Re: [RFC PATCH 0/2] Allow full bridge configuration via sysfs

Bill Nottingham wrote:

Right now, you can configure most bridge device parameters via sysfs.
However, you cannot either:
- add or remove bridge interfaces
- add or remove physical interfaces from a bridge

The attached patch set rectifies this. With this patch set, brctl
(theoretically) becomes completely optional, much like ifenslave is
now for bonding. (In fact, the idea for this patch, and the syntax
used herein, is inspired by the sysfs bonding configuration.)

Both should use netlink instead of extending their sysfs interfaces.
For bridging I have a patch for the bridge device itself, the API
is so far missing support for adding ports though.

Patrick McHardy (kaber@trash.net) said: 

Bill Nottingham wrote:

quoted

Right now, you can configure most bridge device parameters via sysfs.
However, you cannot either:
- add or remove bridge interfaces
- add or remove physical interfaces from a bridge

The attached patch set rectifies this. With this patch set, brctl
(theoretically) becomes completely optional, much like ifenslave is
now for bonding. (In fact, the idea for this patch, and the syntax
used herein, is inspired by the sysfs bonding configuration.)

Both should use netlink instead of extending their sysfs interfaces.
For bridging I have a patch for the bridge device itself, the API
is so far missing support for adding ports though.

How does that improve the situation for bridge devices? Are all
bridging parameters (forward_delay, stp, etc.) going to be configurable
via netlink, or would we still then have multiple tools/interfaces
to configuration? Also, moving bonding configuration to netlink seems
like a step backwards.

Bill

Bill Nottingham wrote:

Patrick McHardy (kaber@trash.net) said:

quoted

Bill Nottingham wrote:

quoted

Right now, you can configure most bridge device parameters via sysfs.
However, you cannot either:
- add or remove bridge interfaces
- add or remove physical interfaces from a bridge

The attached patch set rectifies this. With this patch set, brctl
(theoretically) becomes completely optional, much like ifenslave is
now for bonding. (In fact, the idea for this patch, and the syntax
used herein, is inspired by the sysfs bonding configuration.)

Both should use netlink instead of extending their sysfs interfaces.
For bridging I have a patch for the bridge device itself, the API
is so far missing support for adding ports though.

How does that improve the situation for bridge devices? Are all
bridging parameters (forward_delay, stp, etc.) going to be configurable
via netlink, or would we still then have multiple tools/interfaces
to configuration?

Of course its all going to be configurable via netlink, otherwise
it really wouldn't make sense.

Also, moving bonding configuration to netlink seems
like a step backwards.

Please read up on what the standard interface for network
configuration is, I'm tired of reiterating this once a week.

Patrick McHardy (kaber@trash.net) said: 

quoted

Also, moving bonding configuration to netlink seems
like a step backwards.

Please read up on what the standard interface for network
configuration is

OK, let's see.

In the code: Hm, no TODO or FIXME.

In the included documentation:

Documentation/networking/bonding.txt:
  Module options, modprobe.conf, or 'distro-specific
  configuration file', ifenslave, or sysfs.

Documentation/networking/ip-sysctl.txt:
  sysctl, obviously.

Documentation/networking/generic_netlink.txt
  Hey, here's netlink! Doucmentation points only to a wiki. Referred
  to by zero other included in-kernel documentation.

Well, that's helpful.

Let's try the OSDL web!

http://www.linuxfoundation.org/en/Net:Bridge
  brctl (which uses ioctl and sysfs). And /etc/net.
  
http://www.linuxfoundation.org/en/Net:Bonding
  Module parameters only, including the lovely 'load driver multiple times'
  method. Doesn't even mention sysfs.

http://www.linuxfoundation.org/en/Net:VLAN
  vconfig
  
I could look at wireless network configuration, but I doubt that's going to
help your argument.

I'm tired of reiterating this once a week.

Well, if the documentation that described this as the standard existed,
or wasn't such crap, perhaps you wouldn't have to.

That being said, how is moving from adding a bonding slave from:
  echo "+eth0" >  /sys/class/net/bond0/bonding/slaves to:
to:
  http://www.linuxfoundation.org/en/Net:Generic_Netlink_HOWTO

a worthwhile improvement for the admin? Let's see, a kernel-userspace
protocol with magic message formats. Hey, we reinvented ioctl!

