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 return inetpeer_addr_cmp(a, b) == 0;
87 struct tcpm_hash_bucket {
88 struct tcp_metrics_block __rcu *chain;
91 static struct tcpm_hash_bucket *tcp_metrics_hash __read_mostly;
92 static unsigned int tcp_metrics_hash_log __read_mostly;
94 static DEFINE_SPINLOCK(tcp_metrics_lock);
96 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
97 const struct dst_entry *dst,
103 tm->tcpm_stamp = jiffies;
106 if (dst_metric_locked(dst, RTAX_RTT))
107 val |= 1 << TCP_METRIC_RTT;
108 if (dst_metric_locked(dst, RTAX_RTTVAR))
109 val |= 1 << TCP_METRIC_RTTVAR;
110 if (dst_metric_locked(dst, RTAX_SSTHRESH))
111 val |= 1 << TCP_METRIC_SSTHRESH;
112 if (dst_metric_locked(dst, RTAX_CWND))
113 val |= 1 << TCP_METRIC_CWND;
114 if (dst_metric_locked(dst, RTAX_REORDERING))
115 val |= 1 << TCP_METRIC_REORDERING;
118 msval = dst_metric_raw(dst, RTAX_RTT);
119 tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC;
121 msval = dst_metric_raw(dst, RTAX_RTTVAR);
122 tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC;
123 tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
124 tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
125 tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
127 tm->tcpm_ts_stamp = 0;
128 if (fastopen_clear) {
129 tm->tcpm_fastopen.mss = 0;
130 tm->tcpm_fastopen.syn_loss = 0;
131 tm->tcpm_fastopen.try_exp = 0;
132 tm->tcpm_fastopen.cookie.exp = false;
133 tm->tcpm_fastopen.cookie.len = 0;
137 #define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
139 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
141 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
142 tcpm_suck_dst(tm, dst, false);
145 #define TCP_METRICS_RECLAIM_DEPTH 5
146 #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
148 #define deref_locked(p) \
149 rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
151 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
152 struct inetpeer_addr *saddr,
153 struct inetpeer_addr *daddr,
156 struct tcp_metrics_block *tm;
158 bool reclaim = false;
160 spin_lock_bh(&tcp_metrics_lock);
161 net = dev_net(dst->dev);
163 /* While waiting for the spin-lock the cache might have been populated
164 * with this entry and so we have to check again.
166 tm = __tcp_get_metrics(saddr, daddr, net, hash);
167 if (tm == TCP_METRICS_RECLAIM_PTR) {
172 tcpm_check_stamp(tm, dst);
176 if (unlikely(reclaim)) {
177 struct tcp_metrics_block *oldest;
179 oldest = deref_locked(tcp_metrics_hash[hash].chain);
180 for (tm = deref_locked(oldest->tcpm_next); tm;
181 tm = deref_locked(tm->tcpm_next)) {
182 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
187 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
191 write_pnet(&tm->tcpm_net, net);
192 tm->tcpm_saddr = *saddr;
193 tm->tcpm_daddr = *daddr;
195 tcpm_suck_dst(tm, dst, true);
197 if (likely(!reclaim)) {
198 tm->tcpm_next = tcp_metrics_hash[hash].chain;
199 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
203 spin_unlock_bh(&tcp_metrics_lock);
207 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
211 if (depth > TCP_METRICS_RECLAIM_DEPTH)
212 return TCP_METRICS_RECLAIM_PTR;
216 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
217 const struct inetpeer_addr *daddr,
218 struct net *net, unsigned int hash)
220 struct tcp_metrics_block *tm;
223 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
224 tm = rcu_dereference(tm->tcpm_next)) {
225 if (addr_same(&tm->tcpm_saddr, saddr) &&
226 addr_same(&tm->tcpm_daddr, daddr) &&
227 net_eq(tm_net(tm), net))
231 return tcp_get_encode(tm, depth);
234 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
235 struct dst_entry *dst)
237 struct tcp_metrics_block *tm;
238 struct inetpeer_addr saddr, daddr;
242 saddr.family = req->rsk_ops->family;
243 daddr.family = req->rsk_ops->family;
244 switch (daddr.family) {
246 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
247 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
248 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
250 #if IS_ENABLED(CONFIG_IPV6)
252 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
253 inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
254 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
261 net = dev_net(dst->dev);
262 hash ^= net_hash_mix(net);
263 hash = hash_32(hash, tcp_metrics_hash_log);
265 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
266 tm = rcu_dereference(tm->tcpm_next)) {
267 if (addr_same(&tm->tcpm_saddr, &saddr) &&
268 addr_same(&tm->tcpm_daddr, &daddr) &&
269 net_eq(tm_net(tm), net))
272 tcpm_check_stamp(tm, dst);
276 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
278 struct tcp_metrics_block *tm;
279 struct inetpeer_addr saddr, daddr;
283 if (tw->tw_family == AF_INET) {
284 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
285 inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
286 hash = ipv4_addr_hash(tw->tw_daddr);
288 #if IS_ENABLED(CONFIG_IPV6)
289 else if (tw->tw_family == AF_INET6) {
290 if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) {
291 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
292 inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
293 hash = ipv4_addr_hash(tw->tw_daddr);
295 inetpeer_set_addr_v6(&saddr, &tw->tw_v6_rcv_saddr);
296 inetpeer_set_addr_v6(&daddr, &tw->tw_v6_daddr);
