[systemd-devel] systemd inquiry

[Colin Guthrie]

'Twas brillig, and Colin Guthrie at 10/04/12 13:57 did gyre and gimble:
> 'Twas brillig, and Mark Hounschell at 10/04/12 13:36 did gyre and gimble:
>> On 04/09/2012 08:06 PM, Colin Guthrie wrote:
>>
>>>> Yep, that works. Can the NAutoVTs be set differently on a per target
>>>> basis?
>>>
>>> Not as far as I know, but you should be able to do something similar via
>>> a conflicts directive.
>>>
>>> e.g. if you have NAutoVTs=6 by default you can just put:
>>> Conflicts=autovt at tty1.service autovt at tty2.service ... autovt at tty6.service
>>>
>>> This should prevent then kicking in. That said, I'm really not sure how
>>> much you gain here. They are loaded on demand afterall, so it's not like
>>> they are buring CPU cycles etc. Personally it doesn't seem worth the
>>> bother to me, but maybe you have your reasons :)
>>>
>>
>> Again, thanks for the help. I do have my reasons but they are not really
>> relevant. I will play with the Conflicts directive.
>>
>> I am having another issue with an out of kernel "GPL" device driver not
>> being available "on time" so to say. When the kernel discovers this pci
>> card it loads it's kernel module and sets up the card for use. This
>> takes around 15 seconds per card and there is usually 2-3 of them. When
>> the card is up, running, and ready, the kernel module notifies udev who
>> in turn executes a small script that creates 30 or so different device
>> nodes for use with the card. This little script is not a systemd service
>> nor a sysvinit script. When I use sysvinit, (maybe by luck) all this
>> happens well before any app gets to run in my dedicated run-level. Using
>> systemd it does not appear to. What does the udev-settle.service do? Can
>> it help me here somehow or should I just assume that I will have to turn
>> this script into a systemd service?
> 
> 
> I suspect you want to wait for udev-settle.service before running
> anything that needs it.
> 
> It should ensure that the udev event queue is fully processed and AFAIK,
> this shoudl include your service.
> 
> Note that the default udev-settle timeout is 120s and systemd has a
> higher timeout of 180s on running the unit itself. Depending on your use
> case you may need to copy udev-settle.service to another unit (call it
> what you want) and adjust both the Timeout value in the unit as run by
> systemd and the inner timeout (as a --timeout argument) when calling
> udevadm settle.
> 
> That said, systemd also has a concept of device units. You could create
> a device unit that becomes ready when your udev script is run, that way
> it can be used for ordering without having to run udev-settle.service (I
> believe). See man "systemd.device"
> 
> You can order your device unit "Before=getty at tty1.service" or similar
> such that the login prompt will only appear when the devices are ready.
> Of course depending on how many devices you have you may need to create
> several instances of them linked in your test.target.wants directory.

As Kay has replied also here, I should probably point out that this is
likely wrong.

As your post script makes it's own devices via mknod I'm presuming udev
will be unaware of them and thus the device unit in systemd will simply
not work.

I'm sure Kay will correct me if I'm wrong here.

There are plenty other hacky ways around it tho'. e.g write a service
that runs a special sleep program (create a symlink:
/usr/bin/wait-for-special-dev-nodes to /bin/sleep) and have a unit that
contains:


Type=oneshot
ExecStartPre=-/usr/bin/wait-for-special-dev-nodes 180
ExecStart=/bin/true
RemainAfterExit=true

Then in your script run from udev, just do "killall
wait-for-special-dev-nodes" or similar at the end to kill off the sleep
process and allow that unit to complete and any ordering related to it
to be fed back to other dependant units etc. (note that rather than make
a symlink you could likely call sleep directly and just make your kill
command a bit nicer - i.e. only kill the sleep process that is tagged
with your unit's cgroup - that's likely nicer than a symlink, but it's
harder to explain/test in an email!)

This is still horribly hacky and there are likely nicer ways to do it
(i.e. do whatever Kay says is usually a good rule here!). I just wanted
to highlight to you that there are usually ways to make old stuff at
least play semi-nicely with all the ordering and graphing goodness in
systemd :)


Col

-- 

Colin Guthrie
gmane(at)colin.guthr.ie
http://colin.guthr.ie/

Day Job:
  Tribalogic Limited http://www.tribalogic.net/
Open Source:
  Mageia Contributor http://www.mageia.org/
  PulseAudio Hacker http://www.pulseaudio.org/
  Trac Hacker http://trac.edgewall.org/