Re: [PATCH] powerpc: introduce {cmp}xchg for u8 and u16

[Pan Xinhui <xinhui@linux.vnet.ibm.com>]

Hello Peter,

On 2016年04月12日 22:30, Peter Zijlstra wrote:
> On Sun, Apr 10, 2016 at 10:17:28PM +0800, Pan Xinhui wrote:
>>
>> On 2016年04月08日 15:47, Peter Zijlstra wrote:
>>> On Fri, Apr 08, 2016 at 02:41:46PM +0800, Pan Xinhui wrote:
>>>> From: pan xinhui <xinhui.pan@linux.vnet.ibm.com>
>>>>
>>>> Implement xchg{u8,u16}{local,relaxed}, and
>>>> cmpxchg{u8,u16}{,local,acquire,relaxed}.
>>>>
>>>> Atomic operation on 8-bit and 16-bit data type is supported from power7
>>>
>>> And yes I see nothing P7 specific here, this implementation is for
>>> everything PPC64 afaict, no?
>>>
>> Hello Peter,
>> 	No, it's not for every ppc. So yes, I need add #ifdef here. Thanks for pointing it out.
>> We might need a new config option and let it depend on POWER7/POWER8_CPU or even POWER9...
> 
> Right, I'm not sure if PPC has alternatives, but you could of course
> runtime patch the code from emulated with 32bit ll/sc to native 8/16bit
> ll/sc if present on the current CPU if you have infrastructure for these
> things.
> 
seems interesting. I have no idea about how to runtime patch the code. I will try to learn that.
If so, we need change {cmp}xchg into uninline functions?

>>> Also, note that you don't need explicit 8/16 bit atomics to implement
>>> these. Its fine to use 32bit atomics and only modify half the word.
>>>
>> That is true. But I am a little worried about the performance. It will
>> forbid any other tasks to touch the other half word during the
>> load/reserve, right?
> 
> Well, not forbid, it would just make the LL/SC fail and try again. Other
> archs already implement them this way. See commit 3226aad81aa6 ("sh:
> support 1 and 2 byte xchg") for example.
> 
thanks for your explanation. :)

I wrote one similar patch as you suggested.

I paste the new __xchg_u8's alpha implementation here. it need rewrite to be understood easily...
It does work, but some performance tests are needed later.

static __always_inline unsigned long
__xchg_u8_local(volatile void *p, unsigned char val)
{
        unsigned int prev, prev_mask, tmp, offset, _val, *_p;

        _p = (unsigned int *)round_down((unsigned long)p, sizeof(int));
        _val = val;
        offset = 8 * ( (unsigned long)p - (unsigned long )_p) ;
#ifndef CONFIG_CPU_LITTLE_ENDIAN
        offset = 8 * (sizeof(int) - sizeof(__typeof__(val))) - offset;
#endif
        _val <<= offset;
        prev_mask = ~((unsigned int)(__typeof__ (val))-1 << offset);

        __asm__ __volatile__(
"1:     lwarx   %0,0,%3\n"
"       and %1,%0,%5\n"
"       or %1,%1,%4\n"
        PPC405_ERR77(0,%2)
"       stwcx.  %1,0,%3\n"
"       bne-    1b"
        : "=&r" (prev), "=&r" (tmp), "+m" (*(volatile unsigned int *)_p)
        : "r" (_p), "r" (_val), "r" (prev_mask)
        : "cc", "memory");

        return prev >> offset;
}

>> I am working on the qspinlock implementation on PPC.
>> Your and Waiman's patches are so nice. :)
> 
> Thanks!, last time I looked at PPC spinlocks they could not use things
> like ticket locks because PPC might be a guest and fairness blows etc..
> 
> You're making the qspinlock-paravirt thing work on PPC, or doing
> qspinlock only for bare-metal PPC?
> 
I am making the both work. :)
qspinlock works on PPC now. I am preparing the patches and will send them out in next weeks :)

The paravirt work is a little hard.
currently, there are pv_wait() and pv_kick(). but only pv_kick has the parameter cpu(who will hold the lock as soon as the lock is unlocked). 
We need parameter cpu(who holds the lock now) in pv_wait,too.

thanks
xinhui