Re: [PATCH] powerpc/syscall: Avoid atomic test_and_clear_thread_flag on fast path

["Christophe Leroy (CS GROUP)" <chleroy@kernel.org>]

Le 29/09/2026 à 11:50, Shrikanth Hegde a écrit :
> 
> 
> On 9/29/26 12:47 PM, Mukesh Kumar Chaurasiya (IBM) wrote:
>> In system_call_exception(), test_and_clear_thread_flag(TIF_SYSCALL_RET)
>> performs an atomic bit test-and-clear operation on every system call 
>> entry.
>> This introduces an unnecessary performance regression on the syscall fast
>> path.
>>
> 
> Do you have numbers?
> 
>> Combine the TIF_SYSCALL_RET check with the unlikely error condition from
>> syscall_enter_from_user_mode_randomize_stack() using non-atomic
>> test_thread_flag(). If the flag is set or entry failed, only then 
>> clear the
>> flag via clear_thread_flag() and return the error value.
>>
> 
> It probably needs a fixes tag.

It is not a bug, unless we can figure out a quantified regression.

> 
>> Signed-off-by: Mukesh Kumar Chaurasiya (IBM) <mkchauras@gmail.com>
>> ---
>>   arch/powerpc/kernel/syscall.c | 6 ++----
>>   1 file changed, 2 insertions(+), 4 deletions(-)
>>
>> diff --git a/arch/powerpc/kernel/syscall.c b/arch/powerpc/kernel/ 
>> syscall.c
>> index fbefe1927b10..8529827d697f 100644
>> --- a/arch/powerpc/kernel/syscall.c
>> +++ b/arch/powerpc/kernel/syscall.c
>> @@ -18,14 +18,12 @@ notrace long system_call_exception(struct pt_regs 
>> *regs, unsigned long r0)
>>       long ret;
>>       syscall_fn f;
>> -    if (unlikely(!syscall_enter_from_user_mode_randomize_stack(regs, 
>> &r0))) {
>> +    if (unlikely(!syscall_enter_from_user_mode_randomize_stack(regs, 
>> &r0) ||
>> +            test_thread_flag(TIF_SYSCALL_RET))) {
>>           clear_thread_flag(TIF_SYSCALL_RET);
>>           return syscall_get_error(current, regs);
>>       }
>> -    if (unlikely(test_and_clear_thread_flag(TIF_SYSCALL_RET)))
> 
> I am bit surprised that test_and_clear_bit does
> atomic operations even when bit is not set.

It does (on ppc64 smp)

	sync
   1:	ldarx
	andc
	stdcx.
	bne 1b
	sync

It is optimised to clear a bit that is set. When the bit is already 
unset you still get the full sequence allthough it does nothing.

> 
>> -        return syscall_get_error(current, regs);
>> -
>>       if (unlikely(r0 >= NR_syscalls)) {
>>           if (unlikely(trap_is_unsupported_scv(regs))) {
>>               /* Unsupported scv vector */
>