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>

1 files changed, 36 insertions, 234 deletions

diff --git a/net/ipv4/inet_timewait_sock.c b/net/ipv4/inet_timewait_sock.c
index 118f0f195820..00ec8d5d7e7e 100644
--- a/net/ipv4/inet_timewait_sock.c
+++ b/net/ipv4/inet_timewait_sock.c
@@ -67,9 +67,9 @@ int inet_twsk_bind_unhash(struct inet_timewait_sock *tw,
 }
 
 /* Must be called with locally disabled BHs. */
-static void __inet_twsk_kill(struct inet_timewait_sock *tw,
-			     struct inet_hashinfo *hashinfo)
+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. */
@@ -89,6 +89,8 @@ static void __inet_twsk_kill(struct inet_timewait_sock *tw,
 
 	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)
@@ -168,16 +170,34 @@ void __inet_twsk_hashdance(struct inet_timewait_sock *tw, struct sock *sk,
 }
 EXPORT_SYMBOL_GPL(__inet_twsk_hashdance);
 
-struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int state)
+void tw_timer_handler(unsigned long data)
 {
-	struct inet_timewait_sock *tw =
-		kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
-				 GFP_ATOMIC);
+	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;
@@ -196,13 +216,14 @@ struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int stat
 		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);
-		inet_twsk_dead_node_init(tw);
+
 		__module_get(tw->tw_prot->owner);
 	}
 
@@ -210,139 +231,20 @@ struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk, const int stat
 }
 EXPORT_SYMBOL_GPL(inet_twsk_alloc);
 
