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)->ir_rmt_addr;
219 hash = (__force unsigned int) addr.addr.a4;
221 #if IS_ENABLED(CONFIG_IPV6)
223 *(struct in6_addr *)addr.addr.a6 = inet_rsk(req)->ir_v6_rmt_addr;
224 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
231 net = dev_net(dst->dev);
232 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
234 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
235 tm = rcu_dereference(tm->tcpm_next)) {
236 if (addr_same(&tm->tcpm_addr, &addr))
239 tcpm_check_stamp(tm, dst);
243 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
245 struct tcp_metrics_block *tm;
246 struct inetpeer_addr addr;
250 addr.family = tw->tw_family;
251 switch (addr.family) {
253 addr.addr.a4 = tw->tw_daddr;
254 hash = (__force unsigned int) addr.addr.a4;
256 #if IS_ENABLED(CONFIG_IPV6)
258 *(struct in6_addr *)addr.addr.a6 = tw->tw_v6_daddr;
259 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
267 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
269 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
270 tm = rcu_dereference(tm->tcpm_next)) {
271 if (addr_same(&tm->tcpm_addr, &addr))
277 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
278 struct dst_entry *dst,
281 struct tcp_metrics_block *tm;
282 struct inetpeer_addr addr;
287 addr.family = sk->sk_family;
288 switch (addr.family) {
290 addr.addr.a4 = inet_sk(sk)->inet_daddr;
291 hash = (__force unsigned int) addr.addr.a4;
293 #if IS_ENABLED(CONFIG_IPV6)
295 *(struct in6_addr *)addr.addr.a6 = sk->sk_v6_daddr;
296 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
303 net = dev_net(dst->dev);
304 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
306 tm = __tcp_get_metrics(&addr, net, hash);
308 if (tm == TCP_METRICS_RECLAIM_PTR) {
313 tm = tcpm_new(dst, &addr, hash, reclaim);
315 tcpm_check_stamp(tm, dst);
320 /* Save metrics learned by this TCP session. This function is called
321 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
322 * or goes from LAST-ACK to CLOSE.
324 void tcp_update_metrics(struct sock *sk)
326 const struct inet_connection_sock *icsk = inet_csk(sk);
327 struct dst_entry *dst = __sk_dst_get(sk);
328 struct tcp_sock *tp = tcp_sk(sk);
329 struct tcp_metrics_block *tm;
334 if (sysctl_tcp_nometrics_save || !dst)
337 if (dst->flags & DST_HOST)
341 if (icsk->icsk_backoff || !tp->srtt) {
342 /* This session failed to estimate rtt. Why?
343 * Probably, no packets returned in time. Reset our
346 tm = tcp_get_metrics(sk, dst, false);
347 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
348 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
351 tm = tcp_get_metrics(sk, dst, true);
356 rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
359 /* If newly calculated rtt larger than stored one, store new
360 * one. Otherwise, use EWMA. Remember, rtt overestimation is
361 * always better than underestimation.
363 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
368 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
371 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
377 /* Scale deviation to rttvar fixed point */
382 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
386 var -= (var - m) >> 2;
388 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
391 if (tcp_in_initial_slowstart(tp)) {
392 /* Slow start still did not finish. */
393 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
394 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
395 if (val && (tp->snd_cwnd >> 1) > val)
396 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
399 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
400 val = tcp_metric_get(tm, TCP_METRIC_CWND);
401 if (tp->snd_cwnd > val)
402 tcp_metric_set(tm, TCP_METRIC_CWND,
405 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
406 icsk->icsk_ca_state == TCP_CA_Open) {
407 /* Cong. avoidance phase, cwnd is reliable. */
408 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
409 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
410 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
411 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
412 val = tcp_metric_get(tm, TCP_METRIC_CWND);
413 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
416 /* Else slow start did not finish, cwnd is non-sense,
417 * ssthresh may be also invalid.
