[Stgt-devel] User-mode iSER

[FUJITA Tomonori]

From: Tom Tucker <tom at opengridcomputing.com>
Subject: Re: [Stgt-devel] User-mode iSER
Date: Wed, 02 Aug 2006 10:58:40 -0500

> On Wed, 2006-08-02 at 07:21 +0900, FUJITA Tomonori wrote:
> [...snip...]
> > > 
> > > Also, won't a user-mode approach support virtual target devices more
> > > easily? 
> > 
> > There's no difference. Note that we try to push a small portion of the
> > iSCSI tcp driver into kernel (iSCSI protocol processing). Both
> > approaches perform SCSI protocol processing, I/O in user space.
> > I think that you can easily implement any kind of vitalization, device
> > vitalization (like virtual tape library) and backing vitalization
> > (snapshot, encryption, compression, etc) in user space with both
> > approaches.
> 
> Yes, I think this is true, in the current design, the target device I/O
> is done in user space, so adding a new device type to trunk/usr is just
> as easy. The event that initiates this I/O though comes from the kernel
> over a netlink socket.

We don't use netlink because it turned out that we cannot pass
user-space pointers via netlik due to its philosophy.

The pathch to replace netlink is at:
http://marc.theaimsgroup.com/?l=linux-scsi&m=115413498823780&w=2

The reason why we cannot netlink is at:
http://lkml.org/lkml/2006/4/19/165


> Moving this all to user-space "might" make this perform better and
> simplify the architecture.
>
> Do you think this is true or does it create as many problems as it
> solves?

I'm not sure. That might be true if we think about only target drivers
that can be implemented in user-space (iSCSI tcp for generic NICs,
iSER), but we also need to support target drivers that needs in-kernel
support.

Anyway, as I said in the first mail, we don't know where iSCSI tcp for
generic NICs and iSER drivers live. So just wait for few days.