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>

include/net/sock.h

@@ -1254,6 +1254,7 @@ struct proto {
 	void			(*enter_memory_pressure)(struct sock *sk);
 	void			(*leave_memory_pressure)(struct sock *sk);
 	atomic_long_t		*memory_allocated;	/* Current allocated memory. */
+	int  __percpu		*per_cpu_fw_alloc;
 	struct percpu_counter	*sockets_allocated;	/* Current number of sockets. */
 
 	/*
@@ -1396,22 +1397,48 @@ static inline bool sk_under_memory_pressure(const struct sock *sk)
 	return !!*sk->sk_prot->memory_pressure;
 }
 
+static inline long
+proto_memory_allocated(const struct proto *prot)
+{
+	return max(0L, atomic_long_read(prot->memory_allocated));
+}
+
 static inline long
 sk_memory_allocated(const struct sock *sk)
 {
-	return atomic_long_read(sk->sk_prot->memory_allocated);
+	return proto_memory_allocated(sk->sk_prot);
 }
 
+/* 1 MB per cpu, in page units */
+#define SK_MEMORY_PCPU_RESERVE (1 << (20 - PAGE_SHIFT))
+
 static inline long
 sk_memory_allocated_add(struct sock *sk, int amt)
 {
-	return atomic_long_add_return(amt, sk->sk_prot->memory_allocated);
+	int local_reserve;
+
+	preempt_disable();
+	local_reserve = __this_cpu_add_return(*sk->sk_prot->per_cpu_fw_alloc, amt);
+	if (local_reserve >= SK_MEMORY_PCPU_RESERVE) {
+		__this_cpu_sub(*sk->sk_prot->per_cpu_fw_alloc, local_reserve);
+		atomic_long_add(local_reserve, sk->sk_prot->memory_allocated);
+	}
+	preempt_enable();
+	return sk_memory_allocated(sk);
 }
 
 static inline void
 sk_memory_allocated_sub(struct sock *sk, int amt)
 {
-	atomic_long_sub(amt, sk->sk_prot->memory_allocated);
+	int local_reserve;
+
+	preempt_disable();
+	local_reserve = __this_cpu_sub_return(*sk->sk_prot->per_cpu_fw_alloc, amt);
+	if (local_reserve <= -SK_MEMORY_PCPU_RESERVE) {
+		__this_cpu_sub(*sk->sk_prot->per_cpu_fw_alloc, local_reserve);
+		atomic_long_add(local_reserve, sk->sk_prot->memory_allocated);
+	}
+	preempt_enable();
 }
 
 #define SK_ALLOC_PERCPU_COUNTER_BATCH 16
@@ -1440,12 +1467,6 @@ proto_sockets_allocated_sum_positive(struct proto *prot)
 	return percpu_counter_sum_positive(prot->sockets_allocated);
 }
 
-static inline long
-proto_memory_allocated(struct proto *prot)
-{
-	return atomic_long_read(prot->memory_allocated);
-}
-
 static inline bool
 proto_memory_pressure(struct proto *prot)
 {
@@ -1532,30 +1553,18 @@ int __sk_mem_schedule(struct sock *sk, int size, int kind);
 void __sk_mem_reduce_allocated(struct sock *sk, int amount);
 void __sk_mem_reclaim(struct sock *sk, int amount);
 
-/* We used to have PAGE_SIZE here, but systems with 64KB pages
- * do not necessarily have 16x time more memory than 4KB ones.
- */
-#define SK_MEM_QUANTUM 4096
-#define SK_MEM_QUANTUM_SHIFT ilog2(SK_MEM_QUANTUM)
 #define SK_MEM_SEND	0
 #define SK_MEM_RECV	1
 