297 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
305 hash ^= net_hash_mix(net);
306 hash = hash_32(hash, tcp_metrics_hash_log);
308 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
309 tm = rcu_dereference(tm->tcpm_next)) {
310 if (addr_same(&tm->tcpm_saddr, &saddr) &&
311 addr_same(&tm->tcpm_daddr, &daddr) &&
312 net_eq(tm_net(tm), net))
318 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
319 struct dst_entry *dst,
322 struct tcp_metrics_block *tm;
323 struct inetpeer_addr saddr, daddr;
327 if (sk->sk_family == AF_INET) {
328 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
329 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
330 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
332 #if IS_ENABLED(CONFIG_IPV6)
333 else if (sk->sk_family == AF_INET6) {
334 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
335 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
336 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
337 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
339 inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
340 inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
341 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
348 net = dev_net(dst->dev);
349 hash ^= net_hash_mix(net);
350 hash = hash_32(hash, tcp_metrics_hash_log);
352 tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
353 if (tm == TCP_METRICS_RECLAIM_PTR)
356 tm = tcpm_new(dst, &saddr, &daddr, hash);
358 tcpm_check_stamp(tm, dst);
363 /* Save metrics learned by this TCP session. This function is called
364 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
365 * or goes from LAST-ACK to CLOSE.
367 void tcp_update_metrics(struct sock *sk)
369 const struct inet_connection_sock *icsk = inet_csk(sk);
370 struct dst_entry *dst = __sk_dst_get(sk);
371 struct tcp_sock *tp = tcp_sk(sk);
372 struct tcp_metrics_block *tm;
377 if (sysctl_tcp_nometrics_save || !dst)
380 if (dst->flags & DST_HOST)
384 if (icsk->icsk_backoff || !tp->srtt_us) {
385 /* This session failed to estimate rtt. Why?
386 * Probably, no packets returned in time. Reset our
389 tm = tcp_get_metrics(sk, dst, false);
390 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
391 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
394 tm = tcp_get_metrics(sk, dst, true);
399 rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
400 m = rtt - tp->srtt_us;
402 /* If newly calculated rtt larger than stored one, store new
403 * one. Otherwise, use EWMA. Remember, rtt overestimation is
404 * always better than underestimation.
406 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
411 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
414 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
420 /* Scale deviation to rttvar fixed point */
425 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
429 var -= (var - m) >> 2;
431 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
434 if (tcp_in_initial_slowstart(tp)) {
435 /* Slow start still did not finish. */
436 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
437 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
438 if (val && (tp->snd_cwnd >> 1) > val)
439 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
442 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
443 val = tcp_metric_get(tm, TCP_METRIC_CWND);
444 if (tp->snd_cwnd > val)
445 tcp_metric_set(tm, TCP_METRIC_CWND,
448 } else if (!tcp_in_slow_start(tp) &&
449 icsk->icsk_ca_state == TCP_CA_Open) {
450 /* Cong. avoidance phase, cwnd is reliable. */
451 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
452 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
453 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
454 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
455 val = tcp_metric_get(tm, TCP_METRIC_CWND);
456 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
459 /* Else slow start did not finish, cwnd is non-sense,
460 * ssthresh may be also invalid.
462 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
463 val = tcp_metric_get(tm, TCP_METRIC_CWND);
464 tcp_metric_set(tm, TCP_METRIC_CWND,
465 (val + tp->snd_ssthresh) >> 1);
467 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
468 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
469 if (val && tp->snd_ssthresh > val)
470 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
473 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
474 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
475 if (val < tp->reordering &&
476 tp->reordering != sysctl_tcp_reordering)
477 tcp_metric_set(tm, TCP_METRIC_REORDERING,
481 tm->tcpm_stamp = jiffies;
486 /* Initialize metrics on socket. */
488 void tcp_init_metrics(struct sock *sk)
490 struct dst_entry *dst = __sk_dst_get(sk);
491 struct tcp_sock *tp = tcp_sk(sk);
492 struct tcp_metrics_block *tm;
493 u32 val, crtt = 0; /* cached RTT scaled by 8 */
501 tm = tcp_get_metrics(sk, dst, true);
507 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
508 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
510 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
512 tp->snd_ssthresh = val;
513 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
514 tp->snd_ssthresh = tp->snd_cwnd_clamp;
516 /* ssthresh may have been reduced unnecessarily during.
