Much improved sosend_*() functions

[Andre Oppermann]

John-Mark Gurney wrote:
> Randall Stewart wrote this message on Fri, Sep 29, 2006 at 16:55 -0400:
> 
>>Mike Silbersack wrote:
>>
>>>On Fri, 29 Sep 2006, Andre Oppermann wrote:
>>>
>>>
>>>
>>>>over it an copies the data into the mbufs by using uiomove().  
>>>>sosend_dgram()
>>>>and sosend_generic() are change to use m_uiotombuf() instead of 
>>>>sosend_copyin().
>>>
>>>
>>>Can you do some UDP testing with 512b, 1K, 2K, 4K, 8K, and 16K packets to
>>>see if performance changes there as well?
>>
>>Hmm.. I would think 512b and 1K will not show any
>>improvement.. since they would probably end up either
>>in an mbuf chain.. or a single 2k (or maybe 4k) cluster..
>>... quite a waste.. now if we had 512b and 1k clusters that
>>would be cool...
>>
>>In fact I have always thought we should:
>>
>>a) have no data portion in an mbuf.. just pointers i.e. always
>>   an EXT
>>
>>b) Have a 256/512 and 1k cluster too..
>>
>>This would allow copy by reference no matter what size si
>>being sent...
> 
> 
> IMO it's quite a waste of memory the way we have thigns now, though
> w/ TSO it'll change things...

Receive path != send path.

> w/ 512 byte mbuf and a 2k cluster just to store just 1514 bytes of data,
> that's only 60% effeciency wrt to memory usage...  so, we currently
> waste 40% of memory allocated to mbufs+clusters...  Even reducing
> mbufs back to 128 or 256 would be a big help, though IPSEC I believe
> would have issues...

mbufs are 256 bytes.

> Hmmm.. If we switched clusters to 1536 bytes in size, we'd be able to
> fit 8 in 12k (though I guess for 8k page boxes we'd do 16 in 24k)...  The
> only issue w/ that would be that a few of the clusters would possibly
> split page boundaries...  How much this would effect performance would
> be an interesting question to answer...

Splitting page boundaries is not an option as it may not be physically
contigous.

Just don't overengineer the stuff.  Mbufs are only used temporarily and
a bit theoretical waste is not much a problem (so far at least).

-- 
Andre