-/* sysctl_mem values are in pages, we convert them in SK_MEM_QUANTUM units */
+/* sysctl_mem values are in pages */
 static inline long sk_prot_mem_limits(const struct sock *sk, int index)
 {
-	long val = sk->sk_prot->sysctl_mem[index];
-
-#if PAGE_SIZE > SK_MEM_QUANTUM
-	val <<= PAGE_SHIFT - SK_MEM_QUANTUM_SHIFT;
-#elif PAGE_SIZE < SK_MEM_QUANTUM
-	val >>= SK_MEM_QUANTUM_SHIFT - PAGE_SHIFT;
-#endif
-	return val;
+	return sk->sk_prot->sysctl_mem[index];
 }
 
 static inline int sk_mem_pages(int amt)
 {
-	return (amt + SK_MEM_QUANTUM - 1) >> SK_MEM_QUANTUM_SHIFT;
+	return (amt + PAGE_SIZE - 1) >> PAGE_SHIFT;
 }
 
 static inline bool sk_has_account(struct sock *sk)
@@ -1566,19 +1575,23 @@ static inline bool sk_has_account(struct sock *sk)
 
 static inline bool sk_wmem_schedule(struct sock *sk, int size)
 {
+	int delta;
+
 	if (!sk_has_account(sk))
 		return true;
-	return size <= sk->sk_forward_alloc ||
-		__sk_mem_schedule(sk, size, SK_MEM_SEND);
+	delta = size - sk->sk_forward_alloc;
+	return delta <= 0 || __sk_mem_schedule(sk, delta, SK_MEM_SEND);
 }
 
 static inline bool
 sk_rmem_schedule(struct sock *sk, struct sk_buff *skb, int size)
 {
+	int delta;
+
 	if (!sk_has_account(sk))
 		return true;
-	return size <= sk->sk_forward_alloc ||
-		__sk_mem_schedule(sk, size, SK_MEM_RECV) ||
+	delta = size - sk->sk_forward_alloc;
+	return delta <= 0 || __sk_mem_schedule(sk, delta, SK_MEM_RECV) ||
 		skb_pfmemalloc(skb);
 }
 
@@ -1604,7 +1617,7 @@ static inline void sk_mem_reclaim(struct sock *sk)
 
 	reclaimable = sk->sk_forward_alloc - sk_unused_reserved_mem(sk);
 
-	if (reclaimable >= SK_MEM_QUANTUM)
+	if (reclaimable >= (int)PAGE_SIZE)
 		__sk_mem_reclaim(sk, reclaimable);
 }
 
@@ -1614,49 +1627,24 @@ static inline void sk_mem_reclaim_final(struct sock *sk)
 	sk_mem_reclaim(sk);
 }
 
-static inline void sk_mem_reclaim_partial(struct sock *sk)
-{
-	int reclaimable;
-
-	if (!sk_has_account(sk))
-		return;
-
-	reclaimable = sk->sk_forward_alloc - sk_unused_reserved_mem(sk);
-
-	if (reclaimable > SK_MEM_QUANTUM)
-		__sk_mem_reclaim(sk, reclaimable - 1);
-}
-
 static inline void sk_mem_charge(struct sock *sk, int size)
 {
 	if (!sk_has_account(sk))
 		return;
 	sk->sk_forward_alloc -= size;
 }
 
-/* the following macros control memory reclaiming in sk_mem_uncharge()
+/* the following macros control memory reclaiming in mptcp_rmem_uncharge()
  */
 #define SK_RECLAIM_THRESHOLD	(1 << 21)
 #define SK_RECLAIM_CHUNK	(1 << 20)
 
 static inline void sk_mem_uncharge(struct sock *sk, int size)
 {
-	int reclaimable;
-
 	if (!sk_has_account(sk))
 		return;
 	sk->sk_forward_alloc += size;
-	reclaimable = sk->sk_forward_alloc - sk_unused_reserved_mem(sk);
-
-	/* Avoid a possible overflow.
-	 * TCP send queues can make this happen, if sk_mem_reclaim()
-	 * is not called and more than 2 GBytes are released at once.
-	 *
-	 * If we reach 2 MBytes, reclaim 1 MBytes right now, there is
-	 * no need to hold that much forward allocation anyway.
-	 */
-	if (unlikely(reclaimable >= SK_RECLAIM_THRESHOLD))
-		__sk_mem_reclaim(sk, SK_RECLAIM_CHUNK);
+	sk_mem_reclaim(sk);
 }
 
 /*

include/net/tcp.h

@@ -253,6 +253,8 @@ extern long sysctl_tcp_mem[3];
 #define TCP_RACK_NO_DUPTHRESH    0x4 /* Do not use DUPACK threshold in RACK */
 
 extern atomic_long_t tcp_memory_allocated;
+DECLARE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
+
 extern struct percpu_counter tcp_sockets_allocated;
 extern unsigned long tcp_memory_pressure;
 

include/net/udp.h

@@ -95,6 +95,7 @@ static inline struct udp_hslot *udp_hashslot2(struct udp_table *table,
 extern struct proto udp_prot;
 
 extern atomic_long_t udp_memory_allocated;
+DECLARE_PER_CPU(int, udp_memory_per_cpu_fw_alloc);
 
