Skip to contentSkip to conversation

tcp: exponential backoff in tcp_send_ack()

PATCH In review Eric Dumazet 3 participants 6 messages Sep 30, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover tcp: exponential backoff in tcp_send_ack()

From: Eric Dumazet <[email protected]>

We had outages caused by repeated skb allocation failures in tcp_send_ack()

It is time to add exponential backoff to reduce number of attempts. Before doing so, first patch removes icsk_ack.blocked to make room for a new field (icsk_ack.retry)

Eric Dumazet (2):
inet: remove icsk_ack.blocked
tcp: add exponential backoff in __tcp_send_ack()

 include/net/inet_connection_sock.h |  5 +++--
 net/dccp/timer.c                   |  1 -
 net/ipv4/inet_connection_sock.c    |  2 +-
 net/ipv4/tcp.c                     |  6 ++----
 net/ipv4/tcp_output.c              | 18 ++++++++++--------
 net/ipv4/tcp_timer.c               |  1 -
 6 files changed, 16 insertions(+), 17 deletions(-)
Signature · 1 line
2.28.0.806.g8561365e88-goog
#1 lore raw

PATCH 1/2 inet: remove icsk_ack.blocked

From: Eric Dumazet <[email protected]>

TCP has been using it to work around the possibility of tcp_delack_timer() finding the socket owned by user.

After commit 6f458dfb4092 ("tcp: improve latencies of timer triggered events") we added TCP_DELACK_TIMER_DEFERRED atomic bit for more immediate recovery, so we can get rid of icsk_ack.blocked

This frees space that following patch will reuse.

---
 include/net/inet_connection_sock.h | 4 ++--
 net/dccp/timer.c                   | 1 -
 net/ipv4/inet_connection_sock.c    | 2 +-
 net/ipv4/tcp.c                     | 6 ++----
 net/ipv4/tcp_output.c              | 7 ++-----
 net/ipv4/tcp_timer.c               | 1 -
 6 files changed, 7 insertions(+), 14 deletions(-)
include/net/inet_connection_sock.h +2 -2
index dc763ca9413cc9c6279a59f9d1776cf2dbb1e853..79875f976190750819948425e63dd0309c699050 100644
@@ -110,7 +110,7 @@ struct inet_connection_sock {
 		__u8		  pending;	 /* ACK is pending			   */
 		__u8		  quick;	 /* Scheduled number of quick acks	   */
 		__u8		  pingpong;	 /* The session is interactive		   */
-		__u8		  blocked;	 /* Delayed ACK was blocked by socket lock */
+		/* one byte hole. */
 		__u32		  ato;		 /* Predicted tick of soft clock	   */
 		unsigned long	  timeout;	 /* Currently scheduled timeout		   */
 		__u32		  lrcvtime;	 /* timestamp of last received data packet */
@@ -198,7 +198,7 @@ static inline void inet_csk_clear_xmit_timer(struct sock *sk, const int what)
 		sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 #endif
 	} else if (what == ICSK_TIME_DACK) {
-		icsk->icsk_ack.blocked = icsk->icsk_ack.pending = 0;
+		icsk->icsk_ack.pending = 0;
 #ifdef INET_CSK_CLEAR_TIMERS
 		sk_stop_timer(sk, &icsk->icsk_delack_timer);
 #endif
net/dccp/timer.c +0 -1
index 927c796d76825439a35c4deb3fb2e45e4313f9b3..a934d293237366aeca87bd3c32241880639291c5 100644
@@ -176,7 +176,6 @@ static void dccp_delack_timer(struct timer_list *t)
 	bh_lock_sock(sk);
 	if (sock_owned_by_user(sk)) {
 		/* Try again later. */
-		icsk->icsk_ack.blocked = 1;
 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
 		sk_reset_timer(sk, &icsk->icsk_delack_timer,
 			       jiffies + TCP_DELACK_MIN);
net/ipv4/inet_connection_sock.c +1 -1
index b457dd2d6c75b2f63bc7849474ac909adb14d603..4148f5f78f313cde1e0596b9eb3696df16e3f990 100644
@@ -564,7 +564,7 @@ void inet_csk_clear_xmit_timers(struct sock *sk)
 {
 	struct inet_connection_sock *icsk = inet_csk(sk);
 
