2 * tcp_diag.c Module for monitoring TCP sockets.
4 * Version: $Id: tcp_diag.c,v 1.3 2002/02/01 22:01:04 davem Exp $
6 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
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.
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>
25 #include <net/inet_common.h>
27 #include <linux/inet.h>
28 #include <linux/stddef.h>
30 #include <linux/tcp_diag.h>
32 static struct sock *tcpnl;
35 #define TCPDIAG_PUT(skb, attrtype, attrlen) \
36 ({ int rtalen = RTA_LENGTH(attrlen); \
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; \
44 static int tcpdiag_fill(struct sk_buff *skb, struct sock *sk,
45 int ext, u32 pid, u32 seq)
47 struct inet_opt *inet = inet_sk(sk);
48 struct tcp_opt *tp = tcp_sk(sk);
51 struct tcp_info *info = NULL;
52 struct tcpdiag_meminfo *minfo = NULL;
53 struct tcpvegas_info *vinfo = NULL;
54 unsigned char *b = skb->tail;
56 nlh = NLMSG_PUT(skb, pid, seq, TCPDIAG_GETSOCK, sizeof(*r));
58 if (sk->sk_state != TCP_TIME_WAIT) {
59 if (ext & (1<<(TCPDIAG_MEMINFO-1)))
60 minfo = TCPDIAG_PUT(skb, TCPDIAG_MEMINFO, sizeof(*minfo));
61 if (ext & (1<<(TCPDIAG_INFO-1)))
62 info = TCPDIAG_PUT(skb, TCPDIAG_INFO, sizeof(*info));
64 if (tcp_is_vegas(tp) && (ext & (1<<(TCPDIAG_VEGASINFO-1))))
65 vinfo = TCPDIAG_PUT(skb, TCPDIAG_VEGASINFO, sizeof(*vinfo));
67 r->tcpdiag_family = sk->sk_family;
68 r->tcpdiag_state = sk->sk_state;
70 r->tcpdiag_retrans = 0;
72 r->id.tcpdiag_if = sk->sk_bound_dev_if;
73 r->id.tcpdiag_cookie[0] = (u32)(unsigned long)sk;
74 r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
76 if (r->tcpdiag_state == TCP_TIME_WAIT) {
77 struct tcp_tw_bucket *tw = (struct tcp_tw_bucket*)sk;
78 long tmo = tw->tw_ttd - jiffies;
82 r->id.tcpdiag_sport = tw->tw_sport;
83 r->id.tcpdiag_dport = tw->tw_dport;
84 r->id.tcpdiag_src[0] = tw->tw_rcv_saddr;
85 r->id.tcpdiag_dst[0] = tw->tw_daddr;
86 r->tcpdiag_state = tw->tw_substate;
88 r->tcpdiag_expires = (tmo*1000+HZ-1)/HZ;
89 r->tcpdiag_rqueue = 0;
90 r->tcpdiag_wqueue = 0;
94 if (r->tcpdiag_family == AF_INET6) {
95 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
96 &tw->tw_v6_rcv_saddr);
97 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
101 nlh->nlmsg_len = skb->tail - b;
105 r->id.tcpdiag_sport = inet->sport;
106 r->id.tcpdiag_dport = inet->dport;
107 r->id.tcpdiag_src[0] = inet->rcv_saddr;
108 r->id.tcpdiag_dst[0] = inet->daddr;
111 if (r->tcpdiag_family == AF_INET6) {
112 struct ipv6_pinfo *np = inet6_sk(sk);
114 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
116 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
121 #define EXPIRES_IN_MS(tmo) ((tmo-jiffies)*1000+HZ-1)/HZ
123 if (tp->pending == TCP_TIME_RETRANS) {
124 r->tcpdiag_timer = 1;
125 r->tcpdiag_retrans = tp->retransmits;
126 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
127 } else if (tp->pending == TCP_TIME_PROBE0) {
128 r->tcpdiag_timer = 4;
129 r->tcpdiag_retrans = tp->probes_out;
130 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
131 } else if (timer_pending(&sk->sk_timer)) {
132 r->tcpdiag_timer = 2;
133 r->tcpdiag_retrans = tp->probes_out;
134 r->tcpdiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
