Re: [PATCH 00/30] C/R OpenVZ/Virtuozzo style

[Alexey Dobriyan <adobriyan@gmail.com>]

On Tue, Apr 14, 2009 at 01:46:36AM -0400, Oren Laadan wrote:
> Some meta comments about this patch set:
> 
> * Patches 1-9 are cleanups, unrelated to checkpoint/restart. They
> deserve a separate thread.

They will be sent separatedly.

> * You barely take locks or reference counts to objects that you
> later refer to. What if something really bad happens ?

I'm thinking right now what is needed and isn't.

Tasks are frozen so unmap VMA in the middle of dump can't happen.

Refcounts aren't needed  because tasks are frozen but alive and pin
everything needed. This is during dump part.

What is needed is impossibility of a task to dissapear in frozen state.

On restart, I probably leak refcount to every object, but haven't
checked just it.

> * (contd) If you don't take locks, then you at the very least need
> to rely on the container remaining frozen for the duration of the
> operation (during checkpoint).

> * (contd) Still with locks and references, during restart you can't
> even freeze the container, so need extra logic to prevent bad things
> (e.g. OOM killer, signal or ptrace from parent container etc).
> 
> * What is the rationale behind doing the freeze/unfreeze from within
> sys_checkpoint/sys_restart ?   Instead of you let userspace do it
> (and only verify in kernel) you gain, again, flexibility. For example,
> you want to also snapshot the filesystem, then userspace will do
> something like:  freeze container -> snapshot filesystem -> checkpoint
> -> thaw container.

This is incomplete part. But, yes, freeze, dump, thaw/kill as separate
actions make sense.

	checkpoint(CR_CPT_FREEZE);
	[rsync fs]
	checkpoint(CR_CPT_DUMP|CR_CPT_KILL);

with check that CR_CPT_THAW doesn't happen during dump.

> * A plethora of "FIXME" comments ...
> 
> Alexey Dobriyan wrote:
> > This is to show how we see C/R and to provoke discussion on number of
> > important issues (mounts, ...).
> 
> Quoting your patch:
> ---
> This is one big FIXME:
> 	What to do with overmounted files?
> 	What to do with mounts at all, who should restore them?
> 
> just restore something to not oops on task exit

This is just a skeleton for mnt_ns checkpoint, not seriously suggested
to merge it.

> > This is small part of long-awaited to be cleanuped code.
> > 
> > It's able to restore busyloop on i386 and x86_64 and restore i386
> > busyloop on x86_64. It wasn't tested much more than that.
> 
> Oh .. I really wish you'd sent a x86_64 patch our way, too ;)
> 
> > 
> > I'm currently starting formal testsuite, otherwise it's whack-a-mole
> > game and formal TODO list (a huge one).
> > 
> 
> So I'm still struggling to see the major different in the approaches
> that would justify throwing away our hard worked, reviewed, tested
> and functional code, and take this - similar in design, largely
> incomplete and unreviewed code.