lkml.org 
[lkml]   [2012]   [Jul]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH 3.5 2/2] seccomp: Future-proof against silly tracers
From
On Tue, Jul 17, 2012 at 9:13 PM, Will Drewry <wad@chromium.org> wrote:
> On Tue, Jul 17, 2012 at 6:19 PM, Andy Lutomirski <luto@amacapital.net> wrote:
>> Currently, if a tracer changes a syscall nr to __NR_future_enosys,
>> behavior will differ between kernels that know about
>> __NR_future_enosys (and return -ENOSYS) and older kernels (which
>> return the value from pt_regs). This is silly; we should just
>> return -ENOSYS.
>>
>> This is unlikely to ever happen on x86 because the return value in
>> pt_regs starts out as -ENOSYS, but a silly tracer can change that.
>>
>> Signed-off-by: Andy Lutomirski <luto@amacapital.net>
>> Cc: Will Drewry <wad@chromium.org>
>> ---
>> arch/x86/include/asm/syscall.h | 11 +++++++++++
>> kernel/seccomp.c | 15 +++++++++++++++
>> 2 files changed, 26 insertions(+), 0 deletions(-)
>>
>> diff --git a/arch/x86/include/asm/syscall.h b/arch/x86/include/asm/syscall.h
>> index 1ace47b..8191e057 100644
>> --- a/arch/x86/include/asm/syscall.h
>> +++ b/arch/x86/include/asm/syscall.h
>
> Since this is used in an arch-wide location, the prototype and comment
> should be in asm-generic/syscall.h too.
>
>> @@ -70,6 +70,17 @@ static inline void syscall_set_return_value(struct task_struct *task,
>> regs->ax = (long) error ?: val;
>> }
>>
>> +static inline bool syscall_is_nr_future(struct task_struct *task,
>> + struct pt_regs *regs)
>> +{
>> +#ifdef CONFIG_IA32_EMULATION
>> + if (task_thread_info(task)->status & TS_COMPAT)
>> + return syscall_get_nr(task, regs) > __NR_ia32_syscall_max;
>> +#endif
>> +
>> + return syscall_get_nr(task, regs) > __NR_syscall_max;
>
> I'm not sure how easy this will be to implement on some of the arches
> where this data isn't bubbled up. It'd be good if some non-x86 arch
> maintainers chimed in (since x86 is easy and already works as expected
> :).
>

Since x86 always returns -ENOSYS with an invalid syscall and only x86
supports seccomp filter in 3.5, I'd propose pushing this off for 3.6+
to get more feedback from the relevant parties. Not doing it now
doesn't expose any users of 3.5 to any sort of changing ABI.

(Otherwise, it seems fine but may make adding new arches slightly more
onerous, but I suspect ftrace needs this sort of info too as it
spreads to other arches!)

>> +}
>> +
>> #ifdef CONFIG_X86_32
>>
>> static inline void syscall_get_arguments(struct task_struct *task,
>> diff --git a/kernel/seccomp.c b/kernel/seccomp.c
>> index 5af44b5..bd7527d 100644
>> --- a/kernel/seccomp.c
>> +++ b/kernel/seccomp.c
>> @@ -429,6 +429,21 @@ int __secure_computing(int this_syscall)
>> */
>> if (fatal_signal_pending(current))
>> break;
>> +
>> + if (syscall_is_nr_future(current, regs)) {
>> + /*
>> + * If the tracer selects a system call that
>> + * only future kernels know about, we still need
>> + * to execute that syscall by returning
>> + * -ENOSYS. (On x86, this only matters if the
>> + * tracer changed the return value, which would
>> + * be silly, but user code can be silly.)
>> + */
>> + syscall_set_return_value(current, regs,
>> + -ENOSYS, 0);
>> + goto skip;
>> + }
>> +
>> if (syscall_get_nr(current, regs) < 0)
>> goto skip; /* Explicit request to skip. */
>>
>> --
>> 1.7.7.6
>>
>
> thanks!
> will


\
 
 \ /
  Last update: 2012-07-18 21:21    [from the cache]
©2003-2011 Jasper Spaans