Re: [PATCH] spi/spi_s3c64xx: Move to subsys_initcall()

[Grant Likely <grant.likely@secretlab.ca>]

On Wed, Sep 08, 2010 at 05:22:51PM +0100, Mark Brown wrote:
> On Wed, Sep 08, 2010 at 10:12:45AM -0600, Grant Likely wrote:
> 
> [reflowed into 80 columns]

Sorry about that.  I'm in the process of switching to mutt+vim for my
mailer.  gmail is just not efficient enough when managing patches.

> > ... but it seems to me that there is a systemic problem in the way the
> > driver model is being used if SPI (and I2C) bus drivers need to be
> > 'special' in this regard.  What are the ordering requirements on things
> > like PMICs?  (My uneducated assumption is that other devices depend on
> > them being enabled before being probed.)  Would it be better to have a
> 
> Pretty much this, yes - if you might want to turn on your supplies
> when you're probed it's rather helpful if the subsystem needed to
> control the regulators is available when you probe.
> 
> > mechanism to defer probing on certain devices until other devices are
> > probed?  Then the relationships could be made explicit instead of the
> > rather coarse-grained (and potentially fragile) approach of changing the
> > init order.
> 
> That would be much nicer (and this is far from the only case where we
> need to deal with it) but it's a substantial change to core kernel
> infrastructure so it's being worked around like this.  Given how all the
> discussions about sorting and dependencies for suspend and resume went
> it might be an uphill struggle to do much more, especially while we are
> able to continue to deal with things fairly easily using the current
> infrastructure.
> 
> Doing dependencies could get pretty complicated, especially once you
> handle optional dependencies ("is this missing because it didn't probe
> yet or because it's just not there?") so it's not entirely clear to me
> that it's worth the hassle.

I think it might be doable.  I had a similar problem with Ethernet
MACs and PHYs where the PHY was on a completely separate bus from the
MAC with zero guarantees on probe order.  I had some code that made it
simple to use a bus notifier to defer MAC initialization until the
required phy turned up and was probed.  I eventually abandoned it
because accessing the PHY could be deferred until .ndo_open() time.
However, it would be easy to resurrect, and might be a reasonable
solution.  At the very least it is worth an investigation.

g.