419 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
420 val = tcp_metric_get(tm, TCP_METRIC_CWND);
421 tcp_metric_set(tm, TCP_METRIC_CWND,
422 (val + tp->snd_ssthresh) >> 1);
424 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
425 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
426 if (val && tp->snd_ssthresh > val)
427 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
430 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
431 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
432 if (val < tp->reordering &&
433 tp->reordering != sysctl_tcp_reordering)
434 tcp_metric_set(tm, TCP_METRIC_REORDERING,
438 tm->tcpm_stamp = jiffies;
443 /* Initialize metrics on socket. */
445 void tcp_init_metrics(struct sock *sk)
447 struct dst_entry *dst = __sk_dst_get(sk);
448 struct tcp_sock *tp = tcp_sk(sk);
449 struct tcp_metrics_block *tm;
450 u32 val, crtt = 0; /* cached RTT scaled by 8 */
458 tm = tcp_get_metrics(sk, dst, true);
464 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
465 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
467 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
469 tp->snd_ssthresh = val;
470 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
471 tp->snd_ssthresh = tp->snd_cwnd_clamp;
473 /* ssthresh may have been reduced unnecessarily during.
474 * 3WHS. Restore it back to its initial default.
476 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
478 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
479 if (val && tp->reordering != val) {
480 tcp_disable_fack(tp);
481 tcp_disable_early_retrans(tp);
482 tp->reordering = val;
485 crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
488 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
489 * to seed the RTO for later data packets because SYN packets are
490 * small. Use the per-dst cached values to seed the RTO but keep
491 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
492 * Later the RTO will be updated immediately upon obtaining the first
493 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
494 * influences the first RTO but not later RTT estimation.
496 * But if RTT is not available from the SYN (due to retransmits or
497 * syn cookies) or the cache, force a conservative 3secs timeout.
499 * A bit of theory. RTT is time passed after "normal" sized packet
500 * is sent until it is ACKed. In normal circumstances sending small
501 * packets force peer to delay ACKs and calculation is correct too.
502 * The algorithm is adaptive and, provided we follow specs, it
503 * NEVER underestimate RTT. BUT! If peer tries to make some clever
504 * tricks sort of "quick acks" for time long enough to decrease RTT
505 * to low value, and then abruptly stops to do it and starts to delay
506 * ACKs, wait for troubles.
508 if (crtt > tp->srtt) {
509 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
511 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
512 } else if (tp->srtt == 0) {
513 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
514 * 3WHS. This is most likely due to retransmission,
515 * including spurious one. Reset the RTO back to 3secs
516 * from the more aggressive 1sec to avoid more spurious
519 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
520 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
522 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
523 * retransmitted. In light of RFC6298 more aggressive 1sec
524 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
525 * retransmission has occurred.
527 if (tp->total_retrans > 1)
530 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
531 tp->snd_cwnd_stamp = tcp_time_stamp;
534 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
536 struct tcp_metrics_block *tm;
543 tm = __tcp_get_metrics_req(req, dst);
546 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
547 (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
552 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
561 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
563 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
565 struct tcp_metrics_block *tm;
568 tm = tcp_get_metrics(sk, dst, true);
570 struct tcp_sock *tp = tcp_sk(sk);
572 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
573 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
574 tp->rx_opt.ts_recent = tm->tcpm_ts;
579 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
581 /* VJ's idea. Save last timestamp seen from this destination and hold
582 * it at least for normal timewait interval to use for duplicate
583 * segment detection in subsequent connections, before they enter
584 * synchronized state.
