On 2023-05-05 19:42, Florent Revest wrote:
On Thu, May 4, 2023 at 10:30 PM Alexey Izbyshev <izbyshev@xxxxxxxxx>
wrote:
On 2023-05-04 20:09, Florent Revest wrote:
> Add some tests to cover the new PR_MDWE_NO_INHERIT flag of the
> PR_SET_MDWE prctl.
>
> Signed-off-by: Florent Revest <revest@xxxxxxxxxxxx>
> ---
> tools/testing/selftests/mm/mdwe_test.c | 95 ++++++++++++++++++++++++--
> 1 file changed, 89 insertions(+), 6 deletions(-)
>
> diff --git a/tools/testing/selftests/mm/mdwe_test.c
> b/tools/testing/selftests/mm/mdwe_test.c
> index 91aa9c3099e7..9f08ed1b99ae 100644
> --- a/tools/testing/selftests/mm/mdwe_test.c
> +++ b/tools/testing/selftests/mm/mdwe_test.c
> @@ -22,6 +22,8 @@
>
> TEST(prctl_flags)
> {
> + EXPECT_LT(prctl(PR_SET_MDWE, PR_MDWE_NO_INHERIT, 0L, 0L, 7L), 0);
> +
PR_MDWE_NO_INHERIT is defined to an int constant, so passing it to
prctl() without a cast to long or similar may produce wrong code on
64-bit targets (ABIs typically don't require the compiler to clear the
upper 32 bits of a 64-bit register when passing a 32-bit argument, so
va_arg(arg, unsigned long) in prctl() implementation might get junk).
Ah, good catch Alexey! :)
Arguably, defining PR_MDWE_* to plain int constants is a bug, or at
least a footgun for users of uapi headers.
As part of the next version of this series, I'm happy to:
1- change the existing PR_MDWE_REFUSE_EXEC_GAIN to 1UL
2- introduce PR_MDWE_NO_INHERIT as 2UL
Yes, I think it's the right thing to do. I suggest to spell them as (1UL
<< 0), etc. for consistency with all other unsigned long flags in the
header.
But I'm surprised that most of the macros in
include/uapi/linux/prctl.h are the same sort of footguns already ?
Hasn't it been an issue for other prctls yet ?
Yes, they are. I'm not aware of a public discussion of this specific
issue, but note that at least for some prctl() options the kernel
doesn't care about upper bits because arguments are truncated before
doing anything else with them (e.g. for PR_SCHED_CORE raw prctl()
arguments are implicitly converted to what sched_core_share_pid()
expects). Also, actually getting junk in the upper bits might not always
be easy (e.g. on x86-64 all or almost all instructions with r32
destination operand clear the upper bits). Unfortunately, I don't have a
better answer than this.
Alexey