 /* sysctl variables for udp */
 extern long sysctl_udp_mem[3];

net/core/datagram.c

@@ -320,7 +320,6 @@ EXPORT_SYMBOL(skb_recv_datagram);
 void skb_free_datagram(struct sock *sk, struct sk_buff *skb)
 {
 	consume_skb(skb);
-	sk_mem_reclaim_partial(sk);
 }
 EXPORT_SYMBOL(skb_free_datagram);
 
@@ -336,7 +335,6 @@ void __skb_free_datagram_locked(struct sock *sk, struct sk_buff *skb, int len)
 	slow = lock_sock_fast(sk);
 	sk_peek_offset_bwd(sk, len);
 	skb_orphan(skb);
-	sk_mem_reclaim_partial(sk);
 	unlock_sock_fast(sk, slow);
 
 	/* skb is now orphaned, can be freed outside of locked section */
@@ -396,7 +394,6 @@ int skb_kill_datagram(struct sock *sk, struct sk_buff *skb, unsigned int flags)
 				      NULL);
 
 	kfree_skb(skb);
-	sk_mem_reclaim_partial(sk);
 	return err;
 }
 EXPORT_SYMBOL(skb_kill_datagram);

net/core/sock.c

@@ -991,7 +991,7 @@ EXPORT_SYMBOL(sock_set_mark);
 static void sock_release_reserved_memory(struct sock *sk, int bytes)
 {
 	/* Round down bytes to multiple of pages */
-	bytes &= ~(SK_MEM_QUANTUM - 1);
+	bytes = round_down(bytes, PAGE_SIZE);
 
 	WARN_ON(bytes > sk->sk_reserved_mem);
 	sk->sk_reserved_mem -= bytes;
@@ -1028,9 +1028,9 @@ static int sock_reserve_memory(struct sock *sk, int bytes)
 		mem_cgroup_uncharge_skmem(sk->sk_memcg, pages);
 		return -ENOMEM;
 	}
-	sk->sk_forward_alloc += pages << SK_MEM_QUANTUM_SHIFT;
+	sk->sk_forward_alloc += pages << PAGE_SHIFT;
 
-	sk->sk_reserved_mem += pages << SK_MEM_QUANTUM_SHIFT;
+	sk->sk_reserved_mem += pages << PAGE_SHIFT;
 
 	return 0;
 }
@@ -2987,7 +2987,6 @@ int __sk_mem_raise_allocated(struct sock *sk, int size, int amt, int kind)
 
 	return 0;
 }
-EXPORT_SYMBOL(__sk_mem_raise_allocated);
 
 /**
  *	__sk_mem_schedule - increase sk_forward_alloc and memory_allocated
@@ -3003,10 +3002,10 @@ int __sk_mem_schedule(struct sock *sk, int size, int kind)
 {
 	int ret, amt = sk_mem_pages(size);
 
-	sk->sk_forward_alloc += amt << SK_MEM_QUANTUM_SHIFT;
+	sk->sk_forward_alloc += amt << PAGE_SHIFT;
 	ret = __sk_mem_raise_allocated(sk, size, amt, kind);
 	if (!ret)
-		sk->sk_forward_alloc -= amt << SK_MEM_QUANTUM_SHIFT;
+		sk->sk_forward_alloc -= amt << PAGE_SHIFT;
 	return ret;
 }
 EXPORT_SYMBOL(__sk_mem_schedule);
@@ -3029,17 +3028,16 @@ void __sk_mem_reduce_allocated(struct sock *sk, int amount)
 	    (sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)))
 		sk_leave_memory_pressure(sk);
 }
-EXPORT_SYMBOL(__sk_mem_reduce_allocated);
 
