Skip to contentSkip to conversation

tcp: enable mid stream window clamp

PATCH Superseded Neil Spring 3 participants 5 messages Aug 19, 2021 – Aug 25, 2021

Subsystems (3)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 tcp: enable mid stream window clamp

The TCP_WINDOW_CLAMP socket option is defined in tcp(7) to "Bound the size of the advertised window to this value." Window clamping is distributed across two variables, window_clamp ("Maximal window to advertise" in tcp.h) and rcv_ssthresh ("Current window clamp").

This patch updates the function where the window clamp is set to also reduce the current window clamp, rcv_sshthresh, if needed. With this, setting the TCP_WINDOW_CLAMP option has the documented effect of limiting the window.

---
v2: - fix email formatting

 net/ipv4/tcp.c | 2 ++
 1 file changed, 2 insertions(+)
net/ipv4/tcp.c +2 -0
index f931def6302e..2dc6212d5888 100644
@@ -3338,6 +3338,8 @@ int tcp_set_window_clamp(struct sock *sk, int val)
 	} else {
 		tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
 			SOCK_MIN_RCVBUF / 2 : val;
+		tp->rcv_ssthresh = min(tp->rcv_ssthresh,
+				       tp->window_clamp);
 	}
 	return 0;
 }
Signature · 1 line
2.30.2
#1 lore raw
Quoted 22 lines · Neil Spring

The TCP_WINDOW_CLAMP socket option is defined in tcp(7) to "Bound the size of the advertised window to this value." Window clamping is distributed across two variables, window_clamp ("Maximal window to advertise" in tcp.h) and rcv_ssthresh ("Current window clamp").

This patch updates the function where the window clamp is set to also reduce the current window clamp, rcv_sshthresh, if needed. With this, setting the TCP_WINDOW_CLAMP option has the documented effect of limiting the window.

---
v2: - fix email formatting

 net/ipv4/tcp.c | 2 ++
 1 file changed, 2 insertions(+)
net/ipv4/tcp.c +0 -0
index f931def6302e..2dc6212d5888 100644

net/ipv4/tcp.c int tcp_set_window_clamp(struct sock *sk, int val) · line 3342

Quoted 5 lines
net/ipv4/tcp.c +2 -0
@@ -3338,6 +3338,8 @@ int tcp_set_window_clamp(struct sock *sk, int val)
 	} else {
 		tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
 			SOCK_MIN_RCVBUF / 2 : val;
+		tp->rcv_ssthresh = min(tp->rcv_ssthresh,
+				       tp->window_clamp);

This fits in a single line I think.

		tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);

net/ipv4/tcp.c int tcp_set_window_clamp(struct sock *sk, int val) · line 3345

Quoted 4 lines
net/ipv4/tcp.c +0 -0
 	}
 	return 0;
 }

Hi Neil

Can you provide a packetdrill test showing the what the new expected behavior is ?

It is not really clear why you need this.

Also if we are unable to increase tp->rcv_ssthresh, this means the following sequence will not work as we would expect :

+0 setsockopt(5, IPPROTO_TCP, TCP_WINDOW_CLAMP, [10000], 4) = 0 +0 setsockopt(5, IPPROTO_TCP, TCP_WINDOW_CLAMP, [100000], 4) = 0

Thanks.

#2 lore raw
Quoted 18 lines · Eric Dumazet
Quoted 17 lines · Neil Spring

The TCP_WINDOW_CLAMP socket option is defined in tcp(7) to "Bound the size of the advertised window to this value."  Window clamping is distributed across two variables, window_clamp ("Maximal window to advertise" in tcp.h) and rcv_ssthresh ("Current window clamp").

This patch updates the function where the window clamp is set to also reduce the current window clamp, rcv_sshthresh, if needed.  With this, setting the TCP_WINDOW_CLAMP option has the documented effect of limiting the window.

