Crash on armv7-a using KASAN

[Mark Rutland]

On Tue, Oct 15, 2024 at 07:28:06PM +0200, Ard Biesheuvel wrote:
> On Tue, 15 Oct 2024 at 18:27, Mark Rutland <mark.rutland at arm.com> wrote:
> >
> > On Tue, Oct 15, 2024 at 06:07:00PM +0200, Ard Biesheuvel wrote:
> > > On Tue, 15 Oct 2024 at 17:26, Mark Rutland <mark.rutland at arm.com> wrote:
> > > > Looking some more, I don't see how VMAP_STACK guarantees that the
> > > > old/active stack is mapped in the new mm when switching from the old mm
> > > > to the new mm (which happens before __switch_to()).
> > > >
> > > > Either I'm missing something, or we have a latent bug. Maybe we have
> > > > some explicit copying/prefaulting elsewhere I'm missing?
> > >
> > > We bump the vmalloc_seq counter for that. Given that the top-level
> > > page table can only gain entries covering the kernel space, this
> > > should be sufficient for the old task's stack to be mapped in the new
> > > task's page tables.
> >
> > Ah, yep -- I had missed that. Thanks for the pointer!
> >
> > From a superficial look, it sounds like it should be possible to extend
> > that to also handle the KASAN shadow of the vmalloc area (which
> > __check_vmalloc_seq() currently doesn't copy), but I'm not sure of
> > exactly when we initialise the shadow for a vmalloc allocation relative
> > to updating vmalloc_seq.
> >
> 
> Indeed. It appears both vmalloc_seq() and arch_sync_kernel_mappings()
> need to take the vmalloc shadow into account specifically. And we may
> also need the dummy read from the stack's shadow in __switch_to - I am
> pretty sure I added that for a reason.

I believe that's necessary for the lazy TLB switch, at least for SMP:

	// CPU 0			// CPU 1

	<< switches to task X's mm >>

					<< creates kthread task Y >>
					<< maps task Y's new stack >>
					<< maps task Y's new shadow >>

					// Y switched out
					context_switch(..., Y, ..., ...);

	// Switch from X to Y
	context_switch(..., X, Y, ...) {
		// prev = X
		// next = Y

		if (!next->mm) { 
			// Y has no mm
			// No switch_mm() here
			// ... so no check_vmalloc_seq()
		} else {
			// not taken
		}

		...

		// X's mm still lacks Y's stack + shadow here

		switch_to(prev, next, prev);
	}

... so probably worth a comment that we're faulting in the new
stack+shadow for for lazy tlb when switching to a task with no mm?

In the lazy tlb case the current/old mappings don't disappear from the
active mm, and so we don't need to go add those to the new mm, which is what
we need check_vmalloc_seq() for.

Mark.