KASAN vs realloc

["Maciej Żenczykowski" <maze@google.com>]

We've got internal reports (b/467571011 - from CC'ed Samsung
developer) that kasan realloc is broken for sizes that are not a
multiple of the granule.  This appears to be triggered during Android
bootup by some ebpf program loading operations (a struct is 88 bytes
in size, which is a multiple of 8, but not 16, which is the granule
size).

(this is on 6.18 with
https://lore.kernel.org/all/38dece0a4074c43e48150d1e242f8242c73bf1a5.1764874575.git.m.wieczorretman@pm.me/
already included)

joonki.min@samsung-slsi.corp-partner.google.com summarized it as
"When newly requested size is not bigger than allocated size and old
size was not 16 byte aligned, it failed to unpoison extended area."

and *very* rough comment:

Right. "size - old_size" is not guaranteed 16-byte alignment in this case.

I think we may unpoison 16-byte alignment size, but it allowed more
than requested :(

I'm not sure that's right approach.

if (size <= alloced_size) {
- kasan_unpoison_vmalloc(p + old_size, size - old_size,
+               kasan_unpoison_vmalloc(p + old_size, round_up(size -
old_size, KASAN_GRANULE_SIZE),
      KASAN_VMALLOC_PROT_NORMAL |
      KASAN_VMALLOC_VM_ALLOC |
      KASAN_VMALLOC_KEEP_TAG);
/*
* No need to zero memory here, as unused memory will have
* already been zeroed at initial allocation time or during
* realloc shrink time.
*/
- vm->requested_size = size;
+               vm->requested_size = round_up(size, KASAN_GRANULE_SIZE);

my personal guess is that

But just above the code you quoted in mm/vmalloc.c I see:
        if (size <= old_size) {
...
                kasan_poison_vmalloc(p + size, old_size - size);

is also likely wrong?? Considering:

mm/kasan/shadow.c

void __kasan_poison_vmalloc(const void *start, unsigned long size)
{
        if (!is_vmalloc_or_module_addr(start))
                return;

        size = round_up(size, KASAN_GRANULE_SIZE);
        kasan_poison(start, size, KASAN_VMALLOC_INVALID, false);
}

This doesn't look right - if start isn't a multiple of the granule.

--
Maciej Żenczykowski, Kernel Networking Developer @ Google