On Wed, Jan 22, 2020 at 10:29:05AM +0100, Greg Kroah-Hartman wrote:
> From: Christian Brauner <christian.brauner@xxxxxxxxxx>
>
> commit 6b3ad6649a4c75504edeba242d3fd36b3096a57f upstream.
>
> Commit 69f594a38967 ("ptrace: do not audit capability check when outputing /proc/pid/stat")
> introduced the ability to opt out of audit messages for accesses to various
> proc files since they are not violations of policy. While doing so it
> somehow switched the check from ns_capable() to
> has_ns_capability{_noaudit}(). That means it switched from checking the
> subjective credentials of the task to using the objective credentials. This
> is wrong since. ptrace_has_cap() is currently only used in
> ptrace_may_access() And is used to check whether the calling task (subject)
> has the CAP_SYS_PTRACE capability in the provided user namespace to operate
> on the target task (object). According to the cred.h comments this would
> mean the subjective credentials of the calling task need to be used.
> This switches ptrace_has_cap() to use security_capable(). Because we only
> call ptrace_has_cap() in ptrace_may_access() and in there we already have a
> stable reference to the calling task's creds under rcu_read_lock() there's
> no need to go through another series of dereferences and rcu locking done
> in ns_capable{_noaudit}().
>
> As one example where this might be particularly problematic, Jann pointed
> out that in combination with the upcoming IORING_OP_OPENAT feature, this
> bug might allow unprivileged users to bypass the capability checks while
> asynchronously opening files like /proc/*/mem, because the capability
> checks for this would be performed against kernel credentials.
>
> To illustrate on the former point about this being exploitable: When
> io_uring creates a new context it records the subjective credentials of the
> caller. Later on, when it starts to do work it creates a kernel thread and
> registers a callback. The callback runs with kernel creds for
> ktask->real_cred and ktask->cred. To prevent this from becoming a
> full-blown 0-day io_uring will call override_cred() and override
> ktask->cred with the subjective credentials of the creator of the io_uring
> instance. With ptrace_has_cap() currently looking at ktask->real_cred this
> override will be ineffective and the caller will be able to open arbitray
> proc files as mentioned above.
> Luckily, this is currently not exploitable but will turn into a 0-day once
> IORING_OP_OPENAT{2} land in v5.6. Fix it now!
>
> Cc: Oleg Nesterov <oleg@xxxxxxxxxx>
> Cc: Eric Paris <eparis@xxxxxxxxxx>
> Cc: stable@xxxxxxxxxxxxxxx
> Reviewed-by: Kees Cook <keescook@xxxxxxxxxxxx>
> Reviewed-by: Serge Hallyn <serge@xxxxxxxxxx>
> Reviewed-by: Jann Horn <jannh@xxxxxxxxxx>
> Fixes: 69f594a38967 ("ptrace: do not audit capability check when outputing /proc/pid/stat")
> Signed-off-by: Christian Brauner <christian.brauner@xxxxxxxxxx>
> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
>
> ---
> kernel/ptrace.c | 15 ++++++++++-----
> 1 file changed, 10 insertions(+), 5 deletions(-)
>
> --- a/kernel/ptrace.c
> +++ b/kernel/ptrace.c
> @@ -258,12 +258,17 @@ static int ptrace_check_attach(struct ta
> return ret;
> }
>
> -static int ptrace_has_cap(struct user_namespace *ns, unsigned int mode)
> +static bool ptrace_has_cap(const struct cred *cred, struct user_namespace *ns,
> + unsigned int mode)
> {
> + int ret;
> +
> if (mode & PTRACE_MODE_NOAUDIT)
> - return has_ns_capability_noaudit(current, ns, CAP_SYS_PTRACE);
> + ret = security_capable(cred, ns, CAP_SYS_PTRACE);
> else
> - return has_ns_capability(current, ns, CAP_SYS_PTRACE);
> + ret = security_capable(cred, ns, CAP_SYS_PTRACE);
> +
> + return ret == 0;
This results in
if (condition)
do_something;
else
do_the_same;
Is that really correct ? The upstream patch calls security_capable()
with additional CAP_OPT_NOAUDIT vs. CAP_OPT_NONE parameter, which does
make sense. But I don't really see the benefit of the change above.
Guenter
