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 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr,
26 struct net *net, unsigned int hash);
28 struct tcp_fastopen_metrics {
30 u16 syn_loss:10; /* Recurring Fast Open SYN losses */
31 unsigned long last_syn_loss; /* Last Fast Open SYN loss */
32 struct tcp_fastopen_cookie cookie;
35 struct tcp_metrics_block {
36 struct tcp_metrics_block __rcu *tcpm_next;
37 struct inetpeer_addr tcpm_addr;
38 unsigned long tcpm_stamp;
42 u32 tcpm_vals[TCP_METRIC_MAX + 1];
43 struct tcp_fastopen_metrics tcpm_fastopen;
45 struct rcu_head rcu_head;
48 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
49 enum tcp_metric_index idx)
51 return tm->tcpm_lock & (1 << idx);
54 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
55 enum tcp_metric_index idx)
57 return tm->tcpm_vals[idx];
60 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
61 enum tcp_metric_index idx)
63 return msecs_to_jiffies(tm->tcpm_vals[idx]);
66 static void tcp_metric_set(struct tcp_metrics_block *tm,
67 enum tcp_metric_index idx,
70 tm->tcpm_vals[idx] = val;
73 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
74 enum tcp_metric_index idx,
77 tm->tcpm_vals[idx] = jiffies_to_msecs(val);
80 static bool addr_same(const struct inetpeer_addr *a,
81 const struct inetpeer_addr *b)
83 const struct in6_addr *a6, *b6;
85 if (a->family != b->family)
87 if (a->family == AF_INET)
88 return a->addr.a4 == b->addr.a4;
90 a6 = (const struct in6_addr *) &a->addr.a6[0];
91 b6 = (const struct in6_addr *) &b->addr.a6[0];
93 return ipv6_addr_equal(a6, b6);
96 struct tcpm_hash_bucket {
97 struct tcp_metrics_block __rcu *chain;
100 static DEFINE_SPINLOCK(tcp_metrics_lock);
102 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst,
107 tm->tcpm_stamp = jiffies;
110 if (dst_metric_locked(dst, RTAX_RTT))
111 val |= 1 << TCP_METRIC_RTT;
112 if (dst_metric_locked(dst, RTAX_RTTVAR))
113 val |= 1 << TCP_METRIC_RTTVAR;
114 if (dst_metric_locked(dst, RTAX_SSTHRESH))
115 val |= 1 << TCP_METRIC_SSTHRESH;
116 if (dst_metric_locked(dst, RTAX_CWND))
117 val |= 1 << TCP_METRIC_CWND;
118 if (dst_metric_locked(dst, RTAX_REORDERING))
119 val |= 1 << TCP_METRIC_REORDERING;
122 tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
123 tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
124 tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
125 tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
126 tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
128 tm->tcpm_ts_stamp = 0;
129 if (fastopen_clear) {
130 tm->tcpm_fastopen.mss = 0;
131 tm->tcpm_fastopen.syn_loss = 0;
132 tm->tcpm_fastopen.cookie.len = 0;
136 #define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
138 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
140 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
141 tcpm_suck_dst(tm, dst, false);
144 #define TCP_METRICS_RECLAIM_DEPTH 5
145 #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
147 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
148 struct inetpeer_addr *addr,
151 struct tcp_metrics_block *tm;
153 bool reclaim = false;
155 spin_lock_bh(&tcp_metrics_lock);
156 net = dev_net(dst->dev);
158 /* While waiting for the spin-lock the cache might have been populated
159 * with this entry and so we have to check again.
