[PATCH v2] tcp: Disable Nagle's algorithm (set TCP_NODELAY) on all sockets

newer
[PATCH] vhost_user: Drop packet...

Stefano Brivio

20 Jan 2025 20 Jan '25

6:28 p.m.

Following up on 725acd111ba3 ("tcp_splice: Set (again) TCP_NODELAY on both sides"), David argues that, in general, we don't know what kind of TCP traffic we're dealing with, on any side or path. TCP segments might have been delivered to our socket with a PSH flag, but we don't have a way to know about it. Similarly, the guest might send us segments with PSH or URG set, but we don't know if we should generally TCP_CORK sockets and uncork on those flags, because that would assume they're running a Linux kernel (and a particular version of it) matching the kernel that delivers outbound packets for us. Given that we can't make any assumption and everything might very well be interactive traffic, disable Nagle's algorithm on all non-spliced sockets as well. After all, John Nagle himself is nowadays recommending that delayed ACKs should never be enabled together with his algorithm, but we don't have a practical way to ensure that our environment is free from delayed ACKs (TCP_QUICKACK is not really usable for this purpose): https://news.ycombinator.com/item?id=34180239 Suggested-by: David Gibson Signed-off-by: Stefano Brivio --- v2: Set TCP_NODELAY on inbound socket after accept4(), not on the listening sockets, and change failure message to debug() instead of trace() tcp.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+) diff --git a/tcp.c b/tcp.c index a012b81..4d6a6b3 100644 --- a/tcp.c +++ b/tcp.c @@ -756,6 +756,19 @@ static void tcp_sock_set_bufsize(const struct ctx *c, int s) trace("TCP: failed to set SO_SNDBUF to %i", v); } +/** + * tcp_sock_set_nodelay() - Set TCP_NODELAY option (disable Nagle's algorithm) + * @s: Socket, can be -1 to avoid check in the caller + */ +static void tcp_sock_set_nodelay(int s) +{ + if (s == -1) + return; + + if (setsockopt(s, SOL_TCP, TCP_NODELAY, &((int){ 1 }), sizeof(int))) + debug("TCP: failed to set TCP_NODELAY on socket %i", s); +} + /** * tcp_update_csum() - Calculate TCP checksum * @psum: Unfolded partial checksum of the IPv4 or IPv6 pseudo-header @@ -1285,6 +1298,7 @@ static int tcp_conn_new_sock(const struct ctx *c, sa_family_t af) return -errno; tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s); return s; } @@ -2058,6 +2072,7 @@ void tcp_listen_handler(const struct ctx *c, union epoll_ref ref, goto cancel; tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s); /* FIXME: When listening port has a specific bound address, record that * as our address -- 2.43.0

Show replies by date

David Gibson

21 Jan 21 Jan

4:01 a.m.

On Mon, Jan 20, 2025 at 06:28:16PM +0100, Stefano Brivio wrote:

...

Following up on 725acd111ba3 ("tcp_splice: Set (again) TCP_NODELAY on both sides"), David argues that, in general, we don't know what kind of TCP traffic we're dealing with, on any side or path.

TCP segments might have been delivered to our socket with a PSH flag, but we don't have a way to know about it.

Similarly, the guest might send us segments with PSH or URG set, but we don't know if we should generally TCP_CORK sockets and uncork on those flags, because that would assume they're running a Linux kernel (and a particular version of it) matching the kernel that delivers outbound packets for us.

Given that we can't make any assumption and everything might very well be interactive traffic, disable Nagle's algorithm on all non-spliced sockets as well.

After all, John Nagle himself is nowadays recommending that delayed ACKs should never be enabled together with his algorithm, but we don't have a practical way to ensure that our environment is free from delayed ACKs (TCP_QUICKACK is not really usable for this purpose):

https://news.ycombinator.com/item?id=34180239

Suggested-by: David Gibson Signed-off-by: Stefano Brivio --- v2: Set TCP_NODELAY on inbound socket after accept4(), not on the listening sockets, and change failure message to debug() instead of trace()

Uh.. you've moved it to after accept() for incoming connections, but it looks like you've accidentally removed the call for outgoing connections (tcp_conn_new_sock()).

...

tcp.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)

diff --git a/tcp.c b/tcp.c index a012b81..4d6a6b3 100644 --- a/tcp.c +++ b/tcp.c @@ -756,6 +756,19 @@ static void tcp_sock_set_bufsize(const struct ctx *c, int s) trace("TCP: failed to set SO_SNDBUF to %i", v); }

+/** + * tcp_sock_set_nodelay() - Set TCP_NODELAY option (disable Nagle's algorithm) + * @s: Socket, can be -1 to avoid check in the caller + */ +static void tcp_sock_set_nodelay(int s) +{ + if (s == -1) + return; + + if (setsockopt(s, SOL_TCP, TCP_NODELAY, &((int){ 1 }), sizeof(int))) + debug("TCP: failed to set TCP_NODELAY on socket %i", s); +} + /** * tcp_update_csum() - Calculate TCP checksum * @psum: Unfolded partial checksum of the IPv4 or IPv6 pseudo-header @@ -1285,6 +1298,7 @@ static int tcp_conn_new_sock(const struct ctx *c, sa_family_t af) return -errno;

tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s);

return s; } @@ -2058,6 +2072,7 @@ void tcp_listen_handler(const struct ctx *c, union epoll_ref ref, goto cancel;

tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s);

/* FIXME: When listening port has a specific bound address, record that * as our address

-- David Gibson (he or they) | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you, not the other way | around. http://www.ozlabs.org/~dgibson

Stefano Brivio

9:40 a.m.

