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 unsigned long last_syn_loss; /* Last Fast Open SYN loss */
33 struct tcp_fastopen_cookie cookie;
36 struct tcp_metrics_block {
37 struct tcp_metrics_block __rcu *tcpm_next;
38 struct inetpeer_addr tcpm_saddr;
39 struct inetpeer_addr tcpm_daddr;
40 unsigned long tcpm_stamp;
44 u32 tcpm_vals[TCP_METRIC_MAX + 1];
45 struct tcp_fastopen_metrics tcpm_fastopen;
47 struct rcu_head rcu_head;
50 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
51 enum tcp_metric_index idx)
53 return tm->tcpm_lock & (1 << idx);
56 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
57 enum tcp_metric_index idx)
59 return tm->tcpm_vals[idx];
62 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
63 enum tcp_metric_index idx)
65 return msecs_to_jiffies(tm->tcpm_vals[idx]);
68 static void tcp_metric_set(struct tcp_metrics_block *tm,
69 enum tcp_metric_index idx,
72 tm->tcpm_vals[idx] = val;
75 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
76 enum tcp_metric_index idx,
79 tm->tcpm_vals[idx] = jiffies_to_msecs(val);
82 static bool addr_same(const struct inetpeer_addr *a,
83 const struct inetpeer_addr *b)
85 const struct in6_addr *a6, *b6;
87 if (a->family != b->family)
89 if (a->family == AF_INET)
90 return a->addr.a4 == b->addr.a4;
92 a6 = (const struct in6_addr *) &a->addr.a6[0];
93 b6 = (const struct in6_addr *) &b->addr.a6[0];
95 return ipv6_addr_equal(a6, b6);
98 struct tcpm_hash_bucket {
99 struct tcp_metrics_block __rcu *chain;
102 static DEFINE_SPINLOCK(tcp_metrics_lock);
104 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst,
109 tm->tcpm_stamp = jiffies;
112 if (dst_metric_locked(dst, RTAX_RTT))
113 val |= 1 << TCP_METRIC_RTT;
114 if (dst_metric_locked(dst, RTAX_RTTVAR))
115 val |= 1 << TCP_METRIC_RTTVAR;
116 if (dst_metric_locked(dst, RTAX_SSTHRESH))
117 val |= 1 << TCP_METRIC_SSTHRESH;
118 if (dst_metric_locked(dst, RTAX_CWND))
119 val |= 1 << TCP_METRIC_CWND;
120 if (dst_metric_locked(dst, RTAX_REORDERING))
121 val |= 1 << TCP_METRIC_REORDERING;
124 tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
125 tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
126 tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
127 tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
128 tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
130 tm->tcpm_ts_stamp = 0;
131 if (fastopen_clear) {
132 tm->tcpm_fastopen.mss = 0;
133 tm->tcpm_fastopen.syn_loss = 0;
134 tm->tcpm_fastopen.cookie.len = 0;
138 #define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
140 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
142 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
143 tcpm_suck_dst(tm, dst, false);
146 #define TCP_METRICS_RECLAIM_DEPTH 5
147 #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
149 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
150 struct inetpeer_addr *saddr,
151 struct inetpeer_addr *daddr,
154 struct tcp_metrics_block *tm;
156 bool reclaim = false;
158 spin_lock_bh(&tcp_metrics_lock);
159 net = dev_net(dst->dev);
161 /* While waiting for the spin-lock the cache might have been populated
162 * with this entry and so we have to check again.
