lkml.org 
[lkml]   [2008]   [Oct]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 0/9] OpenVZ kernel based checkpointing/restart
Quoting Oren Laadan (orenl@cs.columbia.edu):
>
>
> Serge E. Hallyn wrote:
> > Quoting Daniel Lezcano (dlezcano@fr.ibm.com):
> >> Oren Laadan wrote:
> >>> Daniel Lezcano wrote:
> >>>> Louis Rilling wrote:
> >>>>> On Fri, Oct 17, 2008 at 04:33:03PM -0700, Dave Hansen wrote:
> >>>>>> On Wed, 2008-09-03 at 14:57 +0400, Andrey Mirkin wrote:
> >>>>>>> This patchset introduces kernel based checkpointing/restart as it is
> >>>>>>> implemented in OpenVZ project. This patchset has limited functionality and
> >>>>>>> are able to checkpoint/restart only single process. Recently Oren Laaden
> >>>>>>> sent another kernel based implementation of checkpoint/restart. The main
> >>>>>>> differences between this patchset and Oren's patchset are:
> >>>>>> Hi Andrey,
> >>>>>>
> >>>>>> I'm curious what you want to happen with this patch set. Is there
> >>>>>> something specific in Oren's set that deficient which you need
> >>>>>> implemented? Are there some technical reasons you prefer this code?
> >>>>> To be fair, and since (IIRC) the initial intent was to start with OpenVZ's
> >>>>> approach, shouldn't Oren answer the same questions with respect to Andrey's
> >>>>> patchset?
> >>>>>
> >>>>> I'm afraid that we are forgetting to take the best from both approaches...
> >>>> I agree with Louis.
> >>>>
> >>>> I played with Oren's patchset and tryed to port it on x86_64. I was able
> >>>> to sys_checkpoint/sys_restart but if you remove the restoring of the
> >>>> general registers, the restart still works. I am not an expert on asm,
> >>>> but my hypothesis is when we call sys_checkpoint the registers are saved
> >>>> on the stack by the syscall and when we restore the memory of the
> >>>> process, we restore the stack and the stacked registers are restored
> >>>> when exiting the sys_restart. That make me feel there is an important
> >>>> gap between external checkpoint and internal checkpoint.
> >>> This is a misconception: my patches are not "internal checkpoint". My
> >>> patches are basically "external checkpoint" by design, which *also*
> >>> accommodates self-checkpointing (aka internal). The same holds for the
> >>> restart. The implementation is demonstrated with "self-checkpoint" to
> >>> avoid complicating things at this early stage of proof-of-concept.
> >> Yep, I read your patchset :)
> >>
> >> I just want to clarify what we want to demonstrate with this patchset
> >> for the proof-of-concept ? A self CR does not show what are the
> >> complicate parts of the CR, we are just showing we can dump the memory
> >> from the kernel and do setcontext/getcontext.
> >>
> >> We state at the container mini-summit on an approach:
> >>
> >> 1. Pre-dump
> >> 2. Freeze the container
> >> 3. Dump
> >> 4. Thaw/Kill the container
> >> 5. Post-dump
> >>
> >> We already have the freezer, and we can forget for now pre-dump and
> >> post-dump.
> >>
> >> IMHO, for the proof-of-concept we should do a minimal CR (like you did),
> >> but conforming with these 5 points, but that means we have to do an
> >> external checkpoint.
> >
> > Right, Oren, iiuc you are insisting that 'external checkpoint' and
> > 'multiple task checkpoint' are the same thing. But they aren't.
> > Rather, I think that what we say is 'multiple tasks c/r' is what you say
> > should be done from user-space :)
>
> Then I don't explain myself clearly :)
>
> The only thing I consider doing in user space is the creation of
> the container, the namespaces and the processes.

That I understand.

> I argue that "external checkpoint of a single process" is very few
> lines of code away from "self checkpoint" that is in v7.
>
> I'm not sure how you define "external restart" ? eventually, the

If I ever said external restart, I actually meant external checkpoint.
I understand that a task should call sys_restart() itself.

> processes restart themselves. It is a question of how the processes
> are created to begin with.
>
> >
> > So particularly given that your patchset seems to be in good shape,
> > I'd like to see external checkpoint explicitly supported. Please
> > just call me a dunce if v7 already works for that.
> >
>
> It seems like you want a single process to checkpoint a single (other)
> process, and then a single process to start a single (other) process.

Yup.

> I tried to explicitly avoid dealing with the container (user space ?
> kernel space ?) and with creating new processes (user space ? kernel
> space ?).

And that's the right thing to do. But:

> Nevertheless, it's the _same_ code. All that is needed is to make the

I was under the impression that sys_checkpoint() on some other task's
pid and then restarting with that image would fail right now.

> checkpoint syscall refer to the other task instead of self, and the
> restart should create a container and fork there, then call sys_restart.
>
> I guess instead of repeating this argument over, I will go ahead and
> post a patch on top of v7 to demonstrate this (without a container,

Cool, thanks!

> therefore without preserving the original pid).

Yes, as i believe you said in another email earlier today, we have not
decided about how to restore the pid. Eric continues to argue for
playing games with /proc/sys/kernel/pid_max.

-serge


\
 
 \ /
  Last update: 2008-10-21 03:01    [from the cache]
©2003-2011 Jasper Spaans