Merge branch 'net-reduce-tcp_memory_allocated-inflation'

Eric Dumazet says:

====================
net: reduce tcp_memory_allocated inflation

Hosts with a lot of sockets tend to hit so called TCP memory pressure,
leading to very bad TCP performance and/or OOM.

The problem is that some TCP sockets can hold up to 2MB of 'forward
allocations' in their per-socket cache (sk->sk_forward_alloc),
and there is no mechanism to make them relinquish their share
under mem pressure.
Only under some potentially rare events their share is reclaimed,
one socket at a time.

In this series, I implemented a per-cpu cache instead of a per-socket one.

Each CPU has a +1/-1 MB (256 pages on x86) forward alloc cache, in order
to not dirty tcp_memory_allocated shared cache line too often.

We keep sk->sk_forward_alloc values as small as possible, to meet
memcg page granularity constraint.

Note that memcg already has a per-cpu cache, although MEMCG_CHARGE_BATCH
is defined to 32 pages, which seems a bit small.

Note that while this cover letter mentions TCP, this work is generic
and supports TCP, UDP, DECNET, SCTP.
====================

Link: https://lore.kernel.org/r/20220609063412.2205738-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

1 files changed, 9 insertions, 5 deletions

diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index aa9f2ec3dc46..6172b4750a88 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -125,6 +125,8 @@ EXPORT_SYMBOL(sysctl_udp_mem);
 
 atomic_long_t udp_memory_allocated ____cacheline_aligned_in_smp;
 EXPORT_SYMBOL(udp_memory_allocated);
+DEFINE_PER_CPU(int, udp_memory_per_cpu_fw_alloc);
+EXPORT_PER_CPU_SYMBOL_GPL(udp_memory_per_cpu_fw_alloc);
 
 #define MAX_UDP_PORTS 65536
 #define PORTS_PER_CHAIN (MAX_UDP_PORTS / UDP_HTABLE_SIZE_MIN)
@@ -1461,11 +1463,11 @@ static void udp_rmem_release(struct sock *sk, int size, int partial,
 
 
 	sk->sk_forward_alloc += size;
-	amt = (sk->sk_forward_alloc - partial) & ~(SK_MEM_QUANTUM - 1);
+	amt = (sk->sk_forward_alloc - partial) & ~(PAGE_SIZE - 1);
 	sk->sk_forward_alloc -= amt;
 
 	if (amt)
-		__sk_mem_reduce_allocated(sk, amt >> SK_MEM_QUANTUM_SHIFT);
+		__sk_mem_reduce_allocated(sk, amt >> PAGE_SHIFT);
 
 	atomic_sub(size, &sk->sk_rmem_alloc);
 
@@ -1558,7 +1560,7 @@ int __udp_enqueue_schedule_skb(struct sock *sk, struct sk_buff *skb)
 	spin_lock(&list->lock);
 	if (size >= sk->sk_forward_alloc) {
 		amt = sk_mem_pages(size);
-		delta = amt << SK_MEM_QUANTUM_SHIFT;
+		delta = amt << PAGE_SHIFT;
 		if (!__sk_mem_raise_allocated(sk, delta, amt, SK_MEM_RECV)) {
 			err = -ENOBUFS;
 			spin_unlock(&list->lock);
@@ -2946,6 +2948,8 @@ struct proto udp_prot = {
 	.psock_update_sk_prot	= udp_bpf_update_proto,
 #endif
 	.memory_allocated	= &udp_memory_allocated,
+	.per_cpu_fw_alloc	= &udp_memory_per_cpu_fw_alloc,
+
 	.sysctl_mem		= sysctl_udp_mem,
 	.sysctl_wmem_offset	= offsetof(struct net, ipv4.sysctl_udp_wmem_min),
 	.sysctl_rmem_offset	= offsetof(struct net, ipv4.sysctl_udp_rmem_min),
@@ -3263,8 +3267,8 @@ EXPORT_SYMBOL(udp_flow_hashrnd);
 
 static void __udp_sysctl_init(struct net *net)
 {
-	net->ipv4.sysctl_udp_rmem_min = SK_MEM_QUANTUM;
-	net->ipv4.sysctl_udp_wmem_min = SK_MEM_QUANTUM;
+	net->ipv4.sysctl_udp_rmem_min = PAGE_SIZE;
+	net->ipv4.sysctl_udp_wmem_min = PAGE_SIZE;
 
 #ifdef CONFIG_NET_L3_MASTER_DEV
 	net->ipv4.sysctl_udp_l3mdev_accept = 0;