Re: [PATCH] 3/5 explicit-iopl

[Zachary Amsden <zach@vmware.com>]

Andi Kleen wrote:

>zach@vmware.com writes:
>  
>
>>Unfortunately, this added one field to the thread_struct.  But as a bonus, on
>>P4, the fastest time measured for switch_to() went from 312 to 260 cycles, a
>>win of about 17% in the fast case through this performance critical path.
>>    
>>
>
>Cool! Definitely want this on x86-64 too.
>  
>

Well... maybe.  On Opteron and/or Intel EMT it may not be a win.  The 
cost of the branch could overtake the cost of the POPF (that's the 
expensive one).  Grrr.

>Can we perhaps get rid of the PUSHF/POPF in the SYSENTER syscall path too?
>iirc they were only for single stepping. But SYSENTER doesn't disable
>single stepping, so the debug handler could detect this and set
>some magic flag that restores it on syscall exit.
>  
>

A context switch requires IRET, which requires the flags to be saved, so 
you can't eliminate the pushf (*) IIRC, the popf is already omitted.  
Many of these patches may be beneficial to x86-64, but. unfortunately 
the performance deltas may not translate.  Lets hope they do!  
Unfortunately, that requires re-measuring the cost of switch_to(), which 
was quite amusing to do.  I can send you diffs if you're interested, but 
using printk around this path turned out to be a really bad idea ;)  I 
really would like to bring some of the cleanup and performance work I've 
done on i386 over to x86_64 as well, but that is still probably a couple 
of weeks out.  If you can't wait, you're welcome to port pieces you 
like!  Let me know.

(*) Well, you could.  It's just that system calls would have to clobber 
flags - hmm.. sysenter based calls already do. But I'm not 100% sure 
there isn't some bogon case where kernel preemption could cause you a 
problem.  Keeping around the fake IRET frame still appears to be a good 
thing to do just for the benefit of ptrace / debug functionality.  PUSHF 
is cheap on every core I have measured on.

Zach