-/* Returns non-zero if quota exceeded.  */
-static int inet_twdr_do_twkill_work(struct inet_timewait_death_row *twdr,
-				    const int slot)
-{
-	struct inet_timewait_sock *tw;
-	unsigned int killed;
-	int ret;
-
-	/* NOTE: compare this to previous version where lock
-	 * was released after detaching chain. It was racy,
-	 * because tw buckets are scheduled in not serialized context
-	 * in 2.3 (with netfilter), and with softnet it is common, because
-	 * soft irqs are not sequenced.
-	 */
-	killed = 0;
-	ret = 0;
-rescan:
-	inet_twsk_for_each_inmate(tw, &twdr->cells[slot]) {
-		__inet_twsk_del_dead_node(tw);
-		spin_unlock(&twdr->death_lock);
-		__inet_twsk_kill(tw, twdr->hashinfo);
-#ifdef CONFIG_NET_NS
-		NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITED);
-#endif
-		inet_twsk_put(tw);
-		killed++;
-		spin_lock(&twdr->death_lock);
-		if (killed > INET_TWDR_TWKILL_QUOTA) {
-			ret = 1;
-			break;
-		}
-
-		/* While we dropped twdr->death_lock, another cpu may have
-		 * killed off the next TW bucket in the list, therefore
-		 * do a fresh re-read of the hlist head node with the
-		 * lock reacquired.  We still use the hlist traversal
-		 * macro in order to get the prefetches.
-		 */
-		goto rescan;
-	}
-
-	twdr->tw_count -= killed;
-#ifndef CONFIG_NET_NS
-	NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITED, killed);
-#endif
-	return ret;
-}
-
-void inet_twdr_hangman(unsigned long data)
-{
-	struct inet_timewait_death_row *twdr;
-	unsigned int need_timer;
-
-	twdr = (struct inet_timewait_death_row *)data;
-	spin_lock(&twdr->death_lock);
-
-	if (twdr->tw_count == 0)
-		goto out;
-
-	need_timer = 0;
-	if (inet_twdr_do_twkill_work(twdr, twdr->slot)) {
-		twdr->thread_slots |= (1 << twdr->slot);
-		schedule_work(&twdr->twkill_work);
-		need_timer = 1;
-	} else {
-		/* We purged the entire slot, anything left?  */
-		if (twdr->tw_count)
-			need_timer = 1;
-		twdr->slot = ((twdr->slot + 1) & (INET_TWDR_TWKILL_SLOTS - 1));
-	}
-	if (need_timer)
-		mod_timer(&twdr->tw_timer, jiffies + twdr->period);
-out:
-	spin_unlock(&twdr->death_lock);
-}
-EXPORT_SYMBOL_GPL(inet_twdr_hangman);
-
-void inet_twdr_twkill_work(struct work_struct *work)
-{
-	struct inet_timewait_death_row *twdr =
-		container_of(work, struct inet_timewait_death_row, twkill_work);
-	int i;
-
-	BUILD_BUG_ON((INET_TWDR_TWKILL_SLOTS - 1) >
-			(sizeof(twdr->thread_slots) * 8));
-
-	while (twdr->thread_slots) {
-		spin_lock_bh(&twdr->death_lock);
-		for (i = 0; i < INET_TWDR_TWKILL_SLOTS; i++) {
-			if (!(twdr->thread_slots & (1 << i)))
-				continue;
-
-			while (inet_twdr_do_twkill_work(twdr, i) != 0) {
-				if (need_resched()) {
-					spin_unlock_bh(&twdr->death_lock);
-					schedule();
-					spin_lock_bh(&twdr->death_lock);
-				}
-			}
-
-			twdr->thread_slots &= ~(1 << i);
-		}
-		spin_unlock_bh(&twdr->death_lock);
-	}
-}
-EXPORT_SYMBOL_GPL(inet_twdr_twkill_work);
-
 /* 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,
-			  struct inet_timewait_death_row *twdr)
+void inet_twsk_deschedule(struct inet_timewait_sock *tw)
 {
-	spin_lock(&twdr->death_lock);
-	if (inet_twsk_del_dead_node(tw)) {
-		inet_twsk_put(tw);
-		if (--twdr->tw_count == 0)
-			del_timer(&twdr->tw_timer);
-	}
-	spin_unlock(&twdr->death_lock);
-	__inet_twsk_kill(tw, twdr->hashinfo);
+	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,
-		       struct inet_timewait_death_row *twdr,
-		       const int timeo, const int timewait_len)
+void inet_twsk_schedule(struct inet_timewait_sock *tw, const int timeo)
 {
-	struct hlist_head *list;
-	int slot;
-
 	/* timeout := RTO * 3.5
 	 *
 	 * 3.5 = 1+2+0.5 to wait for two retransmits.
@@ -367,115 +269,15 @@ void inet_twsk_schedule(struct inet_timewait_sock *tw,
 	 * is greater than TS tick!) and detect old duplicates with help
 	 * of PAWS.
 	 */
-	slot = (timeo + (1 << INET_TWDR_RECYCLE_TICK) - 1) >> INET_TWDR_RECYCLE_TICK;
 
