Re: [BUG] SLOB's krealloc() seems bust

[Matt Mackall <mpm@selenic.com>]

On Tue, 2008-10-07 at 20:31 +0300, Pekka Enberg wrote:
> Hi Matt,
> 
> On Tue, Oct 7, 2008 at 8:13 PM, Matt Mackall <mpm@selenic.com> wrote:
> >> > @@ -515,7 +515,7 @@
> >> >
> >> >        sp = (struct slob_page *)virt_to_page(block);
> >> >        if (slob_page(sp))
> >> > -               return ((slob_t *)block - 1)->units + SLOB_UNIT;
> >> > +               return (((slob_t *)block - 1)->units - 1) * SLOB_UNIT;
> >>
> >> Hmm. I don't understand why we do the "minus one" thing here. Aren't
> >> we underestimating the size now?
> >
> > The first -1 takes us to the object header in front of the object
> > pointer. The second -1 subtracts out the size of the header.
> >
> > But it's entirely possible I'm off by one, so I'll double-check. Nick?
> 
> Yeah, I was referring to the second subtraction. Looking at
> slob_page_alloc(), for example, we compare the return value of
> slob_units() to SLOB_UNITS(size), so I don't think we count the header
> in ->units. I mean, we ought to be seeing the subtraction elsewhere in
> the code as well, no?

Ok, I've looked a bit closer at it and I think we need a different fix.

The underlying allocator, slob_alloc, takes a size in bytes and returns
an object of that size, with the first word containing the number of
slob_t units.

kmalloc calls slob_alloc after adding on some space for header and
architecture padding. This space is not necessarily 1 slob unit:

        unsigned int *m;
        int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
...
                m = slob_alloc(size + align, gfp, align, node);
                *m = size;
 	        return (void *)m + align;

Note that we overwrite the header with our own size -in bytes-. 
kfree does the reverse:

                int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
		unsigned int *m = (unsigned int *)(block - align);
                slob_free(m, *m + align);

That second line is locating the kmalloc header. All looks good.

The MINALIGN business was introduced by Nick with:

 slob: improved alignment handling

but seems to have missed ksize, which should now be doing the following
to match:

diff -r 5e32b09a1b2b mm/slob.c
--- a/mm/slob.c	Fri Oct 03 14:04:43 2008 -0500
+++ b/mm/slob.c	Tue Oct 07 18:05:15 2008 -0500
@@ -514,9 +514,11 @@
 		return 0;
 
 	sp = (struct slob_page *)virt_to_page(block);
-	if (slob_page(sp))
-		return ((slob_t *)block - 1)->units + SLOB_UNIT;
-	else
+	if (slob_page(sp)) {
+		int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
+		unsigned int *m = (unsigned int *)(block - align);
+		return SLOB_UNITS(*m); /* round up */
+	} else
 		return sp->page.private;
 }
 

That leaves the question of why this morning's patch worked at all,
given that it was based on how SLOB worked before Nick's patch. But I
haven't finished working through that. Peter, can I get you to test the
above?

-- 
Mathematics is the supreme nostalgia of our time.