Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
Hi Lennart, Moreover, if we give people this feature I'm pretty sure we'll get lots of people expecting it to work also for any other sort of name and getting confused when it doesn't. Well, this is something we can fix by documentation, no? Or maybe name the match option differently, maybe OriginalName= or KernelName= or so, and then only matching interfaces where you know that the name was selected by userspace in the first place? I like the idea of OriginalName, much less likely to get people confused. I now implemented that, with the restriction that we cannot match on renamed names. For now I left it open to match on ethX style names, as people in principle could do sensible things like OriginalName=eth* or even OriginalName=eth0 when we know there is only one interface. One thing to consider would be to disallow renaming from a .link file where the OriginalName was used to match. That way we don't have the somewhat odd situation that a .link file can only be applied once (we do not remember the original name, so cannot match on that the second time around, as that would be a mess)... Maybe we should even store the original name in a udev property, so that we can make this fully idempotent simply because we can always check this new property for the original name passed down from the kernel? you do realize that once the kernel renamed the name, it is free to reuse that name again. So eth0 as OriginalName can then be present multiple times. I think that you can only have OriginalName matching if a renaming has never occurred. Otherwise this is really not deterministic. Since foo0 (renamed from eth0) and and bar0 (renamed from eth0) will both match OriginalName=eth0. The difference is only timing and that is inhering racy of course. Regards Marcel ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
On 5 Dec 2014 10:07, Marcel Holtmann mar...@holtmann.org wrote: Hi Lennart, Moreover, if we give people this feature I'm pretty sure we'll get lots of people expecting it to work also for any other sort of name and getting confused when it doesn't. Well, this is something we can fix by documentation, no? Or maybe name the match option differently, maybe OriginalName= or KernelName= or so, and then only matching interfaces where you know that the name was selected by userspace in the first place? I like the idea of OriginalName, much less likely to get people confused. I now implemented that, with the restriction that we cannot match on renamed names. For now I left it open to match on ethX style names, as people in principle could do sensible things like OriginalName=eth* or even OriginalName=eth0 when we know there is only one interface. One thing to consider would be to disallow renaming from a .link file where the OriginalName was used to match. That way we don't have the somewhat odd situation that a .link file can only be applied once (we do not remember the original name, so cannot match on that the second time around, as that would be a mess)... Maybe we should even store the original name in a udev property, so that we can make this fully idempotent simply because we can always check this new property for the original name passed down from the kernel? you do realize that once the kernel renamed the name, it is free to reuse that name again. So eth0 as OriginalName can then be present multiple times. I think that you can only have OriginalName matching if a renaming has never occurred. Otherwise this is really not deterministic. Since foo0 (renamed from eth0) and and bar0 (renamed from eth0) will both match OriginalName=eth0. The difference is only timing and that is inhering racy of course. Well, matching on kernel names is not sane, and I put a warning in the manpage. The reason I left it in was that you could match on eth0 of you know it is the only ethX device on your system, or you could match on eth* which would also be deterministic. Cheers, Tom ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
On Fri, 05.12.14 10:07, Marcel Holtmann (mar...@holtmann.org) wrote: Hi Lennart, Moreover, if we give people this feature I'm pretty sure we'll get lots of people expecting it to work also for any other sort of name and getting confused when it doesn't. Well, this is something we can fix by documentation, no? Or maybe name the match option differently, maybe OriginalName= or KernelName= or so, and then only matching interfaces where you know that the name was selected by userspace in the first place? I like the idea of OriginalName, much less likely to get people confused. I now implemented that, with the restriction that we cannot match on renamed names. For now I left it open to match on ethX style names, as people in principle could do sensible things like OriginalName=eth* or even OriginalName=eth0 when we know there is only one interface. One thing to consider would be to disallow renaming from a .link file where the OriginalName was used to match. That way we don't have the somewhat odd situation that a .link file can only be applied once (we do not remember the original name, so cannot match on that the second time around, as that would be a mess)... Maybe we should even store the original name in a udev property, so that we can make this fully idempotent simply because we can always check this new property for the original name passed down from the kernel? you do realize that once the kernel renamed the name, it is free to reuse that name again. So eth0 as OriginalName can then be present multiple times. I think that you can only have OriginalName matching if a renaming has never occurred. Otherwise this is really not deterministic. Since foo0 (renamed from eth0) and and bar0 (renamed from eth0) will both match OriginalName=eth0. The difference is only timing and that is inhering racy of course. Well, sure, if you use OriginalName= for such names, it's your own fault. I think using OriginalName= is primarily useful for things like veth links where the names are chosen by userspace anyway. Hmm, now that the naming policy is exposed by the kernel, can we detect the cases where naming happens like eth0, eth1, and print a nice warning or so, if people then match against that? Lennart -- Lennart Poettering, Red Hat ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