136 r->tcpdiag_timer = 0;
137 r->tcpdiag_expires = 0;
141 r->tcpdiag_rqueue = tp->rcv_nxt - tp->copied_seq;
142 r->tcpdiag_wqueue = tp->write_seq - tp->snd_una;
143 r->tcpdiag_uid = sock_i_uid(sk);
144 r->tcpdiag_inode = sock_i_ino(sk);
147 minfo->tcpdiag_rmem = atomic_read(&sk->sk_rmem_alloc);
148 minfo->tcpdiag_wmem = sk->sk_wmem_queued;
149 minfo->tcpdiag_fmem = sk->sk_forward_alloc;
150 minfo->tcpdiag_tmem = atomic_read(&sk->sk_wmem_alloc);
154 u32 now = tcp_time_stamp;
156 info->tcpi_state = sk->sk_state;
157 info->tcpi_ca_state = tp->ca_state;
158 info->tcpi_retransmits = tp->retransmits;
159 info->tcpi_probes = tp->probes_out;
160 info->tcpi_backoff = tp->backoff;
161 info->tcpi_options = 0;
163 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
165 info->tcpi_options |= TCPI_OPT_SACK;
167 info->tcpi_options |= TCPI_OPT_WSCALE;
168 info->tcpi_snd_wscale = tp->snd_wscale;
169 info->tcpi_rcv_wscale = tp->rcv_wscale;
171 info->tcpi_snd_wscale = 0;
172 info->tcpi_rcv_wscale = 0;
174 if (tp->ecn_flags&TCP_ECN_OK)
175 info->tcpi_options |= TCPI_OPT_ECN;
177 info->tcpi_rto = (1000000*tp->rto)/HZ;
178 info->tcpi_ato = (1000000*tp->ack.ato)/HZ;
179 info->tcpi_snd_mss = tp->mss_cache;
180 info->tcpi_rcv_mss = tp->ack.rcv_mss;
182 info->tcpi_unacked = tp->packets_out;
183 info->tcpi_sacked = tp->sacked_out;
184 info->tcpi_lost = tp->lost_out;
185 info->tcpi_retrans = tp->retrans_out;
186 info->tcpi_fackets = tp->fackets_out;
188 info->tcpi_last_data_sent = ((now - tp->lsndtime)*1000)/HZ;
189 info->tcpi_last_ack_sent = 0;
190 info->tcpi_last_data_recv = ((now - tp->ack.lrcvtime)*1000)/HZ;
191 info->tcpi_last_ack_recv = ((now - tp->rcv_tstamp)*1000)/HZ;
193 info->tcpi_pmtu = tp->pmtu_cookie;
194 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
195 info->tcpi_rtt = ((1000000*tp->srtt)/HZ)>>3;
196 info->tcpi_rttvar = ((1000000*tp->mdev)/HZ)>>2;
197 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
198 info->tcpi_snd_cwnd = tp->snd_cwnd;
199 info->tcpi_advmss = tp->advmss;
200 info->tcpi_reordering = tp->reordering;
204 vinfo->tcpv_enabled = tp->vegas.doing_vegas_now;
205 vinfo->tcpv_rttcnt = tp->vegas.cntRTT;
206 vinfo->tcpv_rtt = tp->vegas.baseRTT;
207 vinfo->tcpv_minrtt = tp->vegas.minRTT;
210 nlh->nlmsg_len = skb->tail - b;
214 skb_trim(skb, b - skb->data);
218 extern struct sock *tcp_v4_lookup(u32 saddr, u16 sport, u32 daddr, u16 dport, int dif);
220 extern struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
221 struct in6_addr *daddr, u16 dport,
225 static int tcpdiag_get_exact(struct sk_buff *in_skb, const struct nlmsghdr *nlh)
229 struct tcpdiagreq *req = NLMSG_DATA(nlh);
232 if (req->tcpdiag_family == AF_INET) {
233 sk = tcp_v4_lookup(req->id.tcpdiag_dst[0], req->id.tcpdiag_dport,
234 req->id.tcpdiag_src[0], req->id.tcpdiag_sport,
238 else if (req->tcpdiag_family == AF_INET6) {
239 sk = tcp_v6_lookup((struct in6_addr*)req->id.tcpdiag_dst, req->id.tcpdiag_dport,
240 (struct in6_addr*)req->id.tcpdiag_src, req->id.tcpdiag_sport,
252 if ((req->id.tcpdiag_cookie[0] != TCPDIAG_NOCOOKIE ||
253 req->id.tcpdiag_cookie[1] != TCPDIAG_NOCOOKIE) &&
