[PATCH v6 22/22] of/platform: Defer probes of registered devices

[Rafael J. Wysocki]

On Thursday, October 22, 2015 04:27:10 PM Scott Wood wrote:
> On Thu, 2015-10-22 at 15:04 +0200, Tomeu Vizoso wrote:
> > On 22 October 2015 at 00:51, Scott Wood <scottwood at freescale.com> wrote:
> > > On Wed, 2015-10-21 at 08:44 -0500, Rob Herring wrote:
> > > > On Wed, Oct 21, 2015 at 12:54 AM, Scott Wood <scottwood at freescale.com>
> > > > wrote:
> > > > > On Mon, 2015-09-21 at 16:03 +0200, Tomeu Vizoso wrote:
> > > > > > Instead of trying to match and probe platform and AMBA devices right
> > > > > > after each is registered, delay their probes until 
> > > > > > device_initcall_sync.
> > > > > > 
> > > > > > This means that devices will start probing once all built-in drivers
> > > > > > have registered, and after all platform and AMBA devices from the DT
> > > > > > have been registered already.
> > > > > > 
> > > > > > This allows us to prevent deferred probes by probing dependencies on
> > > > > > demand.
> > > > > > 
> > > > > > Signed-off-by: Tomeu Vizoso <tomeu.vizoso at collabora.com>
> > > > > > ---
> > > > > > 
> > > > > > Changes in v4:
> > > > > > - Also defer probes of AMBA devices registered from the DT as they 
> > > > > > can
> > > > > >   also request resources.
> > > > > > 
> > > > > >  drivers/of/platform.c | 11 ++++++++---
> > > > > >  1 file changed, 8 insertions(+), 3 deletions(-)
> > > > > 
> > > > > This breaks arch/powerpc/sysdev/fsl_pci.c.  The PCI bus is an OF 
> > > > > platform
> > > > > device, and it must be probed before pcibios_init() which is a
> > > > > subsys_initcall(), or else the PCI bus never gets scanned.
> > > > 
> > > > Thanks for the report. This is probably getting dropped, but it could
> > > > be disabled for PPC.
> > > 
> > > I don't think that adding another arbitrary arch difference would be the
> > > right solution.
> > 
> > I think Rob meant temporarily disable it while things get fixed. At
> > least, 
> 
> So, what is the permanent fix for the swiotlb issue (or more generally, the 
> inability to have a late_initcall that runs after non-module, non-hotplug 
> platform devices have been probed)?
> 
> > I don't see any reason why PPC wouldn't benefit from this
> > series.
> 
> It's not clear to me what the benefit of this is at all, much less for PPC.   
> What is the fundamental problem with deferred probes?  In the cover letter 
> you say this change saves 2.3 seconds, but where is that time being consumed? 
>  Are the drivers taking too long in their probe function trying to initialize 
> and then deferring, rather than checking for dependencies up front?  Or are 
> there really so many devices and such a pessimal ordering that most of the 
> time is spent iterating through and reordering the list, with each defer 
> happening quickly?
> 
> Even if something different does need to be done at this level, forcing all 
> OF platform devices to be probed at the late_initcall level seems quite 
> intrusive.

Totally agreed.

> You limited it to OF because people complained that other things 
> will break.

Right.

And I'm not sure why that was regarded as a good enough reason to do it.

> Things still broke.

Yes, they did.

> Surely there's a better way to address the 
> problem.  Can't the delay be requested by drivers that might otherwise need 
> to defer (which could be done incrementally, focusing on the worst 
> performance problems), rather than enabling it for everything?

Well, I was suggesting to use an opt-in flag there, but I'm not sure if Tomeu
took that into consideration.

In any case, probing is just one aspect of a deeper issue, which is that
we have no way to represent functional dependencies between devices.

Thanks,
Rafael