517 * 3WHS. Restore it back to its initial default.
519 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
521 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
522 if (val && tp->reordering != val) {
523 tcp_disable_fack(tp);
524 tcp_disable_early_retrans(tp);
525 tp->reordering = val;
528 crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
531 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
532 * to seed the RTO for later data packets because SYN packets are
533 * small. Use the per-dst cached values to seed the RTO but keep
534 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
535 * Later the RTO will be updated immediately upon obtaining the first
536 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
537 * influences the first RTO but not later RTT estimation.
539 * But if RTT is not available from the SYN (due to retransmits or
540 * syn cookies) or the cache, force a conservative 3secs timeout.
542 * A bit of theory. RTT is time passed after "normal" sized packet
543 * is sent until it is ACKed. In normal circumstances sending small
544 * packets force peer to delay ACKs and calculation is correct too.
545 * The algorithm is adaptive and, provided we follow specs, it
546 * NEVER underestimate RTT. BUT! If peer tries to make some clever
547 * tricks sort of "quick acks" for time long enough to decrease RTT
548 * to low value, and then abruptly stops to do it and starts to delay
549 * ACKs, wait for troubles.
551 if (crtt > tp->srtt_us) {
552 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
553 crtt /= 8 * USEC_PER_MSEC;
554 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
555 } else if (tp->srtt_us == 0) {
556 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
557 * 3WHS. This is most likely due to retransmission,
558 * including spurious one. Reset the RTO back to 3secs
559 * from the more aggressive 1sec to avoid more spurious
562 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
563 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
565 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
567 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
568 * retransmitted. In light of RFC6298 more aggressive 1sec
569 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
570 * retransmission has occurred.
572 if (tp->total_retrans > 1)
575 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
576 tp->snd_cwnd_stamp = tcp_time_stamp;
579 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst,
580 bool paws_check, bool timestamps)
582 struct tcp_metrics_block *tm;
589 tm = __tcp_get_metrics_req(req, dst);
592 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
593 ((s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW ||
599 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
608 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
610 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
612 struct tcp_metrics_block *tm;
615 tm = tcp_get_metrics(sk, dst, true);
617 struct tcp_sock *tp = tcp_sk(sk);
619 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
620 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
621 tp->rx_opt.ts_recent = tm->tcpm_ts;
626 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
628 /* VJ's idea. Save last timestamp seen from this destination and hold
629 * it at least for normal timewait interval to use for duplicate
630 * segment detection in subsequent connections, before they enter
631 * synchronized state.
633 bool tcp_remember_stamp(struct sock *sk)
635 struct dst_entry *dst = __sk_dst_get(sk);
639 struct tcp_metrics_block *tm;
642 tm = tcp_get_metrics(sk, dst, true);
644 struct tcp_sock *tp = tcp_sk(sk);
646 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
647 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
648 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
649 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
650 tm->tcpm_ts = tp->rx_opt.ts_recent;
659 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
661 struct tcp_metrics_block *tm;
665 tm = __tcp_get_metrics_tw(tw);
667 const struct tcp_timewait_sock *tcptw;
668 struct sock *sk = (struct sock *) tw;
670 tcptw = tcp_twsk(sk);
671 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
672 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
673 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
674 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
675 tm->tcpm_ts = tcptw->tw_ts_recent;
684 static DEFINE_SEQLOCK(fastopen_seqlock);
686 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
687 struct tcp_fastopen_cookie *cookie,
688 int *syn_loss, unsigned long *last_syn_loss)
690 struct tcp_metrics_block *tm;
693 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
695 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
699 seq = read_seqbegin(&fastopen_seqlock);
702 *cookie = tfom->cookie;
703 if (cookie->len <= 0 && tfom->try_exp == 1)
