[lxc-devel] [PATCH] lxc-checkpoint: add pre-checkpoint

[Tycho Andersen]

On Tue, Jun 30, 2015 at 05:09:40PM +0300, Ruslan Kuprieiev wrote:
> Hi Tycho,
> 
> On 06/30/2015 04:50 PM, Tycho Andersen wrote:
> >Hey Ruslan,
> >
> >On Fri, Jun 26, 2015 at 11:24:32AM +0300, Ruslan Kuprieiev wrote:
> >>Drop this one, please.
> >I'm assuming you're probably going to send another version at some
> >point, a question below.
> >
> >>>diff --git a/src/lxc/lxccontainer.h b/src/lxc/lxccontainer.h
> >>>index d60e19a..1faded2 100644
> >>>--- a/src/lxc/lxccontainer.h
> >>>+++ b/src/lxc/lxccontainer.h
> >>>@@ -773,7 +773,7 @@ struct lxc_container {
> >>>  	 * \return \c true on success, else \c false.
> >>>  	 * present at compile time).
> >>>  	 */
> >>>-	bool (*checkpoint)(struct lxc_container *c, char *directory, bool stop, bool verbose);
> >>>+	bool (*checkpoint)(struct lxc_container *c, char *directory, char *prev_dir, bool stop, bool verbose);
> >Here we're making an ABI change, and I'm not sure what the protocol in
> >LXC for this (Stéphane or Serge can tell us I'm sure :). Whatever the
> >case, we'll have to do some tap dancing here. It may (?) be worth
> >turning this into an argument struct with version information to avoid
> >this in the future, depending on how we solve this.
> 
> Neither am I happy about changing abi in such a way. That's one of the
> reasons
> why I asked to drop this patch for now =).
> 
> But looks like there is no other way but to change abi, as current arguments
> for do_checkpoint do not satisfy our needs, if we want teach lxc-checkpoint
> to do anything more advanced than just plain dump.

Yes, unfortunately I didn't do a good job of future proofing the API.

> Struct for options would be nice, I agree. But I actually thought about
> using
> libcriu's options to set\get options without adding any additional mediator
> structures. Though, It would require modifying libcriu to actually return
> void *
> with options in it as well as adding criu_get-ers and fixing criu_set-ers to
> be
> able to pass options struct as an argument.

I see, and then just porting liblxc to use libcriu and allowing users
to pass in a pre-initialized opaque criu options pointer? That also
seems reasonable to me.

FWIW, I think the only way to avoid an ABI break here would be to
implement ->checkpoint2(), which kind of sucks. I'm not sure how much
work it is to bump the soname, but maybe that could be a temporary
solution. If it's not too hard, then we could do the above now and not
have to deal with it again.

Tycho

> >Anyway, I just thought I'd get the discussion started.
> >
> >Tycho
> >_______________________________________________
> >lxc-devel mailing list
> >lxc-devel at lists.linuxcontainers.org
> >http://lists.linuxcontainers.org/listinfo/lxc-devel
> 
> _______________________________________________
> lxc-devel mailing list
> lxc-devel at lists.linuxcontainers.org
> http://lists.linuxcontainers.org/listinfo/lxc-devel