config files and how to have persistent Linux kernel Driver/File System configuration info saved

[Dave Chinner]

On Thu, Jun 28, 2018 at 06:24:59PM -0500, Steve French wrote:
> On Thu, Jun 28, 2018 at 6:21 PM ronnie sahlberg
> <ronniesahlberg at gmail.com> wrote:
> >
> > On Fri, Jun 29, 2018 at 8:58 AM, Theodore Y. Ts'o via samba-technical
> > <samba-technical at lists.samba.org> wrote:
> > > On Thu, Jun 28, 2018 at 05:37:15PM -0500, Steve French wrote:
> > >> Ronnie brought up an interesting point about the problems consistently
> > >> configuring file systems (or any Linux module for that matter) so that
> > >> reboot doesn't wipe away security or performance tuning changes.
> > >
> > > In general it's considered best practice to make the file system
> > > auto-tune itself as much as possible, because the sad fact is that
> > > 99.9999% of the customers aren't going to bother to add any tuning
> > > parameters.  So there hasn't been a push to try to create something
> > > more complex, because it's generally not needed.
> >
> > True, but in these cases I think we are more looking at server or
> > mountpoint specific options than
> > actual fs tuning.
> >
> > For example nfsmount.conf can be used to say "only use NFSv4 when
> > accessing server abc" etc.
> > For the case of CIFS I could imagine that an administrator might want
> > to set "disable smb1 protocol globally"
> 
> Or perhaps
>   "disable smb1 on " ... various public networks but allow it on
> private networks

The way the policy is configured depends on the mechanism used to
configure the policy. If it's a sysctl or a mount option, then we've
already got everything we need. If it's something dynamic in sysfs,
then I think you're on your own.

FYI, I have been looking at making sysctl be able to work on /sys
rather than just /proc/sys (I have a 10 line hack to enable it) so
we could re-use it with custom per-mount error config files in
/etc/xfs/ for XFS that we inject based on a uevent delivered to
udev. It works, but the fact is modifying sysctl in this way exposes
it to a whole bunch of stuff sysctl doesn't understand, shouldn't be
accessing and/or trying modify. i.e. sysctl is disturbingly dumb,
and it gets away with it because of it's restricted scope and API
presented by /proc/sys.

So, really, I'm probably just going roll our own sysfs config file
mechanism into xfs_spaceman (probably based on the new config file
parser we have for mkfs.xfs) and hide the mess with a nice, simple
xfs_admin interface for udev to call. i.e. roll our own :)

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com