Merge branch 'tcp_mem_pressure'

Eric Dumazet says:

====================
tcp: better handling of memory pressure

When testing commit 790ba456 ("tcp: set SOCK_NOSPACE under memory
pressure") using edge triggered epoll applications, I found various
issues under memory pressure and thousands of active sockets.

This patch series is a first round to solve these issues, in send
and receive paths. There are probably other fixes needed, but
with this series, my tests now all succeed.

v2: fix typo in "allow one skb to be received per socket under memory pressure",
as spotted by Jason Baron.
====================
Acked-by: Jason Baron <jbaron@akamai.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/net/sock.h

@@ -1368,7 +1368,7 @@ static inline struct inode *SOCK_INODE(struct socket *socket)
  * Functions for memory accounting
  */
 int __sk_mem_schedule(struct sock *sk, int size, int kind);
-void __sk_mem_reclaim(struct sock *sk);
+void __sk_mem_reclaim(struct sock *sk, int amount);
 
 #define SK_MEM_QUANTUM ((int)PAGE_SIZE)
 #define SK_MEM_QUANTUM_SHIFT ilog2(SK_MEM_QUANTUM)
@@ -1409,7 +1409,7 @@ static inline void sk_mem_reclaim(struct sock *sk)
 	if (!sk_has_account(sk))
 		return;
 	if (sk->sk_forward_alloc >= SK_MEM_QUANTUM)
-		__sk_mem_reclaim(sk);
+		__sk_mem_reclaim(sk, sk->sk_forward_alloc);
 }
 
 static inline void sk_mem_reclaim_partial(struct sock *sk)
@@ -1417,7 +1417,7 @@ static inline void sk_mem_reclaim_partial(struct sock *sk)
 	if (!sk_has_account(sk))
 		return;
 	if (sk->sk_forward_alloc > SK_MEM_QUANTUM)
-		__sk_mem_reclaim(sk);
+		__sk_mem_reclaim(sk, sk->sk_forward_alloc - 1);
 }
 
 static inline void sk_mem_charge(struct sock *sk, int size)

include/net/tcp.h

@@ -286,6 +286,14 @@ extern atomic_long_t tcp_memory_allocated;
 extern struct percpu_counter tcp_sockets_allocated;
 extern int tcp_memory_pressure;
 
+/* optimized version of sk_under_memory_pressure() for TCP sockets */
+static inline bool tcp_under_memory_pressure(const struct sock *sk)
+{
+	if (mem_cgroup_sockets_enabled && sk->sk_cgrp)
+		return !!sk->sk_cgrp->memory_pressure;
+
+	return tcp_memory_pressure;
+}
 /*
  * The next routines deal with comparing 32 bit unsigned ints
  * and worry about wraparound (automatic with unsigned arithmetic).
@@ -311,6 +319,8 @@ static inline bool tcp_out_of_memory(struct sock *sk)
 	return false;
 }
 
+void sk_forced_mem_schedule(struct sock *sk, int size);
+
 static inline bool tcp_too_many_orphans(struct sock *sk, int shift)
 {
 	struct percpu_counter *ocp = sk->sk_prot->orphan_count;

net/core/sock.c

@@ -2069,12 +2069,13 @@ EXPORT_SYMBOL(__sk_mem_schedule);
 /**
  *	__sk_reclaim - reclaim memory_allocated
  *	@sk: socket
+ *	@amount: number of bytes (rounded down to a SK_MEM_QUANTUM multiple)
  */
-void __sk_mem_reclaim(struct sock *sk)
+void __sk_mem_reclaim(struct sock *sk, int amount)
 {
-	sk_memory_allocated_sub(sk,
-				sk->sk_forward_alloc >> SK_MEM_QUANTUM_SHIFT);
-	sk->sk_forward_alloc &= SK_MEM_QUANTUM - 1;
+	amount >>= SK_MEM_QUANTUM_SHIFT;
+	sk_memory_allocated_sub(sk, amount);
+	sk->sk_forward_alloc -= amount << SK_MEM_QUANTUM_SHIFT;
 
 	if (sk_under_memory_pressure(sk) &&
 	    (sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)))

net/ipv4/tcp.c

@@ -815,9 +815,20 @@ struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp)
 	/* The TCP header must be at least 32-bit aligned.  */
 	size = ALIGN(size, 4);
 
+	if (unlikely(tcp_under_memory_pressure(sk)))
+		sk_mem_reclaim_partial(sk);
+
 	skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
-	if (skb) {
-		if (sk_wmem_schedule(sk, skb->truesize)) {
+	if (likely(skb)) {
+		bool mem_schedule;
+
+		if (skb_queue_len(&sk->sk_write_queue) == 0) {
+			mem_schedule = true;
+			sk_forced_mem_schedule(sk, skb->truesize);
+		} else {
+			mem_schedule = sk_wmem_schedule(sk, skb->truesize);
+		}
+		if (likely(mem_schedule)) {
 			skb_reserve(skb, sk->sk_prot->max_header);
 			/*
 			 * Make sure that we have exactly size bytes
@@ -3057,11 +3068,12 @@ __setup("thash_entries=", set_thash_entries);
 
 static void __init tcp_init_mem(void)
 {
-	unsigned long limit = nr_free_buffer_pages() / 8;
+	unsigned long limit = nr_free_buffer_pages() / 16;
+
 	limit = max(limit, 128UL);
-	sysctl_tcp_mem[0] = limit / 4 * 3;
-	sysctl_tcp_mem[1] = limit;
-	sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;
+	sysctl_tcp_mem[0] = limit / 4 * 3;		/* 4.68 % */
+	sysctl_tcp_mem[1] = limit;			/* 6.25 % */
+	sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;	/* 9.37 % */
 }
 
 void __init tcp_init(void)

net/ipv4/tcp_input.c

@@ -359,7 +359,7 @@ static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
 	/* Check #1 */
 	if (tp->rcv_ssthresh < tp->window_clamp &&
 	    (int)tp->rcv_ssthresh < tcp_space(sk) &&
-	    !sk_under_memory_pressure(sk)) {
+	    !tcp_under_memory_pressure(sk)) {
 		int incr;
 
 		/* Check #2. Increase window, if skb with such overhead
@@ -446,7 +446,7 @@ static void tcp_clamp_window(struct sock *sk)
 
 	if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
 	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
-	    !sk_under_memory_pressure(sk) &&
+	    !tcp_under_memory_pressure(sk) &&
 	    sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
 		sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
 				    sysctl_tcp_rmem[2]);
@@ -4507,10 +4507,12 @@ static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
 
 		if (eaten <= 0) {
 queue_and_out:
-			if (eaten < 0 &&
-			    tcp_try_rmem_schedule(sk, skb, skb->truesize))
-				goto drop;
-
+			if (eaten < 0) {
+				if (skb_queue_len(&sk->sk_receive_queue) == 0)
+					sk_forced_mem_schedule(sk, skb->truesize);
+				else if (tcp_try_rmem_schedule(sk, skb, skb->truesize))
+					goto drop;
+			}
 			eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
 		}
 		tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq);
@@ -4781,7 +4783,7 @@ static int tcp_prune_queue(struct sock *sk)
 
 	if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
 		tcp_clamp_window(sk);
-	else if (sk_under_memory_pressure(sk))
+	else if (tcp_under_memory_pressure(sk))
 		tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
 
 	tcp_collapse_ofo_queue(sk);
@@ -4825,7 +4827,7 @@ static bool tcp_should_expand_sndbuf(const struct sock *sk)
 		return false;
 
 	/* If we are under global TCP memory pressure, do not expand.  */
-	if (sk_under_memory_pressure(sk))
+	if (tcp_under_memory_pressure(sk))
 		return false;
 
 	/* If we are under soft global TCP memory pressure, do not expand.  */

net/ipv4/tcp_output.c

@@ -2392,7 +2392,7 @@ u32 __tcp_select_window(struct sock *sk)
 	if (free_space < (full_space >> 1)) {
 		icsk->icsk_ack.quick = 0;
 
-		if (sk_under_memory_pressure(sk))
+		if (tcp_under_memory_pressure(sk))
 			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
 					       4U * tp->advmss);
 
@@ -2816,8 +2816,10 @@ void tcp_xmit_retransmit_queue(struct sock *sk)
  * connection tear down and (memory) recovery.
  * Otherwise tcp_send_fin() could be tempted to either delay FIN
  * or even be forced to close flow without any FIN.
+ * In general, we want to allow one skb per socket to avoid hangs
+ * with edge trigger epoll()
  */
-static void sk_forced_wmem_schedule(struct sock *sk, int size)
+void sk_forced_mem_schedule(struct sock *sk, int size)
 {
 	int amt, status;
 
@@ -2841,7 +2843,7 @@ void tcp_send_fin(struct sock *sk)
 	 * Note: in the latter case, FIN packet will be sent after a timeout,
 	 * as TCP stack thinks it has already been transmitted.
 	 */
-	if (tskb && (tcp_send_head(sk) || sk_under_memory_pressure(sk))) {
+	if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
 coalesce:
 		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
 		TCP_SKB_CB(tskb)->end_seq++;
@@ -2864,7 +2866,7 @@ void tcp_send_fin(struct sock *sk)
 			return;
 		}
 		skb_reserve(skb, MAX_TCP_HEADER);
-		sk_forced_wmem_schedule(sk, skb->truesize);
+		sk_forced_mem_schedule(sk, skb->truesize);
 		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
 		tcp_init_nondata_skb(skb, tp->write_seq,
 				     TCPHDR_ACK | TCPHDR_FIN);

net/ipv4/tcp_timer.c

@@ -247,7 +247,7 @@ void tcp_delack_timer_handler(struct sock *sk)
 	}
 
 out:
-	if (sk_under_memory_pressure(sk))
+	if (tcp_under_memory_pressure(sk))
 		sk_mem_reclaim(sk);
 }