diff mbox series

[net-next,2/2] net: revert "Update RFS target at poll for tcp/udp"

Message ID 20180104024711.257600-2-soheil.kdev@gmail.com
State Accepted, archived
Delegated to: David Miller
Headers show
Series [net-next,1/2] ip: do not set RFS core on error queue reads | expand

Commit Message

Soheil Hassas Yeganeh Jan. 4, 2018, 2:47 a.m. UTC 
From: Soheil Hassas Yeganeh <soheil@google.com>

On multi-threaded processes, one common architecture is to have
one (or a small number of) threads polling sockets, and a
considerably larger pool of threads reading form and writing to the
sockets. When we set RPS core on tcp_poll() or udp_poll() we essentially
steer all packets of all the polled FDs to one (or small number of)
cores, creaing a bottleneck and/or RPS misprediction.

Another common architecture is to shard FDs among threads pinned
to cores. In such a setting, setting RPS core in tcp_poll() and
udp_poll() is redundant because the RFS core is correctly
set in recvmsg and sendmsg.

Thus, revert the following commit:
c3f1dbaf6e28 ("net: Update RFS target at poll for tcp/udp").

Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
---
 net/ipv4/tcp.c | 2 --
 net/ipv4/udp.c | 2 --
 2 files changed, 4 deletions(-)

Comments

David Miller Jan. 5, 2018, 4:15 p.m. UTC | #1 
From: Soheil Hassas Yeganeh <soheil.kdev@gmail.com>
Date: Wed,  3 Jan 2018 21:47:11 -0500

> From: Soheil Hassas Yeganeh <soheil@google.com>
> 
> On multi-threaded processes, one common architecture is to have
> one (or a small number of) threads polling sockets, and a
> considerably larger pool of threads reading form and writing to the
> sockets. When we set RPS core on tcp_poll() or udp_poll() we essentially
> steer all packets of all the polled FDs to one (or small number of)
> cores, creaing a bottleneck and/or RPS misprediction.
> 
> Another common architecture is to shard FDs among threads pinned
> to cores. In such a setting, setting RPS core in tcp_poll() and
> udp_poll() is redundant because the RFS core is correctly
> set in recvmsg and sendmsg.
> 
> Thus, revert the following commit:
> c3f1dbaf6e28 ("net: Update RFS target at poll for tcp/udp").
> 
> Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
> Signed-off-by: Willem de Bruijn <willemb@google.com>
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> Signed-off-by: Neal Cardwell <ncardwell@google.com>

Applied.
diff mbox series

Patch

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 7ac583a2b9fe..f68cb33d50d1 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -498,8 +498,6 @@  unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 	const struct tcp_sock *tp = tcp_sk(sk);
 	int state;
 
-	sock_rps_record_flow(sk);
-
 	sock_poll_wait(file, sk_sleep(sk), wait);
 
 	state = inet_sk_state_load(sk);
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index e9c0d1e1772e..db72619e07e4 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -2490,8 +2490,6 @@  unsigned int udp_poll(struct file *file, struct socket *sock, poll_table *wait)
 	if (!skb_queue_empty(&udp_sk(sk)->reader_queue))
 		mask |= POLLIN | POLLRDNORM;
 
-	sock_rps_record_flow(sk);
-
 	/* Check for false positives due to checksum errors */
 	if ((mask & POLLRDNORM) && !(file->f_flags & O_NONBLOCK) &&
 	    !(sk->sk_shutdown & RCV_SHUTDOWN) && first_packet_length(sk) == -1)