1 #include <linux/rcupdate.h>
2 #include <linux/spinlock.h>
3 #include <linux/jiffies.h>
4 #include <linux/module.h>
5 #include <linux/cache.h>
6 #include <linux/slab.h>
7 #include <linux/init.h>
9 #include <linux/hash.h>
10 #include <linux/tcp_metrics.h>
11 #include <linux/vmalloc.h>
13 #include <net/inet_connection_sock.h>
14 #include <net/net_namespace.h>
15 #include <net/request_sock.h>
16 #include <net/inetpeer.h>
21 #include <net/genetlink.h>
23 int sysctl_tcp_nometrics_save __read_mostly;
25 struct tcp_fastopen_metrics {
27 u16 syn_loss:10; /* Recurring Fast Open SYN losses */
28 unsigned long last_syn_loss; /* Last Fast Open SYN loss */
29 struct tcp_fastopen_cookie cookie;
32 struct tcp_metrics_block {
33 struct tcp_metrics_block __rcu *tcpm_next;
34 struct inetpeer_addr tcpm_addr;
35 unsigned long tcpm_stamp;
39 u32 tcpm_vals[TCP_METRIC_MAX + 1];
40 struct tcp_fastopen_metrics tcpm_fastopen;
42 struct rcu_head rcu_head;
45 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
46 enum tcp_metric_index idx)
48 return tm->tcpm_lock & (1 << idx);
51 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
52 enum tcp_metric_index idx)
54 return tm->tcpm_vals[idx];
57 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
58 enum tcp_metric_index idx)
60 return msecs_to_jiffies(tm->tcpm_vals[idx]);
63 static void tcp_metric_set(struct tcp_metrics_block *tm,
64 enum tcp_metric_index idx,
67 tm->tcpm_vals[idx] = val;
70 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
71 enum tcp_metric_index idx,
74 tm->tcpm_vals[idx] = jiffies_to_msecs(val);
77 static bool addr_same(const struct inetpeer_addr *a,
78 const struct inetpeer_addr *b)
80 const struct in6_addr *a6, *b6;
82 if (a->family != b->family)
84 if (a->family == AF_INET)
85 return a->addr.a4 == b->addr.a4;
87 a6 = (const struct in6_addr *) &a->addr.a6[0];
88 b6 = (const struct in6_addr *) &b->addr.a6[0];
90 return ipv6_addr_equal(a6, b6);
93 struct tcpm_hash_bucket {
94 struct tcp_metrics_block __rcu *chain;
97 static DEFINE_SPINLOCK(tcp_metrics_lock);
99 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst,
104 tm->tcpm_stamp = jiffies;
107 if (dst_metric_locked(dst, RTAX_RTT))
108 val |= 1 << TCP_METRIC_RTT;
109 if (dst_metric_locked(dst, RTAX_RTTVAR))
110 val |= 1 << TCP_METRIC_RTTVAR;
111 if (dst_metric_locked(dst, RTAX_SSTHRESH))
112 val |= 1 << TCP_METRIC_SSTHRESH;
113 if (dst_metric_locked(dst, RTAX_CWND))
114 val |= 1 << TCP_METRIC_CWND;
115 if (dst_metric_locked(dst, RTAX_REORDERING))
116 val |= 1 << TCP_METRIC_REORDERING;
119 tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
120 tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
121 tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
122 tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
123 tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
125 tm->tcpm_ts_stamp = 0;
126 if (fastopen_clear) {
127 tm->tcpm_fastopen.mss = 0;
128 tm->tcpm_fastopen.syn_loss = 0;
129 tm->tcpm_fastopen.cookie.len = 0;
133 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
134 struct inetpeer_addr *addr,
138 struct tcp_metrics_block *tm;
141 spin_lock_bh(&tcp_metrics_lock);
142 net = dev_net(dst->dev);
143 if (unlikely(reclaim)) {
144 struct tcp_metrics_block *oldest;
146 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
147 for (tm = rcu_dereference(oldest->tcpm_next); tm;
148 tm = rcu_dereference(tm->tcpm_next)) {
149 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
154 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
158 tm->tcpm_addr = *addr;
160 tcpm_suck_dst(tm, dst, true);
162 if (likely(!reclaim)) {
163 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
164 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
168 spin_unlock_bh(&tcp_metrics_lock);
