG_SPACE(sizeof(struct tcpdiagreq)))
431 bc = (struct rtattr*)(r+1);
432
433 s_i = cb->args[1];
434 s_num = num = cb->args[2];
435
436 if (cb->args[0] == 0) {
437 if (!(r->tcpdiag_states&(TCPF_LISTEN|TCPF_SYN_RECV)))
438 goto skip_listen_ht;
439 tcp_listen_lock();
440 for (i = s_i; i < TCP_LHTABLE_SIZE; i++) {
441 struct sock *sk = tcp_listening_hash[i];
442
443 if (i > s_i)
444 s_num = 0;
445
446 for (sk = tcp_listening_hash[i], num = 0;
447 sk != NULL;
448 sk = sk->next, num++) {
449 if (num < s_num)
450 continue;
451 if (!(r->tcpdiag_states&TCPF_LISTEN) ||
452 r->id.tcpdiag_dport)
453 continue;
454 if (r->id.tcpdiag_sport != sk->sport && r->id.tcpdiag_sport)
455 continue;
456 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
457 continue;
458 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
459 NETLINK_CB(cb->skb).pid,
460 cb->nlh->nlmsg_seq) <= 0) {
461 tcp_listen_unlock();
462 goto done;
463 }
464 }
465 }
466 tcp_listen_unlock();
467skip_listen_ht:
468 cb->args[0] = 1;
469 s_i = num = s_num = 0;
470 }
471
472 if (!(r->tcpdiag_states&~(TCPF_LISTEN|TCPF_SYN_RECV)))
473 return skb->len;
474
475 for (i = s_i; i < tcp_ehash_size; i++) {
476 struct tcp_ehash_bucket *head = &tcp_ehash[i];
477 struct sock *sk;
478
479 if (i > s_i)
480 s_num = 0;
481
482 read_lock_bh(&head->lock);
483
484 for (sk = head->chain, num = 0;
485 sk != NULL;
486 sk = sk->next, num++) {
487 if (num < s_num)
488 continue;
489 if (!(r->tcpdiag_states&(1<<sk->state)))
490 continue;
491 if (r->id.tcpdiag_sport != sk->sport && r->id.tcpdiag_sport)
492 continue;
493 if (r->id.tcpdiag_dport != sk->dport && r->id.tcpdiag_dport)
494 continue;
495 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
496 continue;
497 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
498 NETLINK_CB(cb->skb).pid,
499 cb->nlh->nlmsg_seq) <= 0) {
500 read_unlock_bh(&head->lock);
501 goto done;
502 }
503 }
504
505 if (r->tcpdiag_states&TCPF_TIME_WAIT) {
506 for (sk = tcp_ehash[i+tcp_ehash_size].chain;
507 sk != NULL;
508 sk = sk->next, num++) {
509 if (num < s_num)
510 continue;
511 if (!(r->tcpdiag_states&(1<<sk->zapped)))
512 continue;
513 if (r->id.tcpdiag_sport != sk->sport && r->id.tcpdiag_sport)
514 continue;
515 if (r->id.tcpdiag_dport != sk->dport && r->id.tcpdiag_dport)
516 continue;
517 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
518 continue;
519 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
520 NETLINK_CB(cb->skb).pid,
521 cb->nlh->nlmsg_seq) <= 0) {
522 read_unlock_bh(&head->lock);
523 goto done;
524 }
525 }
526 }
527 read_unlock_bh(&head->lock);
528 }
529
530done:
531 cb->args[1] = i;
532 cb->args[2] = num;
533 return skb->len;
534}
535
536static int tcpdiag_dump_done(struct netlink_callback *cb)
537{
538 return 0;
539}
540
541
542static __inline__ int
543tcpdiag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
544{
545 if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
546 return 0;
547
548 if (nlh->nlmsg_type != TCPDIAG_GETSOCK)
549 goto err_inval;
550
551 if (NLMSG_LENGTH(sizeof(struct tcpdiagreq)) > skb->len)
552 goto err_inval;
553
554 if (nlh->nlmsg_flags&NLM_F_DUMP) {
555 if (nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(struct tcpdiagreq))) {
556 struct rtattr *rta = (struct rtattr*)(NLMSG_DATA(nlh) + sizeof(struct tcpdiagreq));
557 if (rta->rta_type != TCPDIAG_REQ_BYTECODE ||
558 rta->rta_len < 8 ||
559 rta->rta_len > nlh->nlmsg_len - NLMSG_SPACE(sizeof(struct tcpdiagreq)))
560 goto err_inval;
561 if (tcpdiag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
562 goto err_inval;
563 }
564 return netlink_dump_start(tcpnl, skb, nlh,
565 tcpdiag_dump,
566 tcpdiag_dump_done);
567 } else {
568 return tcpdiag_get_exact(skb, nlh);
569 }
570
571err_inval:
572 return -EINVAL;
573}
574
575
576extern __inline__ void tcpdiag_rcv_skb(struct sk_buff *skb)
577{
578 int err;
579 struct nlmsghdr * nlh;
580
581 if (skb->len >= NLMSG_SPACE(0)) {
582 nlh = (struct nlmsghdr *)skb->data;
583 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
584 return;
585 err = tcpdiag_rcv_msg(skb, nlh);
586 if (err)
587 netlink_ack(skb, nlh, err);
588 }
589}
590
591static void tcpdiag_rcv(struct sock *sk, int len)
592{
593 struct sk_buff *skb;
594
595 while ((skb = skb_dequeue(&sk->receive_queue)) != NULL) {
596 tcpdiag_rcv_skb(skb);
597 kfree_skb(skb);
598 }
599}
600
601static int __init tcpdiag_init(void)
602{
603 tcpnl = netlink_kernel_create(NETLINK_TCPDIAG, tcpdiag_rcv);
604 if (tcpnl == NULL)
605 return -EBUSY;
606 return 0;
607}
608
609static void __exit tcpdiag_exit(void)
610{
611 printk(KERN_INFO "Caution: unloading tcp_diag is not very well supported. Nothing to worry, but yet.\n");
612 if (tcpnl)
613 sock_release(tcpnl->socket);
614}
615
616module_init(tcpdiag_init);
617module_exit(tcpdiag_exit);
618
619/*
620 * Local variables:
621 * compile-command: "gcc -DMOPS -DMODULE -D__KERNEL__ -I../include -I/usr/src/linux/include -Wall -Wstrict-prototypes -O2 -c tcp_diag.c"
622 * End:
623 */
Upstream mirror of iproute2