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 *saddr,
26 const struct inetpeer_addr *daddr,
27 struct net *net, unsigned int hash);
29 struct tcp_fastopen_metrics {
31 u16 syn_loss:10, /* Recurring Fast Open SYN losses */
32 try_exp:2; /* Request w/ exp. option (once) */
33 unsigned long last_syn_loss; /* Last Fast Open SYN loss */
34 struct tcp_fastopen_cookie cookie;
37 /* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
38 * Kernel only stores RTT and RTTVAR in usec resolution
40 #define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
42 struct tcp_metrics_block {
43 struct tcp_metrics_block __rcu *tcpm_next;
44 possible_net_t tcpm_net;
45 struct inetpeer_addr tcpm_saddr;
46 struct inetpeer_addr tcpm_daddr;
47 unsigned long tcpm_stamp;
51 u32 tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
52 struct tcp_fastopen_metrics tcpm_fastopen;
54 struct rcu_head rcu_head;
57 static inline struct net *tm_net(struct tcp_metrics_block *tm)
59 return read_pnet(&tm->tcpm_net);
62 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
63 enum tcp_metric_index idx)
65 return tm->tcpm_lock & (1 << idx);
68 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
69 enum tcp_metric_index idx)
71 return tm->tcpm_vals[idx];
74 static void tcp_metric_set(struct tcp_metrics_block *tm,
75 enum tcp_metric_index idx,
78 tm->tcpm_vals[idx] = val;
81 static bool addr_same(const struct inetpeer_addr *a,
82 const struct inetpeer_addr *b)
84 if (a->family != b->family)
86 if (a->family == AF_INET)
87 return a->addr.a4 == b->addr.a4;
88 return ipv6_addr_equal(&a->addr.in6, &b->addr.in6);
91 struct tcpm_hash_bucket {
92 struct tcp_metrics_block __rcu *chain;
95 static struct tcpm_hash_bucket *tcp_metrics_hash __read_mostly;
96 static unsigned int tcp_metrics_hash_log __read_mostly;
98 static DEFINE_SPINLOCK(tcp_metrics_lock);
100 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
101 const 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 msval = dst_metric_raw(dst, RTAX_RTT);
123 tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC;
125 msval = dst_metric_raw(dst, RTAX_RTTVAR);
126 tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC;
127 tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
128 tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
129 tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
131 tm->tcpm_ts_stamp = 0;
132 if (fastopen_clear) {
133 tm->tcpm_fastopen.mss = 0;
134 tm->tcpm_fastopen.syn_loss = 0;
135 tm->tcpm_fastopen.try_exp = 0;
136 tm->tcpm_fastopen.cookie.exp = false;
137 tm->tcpm_fastopen.cookie.len = 0;
141 #define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
143 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
145 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
146 tcpm_suck_dst(tm, dst, false);
149 #define TCP_METRICS_RECLAIM_DEPTH 5
150 #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
152 #define deref_locked(p) \
153 rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
155 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
156 struct inetpeer_addr *saddr,
157 struct inetpeer_addr *daddr,
160 struct tcp_metrics_block *tm;
162 bool reclaim = false;
164 spin_lock_bh(&tcp_metrics_lock);
165 net = dev_net(dst->dev);
167 /* While waiting for the spin-lock the cache might have been populated
168 * with this entry and so we have to check again.