705 *syn_loss = tfom->syn_loss;
706 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
707 } while (read_seqretry(&fastopen_seqlock, seq));
712 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
713 struct tcp_fastopen_cookie *cookie, bool syn_lost,
716 struct dst_entry *dst = __sk_dst_get(sk);
717 struct tcp_metrics_block *tm;
722 tm = tcp_get_metrics(sk, dst, true);
724 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
726 write_seqlock_bh(&fastopen_seqlock);
729 if (cookie && cookie->len > 0)
730 tfom->cookie = *cookie;
731 else if (try_exp > tfom->try_exp &&
732 tfom->cookie.len <= 0 && !tfom->cookie.exp)
733 tfom->try_exp = try_exp;
736 tfom->last_syn_loss = jiffies;
739 write_sequnlock_bh(&fastopen_seqlock);
744 static struct genl_family tcp_metrics_nl_family = {
745 .id = GENL_ID_GENERATE,
747 .name = TCP_METRICS_GENL_NAME,
748 .version = TCP_METRICS_GENL_VERSION,
749 .maxattr = TCP_METRICS_ATTR_MAX,
753 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
754 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
755 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
756 .len = sizeof(struct in6_addr), },
757 /* Following attributes are not received for GET/DEL,
758 * we keep them for reference
761 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
762 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
763 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
764 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
765 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
766 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
767 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
768 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
769 .len = TCP_FASTOPEN_COOKIE_MAX, },
773 /* Add attributes, caller cancels its header on failure */
774 static int tcp_metrics_fill_info(struct sk_buff *msg,
775 struct tcp_metrics_block *tm)
780 switch (tm->tcpm_daddr.family) {
782 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
783 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
784 goto nla_put_failure;
785 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
786 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
787 goto nla_put_failure;
790 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
791 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
792 goto nla_put_failure;
793 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
794 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
795 goto nla_put_failure;
798 return -EAFNOSUPPORT;
801 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
802 jiffies - tm->tcpm_stamp) < 0)
803 goto nla_put_failure;
804 if (tm->tcpm_ts_stamp) {
805 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
806 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
807 goto nla_put_failure;
808 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
810 goto nla_put_failure;
816 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
818 goto nla_put_failure;
819 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
820 u32 val = tm->tcpm_vals[i];
824 if (i == TCP_METRIC_RTT) {
825 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
827 goto nla_put_failure;
829 val = max(val / 1000, 1U);
831 if (i == TCP_METRIC_RTTVAR) {
832 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
834 goto nla_put_failure;
836 val = max(val / 1000, 1U);
838 if (nla_put_u32(msg, i + 1, val) < 0)
839 goto nla_put_failure;
843 nla_nest_end(msg, nest);
845 nla_nest_cancel(msg, nest);
849 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
853 seq = read_seqbegin(&fastopen_seqlock);
854 tfom_copy[0] = tm->tcpm_fastopen;
855 } while (read_seqretry(&fastopen_seqlock, seq));
859 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
861 goto nla_put_failure;
862 if (tfom->syn_loss &&
863 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
864 tfom->syn_loss) < 0 ||
865 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
866 jiffies - tfom->last_syn_loss) < 0))
867 goto nla_put_failure;
868 if (tfom->cookie.len > 0 &&
869 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
870 tfom->cookie.len, tfom->cookie.val) < 0)
871 goto nla_put_failure;
880 static int tcp_metrics_dump_info(struct sk_buff *skb,
881 struct netlink_callback *cb,
882 struct tcp_metrics_block *tm)
886 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
887 &tcp_metrics_nl_family, NLM_F_MULTI,
888 TCP_METRICS_CMD_GET);
892 if (tcp_metrics_fill_info(skb, tm) < 0)
893 goto nla_put_failure;
895 genlmsg_end(skb, hdr);
899 genlmsg_cancel(skb, hdr);
903 static int tcp_metrics_nl_dump(struct sk_buff *skb,
904 struct netlink_callback *cb)
906 struct net *net = sock_net(skb->sk);
907 unsigned int max_rows = 1U << tcp_metrics_hash_log;
908 unsigned int row, s_row = cb->args[0];
909 int s_col = cb->args[1], col = s_col;
911 for (row = s_row; row < max_rows; row++, s_col = 0) {
912 struct tcp_metrics_block *tm;
