Re: [BUG] Usersapce MTE error with allocation tag 0 when low on memory

From: Catalin Marinas
Date: Thu Mar 30 2023 - 13:36:18 EST


On Thu, Mar 30, 2023 at 02:56:50PM +0100, Steven Price wrote:
> > On Wed, Mar 29, 2023 at 02:55:49AM +0000, Qun-wei Lin (林群崴) wrote:
> >> Having compared the differences between Kernel-5.15 and Kernel-6.1,
> >> We found the order of swap_free() and set_pte_at() is changed in
> >> do_swap_page().
> >>
> >> When fault in, do_swap_page() will call swap_free() first:
> >> do_swap_page() -> swap_free() -> __swap_entry_free() ->
> >> free_swap_slot() -> swapcache_free_entries() -> swap_entry_free() ->
> >> swap_range_free() -> arch_swap_invalidate_page() ->
> >> mte_invalidate_tags_area() -> mte_invalidate_tags() -> xa_erase()
> >>
> >> and then call set_pte_at():
> >> do_swap_page() -> set_pte_at() -> __set_pte_at() -> mte_sync_tags() ->
> >> mte_sync_page_tags() -> mte_restore_tags() -> xa_load()
> >>
> >> This means that the swap slot is invalidated before pte mapping, and
> >> this will cause the mte tag in XArray to be released before tag
> >> restore.
>
> This analysis looks correct to me. The MTE swap code works on the
> assumption that the set_pte_at() will restore the tags to the page
> before the swap entry is removed. The reordering which has happened
> since has broken this assumption and as you observed can cause the tags
> to be unavailable by the time set_pte_at() is called.
>
> >> After I moved swap_free() to the next line of set_pte_at(), the problem
> >> is disappeared.
> >>
> >> We suspect that the following patches, which have changed the order, do
> >> not consider the mte tag restoring in page fault flow:
> >> https://lore.kernel.org/all/20220131162940.210846-5-david@xxxxxxxxxx/
>
> I'm not sure I entirely follow the reasoning in this patch, so I'm not
> sure whether it's safe to just move swap_free() down to below
> set_pte_at() or if that reintroduces the information leak.
>
> I also wonder if sparc has a similar issue as the arch_do_swap()
> callback is located next to set_pte_at().

SPARC has a potential race here since the page is made visible to the
user but the tags are not restored yet (I raised this before). But even
ignoring this small window, arch_do_swap() needs to have the metadata
available.

> >> Any suggestion is appreciated.
>
> The other possibility is to add a(nother) callback for MTE in
> arch_do_swap() that calls mte_restore_tags() on the page before the
> swap_free() call rather than depending on the hook in set_pte_at().

I think we should move arch_do_swap_page() earlier before swap_free()
and in arm64 we copy the tags to pte_page(pte). I don't think SPARC
would have any issues with this change (and it also fixes their race).

--
Catalin