7:02 a.m.
On Tue, Sep 26, 2023 at 11:23:51AM +0100, Daniel P. Berrangé wrote:
On Tue, Sep 26, 2023 at 11:09:44AM +0200, Pavel Hrdina wrote:
> On Mon, Sep 25, 2023 at 08:58:24PM +0200, Andrea Bolognani wrote:
> > + 'service_unit_extra': [
> > + 'Wants=systemd-machined.service',
> > + 'After=systemd-machined.service',
> > + 'After=remote-fs.target',
> > + ],
> > + 'service_service_extra': [
> > + 'KillMode=process',
> > + '# Raise hard limits to match behaviour of systemd >= 240.',
> > + '# During startup, daemon will set soft limit to match hard
limit',
> > + '# per systemd recommendations',
> > + 'LimitNOFILE=1024:524288',
> > + '# The cgroups pids controller can limit the number of tasks started
by',
> > + '# the daemon, which can limit the number of domains for some
hypervisors.',
> > + '# A conservative default of 8 tasks per guest results in a TasksMax
of',
> > + '# 32k to support 4096 guests.',
> > + 'TasksMax=32768',
> > + '# With cgroups v2 there is no devices controller anymore, we have
to use',
> > + '# eBPF to control access to devices. In order to do that we create
a eBPF',
> > + '# hash MAP which locks memory. The default map size for 64 devices
together',
> > + '# with program takes 12k per guest. After rounding up we will get
64M to',
> > + '# support 4096 guests.',
> > + 'LimitMEMLOCK=64M',
> > + ],
>
> This feels wrong to have it in meson.build file. In addition it is the
> same as for virtlxcd and virtqemud so we are basically duplicating the
> data and which makes it easy to make inconsistent changes not affecting
> all places.
>
> IMHO it would be better to have additional file that will be included
> into the template for services where we need it.
>
> I'm not sure about the `service_unit_extra` as well if we want to have
> it in meson.build files as it is not strictly related to the build
> process and there is more data compared to the old `deps`.
If anything I'd reverse the model. The 'virtchd.service.in' file
should be the primary template, the common bits the injected data.
ie
cat virtchd.service.in
[Unit]
Description=Virtualization Cloud-Hypervisor daemon
::common-unit::
Wants=systemd-machined.service
After=remote-fs.target
After=systemd-machined.service
Documentation=man:virtchd(8)
[Service]
::common-service::
KillMode=process
# Raise hard limits to match behaviour of systemd >= 240.
# During startup, daemon will set soft limit to match hard limit
# per systemd recommendations
LimitNOFILE=1024:524288
# The cgroups pids controller can limit the number of tasks started by
# the daemon, which can limit the number of domains for some hypervisors.
# A conservative default of 8 tasks per guest results in a TasksMax of
# 32k to support 4096 guests.
TasksMax=32768
# With cgroups v2 there is no devices controller anymore, we have to use
# eBPF to control access to devices. In order to do that we create a eBPF
# hash MAP which locks memory. The default map size for 64 devices together
# with program takes 12k per guest. After rounding up we will get 64M to
# support 4096 guests.
LimitMEMLOCK=64M
[Install]
::common-install::
This doesn't address the problem with duplication that Pavel pointed
out.
I don't think it helps much with not storing additional data inside
the build system, unless we want to store the contents of the various
common snippets in separate files? Something like
common_service = fs.read('common_service.inc')
unit_conf = configuration_data({
'common_service' = common_service,
})
We'd have to fake fs.read() because it was introduced in 0.57 though.
And we'd have to run the contents of the common parts through
variable substitution anyway, because they will contain a bunch of
lines like
Also=@service@.socket
Also=@service(a)-ro.socket
Also=@service(a)-admin.socket
I'm not sure the result would look much better, but I can give it a
try.
arguably we don't even need the '::common-XXX::' lines in
there. We can
simply see the headers [Unit], [Service], etc and inject the common
bits under each header.
I think markers make things both easier to implement and more obvious
(whoever looks at the file can immediately tell that some sort of
post-processing is going to happen and probably even make a fairly
accurate guess as what it will entail), so I'd prefer having them.
But this is a fairly minor detail compared to the above.
--
Andrea Bolognani / Red Hat / Virtualization