Commit 51c5d0c4 authored by David S. Miller's avatar David S. Miller

tcp: Maintain dynamic metrics in local cache.

Maintain a local hash table of TCP dynamic metrics blobs.

Computed TCP metrics are no longer maintained in the route metrics.

The table uses RCU and an extremely simple hash so that it has low
latency and low overhead.  A simple hash is legitimate because we only
make metrics blobs for fully established connections.

Some tweaking of the default hash table sizes, metric timeouts, and
the hash chain length limit certainly could use some tweaking.  But
the basic design seems sound.

With help from Eric Dumazet and Joe Perches.
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent ab92bb2f
......@@ -7,6 +7,7 @@
#include <net/inet_frag.h>
struct tcpm_hash_bucket;
struct ctl_table_header;
struct ipv4_devconf;
struct fib_rules_ops;
......@@ -39,6 +40,8 @@ struct netns_ipv4 {
struct sock **icmp_sk;
struct sock *tcp_sock;
struct inet_peer_base *peers;
struct tcpm_hash_bucket *tcp_metrics_hash;
unsigned int tcp_metrics_hash_mask;
struct netns_frags frags;
#ifdef CONFIG_NETFILTER
struct xt_table *iptable_filter;
......
......@@ -389,6 +389,7 @@ extern void tcp_enter_loss(struct sock *sk, int how);
extern void tcp_clear_retrans(struct tcp_sock *tp);
extern void tcp_update_metrics(struct sock *sk);
extern void tcp_init_metrics(struct sock *sk);
extern void tcp_metrics_init(void);
extern bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst);
extern void tcp_disable_fack(struct tcp_sock *tp);
extern void tcp_close(struct sock *sk, long timeout);
......
......@@ -3563,6 +3563,8 @@ void __init tcp_init(void)
pr_info("Hash tables configured (established %u bind %u)\n",
tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
tcp_metrics_init();
tcp_register_congestion_control(&tcp_reno);
memset(&tcp_secret_one.secrets[0], 0, sizeof(tcp_secret_one.secrets));
......
#include <linux/rcupdate.h>
#include <linux/spinlock.h>
#include <linux/jiffies.h>
#include <linux/bootmem.h>
#include <linux/module.h>
#include <linux/cache.h>
#include <linux/slab.h>
#include <linux/init.h>
#include <linux/tcp.h>
#include <net/inet_connection_sock.h>
#include <net/net_namespace.h>
#include <net/request_sock.h>
#include <net/inetpeer.h>
#include <net/sock.h>
#include <net/ipv6.h>
#include <net/dst.h>
#include <net/tcp.h>
int sysctl_tcp_nometrics_save __read_mostly;
enum tcp_metric_index {
TCP_METRIC_RTT,
TCP_METRIC_RTTVAR,
TCP_METRIC_SSTHRESH,
TCP_METRIC_CWND,
TCP_METRIC_REORDERING,
/* Always last. */
TCP_METRIC_MAX,
};
struct tcp_metrics_block {
struct tcp_metrics_block __rcu *tcpm_next;
struct inetpeer_addr tcpm_addr;
unsigned long tcpm_stamp;
u32 tcpm_lock;
u32 tcpm_vals[TCP_METRIC_MAX];
};
static bool tcp_metric_locked(struct tcp_metrics_block *tm,
enum tcp_metric_index idx)
{
return tm->tcpm_lock & (1 << idx);
}
static u32 tcp_metric_get(struct tcp_metrics_block *tm,
enum tcp_metric_index idx)
{
return tm->tcpm_vals[idx];
}
static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
enum tcp_metric_index idx)
{
return msecs_to_jiffies(tm->tcpm_vals[idx]);
}
static void tcp_metric_set(struct tcp_metrics_block *tm,
enum tcp_metric_index idx,
u32 val)
{
tm->tcpm_vals[idx] = val;
}
static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
enum tcp_metric_index idx,
u32 val)
{
tm->tcpm_vals[idx] = jiffies_to_msecs(val);
}
static bool addr_same(const struct inetpeer_addr *a,
const struct inetpeer_addr *b)
{
const struct in6_addr *a6, *b6;
if (a->family != b->family)
return false;
if (a->family == AF_INET)
return a->addr.a4 == b->addr.a4;
a6 = (const struct in6_addr *) &a->addr.a6[0];
b6 = (const struct in6_addr *) &b->addr.a6[0];
return ipv6_addr_equal(a6, b6);
}
struct tcpm_hash_bucket {
struct tcp_metrics_block __rcu *chain;
};
static DEFINE_SPINLOCK(tcp_metrics_lock);
static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst)
{
u32 val;
val = 0;
if (dst_metric_locked(dst, RTAX_RTT))
val |= 1 << TCP_METRIC_RTT;