 /**
  *	__sk_mem_reclaim - reclaim sk_forward_alloc and memory_allocated
  *	@sk: socket
- *	@amount: number of bytes (rounded down to a SK_MEM_QUANTUM multiple)
+ *	@amount: number of bytes (rounded down to a PAGE_SIZE multiple)
  */
 void __sk_mem_reclaim(struct sock *sk, int amount)
 {
-	amount >>= SK_MEM_QUANTUM_SHIFT;
-	sk->sk_forward_alloc -= amount << SK_MEM_QUANTUM_SHIFT;
+	amount >>= PAGE_SHIFT;
+	sk->sk_forward_alloc -= amount << PAGE_SHIFT;
 	__sk_mem_reduce_allocated(sk, amount);
 }
 EXPORT_SYMBOL(__sk_mem_reclaim);
@@ -3798,6 +3796,10 @@ int proto_register(struct proto *prot, int alloc_slab)
 		pr_err("%s: missing sysctl_mem\n", prot->name);
 		return -EINVAL;
 	}
+	if (prot->memory_allocated && !prot->per_cpu_fw_alloc) {
+		pr_err("%s: missing per_cpu_fw_alloc\n", prot->name);
+		return -EINVAL;
+	}
 	if (alloc_slab) {
 		prot->slab = kmem_cache_create_usercopy(prot->name,
 					prot->obj_size, 0,

net/decnet/af_decnet.c

@@ -149,6 +149,7 @@ static DEFINE_RWLOCK(dn_hash_lock);
 static struct hlist_head dn_sk_hash[DN_SK_HASH_SIZE];
 static struct hlist_head dn_wild_sk;
 static atomic_long_t decnet_memory_allocated;
+static DEFINE_PER_CPU(int, decnet_memory_per_cpu_fw_alloc);
 
 static int __dn_setsockopt(struct socket *sock, int level, int optname,
 		sockptr_t optval, unsigned int optlen, int flags);
@@ -454,7 +455,10 @@ static struct proto dn_proto = {
 	.owner			= THIS_MODULE,
 	.enter_memory_pressure	= dn_enter_memory_pressure,
 	.memory_pressure	= &dn_memory_pressure,
+
 	.memory_allocated	= &decnet_memory_allocated,
+	.per_cpu_fw_alloc	= &decnet_memory_per_cpu_fw_alloc,
+
 	.sysctl_mem		= sysctl_decnet_mem,
 	.sysctl_wmem		= sysctl_decnet_wmem,
 	.sysctl_rmem		= sysctl_decnet_rmem,

net/ipv4/tcp.c

@@ -294,6 +294,8 @@ EXPORT_SYMBOL(sysctl_tcp_mem);
 
 atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp;	/* Current allocated memory. */
 EXPORT_SYMBOL(tcp_memory_allocated);
+DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
+EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
 
 #if IS_ENABLED(CONFIG_SMC)
 DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
@@ -856,9 +858,6 @@ struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 {
 	struct sk_buff *skb;
 
-	if (unlikely(tcp_under_memory_pressure(sk)))
-		sk_mem_reclaim_partial(sk);
-
 	skb = alloc_skb_fclone(size + MAX_TCP_HEADER, gfp);
 	if (likely(skb)) {
 		bool mem_scheduled;
@@ -2762,8 +2761,6 @@ void __tcp_close(struct sock *sk, long timeout)
 		__kfree_skb(skb);
 	}
 
-	sk_mem_reclaim(sk);
-
 	/* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
 	if (sk->sk_state == TCP_CLOSE)
 		goto adjudge_to_death;
@@ -2871,7 +2868,6 @@ void __tcp_close(struct sock *sk, long timeout)
 		}
 	}
 	if (sk->sk_state != TCP_CLOSE) {
-		sk_mem_reclaim(sk);
 		if (tcp_check_oom(sk, 0)) {
 			tcp_set_state(sk, TCP_CLOSE);
 			tcp_send_active_reset(sk, GFP_ATOMIC);
@@ -2949,7 +2945,6 @@ void tcp_write_queue_purge(struct sock *sk)
 	}
 	tcp_rtx_queue_purge(sk);
 	INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
-	sk_mem_reclaim(sk);
 	tcp_clear_all_retrans_hints(tcp_sk(sk));
 	tcp_sk(sk)->packets_out = 0;
 	inet_csk(sk)->icsk_backoff = 0;
@@ -4661,11 +4656,11 @@ void __init tcp_init(void)
 	max_wshare = min(4UL*1024*1024, limit);
 	max_rshare = min(6UL*1024*1024, limit);
 
-	init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
+	init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
 	init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
 	init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
 
-	init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
+	init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
 	init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
 	init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);