Re: [PATCH 1/6] seccomp: Introduce SECCOMP_PIN_ARCHITECTURE

[Kees Cook]

On Thu, Sep 24, 2020 at 02:41:36AM +0200, Jann Horn wrote:
> On Thu, Sep 24, 2020 at 1:29 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > For systems that provide multiple syscall maps based on audit
> > architectures (e.g. AUDIT_ARCH_X86_64 and AUDIT_ARCH_I386 via
> > CONFIG_COMPAT) or via syscall masks (e.g. x86_x32), allow a fast way
> > to pin the process to a specific syscall table, instead of needing
> > to generate all filters with an architecture check as the first filter
> > action.
> >
> > This creates the internal representation that seccomp itself can use
> > (which is separate from the filters, which need to stay runtime
> > agnostic). Additionally paves the way for constant-action bitmaps.
> 
> I don't really see the point in providing this UAPI - the syscall
> number checking will probably have much more performance cost than the
> architecture number check, and it's not like this lets us avoid the
> check, we're just moving it over into C code.

It's desirable for libseccomp and is a request from systemd (which is,
at this point, the largest seccomp user I know of), as they have no way
to force an arch without doing it in filters, which doesn't help much
with reducing filter runtime.

> 
> > Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
> > ---
> >  include/linux/seccomp.h                       |  9 +++
> >  include/uapi/linux/seccomp.h                  |  1 +
> >  kernel/seccomp.c                              | 79 ++++++++++++++++++-
> >  tools/testing/selftests/seccomp/seccomp_bpf.c | 33 ++++++++
> >  4 files changed, 120 insertions(+), 2 deletions(-)
> >
> > diff --git a/include/linux/seccomp.h b/include/linux/seccomp.h
> > index 02aef2844c38..0be20bc81ea9 100644
> > --- a/include/linux/seccomp.h
> > +++ b/include/linux/seccomp.h
> > @@ -20,12 +20,18 @@
> >  #include <linux/atomic.h>
> >  #include <asm/seccomp.h>
> >
> > +#define SECCOMP_ARCH_IS_NATIVE         1
> > +#define SECCOMP_ARCH_IS_COMPAT         2
> 
> FYI, mips has three different possible "arch" values (per kernel build
> config; the __AUDIT_ARCH_LE flag can also be set, but that's fixed
> based on the config):
> 
>  - AUDIT_ARCH_MIPS
>  - AUDIT_ARCH_MIPS | __AUDIT_ARCH_64BIT
>  - AUDIT_ARCH_MIPS | __AUDIT_ARCH_64BIT | __AUDIT_ARCH_CONVENTION_MIPS64_N32
> 
> But I guess we can deal with that once someone wants to actually add
> support for this on mips.

Yup!

> 
> > +#define SECCOMP_ARCH_IS_MULTIPLEX      3
> 
> Why should X32 be handled specially? If the seccomp filter allows

Because it's a masked lookup into a separate table; the syscalls don't
map to x86_64's table; so for seccomp to correctly figure out which
bitmap to use, it has to do this decoding.

> specific syscalls (as it should), we don't have to care about X32.
> Only in weird cases where the seccomp filter wants to deny specific
> syscalls (a horrible idea), X32 is a concern, and in such cases, the
> userspace code can generate a single conditional jump to deal with it.

I feel like I must not understand what you mean. The x32-aware seccomp
filters are using syscall tests with 0x40000000 included in the values.
So seccomp's bitmap cannot handle this because it must know how many
syscalls to include in a linearly-allocated bitmap.

> And when seccomp is used properly to allow specific syscalls, the
> kernel will just waste time uselessly checking this X32 stuff.

It not measurable in my tests -- seccomp_data::nr is rather hot in the
cache. ;) That said, if it's unwanted, then CONFIG_X86_X32=n is the way
to go.

> [...]
> > diff --git a/kernel/seccomp.c b/kernel/seccomp.c
> [...]
> > +static long seccomp_pin_architecture(void)
> > +{
> > +#ifdef SECCOMP_ARCH
> > +       struct task_struct *task = current;
> > +
> > +       u8 arch = seccomp_get_arch(syscall_get_arch(task),
> > +                                  syscall_get_nr(task, task_pt_regs(task)));
> > +
> > +       /* How did you even get here? */
> 
> Via a racing TSYNC, that's how.

Yes; thanks. This will need to take &current->sighand->siglock.

> 
> > +       if (task->seccomp.arch && task->seccomp.arch != arch)
> > +               return -EBUSY;
> > +
> > +       task->seccomp.arch = arch;
> > +#endif
> > +       return 0;
> > +}
> 
> Why does this return 0 if SECCOMP_ARCH is not defined? That suggests
> to userspace that we have successfully pinned the ABI, even though
> we're actually unable to do so.

Yup; thanks for the catch. This is a logical leftover from the RFC. This
should be, I think:

+       task->seccomp.arch = arch;
+       return 0;
+#else
+	return -EINVAL;
+#endif


-- 
Kees Cook