[PATCH v6 1/5] kasan: support backing vmalloc space with real shadow memory

[Mark Rutland]

On Tue, Sep 03, 2019 at 12:32:49AM +1000, Daniel Axtens wrote:
> Hi Mark,
> 
> >> +static int kasan_depopulate_vmalloc_pte(pte_t *ptep, unsigned long addr,
> >> +					void *unused)
> >> +{
> >> +	unsigned long page;
> >> +
> >> +	page = (unsigned long)__va(pte_pfn(*ptep) << PAGE_SHIFT);
> >> +
> >> +	spin_lock(&init_mm.page_table_lock);
> >> +
> >> +	if (likely(!pte_none(*ptep))) {
> >> +		pte_clear(&init_mm, addr, ptep);
> >> +		free_page(page);
> >> +	}
> >> +	spin_unlock(&init_mm.page_table_lock);
> >> +
> >> +	return 0;
> >> +}
> >
> > There needs to be TLB maintenance after unmapping the page, but I don't
> > see that happening below.
> >
> > We need that to ensure that errant accesses don't hit the page we're
> > freeing and that new mappings at the same VA don't cause a TLB conflict
> > or TLB amalgamation issue.
> 
> Darn it, I knew there was something I forgot to do! I thought of that
> over the weekend, didn't write it down, and then forgot it when I went
> to respin the patches. You're totally right.
> 
> >
> >> +/*
> >> + * Release the backing for the vmalloc region [start, end), which
> >> + * lies within the free region [free_region_start, free_region_end).
> >> + *
> >> + * This can be run lazily, long after the region was freed. It runs
> >> + * under vmap_area_lock, so it's not safe to interact with the vmalloc/vmap
> >> + * infrastructure.
> >> + */
> >
> > IIUC we aim to only free non-shared shadow by aligning the start
> > upwards, and aligning the end downwards. I think it would be worth
> > mentioning that explicitly in the comment since otherwise it's not
> > obvious how we handle races between alloc/free.
> >
> 
> Oh, I will need to think through that more carefully.
> 
> I think the vmap_area_lock protects us against alloc/free races.

AFAICT, on the alloc side we only hold the vmap_area_lock while
allocating the area in __get_vm_area_node(), but we don't holding the
vmap_area_lock while we populate the page tables for the shadow in
kasan_populate_vmalloc().

So I believe that kasan_populate_vmalloc() can race with
kasan_release_vmalloc().

> I think alignment operates at least somewhat as you've described, and
> while it is important for correctness, I'm not sure I'd say it
> prevented races? I will double check my understanding of
> vmap_area_lock, and I agree the comment needs to be much clearer.

I had assumed that you were trying to only free pages which were
definitely not shared (for which there couldn't possibly be a race to
allocate), by looking at the sibling areas to see if they potentially
overlapped.

Was that not the case?

Thanks,
Mark.