161 tm = __tcp_get_metrics(addr, net, hash);
162 if (tm == TCP_METRICS_RECLAIM_PTR) {
167 tcpm_check_stamp(tm, dst);
171 if (unlikely(reclaim)) {
172 struct tcp_metrics_block *oldest;
174 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
175 for (tm = rcu_dereference(oldest->tcpm_next); tm;
176 tm = rcu_dereference(tm->tcpm_next)) {
177 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
182 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
186 tm->tcpm_addr = *addr;
188 tcpm_suck_dst(tm, dst, true);
190 if (likely(!reclaim)) {
191 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
192 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
196 spin_unlock_bh(&tcp_metrics_lock);
200 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
204 if (depth > TCP_METRICS_RECLAIM_DEPTH)
205 return TCP_METRICS_RECLAIM_PTR;
209 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr,
210 struct net *net, unsigned int hash)
212 struct tcp_metrics_block *tm;
215 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
216 tm = rcu_dereference(tm->tcpm_next)) {
217 if (addr_same(&tm->tcpm_addr, addr))
221 return tcp_get_encode(tm, depth);
224 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
225 struct dst_entry *dst)
227 struct tcp_metrics_block *tm;
228 struct inetpeer_addr addr;
232 addr.family = req->rsk_ops->family;
233 switch (addr.family) {
235 addr.addr.a4 = inet_rsk(req)->ir_rmt_addr;
236 hash = (__force unsigned int) addr.addr.a4;
238 #if IS_ENABLED(CONFIG_IPV6)
240 *(struct in6_addr *)addr.addr.a6 = inet_rsk(req)->ir_v6_rmt_addr;
241 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
248 net = dev_net(dst->dev);
249 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
251 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
252 tm = rcu_dereference(tm->tcpm_next)) {
253 if (addr_same(&tm->tcpm_addr, &addr))
256 tcpm_check_stamp(tm, dst);
260 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
262 struct tcp_metrics_block *tm;
263 struct inetpeer_addr addr;
267 addr.family = tw->tw_family;
268 switch (addr.family) {
270 addr.addr.a4 = tw->tw_daddr;
271 hash = (__force unsigned int) addr.addr.a4;
273 #if IS_ENABLED(CONFIG_IPV6)
275 *(struct in6_addr *)addr.addr.a6 = tw->tw_v6_daddr;
276 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
284 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
286 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
287 tm = rcu_dereference(tm->tcpm_next)) {
288 if (addr_same(&tm->tcpm_addr, &addr))
294 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
295 struct dst_entry *dst,
298 struct tcp_metrics_block *tm;
299 struct inetpeer_addr addr;
303 addr.family = sk->sk_family;
304 switch (addr.family) {
306 addr.addr.a4 = inet_sk(sk)->inet_daddr;
307 hash = (__force unsigned int) addr.addr.a4;
309 #if IS_ENABLED(CONFIG_IPV6)
311 *(struct in6_addr *)addr.addr.a6 = sk->sk_v6_daddr;
312 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
319 net = dev_net(dst->dev);
320 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
322 tm = __tcp_get_metrics(&addr, net, hash);
323 if (tm == TCP_METRICS_RECLAIM_PTR)
326 tm = tcpm_new(dst, &addr, hash);
328 tcpm_check_stamp(tm, dst);
333 /* Save metrics learned by this TCP session. This function is called
334 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
335 * or goes from LAST-ACK to CLOSE.
337 void tcp_update_metrics(struct sock *sk)
339 const struct inet_connection_sock *icsk = inet_csk(sk);
340 struct dst_entry *dst = __sk_dst_get(sk);
341 struct tcp_sock *tp = tcp_sk(sk);
342 struct tcp_metrics_block *tm;
347 if (sysctl_tcp_nometrics_save || !dst)
350 if (dst->flags & DST_HOST)
354 if (icsk->icsk_backoff || !tp->srtt) {
355 /* This session failed to estimate rtt. Why?
356 * Probably, no packets returned in time. Reset our
359 tm = tcp_get_metrics(sk, dst, false);
360 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
361 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
364 tm = tcp_get_metrics(sk, dst, true);
369 rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
372 /* If newly calculated rtt larger than stored one, store new
373 * one. Otherwise, use EWMA. Remember, rtt overestimation is
374 * always better than underestimation.
376 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
381 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
384 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
390 /* Scale deviation to rttvar fixed point */
395 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
399 var -= (var - m) >> 2;
401 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
404 if (tcp_in_initial_slowstart(tp)) {
405 /* Slow start still did not finish. */
406 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
407 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
408 if (val && (tp->snd_cwnd >> 1) > val)
409 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
412 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
413 val = tcp_metric_get(tm, TCP_METRIC_CWND);
414 if (tp->snd_cwnd > val)
415 tcp_metric_set(tm, TCP_METRIC_CWND,
418 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
419 icsk->icsk_ca_state == TCP_CA_Open) {
420 /* Cong. avoidance phase, cwnd is reliable. */
421 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
422 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
423 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
424 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
425 val = tcp_metric_get(tm, TCP_METRIC_CWND);
426 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
429 /* Else slow start did not finish, cwnd is non-sense,
430 * ssthresh may be also invalid.