586 bool tcp_remember_stamp(struct sock *sk)
588 struct dst_entry *dst = __sk_dst_get(sk);
592 struct tcp_metrics_block *tm;
595 tm = tcp_get_metrics(sk, dst, true);
597 struct tcp_sock *tp = tcp_sk(sk);
599 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
600 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
601 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
602 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
603 tm->tcpm_ts = tp->rx_opt.ts_recent;
612 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
614 struct tcp_metrics_block *tm;
618 tm = __tcp_get_metrics_tw(tw);
620 const struct tcp_timewait_sock *tcptw;
621 struct sock *sk = (struct sock *) tw;
623 tcptw = tcp_twsk(sk);
624 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
625 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
626 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
627 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
628 tm->tcpm_ts = tcptw->tw_ts_recent;
637 static DEFINE_SEQLOCK(fastopen_seqlock);
639 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
640 struct tcp_fastopen_cookie *cookie,
641 int *syn_loss, unsigned long *last_syn_loss)
643 struct tcp_metrics_block *tm;
646 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
648 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
652 seq = read_seqbegin(&fastopen_seqlock);
655 *cookie = tfom->cookie;
656 *syn_loss = tfom->syn_loss;
657 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
658 } while (read_seqretry(&fastopen_seqlock, seq));
663 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
664 struct tcp_fastopen_cookie *cookie, bool syn_lost)
666 struct tcp_metrics_block *tm;
669 tm = tcp_get_metrics(sk, __sk_dst_get(sk), true);
671 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
673 write_seqlock_bh(&fastopen_seqlock);
676 if (cookie && cookie->len > 0)
677 tfom->cookie = *cookie;
680 tfom->last_syn_loss = jiffies;
683 write_sequnlock_bh(&fastopen_seqlock);
688 static struct genl_family tcp_metrics_nl_family = {
689 .id = GENL_ID_GENERATE,
691 .name = TCP_METRICS_GENL_NAME,
692 .version = TCP_METRICS_GENL_VERSION,
693 .maxattr = TCP_METRICS_ATTR_MAX,
697 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
698 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
699 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
700 .len = sizeof(struct in6_addr), },
701 /* Following attributes are not received for GET/DEL,
702 * we keep them for reference
705 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
706 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
707 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
708 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
709 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
710 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
711 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
712 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
713 .len = TCP_FASTOPEN_COOKIE_MAX, },
717 /* Add attributes, caller cancels its header on failure */
718 static int tcp_metrics_fill_info(struct sk_buff *msg,
719 struct tcp_metrics_block *tm)
724 switch (tm->tcpm_addr.family) {
726 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
727 tm->tcpm_addr.addr.a4) < 0)
728 goto nla_put_failure;
731 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
732 tm->tcpm_addr.addr.a6) < 0)
733 goto nla_put_failure;
736 return -EAFNOSUPPORT;
739 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
740 jiffies - tm->tcpm_stamp) < 0)
741 goto nla_put_failure;
742 if (tm->tcpm_ts_stamp) {
743 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
744 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
745 goto nla_put_failure;
746 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
748 goto nla_put_failure;
754 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
756 goto nla_put_failure;
757 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
758 if (!tm->tcpm_vals[i])
760 if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
761 goto nla_put_failure;
765 nla_nest_end(msg, nest);
767 nla_nest_cancel(msg, nest);
771 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
775 seq = read_seqbegin(&fastopen_seqlock);
776 tfom_copy[0] = tm->tcpm_fastopen;
777 } while (read_seqretry(&fastopen_seqlock, seq));
781 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
783 goto nla_put_failure;
784 if (tfom->syn_loss &&
785 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
786 tfom->syn_loss) < 0 ||
787 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
788 jiffies - tfom->last_syn_loss) < 0))
789 goto nla_put_failure;
790 if (tfom->cookie.len > 0 &&
791 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
792 tfom->cookie.len, tfom->cookie.val) < 0)
793 goto nla_put_failure;
802 static int tcp_metrics_dump_info(struct sk_buff *skb,
803 struct netlink_callback *cb,
804 struct tcp_metrics_block *tm)
808 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
809 &tcp_metrics_nl_family, NLM_F_MULTI,
810 TCP_METRICS_CMD_GET);
814 if (tcp_metrics_fill_info(skb, tm) < 0)