On Fri, Dec 5, 2014 at 2:21 PM, Lennart Poettering lenn...@poettering.net wrote: On Fri, 05.12.14 10:07, Marcel Holtmann (mar...@holtmann.org) wrote: Hi Lennart, Moreover, if we give people this feature I'm pretty sure we'll get lots of people expecting it to work also for any other sort of name and getting confused when it doesn't. Well, this is something we can fix by documentation, no? Or maybe name the match option differently, maybe OriginalName= or KernelName= or so, and then only matching interfaces where you know that the name was selected by userspace in the first place? I like the idea of OriginalName, much less likely to get people confused. I now implemented that, with the restriction that we cannot match on renamed names. For now I left it open to match on ethX style names, as people in principle could do sensible things like OriginalName=eth* or even OriginalName=eth0 when we know there is only one interface. One thing to consider would be to disallow renaming from a .link file where the OriginalName was used to match. That way we don't have the somewhat odd situation that a .link file can only be applied once (we do not remember the original name, so cannot match on that the second time around, as that would be a mess)... Maybe we should even store the original name in a udev property, so that we can make this fully idempotent simply because we can always check this new property for the original name passed down from the kernel? you do realize that once the kernel renamed the name, it is free to reuse that name again. So eth0 as OriginalName can then be present multiple times. I think that you can only have OriginalName matching if a renaming has never occurred. Otherwise this is really not deterministic. Since foo0 (renamed from eth0) and and bar0 (renamed from eth0) will both match OriginalName=eth0. The difference is only timing and that is inhering racy of course. Well, sure, if you use OriginalName= for such names, it's your own fault. I think using OriginalName= is primarily useful for things like veth links where the names are chosen by userspace anyway. Hmm, now that the naming policy is exposed by the kernel, can we detect the cases where naming happens like eth0, eth1, and print a nice warning or so, if people then match against that? I added some warnings to git now. Still need to push some more kernel patches to get more stuff tagged correctly, but at least the systemd side is now good. Cheers, Tom ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
On Thu, Dec 4, 2014 at 3:50 AM, Lennart Poettering lenn...@poettering.net wrote: Tom, I think it would make sense to allow Name= matches in the [Match] section of .link files, no? Hm, so far I hesitated, as the most common scenarios is the one that William has, namely to want to match on the name set by udev and not the one set by the kernel. This we cannot really do in .link files as it would be confusing for admins to figure out exactly what is being matched on, and it could obviously become undecidable/nonterminating depending on how we try to implement it. I mean, most of the times .link files are used to choose the name depending on other fields, but I think in all cases where the name is chosen at creation time of an interface (like for example for veth links), it should be Ok to match links based on the name they already have? So this we could do (i.e., allow name matches if and only if the name is set at creation time). Though, if the name is set at creation time, shouldn't also all the other properties have been? Moreover, if we give people this feature I'm pretty sure we'll get lots of people expecting it to work also for any other sort of name and getting confused when it doesn't. So overall, not really convinced, but if there is a usecase for matching on such limited types of names, I'm not very strongly against it. -t ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
On Thu, 04.12.14 11:20, Tom Gundersen (t...@jklm.no) wrote: I mean, most of the times .link files are used to choose the name depending on other fields, but I think in all cases where the name is chosen at creation time of an interface (like for example for veth links), it should be Ok to match links based on the name they already have? So this we could do (i.e., allow name matches if and only if the name is set at creation time). Though, if the name is set at creation time, shouldn't also all the other properties have been? In nspawn we create veth links, and I *really* *really* don't want to add tons of options there to set the MTU or suchlike. Setting these properties really should be done within networkd I think. I think it would be really useful if we could match on the container's veth links by name, so that a .link file for it could be equally expressive as the .network file already is. Moreover, if we give people this feature I'm pretty sure we'll get lots of people expecting it to work also for any other sort of name and getting confused when it doesn't. Well, this is something we can fix by documentation, no? Or maybe name the match option differently, maybe OriginalName= or KernelName= or so, and then only matching interfaces where you know that the name was selected by userspace in the first place? Lennart -- Lennart Poettering, Red Hat ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