172 #define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
174 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
176 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
177 tcpm_suck_dst(tm, dst, false);
180 #define TCP_METRICS_RECLAIM_DEPTH 5
181 #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
183 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
187 if (depth > TCP_METRICS_RECLAIM_DEPTH)
188 return TCP_METRICS_RECLAIM_PTR;
192 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr,
193 struct net *net, unsigned int hash)
195 struct tcp_metrics_block *tm;
198 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
199 tm = rcu_dereference(tm->tcpm_next)) {
200 if (addr_same(&tm->tcpm_addr, addr))
204 return tcp_get_encode(tm, depth);
207 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
208 struct dst_entry *dst)
210 struct tcp_metrics_block *tm;
211 struct inetpeer_addr addr;
215 addr.family = req->rsk_ops->family;
216 switch (addr.family) {
218 addr.addr.a4 = inet_rsk(req)->rmt_addr;
219 hash = (__force unsigned int) addr.addr.a4;
222 *(struct in6_addr *)addr.addr.a6 = inet6_rsk(req)->rmt_addr;
223 hash = ipv6_addr_hash(&inet6_rsk(req)->rmt_addr);
229 net = dev_net(dst->dev);
230 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
232 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
233 tm = rcu_dereference(tm->tcpm_next)) {
234 if (addr_same(&tm->tcpm_addr, &addr))
237 tcpm_check_stamp(tm, dst);
241 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
243 struct inet6_timewait_sock *tw6;
244 struct tcp_metrics_block *tm;
245 struct inetpeer_addr addr;
249 addr.family = tw->tw_family;
250 switch (addr.family) {
252 addr.addr.a4 = tw->tw_daddr;
253 hash = (__force unsigned int) addr.addr.a4;
256 tw6 = inet6_twsk((struct sock *)tw);
257 *(struct in6_addr *)addr.addr.a6 = tw6->tw_v6_daddr;
258 hash = ipv6_addr_hash(&tw6->tw_v6_daddr);
265 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
267 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
268 tm = rcu_dereference(tm->tcpm_next)) {
269 if (addr_same(&tm->tcpm_addr, &addr))
275 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
276 struct dst_entry *dst,
279 struct tcp_metrics_block *tm;
280 struct inetpeer_addr addr;
285 addr.family = sk->sk_family;
286 switch (addr.family) {
288 addr.addr.a4 = inet_sk(sk)->inet_daddr;
289 hash = (__force unsigned int) addr.addr.a4;
292 *(struct in6_addr *)addr.addr.a6 = inet6_sk(sk)->daddr;
293 hash = ipv6_addr_hash(&inet6_sk(sk)->daddr);
299 net = dev_net(dst->dev);
300 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
302 tm = __tcp_get_metrics(&addr, net, hash);
304 if (tm == TCP_METRICS_RECLAIM_PTR) {
309 tm = tcpm_new(dst, &addr, hash, reclaim);
311 tcpm_check_stamp(tm, dst);
316 /* Save metrics learned by this TCP session. This function is called
317 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
318 * or goes from LAST-ACK to CLOSE.
320 void tcp_update_metrics(struct sock *sk)
322 const struct inet_connection_sock *icsk = inet_csk(sk);
323 struct dst_entry *dst = __sk_dst_get(sk);
324 struct tcp_sock *tp = tcp_sk(sk);
325 struct tcp_metrics_block *tm;
330 if (sysctl_tcp_nometrics_save || !dst)
333 if (dst->flags & DST_HOST)
337 if (icsk->icsk_backoff || !tp->srtt) {
338 /* This session failed to estimate rtt. Why?
339 * Probably, no packets returned in time. Reset our
342 tm = tcp_get_metrics(sk, dst, false);
343 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
344 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
347 tm = tcp_get_metrics(sk, dst, true);
352 rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
355 /* If newly calculated rtt larger than stored one, store new
356 * one. Otherwise, use EWMA. Remember, rtt overestimation is
357 * always better than underestimation.