-	icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
+	icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 
 	sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 	sk_stop_timer(sk, &icsk->icsk_delack_timer);
net/ipv4/tcp.c +2 -4
index 2a8bfa89a5159837e3687e4e0f8cddba7fe54899..ed2805564424a90f003eed867bbed7f5ac4ae833 100644
@@ -1538,10 +1538,8 @@ void tcp_cleanup_rbuf(struct sock *sk, int copied)
 
 	if (inet_csk_ack_scheduled(sk)) {
 		const struct inet_connection_sock *icsk = inet_csk(sk);
-		   /* Delayed ACKs frequently hit locked sockets during bulk
-		    * receive. */
-		if (icsk->icsk_ack.blocked ||
-		    /* Once-per-two-segments ACK was not sent by tcp_input.c */
+
+		if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
 		    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
 		    /*
 		     * If this read emptied read buffer, we send ACK, if
net/ipv4/tcp_output.c +2 -5
index 386978dcd318d84af486d0d1a5bb1786f4a493cf..6bd4e383030ea20441332a30e98fbda8cd90f84a 100644
@@ -3911,11 +3911,8 @@ void tcp_send_delayed_ack(struct sock *sk)
 
 	/* Use new timeout only if there wasn't a older one earlier. */
 	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
-		/* If delack timer was blocked or is about to expire,
-		 * send ACK now.
-		 */
-		if (icsk->icsk_ack.blocked ||
-		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
+		/* If delack timer is about to expire, send ACK now. */
+		if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
 			tcp_send_ack(sk);
 			return;
 		}
net/ipv4/tcp_timer.c +0 -1
index 0c08c420fbc21a98dedf72148ea2a6f85bf3ff7a..6c62b9ea1320d9bbd26ed86b9f41de02fee6c491 100644
@@ -331,7 +331,6 @@ static void tcp_delack_timer(struct timer_list *t)
 	if (!sock_owned_by_user(sk)) {
 		tcp_delack_timer_handler(sk);
 	} else {
-		icsk->icsk_ack.blocked = 1;
 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
 		/* deleguate our work to tcp_release_cb() */
 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
Signature · 1 line
2.28.0.806.g8561365e88-goog
Acked-by on v1 1/2
#2 lore raw

Re: [PATCH net-next 1/2] inet: remove icsk_ack.blocked

Quoted 6 lines · Eric Dumazet

From: Eric Dumazet <[email protected]>

TCP has been using it to work around the possibility of tcp_delack_timer() finding the socket owned by user.

Quoted 7 lines

After commit 6f458dfb4092 ("tcp: improve latencies of timer triggered events") we added TCP_DELACK_TIMER_DEFERRED atomic bit for more immediate recovery, so we can get rid of icsk_ack.blocked

This frees space that following patch will reuse.

Quoted 106 lines
---
 include/net/inet_connection_sock.h | 4 ++--
 net/dccp/timer.c                   | 1 -
 net/ipv4/inet_connection_sock.c    | 2 +-
 net/ipv4/tcp.c                     | 6 ++----
 net/ipv4/tcp_output.c              | 7 ++-----
 net/ipv4/tcp_timer.c               | 1 -
 6 files changed, 7 insertions(+), 14 deletions(-)
include/net/inet_connection_sock.h +2 -2
index dc763ca9413cc9c6279a59f9d1776cf2dbb1e853..79875f976190750819948425e63dd0309c699050 100644
@@ -110,7 +110,7 @@ struct inet_connection_sock {
                __u8              pending;       /* ACK is pending                         */
                __u8              quick;         /* Scheduled number of quick acks         */
                __u8              pingpong;      /* The session is interactive             */
-               __u8              blocked;       /* Delayed ACK was blocked by socket lock */
+               /* one byte hole. */
                __u32             ato;           /* Predicted tick of soft clock           */
                unsigned long     timeout;       /* Currently scheduled timeout            */
                __u32             lrcvtime;      /* timestamp of last received data packet */
@@ -198,7 +198,7 @@ static inline void inet_csk_clear_xmit_timer(struct sock *sk, const int what)
                sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 #endif
        } else if (what == ICSK_TIME_DACK) {
-               icsk->icsk_ack.blocked = icsk->icsk_ack.pending = 0;
+               icsk->icsk_ack.pending = 0;
 #ifdef INET_CSK_CLEAR_TIMERS
                sk_stop_timer(sk, &icsk->icsk_delack_timer);
 #endif
net/dccp/timer.c +0 -1
index 927c796d76825439a35c4deb3fb2e45e4313f9b3..a934d293237366aeca87bd3c32241880639291c5 100644
@@ -176,7 +176,6 @@ static void dccp_delack_timer(struct timer_list *t)
        bh_lock_sock(sk);
        if (sock_owned_by_user(sk)) {
                /* Try again later. */
-               icsk->icsk_ack.blocked = 1;
                __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
                sk_reset_timer(sk, &icsk->icsk_delack_timer,
                               jiffies + TCP_DELACK_MIN);
net/ipv4/inet_connection_sock.c +1 -1
index b457dd2d6c75b2f63bc7849474ac909adb14d603..4148f5f78f313cde1e0596b9eb3696df16e3f990 100644
@@ -564,7 +564,7 @@ void inet_csk_clear_xmit_timers(struct sock *sk)
 {
        struct inet_connection_sock *icsk = inet_csk(sk);
 
-       icsk->icsk_pending = icsk->icsk_ack.pending = icsk->icsk_ack.blocked = 0;
+       icsk->icsk_pending = icsk->icsk_ack.pending = 0;
 
        sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
        sk_stop_timer(sk, &icsk->icsk_delack_timer);
net/ipv4/tcp.c +2 -4
index 2a8bfa89a5159837e3687e4e0f8cddba7fe54899..ed2805564424a90f003eed867bbed7f5ac4ae833 100644
@@ -1538,10 +1538,8 @@ void tcp_cleanup_rbuf(struct sock *sk, int copied)
 
        if (inet_csk_ack_scheduled(sk)) {
                const struct inet_connection_sock *icsk = inet_csk(sk);
-                  /* Delayed ACKs frequently hit locked sockets during bulk
-                   * receive. */
-               if (icsk->icsk_ack.blocked ||
-                   /* Once-per-two-segments ACK was not sent by tcp_input.c */
+
+               if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
                    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
                    /*
                     * If this read emptied read buffer, we send ACK, if
net/ipv4/tcp_output.c +2 -5
index 386978dcd318d84af486d0d1a5bb1786f4a493cf..6bd4e383030ea20441332a30e98fbda8cd90f84a 100644
@@ -3911,11 +3911,8 @@ void tcp_send_delayed_ack(struct sock *sk)
 
        /* Use new timeout only if there wasn't a older one earlier. */
        if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
-               /* If delack timer was blocked or is about to expire,
-                * send ACK now.
-                */
-               if (icsk->icsk_ack.blocked ||
-                   time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
+               /* If delack timer is about to expire, send ACK now. */
+               if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
                        tcp_send_ack(sk);
                        return;
                }
net/ipv4/tcp_timer.c +0 -2
index 0c08c420fbc21a98dedf72148ea2a6f85bf3ff7a..6c62b9ea1320d9bbd26ed86b9f41de02fee6c491 100644
@@ -331,7 +331,6 @@ static void tcp_delack_timer(struct timer_list *t)
        if (!sock_owned_by_user(sk)) {
                tcp_delack_timer_handler(sk);
        } else {
-               icsk->icsk_ack.blocked = 1;
                __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
                /* deleguate our work to tcp_release_cb() */
                if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
--

2.28.0.806.g8561365e88-goog

#3 lore raw

PATCH 2/2 tcp: add exponential backoff in __tcp_send_ack()

In reply to Eric Dumazet

From: Eric Dumazet <[email protected]>

Whenever host is under very high memory pressure, __tcp_send_ack() skb allocation fails, and we setup a 200 ms (TCP_DELACK_MAX) timer before retrying.

On hosts with high number of TCP sockets, we can spend considerable amount of cpu cycles in these attempts, add high pressure on various spinlocks in mm-layer, ultimately blocking threads attempting to free space from making any progress.

This patch adds standard exponential backoff to avoid adding fuel to the fire.

---
 include/net/inet_connection_sock.h |  3 ++-
 net/ipv4/tcp_output.c              | 11 ++++++++---
 2 files changed, 10 insertions(+), 4 deletions(-)
include/net/inet_connection_sock.h +2 -1
index 79875f976190750819948425e63dd0309c699050..7338b3865a2a3d278dc27c0167bba1b966bbda9f 100644
@@ -110,7 +110,7 @@ struct inet_connection_sock {
 		__u8		  pending;	 /* ACK is pending			   */
 		__u8		  quick;	 /* Scheduled number of quick acks	   */
 		__u8		  pingpong;	 /* The session is interactive		   */
-		/* one byte hole. */
+		__u8		  retry;	 /* Number of attempts			   */
 		__u32		  ato;		 /* Predicted tick of soft clock	   */
 		unsigned long	  timeout;	 /* Currently scheduled timeout		   */
 		__u32		  lrcvtime;	 /* timestamp of last received data packet */
@@ -199,6 +199,7 @@ static inline void inet_csk_clear_xmit_timer(struct sock *sk, const int what)
 #endif
 	} else if (what == ICSK_TIME_DACK) {
 		icsk->icsk_ack.pending = 0;
+		icsk->icsk_ack.retry = 0;
 #ifdef INET_CSK_CLEAR_TIMERS
 		sk_stop_timer(sk, &icsk->icsk_delack_timer);
 #endif
net/ipv4/tcp_output.c +8 -3
index 6bd4e383030ea20441332a30e98fbda8cd90f84a..bf48cd73e96787a14b6f9af8beddb1067a7cb8dc 100644
@@ -3941,10 +3941,15 @@ void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
 	buff = alloc_skb(MAX_TCP_HEADER,
 			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
 	if (unlikely(!buff)) {
+		struct inet_connection_sock *icsk = inet_csk(sk);
+		unsigned long delay;
+
+		delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
+		if (delay < TCP_RTO_MAX)
+			icsk->icsk_ack.retry++;
 		inet_csk_schedule_ack(sk);
-		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
-		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
-					  TCP_DELACK_MAX, TCP_RTO_MAX);
+		icsk->icsk_ack.ato = TCP_ATO_MIN;
+		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
 		return;
 	}
 
