linux_dsm_epyc7002/net/ipv4/inet_timewait_sock.c
Eric Dumazet 789f558cfb tcp/dccp: get rid of central timewait timer
Using a timer wheel for timewait sockets was nice ~15 years ago when
memory was expensive and machines had a single processor.

This does not scale, code is ugly and source of huge latencies
(Typically 30 ms have been seen, cpus spinning on death_lock spinlock.)

We can afford to use an extra 64 bytes per timewait sock and spread
timewait load to all cpus to have better behavior.

Tested:

On following test, /proc/sys/net/ipv4/tcp_tw_recycle is set to 1
on the target (lpaa24)

Before patch :

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
419594

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
437171

While test is running, we can observe 25 or even 33 ms latencies.

lpaa24:~# ping -c 1000 -i 0.02 -qn lpaa23
...
1000 packets transmitted, 1000 received, 0% packet loss, time 20601ms
rtt min/avg/max/mdev = 0.020/0.217/25.771/1.535 ms, pipe 2

lpaa24:~# ping -c 1000 -i 0.02 -qn lpaa23
...
1000 packets transmitted, 1000 received, 0% packet loss, time 20702ms
rtt min/avg/max/mdev = 0.019/0.183/33.761/1.441 ms, pipe 2

After patch :

About 90% increase of throughput :

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
810442

lpaa23:~# ./super_netperf 200 -H lpaa24 -t TCP_CC -l 60 -- -p0,0
800992

And latencies are kept to minimal values during this load, even
if network utilization is 90% higher :

lpaa24:~# ping -c 1000 -i 0.02 -qn lpaa23
...
1000 packets transmitted, 1000 received, 0% packet loss, time 19991ms
rtt min/avg/max/mdev = 0.023/0.064/0.360/0.042 ms

