[Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Ralph Sennhauser]

Dear List, Thomas, Gregory

Commit cb4f71c4298853db0c6751b1209e4535956f136c changes the order of
the network interfaces for armada-38x. As a special exception to the
"order by register address" rule says the comment in the dtsi. The
commit messages even calls it a violation.

I can't remember having owned a device were the internal and external
numbering actually matched, so the important bit for me is whatever the
order is it should remain constant.

Distributions like OpenWrt have to fix their code when moving from 4.4
currently to past 4.6 [1]. Worse the so called "wrong ordering" is
actually documented [2]. There are likely more victims out there. In
case it goes unnoticed by the distribution the users lan becomes wan
and vice versa.

It's unfortunate that this commit is already in stable 4.6 and 4.7.

What is your take on this? Should the commit be reverted?

Regards
Ralph

[1]
https://github.com/openwrt/openwrt/blob/master/target/linux/mvebu/base-files/etc/board.d/02_network
[2] https://wiki.openwrt.org/toh/linksys/wrt1x00ac_series#switch_layout

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Thomas Petazzoni]

Hello,

On Sun, 21 Aug 2016 15:11:58 +0200, Ralph Sennhauser wrote:

> Commit cb4f71c4298853db0c6751b1209e4535956f136c changes the order of
> the network interfaces for armada-38x. As a special exception to the
> "order by register address" rule says the comment in the dtsi. The
> commit messages even calls it a violation.
> 
> I can't remember having owned a device were the internal and external
> numbering actually matched, so the important bit for me is whatever the
> order is it should remain constant.
> 
> Distributions like OpenWrt have to fix their code when moving from 4.4
> currently to past 4.6 [1]. Worse the so called "wrong ordering" is
> actually documented [2]. There are likely more victims out there. In
> case it goes unnoticed by the distribution the users lan becomes wan
> and vice versa.

We had many many users getting confused by the fact that the order of
the network interfaces was inverted compared to:

 * The board documentations
 * The U-Boot numbering
 * And to a lesser extent, the vendor kernel

So having things match the documentation numbering was in our opinion
the least confusing thing moving forward. We should have done it
earlier, but we thought that the rule "order by register address" was a
very strong rule.

At this point, reverting the patch is I believe cause more harm than
good. It's going to re-confuse again people.

Regarding the fact that the "wrong numbering if actually documented" is
a fairly specious argument. The OpenWRT Wiki has never been an official
documentation of any sort. I see it as a much more important aspect
that the numbering of the Ethernet interfaces matches the user manual
Marvell provides with its development and evaluation boards. The
OpenWRT Wiki can certainly be fixed accordingly.

Best regards,

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Lennart Sorensen]

On Wed, Aug 24, 2016 at 04:50:11PM +0200, Thomas Petazzoni wrote:
> We had many many users getting confused by the fact that the order of
> the network interfaces was inverted compared to:
> 
>  * The board documentations
>  * The U-Boot numbering
>  * And to a lesser extent, the vendor kernel
> 
> So having things match the documentation numbering was in our opinion
> the least confusing thing moving forward. We should have done it
> earlier, but we thought that the rule "order by register address" was a
> very strong rule.
> 
> At this point, reverting the patch is I believe cause more harm than
> good. It's going to re-confuse again people.

Well wouldn't it only affect the tiny proportian that have moved to a >4.4
kernel?  Looking at LEDE, openWrt, etc, it seems very few have so far.

I don't think that's a good argument.

> Regarding the fact that the "wrong numbering if actually documented" is
> a fairly specious argument. The OpenWRT Wiki has never been an official
> documentation of any sort. I see it as a much more important aspect
> that the numbering of the Ethernet interfaces matches the user manual
> Marvell provides with its development and evaluation boards. The
> OpenWRT Wiki can certainly be fixed accordingly.

That could be a good argument.

Of course this would not be the first linux release to change network
device ordering.  It has happened many times before.

Unfortunately I don't see the eth ports on the marvell in udev, so I
don't know if udev could even help fix the names based on the address
of the port.

On any system I have been part of making in the past, I always added an
ifname attribute to the dtb and made the driver use that.  This problem
is just too silly to put up with.  There needs to be some sane way to
control the names of the interfaces.

-- 
Len Sorensen

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Thomas Petazzoni]

Hello,

On Wed, 24 Aug 2016 12:19:33 -0400, Lennart Sorensen wrote:

> > So having things match the documentation numbering was in our opinion
> > the least confusing thing moving forward. We should have done it
> > earlier, but we thought that the rule "order by register address" was a
> > very strong rule.
> > 
> > At this point, reverting the patch is I believe cause more harm than
> > good. It's going to re-confuse again people.  
> 
> Well wouldn't it only affect the tiny proportian that have moved to a >4.4
> kernel?  Looking at LEDE, openWrt, etc, it seems very few have so far.

