Re: [PATCH] add likely around access_ok for uaccess

[Manfred Spraul <manfred@colorfullife.com>]

Pavel Machek wrote:

>Hi!
>
>  
>
>>>>while trying to figure out why sysv msg is around 30% slower than pipes 
>>>>for data transfers I noticed that gcc's autodetection (3.2.2) guesses 
>>>>the "if(access_ok())" tests in uaccess.h wrong and puts the error memset 
>>>>into the direct path and the copy out of line.
>>>>
>>>>The attached patch adds likely to the tests - any objections? What about 
>>>>the archs except i386?
>>>>        
>>>>
>>>How much speedup did you gain?
>>>      
>>>
>>How much can it hurt?
>>    
>>
>
>The change is obviously okay, I just wanted to know... If it gains 30%
>on sysv messages.. that would be pretty big surprise.
>
No.
I didn't benchmark it at all. I'd expect one or two cycles.
The 30% difference are due to misaligned buffers: The sysvmsg ABI uses
struct msgbuf {
    unsigned long type;
    char data[];
};
msgbuf was aligned, thus data on an not 8-byte aligned address. Thus the 
"rep;movsl" microcode fastpath didn't kick in, and that caused the 
30-50% slowdown.
After manually misaligning the msgbuf structure, and properly aligning 
the kernel buffers, the performance of sysvmsg is now identical to 
pipes: Around 4 k cycles for a one-byte ping-pong, and around 10k cycles 
for 4 kB ping-pong, with a Celeron mobile 1.13 GHz.

I haven't decided yet if a patch to align the kernel buffers is a good 
thing or not - it only helps my benchmark, I'm not aware of a real-world 
app that uses sysvmsg for bulk data transfers.

--
    Manfred