815 goto nla_put_failure;
817 return genlmsg_end(skb, hdr);
820 genlmsg_cancel(skb, hdr);
824 static int tcp_metrics_nl_dump(struct sk_buff *skb,
825 struct netlink_callback *cb)
827 struct net *net = sock_net(skb->sk);
828 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
829 unsigned int row, s_row = cb->args[0];
830 int s_col = cb->args[1], col = s_col;
832 for (row = s_row; row < max_rows; row++, s_col = 0) {
833 struct tcp_metrics_block *tm;
834 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
837 for (col = 0, tm = rcu_dereference(hb->chain); tm;
838 tm = rcu_dereference(tm->tcpm_next), col++) {
841 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
855 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
856 unsigned int *hash, int optional)
860 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
862 addr->family = AF_INET;
863 addr->addr.a4 = nla_get_be32(a);
864 *hash = (__force unsigned int) addr->addr.a4;
867 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
869 if (nla_len(a) != sizeof(struct in6_addr))
871 addr->family = AF_INET6;
872 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
873 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
876 return optional ? 1 : -EAFNOSUPPORT;
879 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
881 struct tcp_metrics_block *tm;
882 struct inetpeer_addr addr;
885 struct net *net = genl_info_net(info);
889 ret = parse_nl_addr(info, &addr, &hash, 0);
893 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
897 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
900 goto nla_put_failure;
902 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
905 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
906 tm = rcu_dereference(tm->tcpm_next)) {
907 if (addr_same(&tm->tcpm_addr, &addr)) {
908 ret = tcp_metrics_fill_info(msg, tm);
916 genlmsg_end(msg, reply);
917 return genlmsg_reply(msg, info);
927 #define deref_locked_genl(p) \
928 rcu_dereference_protected(p, lockdep_genl_is_held() && \
929 lockdep_is_held(&tcp_metrics_lock))
931 #define deref_genl(p) rcu_dereference_protected(p, lockdep_genl_is_held())
933 static int tcp_metrics_flush_all(struct net *net)
935 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
936 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
937 struct tcp_metrics_block *tm;
940 for (row = 0; row < max_rows; row++, hb++) {
941 spin_lock_bh(&tcp_metrics_lock);
942 tm = deref_locked_genl(hb->chain);
945 spin_unlock_bh(&tcp_metrics_lock);
947 struct tcp_metrics_block *next;
949 next = deref_genl(tm->tcpm_next);
950 kfree_rcu(tm, rcu_head);
957 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
959 struct tcpm_hash_bucket *hb;
960 struct tcp_metrics_block *tm;
961 struct tcp_metrics_block __rcu **pp;
962 struct inetpeer_addr addr;
964 struct net *net = genl_info_net(info);
967 ret = parse_nl_addr(info, &addr, &hash, 1);
971 return tcp_metrics_flush_all(net);
973 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
974 hb = net->ipv4.tcp_metrics_hash + hash;
976 spin_lock_bh(&tcp_metrics_lock);
977 for (tm = deref_locked_genl(*pp); tm;
978 pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
979 if (addr_same(&tm->tcpm_addr, &addr)) {
984 spin_unlock_bh(&tcp_metrics_lock);
987 kfree_rcu(tm, rcu_head);
991 static struct genl_ops tcp_metrics_nl_ops[] = {
993 .cmd = TCP_METRICS_CMD_GET,
994 .doit = tcp_metrics_nl_cmd_get,
995 .dumpit = tcp_metrics_nl_dump,
996 .policy = tcp_metrics_nl_policy,
997 .flags = GENL_ADMIN_PERM,
1000 .cmd = TCP_METRICS_CMD_DEL,
1001 .doit = tcp_metrics_nl_cmd_del,
1002 .policy = tcp_metrics_nl_policy,
1003 .flags = GENL_ADMIN_PERM,
1007 static unsigned int tcpmhash_entries;
1008 static int __init set_tcpmhash_entries(char *str)
1015 ret = kstrtouint(str, 0, &tcpmhash_entries);
1021 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1023 static int __net_init tcp_net_metrics_init(struct net *net)
1028 slots = tcpmhash_entries;
1030 if (totalram_pages >= 128 * 1024)
1036 net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1037 size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1039 net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1040 if (!net->ipv4.tcp_metrics_hash)
1041 net->ipv4.tcp_metrics_hash = vzalloc(size);
1043 if (!net->ipv4.tcp_metrics_hash)
1049 static void __net_exit tcp_net_metrics_exit(struct net *net)
1053 for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1054 struct tcp_metrics_block *tm, *next;
1056 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1058 next = rcu_dereference_protected(tm->tcpm_next, 1);
1063 if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1064 vfree(net->ipv4.tcp_metrics_hash);
1066 kfree(net->ipv4.tcp_metrics_hash);
1069 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1070 .init = tcp_net_metrics_init,
1071 .exit = tcp_net_metrics_exit,
1074 void __init tcp_metrics_init(void)
1078 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1081 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1083 ARRAY_SIZE(tcp_metrics_nl_ops));
1085 goto cleanup_subsys;
1089 unregister_pernet_subsys(&tcp_net_metrics_ops);