Why, if netlink is the standard (and it's been around for a long
damn time), was sysfs configuration for bonding added in 2005? Why
was bridge configuration added in 2005, and *extended* in 2006 and
2007? Why were the user-space tools such as brctl ported from ioctl
to sysfs?

Bill

From: Bill Nottingham <redacted>
Date: Mon, 7 Jul 2008 17:34:20 -0400

I could look at wireless network configuration, but I doubt that's going to
help your argument.

Just like any system with age, we have a lot of legacy to
convert over.  But it will happen.

That being said, how is moving from adding a bonding slave from:
  echo "+eth0" >  /sys/class/net/bond0/bonding/slaves to:
to:
  http://www.linuxfoundation.org/en/Net:Generic_Netlink_HOWTO

a worthwhile improvement for the admin? Let's see, a kernel-userspace
protocol with magic message formats. Hey, we reinvented ioctl!

Why, if netlink is the standard (and it's been around for a long
damn time), was sysfs configuration for bonding added in 2005? Why
was bridge configuration added in 2005, and *extended* in 2006 and
2007? Why were the user-space tools such as brctl ported from ioctl
to sysfs?

Because often a lot of shit slips in when someone who understands
the ramifications is too busy or on vacation.

We do want everything to be netlink based.

Why?

Because it means that you can run one monitoring tool to listen
for netlink events and report them to the user for diagnosis.

It means that network configuration events can be sent over
the wire and used remotely at some point.

The latter can never happen as long as we keep adding ad-hoc
config stuff.

On Mon, 07 Jul 2008 14:52:59 -0700 (PDT)
David Miller [off-list ref] wrote:

From: Bill Nottingham <redacted>
Date: Mon, 7 Jul 2008 17:34:20 -0400

quoted

I could look at wireless network configuration, but I doubt that's going to
help your argument.

Just like any system with age, we have a lot of legacy to
convert over.  But it will happen.

quoted

That being said, how is moving from adding a bonding slave from:
  echo "+eth0" >  /sys/class/net/bond0/bonding/slaves to:
to:
  http://www.linuxfoundation.org/en/Net:Generic_Netlink_HOWTO

a worthwhile improvement for the admin? Let's see, a kernel-userspace
protocol with magic message formats. Hey, we reinvented ioctl!

Why, if netlink is the standard (and it's been around for a long
damn time), was sysfs configuration for bonding added in 2005? Why
was bridge configuration added in 2005, and *extended* in 2006 and
2007? Why were the user-space tools such as brctl ported from ioctl
to sysfs?

Because often a lot of shit slips in when someone who understands
the ramifications is too busy or on vacation.

We do want everything to be netlink based.

Why?

Because it means that you can run one monitoring tool to listen
for netlink events and report them to the user for diagnosis.

It means that network configuration events can be sent over
the wire and used remotely at some point.

The latter can never happen as long as we keep adding ad-hoc
config stuff.

There are always historical reasons. In this case it was because I knew
more about sysfs than netlink, and there was no netlink interface for managing
interfaces back in 2005.  Sysfs is okay for simple stuff (set forward-delay to 10seconds),
but it falls down when anything interesting and transactional happens. Think of
sysfs as more an extension of per-device sysctl's or module parameters, rather
than a good configuration interface.

David Miller (davem@davemloft.net) said: 

quoted

Why, if netlink is the standard (and it's been around for a long
damn time), was sysfs configuration for bonding added in 2005? Why
was bridge configuration added in 2005, and *extended* in 2006 and
2007? Why were the user-space tools such as brctl ported from ioctl
to sysfs?

Because often a lot of shit slips in when someone who understands
the ramifications is too busy or on vacation.

Duly noted, will time all patch submissions to land during your
vacations in the future.

More seriously, if there's not a mechanism to prevent ABIs the
kernel doesn't want like this being added, that's a problem.

We do want everything to be netlink based.

Why?

Because it means that you can run one monitoring tool to listen
for netlink events and report them to the user for diagnosis.

It means that network configuration events can be sent over
the wire and used remotely at some point.

The latter can never happen as long as we keep adding ad-hoc
config stuff.

Sure, but it does make them more opaque to the normal user,
leaving them wrapped in the same old ip/brctl/ifenslave/vconfig
tools - for better or worse, people like the discoverability
and obviousness of sysfs.

Bill