359 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
364 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
367 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
373 /* Scale deviation to rttvar fixed point */
378 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
382 var -= (var - m) >> 2;
384 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
387 if (tcp_in_initial_slowstart(tp)) {
388 /* Slow start still did not finish. */
389 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
390 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
391 if (val && (tp->snd_cwnd >> 1) > val)
392 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
395 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
396 val = tcp_metric_get(tm, TCP_METRIC_CWND);
397 if (tp->snd_cwnd > val)
398 tcp_metric_set(tm, TCP_METRIC_CWND,
401 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
402 icsk->icsk_ca_state == TCP_CA_Open) {
403 /* Cong. avoidance phase, cwnd is reliable. */
404 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
405 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
406 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
407 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
408 val = tcp_metric_get(tm, TCP_METRIC_CWND);
409 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
412 /* Else slow start did not finish, cwnd is non-sense,
413 * ssthresh may be also invalid.
415 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
416 val = tcp_metric_get(tm, TCP_METRIC_CWND);
417 tcp_metric_set(tm, TCP_METRIC_CWND,
418 (val + tp->snd_ssthresh) >> 1);
420 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
421 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
422 if (val && tp->snd_ssthresh > val)
423 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
426 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
427 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
428 if (val < tp->reordering &&
429 tp->reordering != sysctl_tcp_reordering)
430 tcp_metric_set(tm, TCP_METRIC_REORDERING,
434 tm->tcpm_stamp = jiffies;
439 /* Initialize metrics on socket. */
441 void tcp_init_metrics(struct sock *sk)
443 struct dst_entry *dst = __sk_dst_get(sk);
444 struct tcp_sock *tp = tcp_sk(sk);
445 struct tcp_metrics_block *tm;
446 u32 val, crtt = 0; /* cached RTT scaled by 8 */
454 tm = tcp_get_metrics(sk, dst, true);
460 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
461 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
463 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
465 tp->snd_ssthresh = val;
466 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
467 tp->snd_ssthresh = tp->snd_cwnd_clamp;
469 /* ssthresh may have been reduced unnecessarily during.
470 * 3WHS. Restore it back to its initial default.
472 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
474 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
475 if (val && tp->reordering != val) {
476 tcp_disable_fack(tp);
477 tcp_disable_early_retrans(tp);
478 tp->reordering = val;
481 crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
484 if (crtt > tp->srtt) {
485 /* Initial RTT (tp->srtt) from SYN usually don't measure
486 * serialization delay on low BW links well so RTO may be
487 * under-estimated. Stay conservative and seed RTO with
488 * the RTTs from past data exchanges, using the same seeding
489 * formula in tcp_rtt_estimator().
491 inet_csk(sk)->icsk_rto = crtt + max(crtt >> 2, tcp_rto_min(sk));
492 } else if (tp->srtt == 0) {
493 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
494 * 3WHS. This is most likely due to retransmission,
495 * including spurious one. Reset the RTO back to 3secs
496 * from the more aggressive 1sec to avoid more spurious
499 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
500 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
502 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
503 * retransmitted. In light of RFC6298 more aggressive 1sec
504 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
505 * retransmission has occurred.
507 if (tp->total_retrans > 1)
510 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
511 tp->snd_cwnd_stamp = tcp_time_stamp;
514 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
516 struct tcp_metrics_block *tm;
523 tm = __tcp_get_metrics_req(req, dst);
526 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
527 (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
532 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
541 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
543 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
545 struct tcp_metrics_block *tm;
548 tm = tcp_get_metrics(sk, dst, true);
550 struct tcp_sock *tp = tcp_sk(sk);
552 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
553 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
554 tp->rx_opt.ts_recent = tm->tcpm_ts;
559 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
561 /* VJ's idea. Save last timestamp seen from this destination and hold
562 * it at least for normal timewait interval to use for duplicate
563 * segment detection in subsequent connections, before they enter
564 * synchronized state.
