Re: [5.3] ucc_geth: Fix to avoid IS_ERR_VALUE abuses and dead code on 64bit systems.

[Arnd Bergmann <arnd@arndb.de>]

On Monday, August 8, 2016 3:49:22 PM CEST David Laight wrote:
> From: Arnd Bergmann
> > Sent: 08 August 2016 16:13
> > 
> > On Monday, August 8, 2016 2:49:11 PM CEST David Laight wrote:
> > >
> > > > If qe_muram_alloc will return any error, Then IS_ERR_VALUE will always
> > > > return 0. it'll not call ucc_fast_free for any failure. Inside 'if code'
> > > > will be a dead code on 64bit. Even qe_muram_addr will return wrong
> > > > virtual address. Which can cause an error.
> > > >
> > > >  kfree((void *)ugeth->tx_bd_ring_offset[i]);
> > >
> > > Erm, kfree() isn't the right function for things allocated by qe_muram_alloc().
> > >
> > > I still thing you need to stop this code using IS_ERR_VALUE() at all.
> > 
> > Those are two separate issues:
> > 
> > a) The ucc_geth driver mixing kmalloc() memory with muram, and assigning
> >    the result to "u32" and "void __iomem *" variables, both of which
> >    are wrong at least half of the time.
> > 
> > b) calling conventions of qe_muram_alloc() being defined in a way that
> >    requires the use of IS_ERR_VALUE(), because '0' is a valid address
> >    here.
> 
> Yep, it is all a big bag of worms...
> '0' being valid is going to make tidying up after failure 'problematic'.
> 
> > The first one can be solved by updating the network driver, ideally
> > by getting rid of the casts and using proper types and accessors,
> > while the second would require updating all users of that interface.
> 
> It might be worth (at least as a compilation option) of embedding the
> 'muram offset' in a structure (passed and returned by value).
> 
> The compiler can then check that the driver code is never be looking
> directly at the value.
>
> For 'b' zero can be made invalid by changing the places where the
> offset is added/subtracted.
> It could even be used to offset the saved physical and virtual
> addresses of the area - so not needing any extra code when the values
> are converted to physical/virtual addresses.

Agreed.

For this driver, we don't actually seem to use the value returned from
the allocation function, only the virtual __iomem address we get after
calling qe_muram_addr(), so it would be a big improvement to just
store the virtual address as a pointer, and wrap the calls
to qe_muram_alloc/qe_muram_addr/qe_muram_free with an appropriate
helper that doesn't even show the offset.

However, I'd also separate the normal kmalloc pointer from the
muram_alloc() pointer because only the latter is __iomem, and
we shouldn't really call MMIO accessor functions on RAM in
portable code.

	Arnd