Re: [PATCH v5 13/25] m68k: add asm/syscall.h

[Dmitry V. Levin]

Hi Geert,

On Wed, Dec 12, 2018 at 01:27:14PM +0100, Geert Uytterhoeven wrote:
> On Wed, Dec 12, 2018 at 1:04 PM Dmitry V. Levin <ldv@xxxxxxxxxxxx> wrote:
> > On Wed, Dec 12, 2018 at 10:43:33AM +0100, Geert Uytterhoeven wrote:
> > > On Wed, Dec 12, 2018 at 10:27 AM Dmitry V. Levin <ldv@xxxxxxxxxxxx> wrote:
> > > > On Wed, Dec 12, 2018 at 10:01:29AM +0100, Geert Uytterhoeven wrote:
> > > > > On Wed, Dec 12, 2018 at 9:55 AM Dmitry V. Levin <ldv@xxxxxxxxxxxx> wrote:
> > > > > > On Mon, Dec 10, 2018 at 04:30:25PM +0300, Dmitry V. Levin wrote:
> > > > > > > On Mon, Dec 10, 2018 at 02:06:28PM +0100, Geert Uytterhoeven wrote:
> > > > > > > > On Mon, Dec 10, 2018 at 1:41 PM Dmitry V. Levin <ldv@xxxxxxxxxxxx> wrote:
> > > > > > > > > On Mon, Dec 10, 2018 at 09:45:42AM +0100, Geert Uytterhoeven wrote:
> > > > > > > > > > On Mon, Dec 10, 2018 at 5:30 AM Dmitry V. Levin <ldv@xxxxxxxxxxxx> wrote:
> > > > > > > > > > > syscall_get_* functions are required to be implemented on all
> > > > > > > > > > > architectures in order to extend the generic ptrace API with
> > > > > > > > > > > PTRACE_GET_SYSCALL_INFO request.
> > > > > > > > > > >
> > > > > > > > > > > This introduces asm/syscall.h on m68k implementing all 5 syscall_get_*
> > > > > > > > > > > functions as documented in asm-generic/syscall.h: syscall_get_nr,
> > > > > > > > > > > syscall_get_arguments, syscall_get_error, syscall_get_return_value,
> > > > > > > > > > > and syscall_get_arch.
> > > > > > > > > > >
> > > > > > > > > > > Cc: Geert Uytterhoeven <geert@xxxxxxxxxxxxxx>
> > > > > > > > > > > Cc: Oleg Nesterov <oleg@xxxxxxxxxx>
> > > > > > > > > > > Cc: Andy Lutomirski <luto@xxxxxxxxxx>
> > > > > > > > > > > Cc: Elvira Khabirova <lineprinter@xxxxxxxxxxxx>
> > > > > > > > > > > Cc: Eugene Syromyatnikov <esyr@xxxxxxxxxx>
> > > > > > > > > > > Cc: linux-m68k@xxxxxxxxxxxxxxxxxxxx
> > > > > > > > > > > Signed-off-by: Dmitry V. Levin <ldv@xxxxxxxxxxxx>
> > > > > > > > > > > ---
> > > > > > > > > > >
> > > > > > > > > > > Notes:
> > > > > > > > > > >     v5: added syscall_get_nr, syscall_get_arguments, syscall_get_error,
> > > > > > > > > > >         and syscall_get_return_value
> > > > > > > > > > >     v1: added syscall_get_arch
> > > > > > > > > >
> > > > > > > > > > > --- /dev/null
> > > > > > > > > > > +++ b/arch/m68k/include/asm/syscall.h
> > > > > > > > > > > @@ -0,0 +1,39 @@
> > > > > > > > > >
> > > > > > > > > > > +static inline void
> > > > > > > > > > > +syscall_get_arguments(struct task_struct *task, struct pt_regs *regs,
> > > > > > > > > > > +                     unsigned int i, unsigned int n, unsigned long *args)
> > > > > > > > > > > +{
> > > > > > > > > > > +       BUG_ON(i + n > 6);
> > > > > > > > > >
> > > > > > > > > > Does this have to crash the kernel?
> > > > > > > > >
> > > > > > > > > This is what most of other architectures do, but we could choose
> > > > > > > > > a softer approach, e.g. use WARN_ON_ONCE instead.
> > > > > > > > >
> > > > > > > > > > Perhaps you can return an error code instead?
> > > > > > > > >
> > > > > > > > > That would be problematic given the signature of this function
> > > > > > > > > and the nature of the potential bug which would most likely be a usage error.
> > > > > > > >
> > > > > > > > Of course to handle that, the function's signature need to be changed.
> > > > > > > > Changing it has the advantage that the error handling can be done at the