if (dst_metric_locked(dst, RTAX_RTTVAR))
val |= 1 << TCP_METRIC_RTTVAR;
if (dst_metric_locked(dst, RTAX_SSTHRESH))
val |= 1 << TCP_METRIC_SSTHRESH;
if (dst_metric_locked(dst, RTAX_CWND))
val |= 1 << TCP_METRIC_CWND;
if (dst_metric_locked(dst, RTAX_REORDERING))
val |= 1 << TCP_METRIC_REORDERING;
tm->tcpm_lock = val;
tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
}
static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
struct inetpeer_addr *addr,
unsigned int hash,
bool reclaim)
{
struct tcp_metrics_block *tm;
struct net *net;
spin_lock_bh(&tcp_metrics_lock);
net = dev_net(dst->dev);
if (unlikely(reclaim)) {
struct tcp_metrics_block *oldest;
oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
for (tm = rcu_dereference(oldest->tcpm_next); tm;
tm = rcu_dereference(tm->tcpm_next)) {
if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
oldest = tm;
}
tm = oldest;
} else {
tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
if (!tm)
goto out_unlock;
}
tm->tcpm_addr = *addr;
tm->tcpm_stamp = jiffies;
tcpm_suck_dst(tm, dst);
if (likely(!reclaim)) {
tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
}
out_unlock:
spin_unlock_bh(&tcp_metrics_lock);
return tm;
}
#define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
{
if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
tcpm_suck_dst(tm, dst);
}
#define TCP_METRICS_RECLAIM_DEPTH 5
#define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
{
if (tm)
return tm;
if (depth > TCP_METRICS_RECLAIM_DEPTH)
return TCP_METRICS_RECLAIM_PTR;
return NULL;
}
static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr,
struct net *net, unsigned int hash)
{
struct tcp_metrics_block *tm;
int depth = 0;
for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
tm = rcu_dereference(tm->tcpm_next)) {
if (addr_same(&tm->tcpm_addr, addr))
break;
depth++;
}
return tcp_get_encode(tm, depth);
}
static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
struct dst_entry *dst)
{
struct tcp_metrics_block *tm;
struct inetpeer_addr addr;
unsigned int hash;
struct net *net;
addr.family = req->rsk_ops->family;
switch (addr.family) {
case AF_INET:
addr.addr.a4 = inet_rsk(req)->rmt_addr;
hash = (__force unsigned int) addr.addr.a4;
break;
case AF_INET6:
*(struct in6_addr *)addr.addr.a6 = inet6_rsk(req)->rmt_addr;
hash = ((__force unsigned int) addr.addr.a6[0] ^
(__force unsigned int) addr.addr.a6[1] ^
(__force unsigned int) addr.addr.a6[2] ^
(__force unsigned int) addr.addr.a6[3]);
break;
default:
return NULL;
}
hash ^= (hash >> 24) ^ (hash >> 16) ^ (hash >> 8);
net = dev_net(dst->dev);
hash &= net->ipv4.tcp_metrics_hash_mask;
for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
tm = rcu_dereference(tm->tcpm_next)) {
if (addr_same(&tm->tcpm_addr, &addr))
break;
}
tcpm_check_stamp(tm, dst);
return tm;
}
static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
struct dst_entry *dst,
bool create)
{
struct tcp_metrics_block *tm;
struct inetpeer_addr addr;
unsigned int hash;
struct net *net;
bool reclaim;
addr.family = sk->sk_family;
switch (addr.family) {
case AF_INET:
addr.addr.a4 = inet_sk(sk)->inet_daddr;
hash = (__force unsigned int) addr.addr.a4;
break;
case AF_INET6:
*(struct in6_addr *)addr.addr.a6 = inet6_sk(sk)->daddr;
hash = ((__force unsigned int) addr.addr.a6[0] ^
(__force unsigned int) addr.addr.a6[1] ^
(__force unsigned int) addr.addr.a6[2] ^
(__force unsigned int) addr.addr.a6[3]);
break;
default:
return NULL;
}
hash ^= (hash >> 24) ^ (hash >> 16) ^ (hash >> 8);
net = dev_net(dst->dev);
hash &= net->ipv4.tcp_metrics_hash_mask;
tm = __tcp_get_metrics(&addr, net, hash);
reclaim = false;
if (tm == TCP_METRICS_RECLAIM_PTR) {
reclaim = true;
tm = NULL;
}
if (!tm && create)
tm = tcpm_new(dst, &addr, hash, reclaim);
else
tcpm_check_stamp(tm, dst);
return tm;
}
/* Save metrics learned by this TCP session. This function is called
* only, when TCP finishes successfully i.e. when it enters TIME-WAIT
* or goes from LAST-ACK to CLOSE.
*/
void tcp_update_metrics(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
const struct inet_connection_sock *icsk = inet_csk(sk);
struct dst_entry *dst = __sk_dst_get(sk);
struct tcp_sock *tp = tcp_sk(sk);
struct tcp_metrics_block *tm;
unsigned long rtt;
u32 val;
int m;
if (sysctl_tcp_nometrics_save)
if (sysctl_tcp_nometrics_save || !dst)
return;
if (dst && (dst->flags & DST_HOST)) {
const struct inet_connection_sock *icsk = inet_csk(sk);
int m;
unsigned long rtt;
if (dst->flags & DST_HOST)
dst_confirm(dst);
rcu_read_lock();
if (icsk->icsk_backoff || !tp->srtt) {
/* This session failed to estimate rtt. Why?
* Probably, no packets returned in time.
* Reset our results.
* Probably, no packets returned in time. Reset our
* results.
*/
if (!(dst_metric_locked(dst, RTAX_RTT)))
dst_metric_set(dst, RTAX_RTT, 0);
return;
}
tm = tcp_get_metrics(sk, dst, false);
if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
tcp_metric_set(tm, TCP_METRIC_RTT, 0);
goto out_unlock;
} else
tm = tcp_get_metrics(sk, dst, true);
if (!tm)
goto out_unlock;
rtt = dst_metric_rtt(dst, RTAX_RTT);
rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
m = rtt - tp->srtt;
/* If newly calculated rtt larger than stored one,
* store new one. Otherwise, use EWMA. Remember,
* rtt overestimation is always better than underestimation.
/* If newly calculated rtt larger than stored one, store new
* one. Otherwise, use EWMA. Remember, rtt overestimation is
* always better than underestimation.
*/
if (!(dst_metric_locked(dst, RTAX_RTT))) {
if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
if (m <= 0)
set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt);
rtt = tp->srtt;
else
set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3));
rtt -= (m >> 3);
tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
}
if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
unsigned long var;
if (m < 0)
m = -m;
......@@ -63,72 +338,94 @@ void tcp_update_metrics(struct sock *sk)
if (m < tp->mdev)
m = tp->mdev;
var = dst_metric_rtt(dst, RTAX_RTTVAR);
var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
if (m >= var)
var = m;
else
var -= (var - m) >> 2;
set_dst_metric_rtt(dst, RTAX_RTTVAR, var);
tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
}
if (tcp_in_initial_slowstart(tp)) {
/* Slow start still did not finish. */
if (dst_metric(dst, RTAX_SSTHRESH) &&
!dst_metric_locked(dst, RTAX_SSTHRESH) &&
(tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1);
if (!dst_metric_locked(dst, RTAX_CWND) &&
tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd);
if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
if (val && (tp->snd_cwnd >> 1) > val)
tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
tp->snd_cwnd >> 1);
}
if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
val = tcp_metric_get(tm, TCP_METRIC_CWND);
if (tp->snd_cwnd > val)
tcp_metric_set(tm, TCP_METRIC_CWND,
tp->snd_cwnd);
}
} else if (tp->snd_cwnd > tp->snd_ssthresh &&
icsk->icsk_ca_state == TCP_CA_Open) {
/* Cong. avoidance phase, cwnd is reliable. */
if (!dst_metric_locked(dst, RTAX_SSTHRESH))
dst_metric_set(dst, RTAX_SSTHRESH,
if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
if (!dst_metric_locked(dst, RTAX_CWND))
dst_metric_set(dst, RTAX_CWND,
(dst_metric(dst, RTAX_CWND) +
tp->snd_cwnd) >> 1);
if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
val = tcp_metric_get(tm, TCP_METRIC_CWND);
tcp_metric_set(tm, RTAX_CWND, (val + tp->snd_cwnd) >> 1);
}
} else {
/* Else slow start did not finish, cwnd is non-sense,
ssthresh may be also invalid.
* ssthresh may be also invalid.
*/
if (!dst_metric_locked(dst, RTAX_CWND))
dst_metric_set(dst, RTAX_CWND,
(dst_metric(dst, RTAX_CWND) +
tp->snd_ssthresh) >> 1);
if (dst_metric(dst, RTAX_SSTHRESH) &&
!dst_metric_locked(dst, RTAX_SSTHRESH) &&
tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH))
dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh);
}
if (!dst_metric_locked(dst, RTAX_REORDERING)) {
if (dst_metric(dst, RTAX_REORDERING) < tp->reordering &&
if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
val = tcp_metric_get(tm, TCP_METRIC_CWND);
tcp_metric_set(tm, TCP_METRIC_CWND,
(val + tp->snd_ssthresh) >> 1);
}
if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
if (val && tp->snd_ssthresh > val)
tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
tp->snd_ssthresh);
}
if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
if (val < tp->reordering &&
tp->reordering != sysctl_tcp_reordering)
dst_metric_set(dst, RTAX_REORDERING, tp->reordering);
tcp_metric_set(tm, TCP_METRIC_REORDERING,
tp->reordering);
}
}
tm->tcpm_stamp = jiffies;
out_unlock:
rcu_read_unlock();
}
/* Initialize metrics on socket. */
void tcp_init_metrics(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
struct dst_entry *dst = __sk_dst_get(sk);
struct tcp_sock *tp = tcp_sk(sk);
struct tcp_metrics_block *tm;
u32 val;
if (dst == NULL)
goto reset;
dst_confirm(dst);
if (dst_metric_locked(dst, RTAX_CWND))
tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
if (dst_metric(dst, RTAX_SSTHRESH)) {
tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
rcu_read_lock();
tm = tcp_get_metrics(sk, dst, true);
if (!tm) {
rcu_read_unlock();
goto reset;
}
if (tcp_metric_locked(tm, TCP_METRIC_CWND))
tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
if (val) {
tp->snd_ssthresh = val;
if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
tp->snd_ssthresh = tp->snd_cwnd_clamp;
} else {
......@@ -137,16 +434,18 @@ void tcp_init_metrics(struct sock *sk)
*/
tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
}
if (dst_metric(dst, RTAX_REORDERING) &&
tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
if (val && tp->reordering != val) {
tcp_disable_fack(tp);
tcp_disable_early_retrans(tp);
tp->reordering = dst_metric(dst, RTAX_REORDERING);
tp->reordering = val;
}
if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0)
val = tcp_metric_get(tm, TCP_METRIC_RTT);
if (val == 0 || tp->srtt == 0) {
rcu_read_unlock();
goto reset;
}
/* Initial rtt is determined from SYN,SYN-ACK.
* The segment is small and rtt may appear much
* less than real one. Use per-dst memory
......@@ -161,14 +460,18 @@ void tcp_init_metrics(struct sock *sk)
* to low value, and then abruptly stops to do it and starts to delay
* ACKs, wait for troubles.
*/
if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) {
tp->srtt = dst_metric_rtt(dst, RTAX_RTT);
val = msecs_to_jiffies(val);
if (val > tp->srtt) {
tp->srtt = val;
tp->rtt_seq = tp->snd_nxt;
}
if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) {
tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR);
val = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
if (val > tp->mdev) {
tp->mdev = val;
tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
}
rcu_read_unlock();
tcp_set_rto(sk);
reset:
if (tp->srtt == 0) {
......@@ -195,8 +498,74 @@ void tcp_init_metrics(struct sock *sk)
bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
{
struct tcp_metrics_block *tm;
bool ret;
if (!dst)
return false;
return dst_metric(dst, RTAX_RTT) ? true : false;
rcu_read_lock();
tm = __tcp_get_metrics_req(req, dst);
if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
ret = true;
else
ret = false;
rcu_read_unlock();
return ret;
}
EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
static unsigned long tcpmhash_entries;
static int __init set_tcpmhash_entries(char *str)
{
ssize_t ret;
if (!str)
return 0;
ret = kstrtoul(str, 0, &tcpmhash_entries);
if (ret)
return 0;
return 1;
}
__setup("tcpmhash_entries=", set_tcpmhash_entries);
static int __net_init tcp_net_metrics_init(struct net *net)
{
int slots, size;
slots = tcpmhash_entries;
if (!slots) {
if (totalram_pages >= 128 * 1024)
slots = 16 * 1024;
else
slots = 8 * 1024;
}
size = slots * sizeof(struct tcpm_hash_bucket);
net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL);
if (!net->ipv4.tcp_metrics_hash)
return -ENOMEM;
net->ipv4.tcp_metrics_hash_mask = (slots - 1);
return 0;
}
static void __net_exit tcp_net_metrics_exit(struct net *net)
{
kfree(net->ipv4.tcp_metrics_hash);
}
static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
.init = tcp_net_metrics_init,
.exit = tcp_net_metrics_exit,
};
void __init tcp_metrics_init(void)
{
register_pernet_subsys(&tcp_net_metrics_ops);
}
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment