Re: [CFT] bsdinstall and zfsboot enhancements

On 12/01/13 07:34, Jilles Tjoelker wrote:
> On Sat, Nov 30, 2013 at 04:36:18PM -0600, Nathan Whitehorn wrote:
>> This took much longer than I'd anticipated, but the patch to init is
>> attached. I chose not to make the changes to init rather than
>> getttyent() and friends in libc, which I am open to revisiting.
> lib/libpam/modules/pam_securetty/pam_securetty.c calls getttynam(3) and
> will not allow root login on a "fake" TTY that getttynam() does not
> know. This module is enabled by default for the "login" service.
>
> So it is probably better to patch libc rather than init.

OK, here's a revised patch. This one is shorter and works by introducing
an "auto" flag (ideas for names appreciated) that means "on" if the line
is an active console and "off" otherwise. Note that the behavior is now:
- ttys marked "off" stay off
- ttys marked "on" stay on
- ttys marked "auto" are enabled iff they are console devices
- ttys not present in /etc/ttys stay off

This behavior change is much easier to implement when doing it in libc
for various structural reasons and allows the terminal type, etc. to be
specified in the usual way.

>> The behavior changes are as follows:
>> If the "console" device in /etc/ttys in marked "on", instead of opening
>> /dev/console, init will loop through the active kernel console devices,
>> and for each will:
>> 1. If the kernel console device is in /etc/ttys and marked "on", it
>> already has a terminal and will be ignored.
>> 2. If marked "off", that is an explicit statement that a console is not
>> wanted and so it will be ignored.
>> 3. If not present in /etc/ttys, init will run getty with whatever
>> parameters "console" has.
> This seems to make sense.
>
>> (3) is the main behavioral change. No changes in behavior will occur if
>> /etc/ttys is not modified. If we turn on "console" by default, it will
>> usually have no effect instead of trying to run multiple gettys, which
>> is new. If we then also comment out the ttyu0 line, instead of marking
>> it "off", the result will be the conditional presence of a login prompt
>> on the first serial port depending on whether it is an active console
>> device for the kernel. I believe this is the behavior we are going for.
> The terminal type for the console entry should probably be changed to
> something other than "unknown" to reduce annoyance.
>
>> Comments and test results would be appreciated.
> As a preparatory patch, you could remove se_index and session_index from
> init. They are only used to warn about a changed slot number in utmp(5)
> which is irrelevant with utmpx. This noise warning would also appear
> in most cases when changing from a "fake" console entry to a real line
> in /etc/ttys. Also, if you do decide to fake ttys entries in init rather
> than libc, the patch to init will be simpler.
>

With the new patch, this is indeed the case: no changes to init are
necessary at all. This does not change any behavior unless explicitly
requested in /etc/ttys, so unless there are any objections in the next
couple days, I will commit it.
-Nathan