Re: [PATCH net-next-2.6] net: __alloc_skb() speedup

[David Miller <davem@davemloft.net>]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Wed, 05 May 2010 14:00:09 +0200

> Sorry, I was thinking about the shinfo part :
> 
> memset(shinfo, 0, offsetof(struct skb_shared_info, dataref));
> 
> offsetof(struct skb_shared_info, dataref) is small enough and we dont
> dirty a full cache line, so maybe I can keep prefetchw(data + size) ?

You do dirty a full line on sparc64, the prefetch invalidate goes a L1
cache line at a time, so 32 bytes.  And this memset() is 40 bytes.
The call to the memset symbol is still generated by gcc for this case.

I think the cutoff for doing it inline is something like 16 bytes on
sparc64, four 64-bit loads and stores.

Unlike x86 these risc chips don't have string-op instructions, and for
sparc64 and powerpc the instructions are fixed in size (4 bytes) so
the inline cost is "(memset_size / word_size) * 4".  Whereas on x86 the
inlining cost is more-or-less fixed.

> If not, in which cases can we use prefetchw() in kernel, if some arches
> dont handle it well ?

It has to be looked at in a case-by-case basis.  There is no simple
answer here.