Re: [PATCH 3/3] x86/kasan: Print original address on #GP

[Jann Horn <jannh@google.com>]

On Wed, Nov 13, 2019 at 11:11 AM Dmitry Vyukov <dvyukov@google.com> wrote:
> On Tue, Nov 12, 2019 at 10:10 PM 'Jann Horn' via kasan-dev
> <kasan-dev@googlegroups.com> wrote:
> > Make #GP exceptions caused by out-of-bounds KASAN shadow accesses easier
> > to understand by computing the address of the original access and
> > printing that. More details are in the comments in the patch.
[...]
> +Andrey, do you see any issues for TAGS mode? Or, Jann, did you test
> it by any chance?

No, I didn't - I don't have anything set up for upstream arm64 testing here.

> > +void kasan_general_protection_hook(unsigned long addr)
> >  {
> > -       if (val == DIE_GPF) {
> > -               pr_emerg("CONFIG_KASAN_INLINE enabled\n");
> > -               pr_emerg("GPF could be caused by NULL-ptr deref or user memory access\n");
> > -       }
> > -       return NOTIFY_OK;
> > -}
> > +       unsigned long orig_addr;
> > +       const char *addr_type;
> > +
> > +       if (addr < KASAN_SHADOW_OFFSET)
> > +               return;
>
> Thinking how much sense it makes to compare addr with KASAN_SHADOW_END...
> If the addr is > KASAN_SHADOW_END, we know it's not a KASAN access,
> but do we ever get GP on canonical addresses?

#GP can occur for various reasons, but on x86-64, if it occurs because
of an invalid address, as far as I know it's always non-canonical. The
#GP handler I wrote will check the address and only call the KASAN
hook if the address is noncanonical (because otherwise the #GP
occurred for some other reason).

> > -static struct notifier_block kasan_die_notifier = {
> > -       .notifier_call = kasan_die_handler,
> > -};
> > +       orig_addr = (addr - KASAN_SHADOW_OFFSET) << KASAN_SHADOW_SCALE_SHIFT;
> > +       /*
> > +        * For faults near the shadow address for NULL, we can be fairly certain
> > +        * that this is a KASAN shadow memory access.
> > +        * For faults that correspond to shadow for low canonical addresses, we
> > +        * can still be pretty sure - that shadow region is a fairly narrow
> > +        * chunk of the non-canonical address space.
> > +        * But faults that look like shadow for non-canonical addresses are a
> > +        * really large chunk of the address space. In that case, we still
> > +        * print the decoded address, but make it clear that this is not
> > +        * necessarily what's actually going on.
> > +        */
> > +       if (orig_addr < PAGE_SIZE)
> > +               addr_type = "dereferencing kernel NULL pointer";
> > +       else if (orig_addr < TASK_SIZE_MAX)
> > +               addr_type = "probably dereferencing invalid pointer";
>
> This is access to user memory, right? In outline mode we call it
> "user-memory-access". We could say about "user" part here as well.

Okay, I'll copy that naming.

> > +       else
> > +               addr_type = "maybe dereferencing invalid pointer";
> > +       pr_alert("%s in range [0x%016lx-0x%016lx]\n", addr_type,
> > +                orig_addr, orig_addr + (1 << KASAN_SHADOW_SCALE_SHIFT) - 1);
>
> "(1 << KASAN_SHADOW_SCALE_SHIFT) - 1)" part may be replaced with
> KASAN_SHADOW_MASK.
> Overall it can make sense to move this mm/kasan/report.c b/c we are
> open-coding a number of things here (e.g. reverse address mapping). If
> another arch will do the same, it will need all of this code too (?).

Alright, I'll try to move it over.