254 ((u32)(unsigned long)sk != req->id.tcpdiag_cookie[0] ||
255 (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.tcpdiag_cookie[1]))
259 rep = alloc_skb(NLMSG_SPACE(sizeof(struct tcpdiagmsg)+
260 sizeof(struct tcpdiag_meminfo)+
261 sizeof(struct tcp_info)+64), GFP_KERNEL);
265 if (tcpdiag_fill(rep, sk, req->tcpdiag_ext,
266 NETLINK_CB(in_skb).pid,
267 nlh->nlmsg_seq) <= 0)
270 err = netlink_unicast(tcpnl, rep, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
276 if (sk->sk_state == TCP_TIME_WAIT)
277 tcp_tw_put((struct tcp_tw_bucket*)sk);
284 static int bitstring_match(const u32 *a1, const u32 *a2, int bits)
286 int words = bits >> 5;
291 if (memcmp(a1, a2, words << 2))
301 mask = htonl((0xffffffff) << (32 - bits));
303 if ((w1 ^ w2) & mask)
311 static int tcpdiag_bc_run(const void *bc, int len, struct sock *sk)
315 struct inet_opt *inet = inet_sk(sk);
316 const struct tcpdiag_bc_op *op = bc;
324 case TCPDIAG_BC_S_GE:
325 yes = inet->num >= op[1].no;
327 case TCPDIAG_BC_S_LE:
328 yes = inet->num <= op[1].no;
330 case TCPDIAG_BC_D_GE:
331 yes = ntohs(inet->dport) >= op[1].no;
333 case TCPDIAG_BC_D_LE:
334 yes = ntohs(inet->dport) <= op[1].no;
336 case TCPDIAG_BC_AUTO:
337 yes = !(sk->sk_userlocks & SOCK_BINDPORT_LOCK);
339 case TCPDIAG_BC_S_COND:
340 case TCPDIAG_BC_D_COND:
342 struct tcpdiag_hostcond *cond = (struct tcpdiag_hostcond*)(op+1);
345 if (cond->port != -1 &&
346 cond->port != (op->code == TCPDIAG_BC_S_COND ?
347 inet->num : ntohs(inet->dport))) {
352 if (cond->prefix_len == 0)
356 if (sk->sk_family == AF_INET6) {
357 struct ipv6_pinfo *np = inet6_sk(sk);
359 if (op->code == TCPDIAG_BC_S_COND)
360 addr = (u32*)&np->rcv_saddr;
362 addr = (u32*)&np->daddr;
366 if (op->code == TCPDIAG_BC_S_COND)
367 addr = &inet->rcv_saddr;
372 if (bitstring_match(addr, cond->addr, cond->prefix_len))
374 if (sk->sk_family == AF_INET6 &&
375 cond->family == AF_INET) {
376 if (addr[0] == 0 && addr[1] == 0 &&
377 addr[2] == htonl(0xffff) &&
378 bitstring_match(addr+3, cond->addr, cond->prefix_len))
397 static int valid_cc(const void *bc, int len, int cc)
400 const struct tcpdiag_bc_op *op = bc;
414 static int tcpdiag_bc_audit(const void *bytecode, int bytecode_len)
416 const unsigned char *bc = bytecode;
417 int len = bytecode_len;
420 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
422 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
424 case TCPDIAG_BC_AUTO:
425 case TCPDIAG_BC_S_COND:
426 case TCPDIAG_BC_D_COND:
427 case TCPDIAG_BC_S_GE:
428 case TCPDIAG_BC_S_LE:
429 case TCPDIAG_BC_D_GE:
430 case TCPDIAG_BC_D_LE:
431 if (op->yes < 4 || op->yes > len+4)
434 if (op->no < 4 || op->no > len+4)
437 !valid_cc(bytecode, bytecode_len, len-op->no))
441 if (op->yes < 4 || op->yes > len+4)
450 return len == 0 ? 0 : -EINVAL;
454 static int tcpdiag_dump(struct sk_buff *skb, struct netlink_callback *cb)
458 struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
459 struct rtattr *bc = NULL;
461 if (cb->nlh->nlmsg_len > 4+NLMSG_SPACE(sizeof(struct tcpdiagreq)))
462 bc = (struct rtattr*)(r+1);
465 s_num = num = cb->args[2];
467 if (cb->args[0] == 0) {
468 if (!(r->tcpdiag_states&(TCPF_LISTEN|TCPF_SYN_RECV)))