164 tm = __tcp_get_metrics(saddr, daddr, net, hash);
165 if (tm == TCP_METRICS_RECLAIM_PTR) {
170 tcpm_check_stamp(tm, dst);
174 if (unlikely(reclaim)) {
175 struct tcp_metrics_block *oldest;
177 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
178 for (tm = rcu_dereference(oldest->tcpm_next); tm;
179 tm = rcu_dereference(tm->tcpm_next)) {
180 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
185 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
189 tm->tcpm_saddr = *saddr;
190 tm->tcpm_daddr = *daddr;
192 tcpm_suck_dst(tm, dst, true);
194 if (likely(!reclaim)) {
195 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
196 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
200 spin_unlock_bh(&tcp_metrics_lock);
204 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
208 if (depth > TCP_METRICS_RECLAIM_DEPTH)
209 return TCP_METRICS_RECLAIM_PTR;
213 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
214 const struct inetpeer_addr *daddr,
215 struct net *net, unsigned int hash)
217 struct tcp_metrics_block *tm;
220 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
221 tm = rcu_dereference(tm->tcpm_next)) {
222 if (addr_same(&tm->tcpm_saddr, saddr) &&
223 addr_same(&tm->tcpm_daddr, daddr))
227 return tcp_get_encode(tm, depth);
230 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
231 struct dst_entry *dst)
233 struct tcp_metrics_block *tm;
234 struct inetpeer_addr saddr, daddr;
238 saddr.family = req->rsk_ops->family;
239 daddr.family = req->rsk_ops->family;
240 switch (daddr.family) {
242 saddr.addr.a4 = inet_rsk(req)->ir_loc_addr;
243 daddr.addr.a4 = inet_rsk(req)->ir_rmt_addr;
244 hash = (__force unsigned int) daddr.addr.a4;
246 #if IS_ENABLED(CONFIG_IPV6)
248 *(struct in6_addr *)saddr.addr.a6 = inet_rsk(req)->ir_v6_loc_addr;
249 *(struct in6_addr *)daddr.addr.a6 = inet_rsk(req)->ir_v6_rmt_addr;
250 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
257 net = dev_net(dst->dev);
258 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
260 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
261 tm = rcu_dereference(tm->tcpm_next)) {
262 if (addr_same(&tm->tcpm_saddr, &saddr) &&
263 addr_same(&tm->tcpm_daddr, &daddr))
266 tcpm_check_stamp(tm, dst);
270 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
272 struct tcp_metrics_block *tm;
273 struct inetpeer_addr saddr, daddr;
277 if (tw->tw_family == AF_INET) {
278 saddr.family = AF_INET;
279 saddr.addr.a4 = tw->tw_rcv_saddr;
280 daddr.family = AF_INET;
281 daddr.addr.a4 = tw->tw_daddr;
282 hash = (__force unsigned int) daddr.addr.a4;
284 #if IS_ENABLED(CONFIG_IPV6)
285 else if (tw->tw_family == AF_INET6) {
286 if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) {
287 saddr.family = AF_INET;
288 saddr.addr.a4 = tw->tw_rcv_saddr;
289 daddr.family = AF_INET;
290 daddr.addr.a4 = tw->tw_daddr;
291 hash = (__force unsigned int) daddr.addr.a4;
293 saddr.family = AF_INET6;
294 *(struct in6_addr *)saddr.addr.a6 = tw->tw_v6_rcv_saddr;
295 daddr.family = AF_INET6;
296 *(struct in6_addr *)daddr.addr.a6 = tw->tw_v6_daddr;
297 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
305 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
307 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
308 tm = rcu_dereference(tm->tcpm_next)) {
309 if (addr_same(&tm->tcpm_saddr, &saddr) &&
310 addr_same(&tm->tcpm_daddr, &daddr))
316 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
317 struct dst_entry *dst,
320 struct tcp_metrics_block *tm;
321 struct inetpeer_addr saddr, daddr;
325 if (sk->sk_family == AF_INET) {
326 saddr.family = AF_INET;
327 saddr.addr.a4 = inet_sk(sk)->inet_saddr;
328 daddr.family = AF_INET;
329 daddr.addr.a4 = inet_sk(sk)->inet_daddr;
330 hash = (__force unsigned int) daddr.addr.a4;
332 #if IS_ENABLED(CONFIG_IPV6)
333 else if (sk->sk_family == AF_INET6) {
334 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
335 saddr.family = AF_INET;
336 saddr.addr.a4 = inet_sk(sk)->inet_saddr;
337 daddr.family = AF_INET;
338 daddr.addr.a4 = inet_sk(sk)->inet_daddr;
339 hash = (__force unsigned int) daddr.addr.a4;
341 saddr.family = AF_INET6;
342 *(struct in6_addr *)saddr.addr.a6 = sk->sk_v6_rcv_saddr;
343 daddr.family = AF_INET6;
344 *(struct in6_addr *)daddr.addr.a6 = sk->sk_v6_daddr;
345 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
352 net = dev_net(dst->dev);
353 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
355 tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
356 if (tm == TCP_METRICS_RECLAIM_PTR)
359 tm = tcpm_new(dst, &saddr, &daddr, hash);
361 tcpm_check_stamp(tm, dst);
366 /* Save metrics learned by this TCP session. This function is called
367 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
368 * or goes from LAST-ACK to CLOSE.
