Re: [RFC][PATCH] x86, syscalls: use SYSCALL_DEFINE() macros for sys_modify_ldt()

From: Brian Gerst
Date: Sat Oct 14 2017 - 02:26:23 EST


On Sat, Oct 14, 2017 at 12:42 AM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> On Fri, Oct 13, 2017 at 4:49 PM, Brian Gerst <brgerst@xxxxxxxxx> wrote:
>> On Fri, Oct 13, 2017 at 5:03 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>>> On Fri, Oct 13, 2017 at 1:39 PM, Dave Hansen
>>> <dave.hansen@xxxxxxxxxxxxxxx> wrote:
>>>>
>>>> I noticed that we don't have tracepoints for sys_modify_ldt(). I
>>>> think that's because we define it directly instead of using the
>>>> normal SYSCALL_DEFINEx() macros.
>>>>
>>>> Is there a reason for that, or were they just missed when the
>>>> macros were created?
>>>
>>> No, and it's a longstanding fsckup that I think you can't fix like
>>> this because...
>>>
>>>>
>>>> Cc: x86@xxxxxxxxxx
>>>> Cc: Andy Lutomirski <luto@xxxxxxxxxx>
>>>>
>>>> ---
>>>>
>>>> b/arch/x86/include/asm/syscalls.h | 2 +-
>>>> b/arch/x86/kernel/ldt.c | 5 +++--
>>>> b/arch/x86/um/ldt.c | 3 ++-
>>>> 3 files changed, 6 insertions(+), 4 deletions(-)
>>>>
>>>> diff -puN arch/x86/kernel/ldt.c~x86-syscall-macros-modify_ldt arch/x86/kernel/ldt.c
>>>> --- a/arch/x86/kernel/ldt.c~x86-syscall-macros-modify_ldt 2017-10-13 13:30:12.802553391 -0700
>>>> +++ b/arch/x86/kernel/ldt.c 2017-10-13 13:30:12.817553391 -0700
>>>> @@ -12,6 +12,7 @@
>>>> #include <linux/string.h>
>>>> #include <linux/mm.h>
>>>> #include <linux/smp.h>
>>>> +#include <linux/syscalls.h>
>>>> #include <linux/slab.h>
>>>> #include <linux/vmalloc.h>
>>>> #include <linux/uaccess.h>
>>>> @@ -294,8 +295,8 @@ out:
>>>> return error;
>>>> }
>>>>
>>>> -asmlinkage int sys_modify_ldt(int func, void __user *ptr,
>>>> - unsigned long bytecount)
>>>> +SYSCALL_DEFINE3(modify_ldt, int , func , void __user * , ptr ,
>>>> + unsigned long , bytecount)
>>>
>>> sys_modify_ldt() returns int, which is wrong, and it's visibly wrong
>>> to 64-bit user code. So I think you need to make sure that the return
>>> value is cast to int in all cases.
>>
>> I don't think there will be a problem here. If 64-bit userspace
>> treats it as an int, it will truncate to 32-bit signed and all is
>> well. If it is treating it as a long, then it's currently broken for
>> errors anyways.
>>
>
> Let me say what I mean more clearly:
>
> The current code is buggy: specifically, a 64-bit modify_ldt() call
> that *fails* will return something like (int)-EFAULT. This is bogus,
> but it's the ABI. There's even a selftest in the kernel tree that
> notices this (although it doesn't check it right now). All that needs
> to happen for this patch to be okay AFAIK is to make sure that we
> preserve that bug instead of accidentally fixing it.

return (unsigned int)ret;

Problem solved.

--
Brian Gerst