[CRIU] [PATCH v4 2/6] files-reg: Make try_clean_ghost() use struct ghost_file ptr

[Andrew Vagin]

On Sat, Feb 06, 2016 at 05:34:38PM +0300, Kirill Tkhai wrote:
> On 06.02.2016 04:03, Andrew Vagin wrote:
> > Kirill, here is one more issue for you. It was introduced in this patch.
> >
> > [root at fc22-vm criu]# ./test/zdtm.py run -t zdtm/live/static/unlink_fstat00 --fault 2 --report report -f h 
> > === Run 1/1 ================
> > 
> > =================== Run zdtm/live/static/unlink_fstat00 in h ===================
> > Start test
> > ./unlink_fstat00 --pidfile=unlink_fstat00.pid --outfile=unlink_fstat00.out --filename=unlink_fstat00.test
> > Dangling ghost file
> > Run criu dump
> > Forcing 2 fault
> > Running zdtm/live/static/unlink_fstat00.hook(--pre-restore)
> > Run criu restore
> > Forcing 2 fault
> > =[log]=> dump/zdtm/live/static/unlink_fstat00/72/1/restore.log
> > ------------------------ grep Error ------------------------
> > (00.014193)     72: Error (cr-restore.c:1606): BUG at cr-restore.c:1606
> > (00.021403) Error (cr-restore.c:1305): 72 killed by signal 11
> > (00.021431) Error (cr-restore.c:2127): Restoring FAILED.
> > ------------------------ ERROR OVER ------------------------
> > Running zdtm/live/static/unlink_fstat00.hook(--fault, restore)
> > ###### Test zdtm/live/static/unlink_fstat00 FAIL at hook --fault restore #######
> > 
> > ##################################### FAIL #####################################
> 
> ghost_files are changed in other process, and the changes are not visible to parent.

How did it work before you changes?

> 
> This patch is to be able to create ghost files under names with a digit suffix,
> if the standard  "%s.cr.%x.ghost" (rfi->path, rfe->remap_id)  is not available.
> So, it's need a different solution.
> 
> We can't check disk state from parent, because mount points are restored only in
> restore_task_with_children() from child. So, the child should report parent,
> which ghosts names it has changed.
> 
> shmem is not infinity; sending the information about changing of names using
> socket is ugly. Do you know better solution?
>  
> > On Mon, Jan 25, 2016 at 07:39:44PM +0300, Kirill Tkhai wrote:
> >> This cleanup will be used in next patch.
> >>
> >> Signed-off-by: Kirill Tkhai <ktkhai at virtuozzo.com>
> >> ---
> >>  files-reg.c |   15 +++++++--------
> >>  1 file changed, 7 insertions(+), 8 deletions(-)
> >>
> >> diff --git a/files-reg.c b/files-reg.c
> >> index 735a15a..648c8d2 100644
> >> --- a/files-reg.c
> >> +++ b/files-reg.c
> >> @@ -432,17 +432,18 @@ int prepare_remaps(void)
> >>  	return ret;
> >>  }
> >>  
> >> -static void try_clean_ghost(struct remap_info *ri)
> >> +static void try_clean_ghost(struct ghost_file *gf)
> >>  {
> >>  	char path[PATH_MAX];
> >>  	int mnt_id, ret;
> >>  
> >> -	mnt_id = ri->rfi->rfe->mnt_id; /* rirfirfe %) */
> >> +	mnt_id = gf->remap.rmnt_id;
> >>  	ret = rst_get_mnt_root(mnt_id, path, sizeof(path));
> >>  	if (ret < 0)
> >>  		return;
> >>  
> >> -	ghost_path(path + ret, sizeof(path) - 1, ri->rfi, ri->rfe);
> >> +	snprintf(path + ret, PATH_MAX - ret, "/%s", gf->remap.rpath);
> >> +
> >>  	if (!unlink(path)) {
> >>  		pr_info(" `- X [%s] ghost\n", path);
> >>  		return;
> >> @@ -455,7 +456,6 @@ static void try_clean_ghost(struct remap_info *ri)
> >>  	 */
> >>  
> >>  	if ((errno == EISDIR)) {
> >> -		strncpy(path + ret, ri->rfi->path, sizeof(path) - 1);
> >>  		if (!rmdir(path)) {
> >>  			pr_info(" `- Xd [%s] ghost\n", path);
> >>  			return;
> >> @@ -467,7 +467,7 @@ static void try_clean_ghost(struct remap_info *ri)
> >>  
> >>  void try_clean_remaps(int ns_fd)
> >>  {
> >> -	struct remap_info *ri;
> >> +	struct ghost_file *gf;
> >>  	int old_ns = -1;
> >>  	int cwd_fd = -1;
> >>  
> >> @@ -497,9 +497,8 @@ void try_clean_remaps(int ns_fd)
> >>  		}
> >>  	}
> >>  
> >> -	list_for_each_entry(ri, &remaps, list)
> >> -		if (ri->rfe->remap_type == REMAP_TYPE__GHOST)
> >> -			try_clean_ghost(ri);
> >> +	list_for_each_entry(gf, &ghost_files, list)
> >> +		try_clean_ghost(gf);
> >>  
> >>  	if (old_ns >= 0) {
> >>  		if (setns(old_ns, CLONE_NEWNS) < 0)
> >>