170 tm = __tcp_get_metrics(saddr, daddr, net, hash);
171 if (tm == TCP_METRICS_RECLAIM_PTR) {
176 tcpm_check_stamp(tm, dst);
180 if (unlikely(reclaim)) {
181 struct tcp_metrics_block *oldest;
183 oldest = deref_locked(tcp_metrics_hash[hash].chain);
184 for (tm = deref_locked(oldest->tcpm_next); tm;
185 tm = deref_locked(tm->tcpm_next)) {
186 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
191 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
195 write_pnet(&tm->tcpm_net, net);
196 tm->tcpm_saddr = *saddr;
197 tm->tcpm_daddr = *daddr;
199 tcpm_suck_dst(tm, dst, true);
201 if (likely(!reclaim)) {
202 tm->tcpm_next = tcp_metrics_hash[hash].chain;
203 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
207 spin_unlock_bh(&tcp_metrics_lock);
211 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
215 if (depth > TCP_METRICS_RECLAIM_DEPTH)
216 return TCP_METRICS_RECLAIM_PTR;
220 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
221 const struct inetpeer_addr *daddr,
222 struct net *net, unsigned int hash)
224 struct tcp_metrics_block *tm;
227 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
228 tm = rcu_dereference(tm->tcpm_next)) {
229 if (addr_same(&tm->tcpm_saddr, saddr) &&
230 addr_same(&tm->tcpm_daddr, daddr) &&
231 net_eq(tm_net(tm), net))
235 return tcp_get_encode(tm, depth);
238 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
239 struct dst_entry *dst)
241 struct tcp_metrics_block *tm;
242 struct inetpeer_addr saddr, daddr;
246 saddr.family = req->rsk_ops->family;
247 daddr.family = req->rsk_ops->family;
248 switch (daddr.family) {
250 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
251 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
252 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
254 #if IS_ENABLED(CONFIG_IPV6)
256 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
257 inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
258 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
265 net = dev_net(dst->dev);
266 hash ^= net_hash_mix(net);
267 hash = hash_32(hash, tcp_metrics_hash_log);
269 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
270 tm = rcu_dereference(tm->tcpm_next)) {
271 if (addr_same(&tm->tcpm_saddr, &saddr) &&
272 addr_same(&tm->tcpm_daddr, &daddr) &&
273 net_eq(tm_net(tm), net))
276 tcpm_check_stamp(tm, dst);
280 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
282 struct tcp_metrics_block *tm;
283 struct inetpeer_addr saddr, daddr;
287 if (tw->tw_family == AF_INET) {
288 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
289 inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
290 hash = ipv4_addr_hash(tw->tw_daddr);
292 #if IS_ENABLED(CONFIG_IPV6)
293 else if (tw->tw_family == AF_INET6) {
294 if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) {
295 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
296 inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
297 hash = ipv4_addr_hash(tw->tw_daddr);
299 inetpeer_set_addr_v6(&saddr, &tw->tw_v6_rcv_saddr);
300 inetpeer_set_addr_v6(&daddr, &tw->tw_v6_daddr);
301 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
309 hash ^= net_hash_mix(net);
310 hash = hash_32(hash, tcp_metrics_hash_log);
312 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
313 tm = rcu_dereference(tm->tcpm_next)) {
314 if (addr_same(&tm->tcpm_saddr, &saddr) &&
315 addr_same(&tm->tcpm_daddr, &daddr) &&
316 net_eq(tm_net(tm), net))
322 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
323 struct dst_entry *dst,
326 struct tcp_metrics_block *tm;
327 struct inetpeer_addr saddr, daddr;
331 if (sk->sk_family == AF_INET) {
332 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
333 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
334 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
336 #if IS_ENABLED(CONFIG_IPV6)
337 else if (sk->sk_family == AF_INET6) {
338 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
339 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
340 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
341 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
343 inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
344 inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
345 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
352 net = dev_net(dst->dev);
353 hash ^= net_hash_mix(net);
354 hash = hash_32(hash, tcp_metrics_hash_log);
356 tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
357 if (tm == TCP_METRICS_RECLAIM_PTR)
360 tm = tcpm_new(dst, &saddr, &daddr, hash);
362 tcpm_check_stamp(tm, dst);
367 /* Save metrics learned by this TCP session. This function is called
368 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
369 * or goes from LAST-ACK to CLOSE.
371 void tcp_update_metrics(struct sock *sk)
373 const struct inet_connection_sock *icsk = inet_csk(sk);
374 struct dst_entry *dst = __sk_dst_get(sk);
375 struct tcp_sock *tp = tcp_sk(sk);
376 struct tcp_metrics_block *tm;
381 if (sysctl_tcp_nometrics_save || !dst)
384 if (dst->flags & DST_HOST)
388 if (icsk->icsk_backoff || !tp->srtt_us) {
389 /* This session failed to estimate rtt. Why?