432 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
433 val = tcp_metric_get(tm, TCP_METRIC_CWND);
434 tcp_metric_set(tm, TCP_METRIC_CWND,
435 (val + tp->snd_ssthresh) >> 1);
437 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
438 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
439 if (val && tp->snd_ssthresh > val)
440 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
443 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
444 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
445 if (val < tp->reordering &&
446 tp->reordering != sysctl_tcp_reordering)
447 tcp_metric_set(tm, TCP_METRIC_REORDERING,
451 tm->tcpm_stamp = jiffies;
456 /* Initialize metrics on socket. */
458 void tcp_init_metrics(struct sock *sk)
460 struct dst_entry *dst = __sk_dst_get(sk);
461 struct tcp_sock *tp = tcp_sk(sk);
462 struct tcp_metrics_block *tm;
463 u32 val, crtt = 0; /* cached RTT scaled by 8 */
471 tm = tcp_get_metrics(sk, dst, true);
477 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
478 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
480 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
482 tp->snd_ssthresh = val;
483 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
484 tp->snd_ssthresh = tp->snd_cwnd_clamp;
486 /* ssthresh may have been reduced unnecessarily during.
487 * 3WHS. Restore it back to its initial default.
489 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
491 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
492 if (val && tp->reordering != val) {
493 tcp_disable_fack(tp);
494 tcp_disable_early_retrans(tp);
495 tp->reordering = val;
498 crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
501 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
502 * to seed the RTO for later data packets because SYN packets are
503 * small. Use the per-dst cached values to seed the RTO but keep
504 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
505 * Later the RTO will be updated immediately upon obtaining the first
506 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
507 * influences the first RTO but not later RTT estimation.
509 * But if RTT is not available from the SYN (due to retransmits or
510 * syn cookies) or the cache, force a conservative 3secs timeout.
512 * A bit of theory. RTT is time passed after "normal" sized packet
513 * is sent until it is ACKed. In normal circumstances sending small
514 * packets force peer to delay ACKs and calculation is correct too.
515 * The algorithm is adaptive and, provided we follow specs, it
516 * NEVER underestimate RTT. BUT! If peer tries to make some clever
517 * tricks sort of "quick acks" for time long enough to decrease RTT
518 * to low value, and then abruptly stops to do it and starts to delay
519 * ACKs, wait for troubles.
521 if (crtt > tp->srtt) {
522 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
524 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
525 } else if (tp->srtt == 0) {
526 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
527 * 3WHS. This is most likely due to retransmission,
528 * including spurious one. Reset the RTO back to 3secs
529 * from the more aggressive 1sec to avoid more spurious
532 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
533 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
535 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
536 * retransmitted. In light of RFC6298 more aggressive 1sec
537 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
538 * retransmission has occurred.
540 if (tp->total_retrans > 1)
543 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
544 tp->snd_cwnd_stamp = tcp_time_stamp;
547 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
549 struct tcp_metrics_block *tm;
556 tm = __tcp_get_metrics_req(req, dst);
559 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
560 (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
565 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
574 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
576 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
578 struct tcp_metrics_block *tm;
581 tm = tcp_get_metrics(sk, dst, true);
583 struct tcp_sock *tp = tcp_sk(sk);
585 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
586 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
587 tp->rx_opt.ts_recent = tm->tcpm_ts;
592 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
594 /* VJ's idea. Save last timestamp seen from this destination and hold
595 * it at least for normal timewait interval to use for duplicate
596 * segment detection in subsequent connections, before they enter
597 * synchronized state.
