On 08/08/2018 01:54 PM, Vlastimil Babka wrote:
> Hmm I have looked at the splats in all the bugs you referenced and the
> Code part always has the de-obfuscation XORs. Then in comment 36 of
> [1] jian-hong says the problem disappeared, and in comment 40 posts
> a config that has CONFIG_SLAB_FREELIST_HARDENED disabled. Earlier
> posting of his config has it enabled and confirms the disassembly.
> Very suspicious, huh.
So I'm looking at 2482ddec670f ("mm: add SLUB free list pointer
obfuscation") from Kees, and one suspicious thing is:
before, prefetch_freepointer() was just:
prefetch(object + s->offset);
after, it is
if (object)
prefetch(freelist_dereference(s, object + s->offset));
Where freelist_dereference() is either a simple dereference of address,
when FREELIST_HARDENED is disabled, or adds those XORs when enabled.
However, this had changed the prefetch intention! Previously it just
prefetched the address (object + s->offset), now it *dereferences it*,
optionally changes the value read with those XORs, and then prefetches
the result.
This unintentionally adds a non-prefetching read from (object +
s->offset), which may fault, and wasn't there before. It's safe from
NULL pointers, but not from bogus pointers, and faults that
get_freepointer_safe() prevents. Note that alone doesn't explain why
disabling SLAB_FREELIST_HARDENED would help, as the dereference is there
unconditionally. But IMHO it's a bug in the commit and minimally it
likely has some performance impact.
