network timeout needed at startup

network timeout needed at startup

[Guillaume Rousse]

When using LDAP (and probably other network resources) to store master
map, the network has to be available at autofs daemon start. As Linux
distributions boot time is currently diminushing (hotplug replaced by
udev, various parallelisation effort), I see various race conditions
occuring: the network service finishes, but the actual network interface
is still in initialisation state, leading to autofs start failure.

The primary problem comes from network service implementation, and
should probably get fixed there, right. However, I also think than
having a configurable network timeout directly in autofs (aka: wait at
least 10 second before considering the server is down) may also help in
other scenarios. For instance, when you start various machine
simultaneously (for instance, after a general power down), and your
server takes longer to boot than your workstation.

With autofs 4, I implemented this in service script. Now than everything
is handled by autoumout daemon, this should get implemented there I think.
-- 
Guillaume Rousse
Projet Estime, INRIA
Domaine de Voluceau
Rocquencourt - B.P. 105
78153 Le Chesnay Cedex - France

Re: network timeout needed at startup

[Ian Kent]

On Wed, 6 Sep 2006, Guillaume Rousse wrote:

> When using LDAP (and probably other network resources) to store master
> map, the network has to be available at autofs daemon start. As Linux
> distributions boot time is currently diminushing (hotplug replaced by
> udev, various parallelisation effort), I see various race conditions
> occuring: the network service finishes, but the actual network interface
> is still in initialisation state, leading to autofs start failure.
> 
> The primary problem comes from network service implementation, and
> should probably get fixed there, right. However, I also think than
> having a configurable network timeout directly in autofs (aka: wait at
> least 10 second before considering the server is down) may also help in
> other scenarios. For instance, when you start various machine
> simultaneously (for instance, after a general power down), and your
> server takes longer to boot than your workstation.

Are you talking about a wait at startup or extending the timeout on mounts 
or adding additional timeout processing for LDAP?

Changing the timeout values for mounts would cause large delays, we've 
been there before with complains about how long it takes for mounts to 
unavailable servers to timeout.

> 
> With autofs 4, I implemented this in service script. Now than everything
> is handled by autoumout daemon, this should get implemented there I think.

At startup in the init script, checking map server availability, we could 
still do such a thing.

Ian

Re: network timeout needed at startup

[Jeff Moyer]

==> Regarding Re: [autofs] network timeout needed at startup; Ian Kent <raven@themaw.net> adds:

raven> On Wed, 6 Sep 2006, Guillaume Rousse wrote:
>> When using LDAP (and probably other network resources) to store master
>> map, the network has to be available at autofs daemon start. As Linux
>> distributions boot time is currently diminushing (hotplug replaced by
>> udev, various parallelisation effort), I see various race conditions
>> occuring: the network service finishes, but the actual network interface
>> is still in initialisation state, leading to autofs start failure.
>> 
>> The primary problem comes from network service implementation, and
>> should probably get fixed there, right. However, I also think than
>> having a configurable network timeout directly in autofs (aka: wait at
>> least 10 second before considering the server is down) may also help in
>> other scenarios. For instance, when you start various machine
>> simultaneously (for instance, after a general power down), and your
>> server takes longer to boot than your workstation.

raven> Are you talking about a wait at startup or extending the timeout on
raven> mounts or adding additional timeout processing for LDAP?

raven> Changing the timeout values for mounts would cause large delays,
raven> we've been there before with complains about how long it takes for
raven> mounts to unavailable servers to timeout.

Consider if your master map is stored in the ldap server.  It sounds to me
like the requested functionality is to be patient when obtaining the
initial map.

>> With autofs 4, I implemented this in service script. Now than everything
>> is handled by autoumout daemon, this should get implemented there I
>> think.

raven> At startup in the init script, checking map server availability, we
raven> could still do such a thing.

True, sounds like a hack, though.  ;)

-Jeff