[PATCH] compat: Fix endian issue in union sigval

[Bamvor Jian Zhang]

On 2015/2/14 19:22, Catalin Marinas wrote:
> On Fri, Feb 13, 2015 at 04:56:29PM -0500, Chris Metcalf wrote:
>> On 2/13/2015 5:44 AM, Catalin Marinas wrote:
>>> On Fri, Feb 13, 2015 at 04:00:43PM +0800, Bamvor Jian Zhang wrote:
>>>> diff --git a/arch/arm64/kernel/signal32.c b/arch/arm64/kernel/signal32.c
>>>> index e299de396e9b..32601939a3c8 100644
>>>> --- a/arch/arm64/kernel/signal32.c
>>>> +++ b/arch/arm64/kernel/signal32.c
>>>> @@ -154,8 +154,7 @@ int copy_siginfo_to_user32(compat_siginfo_t __user *to, const siginfo_t *from)
>>>>  	case __SI_TIMER:
>>>>  		 err |= __put_user(from->si_tid, &to->si_tid);
>>>>  		 err |= __put_user(from->si_overrun, &to->si_overrun);
>>>> -		 err |= __put_user((compat_uptr_t)(unsigned long)from->si_ptr,
>>>> -				   &to->si_ptr);
>>>> +		 err |= __put_user(from->si_int, &to->si_int);
>>>>  		break;
>>>>  	case __SI_POLL:
>>>>  		err |= __put_user(from->si_band, &to->si_band);
>>>> @@ -184,7 +183,7 @@ int copy_siginfo_to_user32(compat_siginfo_t __user *to, const siginfo_t *from)
>>>>  	case __SI_MESGQ: /* But this is */
>>>>  		err |= __put_user(from->si_pid, &to->si_pid);
>>>>  		err |= __put_user(from->si_uid, &to->si_uid);
>>>> -		err |= __put_user((compat_uptr_t)(unsigned long)from->si_ptr, &to->si_ptr);
>>>> +		err |= __put_user(from->si_int, &to->si_int);
>>>>  		break;
>>>>  	case __SI_SYS:
>>>>  		err |= __put_user((compat_uptr_t)(unsigned long)
>>
>> I must be confused here, but I don't see that these do anything different.
>>
>> If we are writing 32 bits to to->si_ptr or to->si_int, either way the high 32 bits
>> are irrelevant.  So whether we read it from from->si_ptr and massage the high bits,
>> or just read it from from->si_int as a straight-up 32-bit quantity, either way it
>> seems we should end up writing the same bits to userspace.
>>
>> I would understand the argument if we were overlaying the si_ptr/si_int union
>> from a kernel-side siginfo_t where si_ptr and si_int are different sizes
>> onto userspace, but it doesn't seem we ever do that.
> 
> That's the problem, "from" above is a kernel siginfo_t while "to" is a
> compat_siginfo_t. The call paths go something like:
> 
> 1. user populates sival_int compat_sigevent and invokes
>    compat_sys_mq_notify()
> 2. kernel get_compat_sigevent() copies compat_sigevent into the native
>    sigevent. compat and native sival_int are the same, no problem so
>    far. The other half of 64-bit sival_ptr is zeroed by a memset in this
>    function (this other half can be top or bottom, depending on
>    endianness)
> 3. signal is about to be delivered to user via arch code. The
>    compat_ptr(from->si_ptr) conversion always takes the least
>    significant part of the native si_ptr. On big endian 64-bit, this is
>    zero because get_compat_sigevent() populated the top part of si_ptr
>    with si_int.
> 
> So delivering such signals to compat user always sets si_int to 0.
> Little endian is fine.
> 
> A similar example is sys_timer_create() which takes a sigevent argument.
> Maybe Bamvor has a test case.
> A similar example is sys_timer_create() which takes a sigevent argument.
> Maybe Bamvor has a test case.
Yeap, this issue is came from glibc rt testcases(tst-cputimer1,
tst-cputimer2, tst-cputimer3, tst-timer4, tst-timer5). The above test
cases include timer_create syscall with sigevent.sigev_notify = SIGEV_THREAD.
They failed when they compiled as armeb elf and run on aarch64_be kernel.
When I try to fix it, I found sys_mq_notify is similar.

regards

bamvor
> (btw, I'm off for a week, I'll follow up when I get back)
>