Re: [PATCH v2 0/5] slab cleanups

[Vlastimil Babka <vbabka@suse.cz>]

On 3/4/22 14:11, Marco Elver wrote:
> On Fri, 4 Mar 2022 at 13:02, Hyeonggon Yoo <42.hyeyoo@gmail.com> wrote:
>>
>> On Fri, Mar 04, 2022 at 12:50:21PM +0100, Marco Elver wrote:
>> > On Fri, 4 Mar 2022 at 07:34, Hyeonggon Yoo <42.hyeyoo@gmail.com> wrote:
>> > >
>> > > Changes from v1:
>> > >         Now SLAB passes requests larger than order-1 page
>> > >         to page allocator.
>> > >
>> > >         Adjusted comments from Matthew, Vlastimil, Rientjes.
>> > >         Thank you for feedback!
>> > >
>> > >         BTW, I have no idea what __ksize() should return when an object that
>> > >         is not allocated from slab is passed. both 0 and folio_size()
>> > >         seems wrong to me.
>> >
>> > Didn't we say 0 would be the safer of the two options?
>> > https://lkml.kernel.org/r/0e02416f-ef43-dc8a-9e8e-50ff63dd3c61@suse.cz
>> >
>>
>> Oh sorry, I didn't understand why 0 was safer when I was reading it.
>>
>> Reading again, 0 is safer because kasan does not unpoison for
>> wrongly passed object, right?
> 
> Not quite. KASAN can tell if something is wrong, i.e. invalid object.
> Similarly, if you are able to tell if the passed pointer is not a
> valid object some other way, you can do something better - namely,
> return 0.

Hmm, but how paranoid do we have to be? Patch 1 converts SLAB to use
kmalloc_large(). So it's now legitimate to have objects allocated by SLAB's
kmalloc() that don't have a slab folio flag set, and their size is
folio_size(). It would be more common than getting a bogus pointer, so
should we return 0 just because a bogus pointer is possible? If we do that,
then KASAN will fail to unpoison legitimate kmalloc_large() objects, no?
What I suggested earlier is we could make the checks more precise - if
folio_size() is smaller or equal order-1 page, then it's bogus because we
only do kmalloc_large() for >order-1. If the object pointer is not to the
beginning of the folio, then it's bogus, because kmalloc_large() returns the
beginning of the folio. Then in these case we return 0, but otherwise we
should return folio_size()?

> The intuition here is that the caller has a pointer to an
> invalid object, and wants to use ksize() to determine its size, and
> most likely access all those bytes. Arguably, at that point the kernel
> is already in a degrading state. But we can try to not let things get
> worse by having ksize() return 0, in the hopes that it will stop
> corrupting more memory. It won't work in all cases, but should avoid
> things like "s = ksize(obj); touch_all_bytes(obj, s)" where the size
> bounds the memory accessed corrupting random memory.
> 
> The other reason is that a caller could actually check the size, and
> if 0, do something else. Few callers will do so, because nobody
> expects that their code has a bug. :-)