Re: [PATCH] usercopy: Remove HARDENED_USERCOPY_PAGESPAN

[Kees Cook]

On Sat, May 11, 2019 at 09:11:42PM -0700, Matthew Wilcox wrote:
> On Sat, May 11, 2019 at 05:03:08PM -0700, Kees Cook wrote:
> > On Fri, May 10, 2019 at 08:41:43PM -0400, Laura Abbott wrote:
> > > On 5/10/19 3:43 PM, Kees Cook wrote:
> > > > This feature continues to cause more problems than it solves[1]. Its
> > > > intention was to check the bounds of page-allocator allocations by using
> > > > __GFP_COMP, for which we would need to find all missing __GFP_COMP
> > > > markings. This work has been on hold and there is an argument[2]
> > > > that such markings are not even the correct signal for checking for
> > > > same-allocation pages. Instead of depending on BROKEN, this just removes
> > > > it entirely. It can be trivially reverted if/when a better solution for
> > > > tracking page allocator sizes is found.
> > > > 
> > > > [1] https://www.mail-archive.com/linux-crypto@xxxxxxxxxxxxxxx/msg37479.html
> > > > [2] https://lkml.kernel.org/r/20190415022412.GA29714@xxxxxxxxxxxxxxxxxxxxxx
> > > 
> > > I agree the page spanning is broken but is it worth keeping the
> > > checks against __rodata __bss etc.?
> > 
> > They're all just white-listing later checks (except RODATA which is
> > doing a cheap RO test which is redundant on any architecture with actual
> > rodata...) so they don't have any value in staying without the rest of
> > the page allocator logic.
> > 
> > > > -	/* Is the object wholly within one base page? */
> > > > -	if (likely(((unsigned long)ptr & (unsigned long)PAGE_MASK) ==
> > > > -		   ((unsigned long)end & (unsigned long)PAGE_MASK)))
> > > > -		return;
> > > > -
> > > > -	/* Allow if fully inside the same compound (__GFP_COMP) page. */
> > > > -	endpage = virt_to_head_page(end);
> > > > -	if (likely(endpage == page))
> > > > -		return;
> > 
> > We _could_ keep the mixed CMA/reserved/neither check if we really wanted
> > to, but that's such a corner case of a corner case, I'm not sure it's
> > worth doing the virt_to_head_page() across the whole span to figure
> > it out.
> 
> I'd delete that first check, because it's a subset of the second check,

It seemed easier to short-circuit with a math test before doing the slightly more expensive virt_to_head_page(end) call. Do you think that's sensible?

> 
> 	/* Is the object wholly within a single (base or compound) page? */
> 	endpage = virt_to_head_page(end);
> 	if (likely(endpage == page))
> 		return;
> 
> 	/*
> 	 * If the start and end are more than MAX_ORDER apart, they must
> 	 * be from separate allocations
> 	 */
> 	if (n >= (PAGE_SIZE << MAX_ORDER))
> 		usercopy_abort("spans too many pages", NULL, to_user, 0, n);
> 
> 	/*
> 	 * If neither page is compound, we can't tell if the object is
> 	 * within a single allocation or not
> 	 */
> 	if (!PageCompound(page) && !PageCompound(endpage))
> 		return;
> 
> > I really wish we had size of allocation reliably held somewhere. We'll
> > need it for doing memory tagging of the page allocator too...
> 
> I think we'll need to store those allocations in a separate data structure
> on the side.  As far as the rest of the kernel is concerned, those struct
> pages belong to them once the page allocator has given them.

Okay, let me work up a page-type refactoring while allocation size can
stay back-burnered.

-- 
Kees Cook