599 bool tcp_remember_stamp(struct sock *sk)
601 struct dst_entry *dst = __sk_dst_get(sk);
605 struct tcp_metrics_block *tm;
608 tm = tcp_get_metrics(sk, dst, true);
610 struct tcp_sock *tp = tcp_sk(sk);
612 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
613 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
614 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
615 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
616 tm->tcpm_ts = tp->rx_opt.ts_recent;
625 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
627 struct tcp_metrics_block *tm;
631 tm = __tcp_get_metrics_tw(tw);
633 const struct tcp_timewait_sock *tcptw;
634 struct sock *sk = (struct sock *) tw;
636 tcptw = tcp_twsk(sk);
637 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
638 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
639 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
640 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
641 tm->tcpm_ts = tcptw->tw_ts_recent;
650 static DEFINE_SEQLOCK(fastopen_seqlock);
652 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
653 struct tcp_fastopen_cookie *cookie,
654 int *syn_loss, unsigned long *last_syn_loss)
656 struct tcp_metrics_block *tm;
659 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
661 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
665 seq = read_seqbegin(&fastopen_seqlock);
668 *cookie = tfom->cookie;
669 *syn_loss = tfom->syn_loss;
670 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
671 } while (read_seqretry(&fastopen_seqlock, seq));
676 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
677 struct tcp_fastopen_cookie *cookie, bool syn_lost)
679 struct dst_entry *dst = __sk_dst_get(sk);
680 struct tcp_metrics_block *tm;
685 tm = tcp_get_metrics(sk, dst, true);
687 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
689 write_seqlock_bh(&fastopen_seqlock);
692 if (cookie && cookie->len > 0)
693 tfom->cookie = *cookie;
696 tfom->last_syn_loss = jiffies;
699 write_sequnlock_bh(&fastopen_seqlock);
704 static struct genl_family tcp_metrics_nl_family = {
705 .id = GENL_ID_GENERATE,
707 .name = TCP_METRICS_GENL_NAME,
708 .version = TCP_METRICS_GENL_VERSION,
709 .maxattr = TCP_METRICS_ATTR_MAX,
713 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
714 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
715 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
716 .len = sizeof(struct in6_addr), },
717 /* Following attributes are not received for GET/DEL,
718 * we keep them for reference
721 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
722 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
723 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
724 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
725 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
726 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
727 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
728 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
729 .len = TCP_FASTOPEN_COOKIE_MAX, },
733 /* Add attributes, caller cancels its header on failure */
734 static int tcp_metrics_fill_info(struct sk_buff *msg,
735 struct tcp_metrics_block *tm)
740 switch (tm->tcpm_addr.family) {
742 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
743 tm->tcpm_addr.addr.a4) < 0)
744 goto nla_put_failure;
747 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
748 tm->tcpm_addr.addr.a6) < 0)
749 goto nla_put_failure;
752 return -EAFNOSUPPORT;
755 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
756 jiffies - tm->tcpm_stamp) < 0)
757 goto nla_put_failure;
758 if (tm->tcpm_ts_stamp) {
759 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
760 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
761 goto nla_put_failure;
762 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
764 goto nla_put_failure;
770 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
772 goto nla_put_failure;
773 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
774 if (!tm->tcpm_vals[i])
776 if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
777 goto nla_put_failure;
781 nla_nest_end(msg, nest);
783 nla_nest_cancel(msg, nest);
787 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
791 seq = read_seqbegin(&fastopen_seqlock);
792 tfom_copy[0] = tm->tcpm_fastopen;
793 } while (read_seqretry(&fastopen_seqlock, seq));
797 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
799 goto nla_put_failure;
800 if (tfom->syn_loss &&
801 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
802 tfom->syn_loss) < 0 ||
803 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
804 jiffies - tfom->last_syn_loss) < 0))
805 goto nla_put_failure;
806 if (tfom->cookie.len > 0 &&
807 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
808 tfom->cookie.len, tfom->cookie.val) < 0)
809 goto nla_put_failure;
818 static int tcp_metrics_dump_info(struct sk_buff *skb,
819 struct netlink_callback *cb,
820 struct tcp_metrics_block *tm)
824 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
825 &tcp_metrics_nl_family, NLM_F_MULTI,
826 TCP_METRICS_CMD_GET);
830 if (tcp_metrics_fill_info(skb, tm) < 0)