Well, just like the for the documentation aspect, you're seeing this
from the OpenWRT/LEDE angle only. Other people are using plenty of
other things.

We knew it would potentially cause some breakage, so it was a
trade-off. I still believe the new arrangement is better, and you've so
far been the only person reporting an issue with this (compared to
numerous people being confused by the original ordering problem).

> > Regarding the fact that the "wrong numbering if actually documented" is
> > a fairly specious argument. The OpenWRT Wiki has never been an official
> > documentation of any sort. I see it as a much more important aspect
> > that the numbering of the Ethernet interfaces matches the user manual
> > Marvell provides with its development and evaluation boards. The
> > OpenWRT Wiki can certainly be fixed accordingly.  
> 
> That could be a good argument.
> 
> Of course this would not be the first linux release to change network
> device ordering.  It has happened many times before.
> 
> Unfortunately I don't see the eth ports on the marvell in udev, so I
> don't know if udev could even help fix the names based on the address
> of the port.

This is more problematic, and something to be investigated. I don't
immediately see why the Marvell network interfaces would not be visible
by udev, but I haven't tested.

> On any system I have been part of making in the past, I always added an
> ifname attribute to the dtb and made the driver use that.  This problem
> is just too silly to put up with.  There needs to be some sane way to
> control the names of the interfaces.

The solution of adding an alias in the DT, and using that to name
network interfaces has already been proposed multiple times, but has
been rejected by the networking maintainer, who suggests to use
userspace tools (udev or something else) to rename network interfaces.
See for example https://patchwork.kernel.org/patch/4122441/, which was
proposed by my colleague Boris Brezillon.

You can always try to propose again a solution, but I doubt it will be
accepted.

Best regards,

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Ralph Sennhauser]

On Wed, 24 Aug 2016 16:50:11 +0200
Thomas Petazzoni <thomas.petazzoni@free-electrons.com> wrote:

> Hello,
> 
> On Sun, 21 Aug 2016 15:11:58 +0200, Ralph Sennhauser wrote:
> 
> > Commit cb4f71c4298853db0c6751b1209e4535956f136c changes the order of
> > the network interfaces for armada-38x. As a special exception to the
> > "order by register address" rule says the comment in the dtsi. The
> > commit messages even calls it a violation.
> > 
> > I can't remember having owned a device were the internal and
> > external numbering actually matched, so the important bit for me is
> > whatever the order is it should remain constant.
> > 
> > Distributions like OpenWrt have to fix their code when moving from
> > 4.4 currently to past 4.6 [1]. Worse the so called "wrong ordering"
> > is actually documented [2]. There are likely more victims out
> > there. In case it goes unnoticed by the distribution the users lan
> > becomes wan and vice versa.  
> 
> We had many many users getting confused by the fact that the order of
> the network interfaces was inverted compared to:
> 
>  * The board documentations
>  * The U-Boot numbering
>  * And to a lesser extent, the vendor kernel
> 

And in how many places this discrepancy was documented? You won't be
able to update them all. Mailing lists, blogs, fora posts and what ever
else. I'd say the damage is done and can't be fixed by simply changing
the order now.

> So having things match the documentation numbering was in our opinion
> the least confusing thing moving forward. We should have done it
> earlier, but we thought that the rule "order by register address" was
> a very strong rule.
> 

Whether strong rule or not, obviously it went in so bending the rule is
at least accepted. It's not what I have an issue with but with the
reordering at this point.

> At this point, reverting the patch is I believe cause more harm than
> good. It's going to re-confuse again people.

I'm clearly not convinced reverting would do more harm than good,
otherwise I wouldn't have brought it up.

> 
> Regarding the fact that the "wrong numbering if actually documented"
> is a fairly specious argument. The OpenWRT Wiki has never been an
> official documentation of any sort. I see it as a much more important
> aspect that the numbering of the Ethernet interfaces matches the user
> manual Marvell provides with its development and evaluation boards.
> The OpenWRT Wiki can certainly be fixed accordingly.
> 

It's not about being able to fix the code or documentation but about
having to do so in the first place.

The nice thing about having the order in the dtb I thought was that it
wont ever change.

Going forward, as we disagree and it's basically a political decision,
whom do we ask to rule here? Linus?

Best regards
Ralph

> Best regards,
> 
> Thomas

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Lennart Sorensen]

On Wed, Aug 24, 2016 at 06:43:34PM +0200, Thomas Petazzoni wrote:
> Well, just like the for the documentation aspect, you're seeing this
> from the OpenWRT/LEDE angle only. Other people are using plenty of
> other things.
> 
> We knew it would potentially cause some breakage, so it was a
> trade-off. I still believe the new arrangement is better, and you've so
> far been the only person reporting an issue with this (compared to
> numerous people being confused by the original ordering problem).

