Re: [PATCH v10 1/6] arm64: mte: Sync tags for pages where PTE is untagged

From: Steven Price
Date: Wed Mar 31 2021 - 06:10:46 EST


On 30/03/2021 11:13, Catalin Marinas wrote:
On Mon, Mar 29, 2021 at 04:55:29PM +0100, Steven Price wrote:
On 26/03/2021 18:56, Catalin Marinas wrote:
On Fri, Mar 12, 2021 at 03:18:57PM +0000, Steven Price wrote:
A KVM guest could store tags in a page even if the VMM hasn't mapped
the page with PROT_MTE. So when restoring pages from swap we will
need to check to see if there are any saved tags even if !pte_tagged().

However don't check pages which are !pte_valid_user() as these will
not have been swapped out.

Signed-off-by: Steven Price <steven.price@xxxxxxx>
---
arch/arm64/include/asm/pgtable.h | 2 +-
arch/arm64/kernel/mte.c | 16 ++++++++++++----
2 files changed, 13 insertions(+), 5 deletions(-)

diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index e17b96d0e4b5..84166625c989 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -312,7 +312,7 @@ static inline void set_pte_at(struct mm_struct *mm, unsigned long addr,
__sync_icache_dcache(pte);
if (system_supports_mte() &&
- pte_present(pte) && pte_tagged(pte) && !pte_special(pte))
+ pte_present(pte) && pte_valid_user(pte) && !pte_special(pte))
mte_sync_tags(ptep, pte);

With the EPAN patches queued in for-next/epan, pte_valid_user()
disappeared as its semantics weren't very clear.

Thanks for pointing that out.

So this relies on the set_pte_at() being done on the VMM address space.
I wonder, if the VMM did an mprotect(PROT_NONE), can the VM still access
it via stage 2? If yes, the pte_valid_user() test wouldn't work. We need
something like pte_present() && addr <= user_addr_max().

AFAIUI the stage 2 matches the VMM's address space (for the subset that has
memslots). So mprotect(PROT_NONE) would cause the stage 2 mapping to be
invalidated and a subsequent fault would exit to the VMM to sort out. This
sort of thing is done for the lazy migration use case (i.e. pages are
fetched as the VM tries to access them).

There's also the protected KVM case which IIUC wouldn't provide any
mapping of the guest memory to the host (or maybe the host still thinks
it's there but cannot access it without a Stage 2 fault). At least in
this case it wouldn't swap pages out and it would be the responsibility
of the EL2 code to clear the tags when giving pages to the guest
(user_mem_abort() must not touch the page).

So basically we either have a valid, accessible mapping in the VMM and
we can handle the tags via set_pte_at() or we leave it to whatever is
running at EL2 in the pKVM case.

For the pKVM case it's up to the EL2 code to hand over suitably scrubbed pages to the guest, and the host doesn't have access to the pages so we (currently) don't have to worry about swap. If swap get implemented it will presumably be up to the EL2 code to package up both the normal data and the MTE tags into an encrypted bundle for the host to stash somewhere.

I don't remember whether we had a clear conclusion in the past: have we
ruled out requiring the VMM to map the guest memory with PROT_MTE
entirely? IIRC a potential problem was the VMM using MTE itself and
having to disable it when accessing the guest memory.

Yes, there are some ugly corner cases if we require the VMM to map with PROT_MTE. Hence patch 5 - an ioctl to allow the VMM to access the tags without having to maintain a PROT_MTE mapping.

Another potential issue (I haven't got my head around it yet) is a race
in mte_sync_tags() as we now defer the PG_mte_tagged bit setting until
after the tags had been restored. Can we have the same page mapped by
two ptes, each attempting to restore it from swap and one gets it first
and starts modifying it? Given that we set the actual pte after setting
PG_mte_tagged, it's probably alright but I think we miss some barriers.

I'm not sure if I've got my head round this one yet either, but you could be right there's a race. This exists without these patches:

CPU 1 | CPU 2
-------------------------+-----------------
set_pte_at() |
--> mte_sync_tags() |
--> test_and_set_bit() |
--> mte_sync_page_tags() | set_pte_at()
[stalls/sleeps] | --> mte_sync_tags()
| --> test_and_set_bit()
| [already set by CPU 1]
| set_pte()
| [sees stale tags]
[eventually wakes up |
and sets tags] |

What I'm struggling to get my head around is whether there's always a sufficient lock held during the call to set_pte_at() to avoid the above. I suspect not because the two calls could be in completely separate processes.

We potentially could stick a lock_page()/unlock_page() sequence in mte_sync_tags(). I just ran a basic test and didn't hit problems with that. Any thoughts?

Also, if a page is not a swap one, we currently clear the tags if mapped
as pte_tagged() (prior to this patch). We'd need something similar when
mapping it in the guest so that we don't leak tags but to avoid any page
ending up with PG_mte_tagged, I think you moved the tag clearing to
user_mem_abort() in the KVM code. I presume set_pte_at() in the VMM
would be called first and then set in Stage 2.

Yes - KVM will perform the equivalent of get_user_pages() before setting the entry in Stage 2, that should end up performing any set_pte_at() calls to populate the VMM's page tables. So the VMM 'sees' the memory before stage 2.

BTW, ignoring virtualisation, can we ever bring a page in from swap on a
PROT_NONE mapping (say fault-around)? It's not too bad if we keep the
metadata around for when the pte becomes accessible but I suspect we
remove it if the page is removed from swap.

There are two stages of bringing data from swap. First is populating the
swap cache by doing the physical read from swap. The second is actually
restoring the page table entries.

When is the page metadata removed? I want to make sure we don't drop it
for some pte attributes.

The tag metadata for swapped pages lives for the same length of time as the swap metadata itself. The swap code already makes sure that the metadata hangs around as long as there are any swap PTEs in existence, so I think everything should be fine here. The arch_swap_invalidate_xxx() calls match up with the frontswap calls as it has the same lifetime requirements.

Steve