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-04-13 16:40:05 -04:00

329 lines
9.5 KiB
C

/*
* INET An implementation of the TCP/IP protocol suite for the LINUX
* operating system. INET is implemented using the BSD Socket
* interface as the means of communication with the user level.
*
* Generic TIME_WAIT sockets functions
*
* From code orinally in TCP
*/
#include <linux/kernel.h>
#include <linux/kmemcheck.h>
#include <linux/slab.h>
#include <linux/module.h>
#include <net/inet_hashtables.h>
#include <net/inet_timewait_sock.h>
#include <net/ip.h>
/**
* inet_twsk_unhash - unhash a timewait socket from established hash
* @tw: timewait socket
*
* unhash a timewait socket from established hash, if hashed.
* ehash lock must be held by caller.
* Returns 1 if caller should call inet_twsk_put() after lock release.
*/
int inet_twsk_unhash(struct inet_timewait_sock *tw)
{
if (hlist_nulls_unhashed(&tw->tw_node))
return 0;
hlist_nulls_del_rcu(&tw->tw_node);
sk_nulls_node_init(&tw->tw_node);
/*
* We cannot call inet_twsk_put() ourself under lock,
* caller must call it for us.
*/
return 1;
}
/**
* inet_twsk_bind_unhash - unhash a timewait socket from bind hash
* @tw: timewait socket
* @hashinfo: hashinfo pointer
*
* unhash a timewait socket from bind hash, if hashed.
* bind hash lock must be held by caller.
* Returns 1 if caller should call inet_twsk_put() after lock release.
*/
int inet_twsk_bind_unhash(struct inet_timewait_sock *tw,
struct inet_hashinfo *hashinfo)
{
struct inet_bind_bucket *tb = tw->tw_tb;
if (!tb)
return 0;
__hlist_del(&tw->tw_bind_node);
tw->tw_tb = NULL;
inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
/*
* We cannot call inet_twsk_put() ourself under lock,
* caller must call it for us.
*/
return 1;
}
/* Must be called with locally disabled BHs. */
static void inet_twsk_kill(struct inet_timewait_sock *tw)
{
struct inet_hashinfo *hashinfo = tw->tw_dr->hashinfo;
struct inet_bind_hashbucket *bhead;
int refcnt;
/* Unlink from established hashes. */
spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
spin_lock(lock);
refcnt = inet_twsk_unhash(tw);
spin_unlock(lock);
/* Disassociate with bind bucket. */
bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), tw->tw_num,
hashinfo->bhash_size)];
spin_lock(&bhead->lock);
refcnt += inet_twsk_bind_unhash(tw, hashinfo);
spin_unlock(&bhead->lock);
BUG_ON(refcnt >= atomic_read(&tw->tw_refcnt));
atomic_sub(refcnt, &tw->tw_refcnt);
atomic_dec(&tw->tw_dr->tw_count);
inet_twsk_put(tw);
}
void inet_twsk_free(struct inet_timewait_sock *tw)
{
struct module *owner = tw->tw_prot->owner;
twsk_destructor((struct sock *)tw);
#ifdef SOCK_REFCNT_DEBUG
pr_debug("%s timewait_sock %p released\n", tw->tw_prot->name, tw);
#endif
kmem_cache_free(tw->tw_prot->twsk_prot->twsk_slab, tw);
module_put(owner);
}
void inet_twsk_put(struct inet_timewait_sock *tw)
{
if (atomic_dec_and_test(&tw->tw_refcnt))
inet_twsk_free(tw);
}
EXPORT_SYMBOL_GPL(inet_twsk_put);
static void inet_twsk_add_node_rcu(struct inet_timewait_sock *tw,
struct hlist_nulls_head *list)
{
hlist_nulls_add_head_rcu(&tw->tw_node, list);
}
static void inet_twsk_add_bind_node(struct inet_timewait_sock *tw,
struct hlist_head *list)
{
hlist_add_head(&tw->tw_bind_node, list);
}
/*
* Enter the time wait state. This is called with locally disabled BH.
* Essentially we whip up a timewait bucket, copy the relevant info into it
* from the SK, and mess with hash chains and list linkage.
*/
void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
struct inet_hashinfo *hashinfo)
{
const struct inet_sock *inet = inet_sk(sk);
const struct inet_connection_sock *icsk = inet_csk(sk);
struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
struct inet_bind_hashbucket *bhead;
/* Step 1: Put TW into bind hash. Original socket stays there too.
Note, that any socket with inet->num != 0 MUST be bound in
binding cache, even if it is closed.
*/
bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), inet->inet_num,
hashinfo->bhash_size)];
spin_lock(&bhead->lock);
tw->tw_tb = icsk->icsk_bind_hash;
WARN_ON(!icsk->icsk_bind_hash);
inet_twsk_add_bind_node(tw, &tw->tw_tb->owners);
spin_unlock(&bhead->lock);
spin_lock(lock);
/*
* Step 2: Hash TW into tcp ehash chain.
* Notes :
* - tw_refcnt is set to 3 because :
* - We have one reference from bhash chain.
* - We have one reference from ehash chain.
* We can use atomic_set() because prior spin_lock()/spin_unlock()
* committed into memory all tw fields.
*/
atomic_set(&tw->tw_refcnt, 1 + 1 + 1);
inet_twsk_add_node_rcu(tw, &ehead->chain);
/* Step 3: Remove SK from hash chain */