Signature · 1 line
2.28.0.806.g8561365e88-goog
Acked-by on v1 2/2
#4 lore raw

Re: [PATCH net-next 2/2] tcp: add exponential backoff in __tcp_send_ack()

Quoted 13 lines · Eric Dumazet

From: Eric Dumazet <[email protected]>

Whenever host is under very high memory pressure, __tcp_send_ack() skb allocation fails, and we setup a 200 ms (TCP_DELACK_MAX) timer before retrying.

On hosts with high number of TCP sockets, we can spend considerable amount of cpu cycles in these attempts, add high pressure on various spinlocks in mm-layer, ultimately blocking threads attempting to free space from making any progress.

Quoted 4 lines

This patch adds standard exponential backoff to avoid adding fuel to the fire.

Quoted 52 lines
---
 include/net/inet_connection_sock.h |  3 ++-
 net/ipv4/tcp_output.c              | 11 ++++++++---
 2 files changed, 10 insertions(+), 4 deletions(-)
include/net/inet_connection_sock.h +2 -1
index 79875f976190750819948425e63dd0309c699050..7338b3865a2a3d278dc27c0167bba1b966bbda9f 100644
@@ -110,7 +110,7 @@ struct inet_connection_sock {
                __u8              pending;       /* ACK is pending                         */
                __u8              quick;         /* Scheduled number of quick acks         */
                __u8              pingpong;      /* The session is interactive             */
-               /* one byte hole. */
+               __u8              retry;         /* Number of attempts                     */
                __u32             ato;           /* Predicted tick of soft clock           */
                unsigned long     timeout;       /* Currently scheduled timeout            */
                __u32             lrcvtime;      /* timestamp of last received data packet */
@@ -199,6 +199,7 @@ static inline void inet_csk_clear_xmit_timer(struct sock *sk, const int what)
 #endif
        } else if (what == ICSK_TIME_DACK) {
                icsk->icsk_ack.pending = 0;
+               icsk->icsk_ack.retry = 0;
 #ifdef INET_CSK_CLEAR_TIMERS
                sk_stop_timer(sk, &icsk->icsk_delack_timer);
 #endif
net/ipv4/tcp_output.c +8 -4
index 6bd4e383030ea20441332a30e98fbda8cd90f84a..bf48cd73e96787a14b6f9af8beddb1067a7cb8dc 100644
@@ -3941,10 +3941,15 @@ void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
        buff = alloc_skb(MAX_TCP_HEADER,
                         sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
        if (unlikely(!buff)) {
+               struct inet_connection_sock *icsk = inet_csk(sk);
+               unsigned long delay;
+
+               delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
+               if (delay < TCP_RTO_MAX)
+                       icsk->icsk_ack.retry++;
                inet_csk_schedule_ack(sk);
-               inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
-               inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
-                                         TCP_DELACK_MAX, TCP_RTO_MAX);
+               icsk->icsk_ack.ato = TCP_ATO_MIN;
+               inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
                return;
        }
 
--

2.28.0.806.g8561365e88-goog

maintainer
#5 lore raw
In reply to Eric Dumazet

From: Eric Dumazet <[email protected]>
Date: Wed, 30 Sep 2020 05:54:55 -0700

Quoted 5 lines

We had outages caused by repeated skb allocation failures in tcp_send_ack()

It is time to add exponential backoff to reduce number of attempts. Before doing so, first patch removes icsk_ack.blocked to make room for a new field (icsk_ack.retry)

Series applied, thanks Eric.