teamd interaction with linux kernel user_carrier_enabled

[George Wilkie]

Hi, like to get your thoughts on this issue we are running into.

Initially, running the lacp runner in teamd. This sends carrier changes 
to the kernel. Kernel sets "user_carrier_enabled" which blocks its 
normal carrier checks.

Then change teamd runner from lacp to loadbalance. When lacp runner 
exits, carrier is set to down in kernel. lacp and loadbalance runners 
both use same kernel mode "loadbalance", so starting loadbalance runner 
does not change kernel mode. Kernel still has "user_carrier_enabled" and 
team interface stays down due to NO_CARRIER.

It seems the only way to clear kernel "user_carrier_enabled" is to 
change the mode. Should there be a new option that teamd can use to tell 
kernel to clear "user_carrier_enabled" when lacp runner terminates? Or 
any other suggestion?

Thanks.

[Jiri Pirko]

Tue, Mar 28, 2017 at 06:51:09PM CEST, [e-mail] wrote:
> Hi, like to get your thoughts on this issue we are running into.
>
> Initially, running the lacp runner in teamd. This sends carrier changes to
> the kernel. Kernel sets "user_carrier_enabled" which blocks its normal
> carrier checks.
>
> Then change teamd runner from lacp to loadbalance. When lacp runner exits,

Hmm. What exactly are you doing, could you send a list of commands?
I suspect some oddities :)
> carrier is set to down in kernel. lacp and loadbalance runners both use same
> kernel mode "loadbalance", so starting loadbalance runner does not change
> kernel mode. Kernel still has "user_carrier_enabled" and team interface stays
> down due to NO_CARRIER.

Hmm. I think that the correct solution is to teach the runners not to
depend on kernel defaults. So in this care the loadbalance runner would
set user_carrier_enabled to false during init (not during takeover)
> It seems the only way to clear kernel "user_carrier_enabled" is to change the
> mode. Should there be a new option that teamd can use to tell kernel to clear
> "user_carrier_enabled" when lacp runner terminates? Or any other suggestion?
>
> Thanks.
> _______________________________________________
> libteam mailing list -- [e-mail]
> To unsubscribe send an email to [e-mail]

[George Wilkie]

On 03/30/2017 04:18 PM, Jiri Pirko wrote:
> Tue, Mar 28, 2017 at 06:51:09PM CEST, [e-mail] wrote:
> > Hi, like to get your thoughts on this issue we are running into.
> >
> > Initially, running the lacp runner in teamd. This sends carrier changes to
> > the kernel. Kernel sets "user_carrier_enabled" which blocks its normal
> > carrier checks.
> >
> > Then change teamd runner from lacp to loadbalance. When lacp runner exits,
>
> Hmm. What exactly are you doing, could you send a list of commands?
> I suspect some oddities :)

1. start off in LACP mode - NO-CARRIER expected as no LACP on other side.
root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy 
--take-over --config='{"runner":{"name":"lacp"}}'
This program is not intended to be run as root.
root@debian9:~# ip link set dev ens9 master team0
root@debian9:~# ip link set dev ens10 master team0
root@debian9:~# ip link set dev team0 up
root@debian9:~# ip link show team0
8: team0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc noqueue 
state DOWN mode DEFAULT group default qlen 1000
     link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
root@debian9:~#

2. change to balanced mode - still NO-CARRIER
root@debian9:~# ip link set dev team0 down
root@debian9:~# ip link set dev ens9 nomaster
root@debian9:~# ip link set dev ens10 nomaster
root@debian9:~# teamd --team-dev team0 --kill
root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy 
--take-over --config='{"runner":{"name":"loadbalance"}}'
This program is not intended to be run as root.
root@debian9:~# ip link set dev ens9 master team0
root@debian9:~# ip link set dev ens10 master team0
root@debian9:~# ip link set dev team0 up
root@debian9:~# ip link show team0
8: team0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc noqueue 
state DOWN mode DEFAULT group default qlen 1000
     link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
root@debian9:~#

