Re: [PATCH] locking/qrwlock: fix write unlock issue in big endian

[xinhui <xinhui.pan@linux.vnet.ibm.com>]

On 2016年06月14日 18:40, Will Deacon wrote:
> On Tue, Jun 14, 2016 at 02:11:48PM +0800, xinhui wrote:
>>
>> On 2016年06月08日 17:22, Will Deacon wrote:
>>> On Thu, Jun 02, 2016 at 06:09:08PM +0800, Pan Xinhui wrote:
>>>> strcut __qrwlock has different layout in big endian machine. we need set
>>>> the __qrwlock->wmode to NULL, and the address is not &lock->cnts in big
>>>> endian machine.
>>>>
>>>> Do as what read unlock does. we are lucky that the __qrwlock->wmode's
>>>> val is _QW_LOCKED.
>>>
>>> Doesn't this have wider implications for the qrwlocks, for example:
>>>
>>>    while ((cnts & _QW_WMASK) == _QW_LOCKED) { ... }
>>>
>>> would actually end up looking at the wrong field of the lock?
>>>
>> I does not clearly understand your idea. :(
>
> That's because I'm talking rubbish :) Sorry, I completely confused myself.
> Locking is bad enough on its own, but add big-endian to the mix and I'm
> all done.
>
>>> Shouldn't we just remove the #ifdef __LITTLE_ENDIAN stuff from __qrwlock,
>>> given that all the struct members are u8?
>>>
>> No. that makes codes complex. for example
>>
>> struct __qrwlock lock;
>>
>> WRITE_ONCE(lock->wmode, _QW_WAITING);
>> if (atomic_(&lock->cnts) == _QW_WAITING) {
>> 	do_something();
>> }
>>
>> IF you remove the  #ifdef __LITTLE_ENDIAN stuff from __qrwlock.
>> codes above obviously will break. And we already have such code.
>
> I was wondering more along the lines of having one definition of the data
> structure, but then defining _QW_* differently depending on endianness
> (i.e. add a << 24 when big-endian). That way queued_write_unlock can
make sense. And I review all the code, there is not much code to be changed.
I will work out one patch based on your idea :)

> stay like it is (having an arch override to handle the big-endian case
> is incredibly ugly).
>
I admit that. HOWEVER from the view of performance, having an arch override is acceptable.

thanks
xinhui
> Will
>