Subject: Re: [PATCH] SCSI driver for VMware's virtual HBA - V4.
From: Alok Kataria <akataria@vmware.com>
Reply-To: akataria@vmware.com
To: Anthony Liguori <anthony@codemonkey.ws>
Cc: James Bottomley <James.Bottomley@suse.de>,
       Rolf Eike Beer <eike-kernel@sf-tec.de>,
       Brian King <brking@linux.vnet.ibm.com>,
       "Chetan.Loke@Emulex.Com" <Chetan.Loke@Emulex.Com>,
       "linux-scsi@vger.kernel.org" <linux-scsi@vger.kernel.org>,
       LKML <linux-kernel@vger.kernel.org>,
       Andrew Morton <akpm@linux-foundation.org>,
       "pv-drivers@vmware.com" <pv-drivers@vmware.com>,
       virtualization <virtualization@lists.linux-foundation.org>
In-Reply-To: <4AA8284A.3010908@codemonkey.ws>
References: <1252458903.24914.73.camel@ank32.eng.vmware.com>
	 <4AA8177A.2020606@codemonkey.ws>
	 <1252533093.24633.40.camel@ank32.eng.vmware.com>
	 <4AA8284A.3010908@codemonkey.ws>
Content-Type: text/plain
Organization: VMware INC.
Date: Thu, 10 Sep 2009 16:43:43 -0700
Message-Id: <1252626223.3899.64.camel@ank32.eng.vmware.com>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3853
Lines: 86

Hi Anthony,

On Wed, 2009-09-09 at 15:12 -0700, Anthony Liguori wrote:
> Alok Kataria wrote:
> > I see your point, but the ring logic or the ABI that we use to
> > communicate between the hypervisor and guest is not shared between our
> > storage and network drivers. As a result, I don't see any benefit of
> > separating out this ring handling mechanism, on the contrary it might
> > just add some overhead of translating between various layers for our
> > SCSI driver.
> >   
> 
> But if you separate out the ring logic, it allows the scsi logic to be 
> shared by other paravirtual device drivers.  This is significant and 
> important from a Linux point of view.
> 
> There is almost nothing vmware specific about the vast majority of your 
> code.  If you split out the scsi logic, then you will receive help 
> debugging, adding future features, and improving performance from other 
> folks interested in this.  In the long term, it will make your life 
> much, much easier by making the driver relevant to a wider audience :-)

>From what you are saying, it seems that for that matter any physical
SCSI HBA's driver could be converted to use the virtio interface;
doesn't each and every driver have something like a ring/queue & I/O
register mechanism to talk to the device ?

Also, why would you add overhead for translation layers between APIs or
data structures just for the sake of it ? I guess you would say it helps
code re-usability, but I fail to see how much of a benefit that is. The
vast majority of the 1500 odd lines of the driver are still very
specific and tied to our PCI device and register interface.

I will just like to re-iterate this once again, this driver should be  
treated no different than a hardware SCSI HBA driver, albeit a very  
simple HBA. We export a PCI device as any other physical HBA, also the
driver talks to the device (emulation) through device IO regisers
without any hypercalls.

As for the virt_scsi layer, we can evaluate it whenever it is ready for
upstream and then take a more informed decision whether we should switch
to using it.

> 
> > Having said that, I will like to add that yes if in some future
> > iteration of our paravirtualized drivers, if we decide to share this
> > ring mechanism for our various device drivers this might be an
> > interesting proposition. 
> >   
> 
> I am certainly not the block subsystem maintainer, but I'd hate to see 
> this merged without any attempt at making the code more reusable.  If 
> adding the virtio layering is going to somehow hurt performance, break 
> your ABI, or in some way limit you, then that's something to certainly 
> discuss and would be valid concerns.  That said, I don't think it's a 
> huge change to your current patch and I don't see any obvious problems 
> it would cause.
> 

I will also like to add that, this is just a driver and is isolated from
the rest of the core of the kernel. The driver is not doing anything
improper either and is using the SCSI stack the way that any other SCSI
driver would use it. 
In earlier cases, when there were changes to the core kernel parts (e.g.
VMI, hypervisor_init- the tsc freq part) VMware did work with the
community to come up with generic interfaces. 

In this case though, I don't think the advantages of using the virtio
interfaces are justified as yet. As and when the virt-scsi layer is
implemented we can re-evaluate our design and use that layer instead.

Holding inclusion of pvscsi driver until the development of virt-scsi
interface is completed doesn't sound right to me. 

Thanks,
Alok

> Regards,
> 
> Anthony Liguori

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/