Re: [ndctl PATCH 0/7] Policy based reconfiguration for daxctl

["Verma, Vishal L" <vishal.l.verma@intel.com>]

On Thu, 2021-09-16 at 15:12 -0700, Dan Williams wrote:
> On Tue, Aug 31, 2021 at 2:05 AM Vishal Verma <vishal.l.verma@intel.com> wrote:
> > 
> > These patches add policy (config file) support to daxctl. The
> > introductory user is daxctl-reconfigure-device. Sysadmins may wish to
> > use daxctl devices as system-ram, but it may be cumbersome to automate
> > the reconfiguration step for every device upon boot.
> > 
> > Introduce a new option for daxctl-reconfigure-device, --check-config.
> > This is at the heart of policy based reconfiguration, as it allows
> > daxctl to look up reconfiguration parameters for a given device from the
> > config system instead of the command line.
> > 
> > Some systemd and udev glue then automates this for every new dax device
> > that shows up, providing a way for the administrator to simply list all
> > the 'system-ram' UUIDs in a config file, and not have to worry about
> > anything else.
> > 
> > An example config file can be:
> > 
> >   # cat /etc/ndctl/daxctl.conf
> 
> Take these comments as provisional until I read through the rest, but
> this is just a reaction to the proposed ini format.

I somehow missed this email originally, and just saw it on lore..

> > 
> >   [auto-online unique_identifier_foo]
> 
> I am thinking this section name should be "reconfigure-device
> unique_identifier_foo" if only because resize might also be something
> someone wants to do, and if other commands get config automation it
> makes it clearer which config snippets apply to which command.

Yep that makes sense - I'll change this.
> 
> >   uuid = 48d8e42c-a2f0-4312-9e70-a837faafe862
> 
> I think this should be called:
> 
> "nvdimm.uuid"
> 
> ...or something like that to make it clear this depends on dax devices
> emitted by libnvdimm, and not those that come from "soft-reserved"
> memory. It also helps distinguish if we ever get UUIDs in the HMAT
> which is something I have been meaning to propose.

Yep makes sense, will change.

> >   mode = system-ram
> 
> I can see this being "mode = devdax" if feature was being used to
> change size or alignment.

Agreed, but that should 'just work' right - especially once we rename
the section name from auto-online to reconfigure-device.

> 
> >   online = true
> >   movable = false
> 
> I wonder if these keys should be prefixed by the mode name:
> 
> system-ram.online = true
> system-ram.movable = false

Hm, maybe, but since the config options feed directly into the commands
params, I figured we can let the command's option parsing throw any
errors for incompatible options. My hope was for config identifiers to
- as far as possible - exactly match command-line options. If we do
make changes like this, I feel for every command that supports config,
the man page is asking for a dedicated config section documenting every
config option it supports. Maybe this is a good idea regardless :)

> 
> ...so it's a bit more self documenting about which parameters are
> sub-options, and delineates them from generic options like size.
> 
> > Any file under '/etc/ndctl/' can be used - all files with a '.conf' suffix
> > will be considered when looking for matches.
> 
> any concern about name collisions between ndctl, daxctl, and cxl-cli
> section names?

Yep good point, changed this to be in /etc/daxctl/, and cxl-cli can get
its own directory too when the time comes.