[CRIU] Fix cpuset restore-in-root bug

[Pavel Emelyanov]

On 10/08/2014 06:40 PM, Tycho Andersen wrote:
> On Wed, Oct 08, 2014 at 06:30:58PM +0400, Pavel Emelyanov wrote:
>> On 10/08/2014 06:25 PM, Pavel Emelyanov wrote:
>>> On 10/08/2014 06:16 PM, Tycho Andersen wrote:
>>>> On Wed, Oct 08, 2014 at 06:02:04PM +0400, Pavel Emelyanov wrote:
>>>>> On 10/08/2014 05:57 PM, Pavel Emelyanov wrote:
>>>>>> On 10/08/2014 02:43 AM, Tycho Andersen wrote:
>>>>>>> Hi Pavel,
>>>>>>>
>>>>>>> On Tue, Oct 07, 2014 at 03:58:39PM -0500, Tycho Andersen wrote:
>>>>>>>> On Tue, Oct 07, 2014 at 11:15:34PM +0400, Pavel Emelyanov wrote:
>>>>>>>>> On 10/07/2014 06:16 PM, Tycho Andersen wrote:
>>>>>>>>>> Hi Pavel,
>>>>>>>>>>
>>>>>>>>>> On Tue, Oct 07, 2014 at 12:58:34PM +0400, Pavel Emelyanov wrote:
>>>>>>>>>>> On 10/06/2014 06:03 PM, Tycho Andersen wrote:
>>>>>>>>>>>> Hi all,
>>>>>>>>>>>>
>>>>>>>>>>>> Based on some discussion, we shouldn't try to do any intelligent copying of any
>>>>>>>>>>>> properties anywhere else in the tree. These cpuset properties are special only
>>>>>>>>>>>> in the sense that they need to be restored before the move_in_cgroup call, and
>>>>>>>>>>>> if the parent cgroups have conflicting options (or empty options), it is best
>>>>>>>>>>>> to just let the restore fail.
>>>>>>>>>>>
>>>>>>>>>>> Patches applied, thanks!
>>>>>>>>>>>
>>>>>>>>>>>> Also, I noticed when making this patch that I am getting a lot of:
>>>>>>>>>>>>
>>>>>>>>>>>> (00.092575)    290: Error (util.c:101): Unable to close fd 1016: Bad file descriptor
>>>>>>>>>>>
>>>>>>>>>>> Ouch. Can you show the full log please?
>>>>>>>>>>
>>>>>>>>>> Yes, these are with the current master:
>>>>>>>>>>
>>>>>>>>>> http://paste.ubuntu.com/8514798/
>>>>>>>>>
>>>>>>>>> Hm... I've run cgroup* tests on recent HEAD and no errors in logs.
>>>>>>>>>
>>>>>>>>>> also, 1016 is the cg service fd, so it seems like maybe a recent
>>>>>>>>>> reordering broke something? I can do a bisect if you take a look at
>>>>>>>>>> the log and nothing obvious pops out.
>>>>>>>>>
>>>>>>>>> Yes, please. If you would be able to shed more light on this, that
>>>>>>>>> would be awesome.
>>>>>>>>
>>>>>>>> Hmm. I tried to bisect with a patch that I know I wasn't seeing this
>>>>>>>> problem with, but even that one has this issue now. Something about my
>>>>>>>> environment must have changed, I guess? I will investigate more and
>>>>>>>> keep you posted.
>>>>>>>
>>>>>>> The patch below fixes the issue for me. Does it make sense? I have no
>>>>>>> idea why we/I didn't see this earlier.
>>>>>>>
>>>>>>> Tycho
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> >From 0f3aad8a78aee56ff7794db33cf76c5d5aae5039 Mon Sep 17 00:00:00 2001
>>>>>>> From: Tycho Andersen <tycho.andersen at canonical.com>
>>>>>>> Date: Tue, 7 Oct 2014 17:25:26 -0500
>>>>>>> Subject: [PATCH] restore: don't race when closing cg yard
>>>>>>>
>>>>>>> If CLONE_FILES is set, we only have one copy of the cg yard fd, so we shouldn't
>>>>>>> all close it. Further, we should wait to close it until after the FORKING stage
>>>>>>> is done, since it is needed during that stage.
>>>>>>>
>>>>>>> If CLONE_FILES is not set, each process can safely close their copy of the fd.
>>>>>>>
