[CRIU] [PATCH] sk-unix: set mntinfo

[Christian Brauner]

On Thu, Sep 15, 2016 at 02:07:47PM +0300, Pavel Emelyanov wrote:
> On 09/14/2016 12:13 AM, Christian Brauner wrote:
> > On Tue, Sep 13, 2016 at 10:51:04PM +0200, Christian Brauner wrote:
> >> On Tue, Sep 13, 2016 at 05:53:59PM +0300, Pavel Emelyanov wrote:
> >>> On 09/07/2016 05:29 PM, Christian Brauner wrote:
> >>>> phys_stat_resolve() call mount_resolve_path() which requires that mntinfo_tree
> >>>> in the ns_id struct is initialized. This is a problem we observed with sockets
> >>>> on btrfs volumes:
> >>>
> >>> Sorry for the late response. See my comment inline.
> >>>
> >>>> Signed-off-by: Christian Brauner <christian.brauner at canonical.com>
> >>>> ---
> >>>>  criu/sk-unix.c | 1 +
> >>>>  1 file changed, 1 insertion(+)
> >>>>
> >>>> diff --git a/criu/sk-unix.c b/criu/sk-unix.c
> >>>> index 99f0b08..a199ca4 100644
> >>>> --- a/criu/sk-unix.c
> >>>> +++ b/criu/sk-unix.c
> >>>> @@ -525,6 +525,7 @@ static int unix_process_name(struct unix_sk_desc *d, const struct unix_diag_msg
> >>>>  			ret = -ENOENT;
> >>>>  			goto out;
> >>>>  		}
> >>>> +		ns->mnt.mntinfo_tree = mntinfo;
> >>>
> >>> I guess that's too late for this assignment. Since this is dump time, then
> >>> ns->mnt.mntinfo_tree should have been initialized in collect_mntns() or
> >>> nearby.
> >>
> >> It is in collect_mntinfo() but it seems calling lookup_ns_by_id() in
> >> unix_process_name() does return an ns_id but ns.mnt->mntinfo_tree is not
> >> initialized. Is mntinfo a global that is identical for all ns_ids in the linked
> >> list? If not how should they be initialized?
> > 
> > Scratch that, collect_mntinfo() is per namespace. The question is when to call
> > it. One option is that as soon as you add a new ns_id you call collect_mntinfo()
> > or you call it *first* when you need it, i.e. you check whether mntinfo_tree ==
> > NULL and if so call collect_mntinfo() on the corresponding ns_id.
> 
> Actually collect_mntinfo() is called for every namespace found already. Can you
> trace back why the ns you see isn't yet collect_mntinfo()-ed at the time the
> unix_process_name() is?

This mainly affects cr-check.c and there only collect_mntinfo() is called. I'd
argue that when unix_process_name() is called and calls lookup_ns_by_id() a
namespace is retrieved for which we did not yet collect_mntinfo() in cr-check.c.
So one possibility is to use collect_mnt_namespaces() instead of
collect_mntinfo().