---
v2: - fix email formatting

  net/ipv4/tcp.c | 2 ++

  1 file changed, 2 insertions(+)
Quoted 13 lines
Quoted 10 lines
net/ipv4/tcp.c +0 -0
index f931def6302e..2dc6212d5888 100644
@@ -3338,6 +3338,8 @@ int tcp_set_window_clamp(struct sock *sk, int val)
         } else {

                tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ? SOCK_MIN_RCVBUF / 2 : val;
+             tp->rcv_ssthresh = min(tp->rcv_ssthresh, +                                    tp->window_clamp);

This fits in a single line I think.
                tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);

I'll fix in v3 in a moment, thanks!

Quoted 8 lines
Quoted 4 lines

        }
        return 0;
  }

Hi Neil

Can you provide a packetdrill test showing the what the new expected behavior is ?

Sure. I submitted a pull request on packetdrill -
https://github.com/google/packetdrill/pull/56 - to document the intended behavior.

Quoted 1 line

It is not really clear why you need this.

The observation is that this option is currently ineffective at limiting once the connection is exchanging data. I interpret this as a result of only looking at the window clamp when growing rcv_ssthresh, not as a key to reduce this limit.

The packetdrill example will fail at the point where an ack should have a reduced window due to the clamp.

Quoted 5 lines

Also if we are unable to increase tp->rcv_ssthresh, this means the following sequence will not work as we would expect :

+0 setsockopt(5, IPPROTO_TCP, TCP_WINDOW_CLAMP, [10000], 4) = 0 +0 setsockopt(5, IPPROTO_TCP, TCP_WINDOW_CLAMP, [100000], 4) = 0

The packetdrill shows that raising the window clamp works:
tcp_grow_window takes over and raises the window quickly, but I'll add a specific test for this sequence (with no intervening data) to confirm.

Quoted 1 line

Thanks.

Thanks Eric!

-neil

#3 lore raw
Quoted 62 lines · Neil Spring
Quoted 30 lines · Eric Dumazet
Quoted 22 lines · Neil Spring

The TCP_WINDOW_CLAMP socket option is defined in tcp(7) to "Bound the size of the advertised window to this value." Window clamping is distributed across two variables, window_clamp ("Maximal window to advertise" in tcp.h) and rcv_ssthresh ("Current window clamp").

This patch updates the function where the window clamp is set to also reduce the current window clamp, rcv_sshthresh, if needed. With this, setting the TCP_WINDOW_CLAMP option has the documented effect of limiting the window.

---
v2: - fix email formatting

 net/ipv4/tcp.c | 2 ++
 1 file changed, 2 insertions(+)
net/ipv4/tcp.c +0 -0
index f931def6302e..2dc6212d5888 100644

net/ipv4/tcp.c int tcp_set_window_clamp(struct sock *sk, int val) · line 3342

Quoted 5 lines
net/ipv4/tcp.c +2 -0
@@ -3338,6 +3338,8 @@ int tcp_set_window_clamp(struct sock *sk, int val)
       } else {
               tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
                       SOCK_MIN_RCVBUF / 2 : val;
+             tp->rcv_ssthresh = min(tp->rcv_ssthresh,
+                                    tp->window_clamp);

This fits in a single line I think.

Quoted 1 line
                tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);

I'll fix in v3 in a moment, thanks!

Quoted 8 lines

net/ipv4/tcp.c int tcp_set_window_clamp(struct sock *sk, int val) · line 3345

Quoted 4 lines
net/ipv4/tcp.c +0 -0
       }
       return 0;
 }

Hi Neil

Can you provide a packetdrill test showing the what the new expected behavior is ?

Sure. I submitted a pull request on packetdrill -
https://github.com/google/packetdrill/pull/56 - to document the intended behavior.

Quoted 1 line

It is not really clear why you need this.

The observation is that this option is currently ineffective at limiting once the connection is exchanging data. I interpret this as a result of only looking at the window clamp when growing rcv_ssthresh, not as a key to reduce this limit.

The packetdrill example will fail at the point where an ack should have a reduced window due to the clamp.

Quoted 2 lines

Also if we are unable to increase tp->rcv_ssthresh, this means the following sequence will not work as we would expect :

Quoted 6 lines
Quoted 2 lines

+0 setsockopt(5, IPPROTO_TCP, TCP_WINDOW_CLAMP, [10000], 4) = 0 +0 setsockopt(5, IPPROTO_TCP, TCP_WINDOW_CLAMP, [100000], 4) = 0

The packetdrill shows that raising the window clamp works:
tcp_grow_window takes over and raises the window quickly, but I'll add a specific test for this sequence (with no intervening data) to confirm.

