[CRIU] [PATCH] mnt: add support for external shared mounts

[Pavel Emelyanov]

On 04/06/2015 05:48 PM, Tycho Andersen wrote:
> On Mon, Apr 06, 2015 at 11:26:41AM +0300, Pavel Emelyanov wrote:
>> On 04/03/2015 07:20 PM, Tycho Andersen wrote:
>>> Hi Andrey,
>>>
>>> On Fri, Apr 03, 2015 at 07:00:43PM +0300, Andrew Vagin wrote:
>>>> On Thu, Apr 02, 2015 at 08:22:33PM +0000, Tycho Andersen wrote:
>>>>> Add support for external shared mounts via the --enable-external-masters flag.
>>>>> This flag assumes that the mount is visible to criu via /proc/self/mountinfo
>>>>> and that it will be present at the same location on restore. No attempt to dump
>>>>> or restore the content is made; the external master is bind mounted into the
>>>>> mount ns and whatever is there is what the container sees.
>>>>
>>>> I need time to think about the idea. At the first glance it should work.
>>>>
>>>> But we (I and Pavel) are not sure that we understand what you do:).
>>>>
>>>> Could you show /proc/PID/mountinfo for a container? Why can you not use
>>>> external mounts?
>>>
>>> Because criu dies with "couldn't find sharing" before the external
>>> mount map is consulted. I could use ext-mount-map, but we'd still need
>>> some patch to skip this check, and since we can resolve these
>>> automatically anyway it seems like we should just do that instead.
>>
>> Ouch, so the slavery is set up for external mount points? I've missed that, sorry :(
>> But taking into account my new insight :) doesn't the patch falls logically into
>> two pieces -- the first is auto-assignments for ext-mount-map-s and the second --
>> enabling for slave external bind mounts?
> 
> I guess it depends on how you want the implementation to work.

:) I want it to suit all the parties' requirements, though not duplicate
the functionality and API.

> Currently it doesn't use the ext-mount-map piece, although I think it
> could.

I was under impression that the whole effort is to support internal
mountpoints that happened to be slave to something from the outside.
Now it seems that the idea is to support external slave mounts only.

Am I right with this? If yes, then we should definitely re-use the
external mounts code we have.

> I think we'd still need an extra flag in the protobufs to
> indicate that the mount was an external sharing and would not be
> provided by the user from the CLI on restore, so I'm not sure using
> ext-mount-map really buys us much, but I can make those changes if you
> want.

We already have the ext_mount bit in the mnt_entry. If I get it right,
you will need a) sign that the mount point in question is slave and
b) the path to mountpoint on dump (to use it on restore).

While "a" seem to be auto-assigned on restore, the "b" can be achieved
by introducing the "--ext-mount-map auto" mode. In this mode all the
mountpoints that are thought to be "external" would be treated as if
the called has specifies --ext-mount-map $mountpoint:$root. I.e. -- the
root path will be written in the image just as you suggested in the
your patch.

Does the above work to you?

> In any case, I have the set with the changes you and Andrey suggested
> prepared, I will send that a long shortly.

Thanks :)

-- Pavel