Re: [PATCH -v2] memdup_user(): introduce

[Andrew Morton <akpm@linux-foundation.org>]

On Sat, 7 Mar 2009 08:48:05 -0800 Arjan van de Ven <arjan@infradead.org> wrote:

> On Fri, 6 Mar 2009 15:03:35 -0800
> Andrew Morton <akpm@linux-foundation.org> wrote:
> 
> > >  
> > >  /**
> > > + * memdup_user - duplicate memory region from user space
> > > + *
> > > + * @src: source address in user space
> > > + * @len: number of bytes to copy
> > > + * @gfp: GFP mask to use
> > > + *
> > > + * Returns an ERR_PTR() on failure.
> > > + */
> > > +void *memdup_user(const void __user *src, size_t len, gfp_t gfp)
> > > +{
> > > +	void *p;
> > > +
> > > +	p = kmalloc_track_caller(len, gfp);
> > > +	if (!p)
> > > +		return ERR_PTR(-ENOMEM);
> > > +
> > > +	if (copy_from_user(p, src, len)) {
> > > +		kfree(p);
> > > +		return ERR_PTR(-EFAULT);
> > > +	}
> > > +
> > > +	return p;
> > > +}
> > > +EXPORT_SYMBOL(memdup_user);
> 
> Hi,
> 
> I like the general idea of this a lot; it will make things much less
> error prone (and we can add some sanity checks on "len" to catch the
> standard security holes around copy_from_user usage). I'd even also
> want a memdup_array() like thing in the style of calloc().
> 
> However, I have two questions/suggestions for improvement:
> 
> I would like to question the use of the gfp argument here;
> copy_from_user sleeps, so you can't use GFP_ATOMIC anyway.
> You can't use GFP_NOFS etc, because the pagefault path will happily do
> things that are equivalent, if not identical, to GFP_KERNEL.
> 
> So the only value you can pass in correctly, as far as I can see, is
> GFP_KERNEL. Am I wrong?

True.

> A second thing.. I'd like to have this function return NULL on failure;
> error checking a pointer for NULL is so much easier than testing for
> anything else; the only distinction is -ENOMEM versus -EFAULT, and I'm
> not sure that that is worth the complexity on all callers.

This errno will be returned to userspace.  If the caller guesses wrong,
that's a kernel bug, of the regression variety.