I didn't report it.  I just commented.  I will be affected when LEDE
does move their kernel past 4.4 I suppose, but I imagine something will
be done to deal with it.

I have run into issues with things being reordered on other systems
though, and it sure is annoying, especially when the new behaviour
removes the ability to control the order that was previously there
(that is not what is happening here of course).

> This is more problematic, and something to be investigated. I don't
> immediately see why the Marvell network interfaces would not be visible
> by udev, but I haven't tested.

Well certainly doing udevtrigger -n -v I see no ethernet devices (but
lots of other things).  Looking in sysfs it is possible to dereive which
ethX belongs to which port based on the directory names, but that's
probably not the most convinient manner to deal with it.

> The solution of adding an alias in the DT, and using that to name
> network interfaces has already been proposed multiple times, but has
> been rejected by the networking maintainer, who suggests to use
> userspace tools (udev or something else) to rename network interfaces.
> See for example https://patchwork.kernel.org/patch/4122441/, which was
> proposed by my colleague Boris Brezillon.

Sure, although it seems many embedded systems would rather avoid udev
(even more so after systemd seems to have taken it over), and in this
case I get the impression so far that udev may not even be able to help
on this system.  (although maybe I just did it wrong).

> You can always try to propose again a solution, but I doubt it will be
> accepted.

Yeah me too, hence why I can't be bothered to try.  I have accepted
there are some things the maintainers won't fix so I have a few patches 
I maintain myself and can keep doing so.  They are tiny after all.

-- 
Len Sorensen

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Lennart Sorensen]

On Wed, Aug 24, 2016 at 07:10:04PM +0200, Ralph Sennhauser wrote:
> And in how many places this discrepancy was documented? You won't be
> able to update them all. Mailing lists, blogs, fora posts and what ever
> else. I'd say the damage is done and can't be fixed by simply changing
> the order now.
> 
> Whether strong rule or not, obviously it went in so bending the rule is
> at least accepted. It's not what I have an issue with but with the
> reordering at this point.
> 
> I'm clearly not convinced reverting would do more harm than good,
> otherwise I wouldn't have brought it up.

I would agree, although if you see below I think it maybe something
that user space will have to figure out a way to deal with sooner or
later anyhow.

> It's not about being able to fix the code or documentation but about
> having to do so in the first place.
> 
> The nice thing about having the order in the dtb I thought was that it
> wont ever change.

I wonder, if someone was to build a box with this cpu, and add a PCIe
network device, which order would they get probed in?  Any chance the
PCIe could grab eth0 before the mvneta devices get probed?

I also would not count on this not changing in the future potentially
by accident.  The mmc block devices used to get probed in DTB order, then
that changed at some point and reordering the dtb no longer controlled it,
rather it was whichever device seemed to finish fastests, and then later
that was changed to using the controller's opinion of the device number
(which at least made it predicable again, even if not controllable from
the dtb).

So at the moment the dtb order controls ethernet probe order.  Well if
other things have stopped doing that, why expect ethernet won't some
day stop doing that too?

"Fixing" the port numbering by reordering is hence a hack as far as I
am concerned and not certain to stay working long term.

> Going forward, as we disagree and it's basically a political decision,
> whom do we ask to rule here? Linus?

-- 
Len Sorensen

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Thomas Petazzoni]

Hello,

On Wed, 24 Aug 2016 14:07:27 -0400, Lennart Sorensen wrote:

> > The nice thing about having the order in the dtb I thought was that it
> > wont ever change.  
> 
> I wonder, if someone was to build a box with this cpu, and add a PCIe
> network device, which order would they get probed in?  Any chance the
> PCIe could grab eth0 before the mvneta devices get probed?

Depends on the network driver I believe. But with an e1000e NIC plugged
in a PCIe slot, it indeed gets assigned as eth0, and the internal
mvneta devices get eth1, eth2, etc.

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Thomas Petazzoni]

Hello,

On Wed, 24 Aug 2016 19:10:04 +0200, Ralph Sennhauser wrote:

> Going forward, as we disagree and it's basically a political decision,
> whom do we ask to rule here? Linus?

I don't think Linus will care about random issues on a random
platform :-)

The people who can take this decision are rather the maintainers of the
platform itself, or possibly the arm-soc maintainers if you still don't
like what the platform maintainers decided.

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Lennart Sorensen]

On Wed, Aug 24, 2016 at 08:14:44PM +0200, Thomas Petazzoni wrote:
> Depends on the network driver I believe. But with an e1000e NIC plugged
> in a PCIe slot, it indeed gets assigned as eth0, and the internal
> mvneta devices get eth1, eth2, etc.

Which of course means the change does not actually ensure the port
ordering matches the marvell documentation or u-boot.  It only handles
the relative order of the ports.  For now.

So since it doesn't actually work, maybe reverting it so it no longer
violates the dtb ordeting rule makes sense.

