Message-ID: <20090922152520.GA9154@ovro.caltech.edu>
Date:	Tue, 22 Sep 2009 08:25:20 -0700
From:	"Ira W. Snyder" <iws@...o.caltech.edu>
To:	Avi Kivity <avi@...hat.com>
Cc:	Gregory Haskins <gregory.haskins@...il.com>,
	"Michael S. Tsirkin" <mst@...hat.com>, netdev@...r.kernel.org,
	virtualization@...ts.linux-foundation.org, kvm@...r.kernel.org,
	linux-kernel@...r.kernel.org, mingo@...e.hu, linux-mm@...ck.org,
	akpm@...ux-foundation.org, hpa@...or.com,
	Rusty Russell <rusty@...tcorp.com.au>, s.hetze@...ux-ag.com,
	alacrityvm-devel@...ts.sourceforge.net
Subject: Re: [PATCHv5 3/3] vhost_net: a kernel-level virtio server

On Tue, Sep 22, 2009 at 12:43:36PM +0300, Avi Kivity wrote:
> On 09/22/2009 12:43 AM, Ira W. Snyder wrote:
> >
> >> Sure, virtio-ira and he is on his own to make a bus-model under that, or
> >> virtio-vbus + vbus-ira-connector to use the vbus framework.  Either
> >> model can work, I agree.
> >>
> >>      
> > Yes, I'm having to create my own bus model, a-la lguest, virtio-pci, and
> > virtio-s390. It isn't especially easy. I can steal lots of code from the
> > lguest bus model, but sometimes it is good to generalize, especially
> > after the fourth implemention or so. I think this is what GHaskins tried
> > to do.
> >    
> 
> Yes.  vbus is more finely layered so there is less code duplication.
> 
> The virtio layering was more or less dictated by Xen which doesn't have 
> shared memory (it uses grant references instead).  As a matter of fact 
> lguest, kvm/pci, and kvm/s390 all have shared memory, as you do, so that 
> part is duplicated.  It's probably possible to add a virtio-shmem.ko 
> library that people who do have shared memory can reuse.
> 

Seems like a nice benefit of vbus.

> > I've given it some thought, and I think that running vhost-net (or
> > similar) on the ppc boards, with virtio-net on the x86 crate server will
> > work. The virtio-ring abstraction is almost good enough to work for this
> > situation, but I had to re-invent it to work with my boards.
> >
> > I've exposed a 16K region of memory as PCI BAR1 from my ppc board.
> > Remember that this is the "host" system. I used each 4K block as a
> > "device descriptor" which contains:
> >
> > 1) the type of device, config space, etc. for virtio
> > 2) the "desc" table (virtio memory descriptors, see virtio-ring)
> > 3) the "avail" table (available entries in the desc table)
> >    
> 
> Won't access from x86 be slow to this memory (on the other hand, if you 
> change it to main memory access from ppc will be slow... really depends 
> on how your system is tuned.
> 

Writes across the bus are fast, reads across the bus are slow. These are
just the descriptor tables for memory buffers, not the physical memory
buffers themselves.

These only need to be written by the guest (x86), and read by the host
(ppc). The host never changes the tables, so we can cache a copy in the
guest, for a fast detach_buf() implementation (see virtio-ring, which
I'm copying the design from).

The only accesses are writes across the PCI bus. There is never a need
to do a read (except for slow-path configuration).

> > Parts 2 and 3 are repeated three times, to allow for a maximum of three
> > virtqueues per device. This is good enough for all current drivers.
> >    
> 
> The plan is to switch to multiqueue soon.  Will not affect you if your 
> boards are uniprocessor or small smp.
> 

Everything I have is UP. I don't need extreme performance, either.
40MB/sec is the minimum I need to reach, though I'd like to have some
headroom.

For reference, using the CPU to handle data transfers, I get ~2MB/sec
transfers. Using the DMA engine, I've hit about 60MB/sec with my
"crossed-wires" virtio-net.

> > I've gotten plenty of email about this from lots of interested
> > developers. There are people who would like this kind of system to just
> > work, while having to write just some glue for their device, just like a
> > network driver. I hunch most people have created some proprietary mess
> > that basically works, and left it at that.
> >    
> 
> So long as you keep the system-dependent features hookable or 
> configurable, it should work.
> 
> > So, here is a desperate cry for help. I'd like to make this work, and
> > I'd really like to see it in mainline. I'm trying to give back to the
> > community from which I've taken plenty.
> >    
> 
> Not sure who you're crying for help to.  Once you get this working, post 
> patches.  If the patches are reasonably clean and don't impact 
> performance for the main use case, and if you can show the need, I 
> expect they'll be merged.
> 

In the spirit of "post early and often", I'm making my code available,
that's all. I'm asking anyone interested for some review, before I have
to re-code this for about the fifth time now. I'm trying to avoid
Haskins' situation, where he's invented and debugged a lot of new code,
and then been told to do it completely differently.

Yes, the code I posted is only compile-tested, because quite a lot of
code (kernel and userspace) must be working before anything works at
all. I hate to design the whole thing, then be told that something
fundamental about it is wrong, and have to completely re-write it.

Thanks for the comments,
Ira

> -- 
> error compiling committee.c: too many arguments to function
> 
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists