[RFC v7][PATCH 2/9] General infrastructure for checkpoint restart

Oren Laadan orenl at cs.columbia.edu
Mon Oct 27 14:51:20 PDT 2008



Matt Helsley wrote:
> On Mon, 2008-10-27 at 13:11 -0400, Oren Laadan wrote:
>> Dave Hansen wrote:
>>> On Mon, 2008-10-27 at 07:03 -0400, Oren Laadan wrote:
>>>>> In our implementation, we simply refused to checkpoint setid
>>>> programs.
>>>>
>>>> True. And this works very well for HPC applications.
>>>>
>>>> However, it doesn't work so well for server applications, for
>>>> instance.
>>>>
>>>> Also, you could use file system snapshotting to ensure that the file
>>>> system view does not change, and still face the same issue.
>>>>
>>>> So I'm perfectly ok with deferring this discussion to a later time :)
>>> Oren, is this a good place to stick a process_deny_checkpoint()?  Both
>>> so we refuse to checkpoint, and document this as something that has to
>>> be addressed later?
>> why refuse to checkpoint ?
> 
> 	If most setuid programs hold privileged resources for extended periods
> of time after dropping privileges then it seems like a good idea to
> refuse to checkpoint. Restart of those programs would be quite
> unreliable unless/until we find a nice solution.
> 
>> if I'm root, and I want to checkpoint, and later restart, my sshd server
>> (assuming we support listening sockets) - then why not ?
>> we can just let it be, and have the restart fail (if it isn't root that
>> does the restart); perhaps add something like warn_checkpoint() (similar
>> to deny, but only warns) ?
> 
> 	How will folks not specializing in checkpoint/restart know when to use
> this as opposed to deny?
> 
> 	Instead, how about a flag to sys_checkpoint() -- DO_RISKY_CHECKPOINT --
> which checkpoints despite !may_checkpoint?

I also agree with Matt - so we have a quorum :)

so just to clarify: sys_checkpoint() is to fail (with what error ?) if the
deny-checkpoint test fails.

however, if the user is risky, she can specify CR_CHECKPOINT_RISKY to force
an attempt to checkpoint as is.

does this sound right ?

Oren.


More information about the Containers mailing list