>>>>>>> Signed-off-by: Tycho Andersen <tycho.andersen at canonical.com>
>>>>>>> ---
>>>>>>>  cr-restore.c | 17 +++++++++--------
>>>>>>>  1 file changed, 9 insertions(+), 8 deletions(-)
>>>>>>>
>>>>>>> diff --git a/cr-restore.c b/cr-restore.c
>>>>>>> index fd35bef..c12d28e 100644
>>>>>>> --- a/cr-restore.c
>>>>>>> +++ b/cr-restore.c
>>>>>>> @@ -1461,14 +1461,6 @@ static int restore_task_with_children(void *_arg)
>>>>>>>  	if (create_children_and_session())
>>>>>>>  		goto err_fini_mnt;
>>>>>>>  
>>>>>>> -	/*
>>>>>>> -	 * This must be done after forking to allow child
>>>>>>> -	 * to get the cgroup fd so it can move into the
>>>>>>> -	 * correct /tasks file if it is in a different cgroup
>>>>>>> -	 * set than its parent
>>>>>>> -	 */
>>>>>>> -	close_service_fd(CGROUP_YARD);
>>>>>>> -
>>>>>>>  	if (restore_task_mnt_ns(current))
>>>>>>>  		goto err_fini_mnt;
>>>>>>>  
>>>>>>> @@ -1480,6 +1472,15 @@ static int restore_task_with_children(void *_arg)
>>>>>>>  	if (restore_finish_stage(CR_STATE_FORKING) < 0)
>>>>>>>  		goto err_fini_mnt;
>>>>>>>  
>>>>>>> +	/*
>>>>>>> +	 * This must be done after forking to allow child
>>>>>>> +	 * to get the cgroup fd so it can move into the
>>>>>>> +	 * correct /tasks file if it is in a different cgroup
>>>>>>> +	 * set than its parent
>>>>>>> +	 */
>>>>>>> +	if (!(ca->clone_flags & CLONE_FILES) || current->parent == NULL)
>>>>>>> +		close_service_fd(CGROUP_YARD);
>>>>>>
>>>>>> I think what really makes sense here is the CLONE_FILES check, because
>>>>>> all children forking happens in create_children_and_session().
>>>>>>
>>>>>> And it looks like I know why this appeared only recently -- the helper
>>>>>> tasks are now created with CLONE_FILES, so we may have a lot of helpers.
>>>>>
>>>>> After a bit more thinking and reading the code. The place where this close
>>>>> is called also matters -- if we close CGYARD before other tasks forked, the
>>>>> latter ones may fail to obtain this service fd.
>>>>
>>>> Yeah, it isn't necessarily clear from the diff, but I moved it after
>>>> the FORKING stage to avoid this problem.
>>>
>>> Yes, and this is also correct. I'm still thinking about the CLONE_FILES check.
>>>
>>>>> And one more notice -- it looks like we have generic service descriptors 
>>>>> closing problem. Helpers inherit files table from parent but close the
>>>>> descriptors.
>>>>
>>>> By this you mean that there are some other bugs like this as well?
>>>
>>> Should be -- other service descriptors are closed just with close_service_fd,
>>> but seem not to emit such warnings -- I've run the session00 test that generates 
>>> helpers and haven't found any errors like that.
>>
>> Ah, of course. All other close_service_fd-s are not called by helpers :)
>> The restore_one_task() switches the task state and only the code run by
>> alive and stopped has these closes.
>>
>> Can you check whether moving the close for cgyard somewhere inside e.g.
>> close_old_fds helps?
> 
> We can't close it in close_old_fds, that happens too early :(

Yes, you're right. I meant the prepare_fds() one -- it closes proc service
fd and tty one. I think the cgyard can be closed even earlier, but this place
is at least about fds :)

Thanks,
Pavel