Doesn't mean openwrt/lede/etc don't have to deal with the ordering in
the future if async probing takes off.

-- 
Len Sorensen

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Lennart Sorensen]

On Wed, Aug 24, 2016 at 01:10:23PM -0400, Lennart Sorensen wrote:
> Well certainly doing udevtrigger -n -v I see no ethernet devices (but
> lots of other things).  Looking in sysfs it is possible to dereive which
> ethX belongs to which port based on the directory names, but that's
> probably not the most convinient manner to deal with it.

OK, udev DOES work:

# udevadm info -p /sys/class/net/eth0
P: /devices/platform/soc/soc:internal-regs/f1034000.ethernet/net/eth0
E: DEVPATH=/devices/platform/soc/soc:internal-regs/f1034000.ethernet/net/eth0
E: IFINDEX=2
E: INTERFACE=eth0
E: SUBSYSTEM=net

# udevadm info -p /sys/class/net/eth1
P: /devices/platform/soc/soc:internal-regs/f1070000.ethernet/net/eth1
E: DEVPATH=/devices/platform/soc/soc:internal-regs/f1070000.ethernet/net/eth1
E: IFINDEX=3
E: INTERFACE=eth1
E: SUBSYSTEM=net

So it isn't hopeless.

-- 
Len Sorensen

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Thomas Petazzoni]

Hello,

On Wed, 24 Aug 2016 14:27:58 -0400, Lennart Sorensen wrote:
> On Wed, Aug 24, 2016 at 08:14:44PM +0200, Thomas Petazzoni wrote:
> > Depends on the network driver I believe. But with an e1000e NIC plugged
> > in a PCIe slot, it indeed gets assigned as eth0, and the internal
> > mvneta devices get eth1, eth2, etc.  
> 
> Which of course means the change does not actually ensure the port
> ordering matches the marvell documentation or u-boot.  It only handles
> the relative order of the ports.  For now.

Correct.

> So since it doesn't actually work, maybe reverting it so it no longer
> violates the dtb ordeting rule makes sense.

I'll let the platform maintainers decide what's the least
intrusive/problematic option. Both solutions have drawbacks, so it's
really a "political" decision to make here.

> Doesn't mean openwrt/lede/etc don't have to deal with the ordering in
> the future if async probing takes off.

Not only async probing, but also PCIe devices, as you mentioned
earlier :-)

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Ralph Sennhauser]

Hi Thomas

On Wed, 24 Aug 2016 20:15:31 +0200
Thomas Petazzoni <thomas.petazzoni@free-electrons.com> wrote:

> Hello,
> 
> On Wed, 24 Aug 2016 19:10:04 +0200, Ralph Sennhauser wrote:
> 
> > Going forward, as we disagree and it's basically a political
> > decision, whom do we ask to rule here? Linus?  
> 
> I don't think Linus will care about random issues on a random
> platform :-)
> 

Probably not about the ordering of the interfaces per se. Let me ask
instead do you think he would sign off on that commit?

What I do not yet understand is why you not simply carry this patch for
your particular board and firmware. As I see it, there is a good chance
every one else will just carry the revert of it themselves otherwise.

> The people who can take this decision are rather the maintainers of
> the platform itself, or possibly the arm-soc maintainers if you still
> don't like what the platform maintainers decided.
> 
> Thomas

We both have a conflict of interest here, so your offer in an other
message to let the platform maintainers decide is appreciated.

Regards
Ralph

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Lennart Sorensen]

On Wed, Aug 24, 2016 at 09:52:00PM +0200, Thomas Petazzoni wrote:
> I'll let the platform maintainers decide what's the least
> intrusive/problematic option. Both solutions have drawbacks, so it's
> really a "political" decision to make here.

I think the main valid argument for a revert is that it violates the
documented dtb ordering rule.  It also fails to actually do what it was
intended to do since network device probing order really isn't defined
in linux and if you care you should fix it in userspace.  Fixing a
regression would be a side effect, and since the ordering isn't certain
anyhow, anyone that did see a regression was doing it "wrong" already,
although also anyone that saw a benefit from the change was also doing it
"wrong".

> Not only async probing, but also PCIe devices, as you mentioned
> earlier :-)

Yeah they better come up with a safer way to determine which network
device is which from user space.

-- 
Len Sorensen

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Jason Cooper]

All,

On Wed, Aug 24, 2016 at 10:41:02PM +0200, Ralph Sennhauser wrote:
> On Wed, 24 Aug 2016 20:15:31 +0200
> Thomas Petazzoni <thomas.petazzoni@free-electrons.com> wrote:
> > On Wed, 24 Aug 2016 19:10:04 +0200, Ralph Sennhauser wrote:
> > 
> > The people who can take this decision are rather the maintainers of
> > the platform itself, or possibly the arm-soc maintainers if you still
> > don't like what the platform maintainers decided.
> 
> We both have a conflict of interest here, so your offer in an other
> message to let the platform maintainers decide is appreciated.

