Re: [PATCH 1/2] x86-64 TIF flags for debug regs and io bitmap in ctxsw

[Andi Kleen <ak@suse.de>]

On Tuesday 04 July 2006 11:05, Mikael Pettersson wrote:
> On Tue, 4 Jul 2006 01:14:13 -0700, Stephane Eranian wrote:
> >On Tue, Jul 04, 2006 at 09:51:49AM +0200, Arjan van de Ven wrote:
> >> > -		}
> >> > -	}
> >> > +	if (unlikely((task_thread_info(next_p)->flags & _TIF_WORK_CTXSW))
> >> > +	    || test_tsk_thread_flag(prev_p, TIF_IO_BITMAP))
> >> > +		__switch_to_xtra(prev_p, next_p, tss);
> >> 
> >> well isn't this replacing an if() (which isn't cheap but also not
> >> expensive, due to unlikely()) with an atomic operation (which *is*
> >> expensive) ?
> >> 
> >Andi is right. I double checked the test_tsk_thread_flag() and it does not
> >use atomic ops.
> 
> The test_tsk_thread_flag() does not, but what about all the
> other places in the patch where currently unsychronised loads
> or stores of ->io_bitmap_ptr or ->debugreg7 get replaced or
> extended with locked-on-SMP {set,clear}_{tsk_,}thread_flag()
> operations?

They are all slow paths where it doesn't matter.

-Andi