390 * Probably, no packets returned in time. Reset our
393 tm = tcp_get_metrics(sk, dst, false);
394 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
395 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
398 tm = tcp_get_metrics(sk, dst, true);
403 rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
404 m = rtt - tp->srtt_us;
406 /* If newly calculated rtt larger than stored one, store new
407 * one. Otherwise, use EWMA. Remember, rtt overestimation is
408 * always better than underestimation.
410 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
415 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
418 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
424 /* Scale deviation to rttvar fixed point */
429 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
433 var -= (var - m) >> 2;
435 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
438 if (tcp_in_initial_slowstart(tp)) {
439 /* Slow start still did not finish. */
440 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
441 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
442 if (val && (tp->snd_cwnd >> 1) > val)
443 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
446 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
447 val = tcp_metric_get(tm, TCP_METRIC_CWND);
448 if (tp->snd_cwnd > val)
449 tcp_metric_set(tm, TCP_METRIC_CWND,
452 } else if (!tcp_in_slow_start(tp) &&
453 icsk->icsk_ca_state == TCP_CA_Open) {
454 /* Cong. avoidance phase, cwnd is reliable. */
455 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
456 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
457 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
458 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
459 val = tcp_metric_get(tm, TCP_METRIC_CWND);
460 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
463 /* Else slow start did not finish, cwnd is non-sense,
464 * ssthresh may be also invalid.
466 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
467 val = tcp_metric_get(tm, TCP_METRIC_CWND);
468 tcp_metric_set(tm, TCP_METRIC_CWND,
469 (val + tp->snd_ssthresh) >> 1);
471 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
472 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
473 if (val && tp->snd_ssthresh > val)
474 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
477 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
478 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
479 if (val < tp->reordering &&
480 tp->reordering != sysctl_tcp_reordering)
481 tcp_metric_set(tm, TCP_METRIC_REORDERING,
485 tm->tcpm_stamp = jiffies;
490 /* Initialize metrics on socket. */
492 void tcp_init_metrics(struct sock *sk)
494 struct dst_entry *dst = __sk_dst_get(sk);
495 struct tcp_sock *tp = tcp_sk(sk);
496 struct tcp_metrics_block *tm;
497 u32 val, crtt = 0; /* cached RTT scaled by 8 */
505 tm = tcp_get_metrics(sk, dst, true);
511 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
512 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
514 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
516 tp->snd_ssthresh = val;
517 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
518 tp->snd_ssthresh = tp->snd_cwnd_clamp;
520 /* ssthresh may have been reduced unnecessarily during.
521 * 3WHS. Restore it back to its initial default.
523 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
525 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
526 if (val && tp->reordering != val) {
527 tcp_disable_fack(tp);
528 tcp_disable_early_retrans(tp);
529 tp->reordering = val;
532 crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
535 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
536 * to seed the RTO for later data packets because SYN packets are
537 * small. Use the per-dst cached values to seed the RTO but keep
538 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
539 * Later the RTO will be updated immediately upon obtaining the first
540 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
541 * influences the first RTO but not later RTT estimation.
543 * But if RTT is not available from the SYN (due to retransmits or
544 * syn cookies) or the cache, force a conservative 3secs timeout.
546 * A bit of theory. RTT is time passed after "normal" sized packet
547 * is sent until it is ACKed. In normal circumstances sending small
548 * packets force peer to delay ACKs and calculation is correct too.
549 * The algorithm is adaptive and, provided we follow specs, it
550 * NEVER underestimate RTT. BUT! If peer tries to make some clever
551 * tricks sort of "quick acks" for time long enough to decrease RTT
552 * to low value, and then abruptly stops to do it and starts to delay
553 * ACKs, wait for troubles.