Ok, I'm going to jump in here with the caveats that a) I don't mind
being overruled, and b) I'm not going to participate in a never-ending
thread on this topic.  ;-)

So, I'm just a back-seat co-maintainer for mvebu nowadays, my opinion
should be taken with a grain of salt.

>From the kernel-side, there is no guarantee of device naming.  The
kernel simply doesn't have sufficient information at boot time.  This is
why udev, systemd-udev-ntpd-syslog-kitchensink, and others were created.
To read immutable attributes of a device and apply consistent naming to
them based on configuration files.  That's why userspace frequently uses
uuids to locate partitions, rather than /dev/sdX[0-9] nodes.

The devicetree "ordered by address" rule is, in my opinion, an arbitrary
CDO-rule [1].  It doesn't describe the hardware, or a relationship
between them.  As such, it's just as arbitrary as probe order.  It just
happens to be something we can twiddle.  It also depends on dtc
preserving that order.

>From the user side, without udev and friends, shit changed from one
kernel to the next.  That's not good.  I can certainly see the point
that it should have been left alone to begin with.  But we aren't there
today.

So what happens if we revert now?  Right or wrong, in a couple of
months, someone else will complain, asking to revert the revert.
And what will our answer be?  "We did it last time, but not this time."
or "Ok, but gosh-darnit, this is the last time."

To be blunt, I think our best path forward is to just hold our noses and
let it stand as is.  Some will fix their userspace to adapt, others will
carry a patch.  It's more important at this point to be consistent
moving forward.  It's better to hear "Yeah, it fucking changed once."
rather than "I don't know what to expect, it changes every few releases."

thx,

Jason.


[1] CDO: OCD with the letters neatly arranged in alphabetical order.

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Ralph Sennhauser]

Hi Jason.

On Wed, 24 Aug 2016 21:48:36 +0000
Jason Cooper <jason@lakedaemon.net> wrote:

> All,
> 
> On Wed, Aug 24, 2016 at 10:41:02PM +0200, Ralph Sennhauser wrote:
> > On Wed, 24 Aug 2016 20:15:31 +0200
> > Thomas Petazzoni <thomas.petazzoni@free-electrons.com> wrote:  
> > > On Wed, 24 Aug 2016 19:10:04 +0200, Ralph Sennhauser wrote:
> > > 
> > > The people who can take this decision are rather the maintainers
> > > of the platform itself, or possibly the arm-soc maintainers if
> > > you still don't like what the platform maintainers decided.  
> > 
> > We both have a conflict of interest here, so your offer in an other
> > message to let the platform maintainers decide is appreciated.  
> 
> Ok, I'm going to jump in here with the caveats that a) I don't mind
> being overruled, and b) I'm not going to participate in a never-ending
> thread on this topic.  ;-)

I'm also not interested in a never ending thread. It's moot that udev
can't rename to kernel names sanely and we were sold ep34aj17asz98 as
the replacement. Or that tearing apart the casing to replace the wifi
modules with an ethernet one when there are already 5 Gbit ports is not
a case I care about.

Relying on the order might in theory have flaws but works just fine in
practice. Thomas Petazzoni, I, OpenWrt / Lede / etc all do so with
success.

It's also not a side effect from major changes, so it didn't break by
accident but as the subject says deliberately.

> So, I'm just a back-seat co-maintainer for mvebu nowadays, my opinion
> should be taken with a grain of salt.
> 
> From the kernel-side, there is no guarantee of device naming.  The
> kernel simply doesn't have sufficient information at boot time.  This
> is why udev, systemd-udev-ntpd-syslog-kitchensink, and others were
> created. To read immutable attributes of a device and apply
> consistent naming to them based on configuration files.  That's why
> userspace frequently uses uuids to locate partitions, rather
> than /dev/sdX[0-9] nodes.
> 
> The devicetree "ordered by address" rule is, in my opinion, an
> arbitrary CDO-rule [1].  It doesn't describe the hardware, or a
> relationship between them.  As such, it's just as arbitrary as probe
> order.  It just happens to be something we can twiddle.  It also
> depends on dtc preserving that order.
> 

How exceptional is this exception we are talking about? I mean if it's
the only place you might want to change it even in 2 years but you
can't because of the very same rule which was broken here.

> From the user side, without udev and friends, shit changed from one
> kernel to the next.  That's not good.  I can certainly see the point
> that it should have been left alone to begin with.  But we aren't
> there today.
> 

I think if we were at 4.6-rcX reverting would be an easy call. I know
it's more difficult to make this call now.

Ltsi is 4.1, longterm is 4.4, so penetration is probably marginal at
best at this time. From my view the damage caused by a revert would be
less than the damage that will be caused by the commit if left in but we
can't wait much much longer until this changes.