On Tue, 21 Jan 2025 13:31:03 +1030 David Gibson wrote:

...

On Mon, Jan 20, 2025 at 06:28:16PM +0100, Stefano Brivio wrote:

...
Following up on 725acd111ba3 ("tcp_splice: Set (again) TCP_NODELAY on both sides"), David argues that, in general, we don't know what kind of TCP traffic we're dealing with, on any side or path.

TCP segments might have been delivered to our socket with a PSH flag, but we don't have a way to know about it.

Similarly, the guest might send us segments with PSH or URG set, but we don't know if we should generally TCP_CORK sockets and uncork on those flags, because that would assume they're running a Linux kernel (and a particular version of it) matching the kernel that delivers outbound packets for us.

Given that we can't make any assumption and everything might very well be interactive traffic, disable Nagle's algorithm on all non-spliced sockets as well.

After all, John Nagle himself is nowadays recommending that delayed ACKs should never be enabled together with his algorithm, but we don't have a practical way to ensure that our environment is free from delayed ACKs (TCP_QUICKACK is not really usable for this purpose):

https://news.ycombinator.com/item?id=34180239

Suggested-by: David Gibson Signed-off-by: Stefano Brivio --- v2: Set TCP_NODELAY on inbound socket after accept4(), not on the listening sockets, and change failure message to debug() instead of trace()

Uh.. you've moved it to after accept() for incoming connections, but it looks like you've accidentally removed the call for outgoing connections (tcp_conn_new_sock()).

Uh... no? It's here:

...

...
tcp.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)

diff --git a/tcp.c b/tcp.c index a012b81..4d6a6b3 100644 --- a/tcp.c +++ b/tcp.c @@ -756,6 +756,19 @@ static void tcp_sock_set_bufsize(const struct ctx *c, int s) trace("TCP: failed to set SO_SNDBUF to %i", v); }

+/** + * tcp_sock_set_nodelay() - Set TCP_NODELAY option (disable Nagle's algorithm) + * @s: Socket, can be -1 to avoid check in the caller + */ +static void tcp_sock_set_nodelay(int s) +{ + if (s == -1) + return; + + if (setsockopt(s, SOL_TCP, TCP_NODELAY, &((int){ 1 }), sizeof(int))) + debug("TCP: failed to set TCP_NODELAY on socket %i", s); +} + /** * tcp_update_csum() - Calculate TCP checksum * @psum: Unfolded partial checksum of the IPv4 or IPv6 pseudo-header @@ -1285,6 +1298,7 @@ static int tcp_conn_new_sock(const struct ctx *c, sa_family_t af) ^^^^^^^^^^^^^^^^^

...

...
return -errno;

tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s);

^^^^^^^^^^^^^^^^^^^^^^^^

...am I missing something?

...

...
return s; } @@ -2058,6 +2072,7 @@ void tcp_listen_handler(const struct ctx *c, union epoll_ref ref, goto cancel;

tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s);

/* FIXME: When listening port has a specific bound address, record that * as our address

-- Stefano

David Gibson

1:42 p.m.

On Tue, Jan 21, 2025 at 09:40:02AM +0100, Stefano Brivio wrote:

...

On Tue, 21 Jan 2025 13:31:03 +1030 David Gibson wrote:

...
On Mon, Jan 20, 2025 at 06:28:16PM +0100, Stefano Brivio wrote:

...
Following up on 725acd111ba3 ("tcp_splice: Set (again) TCP_NODELAY on both sides"), David argues that, in general, we don't know what kind of TCP traffic we're dealing with, on any side or path.

TCP segments might have been delivered to our socket with a PSH flag, but we don't have a way to know about it.

Similarly, the guest might send us segments with PSH or URG set, but we don't know if we should generally TCP_CORK sockets and uncork on those flags, because that would assume they're running a Linux kernel (and a particular version of it) matching the kernel that delivers outbound packets for us.

Given that we can't make any assumption and everything might very well be interactive traffic, disable Nagle's algorithm on all non-spliced sockets as well.

After all, John Nagle himself is nowadays recommending that delayed ACKs should never be enabled together with his algorithm, but we don't have a practical way to ensure that our environment is free from delayed ACKs (TCP_QUICKACK is not really usable for this purpose):

https://news.ycombinator.com/item?id=34180239

Suggested-by: David Gibson Signed-off-by: Stefano Brivio --- v2: Set TCP_NODELAY on inbound socket after accept4(), not on the listening sockets, and change failure message to debug() instead of trace()

Uh.. you've moved it to after accept() for incoming connections, but it looks like you've accidentally removed the call for outgoing connections (tcp_conn_new_sock()).