On Thu, Dec 4, 2014 at 4:11 PM, Lennart Poettering lenn...@poettering.net wrote: On Thu, 04.12.14 11:20, Tom Gundersen (t...@jklm.no) wrote: I mean, most of the times .link files are used to choose the name depending on other fields, but I think in all cases where the name is chosen at creation time of an interface (like for example for veth links), it should be Ok to match links based on the name they already have? So this we could do (i.e., allow name matches if and only if the name is set at creation time). Though, if the name is set at creation time, shouldn't also all the other properties have been? In nspawn we create veth links, and I *really* *really* don't want to add tons of options there to set the MTU or suchlike. Setting these properties really should be done within networkd I think. I think it would be really useful if we could match on the container's veth links by name, so that a .link file for it could be equally expressive as the .network file already is. Moreover, if we give people this feature I'm pretty sure we'll get lots of people expecting it to work also for any other sort of name and getting confused when it doesn't. Well, this is something we can fix by documentation, no? Or maybe name the match option differently, maybe OriginalName= or KernelName= or so, and then only matching interfaces where you know that the name was selected by userspace in the first place? I like the idea of OriginalName, much less likely to get people confused. I now implemented that, with the restriction that we cannot match on renamed names. For now I left it open to match on ethX style names, as people in principle could do sensible things like OriginalName=eth* or even OriginalName=eth0 when we know there is only one interface. One thing to consider would be to disallow renaming from a .link file where the OriginalName was used to match. That way we don't have the somewhat odd situation that a .link file can only be applied once (we do not remember the original name, so cannot match on that the second time around, as that would be a mess)... Testing and feedback welcome. Does this solve your problem William? Cheers, Tom ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
On Thu, 04.12.14 18:53, Tom Gundersen (t...@jklm.no) wrote: Moreover, if we give people this feature I'm pretty sure we'll get lots of people expecting it to work also for any other sort of name and getting confused when it doesn't. Well, this is something we can fix by documentation, no? Or maybe name the match option differently, maybe OriginalName= or KernelName= or so, and then only matching interfaces where you know that the name was selected by userspace in the first place? I like the idea of OriginalName, much less likely to get people confused. I now implemented that, with the restriction that we cannot match on renamed names. For now I left it open to match on ethX style names, as people in principle could do sensible things like OriginalName=eth* or even OriginalName=eth0 when we know there is only one interface. One thing to consider would be to disallow renaming from a .link file where the OriginalName was used to match. That way we don't have the somewhat odd situation that a .link file can only be applied once (we do not remember the original name, so cannot match on that the second time around, as that would be a mess)... Maybe we should even store the original name in a udev property, so that we can make this fully idempotent simply because we can always check this new property for the original name passed down from the kernel? Lennart -- Lennart Poettering, Red Hat ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
On Thu, Dec 4, 2014 at 8:08 PM, Lennart Poettering lenn...@poettering.net wrote: On Thu, 04.12.14 18:53, Tom Gundersen (t...@jklm.no) wrote: Moreover, if we give people this feature I'm pretty sure we'll get lots of people expecting it to work also for any other sort of name and getting confused when it doesn't. Well, this is something we can fix by documentation, no? Or maybe name the match option differently, maybe OriginalName= or KernelName= or so, and then only matching interfaces where you know that the name was selected by userspace in the first place? I like the idea of OriginalName, much less likely to get people confused. I now implemented that, with the restriction that we cannot match on renamed names. For now I left it open to match on ethX style names, as people in principle could do sensible things like OriginalName=eth* or even OriginalName=eth0 when we know there is only one interface. One thing to consider would be to disallow renaming from a .link file where the OriginalName was used to match. That way we don't have the somewhat odd situation that a .link file can only be applied once (we do not remember the original name, so cannot match on that the second time around, as that would be a mess)... Maybe we should even store the original name in a udev property, so that we can make this fully idempotent simply because we can always check this new property for the original name passed down from the kernel? We got into lots of issues in the past with trying to track the renaming from userspace (which is why NET_NAME_RENAMED and friends now are exposed from the kernel), so I'd rather not get back into that game. The issue is that processes can (and do) rename interfaces outside of udev's control (possibly before udev is running), and we'll then be left with knowing that a device was renamed (as the kernel tells us that much), but not from what. I'd really prefer to keep this dead simple and either say (as now) that the config is only applied once if you rename the device, or say that the name-matching config cannot rename. -t ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
In regards to the OS and syntax, this is NixOS. NixOS uses its own
expression language to configure the entire system, including the network
stack. If a user wanted to configure their system's networking stack, they
would modify the networking.* set of options in their
/etc/nixos/configuration.nix. Currently this is done with a set of
generated scripts run as systemd services with device based dependencies.
Ideally this would be directly translated to networkd syntax.
For example, a user might have this snippet as part of their config for
statically configuring one of their ethernet interfaces:
networking.interfaces.enp4s0 = {
macAddress = de:ad:be:eef:ca:fe;
mtu = 9000;
ip4 = [
{ address = 10.0.0.2; prefixLength = 24; }
{ address = 10.0.0.3; prefixLength = 24; }
];
};
For link configuration, we are currently generating a
network-link-ifname.service which sets the mac address and mtu using
iproute2:
https://github.com/NixOS/nixpkgs/blob/master/nixos/modules/tasks/network-interfaces.nix#L691-L713
The networking options are all defined in:
https://github.com/NixOS/nixpkgs/blob/master/nixos/modules/tasks/network-interfaces.nix
Our current translator to networkd is:
https://github.com/NixOS/nixpkgs/blob/master/nixos/modules/tasks/network-interfaces-systemd.nix
Of course, the problem with just using kernel names is that our use case
assumes that the interface was already renamed by udev. Therefore, our
matching must be based on the post-udev name if it is to be backward
compatible. Ultimately it might be best to continue using our set of
scripts coupled with udev generated systemd device units, given our use
cases.
On Thu, Dec 4, 2014 at 9:53 AM, Tom Gundersen t...@jklm.no wrote:
On Thu, Dec 4, 2014 at 4:11 PM, Lennart Poettering
lenn...@poettering.net wrote:
On Thu, 04.12.14 11:20, Tom Gundersen (t...@jklm.no) wrote:
I mean, most of the times .link files are
used to choose the name depending on other fields, but I think in all
cases where the name is chosen at creation time of an interface (like
for example for veth links), it should be Ok to match links based on
the name they already have?
So this we could do (i.e., allow name matches if and only if the name
is set at creation time). Though, if the name is set at creation time,
shouldn't also all the other properties have been?
In nspawn we create veth links, and I *really* *really* don't want to
add tons of options there to set the MTU or suchlike. Setting these
properties really should be done within networkd I think.
I think it would be really useful if we could match on the container's
veth links by name, so that a .link file for it could be equally
expressive as the .network file already is.
Moreover, if we
give people this feature I'm pretty sure we'll get lots of people
expecting it to work also for any other sort of name and getting
confused when it doesn't.
Well, this is something we can fix by documentation, no?
Or maybe name the match option differently, maybe OriginalName= or
KernelName= or so, and then only matching interfaces where you know
that the name was selected by userspace in the first place?
I like the idea of OriginalName, much less likely to get people
confused. I now implemented that, with the restriction that we cannot
match on renamed names. For now I left it open to match on ethX style
names, as people in principle could do sensible things like
OriginalName=eth* or even OriginalName=eth0 when we know there is
only one interface.
One thing to consider would be to disallow renaming from a .link file
where the OriginalName was used to match. That way we don't have the
somewhat odd situation that a .link file can only be applied once (we
do not remember the original name, so cannot match on that the second
time around, as that would be a mess)...
Testing and feedback welcome. Does this solve your problem William?
Cheers,
Tom
___
systemd-devel mailing list
systemd-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
Thanks for the links William. Looks like most of that should be
covered by networkd. I now also added support for MTU and MACAddress
to be set in the .network files. The intention here is that if you
ever disable a network at runtime, we will revert to the default
(.link-defined) settings. However, runtime configuration is not yet
implemented, so for now the only distinction between setting this in a
.link and a .network is that you can now match on the real ifname.
HTH,
Tom
On Thu, Dec 4, 2014 at 9:07 PM, William Kennington
will...@wkennington.com wrote:
In regards to the OS and syntax, this is NixOS. NixOS uses its own
expression language to configure the entire system, including the network
stack. If a user wanted to configure their system's networking stack, they
would modify the networking.* set of options in their
/etc/nixos/configuration.nix. Currently this is done with a set of generated
scripts run as systemd services with device based dependencies. Ideally this
would be directly translated to networkd syntax.