831 goto nla_put_failure;
833 return genlmsg_end(skb, hdr);
836 genlmsg_cancel(skb, hdr);
840 static int tcp_metrics_nl_dump(struct sk_buff *skb,
841 struct netlink_callback *cb)
843 struct net *net = sock_net(skb->sk);
844 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
845 unsigned int row, s_row = cb->args[0];
846 int s_col = cb->args[1], col = s_col;
848 for (row = s_row; row < max_rows; row++, s_col = 0) {
849 struct tcp_metrics_block *tm;
850 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
853 for (col = 0, tm = rcu_dereference(hb->chain); tm;
854 tm = rcu_dereference(tm->tcpm_next), col++) {
857 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
871 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
872 unsigned int *hash, int optional)
876 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
878 addr->family = AF_INET;
879 addr->addr.a4 = nla_get_be32(a);
880 *hash = (__force unsigned int) addr->addr.a4;
883 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
885 if (nla_len(a) != sizeof(struct in6_addr))
887 addr->family = AF_INET6;
888 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
889 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
892 return optional ? 1 : -EAFNOSUPPORT;
895 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
897 struct tcp_metrics_block *tm;
898 struct inetpeer_addr addr;
901 struct net *net = genl_info_net(info);
905 ret = parse_nl_addr(info, &addr, &hash, 0);
909 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
913 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
916 goto nla_put_failure;
918 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
921 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
922 tm = rcu_dereference(tm->tcpm_next)) {
923 if (addr_same(&tm->tcpm_addr, &addr)) {
924 ret = tcp_metrics_fill_info(msg, tm);
932 genlmsg_end(msg, reply);
933 return genlmsg_reply(msg, info);
943 #define deref_locked_genl(p) \
944 rcu_dereference_protected(p, lockdep_genl_is_held() && \
945 lockdep_is_held(&tcp_metrics_lock))
947 #define deref_genl(p) rcu_dereference_protected(p, lockdep_genl_is_held())
949 static int tcp_metrics_flush_all(struct net *net)
951 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
952 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
953 struct tcp_metrics_block *tm;
956 for (row = 0; row < max_rows; row++, hb++) {
957 spin_lock_bh(&tcp_metrics_lock);
958 tm = deref_locked_genl(hb->chain);
961 spin_unlock_bh(&tcp_metrics_lock);
963 struct tcp_metrics_block *next;
965 next = deref_genl(tm->tcpm_next);
966 kfree_rcu(tm, rcu_head);
973 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
975 struct tcpm_hash_bucket *hb;
976 struct tcp_metrics_block *tm;
977 struct tcp_metrics_block __rcu **pp;
978 struct inetpeer_addr addr;
980 struct net *net = genl_info_net(info);
983 ret = parse_nl_addr(info, &addr, &hash, 1);
987 return tcp_metrics_flush_all(net);
989 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
990 hb = net->ipv4.tcp_metrics_hash + hash;
992 spin_lock_bh(&tcp_metrics_lock);
993 for (tm = deref_locked_genl(*pp); tm;
994 pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
995 if (addr_same(&tm->tcpm_addr, &addr)) {
1000 spin_unlock_bh(&tcp_metrics_lock);
1003 kfree_rcu(tm, rcu_head);
1007 static const struct genl_ops tcp_metrics_nl_ops[] = {
1009 .cmd = TCP_METRICS_CMD_GET,
1010 .doit = tcp_metrics_nl_cmd_get,
1011 .dumpit = tcp_metrics_nl_dump,
1012 .policy = tcp_metrics_nl_policy,
1013 .flags = GENL_ADMIN_PERM,
1016 .cmd = TCP_METRICS_CMD_DEL,
1017 .doit = tcp_metrics_nl_cmd_del,
1018 .policy = tcp_metrics_nl_policy,
1019 .flags = GENL_ADMIN_PERM,
1023 static unsigned int tcpmhash_entries;
1024 static int __init set_tcpmhash_entries(char *str)
1031 ret = kstrtouint(str, 0, &tcpmhash_entries);
1037 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1039 static int __net_init tcp_net_metrics_init(struct net *net)
1044 slots = tcpmhash_entries;
1046 if (totalram_pages >= 128 * 1024)
1052 net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1053 size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1055 net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1056 if (!net->ipv4.tcp_metrics_hash)
1057 net->ipv4.tcp_metrics_hash = vzalloc(size);
1059 if (!net->ipv4.tcp_metrics_hash)
1065 static void __net_exit tcp_net_metrics_exit(struct net *net)
1069 for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1070 struct tcp_metrics_block *tm, *next;
1072 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1074 next = rcu_dereference_protected(tm->tcpm_next, 1);
1079 if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1080 vfree(net->ipv4.tcp_metrics_hash);
1082 kfree(net->ipv4.tcp_metrics_hash);
1085 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1086 .init = tcp_net_metrics_init,
1087 .exit = tcp_net_metrics_exit,
1090 void __init tcp_metrics_init(void)
1094 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1097 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1098 tcp_metrics_nl_ops);
1100 goto cleanup_subsys;
1104 unregister_pernet_subsys(&tcp_net_metrics_ops);