Re: [PATCH 1/2] mfd: mfd-core: Allocate reference counting memory directly to the platform device

[Lee Jones]

On Fri, 18 Oct 2019, Daniel Thompson wrote:

> On Fri, Oct 18, 2019 at 01:26:46PM +0100, Lee Jones wrote:
> > MFD provides reference counting (for the 2 consumers who actually use it!)
> > via mfd_cell's 'usage_count' member.  However, since MFD cells become
> > read-only (const), MFD needs to allocate writable memory and assign it to
> > 'usage_count' before first registration.  It currently does this by
> > allocating enough memory for all requested child devices (yes, even disabled
> > ones - but we'll get to that) and assigning the base pointer plus sub-device
> > index to each device in the cell.
> > 
> > The difficulty comes when trying to free that memory.  During the removal of
> > the parent device, MFD unregisters each child device, keeping a tally on the
> > lowest memory location pointed to by a child device's 'usage_count'.  Once
> > all of the children are unregistered, the lowest memory location must be the
> > base address of the previously allocated array, right?
> > 
> > Well yes, until we try to honour the disabling of devices via Device Tree
> > for instance.  If the first child device in the provided batch is disabled,
> > simply skipping registration (and consequentially deregistration) will mean
> > that the first device's 'usage_count' pointer will not be accounted for when
> > attempting to find the base.  In which case, MFD will assume the first non-
> > disabled 'usage_count' pointer is the base and subsequently attempt to
> > erroneously free it.
> > 
> > We can avoid all of this hoop jumping by simply allocating memory to each
> > single child device before it is considered read-only.  We can then free
> > it on a per-device basis during deregistration.
> > 
> > Signed-off-by: Lee Jones <lee.jones@xxxxxxxxxx>
> > ---
> >  drivers/mfd/mfd-core.c | 42 ++++++++++++++++++------------------------
> >  1 file changed, 18 insertions(+), 24 deletions(-)
> > 
> > diff --git a/drivers/mfd/mfd-core.c b/drivers/mfd/mfd-core.c
> > index 23276a80e3b4..eafdadd58e8b 100644
> > --- a/drivers/mfd/mfd-core.c
> > +++ b/drivers/mfd/mfd-core.c
> > @@ -404,7 +398,7 @@ int mfd_clone_cell(const char *cell, const char **clones, size_t n_clones)
> >  		cell_entry.name = clones[i];
> >  		/* don't give up if a single call fails; just report error */
> >  		if (mfd_add_device(pdev->dev.parent, -1, &cell_entry,
> > -				   cell_entry.usage_count, NULL, 0, NULL))
> > +				   NULL, 0, NULL))
> 
> I think this change is broken.
> 
> Cloned cells are supposed to share the same reference counter as their
> template and this change results in each clone having its own counter.
> That means the "the 2 consumers who actually use it" will both end up
> calling cs5535_mfd_res_enable() (and whichever loses the race will fail
> to probe).
> 
> To be honest it might be easier to move the request_region() into
> cs5535_mfd_probe() and rip out the entire reference counting mechanism
> since at that point it would be unused (the other callers of
> mfd_cell_enable() look safe w/o a counter).

Thanks for the review.  Great point(s).

I will fix this and submit a v2 shortly.

> >  			dev_err(dev, "failed to create platform device '%s'\n",
> >  					clones[i]);
> >  	}

-- 
Lee Jones [æçæ]
Linaro Services Technical Lead
Linaro.org â Open source software for ARM SoCs
Follow Linaro: Facebook | Twitter | Blog