370 void tcp_update_metrics(struct sock *sk)
372 const struct inet_connection_sock *icsk = inet_csk(sk);
373 struct dst_entry *dst = __sk_dst_get(sk);
374 struct tcp_sock *tp = tcp_sk(sk);
375 struct tcp_metrics_block *tm;
380 if (sysctl_tcp_nometrics_save || !dst)
383 if (dst->flags & DST_HOST)
387 if (icsk->icsk_backoff || !tp->srtt) {
388 /* This session failed to estimate rtt. Why?
389 * Probably, no packets returned in time. Reset our
392 tm = tcp_get_metrics(sk, dst, false);
393 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
394 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
397 tm = tcp_get_metrics(sk, dst, true);
402 rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
405 /* If newly calculated rtt larger than stored one, store new
406 * one. Otherwise, use EWMA. Remember, rtt overestimation is
407 * always better than underestimation.
409 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
414 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
417 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
423 /* Scale deviation to rttvar fixed point */
428 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
432 var -= (var - m) >> 2;
434 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
437 if (tcp_in_initial_slowstart(tp)) {
438 /* Slow start still did not finish. */
439 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
440 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
441 if (val && (tp->snd_cwnd >> 1) > val)
442 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
445 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
446 val = tcp_metric_get(tm, TCP_METRIC_CWND);
447 if (tp->snd_cwnd > val)
448 tcp_metric_set(tm, TCP_METRIC_CWND,
451 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
452 icsk->icsk_ca_state == TCP_CA_Open) {
453 /* Cong. avoidance phase, cwnd is reliable. */
454 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
455 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
456 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
457 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
458 val = tcp_metric_get(tm, TCP_METRIC_CWND);
459 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
462 /* Else slow start did not finish, cwnd is non-sense,
463 * ssthresh may be also invalid.
465 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
466 val = tcp_metric_get(tm, TCP_METRIC_CWND);
467 tcp_metric_set(tm, TCP_METRIC_CWND,
468 (val + tp->snd_ssthresh) >> 1);
470 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
471 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
472 if (val && tp->snd_ssthresh > val)
473 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
476 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
477 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
478 if (val < tp->reordering &&
479 tp->reordering != sysctl_tcp_reordering)
480 tcp_metric_set(tm, TCP_METRIC_REORDERING,
484 tm->tcpm_stamp = jiffies;
489 /* Initialize metrics on socket. */
491 void tcp_init_metrics(struct sock *sk)
493 struct dst_entry *dst = __sk_dst_get(sk);
494 struct tcp_sock *tp = tcp_sk(sk);
495 struct tcp_metrics_block *tm;
496 u32 val, crtt = 0; /* cached RTT scaled by 8 */
504 tm = tcp_get_metrics(sk, dst, true);
510 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
511 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
513 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
515 tp->snd_ssthresh = val;
516 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
517 tp->snd_ssthresh = tp->snd_cwnd_clamp;
519 /* ssthresh may have been reduced unnecessarily during.
520 * 3WHS. Restore it back to its initial default.
522 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
524 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
525 if (val && tp->reordering != val) {
526 tcp_disable_fack(tp);
527 tcp_disable_early_retrans(tp);
528 tp->reordering = val;
531 crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
534 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
535 * to seed the RTO for later data packets because SYN packets are
536 * small. Use the per-dst cached values to seed the RTO but keep
537 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
538 * Later the RTO will be updated immediately upon obtaining the first
539 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
540 * influences the first RTO but not later RTT estimation.
542 * But if RTT is not available from the SYN (due to retransmits or
543 * syn cookies) or the cache, force a conservative 3secs timeout.