Sure, raising the window clamping is working (even before your patch)

But after your patch, rcv_ssthresh will still be 10000, instead of something maybe bigger ?

Quoted 6 lines
Quoted 2 lines

Thanks.

Thanks Eric!

-neil

#4 lore raw
Quoted 66 lines · Eric Dumazet
Quoted 65 lines · Neil Spring
Quoted 30 lines · Eric Dumazet
Quoted 22 lines · Neil Spring

The TCP_WINDOW_CLAMP socket option is defined in tcp(7) to "Bound the size of the advertised window to this value." Window clamping is distributed across two variables, window_clamp ("Maximal window to advertise" in tcp.h) and rcv_ssthresh ("Current window clamp").

This patch updates the function where the window clamp is set to also reduce the current window clamp, rcv_sshthresh, if needed. With this, setting the TCP_WINDOW_CLAMP option has the documented effect of limiting the window.

---
v2: - fix email formatting

 net/ipv4/tcp.c | 2 ++
 1 file changed, 2 insertions(+)
net/ipv4/tcp.c +0 -0
index f931def6302e..2dc6212d5888 100644

net/ipv4/tcp.c int tcp_set_window_clamp(struct sock *sk, int val) · line 3342

Quoted 5 lines
net/ipv4/tcp.c +2 -0
@@ -3338,6 +3338,8 @@ int tcp_set_window_clamp(struct sock *sk, int val)
       } else {
               tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
                       SOCK_MIN_RCVBUF / 2 : val;
+             tp->rcv_ssthresh = min(tp->rcv_ssthresh,
+                                    tp->window_clamp);

This fits in a single line I think.

Quoted 1 line
                tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);

I'll fix in v3 in a moment, thanks!

Quoted 8 lines

net/ipv4/tcp.c int tcp_set_window_clamp(struct sock *sk, int val) · line 3345

Quoted 4 lines
net/ipv4/tcp.c +0 -0
       }
       return 0;
 }

Hi Neil

Can you provide a packetdrill test showing the what the new expected behavior is ?

Sure. I submitted a pull request on packetdrill -
https://github.com/google/packetdrill/pull/56 - to document the intended behavior.

Quoted 1 line

It is not really clear why you need this.

The observation is that this option is currently ineffective at limiting once the connection is exchanging data. I interpret this as a result of only looking at the window clamp when growing rcv_ssthresh, not as a key to reduce this limit.

The packetdrill example will fail at the point where an ack should have a reduced window due to the clamp.

Quoted 5 lines

Also if we are unable to increase tp->rcv_ssthresh, this means the following sequence will not work as we would expect :

+0 setsockopt(5, IPPROTO_TCP, TCP_WINDOW_CLAMP, [10000], 4) = 0 +0 setsockopt(5, IPPROTO_TCP, TCP_WINDOW_CLAMP, [100000], 4) = 0

Quoted 8 lines
Quoted 3 lines

The packetdrill shows that raising the window clamp works:
tcp_grow_window takes over and raises the window quickly, but I'll add a specific test for this sequence (with no intervening data) to confirm.

Sure, raising the window clamping is working (even before your patch)

But after your patch, rcv_ssthresh will still be 10000, instead of something maybe bigger ?

Ahh, I think I see, thanks for your patience in repeating the idea. I'll alter the patch to set rcv_ssthresh to:

rcv_ssthresh = min(rcv_wnd, window_clamp)

Walking through your example, assuming the host just advertised 50000 bytes, it will raise rcv_ssthresh immediately to 50000 when raising the window_clamp field from 10000 to 100000.

I think this is consistent with what rcv_ssthresh is meant to do, but I could see a case for eagerly increasing to anywhere between rcv_wnd
+ rcv_mss to rcv_wnd * 2. Happy to update again if that seems useful.

I've updated the packetdrill pull request with the new behavior.

Quoted 6 lines
Quoted 6 lines
Quoted 2 lines

Thanks.

Thanks Eric!

-neil