Uh... no? It's here:

...
...
tcp.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)

diff --git a/tcp.c b/tcp.c index a012b81..4d6a6b3 100644 --- a/tcp.c +++ b/tcp.c @@ -756,6 +756,19 @@ static void tcp_sock_set_bufsize(const struct ctx *c, int s) trace("TCP: failed to set SO_SNDBUF to %i", v); }

+/** + * tcp_sock_set_nodelay() - Set TCP_NODELAY option (disable Nagle's algorithm) + * @s: Socket, can be -1 to avoid check in the caller + */ +static void tcp_sock_set_nodelay(int s) +{ + if (s == -1) + return; + + if (setsockopt(s, SOL_TCP, TCP_NODELAY, &((int){ 1 }), sizeof(int))) + debug("TCP: failed to set TCP_NODELAY on socket %i", s); +} + /** * tcp_update_csum() - Calculate TCP checksum * @psum: Unfolded partial checksum of the IPv4 or IPv6 pseudo-header @@ -1285,6 +1298,7 @@ static int tcp_conn_new_sock(const struct ctx *c, sa_family_t af) ^^^^^^^^^^^^^^^^^

...
...
return -errno;

tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s);

^^^^^^^^^^^^^^^^^^^^^^^^

...am I missing something?

Oops, no, I am. The page break fell at an inconvenient spot which made my eye skate over the relevant hunk, sorry.

...

...
...
return s; } @@ -2058,6 +2072,7 @@ void tcp_listen_handler(const struct ctx *c, union epoll_ref ref, goto cancel;

tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s);

/* FIXME: When listening port has a specific bound address, record that * as our address

-- David Gibson (he or they) | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you, not the other way | around. http://www.ozlabs.org/~dgibson

David Gibson

1:42 p.m.

On Mon, Jan 20, 2025 at 06:28:16PM +0100, Stefano Brivio wrote:

...

Following up on 725acd111ba3 ("tcp_splice: Set (again) TCP_NODELAY on both sides"), David argues that, in general, we don't know what kind of TCP traffic we're dealing with, on any side or path.

TCP segments might have been delivered to our socket with a PSH flag, but we don't have a way to know about it.

Similarly, the guest might send us segments with PSH or URG set, but we don't know if we should generally TCP_CORK sockets and uncork on those flags, because that would assume they're running a Linux kernel (and a particular version of it) matching the kernel that delivers outbound packets for us.

Given that we can't make any assumption and everything might very well be interactive traffic, disable Nagle's algorithm on all non-spliced sockets as well.

After all, John Nagle himself is nowadays recommending that delayed ACKs should never be enabled together with his algorithm, but we don't have a practical way to ensure that our environment is free from delayed ACKs (TCP_QUICKACK is not really usable for this purpose):

https://news.ycombinator.com/item?id=34180239

Suggested-by: David Gibson Signed-off-by: Stefano Brivio

Reviewed-by: David Gibson

...

--- v2: Set TCP_NODELAY on inbound socket after accept4(), not on the listening sockets, and change failure message to debug() instead of trace()

tcp.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)

diff --git a/tcp.c b/tcp.c index a012b81..4d6a6b3 100644 --- a/tcp.c +++ b/tcp.c @@ -756,6 +756,19 @@ static void tcp_sock_set_bufsize(const struct ctx *c, int s) trace("TCP: failed to set SO_SNDBUF to %i", v); }

+/** + * tcp_sock_set_nodelay() - Set TCP_NODELAY option (disable Nagle's algorithm) + * @s: Socket, can be -1 to avoid check in the caller + */ +static void tcp_sock_set_nodelay(int s) +{ + if (s == -1) + return; + + if (setsockopt(s, SOL_TCP, TCP_NODELAY, &((int){ 1 }), sizeof(int))) + debug("TCP: failed to set TCP_NODELAY on socket %i", s); +} + /** * tcp_update_csum() - Calculate TCP checksum * @psum: Unfolded partial checksum of the IPv4 or IPv6 pseudo-header @@ -1285,6 +1298,7 @@ static int tcp_conn_new_sock(const struct ctx *c, sa_family_t af) return -errno;

tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s);

return s; } @@ -2058,6 +2072,7 @@ void tcp_listen_handler(const struct ctx *c, union epoll_ref ref, goto cancel;

tcp_sock_set_bufsize(c, s); + tcp_sock_set_nodelay(s);

/* FIXME: When listening port has a specific bound address, record that * as our address

-- David Gibson (he or they) | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you, not the other way | around. http://www.ozlabs.org/~dgibson

112

Age (days ago)

113

Last active (days ago)

List overview

Download

4 comments

2 participants

participants (2)

David Gibson
Stefano Brivio