3. have to switch to different kernel mode first to clear NO-CARRIER
root@debian9:~# ip link set dev team0 down
root@debian9:~# ip link set dev ens9 nomaster
root@debian9:~# ip link set dev ens10 nomaster
root@debian9:~# teamd --team-dev team0 --kill
root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy 
--take-over --config='{"runner":{"name":"roundrobin"}}'
This program is not intended to be run as root.
root@debian9:~# teamd --team-dev team0 --kill
root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy 
--take-over --config='{"runner":{"name":"loadbalance"}}'
This program is not intended to be run as root.
root@debian9:~# ip link set dev ens9 master team0
root@debian9:~# ip link set dev ens10 master team0
root@debian9:~# ip link set dev team0 up
root@debian9:~# ip link show team0
8: team0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state 
UP mode DEFAULT group default qlen 1000
     link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
root@debian9:~#

This is with debian9, teamd version 1.26-1+b1, kernel 4.9.13-1
> > carrier is set to down in kernel. lacp and loadbalance runners both use same
> > kernel mode "loadbalance", so starting loadbalance runner does not change
> > kernel mode. Kernel still has "user_carrier_enabled" and team interface stays
> > down due to NO_CARRIER.
>
> Hmm. I think that the correct solution is to teach the runners not to
> depend on kernel defaults. So in this care the loadbalance runner would
> set user_carrier_enabled to false during init (not during takeover)

ok, thanks.
> > It seems the only way to clear kernel "user_carrier_enabled" is to change the
> > mode. Should there be a new option that teamd can use to tell kernel to clear
> > "user_carrier_enabled" when lacp runner terminates? Or any other suggestion?
> >
> > Thanks.
> > _______________________________________________
> > libteam mailing list -- [e-mail]
> > To unsubscribe send an email to [e-mail]

[Jiri Pirko]

Fri, Mar 31, 2017 at 01:41:08PM CEST, [e-mail] wrote:
> On 03/30/2017 04:18 PM, Jiri Pirko wrote:
> > Tue, Mar 28, 2017 at 06:51:09PM CEST, [e-mail] wrote:
> > > Hi, like to get your thoughts on this issue we are running into.
> > >
> > > Initially, running the lacp runner in teamd. This sends carrier changes to
> > > the kernel. Kernel sets "user_carrier_enabled" which blocks its normal
> > > carrier checks.
> > >
> > > Then change teamd runner from lacp to loadbalance. When lacp runner exits,
> >
> > Hmm. What exactly are you doing, could you send a list of commands?
> > I suspect some oddities :)
> start off in LACP mode - NO-CARRIER expected as no LACP on other side.
>
> root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> --config='{"runner":{"name":"lacp"}}'
> This program is not intended to be run as root.
> root@debian9:~# ip link set dev ens9 master team0
> root@debian9:~# ip link set dev ens10 master team0
> root@debian9:~# ip link set dev team0 up
> root@debian9:~# ip link show team0
> 8: team0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc noqueue state
> DOWN mode DEFAULT group default qlen 1000
>    link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
> root@debian9:~#
> change to balanced mode - still NO-CARRIER
>
> root@debian9:~# ip link set dev team0 down
> root@debian9:~# ip link set dev ens9 nomaster
> root@debian9:~# ip link set dev ens10 nomaster
> root@debian9:~# teamd --team-dev team0 --kill
> root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over

I think that you might be missing the reason of existence of take over.
It suppose to "take over" the kernel team instance in case for example
the teamd segfaults or something. The intension was to run new teamd
instance taking over the kernel instance with exact same config.

Why are you doing this?
> --config='{"runner":{"name":"loadbalance"}}'
> This program is not intended to be run as root.
> root@debian9:~# ip link set dev ens9 master team0
> root@debian9:~# ip link set dev ens10 master team0
> root@debian9:~# ip link set dev team0 up
> root@debian9:~# ip link show team0
> 8: team0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc noqueue state
> DOWN mode DEFAULT group default qlen 1000
>    link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
> root@debian9:~#
> have to switch to different kernel mode first to clear NO-CARRIER
>
> root@debian9:~# ip link set dev team0 down
> root@debian9:~# ip link set dev ens9 nomaster
> root@debian9:~# ip link set dev ens10 nomaster
> root@debian9:~# teamd --team-dev team0 --kill
> root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> --config='{"runner":{"name":"roundrobin"}}'
> This program is not intended to be run as root.
> root@debian9:~# teamd --team-dev team0 --kill
> root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> --config='{"runner":{"name":"loadbalance"}}'
> This program is not intended to be run as root.
> root@debian9:~# ip link set dev ens9 master team0
> root@debian9:~# ip link set dev ens10 master team0
> root@debian9:~# ip link set dev team0 up
> root@debian9:~# ip link show team0
> 8: team0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP
> mode DEFAULT group default qlen 1000
>    link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
> root@debian9:~#
>
> This is with debian9, teamd version 1.26-1+b1, kernel 4.9.13-1
> > > carrier is set to down in kernel. lacp and loadbalance runners both use same
> > > kernel mode "loadbalance", so starting loadbalance runner does not change
> > > kernel mode. Kernel still has "user_carrier_enabled" and team interface stays
> > > down due to NO_CARRIER.
> >
> > Hmm. I think that the correct solution is to teach the runners not to
> > depend on kernel defaults. So in this care the loadbalance runner would
> > set user_carrier_enabled to false during init (not during takeover)
>
> ok, thanks.
> > > It seems the only way to clear kernel "user_carrier_enabled" is to change the
> > > mode. Should there be a new option that teamd can use to tell kernel to clear
> > > "user_carrier_enabled" when lacp runner terminates? Or any other suggestion?
> > >
> > > Thanks.
> > > _______________________________________________
> > > libteam mailing list -- [e-mail]
> > > To unsubscribe send an email to [e-mail]

[George Wilkie]

On 03/31/2017 04:01 PM, Jiri Pirko wrote:
> Fri, Mar 31, 2017 at 01:41:08PM CEST, [e-mail] wrote:
> > On 03/30/2017 04:18 PM, Jiri Pirko wrote:
> > > Tue, Mar 28, 2017 at 06:51:09PM CEST, [e-mail] wrote:
> > > > Hi, like to get your thoughts on this issue we are running into.
> > > >
> > > > Initially, running the lacp runner in teamd. This sends carrier changes to
> > > > the kernel. Kernel sets "user_carrier_enabled" which blocks its normal
> > > > carrier checks.
> > > >
> > > > Then change teamd runner from lacp to loadbalance. When lacp runner exits,
> > >
> > > Hmm. What exactly are you doing, could you send a list of commands?
> > > I suspect some oddities :)
> > start off in LACP mode - NO-CARRIER expected as no LACP on other side.
> >
> > root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> > --config='{"runner":{"name":"lacp"}}'
> > This program is not intended to be run as root.
> > root@debian9:~# ip link set dev ens9 master team0
> > root@debian9:~# ip link set dev ens10 master team0
> > root@debian9:~# ip link set dev team0 up
> > root@debian9:~# ip link show team0
> > 8: team0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc noqueue state
> > DOWN mode DEFAULT group default qlen 1000
> >     link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
> > root@debian9:~#
> > change to balanced mode - still NO-CARRIER
> >
> > root@debian9:~# ip link set dev team0 down
> > root@debian9:~# ip link set dev ens9 nomaster
> > root@debian9:~# ip link set dev ens10 nomaster
> > root@debian9:~# teamd --team-dev team0 --kill
> > root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
>
> I think that you might be missing the reason of existence of take over.
> It suppose to "take over" the kernel team instance in case for example
> the teamd segfaults or something. The intension was to run new teamd
> instance taking over the kernel instance with exact same config.
>
> Why are you doing this?

Want to change the teamd config without losing the interface config 
(addresses, vlans etc) - want to avoid deleting the team device when 
updating team config.

Cheers.

[Jiri Pirko]

Fri, Mar 31, 2017 at 06:01:46PM CEST, [e-mail] wrote:
> On 03/31/2017 04:01 PM, Jiri Pirko wrote:
> > Fri, Mar 31, 2017 at 01:41:08PM CEST, [e-mail] wrote:
> > > On 03/30/2017 04:18 PM, Jiri Pirko wrote:
> > > > Tue, Mar 28, 2017 at 06:51:09PM CEST, [e-mail] wrote:
> > > > > Hi, like to get your thoughts on this issue we are running into.
> > > > >
> > > > > Initially, running the lacp runner in teamd. This sends carrier changes to
> > > > > the kernel. Kernel sets "user_carrier_enabled" which blocks its normal
> > > > > carrier checks.
> > > > >
> > > > > Then change teamd runner from lacp to loadbalance. When lacp runner exits,
> > > >
> > > > Hmm. What exactly are you doing, could you send a list of commands?
> > > > I suspect some oddities :)
> > > start off in LACP mode - NO-CARRIER expected as no LACP on other side.
> > >
> > > root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> > > --config='{"runner":{"name":"lacp"}}'
> > > This program is not intended to be run as root.
> > > root@debian9:~# ip link set dev ens9 master team0
> > > root@debian9:~# ip link set dev ens10 master team0
> > > root@debian9:~# ip link set dev team0 up
> > > root@debian9:~# ip link show team0
> > > 8: team0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc noqueue state
> > > DOWN mode DEFAULT group default qlen 1000
> > >     link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
> > > root@debian9:~#
> > > change to balanced mode - still NO-CARRIER
> > >
> > > root@debian9:~# ip link set dev team0 down
> > > root@debian9:~# ip link set dev ens9 nomaster
> > > root@debian9:~# ip link set dev ens10 nomaster
> > > root@debian9:~# teamd --team-dev team0 --kill
> > > root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> >
> > I think that you might be missing the reason of existence of take over.
> > It suppose to "take over" the kernel team instance in case for example
> > the teamd segfaults or something. The intension was to run new teamd
> > instance taking over the kernel instance with exact same config.
> >
> > Why are you doing this?
>
> Want to change the teamd config without losing the interface config 
> (addresses, vlans etc) - want to avoid deleting the team device when 
> updating team config.

Is it like a regular runtime thing for you? I would expect that you just
do the configuration once and you use the same one forever.

[George Wilkie]

On 04/03/2017 02:14 PM, Jiri Pirko wrote:
> Fri, Mar 31, 2017 at 06:01:46PM CEST, [e-mail] wrote:
> > On 03/31/2017 04:01 PM, Jiri Pirko wrote:
> > > Fri, Mar 31, 2017 at 01:41:08PM CEST, [e-mail] wrote:
> > > > On 03/30/2017 04:18 PM, Jiri Pirko wrote:
> > > > > Tue, Mar 28, 2017 at 06:51:09PM CEST, [e-mail] wrote:
> > > > > > Hi, like to get your thoughts on this issue we are running into.
> > > > > >
> > > > > > Initially, running the lacp runner in teamd. This sends carrier changes to
> > > > > > the kernel. Kernel sets "user_carrier_enabled" which blocks its normal
> > > > > > carrier checks.
> > > > > >
> > > > > > Then change teamd runner from lacp to loadbalance. When lacp runner exits,
> > > > >
> > > > > Hmm. What exactly are you doing, could you send a list of commands?
> > > > > I suspect some oddities :)
> > > > start off in LACP mode - NO-CARRIER expected as no LACP on other side.
> > > >
> > > > root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> > > > --config='{"runner":{"name":"lacp"}}'
> > > > This program is not intended to be run as root.
> > > > root@debian9:~# ip link set dev ens9 master team0
> > > > root@debian9:~# ip link set dev ens10 master team0
> > > > root@debian9:~# ip link set dev team0 up
> > > > root@debian9:~# ip link show team0
> > > > 8: team0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc noqueue state
> > > > DOWN mode DEFAULT group default qlen 1000
> > > >      link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
> > > > root@debian9:~#
> > > > change to balanced mode - still NO-CARRIER
> > > >
> > > > root@debian9:~# ip link set dev team0 down
> > > > root@debian9:~# ip link set dev ens9 nomaster
> > > > root@debian9:~# ip link set dev ens10 nomaster
> > > > root@debian9:~# teamd --team-dev team0 --kill
> > > > root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> > >
> > > I think that you might be missing the reason of existence of take over.
> > > It suppose to "take over" the kernel team instance in case for example
> > > the teamd segfaults or something. The intension was to run new teamd
> > > instance taking over the kernel instance with exact same config.
> > >
> > > Why are you doing this?
> >
> > Want to change the teamd config without losing the interface config
> > (addresses, vlans etc) - want to avoid deleting the team device when
> > updating team config.
>
> Is it like a regular runtime thing for you? I would expect that you just
> do the configuration once and you use the same one forever.