555 if (crtt > tp->srtt_us) {
556 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
557 crtt /= 8 * USEC_PER_MSEC;
558 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
559 } else if (tp->srtt_us == 0) {
560 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
561 * 3WHS. This is most likely due to retransmission,
562 * including spurious one. Reset the RTO back to 3secs
563 * from the more aggressive 1sec to avoid more spurious
566 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
567 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
569 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
571 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
572 * retransmitted. In light of RFC6298 more aggressive 1sec
573 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
574 * retransmission has occurred.
576 if (tp->total_retrans > 1)
579 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
580 tp->snd_cwnd_stamp = tcp_time_stamp;
583 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst,
584 bool paws_check, bool timestamps)
586 struct tcp_metrics_block *tm;
593 tm = __tcp_get_metrics_req(req, dst);
596 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
597 ((s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW ||
603 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
612 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
614 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
616 struct tcp_metrics_block *tm;
619 tm = tcp_get_metrics(sk, dst, true);
621 struct tcp_sock *tp = tcp_sk(sk);
623 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
624 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
625 tp->rx_opt.ts_recent = tm->tcpm_ts;
630 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
632 /* VJ's idea. Save last timestamp seen from this destination and hold
633 * it at least for normal timewait interval to use for duplicate
634 * segment detection in subsequent connections, before they enter
635 * synchronized state.
637 bool tcp_remember_stamp(struct sock *sk)
639 struct dst_entry *dst = __sk_dst_get(sk);
643 struct tcp_metrics_block *tm;
646 tm = tcp_get_metrics(sk, dst, true);
648 struct tcp_sock *tp = tcp_sk(sk);
650 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
651 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
652 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
653 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
654 tm->tcpm_ts = tp->rx_opt.ts_recent;
663 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
665 struct tcp_metrics_block *tm;
669 tm = __tcp_get_metrics_tw(tw);
671 const struct tcp_timewait_sock *tcptw;
672 struct sock *sk = (struct sock *) tw;
674 tcptw = tcp_twsk(sk);
675 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
676 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
677 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
678 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
679 tm->tcpm_ts = tcptw->tw_ts_recent;
688 static DEFINE_SEQLOCK(fastopen_seqlock);
690 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
691 struct tcp_fastopen_cookie *cookie,
692 int *syn_loss, unsigned long *last_syn_loss)
694 struct tcp_metrics_block *tm;
697 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
699 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
703 seq = read_seqbegin(&fastopen_seqlock);
706 *cookie = tfom->cookie;
707 if (cookie->len <= 0 && tfom->try_exp == 1)
709 *syn_loss = tfom->syn_loss;
710 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
711 } while (read_seqretry(&fastopen_seqlock, seq));
716 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
717 struct tcp_fastopen_cookie *cookie, bool syn_lost,
720 struct dst_entry *dst = __sk_dst_get(sk);
721 struct tcp_metrics_block *tm;
726 tm = tcp_get_metrics(sk, dst, true);
728 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
730 write_seqlock_bh(&fastopen_seqlock);
733 if (cookie && cookie->len > 0)
734 tfom->cookie = *cookie;
735 else if (try_exp > tfom->try_exp &&
736 tfom->cookie.len <= 0 && !tfom->cookie.exp)
737 tfom->try_exp = try_exp;
740 tfom->last_syn_loss = jiffies;
743 write_sequnlock_bh(&fastopen_seqlock);
748 static struct genl_family tcp_metrics_nl_family = {
749 .id = GENL_ID_GENERATE,
751 .name = TCP_METRICS_GENL_NAME,
752 .version = TCP_METRICS_GENL_VERSION,
753 .maxattr = TCP_METRICS_ATTR_MAX,
757 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
