Letting start sshd in an early stage on ARM devices

[Dr. Rolf Jansen]

> Am 19.09.2018 um 10:07 schrieb Ian Lepore <ian at freebsd.org>:
> 
> On Tue, 2018-09-18 at 21:48 -0300, Dr. Rolf Jansen wrote:
>> Last week I had the first official presentation of my ADC/DAC-project 
>> with a BeagleBone Black running FreeBSD 12.0-ALPHA5.
>> 
>> I experienced a problem with the NFS client, which was sort of home
>> made problem, because in the hurry I forgot to deactivate the mount-
>> directive of a nfs share in fstab, and when I started the BBB before
>> the presentation, it was not connected to it's home-network and it
>> hang at that point. I built it into a very tight box and there was no
>> place anymore for the FTDI serial connector, and since sshd is
>> usually loaded as one of the last services, I was effectively locked
>> out of the BBB. Fortunately, I overcame the problem by simulating the
>> nfs share with my notebook and after restart, I was able to fix the
>> fstab right before the actual presentation began, so nobody saw that
>> I had a problem - anyway, I would prefer to never become that nervous
>> again. 
>> 
>> My question is now, why is sshd set to start so very late in the
>> booting process? If we want to put autonomous ARM devices somewhere
>> into the field, then any hick-up in the startup sequence would leave
>> us out-locked.
>> 
>> For testing purposes, I changed the sshd rc script by replacing the
>> line starting with # REQUIRE: ... by:
>> 
>>    # REQUIRE: ipfw
>>    # BEFORE: mountcritremote
>> 
>> Now, sshd starts right after ipfw has been set up, and in case the
>> BBB hangs, e.g. in the course of mounting a nfs share, I am still
>> able to login via ssh and fix most of the issues.
>> 
>> I would like to leave it like this (at least on the headless ARM
>> devices). Are there any risks or hidden problems which I might
>> experience, when having sshd running very early in the boot sequence?
>> Of course after any FreeBSD update, I would need to check the sshd rc
>> script.
>> 
>> Best regards
>> 
>> Rolf
> 
> I suspect you'd get better answers to this on a more generic mailing
> list such as -hackers or -current, because it's not really arm
> specific. Here's what I can figure out just looking at the scripts
> involved...
> 
> The sshd script requires FILESYSTEMS and LOGIN.  The FILESYSTEMS
> requirement seems pretty important... people can't log in until things
> like /usr and /home are available. Using mountcritlocal instead would
> leave out ZFS, and that won't work for a lot of people.  The LOGIN
> requirement is explained like this in the script:
> 
>  # This is a dummy dependency to ensure user services such as xdm,
>  # inetd, cron and kerberos are started after everything else, in case
>  # the administrator has increased the system security level and
>  # wants to delay user logins until the system is (almost) fully
>  # operational.
> 
> That is probably important to some people. Other things that come
> before sshd because of LOGINS also seem important... things like
> syslogd and ntpd (must have valid timestamps on files and in logs
> before allowing users to do things).  Just in general, if you run
> "rcorder /etc/rc.d/* /usr/local/etc/rc.d/*" and look at everything that
> shows up between FILESYSTEMS and LOGIN, you can see that in a general-
> purpose-computer setup, you want most of those things to happen before
> users are allowed onto the system.
> 
> init(8) allows logins as the last thing that happens in system startup,
> after all rc.d scripts are done. Sshd (and similar things such as inetd
> and xdm/slim) are also ways of allowing logins, and it makes sense to
> me that they are also almost the very last thing that happens in rc
> processing. So it seems unlikely that this will ever change on the
> freebsd side.
> 
> Sometimes embedded systems have needs that are completely different
> than general-purpose-computer and it's appropriate to do things like
> start sshd very early, but that's always going to end up being a custom
> modification that has to be made by the creator of the embedded system.

Ian, I appreciate your insights and your response. I completely agree with your final comment, and actually that's the reason why I did not raise a general PR, and I even don't want to ask for a general change in the rc order. My embedded systems come with a number of other customizations, and in the meantime I already put said changes to /etc/rc.d/sshd into the list of sed commands to be applied when setting-up a new system from the scratch. 

That I did not care to much for a general case, you might see from that I put ipfw as the requirement. The desktop folks seem to prefer pf nowadays. So basically, I am fine with the need of customization. My question was more about possible adverse implications in respect to the given embedded use case - one filesystem, two users root/rolf, with home directories on the root fs - in the given respect, opinions of FreeBSD desktop users would withdraw the focus, and therefore I raised the issue on this list. 

You listed the general reasons on why sshd starts almost last. I will not dispute that. However, I see that the ssh daemon which is launched by the rc script is the mother daemon, i.e. the listener and initial first responder on the ssh port, which spawns sshd childs on each login request. My take on this is, that FILESYSTEM and LOGIN need not to be for the mother but must be present for the spawned childs in order somebody with home on late file systems could actually login. If the mother sshd is up and running already but cannot spawn childs because other pre-requisites are not yet ready, so what? The user can't login unless NFS, ZFS, etc. is up and running, and in the current situation he can't either.

Anyway, once again, all this is more from the point of view of headless systems (my FreeBSD servers, and the BBB - I don't have much experience with FreeBSD desktop systems).

I will keep my changes for the time being, if I experience some bad effects, I will report back.

Many thanks best regards

Rolf