tcp: switch orphan_count to bare per-cpu counters

Use of percpu_counter structure to track count of orphaned
sockets is causing problems on modern hosts with 256 cpus
or more.

Stefan Bach reported a serious spinlock contention in real workloads,
that I was able to reproduce with a netfilter rule dropping
incoming FIN packets.

    53.56%  server  [kernel.kallsyms]      [k] queued_spin_lock_slowpath
            |
            ---queued_spin_lock_slowpath
               |
                --53.51%--_raw_spin_lock_irqsave
                          |
                           --53.51%--__percpu_counter_sum
                                     tcp_check_oom
                                     |
                                     |--39.03%--__tcp_close
                                     |          tcp_close
                                     |          inet_release
                                     |          inet6_release
                                     |          sock_close
                                     |          __fput
                                     |          ____fput
                                     |          task_work_run
                                     |          exit_to_usermode_loop
                                     |          do_syscall_64
                                     |          entry_SYSCALL_64_after_hwframe
                                     |          __GI___libc_close
                                     |
                                      --14.48%--tcp_out_of_resources
                                                tcp_write_timeout
                                                tcp_retransmit_timer
                                                tcp_write_timer_handler
                                                tcp_write_timer
                                                call_timer_fn
                                                expire_timers
                                                __run_timers
                                                run_timer_softirq
                                                __softirqentry_text_start

As explained in commit cf86a086a180 ("net/dst: use a smaller percpu_counter
batch for dst entries accounting"), default batch size is too big
for the default value of tcp_max_orphans (262144).

But even if we reduce batch sizes, there would still be cases
where the estimated count of orphans is beyond the limit,
and where tcp_too_many_orphans() has to call the expensive
percpu_counter_sum_positive().

One solution is to use plain per-cpu counters, and have
a timer to periodically refresh this cache.

Updating this cache every 100ms seems about right, tcp pressure
state is not radically changing over shorter periods.

percpu_counter was nice 15 years ago while hosts had less
than 16 cpus, not anymore by current standards.

v2: Fix the build issue for CONFIG_CRYPTO_DEV_CHELSIO_TLS=m,
    reported by kernel test robot <lkp@intel.com>
    Remove unused socket argument from tcp_too_many_orphans()

Fixes: dd24c00191d5 ("net: Use a percpu_counter for orphan_count")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Stefan Bach <sfb@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

4 files changed, 37 insertions, 9 deletions

diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index f25d02ad4a8a..f7fea3a7c5e6 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -1015,7 +1015,7 @@ void inet_csk_destroy_sock(struct sock *sk)
 
 	sk_refcnt_debug_release(sk);
 
-	percpu_counter_dec(sk->sk_prot->orphan_count);
+	this_cpu_dec(*sk->sk_prot->orphan_count);
 
 	sock_put(sk);
 }
@@ -1074,7 +1074,7 @@ static void inet_child_forget(struct sock *sk, struct request_sock *req,
 
 	sock_orphan(child);
 
-	percpu_counter_inc(sk->sk_prot->orphan_count);
+	this_cpu_inc(*sk->sk_prot->orphan_count);
 
 	if (sk->sk_protocol == IPPROTO_TCP && tcp_rsk(req)->tfo_listener) {
 		BUG_ON(rcu_access_pointer(tcp_sk(child)->fastopen_rsk) != req);
diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
index bfb522e51346..75737267746f 100644
--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -598,7 +598,7 @@ bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 	if (ok) {
 		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 	} else {
-		percpu_counter_inc(sk->sk_prot->orphan_count);
+		this_cpu_inc(*sk->sk_prot->orphan_count);
 		inet_sk_set_state(sk, TCP_CLOSE);
 		sock_set_flag(sk, SOCK_DEAD);
 		inet_csk_destroy_sock(sk);
diff --git a/net/ipv4/proc.c b/net/ipv4/proc.c
index b0d3a09dc84e..f30273afb539 100644
--- a/net/ipv4/proc.c
+++ b/net/ipv4/proc.c
@@ -53,7 +53,7 @@ static int sockstat_seq_show(struct seq_file *seq, void *v)
 	struct net *net = seq->private;
 	int orphans, sockets;
 
-	orphans = percpu_counter_sum_positive(&tcp_orphan_count);
+	orphans = tcp_orphan_count_sum();
 	sockets = proto_sockets_allocated_sum_positive(&tcp_prot);
 
 	socket_seq_show(seq);
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 414c179c28e0..c2d9830136d2 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -287,8 +287,8 @@ enum {
 	TCP_CMSG_TS = 2
 };
 
-struct percpu_counter tcp_orphan_count;
-EXPORT_SYMBOL_GPL(tcp_orphan_count);
+DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
+EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
 
 long sysctl_tcp_mem[3] __read_mostly;
 EXPORT_SYMBOL(sysctl_tcp_mem);
@@ -2673,11 +2673,36 @@ void tcp_shutdown(struct sock *sk, int how)
 }
 EXPORT_SYMBOL(tcp_shutdown);
 
+int tcp_orphan_count_sum(void)
+{
+	int i, total = 0;
+
+	for_each_possible_cpu(i)
+		total += per_cpu(tcp_orphan_count, i);
+
+	return max(total, 0);
+}
+
+static int tcp_orphan_cache;
+static struct timer_list tcp_orphan_timer;
+#define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
+
+static void tcp_orphan_update(struct timer_list *unused)
+{
+	WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
+	mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
+}
+
+static bool tcp_too_many_orphans(int shift)
+{
+	return READ_ONCE(tcp_orphan_cache) << shift > sysctl_tcp_max_orphans;
+}
+
 bool tcp_check_oom(struct sock *sk, int shift)
 {
 	bool too_many_orphans, out_of_socket_memory;
 
-	too_many_orphans = tcp_too_many_orphans(sk, shift);
+	too_many_orphans = tcp_too_many_orphans(shift);
 	out_of_socket_memory = tcp_out_of_memory(sk);
 
 	if (too_many_orphans)
@@ -2786,7 +2811,7 @@ adjudge_to_death:
 	/* remove backlog if any, without releasing ownership. */
 	__release_sock(sk);
 
-	percpu_counter_inc(sk->sk_prot->orphan_count);
+	this_cpu_inc(tcp_orphan_count);
 
 	/* Have we already been destroyed by a softirq or backlog? */
 	if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
@@ -4479,7 +4504,10 @@ void __init tcp_init(void)
 		     sizeof_field(struct sk_buff, cb));
 
 	percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
-	percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
+
+	timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
+	mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
+
 	inet_hashinfo_init(&tcp_hashinfo);
 	inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
 			    thash_entries, 21,  /* one slot per 2 MB*/