913 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
916 for (col = 0, tm = rcu_dereference(hb->chain); tm;
917 tm = rcu_dereference(tm->tcpm_next), col++) {
918 if (!net_eq(tm_net(tm), net))
922 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
936 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
937 unsigned int *hash, int optional, int v4, int v6)
943 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
945 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
952 if (nla_len(a) != sizeof(struct in6_addr))
954 in6 = nla_get_in6_addr(a);
955 inetpeer_set_addr_v6(addr, &in6);
957 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
960 return optional ? 1 : -EAFNOSUPPORT;
963 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
964 unsigned int *hash, int optional)
966 return __parse_nl_addr(info, addr, hash, optional,
967 TCP_METRICS_ATTR_ADDR_IPV4,
968 TCP_METRICS_ATTR_ADDR_IPV6);
971 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
973 return __parse_nl_addr(info, addr, NULL, 0,
974 TCP_METRICS_ATTR_SADDR_IPV4,
975 TCP_METRICS_ATTR_SADDR_IPV6);
978 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
980 struct tcp_metrics_block *tm;
981 struct inetpeer_addr saddr, daddr;
984 struct net *net = genl_info_net(info);
989 ret = parse_nl_addr(info, &daddr, &hash, 0);
993 ret = parse_nl_saddr(info, &saddr);
997 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1001 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
1002 info->genlhdr->cmd);
1004 goto nla_put_failure;
1006 hash ^= net_hash_mix(net);
1007 hash = hash_32(hash, tcp_metrics_hash_log);
1010 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
1011 tm = rcu_dereference(tm->tcpm_next)) {
1012 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1013 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1014 net_eq(tm_net(tm), net)) {
1015 ret = tcp_metrics_fill_info(msg, tm);
1023 genlmsg_end(msg, reply);
1024 return genlmsg_reply(msg, info);
1034 static void tcp_metrics_flush_all(struct net *net)
1036 unsigned int max_rows = 1U << tcp_metrics_hash_log;
1037 struct tcpm_hash_bucket *hb = tcp_metrics_hash;
1038 struct tcp_metrics_block *tm;
1041 for (row = 0; row < max_rows; row++, hb++) {
1042 struct tcp_metrics_block __rcu **pp;
1043 spin_lock_bh(&tcp_metrics_lock);
1045 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1046 if (net_eq(tm_net(tm), net)) {
1047 *pp = tm->tcpm_next;
1048 kfree_rcu(tm, rcu_head);
1050 pp = &tm->tcpm_next;
1053 spin_unlock_bh(&tcp_metrics_lock);
1057 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1059 struct tcpm_hash_bucket *hb;
1060 struct tcp_metrics_block *tm;
1061 struct tcp_metrics_block __rcu **pp;
1062 struct inetpeer_addr saddr, daddr;
1064 struct net *net = genl_info_net(info);
1066 bool src = true, found = false;
1068 ret = parse_nl_addr(info, &daddr, &hash, 1);
1072 tcp_metrics_flush_all(net);
1075 ret = parse_nl_saddr(info, &saddr);
1079 hash ^= net_hash_mix(net);
1080 hash = hash_32(hash, tcp_metrics_hash_log);
1081 hb = tcp_metrics_hash + hash;
1083 spin_lock_bh(&tcp_metrics_lock);
1084 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1085 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1086 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1087 net_eq(tm_net(tm), net)) {
1088 *pp = tm->tcpm_next;
1089 kfree_rcu(tm, rcu_head);
1092 pp = &tm->tcpm_next;
1095 spin_unlock_bh(&tcp_metrics_lock);
1101 static const struct genl_ops tcp_metrics_nl_ops[] = {
1103 .cmd = TCP_METRICS_CMD_GET,
1104 .doit = tcp_metrics_nl_cmd_get,
1105 .dumpit = tcp_metrics_nl_dump,
1106 .policy = tcp_metrics_nl_policy,
1109 .cmd = TCP_METRICS_CMD_DEL,
1110 .doit = tcp_metrics_nl_cmd_del,
1111 .policy = tcp_metrics_nl_policy,
1112 .flags = GENL_ADMIN_PERM,
1116 static unsigned int tcpmhash_entries;
1117 static int __init set_tcpmhash_entries(char *str)
1124 ret = kstrtouint(str, 0, &tcpmhash_entries);
1130 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1132 static int __net_init tcp_net_metrics_init(struct net *net)
1137 if (!net_eq(net, &init_net))
1140 slots = tcpmhash_entries;
1142 if (totalram_pages >= 128 * 1024)
1148 tcp_metrics_hash_log = order_base_2(slots);
1149 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1151 tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1152 if (!tcp_metrics_hash)
1153 tcp_metrics_hash = vzalloc(size);
1155 if (!tcp_metrics_hash)
1161 static void __net_exit tcp_net_metrics_exit(struct net *net)
1163 tcp_metrics_flush_all(net);
1166 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1167 .init = tcp_net_metrics_init,
1168 .exit = tcp_net_metrics_exit,
1171 void __init tcp_metrics_init(void)
1175 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1177 panic("Could not allocate the tcp_metrics hash table\n");
1179 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1180 tcp_metrics_nl_ops);
1182 panic("Could not register tcp_metrics generic netlink\n");