Re: [PATCH] Add debugging boundary check to pfn_to_page

[Eric B Munson <emunson@mgebm.net>]

[-- Attachment #1: Type: text/plain, Size: 3380 bytes --]

On Wed, 08 Jun 2011, Dave Hansen wrote:

> On Wed, 2011-06-08 at 15:18 -0400, Eric B Munson wrote:
> > -#define __pfn_to_page(pfn)                             \
> > -({     unsigned long __pfn = (pfn);                    \
> > -       struct mem_section *__sec = __pfn_to_section(__pfn);    \
> > -       __section_mem_map_addr(__sec) + __pfn;          \
> > +#ifdef CONFIG_DEBUG_MEMORY_MODEL
> > +#define __pfn_to_page(pfn)                                             \
> > +({     unsigned long __pfn = (pfn);                                    \
> > +       struct mem_section *__sec = __pfn_to_section(__pfn);            \
> > +       struct page *__page = __section_mem_map_addr(__sec) + __pfn;    \
> > +       WARN_ON(__page->flags == 0);                                    \
> > +       __page;                                                         \
> 
> What was the scenario you're trying to catch here?  If you give a really
> crummy __pfn, you'll probably go off the end of one of the mem_section[]
> arrays, and get garbage back for __sec.  You might also get a NULL back
> from __section_mem_map_addr() if the section is possibly valid, but just
> not present on this particular system.
> 
> I _think_ the only kind of bug this will catch is if you have a valid
> section, with a valid section_mem_map[] but still manage to find
> yourself with an 'struct page' unclaimed by any zone and thus
> uninitialized.

This is the case I was going after.  I will rework for a V2 based on the
feedback here.

> 
> You could catch a lot more cases by being a bit more paranoid:
> 
> void check_pfn(unsigned long pfn)
> {
> 	int nid;
> 	
> 	// hacked in from pfn_to_nid:
> 	// Don't actually do this, add a new helper near pfn_to_nid()
> 	// Can this even fit in the physnode_map?
> 	if (pfn / PAGES_PER_ELEMENT > ARRAY_SIZE(physnode_map))
> 		WARN();
> 
> 	// Is there a valid nid there?
> 	nid = pfn_to_nid(pfn);
> 	if (nid == -1)
> 		WARN();
> 	
> 	// check against NODE_DATA(nid)->node_start_pfn;
> 	// check against NODE_DATA(nid)->node_spanned_pages;
> }
> >  })
> > +#else
> > +#define __pfn_to_page(pfn)                                             \
> > +({     unsigned long __pfn = (pfn);                                    \
> > +       struct mem_section *__sec = __pfn_to_section(__pfn);            \
> > +       __section_mem_map_addr(__sec) + __pfn;  \
> > +})
> > +#endif /* CONFIG_DEBUG_MEMORY_MODEL */ 
> 
> Instead of making a completely new __pfn_to_page() in the debugging
> case, I'd probably do something like this:
> 
> #ifdef CONFIG_DEBUG_MEMORY_MODEL
> #define check_foo(foo) {\
> 	some_check_here(foo);\
> 	WARN_ON(foo->flags);\
> }
> #else
> #define check_foo(foo) do{}while(0)
> #endif;
> 
> #define __pfn_to_page(pfn)                                             \
> ({     unsigned long __pfn = (pfn);                                    \
>        struct mem_section *__sec = __pfn_to_section(__pfn);            \
>        struct page *__page = __section_mem_map_addr(__sec) + __pfn;    \
>        check_foo(page)							\
>        __page;                                                         \
>  })
> 
> That'll make sure that the two copies of __pfn_to_page() don't
> accidentally diverge.  It also makes it a lot easier to read, I think.
> 
> -- Dave
> 

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 490 bytes --]