On Thu, May 20, 2021 at 04:05:46PM +0100, Steven Price wrote:

On 20/05/2021 12:54, Catalin Marinas wrote:

quoted

On Mon, May 17, 2021 at 01:32:35PM +0100, Steven Price wrote:

quoted

diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
index c5d1f3c87dbd..8660f6a03f51 100644
--- a/arch/arm64/kvm/mmu.c
+++ b/arch/arm64/kvm/mmu.c

@@ -822,6 +822,31 @@ transparent_hugepage_adjust(struct kvm_memory_slot *memslot,
 	return PAGE_SIZE;
 }
 
+static int sanitise_mte_tags(struct kvm *kvm, unsigned long size,
+			     kvm_pfn_t pfn)
+{
+	if (kvm_has_mte(kvm)) {
+		/*
+		 * The page will be mapped in stage 2 as Normal Cacheable, so
+		 * the VM will be able to see the page's tags and therefore
+		 * they must be initialised first. If PG_mte_tagged is set,
+		 * tags have already been initialised.
+		 */
+		unsigned long i, nr_pages = size >> PAGE_SHIFT;
+		struct page *page = pfn_to_online_page(pfn);
+
+		if (!page)
+			return -EFAULT;

IIRC we ended up with pfn_to_online_page() to reject ZONE_DEVICE pages
that may be mapped into a guest and we have no idea whether they support
MTE. It may be worth adding a comment, otherwise, as Marc said, the page
wouldn't disappear.

I'll add a comment.

quoted

quoted

+
+		for (i = 0; i < nr_pages; i++, page++) {
+			if (!test_and_set_bit(PG_mte_tagged, &page->flags))
+				mte_clear_page_tags(page_address(page));

We started the page->flags thread and ended up fixing it for the host
set_pte_at() as per the first patch:

https://lore.kernel.org/r/c3293d47-a5f2-ea4a-6730-f5cae26d8a7e@arm.com (local)

Now, can we have a race between the stage 2 kvm_set_spte_gfn() and a
stage 1 set_pte_at()? Only the latter takes a lock. Or between two
kvm_set_spte_gfn() in different VMs? I think in the above thread we
concluded that there's only a problem if the page is shared between
multiple VMMs (MAP_SHARED). How realistic is this and what's the
workaround?

Either way, I think it's worth adding a comment here on the race on
page->flags as it looks strange that here it's just a test_and_set_bit()
while set_pte_at() uses a spinlock.

Very good point! I should have thought about that. I think splitting the
test_and_set_bit() in two (as with the cache flush) is sufficient. While
there technically still is a race which could lead to user space tags
being clobbered:

a) It's very odd for a VMM to be doing an mprotect() after the fact to
add PROT_MTE, or to be sharing the memory with another process which
sets PROT_MTE.

b) The window for the race is incredibly small and the VMM (generally)
needs to be robust against the guest changing tags anyway.

But I'll add a comment here as well:

	/*
	 * There is a potential race between sanitising the
	 * flags here and user space using mprotect() to add
	 * PROT_MTE to access the tags, however by splitting
	 * the test/set the only risk is user space tags
	 * being overwritten by the mte_clear_page_tags() call.
	 */

I think (well, I haven't re-checked), an mprotect() in the VMM ends up
calling set_pte_at_notify() which would call kvm_set_spte_gfn() and that
will map the page in the guest. So the problem only appears between
different VMMs sharing the same page. In principle they can be
MAP_PRIVATE but they'd be CoW so the race wouldn't matter. So it's left
with MAP_SHARED between multiple VMMs.

I think we should just state that this is unsafe and they can delete
each-others tags. If we are really worried, we can export that lock you
added in mte.c.

-- 
Catalin

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel