Messages in this thread Patch in this message | | | From | Nick Piggin <> | Subject | Re: [BUG] SLOB's krealloc() seems bust | Date | Wed, 8 Oct 2008 15:22:30 +1100 |
| |
On Wednesday 08 October 2008 10:08, Matt Mackall wrote: > 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:
Right.
> 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; > }
Yes, I came up with nearly the same patch before reading this
--- linux-2.6/mm/slob.c 2008-10-08 14:43:17.000000000 +1100 +++ suth/mm/slob.c 2008-10-08 15:11:06.000000000 +1100 @@ -514,9 +514,11 @@ size_t ksize(const void *block) return 0;
sp = (struct slob_page *)virt_to_page(block); - if (slob_page(sp)) - return (((slob_t *)block - 1)->units - 1) * SLOB_UNIT; - else + if (slob_page(sp)) { + int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN); + unsigned int *m = (unsigned int *)(block - align); + return *m + align; + } else return sp->page.private; }
However, mine is lifted directly from kfree, wheras you do something a bit different. Hmm, ksize arguably could be used to find the underlying allocated slab size in order to use a little bit more than we'd asked for. So probably we should really just `return *m` (don't round up or add any padding).
> 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?
I didn't have ksize in my slob user test harness, but added a couple of tests in there, and indeed ksize was returning complete garbage both before and after the latest patch to slob. I'd say it was simply luck.
| |