Re: [RFC v6][PATCH 0/9] Kernel based checkpoint/restart

[Oren Laadan]

Cedric Le Goater wrote:
>>> the self checkpoint and self restore syscalls, like Oren is proposing, are 
>>> simpler but they require the process cooperation to be triggered. we could
>>> image doing that in a special signal handler which would allow us to jump
>>> in the right task context. 
>> This description is not accurate:
>>
>> For checkpoint, both implementations use an "external" task to read the state
>> from other tasks. (In my implementation that "other" task can be self).
> 
> which is good, since some applications want to checkpoint themselves and that's
> a way to provide them a generic service.
>  
>> For restart, both implementation expect the restarting process to restore its
>> own state. They differ in that Andrew's patchset also creates that process
>> while mine (at the moment) relies on the existing (self) task.
> 
> hmm, 
> 
> It seems that your patchset relies on the fact that the tasks are checkpointed 
> and restarted at a syscall boundary. right ? I'm might be completely wrong
> on that :)
> 

Yes. I believe openvz too. And probably everyone else as well. I don't
know of a sane way to do it otherwise :o

To be precise, either syscall boundary, or the task was in user space
before being frozen, or (not yet in this implementation) in some special
frozen state like ptrace or vfork.

>> In other words, none of them will require any cooperation on part of the
>> checkpointed tasks, and both will require cooperation on part of the restarting
>> tasks (the latter is easy since we create and fully control these tasks).
> 
> yes.
> 
>>> I don't have any preference but looking at the code of the different patchsets
>>> there are some tricky areas and I'm wondering which path is easier, safer, 
>>> and portable. 
>> I am thinking which path is preferred: create the processes in kernel space
>> (like Andrew's patch does) or in user space (like Zap does). In the mini-summit
>> we agreed in favor of kernel space, but I can still see arguments why user space
>> may be better.
> 
> I'm more familiar with the second algorithm, restarting the process tree in
> user space and let each task restart itself with the sys_restart syscall. But
> that's because I've been working on a C/R framework which freezes tasks on 
> a syscall boundary, which makes a developer's life easy for restart. 
> 
> But as you know, a restarted process resumes its execution where it was 
> checkpointed. So i'm wondering what are the hidden issues with a in-kernel 
> checkpoint and in-kernel restart. To be more precise, why Andrey needs a 
> i386_ret_from_resume  trampoline in : 
> 
> 	http://lkml.org/lkml/2008/9/3/181
> 
> and why don't you ? 
> 

Usually fork() at the child task returns to user space. IIRC, he needs
the child task to return in kernel and invoke his function, which
eventually will invoke the equivalent of do_restart().

So the end result is the same: all restarting tasks restore their own
state in their own contexts by eventually calling do_restart(). If you
create  processes in userspace, then they get there via sys_restart(),
if you create them in the kernel, then they get there directly after
the trampoline.

Oren.

>> (note: I refer strictly to the creation of the processes during restart, not 
>>  how their state is restored).
> 
> OK 
> 
>> any thoughts ?
> 
> thanks Oren,
> 
> C.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/