Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 19 Oct 2012 00:03:41 +0200
From:      Andre Oppermann <andre@freebsd.org>
To:        Navdeep Parhar <np@FreeBSD.org>
Cc:        svn-src-head@freebsd.org, svn-src-all@freebsd.org, src-committers@freebsd.org
Subject:   Re: svn commit: r241703 - head/sys/kern
Message-ID:  <50807CBD.8080703@freebsd.org>
In-Reply-To: <50806F6F.60109@FreeBSD.org>
References:  <201210182022.q9IKMHFa016360@svn.freebsd.org> <50806A10.4070703@freebsd.org> <50806F6F.60109@FreeBSD.org>

next in thread | previous in thread | raw e-mail | index | archive | help
On 18.10.2012 23:06, Navdeep Parhar wrote:
> Hello Andre,
>
> A couple of things if you're poking around in this area...

I didn't really mean to dive too deep into COW socket writes.

> On 10/18/12 13:44, Andre Oppermann wrote:
>> On 18.10.2012 22:22, Andre Oppermann wrote:
>>> Author: andre
>>> Date: Thu Oct 18 20:22:17 2012
>>> New Revision: 241703
>>> URL: http://svn.freebsd.org/changeset/base/241703
>>>
>>> Log:
>>>    Remove double-wrapping of #ifdef ZERO_COPY_SOCKETS within
>>>    zero copy specialized sosend_copyin() helper function.
>>
>> Note that I'm not saying zero copy should be used or is even
>> more performant than the optimized m_uiotombuf() function.
>
> Some time back I played around with a modified m_uiotombuf() that was aware of the mbuf_jumbo_16K
> zone (instead of limiting itself to 4K mbufs).  In some cases it performed better than the stock
> m_uiotombuf. I suspect this change would also help drivers that are unable to deal with long gather
> lists when doing TSO.  But my testing wasn't rigorous enough (I was merely playing around), and the
> drivers I work with can mostly cope with whatever the kernel throws at them.  So nothing came out of
> it.

The jumbo 16K zone is special in that the memory is actually allocated
by contigmalloc to get physically contiguous RAM. After some uptime and
heavy use this may become difficult to obtain. Also contigmalloc has to
hunt for it which may cause quite a bit of overhead.

4K mbufs, actually PAGE_SIZE mbufs, are very easily obtainable and fast.

To be honest I'm not really happy about > PAGE_SIZE mbufs.  They were
introduced at a time when DMA engines were more limited and couldn't
do S/G DMA on receive.

So performance with > PAGE_SIZE mbufs may be a little bit better but
when you approach memory fragmentation after some heavy system usage
it sucks up to the point where it fails most of the time.  PAGE_SIZE
mbufs always perform the same with very little deviation.

In an ideal scenario I'd like to see 9K and 16K mbufs go away and
have the RX DMA ring stitch a packet up out of PAGE_SIZE mbufs.

>> Actually there may be some real bit-rot to zero copy sockets.
>> I've just started looking into it.
>
> I have a cxgbe(4)-specific true zero-copy implementation.  The rx side is in head, the tx side works
> only for blocking sockets (the "easy" case) and I haven't checked it in anywhere.  Take a look at
> t4_soreceive_ddp() and m_mbuftouio_ddp() in sys/dev/cxgbe/t4_ddp.c. They're mostly identical to the
> kernel routines they're based on (read: copy-pasted from).  You may find them of some interest if
> you're working in this area and are thinking of adding zero-copy hooks to the socket implementation.

I'm going to have a look at it think about how to generically support
DDP either way with our socket buffer layout.

Actually that may end up as the golden path. Do away with > PAGE_SIZE
mbufs, sink page flipping COW (incorrectly named ZERO_COPY) and use
DDP for those who need utmost performance (as I said only COW aware
applications gain a bit of speed, unaware may end up much worse).

-- 
Andre




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?50807CBD.8080703>