[CRIU] [PATCH] autofs: show pipe inode in mount options

[Stanislav Kinsburskiy]

08.01.2016 13:58, Ian Kent пишет:
> On Fri, 2016-01-08 at 12:29 +0100, Stanislav Kinsburskiy wrote:
>> 08.01.2016 08:20, Ian Kent пишет:
>>> On Thu, 2016-01-07 at 16:46 +0100, Stanislav Kinsburskiy wrote:
>>>> Good day, gentlemen.
>>>>
>>>> Could you update, what's the status with this patch?
>>>> Without it it's impossible to match process pipe with kernel
>>>> pipe,
>>>> while
>>>> this is "must have" to be able to migrate AutoFS via CRIU.
>>> Right, I did mean to reply to this mail but have been distracted by
>>> family stuff.
>>>
>>> I don't know what CRIU is and people looking at changelog entries
>>> shouldn't need to do a web search to find out.
>>>
>>> Could you change it a little.
>> Fair enough. I'll resend with more descriptive message.
>> But first I would like to clarify to you the problem root and why
>> it's
>> done like this.
>>
>>> I'm also not sure whether to forward this (assuming the description
>>> is
>>> updated a little) to Al or to include it in the series to rename
>>> autofs4 to autofs that I'm hoping to ask be included in linux-next
>>> fairly soon.
>> Here I don't know, what's better. Of course Al can take it as well.
>> But,
>> probably, first would be nice to make sure, that this solution is the
>> best one.
>> Description of the problem is below.
>>
>>> Passing it on to Al will likely interfere with the series coming
>>> from
>>> linux-next so that could be bit of a hassle.
>>>
>>> Another thing I'm wondering about is the order this entry will
>>> appear
>>> at in the options. You order choice is sensible though and autofs
>>> shouldn't have a problem with the inserted option but other
>>> applications might.
>> I should put it at the end, probably?
>>
>>> Finally, and perhaps most importantly, I don't get what your trying
>>> to
>>> do, you also haven't given any clues to that in the patch
>>> dscription.
>>>
>>> IOW how do you expect to use this.
>>>
>>>> 16.12.2015 13:02, Stanislav Kinsburskiy пишет:
>>>>> This is required for CRIU to migrate a mount point, when write
>>>>> end
>>>>> in user
>>>>> space is closed.
>>> Like I said what does this mean.
>>>
>>> autofs doesn't need this when it re-constructs a mount tree from
>>> existing mounts on re-start or after a SIGKILL on the automount
>>> process.
>>>
>>> How is this different and how will it be used?
>>>
>>> The question to be answered here is "is this the best way to do it
>>> and
>>> will it work for the autofs mount types you expect it to"?
>> So, here is a brief description of the problem.
>> To migrate autofs mount, one have to reconstruct control pipe between
>> kernel and autofs master.
>> There are two cases I'm wiling to support:
>> 1) Automount binary (autofs package). This program is very gentle and
>> it
>> doesn't close write end of the pipe after mount.
>> 2) Systemd. This program closes write end of the pipe once the mount
>> is
>> done.
> I must admit I'm having trouble understanding the description.
> Give me a little time with it.
>
> I don't know how systemd works with autofs mounts only that it uses the
> autofs direct mount type.

Systemd closes write end of the pipe after mount.

> autofs uses both indirect and direct mounts and both can have offsets
> (from the kernel POV semantically direct mounts). So there is quite a
> bit to worry about beside the kernel pipe.

It's not about direct or indirects mounts.
It's about process state restore.
With CRIU migration, any task is frozen, then disassembled into pieces 
(dump files), which are used to assemble task exactly in the same state 
in was before dump.
The technology is very complex and uses a lot a different tricky 
techniques to make this possible in userspace to describe all the 
details here.

But below is a bit more information, which, hopefully, will clarify all 
this a little bit more.
One of a process attributed to migrate is "opened files". Pipes also 
belong to this attribute.

To restore a pipe CRIU does the following (a very simplified description):
1) Creates a new pipe.
2) Writes (previously stores in images) its contents via write end.
3) Duplicate pipe descriptors to the fds of the process, which were used 
before dump, if required
4) Send pipe descriptors to other processes, sharing it, via unix socket.
5) Close those pipe descriptors, which are not required (say, this 
process had only read end, while it's child had write end).

Thus in case of restoring and autofs mount of systemd (which, for 
example, closed write end and has read end on fd 40), one have to create 
a pipe (say, appeared with fd 5 and fd 6), fill it with content via fd 
6, duplicate fd 5 into fd 40, call mount with pipe fd 6 and then close fd 6.
This is, yet again, a very simple explanation.

> Anyway, it seems your only concern is the kernel pipe and I wonder why
> you can't just set the mount catatonic (in autofs speak) on save and
> open a new kernel pipe then set the pipefd on the autofs mount on
> restore.

I can't because of a bunch of reasons:
1) It can be migration, thus I don't have autofs mount on destination 
node at all
2) It can be a container, which is stopped after dump (thus mount point 
is destroyed).

> But probably my suggestion is far to simplistic as I get the impression
> you have a process already in a given state which you want to restore.
>
> One thing to keep in mind is that if an autofs mount is not set
> catatonic any access other than the owner process (process group pid)
> will hang unless there is an actual user space process to service the
> callback.
>
> Although I don't know the flow of things that might be important at
> some point.
>
> And if the mount is set catatonic the process needs to set the pipefd
> to take "ownership" which also clears the mount catatonic flag.

The migration is already implemented and sent to CRIU mailing list.
Here is the list, if you are interesting (I use kernel with this patch 
applied):
https://lists.openvz.org/pipermail/criu/2016-January/024749.html

> Anyway, let me think about what you've written for a while.
> Ian

Sure, take your time.