> > > > > > > > caller, in common code, instead of duplicating it for all
> > > > > > > > architectures, possibly
> > > > > > > > leading to different semantics.
> > > > > > >
> > > > > > > Given that *all* current users of syscall_get_arguments specify i == 0
> > > > > > > (and there is an architecture that has BUG_ON(i)),
> > > > > > > it should be really a usage error to get into situation where i + n > 6,
> > > > > > > I wish a BUILD_BUG_ON could be used here instead.
> > > > > > >
> > > > > > > I don't think it worths pushing the change of API just to convert
> > > > > > > a "cannot happen" assertion into an error that would have to be dealt with
> > > > > > > on the caller side.
> > > > > >
> > > > > > I suggest the following BUG_ON replacement for syscall_get_arguments:
> > > > > >
> > > > > > #define SYSCALL_MAX_ARGS 6
> > > > > >
> > > > > > static inline void
> > > > > > syscall_get_arguments(struct task_struct *task, struct pt_regs *regs,
> > > > > >                       unsigned int i, unsigned int n, unsigned long *args)
> > > > > > {
> > > > > >         /*
> > > > > >          * Ideally there should have been
> > > > > >          * BUILD_BUG_ON(i + n > SYSCALL_MAX_ARGS);
> > > > > >          * instead of these checks.
> > > > > >          */
> > > > > >         if (unlikely(i > SYSCALL_MAX_ARGS)) {
> > > > > >                 WARN_ONCE(1, "i > SYSCALL_MAX_ARGS");
> > > > > >                 return;
> > > > >
> > > > > Does this have security implications, as args is an output parameter?
> > > > > I.e. if you don't fill the array, the caller will use whatever is on the stack.
> > > > > Can this ever be passed to userspace, leaking data?
> > > >
> > > > In the current kernel code n is always less or equal to 6,
> > > > but in theory future changes can potentially break the assertion
> > > > and this could lead to leaking data to userspace.
> > >
> > > OK.
> > >
> > > > Do you think we should rather be defensive and add some memsets, e.g.
> > > >
> > > >         if (unlikely(i > SYSCALL_MAX_ARGS)) {
> > > >                 WARN_ONCE(1, "i > SYSCALL_MAX_ARGS");
> > > >                 memset(args, 0, n * sizeof(args[0]));
> > > >                 return;
> > > >         }
> > > >         if (unlikely(n > SYSCALL_MAX_ARGS - i)) {
> > > >                 unsigned int extra = n - (SYSCALL_MAX_ARGS - i);
> > > >
> > > >                 WARN_ONCE(1, "i + n > SYSCALL_MAX_ARGS");
> > > >                 n = SYSCALL_MAX_ARGS - i;
> > > >                 memset(&args[n], 0, extra * sizeof(args[0]));
> > > >         }
> > > > ?
> > >
> > > Yes please.
> > >
> > > But please handle all of that in the generic code, so it doesn't have to be
> > > replicated across all architectures.
> > >
> > > E.g. make syscall_get_arguments() a wrapper in generic code, calling
> > > __syscall_get_arguments() in architecture-specific code.
> > >
> > > And make the latter return int, so it can indicate other failures.
> >
> > Other failures?  What syscall_get_arguments is expected to do
> > if __syscall_get_arguments returned, say, -1?
> 
> Fail. Just like in case of other generic ill conditions it can detect itself.

Sorry, I don't quite follow.  syscall_get_arguments() has no return code,
so all it can possibly do is to zero out args[], e.g.

        if (unlikely(__syscall_get_arguments(task, regs, i, n, args) < 0)) {
                memset(args, 0, n * sizeof(args[0]));
                return;
        }

Do you mean this?


-- 
ldv

Attachment: signature.asc
Description: PGP signature