Re: [PATCH] mm: __nr_to_section - make it safe against overflow

[Nick Piggin]

On Mon, Jan 05, 2009 at 06:28:48PM +0300, Cyrill Gorcunov wrote:
> [Christoph Lameter - Mon, Jan 05, 2009 at 09:10:57AM -0600]
> | On Mon, 5 Jan 2009, Cyrill Gorcunov wrote:
> | 
> | >  	/*
> | >  	 * This is, logically, a pointer to an array of struct
> | > @@ -980,9 +986,12 @@ extern struct mem_section mem_section[NR
> | >
> | >  static inline struct mem_section *__nr_to_section(unsigned long nr)
> | >  {
> | > -	if (!mem_section[SECTION_NR_TO_ROOT(nr)])
> | > +	unsigned long idx = SECTION_NR_TO_ROOT(nr);
> | > +	WARN_ON_ONCE(idx >= NR_SECTION_ROOTS);
> | > +
> | > +	if (idx >=NR_SECTION_ROOTS || !mem_section[idx])
> | >  		return NULL;
> | > -	return &mem_section[SECTION_NR_TO_ROOT(nr)][nr & SECTION_ROOT_MASK];
> | > +	return &mem_section[idx][nr & SECTION_ROOT_MASK];
> | >  }
> | >  extern int __section_nr(struct mem_section* ms);
> | >  extern unsigned long usemap_size(void);
> | 
> | Not that you are adding code to numerous hot code path. Plus this is a
> | frequently used inline. Code size is going to increase if you do this.
> 
> yes, I know, that is why I've changed WARN_ON_ONCE to plain WARN_ON.

Still costs. Putting it under a config option would be nice.


> | I would think that the code does not have the tests because of performance
> | and code size concerns. Can we just say that a sane nr must be passed to
> | __nr_section?
> | 
> 
> If you mean did I test this patch for speed regresson then to be fair --
> no, I didn't. BUT we have a number of macros wich are self protective
> like present_section which is used havily too. On the other hand --
> bad argument passed to __nr_to_section will be (and it is now) really
> harmfull -- since it would allow to reference a memory outside the
> valid bounds. The second -- SECTION_ROOT_MASK wich is fragile, any
> attempt to modify mem_section structure will silently lead to insane
> referencing, that is why it deserve a comment on top of structure.
> 
> Don't know Christoph, if it really that important to not spend a few
> cycles here in a sake of safety -- we could easily drop this patch.

The problem with testing every little slowdown for a speed regression
is that they are just going to be in the noise. But we *know* it will
go slower. The problem is that they add up. We just have to be sensible
about it.

Has there ever been a problem here before? Has it been a problem during
development? (in which case putting it in a .config option might make
sense).
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/