For example, a user might have this snippet as part of their config for
statically configuring one of their ethernet interfaces:
networking.interfaces.enp4s0 = {
macAddress = de:ad:be:eef:ca:fe;
mtu = 9000;
ip4 = [
{ address = 10.0.0.2; prefixLength = 24; }
{ address = 10.0.0.3; prefixLength = 24; }
];
};
For link configuration, we are currently generating a
network-link-ifname.service which sets the mac address and mtu using
iproute2:
https://github.com/NixOS/nixpkgs/blob/master/nixos/modules/tasks/network-interfaces.nix#L691-L713
The networking options are all defined in:
https://github.com/NixOS/nixpkgs/blob/master/nixos/modules/tasks/network-interfaces.nix
Our current translator to networkd is:
https://github.com/NixOS/nixpkgs/blob/master/nixos/modules/tasks/network-interfaces-systemd.nix
Of course, the problem with just using kernel names is that our use case
assumes that the interface was already renamed by udev. Therefore, our
matching must be based on the post-udev name if it is to be backward
compatible. Ultimately it might be best to continue using our set of scripts
coupled with udev generated systemd device units, given our use cases.
On Thu, Dec 4, 2014 at 9:53 AM, Tom Gundersen t...@jklm.no wrote:
On Thu, Dec 4, 2014 at 4:11 PM, Lennart Poettering
lenn...@poettering.net wrote:
On Thu, 04.12.14 11:20, Tom Gundersen (t...@jklm.no) wrote:
I mean, most of the times .link files are
used to choose the name depending on other fields, but I think in all
cases where the name is chosen at creation time of an interface (like
for example for veth links), it should be Ok to match links based on
the name they already have?
So this we could do (i.e., allow name matches if and only if the name
is set at creation time). Though, if the name is set at creation time,
shouldn't also all the other properties have been?
In nspawn we create veth links, and I *really* *really* don't want to
add tons of options there to set the MTU or suchlike. Setting these
properties really should be done within networkd I think.
I think it would be really useful if we could match on the container's
veth links by name, so that a .link file for it could be equally
expressive as the .network file already is.
Moreover, if we
give people this feature I'm pretty sure we'll get lots of people
expecting it to work also for any other sort of name and getting
confused when it doesn't.
Well, this is something we can fix by documentation, no?
Or maybe name the match option differently, maybe OriginalName= or
KernelName= or so, and then only matching interfaces where you know
that the name was selected by userspace in the first place?
I like the idea of OriginalName, much less likely to get people
confused. I now implemented that, with the restriction that we cannot
match on renamed names. For now I left it open to match on ethX style
names, as people in principle could do sensible things like
OriginalName=eth* or even OriginalName=eth0 when we know there is
only one interface.
One thing to consider would be to disallow renaming from a .link file
where the OriginalName was used to match. That way we don't have the
somewhat odd situation that a .link file can only be applied once (we
do not remember the original name, so cannot match on that the second
time around, as that would be a mess)...
Testing and feedback welcome. Does this solve your problem William?
Cheers,
Tom
___
systemd-devel mailing list
systemd-devel@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
Yes, we could support a [Link] section in .network files applying such settings per network. Though that this should really be about overriding the link specific settings, so not sure it fits your usecase precisely... On 24 Nov 2014 20:46, William Kennington will...@wkennington.com wrote: I'd like to be able to set the MAC Address and MTU of interfaces, with just the interface name alone. The current method for matching links seems to be heavily tied to udev ATTRs and I understand given the nature of systemd.link that it might not make sense to Match based on name. However, I need this functionality to support mapping the current OS defined network configuration syntax to networkd. Is there any chance this could be supported in systemd.network instead of systemd.link? Regards, William ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
Re: [systemd-devel] networkd: Support setting mtu / mac address by interface name
On Mon, 24.11.14 11:46, William Kennington (will...@wkennington.com) wrote: I'd like to be able to set the MAC Address and MTU of interfaces, with just the interface name alone. The current method for matching links seems to be heavily tied to udev ATTRs and I understand given the nature of systemd.link that it might not make sense to Match based on name. However, I need this functionality to support mapping the current OS defined network configuration syntax to networkd. Is there any chance this could be supported in systemd.network instead of systemd.link? Tom, I think it would make sense to allow Name= matches in the [Match] section of .link files, no? I mean, most of the times .link files are used to choose the name depending on other fields, but I think in all cases where the name is chosen at creation time of an interface (like for example for veth links), it should be Ok to match links based on the name they already have? Lennart -- Lennart Poettering, Red Hat ___ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
