[CRIU] [PATCH] cg: don't leak properties when cg dir exists

Andrew Vagin avagin at parallels.com
Thu Oct 2 08:35:30 PDT 2014 

On Thu, Oct 02, 2014 at 10:16:48AM -0500, Tycho Andersen wrote:
> On Thu, Oct 02, 2014 at 07:09:32PM +0400, Andrew Vagin wrote:
> > On Thu, Oct 02, 2014 at 09:58:23AM -0500, Tycho Andersen wrote:
> > > On Thu, Oct 02, 2014 at 06:48:05PM +0400, Andrew Vagin wrote:
> > > > On Thu, Oct 02, 2014 at 06:00:17PM +0400, Andrew Vagin wrote:
> > > > > On Wed, Oct 01, 2014 at 08:57:28PM +0000, Tycho Andersen wrote:
> > > > > > We're using NULL as a sentinel here to indicate that we shouldn't restore any
> > > > > > cgroup properties. We should make sure that we don't leak this information and
> > > > > > instead check the n_properties field, which we should also set correctly.
> > > > > 
> > > > > Tycho, I beg you to run tests before sending patches.
> > > > > 
> > > > > It's like brushing teeth before bedtime.
> > > > 
> > > > I found that test passes if it's executed in a non-root cpuset group.
> > > > mkdir /sys/fs/cgroup/cpuset/test
> > > > echo 0 >>  /sys/fs/cgroup/cpuset/test/cpuset.mems
> > > > echo 0-3 >>  /sys/fs/cgroup/cpuset/test/cpuset.cpus
> > > > cho $$ > /sys/fs/cgroup/cpuset/test/tasks
> > > > 
> > > > [root at avagin-fc19-cr criu]# bash test/zdtm.sh static/cgroup00
> > > > ================================= CRIU CHECK =================================
> > > > Looks good.
> > > > Execute zdtm/live/static/cgroup00
> > > > ./cgroup00 --pidfile=cgroup00.pid --outfile=cgroup00.out --dirname=cgroup00.test
> > > > Dump 2708
> > > > Executing pre-restore hook
> > > > Cleaning cgclean.j8Mio8
> > > > Left there is:
> > > > cgroup.clone_children  cgroup.procs  cgroup.sane_behavior  notify_on_release  release_agent  tasks
> > > > Restore
> > > > Check results 2708
> > > > Waiting...
> > > > Executing cleanup hook
> > > > Cleaning cgclean.v4Pew1
> > > > Left there is:
> > > > cgroup.clone_children  cgroup.procs  cgroup.sane_behavior  notify_on_release  release_agent  tasks
> > > > 18:40:35.992:  2708: CMP [name=zdtmtst:/subcg00
> > > > ] vs [name=zdtmtst:/subcg00
> > > > ]
> > > > 18:40:35.992:  2710: CMP [name=zdtmtst:/subcg00/subsubcg
> > > > ] vs [name=zdtmtst:/subcg00/subsubcg
> > > > ]
> > > > 18:40:35.993:  2709: CMP [name=zdtmtst:/subcg00
> > > > ] vs [name=zdtmtst:/subcg00
> > > > ]
> > > > 18:40:35.993:  2708: PASS
> > > > Test: zdtm/live/static/cgroup00, Result: PASS
> > > > ZDTM tests PASS.
> > > > 
> > > > So I have a question. For which group does criu dump and restore
> > > > parameters? It's obvious, the in our case the group doesn't belong to
> > > > the test process.
> > > 
> > > I'm not sure what you mean by this. You put $$ in the cgroup, which
> > > means that the test process will be in the cgroup and thus criu should
> > > dump it.
> > 
> > It isn't obvious for me. Let's imagine the we create a container. We
> > create a new set of cgroups and set parameters for these groups. Then we
> > executing the init process, which forks other processes. But nobody of
> > them can't change parameters for the initial set of cgroups. Is it true?
> > 
> > Now we are going to restore the container from images. In this case we
> > create a ndew set of cgroups,
> 
> Only if we use --cgroup-root to rewrite them; if not, it should
> restore to the same set of cgroups that existed before. Except that
> they may not exist (e.g. if you rebooted between dump/restore, and
> didn't run any lxc commands that created cpuset:/lxc).

I agree that criu should restore configuration for groups, which have
not existed before. Why should criu restore parameters for grouts, which
were not created by criu?

> 
> > configure them and execute "criu restore".
> > And I doesn't expect that criu will change parameters for the initial
> > set of cgroup. Could you explain why it should?
> 
> If you're using --cgroup-root, it should certainly not touch the
> original ones. However, if you're not, it has to re-create them
> because of the case I described above.
> 
> Tycho
> 
> > 
> > > 
> > > > Why do we need to dump parameters of groups, where
> > > > is a root process lives? Should we skip the first level of groups and
> > > > dump parameters of subgroups only?
> > > 
> > > I think so at least for cpuset. The root cpuset.cpus and cpuset.mems
> > > always have "all" the resources in them, so we don't need to do this
> > > copy for the root. I can work on this and send a patch.
> > > 
> > > Tycho
> > > 
> > > > > 
> > > > > [root at avagin-fc19-cr criu]# bash test/zdtm.sh static/cgroup00
> > > > > ================================= CRIU CHECK =================================
> > > > > Looks good.
> > > > > Execute zdtm/live/static/cgroup00
> > > > > ./cgroup00 --pidfile=cgroup00.pid --outfile=cgroup00.out --dirname=cgroup00.test
> > > > > Dump 1457
> > > > > Executing pre-restore hook
> > > > > Cleaning cgclean.OkAnTX
> > > > > Left there is:
> > > > > cgroup.clone_children  cgroup.procs  cgroup.sane_behavior  notify_on_release  release_agent  tasks
> > > > > Restore
> > > > > Test: zdtm/live/static/cgroup00, Result: FAIL
> > > > > ==================================== ERROR ====================================
> > > > > Test: zdtm/live/static/cgroup00, Namespace: 
> > > > > Dump log   : /root/git/criu/test/dump/static/cgroup00/1457/1/dump.log
> > > > > --------------------------------- grep Error ---------------------------------
> > > > > ------------------------------------- END -------------------------------------
> > > > > Restore log: /root/git/criu/test/dump/static/cgroup00/1457/1/restore.log
> > > > > --------------------------------- grep Error ---------------------------------
> > > > > (00.032275)   1457: Error (cgroup.c:943): cg: copying property cpuset.cpus to cpuset////cpuset.cpus failed
> > > > > (00.032319)   1457: Error (cgroup.c:968): cg: copying prop cpuset.cpus failed
> > > > > (00.032336)   1457: Error (cgroup.c:1005): cg: Couldn't copy special cgroup properties
> > > > > (00.032373) Error (cr-restore.c:1820): Restoring FAILED.
> > > > > (00.039606) Error (cr-restore.c:1177): 1457 exited, status=1
> > > > > ------------------------------------- END -------------------------------------
> > > > > ================================= ERROR OVER =================================
> > > > > 
> > > > > Thanks,
> > > > > Andrey
> > > > > 
> > > > > > 
> > > > > > Signed-off-by: Tycho Andersen <tycho.andersen at canonical.com>
> > > > > > ---
> > > > > >  cgroup.c | 15 +++++++--------
> > > > > >  1 file changed, 7 insertions(+), 8 deletions(-)
> > > > > > 
> > > > > > diff --git a/cgroup.c b/cgroup.c
> > > > > > index 26b0a17..ce49d04 100644
> > > > > > --- a/cgroup.c
> > > > > > +++ b/cgroup.c
> > > > > > @@ -1114,12 +1114,7 @@ static int prepare_cgroup_dir_properties(char *path, int off, CgroupDirEntry **e
> > > > > >  		size_t off2 = off;
> > > > > >  
> > > > > >  		off2 += sprintf(path + off, "/%s", e->dir_name);
> > > > > > -		/*
> > > > > > -		 * Check to see if we made e->properties NULL during restore
> > > > > > -		 * because directory already existed and as such we don't want to
> > > > > > -		 * change its properties
> > > > > > -		 */
> > > > > > -		if (e->properties) {
> > > > > > +		if (e->n_properties > 0) {
> > > > > >  			for (j = 0; j < e->n_properties; ++j) {
> > > > > >  				if (restore_cgroup_prop(e->properties[j], path, off2) < 0)
> > > > > >  					return -1;
> > > > > > @@ -1168,7 +1163,7 @@ static int prepare_cgroup_dirs(char *paux, size_t off, CgroupDirEntry **ents, si
> > > > > >  		/*
> > > > > >  		 * Checking to see if file already exists. If not, create it. If
> > > > > >  		 * it does exist, prevent us from overwriting the properties
> > > > > > -		 * later by setting the CgroupDirEntry's properties pointer to NULL
> > > > > > +		 * later by removing the CgroupDirEntry's properties.
> > > > > >  		 */
> > > > > >  		if (access(paux, F_OK) < 0) {
> > > > > >  			if (errno != ENOENT) {
> > > > > > @@ -1181,7 +1176,11 @@ static int prepare_cgroup_dirs(char *paux, size_t off, CgroupDirEntry **ents, si
> > > > > >  			}
> > > > > >  			pr_info("Created dir %s\n", paux);
> > > > > >  		} else {
> > > > > > -			e->properties = NULL;
> > > > > > +			if (e->n_properties > 0) {
> > > > > > +				xfree(e->properties);
> > > > > > +				e->properties = NULL;
> > > > > > +				e->n_properties = 0;
> > > > > > +			}
> > > > > >  			pr_info("Determined dir %s already existed\n", paux);
> > > > > >  		}
> > > > > >  
> > > > > > -- 
> > > > > > 1.9.1
> > > > > > 
> > > > > > _______________________________________________
> > > > > > CRIU mailing list
> > > > > > CRIU at openvz.org
> > > > > > https://lists.openvz.org/mailman/listinfo/criu
> > > > > _______________________________________________
> > > > > CRIU mailing list
> > > > > CRIU at openvz.org
> > > > > https://lists.openvz.org/mailman/listinfo/criu

More information about the CRIU mailing list