> So what happens if we revert now?  Right or wrong, in a couple of
> months, someone else will complain, asking to revert the revert.
> And what will our answer be?  "We did it last time, but not this
> time." or "Ok, but gosh-darnit, this is the last time."
> 

If you use the ordering by address as main argument for the revert there
will be nothing to argue about.

> To be blunt, I think our best path forward is to just hold our noses
> and let it stand as is.  Some will fix their userspace to adapt,
> others will carry a patch.  It's more important at this point to be
> consistent moving forward.  It's better to hear "Yeah, it fucking
> changed once." rather than "I don't know what to expect, it changes
> every few releases."
> 
> thx,
> 
> Jason.
> 
> 
> [1] CDO: OCD with the letters neatly arranged in alphabetical order.

Thanks for sharing your thoughts

Regards
Ralph

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Lennart Sorensen]

On Thu, Aug 25, 2016 at 09:38:39AM +0200, Ralph Sennhauser wrote:
> I'm also not interested in a never ending thread. It's moot that udev
> can't rename to kernel names sanely and we were sold ep34aj17asz98 as
> the replacement. Or that tearing apart the casing to replace the wifi
> modules with an ethernet one when there are already 5 Gbit ports is not
> a case I care about.
> 
> Relying on the order might in theory have flaws but works just fine in
> practice. Thomas Petazzoni, I, OpenWrt / Lede / etc all do so with
> success.

It has worked so far.  I expect it to break at some point.

> It's also not a side effect from major changes, so it didn't break by
> accident but as the subject says deliberately.

In this case that is certainly true.

> How exceptional is this exception we are talking about? I mean if it's
> the only place you might want to change it even in 2 years but you
> can't because of the very same rule which was broken here.

I did a quick look at the dts files, and there are quite a few cases
that are not ordered, and I can't find anything in Documentation that
says they should be sorted.  It just appears that in most cases people
do sort them, since that just makes sense for finding things and it is
as good an order as any other.

> I think if we were at 4.6-rcX reverting would be an easy call. I know
> it's more difficult to make this call now.
> 
> Ltsi is 4.1, longterm is 4.4, so penetration is probably marginal at
> best at this time. From my view the damage caused by a revert would be
> less than the damage that will be caused by the commit if left in but we
> can't wait much much longer until this changes.

Certainly no longterm kernels have this change yet, so it certainly does
seem the impact would be rather minimal.  Usually the rule seems to be
"Don't break user space".  This did potentially break userspace, so then
the question is whether what user space was doing was considred proper
and something that should expect not to break.

> If you use the ordering by address as main argument for the revert there
> will be nothing to argue about.

If only there was a documented place that had that rule.  I can't find it.

-- 
Len Sorensen

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Gregory CLEMENT]

Hi Ralph,
 
 On jeu., août 25 2016, Ralph Sennhauser <ralph.sennhauser@gmail.com> wrote:

> Hi Jason.
>
> On Wed, 24 Aug 2016 21:48:36 +0000
> Jason Cooper <jason@lakedaemon.net> wrote:
>
>> All,
>> 
>> On Wed, Aug 24, 2016 at 10:41:02PM +0200, Ralph Sennhauser wrote:
>> > On Wed, 24 Aug 2016 20:15:31 +0200
>> > Thomas Petazzoni <thomas.petazzoni@free-electrons.com> wrote:  
>> > > On Wed, 24 Aug 2016 19:10:04 +0200, Ralph Sennhauser wrote:
>> > > 
>> > > The people who can take this decision are rather the maintainers
>> > > of the platform itself, or possibly the arm-soc maintainers if
>> > > you still don't like what the platform maintainers decided.  
>> > 
>> > We both have a conflict of interest here, so your offer in an other
>> > message to let the platform maintainers decide is appreciated.  
>> 
>> Ok, I'm going to jump in here with the caveats that a) I don't mind
>> being overruled, and b) I'm not going to participate in a never-ending
>> thread on this topic.  ;-)
>
> I'm also not interested in a never ending thread. It's moot that udev

I am the one who applied this commit.

First it is really unfortunate this problem was not raised when this
patch was discussed especially because openwrt was part of the
discussion:
http://lists.infradead.org/pipermail/linux-arm-kernel/2016-February/411382.html

Then, the main motivation for merging this patch was to ease the work of
people doing bring-up of new board using Armada 38x SoCs. When doing this
we just rely on datasheet and U-Boot and it occurred that the way the Soc
was designed they put the first GMAC at a higher address to the second
one. Respecting this order helps low level development.

However for more advanced configuration we expect that more clever tools
such as udev should be used. (and Lennart confirm that it is still
working).