if (__sk_nulls_del_node_init_rcu(sk))
sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
spin_unlock(lock);
}
EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
void tw_timer_handler(unsigned long data)
{
struct inet_timewait_sock *tw = (struct inet_timewait_sock *)data;
if (tw->tw_kill)
NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITKILLED);
else
NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITED);
inet_twsk_kill(tw);
}
struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk,
struct inet_timewait_death_row *dr,
const int state)
{
struct inet_timewait_sock *tw;
if (atomic_read(&dr->tw_count) >= dr->sysctl_max_tw_buckets)
return NULL;
tw = kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
GFP_ATOMIC);
if (tw) {
const struct inet_sock *inet = inet_sk(sk);
kmemcheck_annotate_bitfield(tw, flags);
tw->tw_dr = dr;
/* Give us an identity. */
tw->tw_daddr = inet->inet_daddr;
tw->tw_rcv_saddr = inet->inet_rcv_saddr;
tw->tw_bound_dev_if = sk->sk_bound_dev_if;
tw->tw_tos = inet->tos;
tw->tw_num = inet->inet_num;
tw->tw_state = TCP_TIME_WAIT;
tw->tw_substate = state;
tw->tw_sport = inet->inet_sport;
tw->tw_dport = inet->inet_dport;
tw->tw_family = sk->sk_family;
tw->tw_reuse = sk->sk_reuse;
tw->tw_hash = sk->sk_hash;
tw->tw_ipv6only = 0;
tw->tw_transparent = inet->transparent;
tw->tw_prot = sk->sk_prot_creator;
atomic64_set(&tw->tw_cookie, atomic64_read(&sk->sk_cookie));
twsk_net_set(tw, sock_net(sk));
setup_timer(&tw->tw_timer, tw_timer_handler, (unsigned long)tw);
/*
* Because we use RCU lookups, we should not set tw_refcnt
* to a non null value before everything is setup for this
* timewait socket.
*/
atomic_set(&tw->tw_refcnt, 0);
__module_get(tw->tw_prot->owner);
}
return tw;
}
EXPORT_SYMBOL_GPL(inet_twsk_alloc);
/* These are always called from BH context. See callers in
* tcp_input.c to verify this.
*/
/* This is for handling early-kills of TIME_WAIT sockets. */
void inet_twsk_deschedule(struct inet_timewait_sock *tw)
{
if (del_timer_sync(&tw->tw_timer))
inet_twsk_kill(tw);
}
EXPORT_SYMBOL(inet_twsk_deschedule);
void inet_twsk_schedule(struct inet_timewait_sock *tw, const int timeo)
{
/* timeout := RTO * 3.5
*
* 3.5 = 1+2+0.5 to wait for two retransmits.
*
* RATIONALE: if FIN arrived and we entered TIME-WAIT state,
* our ACK acking that FIN can be lost. If N subsequent retransmitted
* FINs (or previous seqments) are lost (probability of such event
* is p^(N+1), where p is probability to lose single packet and
* time to detect the loss is about RTO*(2^N - 1) with exponential
* backoff). Normal timewait length is calculated so, that we
* waited at least for one retransmitted FIN (maximal RTO is 120sec).
* [ BTW Linux. following BSD, violates this requirement waiting
* only for 60sec, we should wait at least for 240 secs.
* Well, 240 consumes too much of resources 8)
* ]
* This interval is not reduced to catch old duplicate and
* responces to our wandering segments living for two MSLs.
* However, if we use PAWS to detect
* old duplicates, we can reduce the interval to bounds required
* by RTO, rather than MSL. So, if peer understands PAWS, we
* kill tw bucket after 3.5*RTO (it is important that this number
* is greater than TS tick!) and detect old duplicates with help
* of PAWS.
*/
tw->tw_kill = timeo <= 4*HZ;
if (!mod_timer_pinned(&tw->tw_timer, jiffies + timeo)) {
atomic_inc(&tw->tw_refcnt);
atomic_inc(&tw->tw_dr->tw_count);
}
}
EXPORT_SYMBOL_GPL(inet_twsk_schedule);
void inet_twsk_purge(struct inet_hashinfo *hashinfo,
struct inet_timewait_death_row *twdr, int family)
{
struct inet_timewait_sock *tw;
struct sock *sk;
struct hlist_nulls_node *node;
unsigned int slot;
for (slot = 0; slot <= hashinfo->ehash_mask; slot++) {
struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
restart_rcu:
cond_resched();
rcu_read_lock();
restart:
sk_nulls_for_each_rcu(sk, node, &head->chain) {
if (sk->sk_state != TCP_TIME_WAIT)
continue;
tw = inet_twsk(sk);
if ((tw->tw_family != family) ||
atomic_read(&twsk_net(tw)->count))
continue;
if (unlikely(!atomic_inc_not_zero(&tw->tw_refcnt)))
continue;
if (unlikely((tw->tw_family != family) ||
atomic_read(&twsk_net(tw)->count))) {
inet_twsk_put(tw);
goto restart;
}
rcu_read_unlock();
local_bh_disable();
inet_twsk_deschedule(tw);
local_bh_enable();
inet_twsk_put(tw);
goto restart_rcu;
}
/* If the nulls value we got at the end of this lookup is
* not the expected one, we must restart lookup.
* We probably met an item that was moved to another chain.
*/
if (get_nulls_value(node) != slot)
goto restart;
rcu_read_unlock();
}
}
EXPORT_SYMBOL_GPL(inet_twsk_purge);