758 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
759 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
760 .len = sizeof(struct in6_addr), },
761 /* Following attributes are not received for GET/DEL,
762 * we keep them for reference
765 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
766 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
767 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
768 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
769 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
770 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
771 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
772 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
773 .len = TCP_FASTOPEN_COOKIE_MAX, },
777 /* Add attributes, caller cancels its header on failure */
778 static int tcp_metrics_fill_info(struct sk_buff *msg,
779 struct tcp_metrics_block *tm)
784 switch (tm->tcpm_daddr.family) {
786 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
787 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
788 goto nla_put_failure;
789 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
790 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
791 goto nla_put_failure;
794 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
795 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
796 goto nla_put_failure;
797 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
798 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
799 goto nla_put_failure;
802 return -EAFNOSUPPORT;
805 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
806 jiffies - tm->tcpm_stamp) < 0)
807 goto nla_put_failure;
808 if (tm->tcpm_ts_stamp) {
809 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
810 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
811 goto nla_put_failure;
812 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
814 goto nla_put_failure;
820 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
822 goto nla_put_failure;
823 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
824 u32 val = tm->tcpm_vals[i];
828 if (i == TCP_METRIC_RTT) {
829 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
831 goto nla_put_failure;
833 val = max(val / 1000, 1U);
835 if (i == TCP_METRIC_RTTVAR) {
836 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
838 goto nla_put_failure;
840 val = max(val / 1000, 1U);
842 if (nla_put_u32(msg, i + 1, val) < 0)
843 goto nla_put_failure;
847 nla_nest_end(msg, nest);
849 nla_nest_cancel(msg, nest);
853 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
857 seq = read_seqbegin(&fastopen_seqlock);
858 tfom_copy[0] = tm->tcpm_fastopen;
859 } while (read_seqretry(&fastopen_seqlock, seq));
863 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
865 goto nla_put_failure;
866 if (tfom->syn_loss &&
867 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
868 tfom->syn_loss) < 0 ||
869 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
870 jiffies - tfom->last_syn_loss) < 0))
871 goto nla_put_failure;
872 if (tfom->cookie.len > 0 &&
873 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
874 tfom->cookie.len, tfom->cookie.val) < 0)
875 goto nla_put_failure;
884 static int tcp_metrics_dump_info(struct sk_buff *skb,
885 struct netlink_callback *cb,
886 struct tcp_metrics_block *tm)
890 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
891 &tcp_metrics_nl_family, NLM_F_MULTI,
892 TCP_METRICS_CMD_GET);
896 if (tcp_metrics_fill_info(skb, tm) < 0)
897 goto nla_put_failure;
899 genlmsg_end(skb, hdr);
903 genlmsg_cancel(skb, hdr);
907 static int tcp_metrics_nl_dump(struct sk_buff *skb,
908 struct netlink_callback *cb)
910 struct net *net = sock_net(skb->sk);
911 unsigned int max_rows = 1U << tcp_metrics_hash_log;
912 unsigned int row, s_row = cb->args[0];
913 int s_col = cb->args[1], col = s_col;
915 for (row = s_row; row < max_rows; row++, s_col = 0) {
916 struct tcp_metrics_block *tm;
917 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
920 for (col = 0, tm = rcu_dereference(hb->chain); tm;
921 tm = rcu_dereference(tm->tcpm_next), col++) {
922 if (!net_eq(tm_net(tm), net))
926 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
940 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
941 unsigned int *hash, int optional, int v4, int v6)
947 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
949 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