Also note that in the past we considered to being able to modify the
order of the ethernet device from the device tree by using the
alias. But the idea was rejected by David Miller (the network subsystem
maintainer) because this kind of policy should be done at userspace level.

For these reasons I won't revert this commit.

Gregory


> can't rename to kernel names sanely and we were sold ep34aj17asz98 as
> the replacement. Or that tearing apart the casing to replace the wifi
> modules with an ethernet one when there are already 5 Gbit ports is not
> a case I care about.
>
> Relying on the order might in theory have flaws but works just fine in
> practice. Thomas Petazzoni, I, OpenWrt / Lede / etc all do so with
> success.
>
> It's also not a side effect from major changes, so it didn't break by
> accident but as the subject says deliberately.
>
>> So, I'm just a back-seat co-maintainer for mvebu nowadays, my opinion
>> should be taken with a grain of salt.
>> 
>> From the kernel-side, there is no guarantee of device naming.  The
>> kernel simply doesn't have sufficient information at boot time.  This
>> is why udev, systemd-udev-ntpd-syslog-kitchensink, and others were
>> created. To read immutable attributes of a device and apply
>> consistent naming to them based on configuration files.  That's why
>> userspace frequently uses uuids to locate partitions, rather
>> than /dev/sdX[0-9] nodes.
>> 
>> The devicetree "ordered by address" rule is, in my opinion, an
>> arbitrary CDO-rule [1].  It doesn't describe the hardware, or a
>> relationship between them.  As such, it's just as arbitrary as probe
>> order.  It just happens to be something we can twiddle.  It also
>> depends on dtc preserving that order.
>> 
>
> How exceptional is this exception we are talking about? I mean if it's
> the only place you might want to change it even in 2 years but you
> can't because of the very same rule which was broken here.
>
>> From the user side, without udev and friends, shit changed from one
>> kernel to the next.  That's not good.  I can certainly see the point
>> that it should have been left alone to begin with.  But we aren't
>> there today.
>> 
>
> I think if we were at 4.6-rcX reverting would be an easy call. I know
> it's more difficult to make this call now.
>
> Ltsi is 4.1, longterm is 4.4, so penetration is probably marginal at
> best at this time. From my view the damage caused by a revert would be
> less than the damage that will be caused by the commit if left in but we
> can't wait much much longer until this changes.
>
>> So what happens if we revert now?  Right or wrong, in a couple of
>> months, someone else will complain, asking to revert the revert.
>> And what will our answer be?  "We did it last time, but not this
>> time." or "Ok, but gosh-darnit, this is the last time."
>> 
>
> If you use the ordering by address as main argument for the revert there
> will be nothing to argue about.
>
>> To be blunt, I think our best path forward is to just hold our noses
>> and let it stand as is.  Some will fix their userspace to adapt,
>> others will carry a patch.  It's more important at this point to be
>> consistent moving forward.  It's better to hear "Yeah, it fucking
>> changed once." rather than "I don't know what to expect, it changes
>> every few releases."
>> 
>> thx,
>> 
>> Jason.
>> 
>> 
>> [1] CDO: OCD with the letters neatly arranged in alphabetical order.
>
> Thanks for sharing your thoughts
>
> Regards
> Ralph

-- 
Gregory Clement, Free Electrons
Kernel, drivers, real-time and embedded Linux
development, consulting, training and support.
http://free-electrons.com

Re: [Regression?] Commit cb4f71c429 deliberately changes order of network interfaces

[Ralph Sennhauser]

Hi Gregory

On Fri, 26 Aug 2016 10:43:43 +0200
Gregory CLEMENT <gregory.clement@free-electrons.com> wrote:

> Hi Ralph,
>  
>  On jeu., août 25 2016, Ralph Sennhauser <ralph.sennhauser@gmail.com>
> wrote:
> 
> > Hi Jason.
> >
> > On Wed, 24 Aug 2016 21:48:36 +0000
> > Jason Cooper <jason@lakedaemon.net> wrote:
> >  
> >> All,
> >> 
> >> On Wed, Aug 24, 2016 at 10:41:02PM +0200, Ralph Sennhauser wrote:  
> >> > On Wed, 24 Aug 2016 20:15:31 +0200
> >> > Thomas Petazzoni <thomas.petazzoni@free-electrons.com> wrote:    
> >> > > On Wed, 24 Aug 2016 19:10:04 +0200, Ralph Sennhauser wrote:
> >> > > 
> >> > > The people who can take this decision are rather the
> >> > > maintainers of the platform itself, or possibly the arm-soc
> >> > > maintainers if you still don't like what the platform
> >> > > maintainers decided.    
> >> > 
> >> > We both have a conflict of interest here, so your offer in an
> >> > other message to let the platform maintainers decide is
> >> > appreciated.    
> >> 
> >> Ok, I'm going to jump in here with the caveats that a) I don't mind
> >> being overruled, and b) I'm not going to participate in a
> >> never-ending thread on this topic.  ;-)  
> >
> > I'm also not interested in a never ending thread. It's moot that
> > udev  
> 
> I am the one who applied this commit.
> 
> First it is really unfortunate this problem was not raised when this
> patch was discussed especially because openwrt was part of the
> discussion:
> http://lists.infradead.org/pipermail/linux-arm-kernel/2016-February/411382.html
> 
> Then, the main motivation for merging this patch was to ease the work
> of people doing bring-up of new board using Armada 38x SoCs. When
> doing this we just rely on datasheet and U-Boot and it occurred that
> the way the Soc was designed they put the first GMAC at a higher
> address to the second one. Respecting this order helps low level
> development.
> 
> However for more advanced configuration we expect that more clever
> tools such as udev should be used. (and Lennart confirm that it is
> still working).
> 
> Also note that in the past we considered to being able to modify the
> order of the ethernet device from the device tree by using the
> alias. But the idea was rejected by David Miller (the network
> subsystem maintainer) because this kind of policy should be done at
> userspace level.
> 
> For these reasons I won't revert this commit.
> 