545 * A bit of theory. RTT is time passed after "normal" sized packet
546 * is sent until it is ACKed. In normal circumstances sending small
547 * packets force peer to delay ACKs and calculation is correct too.
548 * The algorithm is adaptive and, provided we follow specs, it
549 * NEVER underestimate RTT. BUT! If peer tries to make some clever
550 * tricks sort of "quick acks" for time long enough to decrease RTT
551 * to low value, and then abruptly stops to do it and starts to delay
552 * ACKs, wait for troubles.
554 if (crtt > tp->srtt) {
555 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
557 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
558 } else if (tp->srtt == 0) {
559 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
560 * 3WHS. This is most likely due to retransmission,
561 * including spurious one. Reset the RTO back to 3secs
562 * from the more aggressive 1sec to avoid more spurious
565 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
566 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
568 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
569 * retransmitted. In light of RFC6298 more aggressive 1sec
570 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
571 * retransmission has occurred.
573 if (tp->total_retrans > 1)
576 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
577 tp->snd_cwnd_stamp = tcp_time_stamp;
580 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
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)
598 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
607 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
609 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
611 struct tcp_metrics_block *tm;
614 tm = tcp_get_metrics(sk, dst, true);
616 struct tcp_sock *tp = tcp_sk(sk);
618 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
619 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
620 tp->rx_opt.ts_recent = tm->tcpm_ts;
625 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
627 /* VJ's idea. Save last timestamp seen from this destination and hold
628 * it at least for normal timewait interval to use for duplicate
629 * segment detection in subsequent connections, before they enter
630 * synchronized state.
632 bool tcp_remember_stamp(struct sock *sk)
634 struct dst_entry *dst = __sk_dst_get(sk);
638 struct tcp_metrics_block *tm;
641 tm = tcp_get_metrics(sk, dst, true);
643 struct tcp_sock *tp = tcp_sk(sk);
645 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
646 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
647 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
648 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
649 tm->tcpm_ts = tp->rx_opt.ts_recent;
658 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
660 struct tcp_metrics_block *tm;
664 tm = __tcp_get_metrics_tw(tw);
666 const struct tcp_timewait_sock *tcptw;
667 struct sock *sk = (struct sock *) tw;
669 tcptw = tcp_twsk(sk);
670 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
671 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
672 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
673 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
674 tm->tcpm_ts = tcptw->tw_ts_recent;
683 static DEFINE_SEQLOCK(fastopen_seqlock);
685 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
686 struct tcp_fastopen_cookie *cookie,
687 int *syn_loss, unsigned long *last_syn_loss)
689 struct tcp_metrics_block *tm;
692 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
694 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
698 seq = read_seqbegin(&fastopen_seqlock);
701 *cookie = tfom->cookie;
702 *syn_loss = tfom->syn_loss;
703 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
704 } while (read_seqretry(&fastopen_seqlock, seq));
709 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
710 struct tcp_fastopen_cookie *cookie, bool syn_lost)
712 struct dst_entry *dst = __sk_dst_get(sk);
713 struct tcp_metrics_block *tm;
718 tm = tcp_get_metrics(sk, dst, true);
720 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
722 write_seqlock_bh(&fastopen_seqlock);
725 if (cookie && cookie->len > 0)
726 tfom->cookie = *cookie;
729 tfom->last_syn_loss = jiffies;
732 write_sequnlock_bh(&fastopen_seqlock);
737 static struct genl_family tcp_metrics_nl_family = {
738 .id = GENL_ID_GENERATE,
740 .name = TCP_METRICS_GENL_NAME,
741 .version = TCP_METRICS_GENL_VERSION,
