Re: de-indirect TCP congestion control

[Stephen Hemminger]

On Mon, 12 Mar 2018 16:05:16 -0400 (EDT)
David Miller  wrote:

> From: Eric Dumazet 
> Date: Mon, 12 Mar 2018 13:03:35 -0700
> 
> > 
> > 
> > On 03/12/2018 12:48 PM, Stephen Hemminger wrote:  
> >> On Mon, 12 Mar 2018 15:04:06 -0400 (EDT)
> >> David Miller  wrote:
> >>   
> >>> From: Stephen Hemminger 
> >>> Date: Mon, 12 Mar 2018 11:45:52 -0700
> >>>  
>  Since indirect calls are expensive, and now even more so, perhaps we
>  should figure out
>  a way to make the default TCP congestion control hooks into direct
>  calls.
>  99% of the users just use the single CC module compiled into the
>  kernel.  
> >>>
> >>> Who is this magic user with only one CC algorithm enabled in their
> >>> kernel?  I want to know who this dude is?
> >>>
> >>> I don't think it's going to help much since people will have I think
> >>> at least two algorithms compiled into nearly everyone's tree.
> >>>
> >>> Distributions will enable everything.
> >>>
> >>> Google is going to have at least two algorithms enabled.
> >>>
> >>> etc. etc. etc.
> >>>
> >>> Getting rid of indirect calls is a fine goal, but the precondition you
> >>> are mentioning to achieve this doesn't seem practical at all.  
> >> What I meant is that kernels with N congestion controls, almost all
> >> traffic
> >> uses the default So that path can be optimized. The example I gave
> >> would
> >> have all the others doing the same indirect call.  
> > 
> > I do not understand. What is default_tcp_ops anyway ?
> > 
> > How changes to /proc/sys/net/ipv4/tcp_congestion_control will impact
> > this ?  
> 
> I'm also confused what is being suggested exactly and how this can
> work. :-)

Trying to directly call the fastpath TCP congestion operations for sockets
using the default.  The problem is how to do this and it gets messy with Kconfig
and CPP to deal with as raw tools. Maybe weak symbols could be used.

Re: de-indirect TCP congestion control

[David Miller]

From: Eric Dumazet 
Date: Mon, 12 Mar 2018 13:03:35 -0700

> 
> 
> On 03/12/2018 12:48 PM, Stephen Hemminger wrote:
>> On Mon, 12 Mar 2018 15:04:06 -0400 (EDT)
>> David Miller  wrote:
>> 
>>> From: Stephen Hemminger 
>>> Date: Mon, 12 Mar 2018 11:45:52 -0700
>>>
 Since indirect calls are expensive, and now even more so, perhaps we
 should figure out
 a way to make the default TCP congestion control hooks into direct
 calls.
 99% of the users just use the single CC module compiled into the
 kernel.
>>>
>>> Who is this magic user with only one CC algorithm enabled in their
>>> kernel?  I want to know who this dude is?
>>>
>>> I don't think it's going to help much since people will have I think
>>> at least two algorithms compiled into nearly everyone's tree.
>>>
>>> Distributions will enable everything.
>>>
>>> Google is going to have at least two algorithms enabled.
>>>
>>> etc. etc. etc.
>>>
>>> Getting rid of indirect calls is a fine goal, but the precondition you
>>> are mentioning to achieve this doesn't seem practical at all.
>> What I meant is that kernels with N congestion controls, almost all
>> traffic
>> uses the default So that path can be optimized. The example I gave
>> would
>> have all the others doing the same indirect call.
> 
> I do not understand. What is default_tcp_ops anyway ?
> 
> How changes to /proc/sys/net/ipv4/tcp_congestion_control will impact
> this ?

I'm also confused what is being suggested exactly and how this can
work. :-)

Re: de-indirect TCP congestion control

[Eric Dumazet]

On 03/12/2018 12:48 PM, Stephen Hemminger wrote:

On Mon, 12 Mar 2018 15:04:06 -0400 (EDT)
David Miller  wrote:


From: Stephen Hemminger 
Date: Mon, 12 Mar 2018 11:45:52 -0700


Since indirect calls are expensive, and now even more so, perhaps we should 
figure out
a way to make the default TCP congestion control hooks into direct calls.
99% of the users just use the single CC module compiled into the kernel.


Who is this magic user with only one CC algorithm enabled in their
kernel?  I want to know who this dude is?

I don't think it's going to help much since people will have I think
at least two algorithms compiled into nearly everyone's tree.

Distributions will enable everything.

Google is going to have at least two algorithms enabled.

etc. etc. etc.

Getting rid of indirect calls is a fine goal, but the precondition you
are mentioning to achieve this doesn't seem practical at all.


What I meant is that kernels with N congestion controls, almost all traffic
uses the default So that path can be optimized. The example I gave would
have all the others doing the same indirect call.



I do not understand. What is default_tcp_ops anyway ?

How changes to /proc/sys/net/ipv4/tcp_congestion_control will impact this ?

Re: de-indirect TCP congestion control

[Stephen Hemminger]

On Mon, 12 Mar 2018 15:04:06 -0400 (EDT)
David Miller  wrote:

> From: Stephen Hemminger 
> Date: Mon, 12 Mar 2018 11:45:52 -0700
> 
> > Since indirect calls are expensive, and now even more so, perhaps we should 
> > figure out
> > a way to make the default TCP congestion control hooks into direct calls.
> > 99% of the users just use the single CC module compiled into the kernel.  
> 
> Who is this magic user with only one CC algorithm enabled in their
> kernel?  I want to know who this dude is?
> 
> I don't think it's going to help much since people will have I think
> at least two algorithms compiled into nearly everyone's tree.
> 
> Distributions will enable everything.
> 
> Google is going to have at least two algorithms enabled.
> 
> etc. etc. etc.
> 
> Getting rid of indirect calls is a fine goal, but the precondition you
> are mentioning to achieve this doesn't seem practical at all.

What I meant is that kernels with N congestion controls, almost all traffic
uses the default So that path can be optimized. The example I gave would
have all the others doing the same indirect call.

Re: de-indirect TCP congestion control

[David Miller]

From: Stephen Hemminger 
Date: Mon, 12 Mar 2018 11:45:52 -0700

> Since indirect calls are expensive, and now even more so, perhaps we should 
> figure out
> a way to make the default TCP congestion control hooks into direct calls.
> 99% of the users just use the single CC module compiled into the kernel.

Who is this magic user with only one CC algorithm enabled in their
kernel?  I want to know who this dude is?

I don't think it's going to help much since people will have I think
at least two algorithms compiled into nearly everyone's tree.

Distributions will enable everything.

Google is going to have at least two algorithms enabled.

etc. etc. etc.

Getting rid of indirect calls is a fine goal, but the precondition you
are mentioning to achieve this doesn't seem practical at all.

de-indirect TCP congestion control

[Stephen Hemminger]

Since indirect calls are expensive, and now even more so, perhaps we should 
figure out
a way to make the default TCP congestion control hooks into direct calls.
99% of the users just use the single CC module compiled into the kernel.

Use some macros (or other stuff) to change indirect call to direct
calls. So that code like:

static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
{
const struct inet_connection_sock *icsk = inet_csk(sk);

icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);

to:

static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
{
const struct tcp_congestion_ops *ops = inet_csk(sk)->icsk_ca_ops;

if (ops == default_tcp_ops)
default_tcp_cong_avoid(s, ack, acked);
else
icsk->icsk_ca_ops->cong_avoid(sk, ack, acked);


The use macros and/or compiler symbols so the builtin (non-module) congestion
control is always default_tcp_XXX.