Re: [PATCH] libxl: a domain can be dying but not shutdown

[Ian Campbell <Ian.Campbell@citrix.com>]

On Wed, 2014-10-22 at 15:08 +0100, Andrew Cooper wrote:
> On 22/10/14 14:52, David Scott wrote:
> > The shutdown code is only present if the domain is shutdown.
> > If we attempt to extract it from the flags from a dying but not
> > shutdown domain then we get values like '255' which is not a
> > valid LIBXL_SHUTDOWN_REASON_. We should use LIBXL_SHUTDOWN_UNKNOWN
> > in this case.
> >
> > Signed-off-by: David Scott <dave.scott@citrix.com>
> >
> > ---
> >
> > This can be tested by running 2 domUs, and having one map pages from
> > the other. I used a vchan connection, so I had a vchan_server granting
> > pages and a vchan_client mapping them. I made sure the client is
> > never going to unmap the pages (I used 'sleep' in a Mirage kernel but
> > 'xl pause' from outside would probably also work) and then I
> > 'xl destroyed' the server. The server domain ends up stuck in the dying
> > state because the client still has a page mapped. The server domain
> > is not shutdown.
> >
> > According to 'xl list':
> >
> > djs@st20:~/djs55/list$ sudo xl list
> > Name                                        ID   Mem VCPUs      State   Time(s)
> > Domain-0                                     0  5278     6     r-----   10971.2
> > fedora                                      12  2048     1     -b----    5470.0
> > (null)                                      21     0     1     -bp--d      25.1
> > vchan_client                                22   256     1     -b----       0.0
> >
> > and according to my test program which calls libxl_list_domain:
> >
> > domain 0 shutdown = 0 dying = 0 shutdown_reason = -1
> > domain 12 shutdown = 0 dying = 0 shutdown_reason = -1
> > domain 21 shutdown = 0 dying = 1 shutdown_reason = 255
> > domain 22 shutdown = 0 dying = 0 shutdown_reason = -1
> >
> > I believe this also manifests transiently during a normal 'xl destroy'.
> >
> > Cheers,
> > Dave
> > ---
> 
> Hmm.  Xen unconditionally sets the shutdown code in info->flags when
> querying for dominfo.
> 
> d->shutdown_code defaults to -1 (which explains the 255, given the
> mask), but only becomes valid once d->is_shutting_down gets set.
> 
> This equates to XEN_DOMINF_shutdown lower in the flags field, which is
> translated to xlinfo->shutdown just ahead of the context below.
> 
> It is explicitly not valid for a dying domain, as a dying domain can be
> dying for many reasons, few of which include a valid shutdown code.

I'm afraid I can't tell whether this constitutes an implicit ack or a
nack of this patch.

> Therefore, other bits of libxl.h could do with correcting, given the
> buggy changeset 4d70c9c5

Do you know of any specific locations? I don't see any relevant uses of
the word "shutdown" or "reason" in libxl.h.

Did you instead mean this from libxl_types.idl:
    # Valid iff (shutdown||dying).
    #
    # Otherwise set to a value guaranteed not to clash with any valid
    # LIBXL_SHUTDOWN_REASON_* constant.
    ("shutdown_reason", libxl_shutdown_reason),
?

Ian.