471 for (i = s_i; i < TCP_LHTABLE_SIZE; i++) {
473 struct hlist_node *node;
479 sk_for_each(sk, node, &tcp_listening_hash[i]) {
480 struct inet_opt *inet = inet_sk(sk);
483 if (!(r->tcpdiag_states&TCPF_LISTEN) ||
486 if (r->id.tcpdiag_sport != inet->sport &&
489 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
491 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
492 NETLINK_CB(cb->skb).pid,
493 cb->nlh->nlmsg_seq) <= 0) {
503 s_i = num = s_num = 0;
506 if (!(r->tcpdiag_states&~(TCPF_LISTEN|TCPF_SYN_RECV)))
509 for (i = s_i; i < tcp_ehash_size; i++) {
510 struct tcp_ehash_bucket *head = &tcp_ehash[i];
512 struct hlist_node *node;
517 read_lock_bh(&head->lock);
520 sk_for_each(sk, node, &head->chain) {
521 struct inet_opt *inet = inet_sk(sk);
525 if (!(r->tcpdiag_states & (1 << sk->sk_state)))
527 if (r->id.tcpdiag_sport != inet->sport &&
530 if (r->id.tcpdiag_dport != inet->dport && r->id.tcpdiag_dport)
532 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
534 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
535 NETLINK_CB(cb->skb).pid,
536 cb->nlh->nlmsg_seq) <= 0) {
537 read_unlock_bh(&head->lock);
543 if (r->tcpdiag_states&TCPF_TIME_WAIT) {
544 sk_for_each(sk, node,
545 &tcp_ehash[i + tcp_ehash_size].chain) {
546 struct inet_opt *inet = inet_sk(sk);
550 if (!(r->tcpdiag_states & (1 << sk->sk_zapped)))
552 if (r->id.tcpdiag_sport != inet->sport &&
555 if (r->id.tcpdiag_dport != inet->dport &&
558 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
560 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
561 NETLINK_CB(cb->skb).pid,
562 cb->nlh->nlmsg_seq) <= 0) {
563 read_unlock_bh(&head->lock);
569 read_unlock_bh(&head->lock);
578 static int tcpdiag_dump_done(struct netlink_callback *cb)
584 static __inline__ int
585 tcpdiag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
587 if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
590 if (nlh->nlmsg_type != TCPDIAG_GETSOCK)
593 if (NLMSG_LENGTH(sizeof(struct tcpdiagreq)) > skb->len)
596 if (nlh->nlmsg_flags&NLM_F_DUMP) {
597 if (nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(struct tcpdiagreq))) {
598 struct rtattr *rta = (struct rtattr*)(NLMSG_DATA(nlh) + sizeof(struct tcpdiagreq));
599 if (rta->rta_type != TCPDIAG_REQ_BYTECODE ||
601 rta->rta_len > nlh->nlmsg_len - NLMSG_SPACE(sizeof(struct tcpdiagreq)))
603 if (tcpdiag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
606 return netlink_dump_start(tcpnl, skb, nlh,
610 return tcpdiag_get_exact(skb, nlh);
618 static inline void tcpdiag_rcv_skb(struct sk_buff *skb)
621 struct nlmsghdr * nlh;
623 if (skb->len >= NLMSG_SPACE(0)) {
624 nlh = (struct nlmsghdr *)skb->data;
625 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
627 err = tcpdiag_rcv_msg(skb, nlh);
628 if (err || nlh->nlmsg_flags & NLM_F_ACK)
629 netlink_ack(skb, nlh, err);
633 static void tcpdiag_rcv(struct sock *sk, int len)
637 while ((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
638 tcpdiag_rcv_skb(skb);
643 void __init tcpdiag_init(void)
645 tcpnl = netlink_kernel_create(NETLINK_TCPDIAG, tcpdiag_rcv);
647 panic("tcpdiag_init: Cannot create netlink socket.");