566 bool tcp_remember_stamp(struct sock *sk)
568 struct dst_entry *dst = __sk_dst_get(sk);
572 struct tcp_metrics_block *tm;
575 tm = tcp_get_metrics(sk, dst, true);
577 struct tcp_sock *tp = tcp_sk(sk);
579 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
580 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
581 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
582 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
583 tm->tcpm_ts = tp->rx_opt.ts_recent;
592 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
594 struct tcp_metrics_block *tm;
598 tm = __tcp_get_metrics_tw(tw);
600 const struct tcp_timewait_sock *tcptw;
601 struct sock *sk = (struct sock *) tw;
603 tcptw = tcp_twsk(sk);
604 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
605 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
606 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
607 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
608 tm->tcpm_ts = tcptw->tw_ts_recent;
617 static DEFINE_SEQLOCK(fastopen_seqlock);
619 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
620 struct tcp_fastopen_cookie *cookie,
621 int *syn_loss, unsigned long *last_syn_loss)
623 struct tcp_metrics_block *tm;
626 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
628 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
632 seq = read_seqbegin(&fastopen_seqlock);
635 *cookie = tfom->cookie;
636 *syn_loss = tfom->syn_loss;
637 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
638 } while (read_seqretry(&fastopen_seqlock, seq));
643 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
644 struct tcp_fastopen_cookie *cookie, bool syn_lost)
646 struct tcp_metrics_block *tm;
649 tm = tcp_get_metrics(sk, __sk_dst_get(sk), true);
651 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
653 write_seqlock_bh(&fastopen_seqlock);
656 tfom->cookie = *cookie;
659 tfom->last_syn_loss = jiffies;
662 write_sequnlock_bh(&fastopen_seqlock);
667 static struct genl_family tcp_metrics_nl_family = {
668 .id = GENL_ID_GENERATE,
670 .name = TCP_METRICS_GENL_NAME,
671 .version = TCP_METRICS_GENL_VERSION,
672 .maxattr = TCP_METRICS_ATTR_MAX,
676 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
677 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
678 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
679 .len = sizeof(struct in6_addr), },
680 /* Following attributes are not received for GET/DEL,
681 * we keep them for reference
684 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
685 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
686 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
687 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
688 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
689 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
690 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
691 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
692 .len = TCP_FASTOPEN_COOKIE_MAX, },
696 /* Add attributes, caller cancels its header on failure */
697 static int tcp_metrics_fill_info(struct sk_buff *msg,
698 struct tcp_metrics_block *tm)
703 switch (tm->tcpm_addr.family) {
705 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
706 tm->tcpm_addr.addr.a4) < 0)
707 goto nla_put_failure;
710 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
711 tm->tcpm_addr.addr.a6) < 0)
712 goto nla_put_failure;
715 return -EAFNOSUPPORT;
718 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
719 jiffies - tm->tcpm_stamp) < 0)
720 goto nla_put_failure;
721 if (tm->tcpm_ts_stamp) {
722 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
723 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
724 goto nla_put_failure;
725 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
727 goto nla_put_failure;
733 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
735 goto nla_put_failure;
736 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
737 if (!tm->tcpm_vals[i])
739 if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
740 goto nla_put_failure;
744 nla_nest_end(msg, nest);
746 nla_nest_cancel(msg, nest);
750 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
754 seq = read_seqbegin(&fastopen_seqlock);
755 tfom_copy[0] = tm->tcpm_fastopen;
756 } while (read_seqretry(&fastopen_seqlock, seq));
760 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
762 goto nla_put_failure;
763 if (tfom->syn_loss &&
764 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
765 tfom->syn_loss) < 0 ||
766 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
767 jiffies - tfom->last_syn_loss) < 0))
768 goto nla_put_failure;
769 if (tfom->cookie.len > 0 &&
770 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
771 tfom->cookie.len, tfom->cookie.val) < 0)
772 goto nla_put_failure;
781 static int tcp_metrics_dump_info(struct sk_buff *skb,
782 struct netlink_callback *cb,
783 struct tcp_metrics_block *tm)
787 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
788 &tcp_metrics_nl_family, NLM_F_MULTI,
789 TCP_METRICS_CMD_GET);
793 if (tcp_metrics_fill_info(skb, tm) < 0)
794 goto nla_put_failure;