I don't think I could say it would be a regular thing done in production -
more like something you do during setup or troubleshooting.

Do you think this will get addressed or is it something we will have to
live with?

Thx.

[Jiri Pirko]

Mon, Apr 03, 2017 at 07:07:31PM CEST, [e-mail] wrote:
> On 04/03/2017 02:14 PM, Jiri Pirko wrote:
> > Fri, Mar 31, 2017 at 06:01:46PM CEST, [e-mail] wrote:
> > > On 03/31/2017 04:01 PM, Jiri Pirko wrote:
> > > > Fri, Mar 31, 2017 at 01:41:08PM CEST, [e-mail] wrote:
> > > > > On 03/30/2017 04:18 PM, Jiri Pirko wrote:
> > > > > > Tue, Mar 28, 2017 at 06:51:09PM CEST, [e-mail] wrote:
> > > > > > > Hi, like to get your thoughts on this issue we are running into.
> > > > > > > 
> > > > > > > Initially, running the lacp runner in teamd. This sends carrier changes to
> > > > > > > the kernel. Kernel sets "user_carrier_enabled" which blocks its normal
> > > > > > > carrier checks.
> > > > > > > 
> > > > > > > Then change teamd runner from lacp to loadbalance. When lacp runner exits,
> > > > > > Hmm. What exactly are you doing, could you send a list of commands?
> > > > > > I suspect some oddities :)
> > > > > start off in LACP mode - NO-CARRIER expected as no LACP on other side.
> > > > >
> > > > > root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> > > > > --config='{"runner":{"name":"lacp"}}'
> > > > > This program is not intended to be run as root.
> > > > > root@debian9:~# ip link set dev ens9 master team0
> > > > > root@debian9:~# ip link set dev ens10 master team0
> > > > > root@debian9:~# ip link set dev team0 up
> > > > > root@debian9:~# ip link show team0
> > > > > 8: team0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc noqueue state
> > > > > DOWN mode DEFAULT group default qlen 1000
> > > > >      link/ether 52:54:00:8b:12:a5 brd ff:ff:ff:ff:ff:ff
> > > > > root@debian9:~#
> > > > > change to balanced mode - still NO-CARRIER
> > > > >
> > > > > root@debian9:~# ip link set dev team0 down
> > > > > root@debian9:~# ip link set dev ens9 nomaster
> > > > > root@debian9:~# ip link set dev ens10 nomaster
> > > > > root@debian9:~# teamd --team-dev team0 --kill
> > > > > root@debian9:~# teamd --team-dev team0 --daemon --no-quit-destroy --take-over
> > > >
> > > > I think that you might be missing the reason of existence of take over.
> > > > It suppose to "take over" the kernel team instance in case for example
> > > > the teamd segfaults or something. The intension was to run new teamd
> > > > instance taking over the kernel instance with exact same config.
> > > >
> > > > Why are you doing this?
> > >
> > > Want to change the teamd config without losing the interface config
> > > (addresses, vlans etc) - want to avoid deleting the team device when
> > > updating team config.
> >
> > Is it like a regular runtime thing for you? I would expect that you just
> > do the configuration once and you use the same one forever.
>
> I don't think I could say it would be a regular thing done in production -
> more like something you do during setup or troubleshooting.

I don't understand why is it a problem to re-set to iface again in this
case.
> Do you think this will get addressed or is it something we will have to
> live with?

Currently I don't really see reason for changing the behaviour.
If anything, I would add a check in takeover for the team driver
instance options. So that would just fail in case you would try to do
takeover with different setup.