Re: [PATCH] secretmem: disable memfd_secret() if arch cannot set direct map

From: Andrew Morton
Date: Tue Oct 01 2024 - 18:04:48 EST


On Tue, 1 Oct 2024 09:00:41 +0100 Patrick Roy <roypat@xxxxxxxxxxxx> wrote:

> Return -ENOSYS from memfd_secret() syscall if !can_set_direct_map().
> This is the case for example on some arm64 configurations, where marking
> 4k PTEs in the direct map not present can only be done if the direct map
> is set up at 4k granularity in the first place (as ARM's
> break-before-make semantics do not easily allow breaking apart
> large/gigantic pages).
>
> More precisely, on arm64 systems with !can_set_direct_map(),
> set_direct_map_invalid_noflush() is a no-op, however it returns success
> (0) instead of an error. This means that memfd_secret will seemingly
> "work" (e.g. syscall succeeds, you can mmap the fd and fault in pages),
> but it does not actually achieve its goal of removing its memory from
> the direct map.
>
> Note that with this patch, memfd_secret() will start erroring on systems
> where can_set_direct_map() returns false (arm64 with
> CONFIG_RODATA_FULL_DEFAULT_ENABLED=n, CONFIG_DEBUG_PAGEALLOC=n and
> CONFIG_KFENCE=n), but that still seems better than the current silent
> failure. Since CONFIG_RODATA_FULL_DEFAULT_ENABLED defaults to 'y', most
> arm64 systems actually have a working memfd_secret() and aren't be
> affected.
>
> >From going through the iterations of the original memfd_secret patch
> series, it seems that disabling the syscall in these scenarios was the
> intended behavior [1] (preferred over having
> set_direct_map_invalid_noflush return an error as that would result in
> SIGBUSes at page-fault time), however the check for it got dropped
> between v16 [2] and v17 [3], when secretmem moved away from CMA
> allocations.
>
> [1]: https://lore.kernel.org/lkml/20201124164930.GK8537@xxxxxxxxxx/
> [2]: https://lore.kernel.org/lkml/20210121122723.3446-11-rppt@xxxxxxxxxx/#t
> [3]: https://lore.kernel.org/lkml/20201125092208.12544-10-rppt@xxxxxxxxxx/

Thanks.

> Fixes: 1507f51255c9 ("mm: introduce memfd_secret system call to create "secret" memory areas")

So I'm thinking this fix should be backported into kernels which
contain 1507f51255c9, agree?