742 .maxattr = TCP_METRICS_ATTR_MAX,
746 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
747 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
748 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
749 .len = sizeof(struct in6_addr), },
750 /* Following attributes are not received for GET/DEL,
751 * we keep them for reference
754 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
755 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
756 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
757 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
758 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
759 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
760 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
761 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
762 .len = TCP_FASTOPEN_COOKIE_MAX, },
766 /* Add attributes, caller cancels its header on failure */
767 static int tcp_metrics_fill_info(struct sk_buff *msg,
768 struct tcp_metrics_block *tm)
773 switch (tm->tcpm_daddr.family) {
775 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
776 tm->tcpm_daddr.addr.a4) < 0)
777 goto nla_put_failure;
778 if (nla_put_be32(msg, TCP_METRICS_ATTR_SADDR_IPV4,
779 tm->tcpm_saddr.addr.a4) < 0)
780 goto nla_put_failure;
783 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
784 tm->tcpm_daddr.addr.a6) < 0)
785 goto nla_put_failure;
786 if (nla_put(msg, TCP_METRICS_ATTR_SADDR_IPV6, 16,
787 tm->tcpm_saddr.addr.a6) < 0)
788 goto nla_put_failure;
791 return -EAFNOSUPPORT;
794 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
795 jiffies - tm->tcpm_stamp) < 0)
796 goto nla_put_failure;
797 if (tm->tcpm_ts_stamp) {
798 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
799 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
800 goto nla_put_failure;
801 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
803 goto nla_put_failure;
809 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
811 goto nla_put_failure;
812 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
813 if (!tm->tcpm_vals[i])
815 if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
816 goto nla_put_failure;
820 nla_nest_end(msg, nest);
822 nla_nest_cancel(msg, nest);
826 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
830 seq = read_seqbegin(&fastopen_seqlock);
831 tfom_copy[0] = tm->tcpm_fastopen;
832 } while (read_seqretry(&fastopen_seqlock, seq));
836 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
838 goto nla_put_failure;
839 if (tfom->syn_loss &&
840 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
841 tfom->syn_loss) < 0 ||
842 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
843 jiffies - tfom->last_syn_loss) < 0))
844 goto nla_put_failure;
845 if (tfom->cookie.len > 0 &&
846 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
847 tfom->cookie.len, tfom->cookie.val) < 0)
848 goto nla_put_failure;
857 static int tcp_metrics_dump_info(struct sk_buff *skb,
858 struct netlink_callback *cb,
859 struct tcp_metrics_block *tm)
863 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
864 &tcp_metrics_nl_family, NLM_F_MULTI,
865 TCP_METRICS_CMD_GET);
869 if (tcp_metrics_fill_info(skb, tm) < 0)
870 goto nla_put_failure;
872 return genlmsg_end(skb, hdr);
875 genlmsg_cancel(skb, hdr);
879 static int tcp_metrics_nl_dump(struct sk_buff *skb,
880 struct netlink_callback *cb)
882 struct net *net = sock_net(skb->sk);
883 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
884 unsigned int row, s_row = cb->args[0];
885 int s_col = cb->args[1], col = s_col;
887 for (row = s_row; row < max_rows; row++, s_col = 0) {
888 struct tcp_metrics_block *tm;
889 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
892 for (col = 0, tm = rcu_dereference(hb->chain); tm;
893 tm = rcu_dereference(tm->tcpm_next), col++) {
896 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
910 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
911 unsigned int *hash, int optional, int v4, int v6)
917 addr->family = AF_INET;
918 addr->addr.a4 = nla_get_be32(a);
920 *hash = (__force unsigned int) addr->addr.a4;
925 if (nla_len(a) != sizeof(struct in6_addr))
927 addr->family = AF_INET6;
928 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
930 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
933 return optional ? 1 : -EAFNOSUPPORT;
936 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
937 unsigned int *hash, int optional)
939 return __parse_nl_addr(info, addr, hash, optional,
940 TCP_METRICS_ATTR_ADDR_IPV4,
941 TCP_METRICS_ATTR_ADDR_IPV6);
944 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
946 return __parse_nl_addr(info, addr, NULL, 0,
947 TCP_METRICS_ATTR_SADDR_IPV4,
948 TCP_METRICS_ATTR_SADDR_IPV6);
951 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
953 struct tcp_metrics_block *tm;
954 struct inetpeer_addr saddr, daddr;
957 struct net *net = genl_info_net(info);
962 ret = parse_nl_addr(info, &daddr, &hash, 0);
966 ret = parse_nl_saddr(info, &saddr);
970 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
974 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
977 goto nla_put_failure;
979 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
982 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
983 tm = rcu_dereference(tm->tcpm_next)) {
984 if (addr_same(&tm->tcpm_daddr, &daddr) &&
985 (!src || addr_same(&tm->tcpm_saddr, &saddr))) {
986 ret = tcp_metrics_fill_info(msg, tm);
994 genlmsg_end(msg, reply);
995 return genlmsg_reply(msg, info);
1005 #define deref_locked_genl(p) \
1006 rcu_dereference_protected(p, lockdep_genl_is_held() && \
1007 lockdep_is_held(&tcp_metrics_lock))
1009 #define deref_genl(p) rcu_dereference_protected(p, lockdep_genl_is_held())
1011 static int tcp_metrics_flush_all(struct net *net)
1013 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
1014 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
1015 struct tcp_metrics_block *tm;
1018 for (row = 0; row < max_rows; row++, hb++) {
1019 spin_lock_bh(&tcp_metrics_lock);
1020 tm = deref_locked_genl(hb->chain);
1023 spin_unlock_bh(&tcp_metrics_lock);
1025 struct tcp_metrics_block *next;
1027 next = deref_genl(tm->tcpm_next);
1028 kfree_rcu(tm, rcu_head);
1035 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1037 struct tcpm_hash_bucket *hb;
1038 struct tcp_metrics_block *tm;
1039 struct tcp_metrics_block __rcu **pp;
1040 struct inetpeer_addr saddr, daddr;
1042 struct net *net = genl_info_net(info);
1044 bool src = true, found = false;
1046 ret = parse_nl_addr(info, &daddr, &hash, 1);
1050 return tcp_metrics_flush_all(net);
1051 ret = parse_nl_saddr(info, &saddr);
1055 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
1056 hb = net->ipv4.tcp_metrics_hash + hash;
1058 spin_lock_bh(&tcp_metrics_lock);
1059 for (tm = deref_locked_genl(*pp); tm; tm = deref_locked_genl(*pp)) {
1060 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1061 (!src || addr_same(&tm->tcpm_saddr, &saddr))) {
1062 *pp = tm->tcpm_next;
1063 kfree_rcu(tm, rcu_head);
1066 pp = &tm->tcpm_next;
1069 spin_unlock_bh(&tcp_metrics_lock);
1075 static const struct genl_ops tcp_metrics_nl_ops[] = {
1077 .cmd = TCP_METRICS_CMD_GET,
1078 .doit = tcp_metrics_nl_cmd_get,
1079 .dumpit = tcp_metrics_nl_dump,
1080 .policy = tcp_metrics_nl_policy,
1081 .flags = GENL_ADMIN_PERM,
1084 .cmd = TCP_METRICS_CMD_DEL,
1085 .doit = tcp_metrics_nl_cmd_del,
1086 .policy = tcp_metrics_nl_policy,
1087 .flags = GENL_ADMIN_PERM,
1091 static unsigned int tcpmhash_entries;
1092 static int __init set_tcpmhash_entries(char *str)
1099 ret = kstrtouint(str, 0, &tcpmhash_entries);
1105 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1107 static int __net_init tcp_net_metrics_init(struct net *net)
1112 slots = tcpmhash_entries;
1114 if (totalram_pages >= 128 * 1024)
1120 net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1121 size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1123 net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1124 if (!net->ipv4.tcp_metrics_hash)
1125 net->ipv4.tcp_metrics_hash = vzalloc(size);
1127 if (!net->ipv4.tcp_metrics_hash)
1133 static void __net_exit tcp_net_metrics_exit(struct net *net)
1137 for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1138 struct tcp_metrics_block *tm, *next;
1140 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1142 next = rcu_dereference_protected(tm->tcpm_next, 1);
1147 if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1148 vfree(net->ipv4.tcp_metrics_hash);
1150 kfree(net->ipv4.tcp_metrics_hash);
1153 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1154 .init = tcp_net_metrics_init,
1155 .exit = tcp_net_metrics_exit,
1158 void __init tcp_metrics_init(void)
1162 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1165 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1166 tcp_metrics_nl_ops);
1168 goto cleanup_subsys;
1172 unregister_pernet_subsys(&tcp_net_metrics_ops);