[systemd-devel] Handling sysctl entries that depend on module presence?

[Lennart Poettering]

On Mon, 19.11.12 08:13, Mantas Mikulėnas (grawity at gmail.com) wrote:

> On Mon, Nov 19, 2012 at 8:01 AM, Mantas Mikulėnas <grawity at gmail.com> wrote:
> > Anything that is written to /sys/module/*/parameters goes to
> > modprobe.d(5) -- again, to be applied when the module is inserted, not
> > at some fixed point:
> >
> >   https://github.com/vodik/powersave/blob/master/modprobe.d/powersave.conf
> >
> > Use sysctl.conf(5) for settings under /proc/sys:
> >
> >   https://github.com/vodik/powersave/blob/master/sysctl.d/powersave.conf
> 
> This reminded me that systemd-sysctl doesn't properly handle
> sysctl.conf entries that depend on the presence of modules loaded by
> modules-load.d...
> 
> In particular, I was having problems with:
> 
>     net.ipv4.tcp_allowed_congestion_control = cubic reno lp
> 
> This setting is required to allow non-root programs (the Transmission
> BitTorrent daemon in particular) to use the "lowest priority"
> congestion control algorithm.
> 
> However, if the "tcp_lp" module is not yet loaded, the write to
> /proc/sys is simply refused with EINVAL (if I remember correctly).
> 
> And if the module is loaded later (or not loaded manually), without
> the sysctl setting being updated, Transmission receives EPERM when it
> calls setsockopt(TCP_CONGESTION).
> 
> This doesn't affect TCPCongestion= in systemd socket units, since root
> is allowed to set any algorithm, and the module is simply loaded on
> demand when systemd asks for it...
> 
> How can this be fixed?

Hmm, whenever a module is loaded this also generates udev events, maybe
for this specific case we should add a udev rule that tries to write the
sysctl again after the module was successfully loaded?

Kay?

Lennart

-- 
Lennart Poettering - Red Hat, Inc.