Re: [RFC][PATCH 1/2] Track in-kernel when we expect checkpoint/restart to work

[Oren Laadan <orenl@cs.columbia.edu>]

Daniel Lezcano wrote:
> Greg Kurz wrote:
>> On Thu, 2008-10-09 at 12:04 -0700, Dave Hansen wrote:
>>> Suggested by Ingo.
>>>
>>> Checkpoint/restart is going to be a long effort to get things working.
>>> We're going to have a lot of things that we know just don't work for
>>> a long time.  That doesn't mean that it will be useless, it just means
>>> that there's some complicated features that we are going to have to
>>> work incrementally to fix.
>>>
>>> This patch introduces a new mechanism to help the checkpoint/restart
>>> developers.  A new function pair: task/process_deny_checkpoint() is
>>> created.  When called, these tell the kernel that we *know* that the
>>> process has performed some activity that will keep it from being
>>> properly checkpointed.
>>>
>>> The 'flag' is an atomic_t for now so that we can have some level
>>> of atomicity and make sure to only warn once.
>>>
>>> For now, this is a one-way trip.  Once a process is no longer
>>> 'may_checkpoint' capable, neither it nor its children ever will be.
>>> This can, of course, be fixed up in the future.  We might want to
>>> reset the flag when a new pid namespace is created, for instance.
>>>
>> Then this patch should be described as:
>>
>> Track in-kernel when we expect checkpoint/restart to fail.
>>
>> By the way, why don't you introduce the reverse operation ?
> 
> I think implementing the reverse operation will be a nightmare, IMHO it 
> is safe to say we deny checkpointing for the process life-cycle either 
> if the created resource was destroyed before we initiate the checkpoint.
> 
> For example, you create a socket, the process becomes uncheckpointable, 
> you close (via sys_close) the socket, you have to track this close to be 
> related to the socket which made the process uncheckpointable in order 
> to make the operation reversible.

I agree that it makes sense to only track transitions in one direction.
Therefore at any given point in time all we'll know is that the process
"may be non-checkpointable", instead of the clear-cut "uncheckpointable"
(webster anyone ?).

The distinction is important, because it may be that the process is,
after all, checkpointable, so users/developers could still try to
perform a checkpoint, should they wish too. The only thing is that
it is not guaranteed to succeed.

In fact, one way to transition back to the "checkpointable" state is
by doing a dry-checkpoint, where no data is saved (/dev/null ?). No
side effects will occur except for a short downtime due to the freeze
period. If the dry-checkpoint completes successfully - we can reset
the non-/un-/not-/a-/dis-checkpointable flag.

> 
> Let's imagine you implement this reverse operation anyway, you have a 
> process which creates a TCP connection, writes data and close the socket 
> (so you are again checkpointable), but in the namespace there is the 
> orphan socket which is not checkpointable yet and you missed this case.
> _______________________________________________
> Containers mailing list
> Containers@lists.linux-foundation.org
> https://lists.linux-foundation.org/mailman/listinfo/containers

Oren.