956 if (nla_len(a) != sizeof(struct in6_addr))
958 in6 = nla_get_in6_addr(a);
959 inetpeer_set_addr_v6(addr, &in6);
961 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
964 return optional ? 1 : -EAFNOSUPPORT;
967 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
968 unsigned int *hash, int optional)
970 return __parse_nl_addr(info, addr, hash, optional,
971 TCP_METRICS_ATTR_ADDR_IPV4,
972 TCP_METRICS_ATTR_ADDR_IPV6);
975 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
977 return __parse_nl_addr(info, addr, NULL, 0,
978 TCP_METRICS_ATTR_SADDR_IPV4,
979 TCP_METRICS_ATTR_SADDR_IPV6);
982 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
984 struct tcp_metrics_block *tm;
985 struct inetpeer_addr saddr, daddr;
988 struct net *net = genl_info_net(info);
993 ret = parse_nl_addr(info, &daddr, &hash, 0);
997 ret = parse_nl_saddr(info, &saddr);
1001 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1005 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
1006 info->genlhdr->cmd);
1008 goto nla_put_failure;
1010 hash ^= net_hash_mix(net);
1011 hash = hash_32(hash, tcp_metrics_hash_log);
1014 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
1015 tm = rcu_dereference(tm->tcpm_next)) {
1016 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1017 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1018 net_eq(tm_net(tm), net)) {
1019 ret = tcp_metrics_fill_info(msg, tm);
1027 genlmsg_end(msg, reply);
1028 return genlmsg_reply(msg, info);
1038 static void tcp_metrics_flush_all(struct net *net)
1040 unsigned int max_rows = 1U << tcp_metrics_hash_log;
1041 struct tcpm_hash_bucket *hb = tcp_metrics_hash;
1042 struct tcp_metrics_block *tm;
1045 for (row = 0; row < max_rows; row++, hb++) {
1046 struct tcp_metrics_block __rcu **pp;
1047 spin_lock_bh(&tcp_metrics_lock);
1049 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1050 if (net_eq(tm_net(tm), net)) {
1051 *pp = tm->tcpm_next;
1052 kfree_rcu(tm, rcu_head);
1054 pp = &tm->tcpm_next;
1057 spin_unlock_bh(&tcp_metrics_lock);
1061 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1063 struct tcpm_hash_bucket *hb;
1064 struct tcp_metrics_block *tm;
1065 struct tcp_metrics_block __rcu **pp;
1066 struct inetpeer_addr saddr, daddr;
1068 struct net *net = genl_info_net(info);
1070 bool src = true, found = false;
1072 ret = parse_nl_addr(info, &daddr, &hash, 1);
1076 tcp_metrics_flush_all(net);
1079 ret = parse_nl_saddr(info, &saddr);
1083 hash ^= net_hash_mix(net);
1084 hash = hash_32(hash, tcp_metrics_hash_log);
1085 hb = tcp_metrics_hash + hash;
1087 spin_lock_bh(&tcp_metrics_lock);
1088 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1089 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1090 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1091 net_eq(tm_net(tm), net)) {
1092 *pp = tm->tcpm_next;
1093 kfree_rcu(tm, rcu_head);
1096 pp = &tm->tcpm_next;
1099 spin_unlock_bh(&tcp_metrics_lock);
1105 static const struct genl_ops tcp_metrics_nl_ops[] = {
1107 .cmd = TCP_METRICS_CMD_GET,
1108 .doit = tcp_metrics_nl_cmd_get,
1109 .dumpit = tcp_metrics_nl_dump,
1110 .policy = tcp_metrics_nl_policy,
1113 .cmd = TCP_METRICS_CMD_DEL,
1114 .doit = tcp_metrics_nl_cmd_del,
1115 .policy = tcp_metrics_nl_policy,
1116 .flags = GENL_ADMIN_PERM,
1120 static unsigned int tcpmhash_entries;
1121 static int __init set_tcpmhash_entries(char *str)
1128 ret = kstrtouint(str, 0, &tcpmhash_entries);
1134 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1136 static int __net_init tcp_net_metrics_init(struct net *net)
1141 if (!net_eq(net, &init_net))
1144 slots = tcpmhash_entries;
1146 if (totalram_pages >= 128 * 1024)
1152 tcp_metrics_hash_log = order_base_2(slots);
1153 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1155 tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1156 if (!tcp_metrics_hash)
1157 tcp_metrics_hash = vzalloc(size);
1159 if (!tcp_metrics_hash)
1165 static void __net_exit tcp_net_metrics_exit(struct net *net)
1167 tcp_metrics_flush_all(net);
1170 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1171 .init = tcp_net_metrics_init,
1172 .exit = tcp_net_metrics_exit,
1175 void __init tcp_metrics_init(void)
1179 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1181 panic("Could not allocate the tcp_metrics hash table\n");
1183 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1184 tcp_metrics_nl_ops);
1186 panic("Could not register tcp_metrics generic netlink\n");