netdev
[Top] [All Lists]

Re: netdev.stats change suggestion

To: Martin Josefsson <gandalf@xxxxxxxxxxxxxx>
Subject: Re: netdev.stats change suggestion
From: Martin Devera <devik@xxxxxx>
Date: Wed, 23 Jan 2002 18:02:04 +0100 (CET)
Cc: Dmitrii Tisnek <dima@xxxxxxxxx>, netdev@xxxxxxxxxxx
In-reply-to: <Pine.LNX.4.21.0201231713140.3159-100000@xxxxxxxxxxxxxx>
Sender: owner-netdev@xxxxxxxxxxx
> the problem with 64bit counters on 32bit systems is that after each
> increase of the low 32bits you have to check for an overflow and if one
> occured then we should increase the high 32bits.

you are right. It is probably not good to use ADC which would
generate one more memory write cycle everytime but decent

        ...
        add ax,mem1
        jc 1f
.section .text.stub
1:      inc mem2 
        jmp 2f
.previous
2:

should do it with low overhead. You can count about half of
cycle on Pentiums for not taken branch. Only it will take one
position in BTB ...
Because every packet eat hundrets of cycles AFAIK the single
JC should not be even measurable.
Or am I missing something ? ;)

devik



<Prev in Thread] Current Thread [Next in Thread>