Look at the Gentoo ebuild for Lennarts definition of works. But yes,
obviously it can among other ways be addressed in userspace. For me
this never was a technical discussion but one about politics.

>From your mail I take the following:
1) It's not a case of sneaking in the change and hoping no one notices
before it hits stable.
2) It "broke" those Linksys devices but as you had an offer for testing
I can only assume fair game.
3) You feel comfortable holding out your neck creating a precedent.

Therefore I respect your decision and wont pursue the issue any further.

Thanks
Ralph


> Gregory
> 
> 
> > can't rename to kernel names sanely and we were sold ep34aj17asz98
> > as the replacement. Or that tearing apart the casing to replace the
> > wifi modules with an ethernet one when there are already 5 Gbit
> > ports is not a case I care about.
> >
> > Relying on the order might in theory have flaws but works just fine
> > in practice. Thomas Petazzoni, I, OpenWrt / Lede / etc all do so
> > with success.
> >
> > It's also not a side effect from major changes, so it didn't break
> > by accident but as the subject says deliberately.
> >  
> >> So, I'm just a back-seat co-maintainer for mvebu nowadays, my
> >> opinion should be taken with a grain of salt.
> >> 
> >> From the kernel-side, there is no guarantee of device naming.  The
> >> kernel simply doesn't have sufficient information at boot time.
> >> This is why udev, systemd-udev-ntpd-syslog-kitchensink, and others
> >> were created. To read immutable attributes of a device and apply
> >> consistent naming to them based on configuration files.  That's why
> >> userspace frequently uses uuids to locate partitions, rather
> >> than /dev/sdX[0-9] nodes.
> >> 
> >> The devicetree "ordered by address" rule is, in my opinion, an
> >> arbitrary CDO-rule [1].  It doesn't describe the hardware, or a
> >> relationship between them.  As such, it's just as arbitrary as
> >> probe order.  It just happens to be something we can twiddle.  It
> >> also depends on dtc preserving that order.
> >>   
> >
> > How exceptional is this exception we are talking about? I mean if
> > it's the only place you might want to change it even in 2 years but
> > you can't because of the very same rule which was broken here.
> >  
> >> From the user side, without udev and friends, shit changed from one
> >> kernel to the next.  That's not good.  I can certainly see the
> >> point that it should have been left alone to begin with.  But we
> >> aren't there today.
> >>   
> >
> > I think if we were at 4.6-rcX reverting would be an easy call. I
> > know it's more difficult to make this call now.
> >
> > Ltsi is 4.1, longterm is 4.4, so penetration is probably marginal at
> > best at this time. From my view the damage caused by a revert would
> > be less than the damage that will be caused by the commit if left
> > in but we can't wait much much longer until this changes.
> >  
> >> So what happens if we revert now?  Right or wrong, in a couple of
> >> months, someone else will complain, asking to revert the revert.
> >> And what will our answer be?  "We did it last time, but not this
> >> time." or "Ok, but gosh-darnit, this is the last time."
> >>   
> >
> > If you use the ordering by address as main argument for the revert
> > there will be nothing to argue about.
> >  
> >> To be blunt, I think our best path forward is to just hold our
> >> noses and let it stand as is.  Some will fix their userspace to
> >> adapt, others will carry a patch.  It's more important at this
> >> point to be consistent moving forward.  It's better to hear "Yeah,
> >> it fucking changed once." rather than "I don't know what to
> >> expect, it changes every few releases."
> >> 
> >> thx,
> >> 
> >> Jason.
> >> 
> >> 
> >> [1] CDO: OCD with the letters neatly arranged in alphabetical
> >> order.  
> >
> > Thanks for sharing your thoughts
> >
> > Regards
> > Ralph  
>