Re: [PATCH] sh: mm: Fix unprotected access to struct device

[Rich Felker <dalias@libc.org>]

On Fri, Apr 20, 2018 at 11:59:13AM +0200, Geert Uytterhoeven wrote:
> Hi Christoph,
> 
> On Fri, Apr 20, 2018 at 10:31 AM, Christoph Hellwig <hch@infradead.org> wrote:
> > On Wed, Apr 18, 2018 at 03:13:14PM +0200, jacopo mondi wrote:
> >> As long as it goes for arch/sh, the only user of dma_alloc_coherent()
> >> is platform_resource_setup_memory(), and it has been fixed by this
> >> patch.
> >
> > Great!
> >
> >> Unfortunately, as Thomas pointed out, there are drivers which calls
> >> into this with the wrong 'struct device' as the sh_eth one he had fixed.
> >
> > Yes, we'll need fixes there.  Other DMA ops implementations also look
> > at struct device, so they generally are buggy.
> >
> >> I would then say that as long as it goes for the NULL case, we should be
> >> fine now.
> >
> > Then I'd say skil that part, please.
> 
> The major reason for keeping the NULL WARN_ON() checks is to make it
> obvious to the developer what is wrong, and fall back to the old behavior.
> 
> Without the checks, the kernel will just crash during early startup,
> without a clue in the (missing) kernel output, usually leading to a
> frustrating bisection experience (if the developer is sufficiently motivated,
> at all).
> 
> Hence my vote for keeping the checks.

Sounds good to me.

Rich