-	spin_lock(&twdr->death_lock);
-
-	/* Unlink it, if it was scheduled */
-	if (inet_twsk_del_dead_node(tw))
-		twdr->tw_count--;
-	else
+	tw->tw_kill = timeo <= 4*HZ;
+	if (!mod_timer_pinned(&tw->tw_timer, jiffies + timeo)) {
 		atomic_inc(&tw->tw_refcnt);
-
-	if (slot >= INET_TWDR_RECYCLE_SLOTS) {
-		/* Schedule to slow timer */
-		if (timeo >= timewait_len) {
-			slot = INET_TWDR_TWKILL_SLOTS - 1;
-		} else {
-			slot = DIV_ROUND_UP(timeo, twdr->period);
-			if (slot >= INET_TWDR_TWKILL_SLOTS)
-				slot = INET_TWDR_TWKILL_SLOTS - 1;
-		}
-		tw->tw_ttd = inet_tw_time_stamp() + timeo;
-		slot = (twdr->slot + slot) & (INET_TWDR_TWKILL_SLOTS - 1);
-		list = &twdr->cells[slot];
-	} else {
-		tw->tw_ttd = inet_tw_time_stamp() + (slot << INET_TWDR_RECYCLE_TICK);
-
-		if (twdr->twcal_hand < 0) {
-			twdr->twcal_hand = 0;
-			twdr->twcal_jiffie = jiffies;
-			twdr->twcal_timer.expires = twdr->twcal_jiffie +
-					      (slot << INET_TWDR_RECYCLE_TICK);
-			add_timer(&twdr->twcal_timer);
-		} else {
-			if (time_after(twdr->twcal_timer.expires,
-				       jiffies + (slot << INET_TWDR_RECYCLE_TICK)))
-				mod_timer(&twdr->twcal_timer,
-					  jiffies + (slot << INET_TWDR_RECYCLE_TICK));
-			slot = (twdr->twcal_hand + slot) & (INET_TWDR_RECYCLE_SLOTS - 1);
-		}
-		list = &twdr->twcal_row[slot];
+		atomic_inc(&tw->tw_dr->tw_count);
 	}
-
-	hlist_add_head(&tw->tw_death_node, list);
-
-	if (twdr->tw_count++ == 0)
-		mod_timer(&twdr->tw_timer, jiffies + twdr->period);
-	spin_unlock(&twdr->death_lock);
 }
 EXPORT_SYMBOL_GPL(inet_twsk_schedule);
 
-void inet_twdr_twcal_tick(unsigned long data)
-{
-	struct inet_timewait_death_row *twdr;
-	int n, slot;
-	unsigned long j;
-	unsigned long now = jiffies;
-	int killed = 0;
-	int adv = 0;
-
-	twdr = (struct inet_timewait_death_row *)data;
-
-	spin_lock(&twdr->death_lock);
-	if (twdr->twcal_hand < 0)
-		goto out;
-
-	slot = twdr->twcal_hand;
-	j = twdr->twcal_jiffie;
-
-	for (n = 0; n < INET_TWDR_RECYCLE_SLOTS; n++) {
-		if (time_before_eq(j, now)) {
-			struct hlist_node *safe;
-			struct inet_timewait_sock *tw;
-
-			inet_twsk_for_each_inmate_safe(tw, safe,
-						       &twdr->twcal_row[slot]) {
-				__inet_twsk_del_dead_node(tw);
-				__inet_twsk_kill(tw, twdr->hashinfo);
-#ifdef CONFIG_NET_NS
-				NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_TIMEWAITKILLED);
-#endif
-				inet_twsk_put(tw);
-				killed++;
-			}
-		} else {
-			if (!adv) {
-				adv = 1;
-				twdr->twcal_jiffie = j;
-				twdr->twcal_hand = slot;
-			}
-
-			if (!hlist_empty(&twdr->twcal_row[slot])) {
-				mod_timer(&twdr->twcal_timer, j);
-				goto out;
-			}
-		}
-		j += 1 << INET_TWDR_RECYCLE_TICK;
-		slot = (slot + 1) & (INET_TWDR_RECYCLE_SLOTS - 1);
-	}
-	twdr->twcal_hand = -1;
-
-out:
-	if ((twdr->tw_count -= killed) == 0)
-		del_timer(&twdr->tw_timer);
-#ifndef CONFIG_NET_NS
-	NET_ADD_STATS_BH(&init_net, LINUX_MIB_TIMEWAITKILLED, killed);
-#endif
-	spin_unlock(&twdr->death_lock);
-}
-EXPORT_SYMBOL_GPL(inet_twdr_twcal_tick);
-
 void inet_twsk_purge(struct inet_hashinfo *hashinfo,
 		     struct inet_timewait_death_row *twdr, int family)
 {
@@ -509,7 +311,7 @@ restart:
 
 			rcu_read_unlock();
 			local_bh_disable();
-			inet_twsk_deschedule(tw, twdr);
+			inet_twsk_deschedule(tw);
 			local_bh_enable();
 			inet_twsk_put(tw);
 			goto restart_rcu;