Re: [PATCH v1] module: Add support for default value for module async_probe

[Luis Chamberlain]

On Fri, Jun 03, 2022 at 04:16:43PM -0700, Saravana Kannan wrote:
> On Fri, Jun 3, 2022 at 8:01 AM Luis Chamberlain <mcgrof@xxxxxxxxxx> wrote:
> >
> > On Thu, Jun 02, 2022 at 10:54:41PM -0700, Saravana Kannan wrote:
> > > Add a module.async_probe kernel command line option that allows enabling
> > > async probing for all modules. When this command line option is used,
> > > there might still be some modules for which we want to explicitly force
> > > synchronous probing, so extend <modulename>.async_probe to take an
> > > optional bool input so that async probing can be disabled for a specific
> > > module.
> > >
> > > Signed-off-by: Saravana Kannan <saravanak@xxxxxxxxxx>
> > > ---
> > >  Documentation/admin-guide/kernel-parameters.txt |  8 ++++++--
> > >  kernel/module/main.c                            | 11 ++++++++++-
> > >  2 files changed, 16 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
> > > index 710b52d87bdd..32083056bd25 100644
> > > --- a/Documentation/admin-guide/kernel-parameters.txt
> > > +++ b/Documentation/admin-guide/kernel-parameters.txt
> > > @@ -1147,8 +1147,12 @@
> > >       nopku           [X86] Disable Memory Protection Keys CPU feature found
> > >                       in some Intel CPUs.
> > >
> > > -     <module>.async_probe [KNL]
> > > -                     Enable asynchronous probe on this module.
> > > +     <module>.async_probe[=<bool>] [KNL]
> > > +                     If no <bool> value is specified or if the value
> > > +                     specified is not a valid <bool>, enable asynchronous
> > > +                     probe on this module.  Otherwise, enable/disable
> > > +                     asynchronous probe on this module as indicated by the
> > > +                     <bool> value.
> >
> > The commit log says a bit more. Can you clarify this on the
> > documentation?
> 
> Oh yeah, forgot to add module.async_probe there! Will do.
> 
> > We should strive slowly towards more async probes. This will take
> > time.
> 
> Agreed.
> 
> > To help with further then a Kconfig option which sets this
> > to a default to true if enabled would be useful so that no kernel
> > parameter is needed at all to set the default. Then you can
> > override the default, and blacklist each driver as well.
> 
> Based on Linus's view in this thread [1] (I see his point), I don't
> think we'll ever enable default async probes for modules  as a compile
> time config. 

OK that's fair. But the position there was in reference to *not*
regress userspace.

If we have new tech we can do things differently if the expectation
is set from the beginning, ie if userspace *does* expect dependencies
to be dealt with differently. And systems which are old can be
deprecated with ACPI legacy flags.

> I think it has to be an explicit decision by whoever
> decides the list of modules being loaded in the system (OEMs in the
> case of Android, end user in the case of a PC?) to enable the default
> to be async probe and then the same entity can decide which modules to
> force sync probe on. So, I'm not sure we want to add a Kconfig for
> this or enable it by default. Let me know what you think. I'll send
> out a v2 with the doc fixes in the meantime.

Yeah makes sense.

> On a related note, I'm working on default async probes for built-in
> drivers,

Cool! I'm very excited to hear how that goes!

> but that's feasible to turn on by default because we can
> synchronize everything before we jump to init.

That's cheap, in that I think we can do better. Sure it works.

> And then
> <module>.async_probe needs to be passed explicitly for any modules we
> want to allow async on.

Sure..

  Luis