Re: fast softint support for sparc64

>>> Eduardo Horvath <[email protected]> wrote

> If you're trying to order the store to the CI_MTX_COUNT with the store to 
> L_CTXSWTCH then I think it should be #StoreStore.  
> 
> Note the comment in membar_ops.S:
> 
> /* These assume Total Store Order (TSO) */
> 
> Um, if you're running TSO then you don't need any membars at all.  We 
> should really be coding this for RMO to get some performance improvements.  
> Otherwise leaving them out will reduce instruction count and CPU cycles.

Ok, I will just leave it out.  Thank you for detailed comment.

> Hm.  Looking more at mi_switch() it doesn't seem all that mi to me.  It 
> appears to be using l->l_ctxswtch as a synchronization flag, but the code 
> where it's set:
> 
>     733 		KASSERT(l->l_ctxswtch == 0);
>     734 		l->l_ctxswtch = 1;
>     735 		l->l_ncsw++;
>     736 		KASSERT((l->l_pflag & LP_RUNNING) != 0);
>     737 		l->l_pflag &= ~LP_RUNNING;
>     738 
> 
> really needs a membar after setting it to 1 to make sure the store 
> completes before continuing.  Well this explains why things didn't work 
> when I fixed up the stuff in libatomic and tried running the kernel RMO.
> 
> OTOH I think you can just leave the membar out.  It should work fine and 
> simplify code maintenance in the future.  As far as I'm concerned extra 
> membars in the wrong places are worse than not having them at all.  The 
> chance someone will go fix the membar correctly in future is pretty close 
> to zero.
> 
> Eduardo

Thanks,

Takeshi Nakayama