796 return genlmsg_end(skb, hdr);
799 genlmsg_cancel(skb, hdr);
803 static int tcp_metrics_nl_dump(struct sk_buff *skb,
804 struct netlink_callback *cb)
806 struct net *net = sock_net(skb->sk);
807 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
808 unsigned int row, s_row = cb->args[0];
809 int s_col = cb->args[1], col = s_col;
811 for (row = s_row; row < max_rows; row++, s_col = 0) {
812 struct tcp_metrics_block *tm;
813 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
816 for (col = 0, tm = rcu_dereference(hb->chain); tm;
817 tm = rcu_dereference(tm->tcpm_next), col++) {
820 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
834 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
835 unsigned int *hash, int optional)
839 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
841 addr->family = AF_INET;
842 addr->addr.a4 = nla_get_be32(a);
843 *hash = (__force unsigned int) addr->addr.a4;
846 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
848 if (nla_len(a) != sizeof(struct in6_addr))
850 addr->family = AF_INET6;
851 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
852 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
855 return optional ? 1 : -EAFNOSUPPORT;
858 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
860 struct tcp_metrics_block *tm;
861 struct inetpeer_addr addr;
864 struct net *net = genl_info_net(info);
868 ret = parse_nl_addr(info, &addr, &hash, 0);
872 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
876 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
879 goto nla_put_failure;
881 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
884 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
885 tm = rcu_dereference(tm->tcpm_next)) {
886 if (addr_same(&tm->tcpm_addr, &addr)) {
887 ret = tcp_metrics_fill_info(msg, tm);
895 genlmsg_end(msg, reply);
896 return genlmsg_reply(msg, info);
906 #define deref_locked_genl(p) \
907 rcu_dereference_protected(p, lockdep_genl_is_held() && \
908 lockdep_is_held(&tcp_metrics_lock))
910 #define deref_genl(p) rcu_dereference_protected(p, lockdep_genl_is_held())
912 static int tcp_metrics_flush_all(struct net *net)
914 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
915 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
916 struct tcp_metrics_block *tm;
919 for (row = 0; row < max_rows; row++, hb++) {
920 spin_lock_bh(&tcp_metrics_lock);
921 tm = deref_locked_genl(hb->chain);
924 spin_unlock_bh(&tcp_metrics_lock);
926 struct tcp_metrics_block *next;
928 next = deref_genl(tm->tcpm_next);
929 kfree_rcu(tm, rcu_head);
936 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
938 struct tcpm_hash_bucket *hb;
939 struct tcp_metrics_block *tm;
940 struct tcp_metrics_block __rcu **pp;
941 struct inetpeer_addr addr;
943 struct net *net = genl_info_net(info);
946 ret = parse_nl_addr(info, &addr, &hash, 1);
950 return tcp_metrics_flush_all(net);
952 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
953 hb = net->ipv4.tcp_metrics_hash + hash;
955 spin_lock_bh(&tcp_metrics_lock);
956 for (tm = deref_locked_genl(*pp); tm;
957 pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
958 if (addr_same(&tm->tcpm_addr, &addr)) {
963 spin_unlock_bh(&tcp_metrics_lock);
966 kfree_rcu(tm, rcu_head);
970 static struct genl_ops tcp_metrics_nl_ops[] = {
972 .cmd = TCP_METRICS_CMD_GET,
973 .doit = tcp_metrics_nl_cmd_get,
974 .dumpit = tcp_metrics_nl_dump,
975 .policy = tcp_metrics_nl_policy,
976 .flags = GENL_ADMIN_PERM,
979 .cmd = TCP_METRICS_CMD_DEL,
980 .doit = tcp_metrics_nl_cmd_del,
981 .policy = tcp_metrics_nl_policy,
982 .flags = GENL_ADMIN_PERM,
986 static unsigned int tcpmhash_entries;
987 static int __init set_tcpmhash_entries(char *str)
994 ret = kstrtouint(str, 0, &tcpmhash_entries);
1000 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1002 static int __net_init tcp_net_metrics_init(struct net *net)
1007 slots = tcpmhash_entries;
1009 if (totalram_pages >= 128 * 1024)
1015 net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1016 size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1018 net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1019 if (!net->ipv4.tcp_metrics_hash)
1020 net->ipv4.tcp_metrics_hash = vzalloc(size);
1022 if (!net->ipv4.tcp_metrics_hash)
1028 static void __net_exit tcp_net_metrics_exit(struct net *net)
1032 for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1033 struct tcp_metrics_block *tm, *next;
1035 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1037 next = rcu_dereference_protected(tm->tcpm_next, 1);
1042 if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1043 vfree(net->ipv4.tcp_metrics_hash);
1045 kfree(net->ipv4.tcp_metrics_hash);
1048 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1049 .init = tcp_net_metrics_init,
1050 .exit = tcp_net_metrics_exit,
1053 void __init tcp_metrics_init(void)
1057 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1060 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1062 ARRAY_SIZE(tcp_metrics_nl_ops));
1064 goto cleanup_subsys;
1068 unregister_pernet_subsys(&tcp_net_metrics_ops);