Re: [PATCH] scsi: allow persistent reservations without CAP_SYS_RAWIO

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Tue, 2012-06-12 at 18:54 +0200, Paolo Bonzini wrote:
> Il 12/06/2012 18:24, Paolo Bonzini ha scritto:
> > Il 12/06/2012 18:21, James Bottomley ha scritto:
> >>>> Persistent reservations commands cannot be issued right now without
> >>>> giving CAP_SYS_RAWIO to the process who wishes to send them.  This
> >>>> is a bit heavy-handed, allow these two commands.
> >>
> >> Why is this heavy handed?  If you remove CAP_SYS_RAWIO, any userspace
> >> process can send these, which would allow any user to completely disrupt
> >> a SAN by injecting spurious reservations ... that doesn't look to be
> >> terribly safe for an operating system running in a data centre.
> > 
> > It is heavy-handed because:
> > 
> > 1) there are still other protections such as DAC (both Unix permissions
> > and ACLs) and SELinux; CAP_SYS_RAWIO is effectively the same as root.
> > 
> > 2) if any user could disrupt the SAN by injecting spurious reservations
> > just by having his laptop's root password, that data centre wouldn't be
> > terribly safe to begin with.
> 
> 3) assume that with this patch user X could disrupt the SAN by injecting
> spurious reservations, e.g. forbidding another user from writing some
> data.  Then they could also destroy those same data even without this
> patch, which is just as disrupting.
> 
> This is because you still need write permission to the device to issue
> reservations.  Read permission will only let you use PERSISTENT RESREVE IN.

I don't think you understand how persistent reservations work.

The first thing I'll say is I agree with Alan.  Unless you can justify
why you want to relax permissions I'm not going to do it.

But secondly, the reason we're so up in arms about SCSI-3 PR is that
there's a feature called reservation by transport ID.  This is used to
reserve multipath devices when one of the paths is down.  Effectively it
allows a PR-OUT command to set a reservation on any LUN with access only
to one of them.  It's definitely a hack in the SCSI standard, but it's
not one that can be controlled by a unix like permission model.  Write
access to *any* LUN allows you to reserve *all* luns.

James